Econ 4413 Poverty and Income Distribution

Reading Package
For
Poverty and Income Distribution
ECON 4413
Abid A. Burki
Fall Semester 2019-20
Lahore University of Management Sciences

School of Humanities, Social Sciences & Law
Econ 4413 – Poverty and Income Distribution
Fall 2019
Professor Abid A. Burki Phone: 042-35608076, Ext. 8076

Room 1-0104, Old Econ Wing, Academic Block E-mail: burki@lums.edu.pk
Office Hours: Mon & Wed, 3:00 – 5:00 pm. Meetings: Mon & Wed, 11:00 am – 12:50pm
Course Description:
This is a course in the development stream designed to understand the theory and practice of
poverty and income distribution in developing countries like Pakistan. The course is structured to
provide students with hands-on experience of measuring various poverty and inequality methods
using Pakistan’s household survey data. The course will provide a review of quantitative methods
to measure poverty and inequality and interpretation of these measures in the context of policy
research and poverty reduction strategies. The course assumes understanding of basic tools of
econometric theory and their applications. The topics to be covered include understanding patterns
of economic growth, link between growth inequality and poverty, concepts and measurement of
poverty, income standards and inequality, income standards and poverty, interpreting inequality
and poverty measures, and understanding determinants of poverty.
Course Distribution:
This is an elective course, which is open to BSc Seniors, BSc Juniors and MS Economics.
Course Prerequisites:
Intermediate microeconomics; Basic Econometrics, or their equivalents.
IT Lab Sessions:
Students are required to attend lab sessions/tutorials to complete home works.
Grading Breakup and Policies:

The course grade will depend on:
Six Home Works: 25%
Research Project: 25%
Mid-term: 25%
Final Exam: 25%
Recommended Texts & Readings:

Foster, J., S. Seth, M. Lokshin, Z. Sajaia (2013). A Unified Approach to Measuring Poverty and
Inequality: Theory and Practice, the World Bank, Washington, D.C.
Haughton, J., Shahidur R. Khandker (2009). Handbook on Poverty and Inequality, World Bank,
Washington, D.C.
1
Other recommended readings are journal articles or book chapters. Additional readings may be
assigned if necessary.
Sessions and Readings:

1. INTRODUCTION
Why do we care about poverty and inequality; dimensions of poverty and inequality.
Reading:
Burki, Abid A. (2017). Dynamics of income inequality in Pakistan. Keynote Address.
Further Readings:
Sen, A. (1992). Inequality Reexamined, Russell Sage Foundation. Chapter 1: Equality of What?
World Bank (2006). Inequalities within Countries: Individuals and Groups, Chapter 2, World
Development Report: Equity and Development. The World Bank, Washington D.C. 28 – 54.
2. INCOME STANDARDS AND INEQUALITY

Basic concepts (density function; cumulative distribution function; quantile function); income
standards (desirable properties; commonly used income standards; growth curves); inequality
measures (desirable properties); commonly used income inequality measures (Atkinson’s ratio;
Gini coefficient, general Entropy measures, dominance and unanimity). Interpreting inequality and
analysis at the national and regional level and decompositions. Kuznets curve; earnings inequality
among the top 1 percent in industrialized economies.
Readings:
Alvaredo, F., A.B. Atkinson, T. Piketty, E. Saez (2013). The Top 1 Percent in International and
Historical Perspective, Journal of Economic Perspectives, 27(3), 3 – 20.
Inequality: Theory and Practice, the World Bank, Washington, D.C., Chapter 2, 45 – 104, 155
– 224.
Further Reading:
Washington, D.C., Chapter 6, 101 – 120.
Mankiw, N. Gregory (2013). Defending the One Percent, Journal of Economic Perspectives, 27(3), 21
– 34.
Song, J., D.J. Price, F. Guvenen, N. Bloom, Til von Wachter (2019). Firming Up Inequality, Quarterly
Journal of Economics, 134(1), 1 – 50.
3. WEALTH INEQUALITY
Alternative measures for the distribution of wealth; measurement and decomposition of wealth
inequality; using proxies for wealth by constructing linear index from asset ownership indicators;
using principal component analysis to derive weight for wealth inequality and z-score to aggregate.
Readings:
Filmer, D., Lant H. Pritchett (2001). Estimating Wealth Effects without Expenditure Data—or Tears:
An Application to Educational Enrollments in States of India, Demography, 38(1), 115 – 132.
2
Picketty, T., E. Saez (2014). Inequality in the Long Run, Science, 344(6186), 838 – 843.
Further Readings:
Davies, James B., Nocole M. Fortin, T. Lemieux (2017). Wealth Inequality: Theory, Measurement and
Decomposition, Canadian Journal of Economics, 50(5), 1224 – 1261.
McKenzie, David J (2005). Measuring Inequality with Asset Indicators, Journal of Population
Economics, 18(2), 229 – 260.
4. ESTIMATING WAGE INEQUALITY

Human capital model; quantile regressions; does education reduce wage inequality; inequality
among the other 99 percent
Readings:
Autor, David H. (2014). Skills, Education, and the Rise of Earnings Inequality among the “other 99
percent”. Science, 344(6186), 843 – 851.
Martins, Pedro S., Pedro T. Pereira (2004). Does Education Reduce Wage Inequality? Quintile
Regression Analysis from 16 Countries, Labor Economics, 11, 355– 371.
Further Readings:
Autor, David H., Alan M. Christopher, L. Smith (2010). The Contribution of the Minimum Wage to
U.S. Wage Inequality over Three Decades: A Reassessment, NBER Working Paper Series
No.16533, 61 pages.
5. INEQUALITY OF OPPORTUNITY
The concept; operationalizing inequality of opportunity; Human Opportunity Index (HOI); coverage
of basic opportunities; measurement of HOI.
Readings:
Brunori, P., Francisco H.G. Ferreira, V. Peragine (2013). Inequality of opportunity, income inequality
and economic mobility: Some international comparisons, IZA Discussion Papers, No. 7155,
Institute for the Study of Labor (IZA), Bonn.
Paes de Barros, R., J. R. M. Vega, J. Saavedra (2008). Measuring Inequality of Opportunities for
Children, Unpublished, World Bank, Washington, D.C.
Further Readings:
Saavedra-Chanduvi, J., J.R. Molinas, F.H.G Ferreira (2009). Measuring Inequality of Opportunities in
Latin America and the Caribbean, the World Bank, Washington, D.C., Chapter 1-3, 23 - 123.
Shaheen, S., M. S. Awan, A.R. Cheema (2016) Measuring Inequality of Opportunity in Pakistan:
Parametric and Non-Parametric Analysis, Pakistan Economic and Social Review, 54(2), 165 –
190.
3
6. INCOME STANDARDS AND POVERTY
Desirable properties, poverty and income standards, properties of poverty measures; headcount
index; poverty gap index; squared poverty gap; Sen index; The Sen-Shorrocks-Thon index; the
Watts index, advantages and disadvantages of each measure, policy relevance of poverty measures.
Interpreting poverty and analysis at the national level and rural/urban decomposition, analysis at
the subnational level,
Reading:
Inequality: Theory and Practice, the World Bank, Washington, D.C., Chapter 2, 105 – 153, 155
– 224.
Further Readings:
Banerjee, T. Picketty (2005). Top Indian Incomes, 1922 – 2000, The World Bank Economic Review,
19(1), 1 – 20 (2005).
Chapters 1 – 5.
7. UNDERSTANDING DETERMINANTS OF POVERTY

What causes poverty; household & individual level characteristics; analyzing determinants of
poverty with regression and problems; solving estimation problems; logistic regressions?
Readings:
Attanasio, Orazio P., Luigi Pistaferri (2016). Consumption Inequality, Journal of Economic
Perspectives, 30(2), 2 – 28.
Burki, Abid A. (2011). Exploring the Links between Inequality, Polarization and Poverty: Empirical
Evidence from Pakistan. SANEI Working Paper Series No.11/04. 47 pages.
Washington, D.C., Chapter 8 & 14.
Ravallion, M., and G. Datt (2002). Why Has Economic Growth been more Pro-poor in Some States of
India than Others? Journal of Development Economics, 68, 381 – 400.
Further Readings:
Nazli, H., E. Whitney, K. Mahrt (2015). Poverty Trends in Pakistan, WIDER Working Paper
2015/136. United Nations University World Institute for Development Economics Research.
8. MULTIDIMENSIONAL POVERTY
Multidimensional poverty; multidimensional standards; polarization.
Readings:
Inequality: Theory and Practice, the World Bank, Washington, D.C., 224 – 244.
Planning Commission (2016). Multidimensional Poverty in Pakistan, Planning Commission,
Government of Pakistan, Islamabad. Chapter 2 and Chapter 3.
4
The Dynamics of Abid Aman Burki
Income Inequality in Department of Economics

Lahore University of
Management Sciences (LUMS)
Pakistan May 10, 2018

Outline
• Introduction.
• Evolution of top incomes in advanced industrialized countries.
• Taxes and top shares in advanced countries.
• Inequalities of income and consumption in Pakistan.
• Intergenerational mobility in Pakistan.
• Educational gap between the rich and poor quintiles.
• Unfair and regressive taxation.
• Conclusions.
5/9/2018 2
Introduction
1 2 3
Income and wealth No clarity about the The reality is that
inequality is widespread ideal level of inequality. simple narratives do not
all over the world. • This decision has to be made capture the wide variety
• If not properly monitored or by political institutions. of country experiences.
addressed, it can lead to • To initiate this process
economic, political and social requires reliable information
disasters. on income and wealth.
5/9/2018 3
Introduction
4 5 6
Inequality does not follow and Powerful forces alternatively Income and wealth inequalities
deterministic process. pushing in the direction of were very high a century ago.
• Both Marx and Kuznets were wrong. rising or shrinking inequality. It has dropped in first half of
• Which one dominates depends on the 20th century.
policies and institutions.
Income & wealth inequality has
increased with globalization.
5/9/2018 4
The share of total annual
income received by top 1%
has more than doubled from
9% in 1976 to 20% in 2011
Evolution of Kuznets has been

Top Incomes turned upside down.
in the US Decline in top share
from WWII to 60s
Source: Alvaredo, F., A.B. Atkinson, T. Picketty, E. Saez (2013), The Top 1 Percent in International and Historical Perspective,
Journal of Economic Perspectives, 27(3), 3 – 20.
5/9/2018 5
Evolution of
top incomes
in different
countries
Shows a strong asymmetric U-
shape. During 80 – 07 period,
top share rose by 135% in US;
only 76% in Canada and 105% in
Australia
5/9/2018 6
Evolution of
top incomes L-shaped pattern
in Europe than U-shaped
and Japan
5/9/2018 7
Strong correlation between
reduction in top tax rates and
increase in top 1% pre-tax income
Taxes and top

shares
Shows top individual tax rates in four countries since 1900. Tax
rates have followed an Inverse U-shaped time-path. Top tax
rate in France only 10 percentage point lower than in 1950; In
US, it was less than half its 1950 value.
Source: Alvaredo, F., A.B. Atkinson, T. Picketty, E. Saez (2013), The Top 1 Percent in International
and Historical Perspective, Journal of Economic Perspectives, 27(3), 3 – 20.
5/9/2018 8
Bottom line:
institutions and
policies matter
in shaping
inequality
Source: World Inequality Report 2018
5/9/2018 9
Inequalities of income/consumption in
Pakistan
Under-representation of
Inequality under-estimated upper income strata.
Most attention on poverty with HIES data.
alleviation; inequality - Non-response.
ignored. HIES sample too small to
estimate inequality. - Replacement HH not from
upper income strata.
All inter-temporal inequality Lack of data is a major hurdle

No tax records data available
measures are fairly stable in the analysis of income &
at the individual level.
over time. wealth inequality in Pakistan.
5/9/2018 10
Probability Density Functions of Pakistan, 2001/02 and 2015-16
0.0016
2001
0.0014 Median 2001
0.0012 2015
Probability density function
Median 2015
0.001
0.0008
0.0006
0.0004
0.0002
0
0 1 2 3 4 5 6 7 8 9 10
Monthly Per Capita Income, thousands
Note: Author’s calculations based on data from HIES 2001-02 and PSLM-HIES 2015-16
Per capita consumption is in constant 2000-01 prices.
5/9/2018 11
Probability Density Function of Urban Probability Density Function of Rural
2001
2001
Median 2001
0.0012 0.0018
Median 2001
2015
2015 0.0016
0.001 Median 2015

Median 2015 0.0014
0.0008 0.0012
0.001
0.0006
0.0008
0.0004 0.0006
0.0004
0.0002
0.0002
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Monthly per capita income, thousands Monthly per capita income, thousands
Per capita consumption is in constant 2000-01 prices.
5/9/2018 12
Mean and Median Per Capita Income, Growth and the Gini Coefficient
Mean Median Gini coefficient
Region 2001-02 2015-16 Growth 2001-02 2015-16 Growth Change

2001-02 2015-16
(PKR) (PKR) (%) (PKR) (PKR) (%) (%)
(1) (2) (3) (4) (5) (6) (7) (8) (9)

Urban 1233 1157 -6.2 952 980 2.8 35.2 26.7 -8.5
Rural 784 1884 140.3 688 1432 108 24.8 35.2 10.4
Total 915 1411 54.3 744 1114 49.7 30.4 32.6 2.2
Note: Author’s calculations based on data from HIES 2001-02 and PSLM-HIES 2015-16. Gini coefficient has been reported on a scale
from 0 to 100 rather than from 0 to 1.
5/9/2018 13
Lorenz curves of Pakistan, 2001-02 versus 2015-16
2001
1
2015
Equality
0.8
0.6
Lorenz curve
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
Cumulative population proportion
5/9/2018 14
Lorenz curve of urban Pakistan, 2001- Lorenz curve of rural Pakistan, 2001-02
02 vs. 2015-16 vs. 2015-16
2001 2015 Equality 2001 2015 Equality
1
1
0.8
0.8
0.6
0.6
Lorenz curve
Lorenz curve
0.4
0.4
0.2
0.2
0
0 0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1 Cumulative population proportion
5/9/2018 15
Partial Means and Partial Mean Ratios
Lower Partial Mean Upper Partial Mean Partial Mean Ratio
10th percentile 20th percentile 80th percentile 90th percentile 90-10
Region 80-20
(PKR) (PKR) (PKR) (PKR) (Richest vs. poorest)
(1) (2) (3) (4) (5) (6)
HIES 2001-02
90.2 82.9
Urban 364.5 469.1 2744.7 3724.2
(10.2 times) (5.8 times)
79.8 71.3
Rural 340.8 401.3 1696.6 1690.9
85.9 77.7
Total 340.7 412.6 1846.2 2414.9
PSLM-HIES 2015-16
80.4 72.4
Urban 523.1 593.0 2149.4 2674.3
88.0 81.2
Rural 673.7 780.0 4159.2 5601.0
86.0 78.7
Total 552.1 630.9 2963.7 3932.0
Note: Author’s calculations based on HIES 2001-02 and PSLM-HIES 2015-16.

5/9/2018 16
Growth Incidence Curves for Pakistan
Total Consumption growth by Percentiles, 2001/02 – 2015/16
2001/2005 2005/2010 2010/2015
6
5 5.3
Annual growth rate (%)
5.1
4.9
4.6 4.6
4 4.4
4.2 4.3
3.9 4.0 3.9 3.9
3.8 3.7 3.8 3.8 3.7 3.8 3.8 3.8
3.7 3.6
3 3.3 3.5 3.4
3.2 3.1 3.1
3.1 3.0 3.0 3.0 3.0 3.0 3.0 3.0
2.9 2.8 2.8
2.6 2.7 2.7 2.7 2.6
2 2.4 2.3
2.1 2.1 2.1 2.1 2.2 2.1
1.9 2.0
1.8
1.6
1 1.3
0
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95
Consumption groups (Percentiles)
Source: Author’s calculations from HIES 2001-02, 2005-06, 2010-11 and 2015-16.
On the horizontal axis, Pakistan’s population is divided into percentiles and sorted in ascending order based on consumption
level of each group. The vertical axis indicates total consumption growth per household member in each percentile between
different time periods.
5/9/2018 17
Total income growth by percentile, 2001-02 to 2015-16
65
Bottom 50% captured 32.4% of total

60
growth. Top 10% captured
24.3% of total growth
55
Percentage
50
45
40
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Source: Author’s calculations from HIES 2001-02, 2005-06, 2010-11 and 2015-16.
5/9/2018 18
Intergenerational Mobility
• Effects of fathers’ earning on sons earnings is measured by inter-generational earning elasticity.

• Earnings of fathers are significantly correlated with their sons’ earnings (Burki, Memon and Mir,
2015).
• Inter-generational earning elasticity estimates reveal that fathers’ earnings status does predict
the earnings of the sons.
• Every 10% increase in fathers’ earnings leads to 2.9% increase in sons’ earnings
• This relationship is found to be stable over the last 11 years.
• If these trends continue, it would take some of these families up to 7 generations” to transition
to upper income quintiles (Javaid, 2016).
5/9/2018 19
Income quintiles of sons born to bottom- Income quintiles of sons born to bottom-
and top-quintile fathers, 2010-11 and top-quintile fathers, 1994-95
60
Born to bottom quintile father Born to top quintile father
Born to bottom quintile father Born to top quintile father
35
50
30
Percentage
40
25
30
PErcentage
20
15
20
10
10
5
0 0
Bottom 2nd 3rd 4th Top Bottom 2nd 3rd 4th Top
Source: Abid A. Burki, Rashid Memon, Khalid Mir (2015), Multiple Inequalities and Policies to Mitigate Inequality Traps in Pakistan, Oxfam Research Report, Lahore
University of Management Sciences & Oxfam Pakistan, March.
Note: These calculations are based on a sub-sample of PSLM-HIES 2010–11 and HIES 1994-95. Quintiles are constructed separately for fathers’ and sons’ age cohorts.
5/9/2018 20
Earnings quintiles over time
Bottom Quintile Top Quintile
200,000
180,000
160,000
140,000
120,000
RUPEES
100,000
80,000
60,000
40,000
20,000
0
1990-91 1992-93 1993-94 1994-95 1996-97 1997-98 1999-00 2001-02 2003-04 2005-06 2006-07 2007-08 2008-09 2009-10 2010-11
YEARS
Source: Based on authors’ calculations using different rounds of the LFS. Quintiles are constructed for the entire sample in any given year.
5/9/2018 21
Educational gap between rich and poor quintile, 1995 vs. 2011
10
6
Father
5 Son
4
Daughter
Mother
3
0
Bottom 1995 Top 1995 Bottom 2011 Top 2011
5/9/2018 22
Unfair and Regressive Taxation
60% of FBR income comes Pakistan has low tax to GDP Services sector contributes
from indirect taxes; shows ratio; remained below 10% in over 55% in GDP, but only
regressive taxation. past two decades. 30% to taxation.
Lack of tax revenue puts

Inflation is an offensive tax on pressure on budget and leads
consumption; it has strong to inflationary monetary
impact on poverty and policies. It has negative
inequality. impact on the distribution of
real income.
5/9/2018 23
Conclusions
• Income inequality in top 1% of incomes in Pakistan is unknown.

• Income & wealth inequality is under-estimated with HIES data.
• Over the past 15 years, urban incomes have fallen and rural incomes have
increased.
• Increase in overall income inequality; rural inequality has increased, urban
inequality has decreased.
• The gap between 90/10 of rural population has increased (from 4 times to 8
times) from 2001 to 2015.
• This gap between urban population has decreased in the same period.
• Inequality traps seem to persist.
5/9/2018 24
Journal of Economic Perspectives—Volume 27, Number 3—Summer 2013—Pages 3–20
The Top 1 Percent in International and

Historical Perspective†
Facundo Alvaredo, Anthony B. Atkinson, Thomas

Piketty, and Emmanuel Saez
F
or three decades, the debate about rising income inequality in the United
States has centered on the dispersion of wages and the increased premium
for skilled/educated workers, attributed in varying proportions to skill-
biased technological change and to globalization (for example, see Katz and Autor
1999 for a survey). In recent years, however, there has been a growing realization
that most of the action has been at the very top. This has attracted a great deal of
public attention (as witnessed by the number of visits to and press citations of our
World Top Incomes Database at http://topincomes.parisschoolofeconomics.eu/)
and has represented a challenge to the economics profession. Stories based on the
supply and demand for skills are not enough to explain the extreme top tail of
the earnings distribution; nor is it enough to look only at earned incomes. Different
approaches are necessary to explain what has happened in the United States over
the past century and also to explain the differing experience in other high-income
countries over recent decades. We begin with the international comparison in the
first section and then turn to the causes and implications of the evolution of top
income shares.
■ Facundo Alvaredo is Research Fellow at Nuffield College and Department of Economics,

Oxford, United Kingdom, and CONICET (Consejo Nacional de Investigaciones Científicas
y Técnicas), Buenos Aires, Argentina, and Affiliate Member, Paris School of Economics, Paris,
France. Anthony B. Atkinson is Fellow of Nuffield College, Oxford, and Centennial Professor at the
London School of Economics, London, United Kingdom. Thomas Piketty is Professor of Economics
at the Paris School of Economics, Paris, France. Emmanuel Saez is Professor of Economics, Univer-
sity of California at Berkeley, United States. Their email addresses are alvaredo@gmail.com,
tony.atkinson@nuffield.ox.ac.uk, piketty@ens.fr, and saez@econ.berkeley.edu, respectively.
†
To access the disclosure statements, visit
http://dx.doi.org/10.1257/jep.27.3.3 doi=10.1257/jep.27.3.3
4 Journal of Economic Perspectives
Figure 1
Top 1 Percent Income Share in the United States
25%
20%
15%
10%
5%
Top 1% income share excluding capital gains
Top 1% income share including capital gains
0%
19
19
19
19
19
19
19
19
19
19
19
19
19
20
20
13
20
27
34
41
48
55
62
69
76
83
90
97
04
11
Source: Source is Piketty and Saez (2003) and the World Top Incomes Database.
Notes: The figure reports the share of total income earned by top 1 percent families in the United States
from 1913 to 2011. Income is defined as pre-tax market income; it excludes government transfers and
nontaxable fringe benefits. The figure reports series including realized capital gains (solid squares)
and series excluding realized capital gains (hollow squares).
We should start by emphasizing the factual importance of the top 1 percent.

It is tempting to dismiss the study of this group as a passing political fad due to
the slogans of the Occupy movement or as the academic equivalent of reality
TV. But the magnitudes are truly substantial. Based on pre-tax and pre-transfer
market income (excluding nontaxable fringe benefits such as health insurance
but including realized capital gains) per family reported on tax returns, the share
of total annual income received by the top 1 percent has more than doubled from
9 percent in 1976 to 20 percent in 2011 (Piketty and Saez, 2003, and the World
Top Incomes Database). There have been rises for other top shares, but these
have been much smaller: during the same period, the share of the group from
95th to 99th percentile rose only by 3 percentage points. The rise in the share of
the top 1 percent has had a noticeable effect on overall income inequality in the
United States (Atkinson, Piketty, and Saez 2011, Section 2.2).
The United States Top 1 Percent in International Perspective
Figure 1 depicts the US top 1 percent income share since 1913. Simon Kuznets
(1955) famously hypothesized that economic growth would first be accompanied by
a rise in inequality and then by a decline in inequality. At first glance, it is tempting
Facundo Alvaredo, Anthony B. Atkinson, Thomas Piketty, and Emmanuel Saez 5
to conclude from Figure 1 that the Kuznets curve has been turned upside-down. But
this suggestion is too facile. After all, the interwar period did not exhibit a secular
downward trend in shares of top incomes. Apart from the bubble of the late 1920s,
the US top 1 percent share was between 15 and 20 percent throughout this time.
At the time of Pearl Harbor in 1941, the share of the top 1 percent was essentially
the same as in 1918. The downward trend in top shares started at the time of World
War II and continued until the end of the 1960s. There was then a sharp reversal
such that the top share is today back in the same range as in the 1920s. Interest-
ingly, the Great Recession of 2008 –2009 does not seem to have reversed the upward
trend. There was a fall in the top 1 percent share in 2008 –2009 but a rebound in
2010. This would be consistent with the experience of the previous economic down-
turn: top income shares fell in 2001–2002 but quickly recovered and returned to the
previous trend in 2003 –2007. Another piece of evidence that is consistent with this
interpretation is the smaller cyclical variation in the series excluding capital gains
(shown by the hollow squares in Figure 1).
Has the US experience been reproduced in other high-income countries?
The evolution of the shares of the top 1 percent is shown for four Anglo-Saxon
countries in Figure 2A and for France, Germany, Sweden, and Japan in Figure 2B
(it should be noted that the estimates for France and the United Kingdom do
not include capital gains, the estimates for Canada, Germany, Japan, and Sweden
include realized capital gains after the year therein shown, and the estimates
for Australia include them only partially and at varying degrees over time). The
other Anglo-Saxon countries—Australia, Canada, and the United Kingdom—all
show a strong asymmetric U-shape. However, the rises were less marked in two
of these countries. Over the period 1980 to 2007, when the top 1 percent share
rose by some 135 percent in the United States and the United Kingdom, it rose
by some 105 percent in Australia and 76 percent in Canada (and by 39 percent
in New Zealand, not shown). The experience is markedly different in continental
Europe and Japan, where the long pattern of income inequality is much closer to
an L-shaped than a U-shaped curve. (Sweden and other Scandinavian countries
such as Norway (not shown) are intermediate cases.)1 There has been some rise in
recent years in the top shares in these countries, but the top 1 percent shares are
not far today from their levels in the late 1940s, whereas in the United States the
share of the top 1 percent is higher by more than a half.
To us, the fact that high-income countries with similar technological and
productivity developments have gone through different patterns of income
inequality at the very top supports the view that institutional and policy differences
play a key role in these transformations. Purely technological stories based solely
upon supply and demand of skills can hardly explain such diverging patterns. What
is more, within countries, we have to explain not only why top shares rose (in the
U-shaped countries) but also why they fell for a sustained period of time earlier in
1
The Swedish top 1 percent share was very high during World War I. The same is observed in Denmark—
see the discussion in Atkinson and Søgaard (2013).
Figure 2
The Evolution of the Shares of the Top 1 Percent in Different Countries
A: Top 1 Percent Income Shares in English-speaking Countries (U-Shape)

25%
United States—including capital gains
Australia
Canada-including capital gains from 1972
20% United Kingdom—families
United Kingdom—adults
15%
10%
5%
0%
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
20
20
20
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
00
05
10
B: Top 1 Percent Income Shares in Continental Europe and Japan (L-Shape)
25%
France
Germany—including capital gains from 1950
Japan—including capital gains from 1947
20% Sweden—including capital gains
15%
10%
5%
0%
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
19
20
20
20
10
15
20
25
30
35
40
45
50
55
60
65
70
75
80
85
90
95
00
05
10
Source: The World Top Incomes Database.

Notes: The figure reports the share of total income earned by the top 1 percent in four English-speaking
countries in panel A, and in four other OECD countries ( Japan and three continental European
countries) in panel B. Income is defined as pre-tax market income. The estimates for Australia include
realized capital gains partially and at varying degrees over time.
The Top 1 Percent in International and Historical Perspective 7
Figure 3
Top Marginal Income Tax Rates, 1900 – 2011
100%
90%
80%
70%
60%
50%
40%
30%
US UK
20%
France Germany
10%
0%
19
19
19
19
19
19
19
19
19
19
20
20
00
10
20
30
40
50
60
70
80
90
00
10
Source: Piketty and Saez (2013, figure 1).
Notes: The figure depicts the top marginal individual income tax rate in the United States, United
Kingdom, France, and Germany since 1900. The tax rate includes only the top statutory individual
income tax rate applying to ordinary income with no tax preference. State income taxes are not included
in the case of the United States. For France, we include both the progressive individual income tax and
the flat rate tax “Contribution Sociale Generalisée.”
the twentieth century. The most obvious policy difference—between countries and
over time—regards taxation, and it is here that we begin.
Taxes and Top Shares

During the twentieth century, top income tax rates have followed an inverse
U-shaped time-path in many countries, as illustrated in Figure 3. In the United
States, top income tax rates were consistently above 60 percent from 1932 to 1981,
and at the start of the 1920s, they were above 70 percent (of course, varying propor-
tions of taxpayers were subject to the top rate). High income tax rates are not just
a feature of the post-World War II period, and their cumulative effect contributed
to the earlier decline in top income shares. While many countries have cut top
tax rates in recent decades, the depth of these cuts has varied considerably. For
example, the top tax rate in France in 2010 was only 10 percentage points lower
than in 1950, whereas the top tax rate in the US was less than half its 1950 value.
Figure 4 plots the changes in top marginal income tax rates (combining
both central and local government income taxes) since the early 1960s against
the changes over that period in top 1 percent income shares for 18 high-income
countries in the World Top Incomes Database. It shows that there is a strong corre-
lation between the reductions in top tax rates and the increases in top 1 percent
Figure 4
Changes in Top Income Shares and Top Marginal Income Tax Rates since 1960
(combining both central and local government income taxes)
10
US
Change in top 1% income share (percentage points)
Elasticity = .47 (.11)

8
6 UK
Ireland
4 Norway
Portugal Canada
Italy
2 Australia
NZ Spain
Denmark
Japan Sweden
0
France Germany
Finland
Netherlands Switzerland
–40 –30 –20 –10 0 10

Change in top marginal income tax rate (percentage points)
Source: Piketty, Saez, and Stantcheva (2011, revised October 2012, figure 3). Source for top income
shares is the World Top Incomes Database. Source for top income tax rates is OECD and country-
specific sources.
Notes: The figure depicts the change in the top 1 percent income share against the change in the top
income tax rate from 1960– 64 to 2005–2009 for 18 OECD countries. If the country does not have
top income share data for those years, we select the first available five years after 1960 and the most
recent 5 years. For the following five countries, the data start after 1960: Denmark (1980), Ireland (1975),
Italy (1974), Portugal (1976), Spain (1981). For Switzerland, the data end in 1995 (they end in 2005 or
after for all the other countries). Top tax rates include both the central and local government top tax
rates. The correlation between those changes is very strong. The elasticity estimates of the ordinary least
squares regression of 4log(top 1% share) on 4log(1 – MTR) based on the depicted dots is 0.47 (0.11).
pre-tax income shares. For example, the United States experienced a reduction
of 47 percentage points in its top income tax rate and a 10 percentage point
increase in its top 1 percent pre-tax income share. By contrast, countries such
as Germany, Spain, or Switzerland, which did not experience any significant top
rate tax cut, did not show increases in top 1 percent income shares. Hence, the
evolution of top tax rates is strongly negatively correlated with changes in pre-tax
income concentration.
This negative correlation can be explained in a variety of ways. As pointed out
originally by Slemrod (1996), it is possible that the rise in top US income shares
occurred because, when top tax rates declined, those with high incomes had less
reason to seek out tax avoidance strategies. This argument has more recently been
used to deny that any real increase in income concentration actually took place—
that it is a pure statistical artifact. Under this scenario, the real US top income shares
were as high in the 1960s as they are today, but a smaller fraction of top incomes
was reported on tax returns. While this factor may have affected the pattern of the
data at certain times—for example, the jump in top US income shares following
the 1986 Tax Reform Act—closer examination of the US case suggests that the tax
avoidance response cannot account for a significant fraction of the long-run surge
in top incomes. Top income shares based on a broader definition of income that
includes realized capital gains, and hence a major portion of avoidance channels,
have increased virtually as much as top income shares based on a narrower defini-
tion of income subject to the progressive tax schedule (see Figure 1 and Piketty,
Saez, and Stantcheva 2011 for a detailed analysis).
The explanation that changes in tax rates in the top tax brackets do lead to
substantive behavioral change has indeed received some support. After noting that
top US incomes surged following the large top marginal tax rate cuts of the 1980s,
Lindsey (1987) and Feldstein (1995) proposed a standard supply-side story whereby
lower tax rates stimulate economic activity among top earners involving more work,
greater entrepreneurship, and the like. In this scenario, lower top tax rates would
lead to more economic activity by the rich and hence more economic growth.
Behavioral change is at the heart of the optimal income tax analysis pioneered
by Mirrlees (1971) and publicly evoked in the debate about top tax rates in the
UK, where the Chancellor of the Exchequer has argued that reducing the top tax
rate below 50 percent (for broadly the top 1 percent) will not reduce revenue.
The standard optimal tax formula (Diamond and Saez 2011) implies, with an elas-
ticity of taxable income of 0.5, that the revenue-maximizing top tax rate would be
57 percent.2 When allowance is made for other taxes levied in the United Kingdom,
such as the payroll tax, this implies a top income tax rate in the United Kingdom of
some 40 percent (Atkinson 2012).
Richer Models of Pay Determination

The optimal tax literature has, however, remained rooted in an oversimpli-
fied model of pay determination that takes no account of developments in labor
economics, and the same applies to the explanations of changing top income
shares. Changes in the pay of a worker are assumed to have no impact on either
the other side of the labor market or on other workers. The worker generates more
output and pay adjusts by the same amount. Each person is an island. However, in
the now-standard models of job-matching, a job emerges as the result of the costly
creation of a vacancy by the employer and of job search by the employee. A match
2
The revenue-maximizing top tax rate formula takes the form τ = 1/(1 + a · e) where a is the Pareto
parameter of the top tail of the income distribution, and e is the elasticity of pre-tax income with respect
to the net-of-tax rate 1 – τ. With e = 0.5 (as estimated from Figure 4) and a = 1.5 (the current Pareto
parameter of the US income distribution), we get τ = 1/(1 + 0.5 · 1.5) = 57 percent.
creates a positive surplus, and there is Nash bargaining over the division of the
surplus, leading to a proportion β going to the worker and (1 – β ) to the employer.
Typically, β is assumed fixed, but it is possible that what we have observed, at least at
the top, is an increase in β, which can lead to changes in the distribution of income.3
Why should β have increased? The extent to which top earners exercised
bargaining power may have interacted with the changes in the tax system. When
top marginal tax rates were very high, the net reward to a highly paid executive for
bargaining for more compensation was modest. When top marginal tax rates fell,
high earners started bargaining more aggressively to increase their compensation.
In this scenario, cuts in top tax rates can increase top income shares—consistent
with the observed trend in Figure 1—but the increases in top 1 percent incomes
now come at the expense of the remaining 99 percent.
One can also weave this notion of greater incentives for bargaining into a
broader scenario, in which the improved information and communications tech-
nology and globalization were increasing the demand for high-skilled labor, and the
deregulation of finance and of other industries was both raising the demand for skill
at the top and changing the rules under which compensation had been calculated
in the past. In this perspective, high marginal tax rates had served as a brake on the
level of surplus extraction in the past, but then this brake was released at the same
time that economic and institutional conditions allowed for higher compensation
at the top of the income distribution (Piketty, Saez, and Stantcheva 2011).
In this scenario, the higher share of income going to the top 1 percent does
not reflect higher economic growth—which is a key difference with the supply-
side scenario. It is even possible that reductions in top marginal tax rates may
have adverse effects on growth, as may be seen if we go back to the theories of
managerial firms and the separation of ownership and control developed by
Oliver E. Williamson, William Baumol, and Robin Marris in the 1960s and 1970s
(for discussion, see Solow 1971). In these models, managers are concerned with
their remuneration (both monetary and nonmonetary) but also with other dimen-
sions such as the scale or rate of growth of their firms, and allocate their effort
accordingly. Where top tax rates were high, there was a low return to effort spent
on negotiating higher pay. Top corporate executives may have concentrated on
securing alternative sources of utility, such as unproductive corporate expenses,
but they may also have ploughed back profits into securing faster expansion than
in the traditional stock market valuation-maximizing firm. Cuts in top tax rates,
however, meant that top executives switched efforts back to securing a larger share
of the profits, in which case increases in remuneration, or bonuses, may have come
at the expense of employment and growth.
The correlation shown in Figure 4 between top marginal tax rates and changes in
top income shares may of course reflect in part coincidence rather than causality. The
political factors that led to top tax rate cuts —such as those by Reagan and Thatcher
3
Kleven, Landais, Saez, and Schultz (2013) find evidence of such bargaining effects in the pay determi-
nation of high earners, using the Danish preferential tax scheme for highly paid immigrants.
in the 1980s in the United States and the United Kingdom—were accompanied by
other legislative changes, such as deregulation, which may have caused top incomes
to rise, not least on account of the impetus they gave to the growth of the financial
services (Philippon and Reshef 2012) and legal services sectors. More generally, the
effects of taxation may interact with other changes, such as those in remuneration
practices. Where there is a surplus to be shared, the division may reflect relative
bargaining strength, as above, but it may also be influenced by social norms. Notions
of fairness, or a “pay code,” may come into play to remove the indeterminacy where
“individual incentives are not by themselves . . . sufficient to determine a unique
equilibrium” (MacLeod and Malcomson 1998, p. 400). A “pay code” limits the extent
to which earnings are individually determined, a situation that both workers and
employers accept on reputational grounds. As argued in Atkinson (2008), there may
be a tipping-point where there is a switch from a high level of adherence to such
a code to a situation where pay becomes largely individually determined. This has
been documented in the case of the United States by Lemieux, MacLeod, and Parent
(2009), who find an increase in the proportion of performance-pay jobs over the
period 1976 to 1998. As they note, the increased extent of performance-pay may be
a channel by which other factors are expressed in greater wage dispersion, and they
stress the effect at the top end of the wage distribution.
Top Tax Rates and Growth

If we look at the aggregate outcomes, we find no apparent correlation between
cuts in top tax rates and growth rates in real per capita GDP (Piketty, Saez, and
Stantcheva 2011). Countries that made large cuts in top tax rates such as the United
Kingdom or the United States have not grown significantly faster than countries that
did not, such as Germany or Switzerland. This lack of correlation is more consistent
with a story that the response of pre-tax top incomes to top tax rates documented in
Figure 4 is due to increased bargaining power or more individualized pay at the top,
rather than increased productive effort. Naturally, cross-country comparisons are
bound to be fragile; exact results vary with the specification, years, and countries.
However, the regression analysis by Piketty, Saez, and Stantcheva (2011), using the
complete time-series data since 1960, shows that the absence of correlation between
economic growth and top tax rates is quite robust. By and large, the bottom line is
that rich countries have all grown at roughly the same rate over the past 40 years—in
spite of huge variations in tax policies.
More specifically, international evidence shows that current pay levels for chief
executive officers across countries are strongly negatively correlated with top tax
rates even controlling for firm’s characteristics and performance, and that this
correlation is stronger in firms with poor governance (Piketty, Saez, and Stantcheva
2011).4 This finding also suggests that the link between top tax rates and pay of chief
4
Governance is measured with an index that combines various governance measures: insider ownership,
institutional ownership, the ratio of independent board directors, whether the CEO is also chairman of
the board, and the average number of board positions held by board members.
executive officers does not run through firm performance but is likely to be due to
bargaining effects.
Such findings have strong implications for top tax rate policies. The optimal
top tax rate rises dramatically if a substantial fraction of the effect of top tax rates
on pre-tax top incomes documented in Figure 4 above is due to wage-bargaining
effects instead of supply-side effects. Using mid-range parameter values where
the response of top earners to top tax rate cuts is three-fifths due to increased
bargaining behavior and two-fifths due to increased productive work, Piketty, Saez,
and Stantcheva (2011) find that the top tax rate could potentially be set as high as
83 percent—as opposed to 57 percent in the pure supply-side model.5
Capital Income and Inheritance
The analysis just cited focused—like much of the literature— on what is

commonly called “earned incomes,” referring to income received in return for work.
But capital income is also an important part of the story. Of course, the distinction
between the two types of income can become blurry in some cases—notably, entre-
preneurial income can have elements of both compensation for work and a return
to capital investment. Here, we define “capital income” as rents, dividends, interest,
and realized capital gains. The decline of top capital incomes is the main driver of
the falls in top income shares that occurred in many countries early in the twentieth
century. For example, from 1916 to 1939, capital income represented 50 percent of
US top 1 percent incomes, whereas by the end of the century from 1987 to 2010,
the share had fallen to one-third (Piketty and Saez 2003, tables A7 and A8). In the
United Kingdom, the corresponding share fell from 60 percent in 1937 to under
20 percent by the end of the century (Atkinson 2007, figure 4.11). At the same
time, it should be borne in mind that these calculations depend on the definition of
taxable incomes. In times past, a number of income tax systems like those in France
and the United Kingdom included imputed rents of homeowners in the income tax
base, but today imputed rents are typically excluded. Where the tax base has been
extended, this has in some cases taken the form of separate taxation (as with real-
ized capital gains in the United Kingdom), so that this element of capital income
is not covered in the income tax data. As a result of these developments, the share
of capital income that is reportable on income tax returns has often significantly
decreased over time.
Earlier we referred to the cumulative effect of progressive taxation. A long
period of high top rates of income taxation, coupled with high top rates of taxation
on the transmission of wealth by inheritance and gift, reduced the capacity of large
5
With wage-bargaining effects, the optimal top tax rate formula becomes τ = (1 + s · a · e)/(1 + a · e)
where s is the fraction of the total behavioral elasticity due to bargaining effects. With a = 1.5, e = 0.5
(as above), and s = 3/5, we obtain τ = 83 percent. In the standard model with no wage-bargaining
effects, we had s = 0 and τ = 57 percent.
Figure 5
Annual Inheritance Flow as a Fraction of Disposable Income, France 1820 –2008
40%
Economic flow (computed from national wealth estimates,
mortality tables and observed age-wealth profiles)
36%
Fiscal flow (computed from observed bequest and gift tax data,
including tax exempt assets)
32%
28%
24%
20%
16%
12%
8%
4%
0%
18
18
18
18
19
19
19
19
19
20
20
40
60
80
00
20
40
60
80
00
Source: Piketty (2011).
Notes: The annual inheritance flow is defined as the total market value of all assets (tangible and financial
assets, net of financial liabilities) transmitted at death or through inter vivos gifts. Disposable income was
as high as 90–95 percent of national income during the 19th century and early 20th century (when taxes
and transfers were almost nonexistent), while it is now about 70 percent of national income.
wealth-holders to sustain their preeminence. The key factor in determining the

capacity to transmit wealth is the difference between the “internal rate of accumula-
tion” (the savings rate times the rate of return net of taxes) and the rate of growth
of the economy. This means that the taxation of income and wealth transfers can
cause the share of top wealth-holders to fall, as in the United Kingdom over the first
three-quarters of the twentieth century (Atkinson and Harrison 1978), contributing
to the downward trajectory of top income shares. Alongside this was the growth
of “popular wealth” owned by the bottom 99 percent. Back in 1908 in the United
Kingdom, the 17th Earl of Derby had a rent roll of some £100,000, which was more
than 1,000 times the average income at the time. Many of these houses are now
owned by their occupiers.
In recent decades, however, the relation between the internal rate of accumula-
tion of wealth holdings and the rate of growth of capital has now been reversed as a
result of the cuts in capital taxation and the decline in the macroeconomic growth
rate (Piketty 2011). As a result, a number of countries are witnessing a return of
inheritance as a major factor. Figure 5 shows the estimates of Piketty (2011) for
France for the period 1820 to 2008 of the annual inheritance flow (the amount
passed on through bequests and gifts inter vivos), ), expressed as percentage of
disposable income.6 Two methods are employed: a constructive calculation from

national wealth figures, mortality rates, and observed age-wealth profiles, and an
estimate based on the estate and gift tax records. The two methods differ in levels
(the fiscal flows are lower), but the time-paths are very similar.
The inheritance flow in France was relatively stable around 20–25 percent
of disposable income throughout the 1820 –1910 period (with a slight upward
trend), before being divided by a factor of about 5 to 6 between 1910 and the
1950s. Since then, it has been rising regularly, with an acceleration of the trend
during the past 30 years. These truly enormous historical variations bring France
back to a situation similar to that of 100 years ago. An annual inheritance flow
around 20 percent of disposable income is very large. It is typically much larger
than the annual flow of new savings and almost as big as the annual flow of
capital income. This implies that inheritance is again becoming a very important
factor of lifetime economic inequality. As shown in Piketty and Saez (2012), in a
world where inheritance is quantitatively significant, those receiving no bequests
will leave smaller-than-average bequests themselves and hence should support
shifting labor taxation toward bequest taxation. In this situation, inheritance taxa-
tion (and more generally capital taxation, given capital market imperfections)
becomes a powerful and desirable tool for redistribution toward those receiving
no inheritance.
The return of inherited wealth may well differ in magnitude across coun-
tries. The historical series available so far regarding the inheritance flows are
too scarce to reach firm conclusions. Existing estimates suggest that the French
U-shaped pattern also applies to Germany, and to a lesser extent to the United
Kingdom and the United States (Atkinson 2013; Schinke 2012; see Piketty and
Zucman, forthcoming, for a survey). Such variations could be due to differences
in pension systems and the share of private wealth that is annuitized (and there-
fore nontransmissible). From a theoretical perspective, it is unclear however why
there should be much crowding out between lifecycle wealth and transmissible
wealth in an open economy (that is, the fact that individuals save more for their
pension should not make them save less for their children; the extra pension
wealth coming from the lifecycle motive should be invested abroad). It could
be that there are differences in tastes for wealth transmission. Maybe wealthy
individuals in the United Kingdom and in the United States have less taste for
bequest than their French and German counterparts. However it should be kept
in mind that there are important data problems (in particular, wealth surveys tend
to vastly underestimate inheritance receipts), which could partly explain why the
rise of inheritance flows in the recent period appears to be more limited in some
6
It is critical to include both bequests (wealth transmitted at death) and gifts (wealth transmitted inter
vivos) in our definition of inheritance, first because gifts have always represented a large fraction of total
wealth transmission, and second because this fraction has changed a lot over time.
Figure 6
Private Wealth/National Income Ratios, 1870 – 2010
800%
United States
700% Europe
600%
500%
400%
300%
200%
100%
18
18
18
19
19
19
19
19
19
19
19
19
19
20
20
70
80
90
00
10
20
30
40
50
60
70
80
90
00
10
Source: Piketty and Zucman (2013).
Notes: Europe is the (unweighted) average of France, Germany, and the United Kingdom. Private wealth
is defined as the sum of nonfinancial assets, financial assets, minus financial liabilities in the household
and nonprofit sectors.
countries than in others.7 Another source of difference between countries could

come from variations in the total magnitude of wealth accumulation. There may
in this respect be an important difference between the United States and Europe,
as is indeed suggested when we look at total private wealth (expressed as a ratio to
national income), shown in Figure 6 (see Piketty and Zucman, 2013, for a discus-
sion on the differences between private and national wealth).
As may be seen from Figure 6, the twentieth century has seen a U-shaped
time-path in the ratio of private wealth to national income that is more marked in
Europe than in the United States. Private wealth in Europe was around six times
7
In particular, the smaller rise of the UK inheritance flow (as compared to France and Germany) is entirely
due to the much smaller rise of recorded inter vivos gifts, which according to fiscal data barely rose in the
United Kingdom during recent decades, while they have become almost as large as bequests in France and
Germany. This might simply be due to the fact that gifts are not properly recorded by the UK tax adminis-
tration (Atkinson 2013). In the United States, due to the limitations of federal fiscal data on bequests and
gifts, scholars often use retrospective wealth survey data. The problem is that in countries with exhaustive
administrative data on bequests and gifts (such as France, and to some extent Germany), survey-based
self-reported flows appear to be less than 50 percent of fiscal flows. This probably contributes to explaining
the low level of inheritance receipts found in a number of US studies. An example of such a study is Wolff
and Gittleman (2011); one additional bias in this study is that inherited assets are valued using asset prices
at the time these assets were transmitted, and no capital gain or income is included.
national income in 1910, and then fell after the World Wars to less than two and
a half times in 1950. In the past 60 years, it has risen sharply to reach more than
five times national income. This pattern suggests that capital is “back” and that the
low wealth–income ratios observed in Europe from the 1950s to the 1970s were
an anomaly. This can be well accounted for by the long-run wealth accumulation
formula β = s/ /g,, where β is the Harrod–Domar–Solow wealth/income ratio, s is the
saving rate, and g is the growth rate including both real per capita and population
growth. For a given saving rate (say s = 10 percent), you accumulate a lot more
wealth relative to income in the long run when the growth rate is 1.5 to 2 percent
than if the growth rate is 2.5 to 3 percent. Given the large and continuing differ-
ence in population growth rates between Old Europe and the New World, this can
explain not only the long-run changes but also the difference in levels between
Europe and the United States (Piketty 2011; Piketty and Zucman 2013).8
On the other hand, it should be noted that wealth concentration (as opposed to
wealth accumulation) is significantly greater in the United States, where the top
1 percent owns about 35 percent of aggregate wealth (for comparison, the share is
about 20 –25 percent in Europe). So far, existing studies have found that the increase
in US wealth concentration since the 1970s and 1980s has been relatively moderate
in contrast to the huge increase in US income concentration documented above
(Kennickell 2009; Kopczuk and Saez 2004). However, we should be modest about
our ability to measure the trends in top billionaire wealth. With low and diminishing
growth rates and high global returns to capital, the potential for divergence of the
wealth distribution is naturally quite large.
Joint Distribution of Earned and Capital Income

We have discussed earned income and capital income. The last piece of the
puzzle concerns the joint distribution of earned and capital incomes—an aspect that
is rarely given explicit consideration. Yet it is important to know whether the same
people are at the top of both the distribution of capital income and the distribution
of earned income. Suppose that we imagine asking the population first to line up
along one side of a room in increasing order of their earned income and then to
go to the other side of the room and line up in increasing order of their capital
income. How much will they cross over? In the Ricardian class model, the crossing is
complete: the capitalists come at the top in one case and at the bottom in the other.
Has a negative correlation in the nineteenth century been replaced today by a zero
correlation? Or is there a perfect correlation, so that people cross straight over? The
pattern of crossing is given by the copula, which represents the joint distribution
in terms of a function of the ranks in the two distributions of earnings and capital
income. Because the copula compares ranks, it is not affected by whether the distri-
butions themselves are widening or narrowing.
8
In a way, this is equivalent to the explanation based upon lower bequest taste: with higher population
growth and the same bequest taste (per children), the United States should save more. However a significant
part of US population growth historically comes from migration, so this interpretation is not fully accurate.
Table 1
Relation between Top Labor Incomes and Top Capital
Incomes in the United States
Year
1980 2000
A: Percent of top 1% capital incomes in various top labor income groups

Labor income groups:
Top 1% 17% 27%
Top 5% 27% 45%
Top 10% 32% 52%
Top 20% 38% 61%
B: Percent of top 1% labor incomes in various top capital income groups

Capital income groups:
Top 1% 17% 27%
Top 5% 36% 50%
Top 10% 47% 63%
Top 20% 68% 80%
Source: Aaberge, Atkinson, Königs, and Lakner (forthcoming).

Notes: Panel A reports the percent of top 1 percent capital income earners in
various top labor income groups in 1980 (column 1) and 2000 (column 2).
In 2000, 27 percent of top 1 percent capital income earners were also in the
top 1 percent of labor incomes, 45 percent were in the top 5 percent of labor
incomes, etc. Panel B reports the percent of top 1 percent labor income
earners in various top capital income groups in 2000 (column 1) and 1980
(column 2). The computations are based on the public use US tax return
micro-datafiles (see Aaberge et al., forthcoming, for complete details).
What can be learned by considering the copula? Table 1 shows results for the
United States in 2000 and in 1980 based on tax return data analysis from Aaberge,
Atkinson, Königs, and Lakner (forthcoming). Three conclusions may be drawn.
First, the joint distribution is asymmetric. In 2000, of those in the top 1 percent of
capital income, 61 percent were in the top 20 percent of earned income. However,
turning things round, of those in the top 1 percent of earned income, a larger
proportion of 80 percent were in the top 20 percent of capital income. In fact,
63 percent of the top 1 percent of earners were in the top 10 percent of capital
income. Such asymmetry could easily be missed by the use of a measure such as
the correlation coefficient or a parametric form for the copula function. Second,
the degree of association appears strong. Even for capital income, over half of the
top 1 percent find themselves in the top tenth of earners. A quarter are in the top
1 percent for both. Third, the numbers for 1980 are all smaller than their coun-
terparts for 2000. The degree of association increased between 1980 and 2000: in
1980 only 17 percent were in the top 1 percent for both. The proportion of the top
1 percent of earners who were in the top 5 percent of capital income rose from one-
third to one-half, and the reverse proportion rose from 27 to 45 percent.
To understand the changing relationship between earned and capital incomes,

we need to consider the mechanisms that link the two sources. In one direction,
there is the accumulation of wealth out of earned income. Here the opportuni-
ties have changed in Anglo-Saxon countries. A third of a century ago, Kay and
King (1980, p. 59) described the hypothetical position of a senior executive with
a large corporation in the United Kingdom who had saved a quarter of his after-
tax earnings: “[F]eeling . . . that he has been unusually fortunate in his career and
unusually thrifty . . . he may be somewhat surprised to discover that there are in
Britain at least 100,000 people richer than he is.” Today, a chief executive officer
may be both better paid and more able to accumulate. In the other direction, there
is the effect of large family wealth on earnings. In the past, the link may have been
negative, whereas today it may be socially unacceptable to live purely off unearned
income. Wealth/family connections may provide access to high-paying employment
(to assess this, it is necessary to investigate the cross-generation correlation of all
income, not just earnings).
Conclusions
The rise in top income shares in the United States has been dramatic. In seeking
explanations, however, it would be misleading to focus just on the doubling of the
share of income going to the top 1 percent of the US distribution over the past
40 years. We also have to account for the fact that a number of high-income coun-
tries have seen more modest or little increase in top shares. Hence, the explanation
cannot rely solely on forces common to advanced countries, like the impact of new
technologies and globalization on the supply and demand for skills. Moreover, the
explanations have to accommodate the falls in top income shares earlier in the twen-
tieth century that characterize the countries discussed here.
In this paper, we have highlighted four main factors that have contributed to
the growing income shares at the very top of the income distribution, noting that
they may operate to differing extents in the United States and other countries,
particularly in continental Europe. The first is tax policy: top tax rates have moved in
the opposite direction from top pre-tax income shares. The second factor is a richer
view of the labor market, where we have contrasted the standard supply-side model
with the alternative possibility that there may have been changes to bargaining
power and greater individualization of pay. Tax cuts may have led managerial ener-
gies to be diverted to increasing their remuneration at the expense of enterprise
growth and employment. The third factor is capital income. In Europe—but less
so in the United States—private wealth (relative to national income) has followed a
spectacular U-shaped path over time, and inherited wealth may be making a return,
implying that inheritance and capital income taxation will become again central
policy tools for curbing inequality. The final, little-investigated, element is the
correlation between earned income and capital income, which have become more
closely associated in the United States.
■ We are grateful to the journal’s Editor, David Autor, the Managing Editor, Timothy
Taylor, and Coeditors, Chang-Tai Hseih and Ulrike Malmendier, for most helpful comments.
Financial support from the MacArthur Foundation, the Center for Equitable Growth at
UC Berkeley, the Institute for New Economic Thinking at the Oxford Martin School, and the
ESRC-DFID Joint Fund is thankfully acknowledged.
References
Aaberge, Rolf, Anthony B. Atkinson, Sebastian 2013. “The Long-Run History of Income Inequality
Königs, and Christoph Lakner. Forthcoming. in Denmark: Top Incomes from 1870 to 2010.”
“Wages, Capital and Top Incomes.” Unpublished EPRU working paper 2013-01, Economic Policy
paper, not completed. Research Unit, University of Copenhagen.
Alvaredo, Facundo, Anthony B. Atkinson, Diamond, Peter, and Emmanuel Saez. 2011
Thomas Piketty, and Emmanuel Saez. 2011. The “The Case for a Progressive Tax: From Basic
World Top Incomes Database, online at http:// Research to Policy Recommendations.” Journal of
topincomes.g-mond.parisschoolofeconomics.eu/, Economic Perspectives 25(4): 165 –90.
May 15, 2013. Feldstein, Martin. 1995. “The Effect of Marginal
Atkinson, Anthony B. 2007. “The Distribution of Tax Rates on Taxable Income: A Panel Study of the
Top Incomes in the United Kingdom 1908–2000.” 1986 Tax Reform Act.” Journal of Political Economy
In Top Incomes over the Twentieth Century—A Contrast 103(3): 551–72.
between Continental European and English-Speaking Katz, Lawrence F., and David H. Autor. 1999.
Countries, edited by A. B. Atkinson and T. Piketty, “Changes in the Wage Structure and Earnings
82 –140. Oxford University Press. Inequality.” Chap. 26 in Handbook of Labor
Atkinson, Anthony B. 2008. The Changing Economics, edited by O. C. Ashenfelter and D. Card,
Distribution of Earnings in OECD Countries. Oxford Vol. 3A. Amsterdam: North-Holland.
University Press. Kay, John A., and Mervyn A. King. 1980. The
Atkinson, Anthony B. 2012. “The Mirrlees British Tax System. Oxford University Press.
Review and the State of Public Economics.” Journal Kennickell, Arthur B. 2009. “Ponds and
of Economic Literature 50(3): 770 – 80. Streams: Wealth and Income in the U.S., 1989 to
Atkinson, Anthony B. 2013. “Wealth and 2007.” Finance and Economics Discussion Series
Inheritance in Britain from 1896 to the Present.” 2009 -13, Federal Reserve Board, Washington, DC.
Unpublished paper. Kleven, Henrik Jacobsen, Camille Landais,
Atkinson, Anthony B., and Allan J. Harrison. Emmanuel Saez, and Esben Schultz. 2013. “Migra-
1978. Distribution of Personal Wealth in Britain. tion and Wage Effects of Taxing Top Earners:
Cambridge University Press. Evidence from the Foreigners’ Tax Scheme in
Atkinson, Anthony B., and Thomas Piketty. Denmark.” NBER Working Paper 18885.
2007. Top Incomes over the Twentieth Century—A Kopczuk, Wojciech, and Emmanuel Saez.
Contrast between Continental European and English- 2004. “Top Wealth Shares in the United States,
Speaking Countries. Oxford University Press. 1916 –2000: Evidence from Estate Tax Returns.”
Atkinson, Anthony B., and Thomas Piketty. National Tax Journal 57(2, part 2): 445 – 88.
2010. Top Incomes: A Global Perspective. Oxford Kuznets, Simon. 1955. “Economic Growth and
University Press. Income Inequality.” American Economic Review
Atkinson, Anthony B., Thomas Piketty, and 45(1): 1–28.
Emmanuel Saez. 2011. “Top Incomes in the Long Lemieux, Thomas, W. Bentley MacLeod, and
Run of History.” Journal of Economic Literature Daniel Parent. 2009. “Performance Pay and Wage
49(1): 3 –71. Inequality.” Quarterly Journal of Economics 124(1):
Atkinson, Anthony B., and Jakob E. Søgaard. 1– 49.
Lindsey, Lawrence B. 1987. “Individual Taxpayer Piketty, Thomas, Emmanuel Saez, and Stefanie
Response to Tax Cuts: 1982–84: With Implications Stantcheva. 2011. “Optimal Taxation of Top Labor
for the Revenue Maximizing Tax Rate.” Journal of Incomes: A Tale of Three Elasticities.” NBER
Public Economics 33(2): 173 –206. Working Paper 17616. (Forthcoming in American
MacLeod, W. Bentley, and Malcomson, James Economic Journal: Economic Policy.)
M. 1998. “Motivation and Markets.” American Piketty, Thomas, and Gabriel Zucman. 2013.
Economic Review 88(3): 388 – 411. “Capital is Back: Wealth-Income Ratios in Rich
Mirrlees, James A. 1971. “An Exploration in the Countries, 1700–2010.” http://piketty.pse.ens.fr
Theory of Optimum Income Taxation.” Review of /files/PikettyZucman2013WP.pdf.
Economic Studies 38(2): 175 –208. Piketty, Thomas, and Gabriel Zucman.
Philippon, Thomas, and Ariell Reshef. 2012. Forthcoming. “Wealth and Inheritance in the
“Wages and Human Capital in the U.S. Finance Long Run.” In Handbook of Income Distribution,
Industry: 1909 –2006.” Quarterly Journal of Economics Vol. 2, edited by A. Atkinson and F. Bourguignon.
127(4): 1551–1609. Elsevier-North Holland.
Piketty, Thomas. 2011. “On the Long-run Schinke, Christoph. 2012. “Inheritance in
Evolution of Inheritance: France 1820–2050.” Germany 1911 to 2009: A Mortality Multiplier
Quarterly Journal of Economics 126(3): 1071–1131. Approach.” DIW-Berlin SOEP Paper 462.
Piketty, Thomas, and Emmanuel Saez. Slemrod, Joel. 1996. “High Income Families
2003. “Income Inequality in the United States, and the Tax Changes of the 1980s: The Anatomy
1913–1998.” Quarterly Journal of Economics 118(1): of Behavioral Response.” In Empirical Foundations
1–39. (Series updated to 2011 in January 2013; of Household Taxation, edited by Martin Feldstein
available at Saez’s website: http://elsa.berkeley.edu and James M. Poterba, 169 –92. University of
/~saez/.) Chicago.
Piketty, Thomas, and Emmanuel Saez. 2012. Solow, Robert. 1971. “Some Implications of
“A Theory of Optimal Inheritance Taxation.” Alternative Criteria for the Firm.” In The Corporate
CEPR Discussion Paper 9241. (Forthcoming in Economy, edited by R. Marris and A. Wood, 318 – 42.
Econometrica.) Macmillan.
Piketty, Thomas, and Emmanuel Saez. 2013. Wolff, Edward N., and Maury Gittleman. 2011.
“Optimal Labor Income Taxation.” Chap. 9 in “Inheritances and the Distribution of Wealth or
Handbook of Public Economics, Vol. 5, edited by A. Whatever Happened to the Great Inheritance
Auerbach, R. Chetty, M. Feldstein, and E. Saez. Boom?” BLS Working Paper 445, US Bureau of
Elsevier-North Holland. Labor Statistics.
Chapter 2
Income Standards, Inequality,

and Poverty
This chapter complements the introductory chapter by providing a detailed

discussion and more formal analysis of the concepts involved in measuring
income standards, inequality, and poverty. This chapter follows closely the
introduction’s organization. It is divided into four sections. The first sec-
tion introduces notations and basic concepts that will be used throughout
the rest of this chapter. The second and third sections discuss tools and
instruments related to income standards and inequality measures. The
fourth section uses the tools from the second and third sections to construct
poverty measures.
According to Sen’s seminal work (1976a), evaluating poverty within a
society (which may be a country or other geographic region) involves two
steps:
1. Identification, in which individuals are identified as poor or nonpoor

2. Aggregation, in which data about the poor are combined to evaluate
poverty within the society.
However, to identify individuals as poor or nonpoor, we need to select a

space on which their welfare level is to be assessed. The welfare indicator is the
variable for assessing an individual’s welfare level. Thus, evaluating poverty
within a society involves three steps:
45
A Unified Approach to Measuring Poverty and Inequality
1. Space selection, which is described below

2. Identification, in which individuals with welfare levels below the
threshold are classified as poor and individuals with welfare levels
above the threshold are classified as nonpoor
3. Aggregation, our focus, which requires choosing an appropriate aggre-
gation method to measure the poverty level in a society.
In this book, we define the space for evaluating poverty as money metric
and single dimensional. The welfare indicator is either consumption expen-
diture or income:
• An individual’s consumption is the destruction of goods and services

through use by that individual. Consumption expenditure is the overall
consumption of goods and services valued at current prices, regardless
of whether an actual transaction has taken place.
• An individual’s income, in contrast, is the maximum possible expen-
diture the individual is able to spend on consumption of goods and
services, without depleting the assets held.
Whether it is income or consumption expenditure, welfare indica-

tors are constructed by aggregating various components. For example, an
individual’s consumption expenditure is constructed by aggregating the
commodities and services consumed by the individual using the prices paid.
Consumption expenditure as a welfare indicator is more commonly used for
assessing developing countries in Asia and Africa (Deaton and Zaidi 2002).
In contrast, using income as a welfare indicator is common when assessing
Latin American countries.
Although both income and consumption expenditure are used as wel-
fare indicators, consumption expenditure has certain advantages. Income
data, for example, may not lead to an accurate assessment of welfare when
incomes fluctuate significantly. Furthermore, in developing countries,
income data may be difficult to collect, and data accuracy is difficult to ver-
ify because most of the population may be employed in the informal sector.
To work around these problems, many developing countries collect
consumer expenditure survey data, which include detailed information
on goods and services consumed by individuals. Then they use the market
prices to compute the overall consumption expenditure. The surveys ask
about food consumption for several items over a specific reference period,
46
Chapter 2: Income Standards, Inequality, and Poverty
which may be a month or any longer period of time. If the reference period
is short (for example, one month), seasonality concerns may be overcome,
but a shorter reference period may also lead to more noise in the expenditure
data. Noise can be avoided by using a longer reference period, but difficulties
in recollection may bias expenditures downward.1
A person may consume many private and public goods from the long
list of commodities in a consumer expenditure survey. For a private good,
total expenditure is the amount of commodity consumed times that com-
modity’s price. Consumption expenditure for two individuals having the
same consumption patterns and requirements, therefore, should be twice the
consumption expenditure for either of the two.
This straightforward expenditure computation may not be possible when
the consumed commodities are, instead, public goods. Given that public
goods are nonrival and nonexcludable, the same amount of public goods
may be consumed by multiple individuals without additional cost. Multiple
individuals living together and sharing public goods enjoy economies of scale.
Examples of public goods include a radio, a water pump, bulk purchase dis-
counts of food items, and food preparation efficiencies (which may lower the
cost of fuel and time).
Although the goal is to construct a money-metric wealth indicator for
each person, fulfilling that goal may not be straightforward. Most of the
time, data for commodities and services consumed are collected at the
household level. A household typically consists of members with different
characteristics, such as age, sex, and employment status. Usually, an individ-
ual’s welfare indicator is calculated by dividing total household expenditures
by the number of people residing in that household. The result is called the
per capita expenditure.
Analyzing poverty on the basis of per capita expenditure, however,
ignores the fact that different individuals may have different needs.
The cost per person to reach a certain welfare level may be lower in
large households, because large households enjoy certain economies of
scale. For example, a child may not need the same share of income as
an adult member, or the food consumption expenditure may not be the
same across men and women within a household. The minimum income
needed to meet the subsistence needs of a household with four adults
may be much more than the subsistence income needed for a household
with two adults and two children. This intrahousehold allocation can be
adjusted using an equivalence scale tool.
47
There are various types of equivalence scales and economies of scale.

Also, there are different ways of determining these scales, such as evalu-
ating nutritional needs and behavioral needs. Differences in nutritional
needs are derived from various health studies. Data on behavioral needs are
obtained from econometric estimates that are based on observed commodity
allocations.
However, the observed allocation is suspect because what is observed
may not necessarily be what is actually needed. For example, if female chil-
dren are observed to consume less, does this mean that they need less, or are
they just discriminated against? There is no straightforward answer to this
question, unfortunately, because it is beyond the scope of most consumer
expenditure surveys.
Two adult equivalence (AE) scales are more commonly used than oth-
ers. The first is used by the Organisation for Economic Co-operation and
Development (OECD), which we denote by AEOECD. It is defined as
AEOECD = 1 + 0.7(NA − 1) + 0.5NC , (2.1)
where NA is the number of adults in the household, and NC is the number
of children in the household.
This scale actually serves as both an equivalence scale and an economy
of scale. Note that when there is only one adult member in the household,
AEOECD = 1. For a household with two adult members, AEOECD = 1.7
(AEOECD = 2 is incorrect because two adults sharing the same household
are assumed to enjoy economy of scale). For instance, if the actual total
income of a two-member household is Rs 17,000, then the per capita real
income of the household is equivalent to Rs 17,000/1.7 = Rs 10,000 and not
Rs 8,500, as it would be in the per capita case. This is an example of adjust-
ing for economy of scale. For a single parent household with two children,
however, the actual total income of Rs 17,000 is equivalent to a per capita
real income of Rs 8,500 because AEOECD = 1 + 2 × 0.5 = 2.
The second adult equivalent scale is used by the Living Standards
Measurement Study (LSMS), which we denote by AELSMS. It is defined as
AELSMS = (NA + ϱNC)ϑ, (2.2)
where NA is the number of adults in the household, and NC is the number
of children in the household.
In this scale, parameter ϱ measures the cost of a child compared to an
adult. Parameter ϑ captures the effect of economy of scale. Both parameters
48
are positive but not larger than one. When ϱ = 1, then the cost of a child
is equal to the cost of an adult. The lower the value of ϱ, the lower the cost
of each child compared to an adult. Similarly, when ϑ = 1, no economy of
scale is assumed. The lower the value of ϑ, the larger the economy of scale
is assumed to be.
For example, suppose there are five members in a household: three adults
and two children. If a child is assumed to be half as costly as an adult, then
ϱ = 0.5 and ϑ = 0.5. Then AELSMS = (3 + 0.5 × 2)0.5 = 2. Therefore, if the
actual total income of the household is Rs 20,000, then the real per capita
income of the household is equivalent to Rs 10,000. However, if no econ-
omy of scale is assumed and each child is considered as equally expensive as
an adult, then the household’s per capita income is only Rs 4,000.
In the subsequent analysis in this chapter, we assume that we are using
a dataset having all the information required for constructing a welfare
indicator either at the individual level or at the household level. The
dataset may cover the entire population or may just be a collection of
samples from the population. There are other important issues one should
take into account regarding a dataset (such as its survey design, sample
coverage, sample variability, and so on), which are not covered in this
chapter.2
To keep explanations and mathematical formulas simple, we make two
fundamental assumptions. First, we use income as the welfare indicator
and assume that information on income is available for every person in our
dataset. Second, we assume that every household contains only one adult
member. As a result of the second assumption, we do not need to make
any adjustment for the economy of scale and equivalent scale because each
member is an adult and lives in a single-member household. However, the
tools and techniques introduced in this chapter can be easily extended to
situations when the welfare indicator is consumption expenditure and more
than one person lives in a household.
Basic Concepts
Suppose our reference society X consists of N people, where the income of

person n is denoted by xn for all n = 1,2,…,N. Thus, the income distribution
data for society X has N incomes. For the sake of simplicity, we assume these
incomes are ordered so that x1 ≤ x2 ≤ … ≤ xN.
49
There are two different ways to represent an income distribution:
• The simplest income distribution is a vector of incomes. We denote

the society’s vector of incomes as X = (x1,x2,…,xN).
• The second way is to represent the income distribution in terms of
a cumulative distribution function (cdf) in which x is designated an
income distribution. We denote the average, or mean, of all elements
in x by x̄ = (x1 + … + xN)/N. For a large enough sample, the cdf may
be approximated by a density function.
Another, more intuitive, presentation of the cdf is the quantile function,

which is more suitable to our needs. Before moving into the discussion on
measurement, we will discuss these three concepts and examine their signifi-
cance in describing various aspects of an income distribution.
Density Function
An income distribution’s density function reports the percentage of the popu-

lation that falls within an income range. Suppose incomes in distribution x
range from $100 to $100,000, and we want to know what percentage of the
population earns income between $10,000 and $20,000. The answer can
be easily obtained by calculating the area underneath the density function
between $10,000 and $20,000.
Notice that the total area underneath the density function between $100
and $100,000 is 100 percent because incomes of the entire population fall
within this range. Thus, the density function is a frequency distribution that
is normalized by the total population in the distribution.
Figure 2.1 depicts the probability density function of income distribution
x. Recall that the minimum and maximum incomes in distribution x are x1
and xN, respectively. The horizontal axis reports the income and the verti-
cal axis reports the density. We denote the density function of distribution
x by fx, which is a bell-shaped curve in figure 2.1. The total area between x1
and xN underneath the density function fx is 100 percent. The share of the
population with incomes between b' and b'' is the shaded area.
Two interesting statistics may be found in figure 2.1:
• The median is the income in the distribution that divides the entire
population into two equal shares. In the figure, xM is the median of
50
Figure 2.1: Probability Density Function
Density
fx
x1 xMo xM b ′ b ″ xN
Income
distribution x. Hence, 50 percent of the area underneath fx lies to the

right of xM, and the remaining 50 percent lies to the left of xM.
• The mode is the income in the distribution that corresponds to the
largest density (locally). In figure 2.1, the distribution’s mode is
denoted by xMo.
Commonly, income distributions have one mode, but there can
be distributions with more than one mode. A density with two modes
is called bimodal and that with many modes is called multimodal.
When there is more than one mode, a society is understood to be
polarized in different groups according to their achievements. A
polarized society may produce social tensions among different groups,
which increases the chance of social unrest. These issues are discussed
in more detail in chapter 3.
In addition, a density function can be a useful tool for understanding the

skewness of an income distribution. Skewness is a measure of asymmetry in
the distribution of incomes. It arises when most incomes lie on any one side
of the mean of the distribution. If more observations are located to the left
of the distribution’s mean, then the distribution is positively skewed. If more
observations lie to the right of the mean, then the distribution is negatively
51
skewed. If there is an equal number of observations on both sides of the

mean, then there is no skewness, and the distribution is symmetric around
the mean. Income distributions are usually positively skewed.
Cumulative Distribution Function

A cdf, or cumulative distribution function, denotes the proportion of the
population whose income falls below a given level. A cdf may be easily
obtained from a density function and vice versa. For every income reported
on the horizontal axis of figure 2.1, a distribution function reports the area
to the left of the income underneath fx. Because the total area underneath fx
is 100 percent, the highest value that a distribution function can take is 100
percent. We denote the distribution function of x by Fx, and Fx(b) denotes
the percentage of the population whose income is no greater than b.
For example, if the number of people in society X having incomes less
than b is q, then Fx(b) = 100 × q/N. For any two incomes b' and b", Fx(b')
≤ Fx(b") when b' ≤ b" because having income less than b' must also imply
having income less than b". Therefore, a distribution function should not
decrease as income increases.
As seen in figure 2.2, the horizontal axis denotes income and the verti-
cal axis denotes the value of the cumulative distribution function. For xN,
which is the largest income in distribution x, the value of the distribution
function is Fx(xN) = 100 percent because no one in distribution x has an
income above xN.
Figure 2.2: Cumulative Distribution Function

Cumulative distribution
Fx(xN) = 100%
Fx(b ″) Fx
Fx(b ′)
Fx(bM) = 50%
Mean
bM b ′ b ″ xN
Income
52
At median bM, the distribution function’s value is Fx(bM) = 50 percent,

which implies that half the population has an income less than bM. In figure
2.1, the share of the population with income ranging between b' and b" is
represented by the shaded area, which, in figure 2.2, is denoted by the dif-
ference Fx(b") − Fx(b'). A distribution function provides another important
statistic: the mean of the distribution. In figure 2.2, the shaded area to the
left of Fx is the mean x̄ of distribution x.
Quantile Function
A quantile function is the inverse of a cdf. Recall that a distribution function

shows the percentage of the population whose income falls below a given
level of income. The quantile function, however, reports the level of income
below which incomes of a given percentage of the population fall.
We denote the quantile function of distribution x by Qx and by con-
struction Qx = Fx–1, where Fx–1 is the inverse of the cdf Fx. For example, the
level of income below which incomes of 25 percent of the population lie is
Qx(25). If 25 percent of Georgia’s population has income below GEL 2,000,
then QGEO (25) = GEL 2,000.
Figure 2.3 describes the quantile function corresponding to distribution
x. The horizontal axis denotes the population share, or the percentage of
Figure 2.3: Quantile Function
xN
Qx
Income
bM
x1 Mean
0 50 100
Population share (percent)
53
population. The vertical axis denotes the corresponding value of a quantile

function in terms of income. Of course, no one in the society can have any
income above Qx (100). Half of the population has an income less than the
median bM, so Qx (50) = bM. The shaded area underneath the quantile func-
tion is the mean x̄ of the distribution x.
Having introduced these basic concepts, we discuss income standards in
the next section.
Income Standards
An income standard gauges the size of a distribution by summarizing the

entire distribution in a single income level. Some income standards can be
viewed as stylized measures of a society’s overall level of well-being. Others
focus more narrowly on one part of the distribution or have no general wel-
fare interpretation. We begin this section by introducing common proper-
ties that an income standard should satisfy. We denote any income standard
by W and use subscripts to indicate specific measures or indices.
Desirable Properties
An income standard can satisfy several basic properties. We refer to the first
two properties—symmetry and population invariance—as invariance properties
because they describe changes in the distribution that leave the income
standard unaltered. The second pair of properties—weak monotonicity and
the weak transfer principle—are called dominance properties because they
require the income standard to rise (or not fall) when the income distribution
changes in a particular way. Finally, normalization and linear homogeneity are
calibration properties that ensure the income standard is measured by income.
The additional property of subgroup consistency is not a part of the basic prop-
erties, but it is desirable when evaluating income standards of subpopulations.
Symmetry requires that switching two people’s incomes leaves the
income standard evaluation unchanged. In other words, a person should
not be given priority on the basis of his or her identity when calculating a
society’s income standard. Thus, symmetry is also known as anonymity. In
technical terms, symmetry requires the income standard of distribution x to
be equal to the income standard of distribution x', if x' is obtained from x by
a permutation of incomes.
54
What is a permutation of income? An example will explain. Consider

the three-person income vector x = ($10k, $20k, $30k) so that the first,
second, and the third person receive incomes $10k, $20k, and $30k, respec-
tively. If the incomes of the first and second persons are switched, then the
new income vector becomes x' = ($20k, $10k, $30k). This new vector x' is
said to be obtained from vector x' by a permutation of incomes. The sym-
metry property thus can be stated as follows:
Symmetry: If distribution x' is obtained from distribution x by a per-

mutation of incomes, then W(x') = W(x).
The second property is population invariance. This property requires that

the income standard not depend on population size. That is, a replication of
an income vector results in the same income standard as the original sample
vector. Consider the income vector of society X to be x = ($10k, $20k, $30k).
Now suppose three more people join the society with the same income
distribution. The new income vector of society X is x' = ($10k, $10k, $20k,
$20k, $30k, $30k). Society X now has more overall income, but population
invariance requires that the income standard of society X remain unaltered.
What is the implication of population invariance? It allows us to
compare income standards across countries and across time with varying
population sizes. Furthermore, when combined with symmetry, population
invariance allows the income standard to depend only on information found
in a distribution function, which does not include the population size and
the identities of income receivers.
Population Invariance: If vector x' is obtained by replicating vector x

at least once, then W(x') = W(x).
The third property requires that if the income of any person in a society
increases, then the income standard should register an increase, or at least
should not fall. Implicitly, this property assumes that increasing someone’s
income is not harmful to the entire society.
There are two versions of this property. One is weak monotonicity, which
requires that the income standard not fall because of an increase in any-
one’s income. The other version is monotonicity, the stronger version, which
requires that the income standard register an increase if anyone’s income in
the society increases.
55
For vectors x and x', the notation x' > x implies that at least one element
in x' is strictly greater than that in x, and all other elements in x' are no less
than the corresponding elements in x. For example, if x' = ($20k, $10k, $30k)
and x = ($25k, $10k, $30k), then x' > x. However, if x' = ($20k, $10k, $30k)
and x = ($25k, $10k, $25k), then x' ⬎ x because the income of the third
person is lower in x than that in x'.
Weak Monotonicity: If distribution x' is obtained from distribution x

such that x' > x, then W(x') ≥ W(x).
Monotonicity: If distribution x' is obtained from distribution x such
that x' > x, then W(x') > W(x).
Some income standards are occasionally interpreted as social welfare mea-

sures. The fourth property, known as the transfer principle, is the key property that
enables this interpretation. A regressive transfer occurs when income is transferred
from a poorer person to a richer person. The transfer principle requires that a
regressive transfer between two people in a society should lower the income
standard. Conversely, a progressive transfer occurs when income is transferred
from a richer person to a poorer person. The transfer principle requires that a
progressive transfer between two people raise the income standard.
Here is a formal definition of these two kinds of transfers using vector x.
We have already assumed that incomes in x are ordered so that x1 ≤ x2 ≤ … ≤
xN. Let income d be transferred from person n to person m, where n < m and
0 < d < (xm − xn)/2. Denote the post-transfer income vector by x', where all
incomes except those for people n and m are the same as in x, but x'n = xn − d
and xm' = xm + d. Then x' is said to be obtained from x by a regressive transfer.
Now, let income d > 0 be transferred from person m to person n. Denote
the post-transfer income vector by x", where all incomes except those for
people n and m are the same as in x, but x"n = xn + d and xm " = xm − d such
that xm" > xn. Then x" is said to be obtained from x by a progressive transfer.
Consider the following example. Let the two income vectors of society
X at two different points in time be x = ($10k, $20k, $30k) and x' = ($15k,
$20k, $25k), where x' has been obtained from x by transferring $5k from the
third person to the first person. This is a progressive transfer.
Below is the formal statement of the transfer principle property. This
principle is also known as the Pigou-Dalton transfer principle after the
English economists Arthur Cecil Pigou and Hugh Dalton.3
56
Transfer Principle: If distribution x' is obtained from distribution x by

a regressive transfer, then W(x') < W(x). If distribution x" is obtained
from distribution x by a progressive transfer, then W(x") > W(x).
One justification of the transfer principle invokes a utilitarian form of

welfare function that takes welfare to be the average level of (indirect) util-
ity in society and assumes that all utility functions are identical and strictly
increasing (see Atkinson 1970). In this context, the intuitive assumption of
diminishing marginal utility yields the transfer principle. Diminishing marginal
utility requires that the loss to the poorer giver is greater than the gain to
the richer receiver because of a regressive transfer. Hence, overall welfare
falls, or, equivalently, the gain to the poorer receiver is greater than the loss
to the richer giver because of a progressive transfer—hence, welfare rises.
The fifth property is normalization. This property requires that if incomes
are the same across all people in a society, then the income standard should
be represented by that commonly held income. This property is intuitive.
For example, let the income vector of a three-person society be ($20k, $20k,
$20k). Then the income standard should be $20k.
Normalization: For the income distribution, x = (b, b, …, b), W(x) = b.
The sixth property is linear homogeneity. This property requires that if

an income distribution is obtained from another income distribution by
changing the incomes by some proportion, then the income standard should
also change by the same proportion. For example, if everyone’s income in
a society doubles, then the society’s income standard doubles. If everyone’s
income is halved, then the society’s income standard is halved.
Linear Homogeneity: If distribution x' is obtained from distribution x

such that x' = cx where c > 0, then W(x') = cW(x).
Subgroup consistency is the final property presented here. In some empiri-

cal applications, there is a natural concern for certain identifiable popula-
tion subgroups as well as for the overall population. We might be interested,
for instance, in the performances of various states or subregions of a country
to understand how the overall improvement in income standard is distrib-
uted across those regions.
57
When population subgroups are tracked alongside the overall population

value, there is a risk that the income standard could indicate contradictory
or confusing trends. For example, it may be possible that the income stan-
dards of some regions within a country improve while the income standards
of the rest of the country remain the same, but the income standard of the
country as a whole deteriorates. This type of result may cause confusion
because following the regional performances, one would expect the coun-
try’s overall performance to improve.
Thus, a natural consistency property for an income standard might be
that if subgroup population sizes are fixed but incomes are varying, when the
income standard rises in one subgroup and does not fall in the rest, the over-
all population income standard must rise. This property, known as subgroup
consistency, avoids inconsistencies arising from multilevel analyses of this sort.
As an example, suppose the income vector x with population size N is
divided into two subgroup vectors x' with population size N' and x" with
population size N" such that N' + N" = N. Let a new vector y be obtained
from x with the same population size N and its corresponding two subgroups
be y' with population size N' and y" with population size N". The subgroup
consistency property can be stated as follows:4
Subgroup Consistency: Given that the overall population size and

the subgroup population sizes remain unchanged, if W(y') > W(x')
and W(y") ≥ W(x"), then W(y) > W(x).
Having discussed the properties of the income standards, we now discuss

the commonly used income standards. We outline these income standards
and analyze their usefulness in terms of the properties they satisfy.
Commonly Used Income Standards
Four kinds of income standards are in common use: quantile incomes,

partial means, general means, and means based on the maximin approach.
(Among the maximin means, we discuss only the Sen mean in this book.)
We now describe each kind in greater detail.
Quantile Income
Quantile incomes provide information about a specific point on the distri-

bution. They can be directly calculated from a quantile function or a cdf.
58
The quantile income at the pth percentile is the income below which the
incomes of p percent of the population fall. For the income distribution x
with N people, the quantile income at the pth percentile is the income that
is larger than the incomes of the poorest pN/100 people.
We denote the quantile income at the pth percentile of distribu-
tion x by WQI (x; p). For example, if p = 50 percent, then the quantile
income at the pth percentile of distribution x is denoted by WQI (x; 50).
If WQI (x; 50) = $200, then it should be read as 50 percent of the population
in society X earns less than $200. Similarly, if WQI (x; 90) = $1,000, then
90 percent of its population earns less than $1,000.
Commonly reported quantile incomes used when gauging societies’
standard of living are the quantile incomes at the 10th percentile, 20th per-
centile, 50th percentile, 80th percentile, and 90th percentile. A close look
at the quantile income at the 50th percentile reveals that this is the income
below which half of the population of a distribution lies. Therefore, the
quantile at the pth percentile income is just the median of a distribution. For
a particular income distribution where each and every person earns equal
income, the quantile incomes at all percentiles are equal to each other,
ensuring that the quantile incomes satisfy the normalization property.
A quantile function is the most helpful tool for visualizing quantile
incomes. Figure 2.4 shows the quantile function for income distribution x.
Figure 2.4: Quantile Function and the Quantile Incomes
Qx(100) WQI(x;100)
Corresponding value of
a quantile function Qx
Qx(90) WQI(x;90)
Quantile income
Qx
Qx(75) WQI(x;75)
bM WQI(x;50)
Qx(25) WQI(x;25)
Qx(10) WQI(x;10)
0 10 25 50 75 90 100
Population share or percentiles
59
As in figure 2.3, the horizontal axis in figure 2.4 denotes the population share
in percentage, which lies between 0 and 100. The left-hand vertical axis
denotes the corresponding value of a quantile function Qx and the right-hand
vertical axis reports the quantile incomes.
By definition, the quantile income for a certain percentile is the value of
the quantile function at that percentile, so WQI (x; p) = Qx(p). In the figure,
WQI (x; 50) = bM is the median of distribution x. Likewise, WQI (x; 25) and
WQI (x; 75) are the first and the third quartiles of distribution x. The well-
known 10th and 90th percentiles of distribution x are WQI (x; 10) = Qx(10)
and WQI (x; 90) = Qx(90), respectively. Given that a cdf is an inverse of a
quantile function, quantile incomes can also be graphically portrayed and
calculated using a cdf.
What properties do quantile incomes satisfy? It is straightforward to verify
that any quantile income satisfies symmetry, normalization, population invari-
ance, linear homogeneity, and weak monotonicity. However, no quantile income
satisfies the other dominance properties: monotonicity, transfer principle,
and subgroup consistency. Quantile incomes do not satisfy monotonicity
because a person’s income may increase, but as long as it does not surpass a
certain quantile, that quantile income remains unaltered. Similarly, quantile
incomes do not satisfy the transfer principle because they do not change to a
transfer that takes place at a nonrelevant part of the distribution.
The income standards are not subgroup consistent because the quantile
incomes of the subregions may increase, but the overall quantile income may
fall. Consider the following example. Suppose the income vector of society
X is x = ($10k, $20k, $30k, $50k, $60k, $80k) and the income vector of two
subgroups is x' = ($10k, $20k, $30k) and x" = ($50k, $60k, $80k). The 67th
quantile of the three distributions is WQI (x'; 67) = $20k, WQI (x"; 67) = $60k,
and WQI (x; 67) = $50k. Now, suppose the subgroup income vectors over time
become y' = ($10k, $20k, $30k) and y" = ($45k, $65k, $80k). Apparently, the
quantile income at the 67th percentile of the first group does not change, but
that of the second does. In fact, WQI (x'; 67) = WQI (y'; 67) but WQI (y"; 67) >
WQI (x"; 67). What happens to the quantile income at the 67th percentile
of the overall distribution? It turns out that WQI (y; 67) = 45 < WQI (x; 67).
Partial Mean
The next set of commonly used means is the partial means. There are two
types of partial means: lower partial means and upper partial means. A lower
60
partial mean is obtained by finding the mean of the incomes below a specific
percentile cutoff. An upper partial mean is obtained by finding the mean of
incomes above a specific percentile cutoff. Lower partial means are more
commonly used than upper partial means.
The lower partial mean of the pth percentile is the average or mean
income of the bottom p percent of the population. The upper partial mean
of the pth percentile, in contrast, is the average or mean income of the
top (1 – p) percent of the population. We denote the lower partial mean
and upper partial mean of distribution x for percentile p by WLPM(x; p) and
WUPM(x; p), respectively. For example, if p = 50 percent, then the lower par-
tial mean of the pth percentile of distribution x is denoted by WLPM(x; 50).
If WLPM(x; 50) = $100 and WUPM(x; 50) = $10,000, then together they
should be read as the mean income of the bottom 50 percent of the population
is $100, and the mean income of the top 50 percent of the population is $10,000
(see example 2.1).
Example 2.1: Consider the income vector x = ($2k, $4k, $8k, $10k).
The lower partial mean of the 50th percentile of the distribution is
($2k + $4k)/2 = $3k, and that of the 75th percentile of the distribution
is ($2k + $4k + $8k)/3 = $4.7k. In contrast, the upper partial mean
of the 50th percentile of the distribution is ($8k + $10k)/2 = $9k and
that of the 75th percentile of the distribution is $10k.
The following is a graphical description of how partial means can

be calculated using quantile function Qx. The vertical axis of figure 2.5
denotes income, and the horizontal axis denotes population share. There
are two percentiles, p' and p", for describing the lower and upper partial
means. The lower partial mean of the p' percentile population is the
shaded area underneath the quantile function Qx to the left of p' divided
by p'. The lower partial mean is the average income of all people in society
X whose income is less than Qx(p'). Similarly, the upper partial mean of
the p" percentile population is the shaded area underneath the quantile
function Qx to the right of p" divided by (100 – p"). This upper partial
mean is the average income of all people in society X whose income is
larger than Qx(p").
Like the quantile incomes, any partial mean satisfies symmetry, normal-
ization, population invariance, linear homogeneity, and weak monotonicity, but
no partial mean satisfies monotonicity, transfer principle, and subgroup
61
Figure 2.5: Quantile Function and the Partial Means
xN
Qx(p″)
Income
Qx
Qx(p′)
0 p′ p″ 100
Population share
consistency. Like the quantile incomes, one can easily show using a simple
example that partial means do not satisfy subgroup consistency.
Quantile incomes and partial means are crude income standards because
they do not depend on the entire income distribution. Yet they are highly
informative and easy to understand. Especially when income data are miss-
ing for certain parts of the income distribution, these crude income stan-
dards are useful tools for understanding a society’s performance.
In contrast, when rich datasets are available, a study based on quan-
tile incomes and partial means may be limited because they do not reflect
changes in every part of the distribution. For example, if the income
of a person below the median increases—but not by enough to surpass
the median income—then the distribution median does not reflect any
change.
The following income standards are designed to consider the entire
distribution. These income standards will, in most cases, reflect a change in
any part of the distribution.
General Mean
General means are a family of normative income standards. Standards in this

family are normative because the formulation of each measure depends on
62
a parameter denoted by a, which can take any value between − ∞ and + ∞.

Unlike the quantile means and the partial means, general means take into
account the entire income distribution, but emphasize lower or higher incomes
depending on the value of a. Parameter a is familiar in the literature as the
order of general means.
For income distribution x, we denote the general mean of order a by
WGM(x; a). It is defined as
⎧⎛ x a + x a + …+ x a ⎞ 1a
⎪⎪⎜ ⎟⎠ if a ≠ 0.
1 2 N
WGM (x; a ) = ⎨⎝ N (2.3)

⎪ 1
⎪⎩(x1 × x 2 × …× x N ) N if a = 0
Although a may take any value between − ∞ and + ∞, four means in this
family are more well known than others: arithmetic mean, geometric mean,
harmonic mean, and Euclidean mean.
• For a = 1, WGM is known as the arithmetic mean (denoted by WA)

or the average x̄ of all elements in x and can be written as5
x1 + x 2 + L + x N
WA (x) = . (2.4)
N
• For a = 0, WGM becomes the geometric mean (denoted by WG) of all
elements in distribution x and can be expressed as
WG(x) = (x1 × x2 × ... × xN)1/N. (2.5)
If we take a natural logarithm on both sides of equation (2.5), we find
ln x1 + ln x 2 + L + ln x N
WL (x) = ln WG (x) = . (2.6)
N
WL(x) is the average of the logarithm of all incomes in distribution

x. The logarithm of incomes is frequently used for various analyses by
labor economists.
• For a = –1, WGM becomes the harmonic mean (WH) of distribution
x and can be expressed as
−1
⎛ x −1 + x 2−1 + L + x N−1 ⎞
WH (x) = ⎜ 1 ⎟⎠ . (2.7)
⎝ N
63
• Finally, another well-known mean is the Euclidean mean (WE),

obtained when a = 2. The Euclidean mean formula is
1
⎛ x12 + x 22 + L + x 2N ⎞ 2
WE (x) = ⎜ ⎟⎠ . (2.8)
⎝ N
Example 2.2 shows the results of calculating these means for a given
income vector.
• The arithmetic mean of x is ($2k + $4k + $8k + $10k)/4 = $6k.
• The geometric mean of x is ($2k × $4k × $8k × $10k)1/4
= $5.03k.
• The harmonic mean of x is [($2k−1 + $4k−1 + $8k−1 +
$10k−1)/4]−1 = $4.10k.
• The Euclidean mean of x is [($2k2 + $4k2 + $8k2 + $10k2)/4]1/2
= $6.78k.
Having been introduced to the family, one can now understand the
properties of general means and the way they depend on parameter a. All
means in this family satisfy symmetry, normalization, population invariance,
linear homogeneity, monotonicity, and subgroup consistency. Furthermore, for
a < 1, general means satisfy the transfer principle. Thus, the general means
satisfy all the dominance properties introduced earlier. One reason is that,
unlike the quantile means and the partial means, general means consider all
incomes in the distribution.
It is straightforward to show that general means satisfy symmetry, nor-
malization, population invariance, linear homogeneity, and monotonicity.
That general means satisfy subgroup consistency may be verified as follows:
if vector x is divided into subgroup vectors x' and x", then the general mean
of x can be expressed as
WGM(x; a) = WGM((WGM(x'; a), WGM(x"; a)); a). (2.9)
In other words, the general mean of x is the general mean of the general
means of x' and x". Then the monotonicity property ensures that subgroup
consistency is satisfied.
64
Another interesting property of WGM is its monotonic relationship with

parameter a, which requires that the value of WGM increase as a rises and
decrease as a falls. A lower a gives more emphasis to lower values within a
distribution and thus causes WGM to fall. Conversely, a higher a gives more
emphasis to higher values within a distribution, causing the value of WGM
to rise. Technically speaking, WGM(x; a) < WGM(x; a ') for any a < a '. We
refer to this property of general means as increasingness to a. It follows from
this property that WE(x) ≥ WA(x) ≥ WG(x) ≥ WH(x).
There is an exception, however, when the values of general means do
not change as a changes, and this happens when a distribution is degener-
ate. A society’s income distribution is degenerate if all people in that society
have equal incomes. For a degenerate income distribution, all general means
are equal; that is, WGM(x; a) = WGM(x; a ') for all a ≠ a '. Invariance of
general means to degenerate distribution is another way of ensuring that
they satisfy the normalization property.
Given that a ranges from − ∞ to + ∞, what is the range of WGM? Unlike
the value of a, however, WGM is not unbounded. Rather, it has a lower bound
and an upper bound. When a decreases and approaches − ∞, WGM(x; a) con-
verges to the minimum element in x. The society’s income standard in this
case is nothing, but the poorest person’s income is x1. In contrast, when a
increases and approaches + ∞, WGM(x; a) converges toward the maximum
element in x, and the society’s income standard equals the income of the
richest person, xN. Notice, however, that unlike the other general means,
these two extreme income standards—WGM(x; − ∞) and WGM(x; + ∞) —are
not sensitive to the entire distribution. That is, if any element in x other than
x1 and xN changes, these two income standards do not reflect that change.
Figure 2.6 describes the relationship between the family of generalized
means and parameter a. As already discussed, the general mean is the
arithmetic mean at a = 1, the geometric mean at a = 0, the harmonic
mean at a = −1, and the Euclidean mean at a = 2. Values of general means
increase with parameter a. They are bounded below by x1 = min{x} and are
bounded above by xN = max{x}.
One feature we should note carefully is that the general means are
undefined for a < 0 when there is at least one nonpositive element in an
income vector. For example, if an element of x is 0, then for a = −1, we
have (0)−1 = 1/0. Therefore, one requirement for any measure in this family
with a < 0 is that all elements in x be strictly positive.
65
Figure 2.6: Generalized Means and Parameter `
Generalized mean of order

xN
WE(x) WGM(x; α)
WA(x)
WG(x)
WH(x)
x1
–∞ –2 –1 0 1 2 ∞
Parameter
General Means as Welfare Measures

The transfer principle ensures that the general means may be interpreted as
social welfare measures. Actually, the general means for a < 1 are commonly
interpreted as measures of social welfare. This form of welfare function was
considered by Atkinson (1970), who then defined a helpful transforma-
tion of the function called the equally distributed equivalent income (ede).
The utility function that Atkinson assumed to obtain his particular ede was
1 1
U(x n ) = (x n ) a for a < 1 and a ≠ 0 and U(x ) = ln x for a = 0 for all n.
a n n
a
The ede represents the level of income x ede, which, if received by all people
in a society, yields the same welfare level as that of the original income dis-
tribution. Thus, like the general mean itself, the value of ede depends on the
parameter a, and for vector x, the ede of order a is EDE(x; a) = WGM(x; a).
Sen Mean
The usual mean can be reinterpreted as the expected value of a single income
drawn randomly from the population. Now, suppose that instead of a single
income, we were to draw two incomes randomly from the population (with
replacement). If we then evaluated the pair in terms of the lower of the two
incomes, this would lead to the Sen mean, which is defined as the expecta-
tion of the minimum of two randomly drawn incomes.6 These two random
incomes are drawn with replacement, which means that these two incomes
may belong to the same person in a society. If every income in distribution x
66
is compared with every other income in x with replacement, then there are
N2 possible comparisons. Thus, the Sen mean can be defined as
1 N N
WS (x) = ∑ ∑ min{x n′ x n′}.
N 2 n =1 n ′ =1
(2.10)
Because we are using the minimum of the two incomes, this number can be
no higher than the mean, and is generally lower. The Sen mean also empha-
sizes the lower incomes but in a way that differs from the general means with
α < 1, the lower partial means, or the quantile incomes below the median.
There is a straightforward way of calculating the Sen mean for an income
vector—by creating an N × N matrix that has a cell for every possible pair of
incomes and placing the lower value of the two incomes in the cell. Adding
all the entries and dividing by the number of entries (N2) to obtain their
mean provides the Sen mean. Consider example 2.3 to better understand
this way of calculating the Sen mean.
First, we construct the following matrix:
x $2k $4k $8k $10k

$2k $2k $2k $2k $2k
$4k $2k $4k $4k $4k
$8k $2k $4k $8k $8k
$10k $2k $4k $8k $10k
Each cell in this 4 × 4 matrix is the minimum of the top row and
the left column, both of which represent the ordered income vector x.
The Sen mean is the average of all elements in the matrix. Thus,
1
WS (x) = (7 × $2k + 5 × $4k + 3 × $8k + 1 × $10k) = $4.25k.
42
The Sen mean of x is lower than the arithmetic mean of x,
which is $6k.
There is another interesting way of understanding the Sen mean—the

weighted average of all elements of an income distribution—where the
weight on each element depends on the rank of the corresponding element.
Recall that we assumed x1 ≤ x2 ≤ … ≤ xN for distribution x so that the Nth
67
person has the highest income and the first person has the lowest income.
Thus, element xN receives the highest rank and element x1 receives the low-
est rank. The Sen mean attaches the highest weight to the lowest income,
the second-highest weight to the second-lowest income, and the lowest
weight to the highest income.
For distribution x, the Sen mean can be expressed as WS(x) = a1x1 + … +
aNxN, where aN = (2(N − n) + 1)/N2 for all n. Thus, the weight attached
to the highest income xN is aN = 1/N2; the weight attached to the second-
highest income xN–1 is aN–1 = 3/N2; and the weight attached to the lowest
income x1 is a1 = (2N − 1)/N2. The weight attached to the richest income
in the example above ($10k) is 1/16, whereas the weight attached to the
poorest income ($2k) is 7/16. Notice that the weights sum to one, that is,
1 N2
a1 + a 2 + ... + a N = (1 + 3 + 5 + ... + (2 N − 1) = = 1. (2.11)
N2 N2
Thus, the Sen mean can also be expressed as
1
WS (x) = ((2N − 1)x1 + (2N − 3)x 2 + L + 3x N −1 + x N ). (2.12)
N2
The Sen mean satisfies symmetry, normalization, population invariance, lin-

ear homogeneity, monotonicity, and the transfer principle. It does not, however,
satisfy subgroup consistency, which means it is possible that the Sen mean
of one region increases while the Sen mean for the other regions remains
the same and the overall Sen mean falls.
This failure to satisfy subgroup consistency can be shown using a simple
example. Suppose the income vector of society X is x = ($4k, $5k, $6k, $7k,
$14k, $16k) and the income vectors of two subregions are x' = ($4k, $5k,
$7k) and x" = ($6k, $14k, $16k). The Sen means of these three income
vectors are WS(x) = $6.22k, WS(x') = $4.67k, and WS(x") = $9.78k. Now,
suppose the income vector of society X changes to y = ($3.4k, $6.1k,
$6k, $6.5k, $14k, $16k) so that the income vector of the first subgroup
changes to y' = ($3.4k, $6.1k, $6.5k), whereas that of the other subgroup
remains unaltered such that y" = x". Note that the overall mean income
and the mean income of both groups remain unchanged. The Sen means
of the three income vectors become WS(y) = $6.24k, WS(y') = $4.64k, and
WS(y") = $9.78k. Clearly, the Sen mean of the first subgroup decreases
while that of the second subgroup remains the same; yet the overall Sen
68
mean goes up. This feature of the Sen mean is inherited by the inequality
and poverty measures that are based on the Sen mean—the famous Gini
coefficient and the Sen-Shorrocks-Thon index of poverty.
Finally, unlike Atkinson, Sen suggested going beyond the utilitarian
form. His key nonutilitarian example, the Sen mean, can be viewed as both
an ede and a general welfare function, because it satisfies the transfer prin-
ciple. If we denote the Sen ede as EDES(x), then EDES(x) = WS(x).
During our subsequent discussion in this chapter, we will see that these
five means (arithmetic, geometric, harmonic, Euclidean, and Sen) and their
various functional forms are often used in the measurement of welfare,
inequality, and poverty.
Dominance and Unanimity
An income standard provides a point estimate of the evaluation of a certain

income distribution. We might ask one obvious question: Does the direc-
tion of comparison between distributions in a given point in time, or even
across time, using one income standard continue to hold for other income
standards? Let us clarify this concern with a few examples.
Consider two income vectors x = ($4k, $5k, $6k, $7k, $14k, $16k) and
y = ($3k, $5k, $6k, $9k, $14k, $16k). If we use arithmetic mean WA as an
income standard, then WA(x) = 8.7 and WA(y) = 8.8. Clearly, distribution
y has higher mean income than distribution x. What if we, instead, use the
Sen mean? We get WS(x) = 6.22 and WS(y) = 6.19. Thus, according to the
Sen mean, distribution x has higher welfare than distribution y.
How do the geometric mean and the Euclidean mean of these two
vectors compare? According to the geometric mean, distribution x has
higher welfare than distribution y because WG(x) = 7.57 and WG(y) = 7.52.
According to the Euclidean mean, distribution y has higher welfare than
distribution x because WE(x) = 9.81 and WE(y) = 10.02. What we see from
these comparisons is that different income standards rank two distributions
differently.
Are there situations when the various income standards agree with each
other? This question leads to a discussion of dominance and unanimity. If
there is a situation where we find a dominance relation holding between
two distributions, then there is no need to use different income standards to
evaluate that situation because all income standards would agree. If there is
no unanimous relation, then certain curves may help in understanding the
69
source of ambiguity. Thus, conducting a dominance analysis that is based on

these curves should be the first step in welfare comparison.
A second important motivation for dominance analysis might be focus,
or an identified concern with different parts of the distribution. Has the
rapid growth for the higher-income group been matched by growth of the
middle-income group or the lower-income group? We spend some time in
this subsection finding answers to these questions by plotting entire classes
of income standards using the various curves to be defined next. If one curve
always remains above another curve, then all income standards in that class
agree in ranking—for example, two income distributions. However, if the
curves cross, then situations may arise in which different income standards
in the same class disagree with each other.
A first such curve is the quantile function itself, which simultaneously
depicts incomes from lowest to highest. When all income quantiles are
the same, then one income distribution always lies above another income
distribution. When two distributions never cross, the situation is known as
first-order stochastic dominance (FSD). An income distribution x first order
stochastically dominates another distribution y, denoted by x FSD y, if and
only if (a) no portion of x’s quantile function lies below y’s quantile func-
tion and (b) at least some part of x’s quantile function lies above y’s quantile
function. Let us denote quantile function using the notations introduced
earlier. So x’s quantile function is denoted by Qx and that of y is denoted by
Qy. Then, the definition of FSD is as follows:
First-Order Stochastic Dominance: Distribution x first order stochasti-

cally dominates another distribution y if and only if Qx(p) ≥ Qy(p) for
all p in the range [0,100] and Qx(p) > Qy(p) for some p.
The concept of FSD may also be understood in terms of cumulative

distribution functions. Recall that a quantile function is just an inverse
of a cdf. Using the notations introduced earlier, we denote the cdf of x by
Fx and that of y by Fy. The formal definition of FSD in terms of cdfs is as
follows:
First-Order Stochastic Dominance: Distribution x first order

stochastically dominates another distribution y if and only if Fx(b) ≤
Fy(b) for all b in the range [0, ∞] and Fx(b) < Fy(b) for some b.
70
FSD ensures higher welfare according to every utilitarian welfare func-

tion with identical, increasing utility functions. The robustness implied by
an unambiguous comparison of quantile functions extends to all income stan-
dards and all symmetric welfare functions for which “more is better.” However,
if the resulting curves cross, the final judgment is contingent on which income
standard is selected. Even in this case, the quantile function can be helpful in
identifying the winning and losing portions of the distribution.
Figure 2.7 depicts the situation where x FSD y. Panel a shows the FSD
by quantile functions, and panel b shows the FSD by cdfs. In panel a, the
quantile function of x lies completely above that of y, which means that
every quantile income of distribution x is larger than the corresponding
quantile income of distribution y, so x FSD y. The same argument applies to
the cdfs in panel b, where the cdf of x lies to the right of y. Later, we will find
the concept of FSD that is based on cdfs useful, especially in poverty analysis.
The generalized Lorenz (GL) curve is a second curve that is useful for
dominance analysis. The generalized Lorenz curve graphs the area under the
quantile function up to each percent p of the population. Thus, any point
on a generalized Lorenz curve is the cumulative mean income held by the
bottom p percent of the population. We denote the generalized Lorenz func-
tion of distribution x by GLx, and that for the p percent of the population
by GLx(p). By construction, for income distribution x, GLx(100) = WA(x)
and GLx (0) = 0.
Figure 2.7: First-Order Stochastic Dominance Using Quantile Functions and Cumulative
Distribution Functions
a. Quantile function b. Cumulative distribution function

100
Fy
Income
Qx Fx
Qy
0 100 0
Population share Income
71
Figure 2.8 describes the construction of a generalized Lorenz curve from

a quantile function of a five-person income vector x = ($10k, $15k, $20k,
$25k, $30k). There are five percentiles: 20th, 40th, 60th, 80th, and 100th. In
panel a, we outline the quantile function of x, Qx. In panel b, we report the
generalized Lorenz curve of x, GLx. The mean of distribution x is WA(x) = 20.
A point on the generalized Lorenz curve denotes the area underneath the
quantile function for the corresponding percentile of the population. Up to
the 20th percentile of the population, the area under Qx is the area A.
In panel b, the corresponding value of GLx for the 20th percentile of the
population is denoted by point I. Thus, the value at point I is A/100 = 10 ×
20/100 = 2. Similarly, the value of GLx for the 40th percentile of the popula-
tion is denoted by point II, and the value at point II is (A+B)/100 = (10+15)
× 20/100 = 5. Repeating this approach, we find that the value of GLx for
the 100th percentile of the population is denoted by point V, and the
value at point V is (A + B + C + D + E)/100 = (10 + 15 + 20 + 25 + 30)
× 20/100 = 20. Note that the value at point V is the same as the mean of
distribution x, WA(x).
The generalized Lorenz curve is closely linked with lower partial means
(see Shorrocks 1983). Recall from our earlier discussion that the lower
partial mean for a certain percentile of population p is the area underneath
Figure 2.8: Quantile Function and Generalized Lorenz Curve
a. Quantile function b. Generalized Lorenz curve
30
25
Q
x
20 20 V
WA(x)
Income
Income
E
15 14 IV
D
C
10 III GLx
B 9
A II
5
I
2
0 20 40 60 80 100 0 20 40 60 80 100
Population share Population share
72
the quantile function divided by the percentile itself. Thus, the height of
the generalized Lorenz curve at any percentile of population p is the lower
partial mean times p itself, because the height of the generalized mean is the
area underneath the quantile function at corresponding percentile p, that is,
GLx(p) = pWLPM(x; p). If income distribution x has a large enough sample
size, the generalized Lorenz curve takes a form similar to the one described
in figure 2.9.
The horizontal axis in figure 2.9 shows the population share, and the ver-
tical axis denotes the height of the generalized Lorenz curve by income. The
generalized Lorenz curve for distribution x is denoted by GLx. The maximum
height of GLx is WA(x). The height of GLx for the 50th percentile of the
population is GLx(0.5).
If the total income in distribution x is distributed equally across all
people in the society and distribution y is obtained, then the generalized
Lorenz curve GLy becomes a straight line. The maximum height of GLy is
also WA(x), because redistribution of incomes does not change the mean
income. Notice that the height of GLy is higher than the height of GLx
for every percentile p. This implies that every partial mean of distribution
y is larger than the corresponding partial mean of distribution x. Thus, two
generalized Lorenz curves of this sort show a dominance relation between two
distributions in terms of partial means.
Figure 2.9: Generalized Lorenz Curve
GLx(100) WA(x)
GLy
Income
GLy(50) WA(x)/2
GLx
WA(x ′)
GLx(50)
GLx ′ (50) GLx ′
0 50 100
Population share
73
All partial means agree that distribution y has higher welfare than dis-
tribution x. Similarly, if there is another distribution x' whose generalized
Lorenz curve, GLx', lies completely below GLx (also shown in figure 2.9),
then all partial means agree that distribution x has higher welfare than dis-
tribution x'. The heights of the generalized Lorenz curves for distributions
y, x, and x' at the 50th percentile are GLy(50), GLx(50), and GLy'(50),
respectively. The generalized Lorenz curve represents second-order stochas-
tic dominance, which signals higher welfare according to every utilitarian
welfare function with identical and increasing utility function exhibiting
diminishing marginal utility. Example 2.4 provides a practical illustration of
generalized Lorenz calculations. The generalized Lorenz curve is also closely
related to the Sen mean. For distribution x, the Sen mean, WS(x), is twice
the area underneath GLx.
Example 2.4: Suppose per capita income in India is Rs 25,000. If only

3 percent of this mean income is received by the poorest 20 percent
of the population, then GLInd(20) = Rs 750.
Suppose incomes in India were redistributed, thereby keeping
the average income unaltered so that everyone in India has identical
income. Let us denote this income distribution by y. Then the
cumulative average income received by the poorest 20 percent of the
population is 20 percent and GLy(20) = Rs 5,000. Thus, GLy(20) –
GLInd(20) = Rs 5,000 – Rs 750 = Rs 4,250. The loss of welfare
because of unequal distribution of income for the poorest 20 percent
of the population is Rs 4,250. In relative terms, the loss of welfare is
4,250/5,000 = 85 percent.
However, note that the loss presented in terms of the height of the
generalized Lorenz curve is not the potential loss in the mean income
of the poorest 20 percent of the population. The mean income of the
poorest 20 percent of the population is GLInd(20)/0.2 = Rs 3,750.
Had income been equally distributed, the mean income of the
poorest 20 percent would have been Rs 25,000. In that scenario, the
potential loss of mean income is Rs 21,250. But in a relative sense,
the percentage loss in mean income is 25,205/25,000 = 85 percent,
which is the same as the percentage loss in terms of the height of the
generalized Lorenz curve. In fact, the percentage loss of welfare using
the height of the generalized Lorenz curve is always the same as the
percentage loss of mean income of that percentile.
74
Finally, a third curve depicts the general mean levels as parameter a varies.
We call this curve a general mean curve. This curve has already been outlined
in figure 2.6, where it is increasing in α; tends to the minimum income for
very low a ; rises through the harmonic, geometric, arithmetic, and Euclidean
means; and tends toward the maximum income as α becomes very large.
Why is this curve useful? At the beginning of this subsection, an example
showed that different generalized means may rank an income distribution
differently. So the general mean curve is useful for determining (a) whether
a given comparison of general means is robust to the choice of any income
standard from the entire class of general means, and, if not, (b) which of the
income standards is higher or lower.
General mean curves are also related to the quantile function and the
generalized Lorenz curve. A higher quantile function will always yield a
higher general mean curve, and a higher generalized Lorenz curve will raise
the general mean curve for a < 1, or the general means that favor the low
incomes. The general mean curve concept will be particularly relevant to
our later discussions of Atkinson’s inequality measure.
Growth Curves
Some analyses go beyond the ordinal question (Which distribution is

larger?) to consider the cardinal question: How much larger in percentage
terms is one distribution than another? This question is especially salient
when the two distributions are associated with the same population at two
points in time. Thus, the second question follows: At what percentage rate
did the income standard grow?
The most common and well-known way of understanding growth is by
the growth of per capita income or mean income. The arithmetic mean is
the income standard involved in this case. However, the defining proper-
ties of an income standard ensure that its rate of growth is a meaningful
number that can be compared with the growth rates of other income stan-
dards, either for robustness purposes or for an understanding of the quality
of growth.
As in our use of various curves in dominance analysis, we may also use
different growth curves to understand how robust the growth of an income
standard is and to understand whether the growth is of meaningful quality.
A growth curve depicts the rates of growth across an entire class of income
standards, in which the standards are ordered from lowest to highest.
75
In fact, each of the three dominance curves presented earlier suggests an

associated growth curve. First, the growth incidence curve assesses how the
quantile incomes are changing over time. Second, the generalized Lorenz
growth curve indicates how the lower partial means are changing over time.
Finally, the general mean growth curve plots the rate of growth of each general
mean over time against parameter a. In the remainder of this section, we
discuss the concepts of these different growth curves in greater detail.
Growth Incidence Curve
We start with the growth incidence curve. Consider two income distributions,
x and y, at two different periods of time, where x is the initial income distri-
bution. The quantile incomes of distribution x and distribution y at percen-
tile p are denoted by WQI(x; p) and WQI(y; p), respectively. The growth of
quantile income at percentile p is denoted by
WQI (y; p) − WQI (x; p)

g QI (x, y; p) = × 100%. (2.13)
WQI (x; p)
If every quantile registers an increase over time, then gQI(x, y; p) > for
all p. The curve’s height at p = 50 percent gives the median income’s growth
rate. Note that no part of this growth curve provides any information about
the growth of mean income. Varying p allows us to examine whether this
growth rate is robust to the choice of income standard, or whether the low-
income standards grew at a different rate than the rest.
Figure 2.10 depicts the growth curves of quantile incomes. The vertical
axis denotes the growth rate of quantile income and the horizontal axis denotes
the cumulative population share. Suppose there are two societies, X and X'.
The income distributions of society X at two different points in time are
x and y, while those of society X' are x' and y'. The dashed growth curve
gQI(x, y) denotes the quantile income growth rates of society X over time,
whereas the dotted growth curve gQI(x', y') denotes the quantile income
growth rates of society X' over time.
Suppose the growth rates of mean income across these two distributions
are the same and are denoted by –g > 0. Thus, the solid horizontal line at
–g denotes the growth rate if the growth rate had been the same for all per-
centiles or the cumulative population share.
76
Figure 2.10: Growth Incidence Curves
Growth rate of quantile income
A
B
g
gQI(x,y)
A′ B′ gQI(x ′,y ′)
0 20 40 60 80 100
Cumulative population share
What information do these two growth curves provide? Growth between

x and y is pro-poor in the sense that lower quantile incomes have positive
growth, whereas the upper quantile incomes have negative growth. Growth
between x' and y', in contrast, is not pro-poor because lower quantile
incomes have negative income growth, whereas upper quantile incomes
have positive growth. In society X, the growth rate of income for the 20th
percentile is much higher than that of the 40th percentile, as denoted by
point A and point B, respectively. Note that the growths are higher than the
mean growth rates. In society X', however, the income growth rate for the
20th percentile is almost the same as that of the 40th percentile, as denoted
by point A' and point B', respectively. We will discuss pro-poor growth in
greater detail in the poverty section of this chapter.
Generalized Lorenz Growth Curve
The next growth curve is the generalized Lorenz growth curve. Consider the
two income distributions, x and y, used previously. The lower partial means
of distribution x and distribution y at percentile p are denoted by WLPM
(x; p) and WLPM(y; p), respectively. The growth of partial means at percen-
tile p is denoted by
77
WLPM (y; p) − WLPM (x; p)

g LPM (x, y; p) = × 100%. (2.14)
WLPM (x; p)
If every quantile income registers an increase over time, then gLPM(x, y;

p) > 0 for all p. Given that GLx(p) = pWLPM(x; p), the growth of the lower
partial mean at a certain percentile is equal to the growth of the general-
ized Lorenz curve at that percentile. So the height of the generalized Lorenz
growth curve at p = 20 percent is the rate at which the mean income of the
lowest 20 percent of the population changed over time.
Unlike the growth incidence curve, this curve provides information
about the growth rate of mean income, which is the height of the curve at
p = 100 percent. Again, varying p allows us to examine whether this growth
rate is robust to the choice of income standard, or whether the low-income
standards grew at a different rate than that of the rest. If the growth rates of
the lower-income standards are found to be lower than the mean income,
then overall growth, indeed, has not been pro-poor. However, if all lesser
“lower partial means” grow at a faster rate than the higher “lower partial
means,” then growth is assumed to be pro-poor.
Figure 2.11 depicts the growth curves of lower partial mean incomes.
The vertical axis denotes the growth rate of lower partial mean income, and
the horizontal axis denotes the cumulative population share. Following the
same notations as the growth incidence curve, suppose that there are two
Figure 2.11: Growth Rate of Lower Partial Mean Income

Growth rate of partial mean income
C
D gLPM(x,y)
g
C′ D′
gLPM(x ′,y ′)
0 20 40 60 80 100
Cumulative population share
78
societies, X and X'. The income distributions of society X at two differ-

ent points in time are x and y, whereas those of society X' are x' and y'.
The dashed growth curve gLPM(x, y) denotes growth rates of lower partial
mean income of society X over time, whereas the dotted growth curve
gLPM(x', y') denotes growth rates of lower partial mean income of society
X' over time.
are the same and are denoted by g– > 0. Thus, the solid horizontal line at
g– denotes the growth rate if the growth rate had been the same for all per-
centiles or the cumulative population share.
x and y is pro-poor in the sense that mean incomes of the population’s bot-
tom percentiles have positive growth, whereas mean incomes of the popula-
tion’s upper percentiles have negative growth. Growth between x' and y', in
contrast, is not pro-poor because mean incomes of the population’s bottom
percentiles have negative income growth, whereas mean incomes of the
population’s upper percentiles have positive growth.
In society X, the growth rates of the mean income of the bottom 20th
percentile of the population and that of the bottom 40th percentile of the
population are denoted by point C and point D, respectively. In society X',
however, the growth rate of mean income of the bottom 20th percentile of
the population and that of the bottom 40th percentile of the population are
denoted by point C' and point D', respectively. Note that growth of mean
income is the growth at the 100th percentile income where the two growth
curves meet because they have been assumed to have the same growth rate
of mean income.
General Mean Growth Curve
The final of the three growth curves is the general mean growth curve.
Considering the income distributions x and y discussed previously, we
denote the general mean of order a of distribution x and distribution y by
WGM(x; a) and WGM(y; a), respectively. The growth of general mean of
order a is denoted by
WGM (y; a) − WGM (x;a)

g GM (x, y; a) = × 100%. (2.15)
WGM (x; a)
79
When every general mean registers an increase over time, gGM(x,y; a) > 0.
When a = 1, the curve’s height is the usual mean income growth rate. This
rate is equal to the growth of the generalized Lorenz growth curve at p = 100
percent. At a = 0 the curve shows the growth rate for the geometric mean,
and so forth. As we will see later, each of these growth curves can help
in understanding the link between growth and change in inequality
over time.
Figure 2.12 shows the growth curves of general mean incomes. The verti-
cal axis denotes the growth rate of general mean income, and the horizontal
axis denotes the values of parameter a. Following the same notations as the
previous two growth incidence curves, suppose that there are two societies,
X and X'. Income distributions of society X at two different points in time
are x and y, whereas those of society X' are x' and y'. The dashed growth
curve gGM(x, y) denotes the growth rates of general mean income of soci-
ety X over time, whereas the dotted growth curve gGM(x', y') denotes the
growth rates of general mean income of society X' over time.
–
are the same and are denoted by g > 0. Thus, the solid horizontal line at
–
g denotes the growth rate if the growth rate had been the same for all a.
x and y is pro-poor in the sense that general means for lower a, which focus
more on the lower end of the distribution, have positive growth, whereas
Figure 2.12: General Mean Growth Curves

Growth rate of general mean income
gG ( ′)
M x,y) ′,y
(x
M
gG
g
–∞ –2 –1 0 1 2 ∞
Parameter
80
general means for larger a have negative growth. Growth between x' and y', in
contrast, is not pro-poor because the general means for lower a have negative
income growth whereas the general means for larger a have positive growth.
The mean income growth rates are the heights of the two growth curves at
a = 1, which are equal by assumption for this example. Heights at a = 0 and
a = –1 are growth rates of the geometric and harmonic means, respectively.
Inequality Measures
The second aspect of a distribution is spread, which is evaluated using a

numerical inequality measure, assigning each distribution a number that
indicates its inequality level. There are two ways of understanding and inter-
preting an income inequality measure. One way is through the properties it
satisfies. The other way is by using a fundamental link between inequality
measures and income standards. We begin with the first approach by out-
lining the desirable properties an inequality measure should satisfy. In this
section, any inequality measure is denoted by the notation I. Specific indices
are denoted by using corresponding subscripts.
An inequality measure should satisfy four basic properties: symmetry, popula-

tion invariance, scale invariance, and the transfer principle. Like income stan-
dards, these properties may be classified into categories. Invariance properties
leave the inequality measures invariant to certain changes in the dataset, and
they include symmetry, population invariance, scale invariance, and normaliza-
tion. The normalization property calibrates the measure’s value when there is
no inequality. Dominance properties cause inequality measures to change in
a particular direction. Properties in this category include the transfer principle
and transfer sensitivity. Other properties, such as subgroup consistency and
additive decomposability, are compositional properties relating subgroups and
overall inequality levels. Most of these properties are similar in interpreta-
tion to the corresponding properties of income standards.
The first property, symmetry, requires that switching the income levels of
two people leaves the evaluation of a society’s inequality unchanged. In other
words, a person should not be given priority on the basis of his or her identity
when evaluating a society’s inequality. In more technical terms, it requires
81
the inequality measure of distribution x to be equal to the inequality measure

of another distribution x' if x' is obtained from x by a permutation of incomes.
For example, recall the three-person income vector x = ($10k, $20k,
$30k) so that the first, second, and the third persons receive incomes $10k,
$20k, and $30k, respectively. If the incomes of the first and second persons
are switched, then the new income vector becomes x' = ($20k, $10k, $30k).
This new vector x' is said to be obtained from vector x by a permutation of
incomes.
Symmetry: If distribution x' is obtained from distribution x by

permutation of incomes, then I(x') = I(x).
The second property, population invariance, requires that the level of

inequality within a society is invariant to population size, in the sense that
a replication of an income vector results in the same inequality level as the
original sample vector. What is the implication of this property? Consider
the income vector of society X, x = ($10k, $20k, $30k). Now, suppose three
more people join the society with the same income distribution so that the
new income vector of society X is x' = ($10k, $10k, $20k, $20k, $30k, $30k).
The population invariance property requires that the inequality level in society
X remain unaltered. This property allows us to compare the inequality level
across countries and across time with varying population sizes. Furthermore,
population invariance allows the inequality measure to depend on a distribu-
tion function, which normalizes the population size to one.
Population invariance: If a vector x' is obtained by replicating vector

x at least once, then I(x') = I(x).
The third property, scale invariance, requires that if an income distribu-

tion is obtained from another distribution by scaling all incomes by the same
factor, then the inequality level should remain unchanged. For example,
if everyone’s income in a society is doubled or halved, then the level of
inequality of the society does not change. The scale invariance property
ensures that the inequality being measured is a purely relative concept and
is independent of the distribution’s size.
Scale invariance is analogous to the linear homogeneity property for
income standards, which ensures that the relative status of every person
82
remains unchanged when compared to the income standard, even after all
incomes are scaled up or down by the same factor. This similarity supports
the idea that the relative inequality level remains unchanged.7
Scale Invariance: If distribution x' is obtained from distribution x'

such that x' = cx, where c > 0, then I(x') = I(x).
The fourth property, normalization, requires that if incomes are the same
across all people in a society, then no inequality exists within the society
and the inequality measure should be zero. Normalization is a natural
property. For example, if the income vector of a three-person society is
($20k, $20k, $20k), then the inequality measure should be zero. Even if
everyone’s income increases 10-fold and the new income vector is ($200k,
$200k, $200k), the inequality measure should still be zero.
Normalization: For the income distribution x = (b, b ,..., b), I(x) = 0.
The fifth property is the transfer principle, which requires that a regressive
transfer between two people in a society should increase inequality and a
progressive transfer between two people should reduce inequality. Regressive
and progressive transfers were defined earlier for income standards.
Transfer Principle: If distribution x' is obtained from distribution x

by a regressive transfer, then I(x') > I(x). If distribution x" is obtained
from distribution x by a progressive transfer, then I(x") < I(x).
In inequality measurement, there is also a weaker version of the transfer

principle, which requires that a regressive transfer between two people in a
society not decrease inequality and that a progressive transfer between two
people not increase inequality. Thus, the weaker principle allows the pos-
sibility that the level of inequality may remain unaltered because of progres-
sive or regressive transfers.
Weak Transfer Principle: If distribution x' is obtained from

distribution x by a regressive transfer, then I(x') ≥ I(x). If distribution
x" is obtained from distribution x by a progressive transfer, then
I(x") ≤ I(x).
83
The transfer principle requires an inequality measure to decrease if the

transfer is progressive. However, it does not specify the amount by which
inequality should fall, and it is not concerned with the part of the distribu-
tion where the transfer is taking place. The same amount may be transferred
between two poor people or between two rich people. Should the transfer
have the same effect on the inequality measure no matter where it takes
place? Consider the four-person income vector x = ($100, $200, $10,000,
$20,000). First, suppose $20 is transferred from the second person to the
first person. The post-transfer income vector is x' = ($120, $180, $10,000,
$20,000). Thus, transferring 10 percent of the second person’s income has
increased the first person’s income by 20 percent.
Now, suppose instead that the same $20 transfer takes place between the
third and the fourth person. The post-transfer income vector is x" = ($100,
$200, $10,020, $19,980), where transferring 0.1 percent of the fourth person’s
income has increased the third person’s income by 0.2 percent. This transfer
makes hardly any difference in the large incomes of the two richer people.
It may seem that a transfer of the same amount between two poor people
and two rich people should not have the same effect on the overall inequal-
ity. However, the sixth property, transfer sensitivity, requires an inequality
measure be more sensitive to transfers at the lower end of the distribution. In
other words, this property requires that the inequality measure change more
if a transfer takes place between two poor people than if the same amount of
transfer takes place between two rich people the same distance apart.
Suppose the initial income distribution is x = (x1, x2, x3, x4), where x1 < x2
< x3 < x4, x2 – x1 = x4 – x3 > 0. Note that the distance between x1 and x2 is the
same as the distance between x3 and x4. Suppose distribution x' is obtained
from distribution x by a progressive transfer of amount d < (x2 – x1)/2 between
x2 and x1, that is, x' = (x1 + d, x2 – d, x3, x4), and distribution x" is obtained
from distribution x by a progressive transfer of the same amount d between
x3 and x4, that is, x" = (x1, x2, x3 + d, x4 – d ). Thus, the same amount of
progressive transfer has been made between two poorer people and two richer
people, who are equally distant from each other. Both x' and x" are more
equal than x according to the transfer principle, but can we compare x' and
x"? The answer is yes. In fact, any transfer sensitive inequality measure should
judge distribution x' as more equal than distribution x". Shorrocks and Foster
(1987) have reinterpreted the transfer sensitivity property in terms of favor-
able composite transfer (FACT). When a distribution is obtained from another
84
distribution by a progressive transfer at the lower end of a distribution and

simultaneously by a regressive transfer at the upper end of the same distribu-
tion, such that the variance remains unchanged, then the latter distribution
is stated to be obtained from the former distribution by FACT. Thus, the
transfer sensitivity property may be stated as follows:
Transfer Sensitivity: If distribution x' is obtained from distribution x

by FACT, then I(x') < I(x').
When one distribution is obtained from another distribution by FACT,

then the corresponding Lorenz curves intersect each other. In this case, the
transfer principle cannot rank two distributions. However, if a Lorenz curve
crosses the Lorenz curve of another distribution once from above, and the
coefficient of variation (standard deviation divided by the mean) of the former
distribution is no higher than that of the latter distribution, then all transfer
sensitive measures agree that the former distribution has less inequality.8
The seventh property is subgroup consistency, which is conceptually the
same as the corresponding property for income standards. This property
requires that if the sizes and means of a subgroup population are fixed, then
overall inequality must rise when the inequality level rises in one subgroup
and does not fall in the rest of the subgroups.
For example, suppose that income vector x with population size N is
divided into two subgroup vectors: x' with population size N' and x" with
population size N" such that N' + N" = N. Let a new vector, y, be obtained
from x with the same population size N, and let its two subgroups be denoted
by y' with population size N' and y" with population size N". The subgroup
consistency property can be stated as follows:
Subgroup Consistency: Given that subgroup population sizes and

subgroup means remain unchanged, if I(y') > I(x') and I(y") ≥ I(x"),
then I(y) > I(x).
There is a closely related property that is often useful for understanding

how much of the overall inequality can be attributed to inequality within
subgroups and how much can be attributed to inequality across subgroups,
given a collection of population subgroups. For example, the population of
a country may be divided across various subgroups, such as across rural and
85
urban areas, states, provinces, and other geographic regions; across ethnic
and religious groups; across genders; or across age groups. One may want to
evaluate the source of inequality, such as whether overall income inequality
is due to unequal income distribution within sex or unequal income distri-
bution across sex.
The eighth property is additive decomposability, which requires overall
inequality to be expressed as a sum of within-group inequality and between-
group inequality. Within-group inequality is a weighted sum of subgroup
inequalities. Between-group inequality is the inequality level obtained when
every person within each subgroup receives the subgroup’s mean income.
Kanbur (2006) discussed the policy significance of this type of inequality
decomposition. It is often found that the contribution of the between-group
term is much lower than the within-group term, and, thus, policy priority
is directed toward ameliorating within-group rather than between-group
inequality. These types of policy conclusions should be carefully drawn,
because the lower between-group term may receive much larger social
weight than its within-group counterpart. Also, the between-group term’s
share of overall inequality may increase as the number of groups increases.
How to incorporate these issues into inequality measurement requires fur-
ther research, and solving these issues is beyond the scope of this book.
However, if the policy interest is in understanding how the between-group
inequality as a share of total inequality has changed over time for a fixed
number of groups, then the decomposability property is very useful.
To formally outline the additive decomposability property, we will use
two groups to simplify the interpretation, but the definition can be extended
to any number of groups. Suppose the income vector x with population size
N is divided into two subgroup vectors: x' with population size N' and x"
with population size N" such that N' + N" = N. Let us denote the means of
these three vectors by x̄, x̄', and x̄". The additive decomposability property
can be stated as follows:
Additive Decomposability: If income distribution x is divided into

two subgroup distributions x' and x", then I(x) = W'I(x') + W"I(x") +
I(x̄',x̄"), where W' and W" are weights.
The between-group contribution is I(x̄', x̄")/I(x) and the within-group

contribution is [W' I(x') + W" I(x")]/I(x), as seen in example 2.5.
86
Example 2.5: Consider the five-person income vector x = ($10k, $15k,

$20k, $25k, $30k), which is divided into two subgroups x' = ($10k,
$30k) and x" = ($15k, $20k, $25k). The mean of x' is x̄' = $20k, and
the mean of x" is also x̄" = $20k. Let an additively decomposable
inequality index I be used to estimate the inequality level. The total
within-group inequality is W'I(x') + W"I(x"). However, there is no
between-group inequality in this case, because the mean incomes
of both groups are equal. So the between-group contribution I(x̄',
x̄")/I(x) is 0.
Inequality and Income Standards
There is a second way of understanding inequality measures: through

income standards. This, in fact, relies on an intuitive link between inequal-
ity measures and pairs of income standards: a and b. Let a be the smaller
income standard, and let b be the larger income standard. It is natural
to measure inequality in terms of the relative distance between a and b,
such as I = (b − a)/b, or some other increasing function of the ratio b/a.
Indeed, scale invariance and the weak transfer principle essentially require
this form for the measure. We will find in our subsequent discussions that
virtually all inequality measures in common use are based on twin income
standards.
Commonly Used Inequality Measures
Commonly used inequality measures are mostly related to the five kinds
of income standards we discussed earlier. The inequality measures that we
discuss in this section are quantile ratios, partial mean ratios, Gini coefficient,
Atkinson’s class of inequality measures, and generalized entropy measures.
Quantile Ratio
A quantile ratio compares incomes of higher and lower quantile incomes.

Inequality across quantile incomes provides a useful way to understand
income dispersion across the distribution. Because no quantile ratio considers
the entire distribution, this measure is a crude way of presenting inequality.
87
For income distribution x, let the quantile income at the pth percentile
be denoted by WQI(x; p), and let the quantile income at the p'th percentile
be denoted by WQI(x; p'), such that p > p'. A quantile ratio is commonly
reported as a ratio of the larger quantile income to the smaller quantile
income. However, this view leads the values of inequality measures to range
from one to ∞. This range is not comparable to other inequality measures,
which commonly range from zero to one. In this book, we formulate the
quantile ratio in such a way that it ranges from zero to one. The p/p' quantile
ratio is represented by the following formula:
WQI (x; p) − WQI (x; p ′) WQI (x; p ′)

IQR (x; p / p ′) = = 1− . (2.16)
WQI (x; p) WQI (x; p)
In this case, the quantile income at the pth percentile WQI(x; p) is the
higher income standard, and the quantile income at the p'th percentile
WQI(x; p') is the lower income standard.
The higher the quantile ratio, the higher the level of inequality across
two percentiles of the population in the society. A quantile ratio is zero
when both the upper and the lower quantile incomes are equal. A quantile
ratio reaches its maximum value of one when the lower quantile income
WQI(x; p') is zero. This means that no one in the lower percentile earns
any income and that the upper quantile income is positive. Note that if all
people in the society have equal incomes, then any quantile ratio is zero.
However, a quantile ratio of zero does not necessarily mean that incomes are
equally distributed across everyone in the society.
The quantile ratios used most often include the 90/10 ratio, 80/20 ratio,
50/10 ratio, and 90/50 ratio. The 90/10 ratio, for example, captures the dis-
tance between the quantile income at the 90th percentile and the quantile
income at the pth percentile as a proportion of the quantile income at the
10th percentile. How should the number IQR(x; 90/10) = 0.9 be interpreted?
There are, in fact, several ways to interpret the number:
• The number may be directly read as the gap between the lowest
income of the richest 10 percent and the highest income of the poorest
10 percent of the population, being 90 percent of the lowest income of
the richest 10 percent of the population.
• The number may be seen as the highest income of the poorest
10 percent of the population, being 10 percent (1 − 0.9 = 0.1) of the
lowest income of the richest 10 percent of the population.
88
• The number can be interpreted as the lowest income of the richest

10 percent of the population, being 10 times (1/(1 – 0.9)) larger
than the highest income of the poorest 10 percent of the population.
Similarly, IQR(x; 90/50) = 0.75 implies that the lowest income of the
richest 10 percent of the population is 1/(1 − 0.75) = 4 times larger
than the highest income of the poorest 50 percent of the population.
Quantile ratios may be classified into three categories: upper end quantile
ratio, lower end quantile ratio, and mixed quantile ratio. The first two categories
capture inequality within any one side of the median, and the third category
captures inequality in one side of the median versus that of the other side of
the median. For example, IQR(x; 90/50) is an upper end quantile ratio, and
IQR(x; 50/10) is a lower end quantile ratio, whereas IQR(x; 90/10) is a mixed
quantile ratio.
What properties does a quantile ratio satisfy? A quantile ratio, as defined
earlier, satisfies symmetry, normalization, population invariance, and scale
invariance. Thus, a quantile ratio satisfies all four invariance properties.
What about the dominance properties? It turns out that a quantile ratio
satisfies none of the dominance properties.
The following example shows that a quantile ratio does not satisfy the weak
transfer principle. Suppose the highest income of the poorest 10 percent of
the population is $100 and the lowest income of the richest 10 percent of the
population is $2,000. Then IQR(x; 90/10) = ($2,000 − $100)/$2,000 = 0.95.
Now, suppose that a regressive transfer takes place between the poorest
person in the society and the richest person among the poorest 10 percent
of the population such that the highest income in that group increases to
$120. Then the post-transfer quantile ratio is IQR(x; 90/10) = ($2,000 −
$120)/$2,000 = 0.94.
Therefore, the quantile ratio shows a decrease in inequality even when
a regressive transfer has taken place. If a quantile ratio does not satisfy the
weak transfer principle, then it cannot satisfy its stronger version—the
transfer principle, or transfer sensitivity. The quantile ratios are not addi-
tively decomposable and also do not satisfy subgroup consistency.
Partial Mean Ratio
A partial mean ratio is an inequality measure comparing an upper partial

mean and a lower partial mean. Like quantile ratios, no partial mean ratio
89
considers the entire income distribution; thus, it is also a crude way of

understanding inequality.
For income distribution x, let the upper partial mean for percentile p
be denoted by WUPM(x; p) and the lower partial mean for percentile p' be
denoted by WLPM(x; p'). A partial mean ratio is also commonly reported
as a ratio of both partial means ranging from one to ∞. However, as with
the quantile ratio, we formulate the partial mean ratio in such a way that
it ranges from zero to one. The p/p' partial mean ratio is represented by the
following formula:
WUPM (x; p) − WLPM (x; p′ ) W (x; p′ )
IPMR (x; p / p′ ) = = 1 − LPM . (2.17)
WUPM (x; p) WUPM (x; p)
The higher the partial mean ratio, the higher the level of inequality
across two percentiles of a society’s population. A partial mean ratio is
zero when both upper and lower partial mean incomes are equal. A quan-
tile ratio reaches its maximum value of one when the lower partial mean
income WLPM(x; p') is zero and the upper partial mean income is positive.
Note that if all people in the society have equal incomes, then any partial
mean ratio is zero. However, a partial mean ratio of zero does not necessarily
imply that incomes are equally distributed across all people in the society.
The most well-known partial mean ratio was devised by Simon Kuznets
and is known as the Kuznets ratio. It is based on two income standards: the
mean of the poorest 20 percent of the population and the mean of the rich-
est 40 percent of the population. Using our formulation, the Kuznets ratio
equivalent inequality measure of distribution x is denoted by IPMR(x; 20/40).
How should the number IPMR(x; 20/40) = 0.8 be interpreted? Again, there
are several ways to interpret this measure:
• The difference in mean income between the richest 20 percent of

the population and the poorest 40 percent of the population is
80 percent of the mean income of the richest 20 percent of the
population.
• The mean income of the poorest 40 percent of the population is
(1 − 0.8) = 0.2 or 20 percent or one-fifth of the mean income of the
richest 20 percent of the population.
• The mean income of the richest 20 percent of the population is
1/(1 − 0.8) = 5 times larger than the mean income of the poorest
40 percent of the population.
90
What properties does a partial mean ratio satisfy? A partial mean ratio,
as defined in equation (2.17), satisfies symmetry, normalization, population
invariance, and scale invariance. Thus, a partial mean ratio satisfies all four
invariance properties. What about the dominance properties? A quantile
ratio satisfies the weak transfer principle but does not satisfy the transfer
principle, transfer sensitivity, and subgroup consistency. It does not satisfy
the transfer principle because some regressive and progressive transfers may
leave the inequality measure unchanged, since a partial mean ratio does not
consider the entire income distribution.
Atkinson’s Class of Inequality Measures
Atkinson’s class of inequality measures, developed by Sir Anthony Atkinson,

is based on general means (see Atkinson 1970). All inequality measures in
this family are constructed by comparing the arithmetic mean and another
income standard from the family of general means. Recall that each mea-
sure’s formulation in the general means family depends on a parameter
denoted by a, which can take any value between − ∞ and + ∞.
In the Atkinson family of inequality measures, a is called the inequality
aversion parameter. The lower the value of a, the higher a society’s aver-
sion toward inequality. In other words, the more averse a society is toward
inequality across the population, the more emphasis it gives to lower
incomes in the distribution by choosing a lower value of a. The Atkinson
class of inequality measures for a < 1 may be expressed as
WA (x) − WGM (x; a) W (x; a)

IA (x; a) = = 1 − GM . (2.18)
WA (x) WA (x)
The Atkinson index of order a is the difference between the arithmetic

mean and the general mean of order a divided by the arithmetic mean. Any
Atkinson index lies between zero and one, and inequality increases as the
index moves from zero to one. The minimum level of inequality, zero, is
obtained when the total income is equally distributed across everyone in the
society. Unlike the quantile ratios and the partial mean ratios, if IA(x; a) = 0
for any a < 1, then, by implication, the total income in the society is equally
distributed. This is because any inequality measure in this family is con-
structed by considering the entire distribution.
91
We already know from our discussion of income standards that the value
of general means falls as a falls and vice versa. As a decreases, the distance
between WA(x) and WGM(x; a) increases, implying that IA increases as a
falls for a particular income distribution. Among the entire class of mea-
sures, three are used more frequently: a = 0, a = –1, and a = –2. For a = 0,
the general mean takes the form of the geometric mean. The corresponding
Atkinson’s inequality measure for distribution x is expressed as
WA (x) − WG (x) W (x)

IA (x; 0) = = 1− G . (2.19)
WA (x) WA (x)
For a = –1, the general mean is known as the harmonic mean. The cor-
responding Atkinson’s inequality measure for distribution x is expressed as
WA (x) − WH (x) W (x)

IA (x; −1) = = 1− H . (2.20)
WA (x) WA (x)
For a = –2, the general mean has no such name, and we will call it
simply WGM(X; –2). The corresponding Atkinson’s inequality measure for
distribution x is expressed as
WA (x) − WGM (x; −2) W (x; −2)

IA (x; −2) = = 1 − GM . (2.21)
WA (x) WA (x)
Following the relationship between the Atkinson’s class of inequality

measures and parameter a, we can state that IA(x; −2) < IA(x; −1) < IA(x; 0)
unless all incomes in distribution x are equal (see example 2.6).
Example 2.6: Consider the income vector x = ($2k, $4k, $8k, $10k)
used previously in the general means example. The arithmetic mean
is WA(x) = $6k, the geometric mean is WG(x) = $5.03k, the harmonic
mean is WH(x) = $4.10k, and WGM(x; –2) = $3.44k.
Thus,
IA(x; 0) = ($6k − $5.03k)/$6k = 0.162.
IA(x; −1) = ($6k − $4.10k)/$6k = 0.317.
IA(x; −2) = ($6k − $3.44k)/$6k = 0.427.
What is the interpretation of the number IA(x; 0) = 0.162? First, note

that IA(x; 0) is based on two income standards: the arithmetic mean of x
92
and the geometric mean of x. The arithmetic mean represents the level
of welfare obtained when the overall income is distributed equally across
everyone in the society. This is an ideal situation when there is no inequality
in the society.
The geometric mean, in contrast, is the equally distributed equivalent
(ede) income, which, if received by everyone in the society, would yield
the same welfare level as in x for the degree of inequality aversion a = 0. So
IA(x; 0) = 0.162 implies that the loss of welfare because of inequality in dis-
tribution x is 16.2 percent of what the welfare level would be if the overall
income had been equally distributed.
Suppose the society becomes more averse to inequality and a is reduced
to −1. In this case, the equally distributed equivalent income is the har-
monic mean of x. The loss of total welfare because of unequal distribution
increases from 16.2 percent to 31.7 percent. Likewise, the percentage loss
of welfare would increase to 42.7 percent if the society became even more
averse to inequality and a fell to −2.
What properties does any index in this family satisfy? Any measure in
this family satisfies all four invariance properties: symmetry, population invari-
ance, scale invariance, and normalization. In addition, unlike the quantile
ratios and the partial mean ratios, measures in this class satisfy the transfer
principle, transfer sensitivity, and subgroup consistency.
If distribution x' is obtained from distribution x by at least one regres-
sive transfer, then the level of inequality in x' is strictly higher than that
in x. Furthermore, if transfers take place between poor people, then the
inequality measure changes more than if the same amounts of transfers take
place among rich people. Finally, because these measures satisfy subgroup
consistency, they do not lead to any inconsistent results while decomposing
across subgroups. If inequality in certain subgroups increases while inequal-
ity in the others does not fall, then overall inequality increases. However,
measures in this class are not additively decomposable.
Gini Coefficient
The Gini coefficient, developed by Italian statistician Corrado Gini (1912),

is the most commonly used inequality measure. It measures the average dif-
ference between pairs of incomes in a distribution, relative to the distribu-
tion’s mean. The most common formulation of the Gini coefficient for the
distribution x is
93
N N
1
IGini (x) = ∑ ∑ x n − x n′ .
2N 2 × WA (x) n =1 n’=1
(2.22)
Note that equation (2.22) may be broken into two components: WA(x)
(the mean of the distribution) and (∑N n=1∑ n'=1|xn – xn'|)/2N (the average
N 2
difference between pairs of incomes). The second component is divided

by its number of elements, 2N2. There are 2N2 elements because each ele-
ment in x is compared with another element in x including itself twice.
This original Gini coefficient formula can be simplified further. The second
component of the Gini coefficient can be written as
1 N N 1 N N
∑ ∑ n n′ A
2 N 2 n =1 n ′ =1
x − x = W (x) − ∑ ∑ min {x n , x n′}= WA (x) − Ws (x), (2.23)
N 2 n = 1 n ′= 1
where WS(x) is the Sen mean of distribution x. Therefore, the Gini coefficient
may be simply formulated by using the arithmetic mean and the Sen mean.
Like any measure in Atkinson’s class, the Gini coefficient can be expressed as
WA (x) − WS (x) W (x)

IGini (x) = = 1− S . (2.24)
WA (x) WA (x)
Thus, the Gini coefficient is the difference between the arithmetic

mean and the Sen mean divided by the arithmetic mean. The coef-
ficient lies between zero and one, and inequality increases as the index
moves from zero to one. The minimum inequality level, zero, is obtained
when income is equally distributed across everyone in the society.
Like Atkinson’s measures, if IGini(x) = 0, then, by implication, income in
the society is equally distributed. Again, this is because any inequality
measure in this family is constructed by considering the entire distribution
(see example 2.7).
What is the interpretation of IGini(x) = 0.292? First, IGini(x) is based on
two income standards: the arithmetic mean of x and the Sen mean of x.
The arithmetic mean represents the level of welfare obtained when the
overall income is distributed equally across all people in the society. This is
an ideal situation when there is no inequality in the society. The Sen mean,
in contrast, is an ede income, which, if received by everyone in the society,
would yield the same welfare level as in x. So IGini(x) = 0.292 implies that
the loss of welfare because of inequality in distribution x is 29.2 percent of
94
that we used previously. First, we calculate the Gini coefficient using
the formulation in equation (2.22). It can be easily verified that
WA(x) = $6k. The second component is
1
( 2 − 2 + 2 − 4 + 2 − 8 + 2 − 10 + 4 − 2 + 4 − 4 + 4 − 8 + 4 − 10
2 × 42
+ 8 − 2 + 8 − 4 + 8 − 8 + 8 − 10 + 10 − 2 + 10 − 4 + 10 − 8 + 10 − 10 )
1
= (0 + 2 + 4 + 6 + 8 + 2 + 0 + 4 + 6 + 6 + 4 + 0 + 2 + 8 + 6 + 2 + 0)
32
56
= = 1.75.
32
Thus, IGini(x) = 1.75/6 = 0.292.

Next, we calculate the Gini coefficient using equation (2.24). The
Sen mean of distribution x is WS(x) = $4.25k. Thus, IGini(x) = ($6k −
$4.25k)/$6k = 1.75/6 = 0.292.
the welfare level if overall income had been equally distributed. We will see
later that the Gini coefficient has an interesting relationship with the well-
known Lorenz curve.
The Gini coefficient satisfies all invariance properties: symmetry, population
invariance, scale invariance, and normalization. In addition, it satisfies the transfer
principle. If distribution x' is obtained from distribution x by at least one regres-
sive transfer, then the level of inequality in x' is strictly higher than that in x.
However, the Gini coefficient is neither transfer sensitive nor subgroup con-
sistent. It is not transfer sensitive because the Gini coefficient changes by the
same amount whether transfers take place between poor people or between
rich people. That the Gini coefficient is not subgroup consistent means that if
the inequality in some subgroups increases while inequality in other subgroups
does not fall, then the overall inequality may register a decrease.
The following is an example showing that the Gini coefficient is neither
transfer sensitive nor subgroup consistent. Consider the vector x = ($4k,
$5k, $6k, $7k, $14k, $16k). If a progressive transfer of $0.5k takes place
between the first person and the second person, then x' = ($4.5k, $4.5k, $6k,
$7k, $14k, $16k). If a progressive transfer of the same amount takes place
95
between the two richer people, then x" = ($4k, $5k, $6.5k, $6.5k, $14k,
$16k). As a result, IGini(x') = IGini(x") = 0.279. Thus, the Gini coefficient
cannot distinguish between these two transfers.
The next example shows that the Gini coefficient is not subgroup consis-
tent. We use the same example that we used to show that the Sen mean does
not satisfy subgroup consistency. The original income vector x = ($4k, $5k,
$6k, $7k, $14k, $16k) becomes, over time, y = ($3.4k, $6.1k, $6k, $6.5k,
$14k, $16k). The income vector of the first subgroup x' = ($4k, $5k, $7k)
becomes y' = ($3.4k, $6.1k, $6.5k), whereas the income vector of the sec-
ond subgroup remains unaltered. The Sen mean of the first group falls from
WS(x') = $4.67k to WS(y') = $4.64k, whereas the mean income remains
unchanged at WA(x') = WA(y') = $5.33k. So the inequality of the first
group increases from IGini(x') = 0.125 to IGini(y') = 0.129. What happens
to the overall inequality? It turns out that the overall Sen mean increases
from WS(x) = $6.22k to WS(y) = $6.24k, whereas the overall mean income
remains unchanged at WA(x) = WA(y) = $8.67k. The overall inequality
decreases from IGini(x) = 0.282 to IGini(y) = 0.280.
However, unlike the Atkinson class of measures, the Gini coefficient is
additively decomposable, but with an added residual term. If distribution x is
divided into population subgroups x' with population size N' and x" with
population size N", then the decomposition formula of the Gini coefficient is
IGini(x) = w'IGini(x')+w" IGini(x") + IGini(x–', x– ") – residual, (2.25)
where the weights are w' = (N'/N)2(x̄'/x̄) and w" = (N"/N)2(x̄"/x̄). Note,
however, that the weights may not sum to one. The residual term is not zero
if and only if the groups’ income ranges overlap. If we consider the example
above, where the income vector x = ($4k, $5k, $6k, $7k, $14k, $16k) is
divided into two subgroup vectors: x' = ($4k, $5k, $7k) and x" = ($6k, $14k,
$16k). These vectors overlap as $7k > $6k. Thus, the residual term will
not vanish. However, if the two subgroups were x' = ($4k, $5k, $6k) and
x" = ($7k, $14k, $16k), then the residual term would be zero.9
Generalized Entropy Measures
The final inequality measures we consider are in the class of generalized

entropy measures. Two well-known Theil measures are also in this class. The
common formula for the generalized entropy measures of order a for any
distribution x is
96
⎧ 1 ⎡1 N ⎛ xn ⎞
a
⎤
⎪ ⎢ ∑ n =1 ⎜ ⎟ − 1⎥ if a ≠ 0,1
⎪ a(a − 1) ⎢⎣ N ⎝ x⎠ ⎦⎥
⎪
⎪1 N x ⎛x ⎞
IGE (x; a) = ⎨ ∑ n =1 n ln ⎜ n ⎟ if a = 1. (2.26)
⎪N x ⎝ x⎠
⎪1 ⎛ x⎞
⎪ ∑ n =1 ln ⎜ ⎟
N
if a = 0
⎪⎩ N ⎝ xn ⎠
At first glance, the formula above looks complicated. However, measures

in this class are closely related to general means. Every index in this class,
except one, can be expressed as a function of the arithmetic mean and the
general mean of order a. For a ≠ 0, 1, the class of generalized entropy mea-
sures can be written as
1 ⎛ ⎢⎣WGM (x; a ')⎥⎦a − ⎢⎣WA (x)⎥⎦a ⎞
IGE (x; a ) = ⎜ a ⎟, (2.27)
a (a − 1) ⎝ [ W (x)] ⎠
A
where we replace the term x̄ by WA(x) (the arithmetic mean), and where
WGM(x; a) denotes the general mean of order a. Thus, a generalized
entropy measure for any a ≠ 0,1 may be easily calculated once we know the
arithmetic mean and the general mean of order a.
For a = 1, the generalized entropy is Theil’s first measure of inequality
and can be written as
1 N xn ⎛ xn ⎞
IT1 (x) = ∑ ln ⎜
N n =1 WA (x) ⎝ WA (x) ⎟⎠
. (2.28)
This is the only measure in this class that cannot be expressed as a function
of general means and does not have a natural twin-standards representation.
For a = 0, the generalized entropy index is Theil’s second measure of
inequality, which is also known as the mean logarithmic deviation and can be
expressed as a function of the arithmetic mean, WA(x), and the geometric
mean, WG(x), as follows:
WA (x)
IT 2 (x) = ln WA (x) − ln WG (x) = ln . (2.29)
WG (x)
Besides the two Theil measures, the other commonly used measure in
the entropy class is the index for a = 2, which is closely related to the coef-
ficient of variation (CV). The CV is the ratio of the standard deviation and
97
mean. For a = 2, the general entropy measure is half the CV squared and
can be expressed as
2 2
1 ⎢⎣WE (x)⎢⎣ − ⎢⎣WA (x)⎢⎣ 1 Var(x) CV 2
IGE (x; 2) = 2
= 2
= , (2.30)
2 2 2
⎣⎢ WA (x)⎣⎢ ⎣⎢ WA (x)⎢⎣
where Var(x) is the variance of the distribution x, which is the square
of its standard deviation. In equation (2.29), WE(x) is the Euclidean
1
mean, as in equation (2.8) and [WE (x)]2 = ∑ n =1 x 2n . Clearly,
N
1 N
⎢⎣WE (x)⎢⎣ − ⎢⎣WA (x)⎢⎣ = ∑ n =1 x n − x is the variance of x (see example 2.8).
2 2 N 2 2
that we used in the general means example. The arithmetic mean
is WA(x) = $6k, the geometric mean is WG(x) = $5.03k, and the
Euclidean mean is WE(x) = $6.78k.
We now calculate the two Theil inequality measures and the
squared coefficient of variation:
IGE(x; 2) = ([WE(x)]2 − [WA(x)]2)/(2[WA(x)2] = (6.782 − 62)/(2 × 62)
= 0.279.
IT2(x) = ln[WA(x)/WG(x)] = ln [$6k/$5.03k] = 0.176.
The calculation of Theil’s first measure is not as straightforward
as that of the previous two measures. However, it can be calculated
using the following steps. First, create a new vector from vector x by
dividing every element by the mean of x as (2/6, 4/6, 8/6, 10/6). Then
1 ⎡ 2 ⎛ 2 ⎞ 4 ⎛ 4 ⎞ 8 ⎛ 8 ⎞ 10 ⎛ 10 ⎞ ⎤
IT1 (x) = ln ⎜ ⎟ + ln ⎜ ⎟ + ln ⎜ ⎟ + ln ⎜ ⎟ = 0.15.
4 ⎢⎣ 6 ⎝ 6 ⎠ 6 ⎝ 6 ⎠ 6 ⎝ 6 ⎠ 6 ⎝ 6 ⎠ ⎥⎦ .
Having introduced the measures in the generalized entropy class, now

we try to understand their behavior. First, what is the range of any measure
in this class? The lower bound of any measure in this class is zero, which is
obtained when incomes in a society are equally distributed across all people.
However, unlike the Atkinson’s measures and the Gini coefficient, general-
ized entropy measures may not necessarily be bounded above by one.
Next, how do the measures in this class relate to the parameter? There
are, in fact, three distinct ranges: a lower range a < 1, an upper range a > 1,
98
and a limiting case where a = 1. For the lower range, a < 1, measures in this
class are monotonic transformations of the Atkinson’s class of measures and
can be written as
⎧ ⎢⎣1 − IA (x ;a − 1)⎥⎦a
⎪ if a ≠ 0, a < 1
⎪ a (a − 1)
IGE (x; a ) = ⎨ , (2.31)
⎪ ln 1
if a = 0
⎪⎩ 1 − IA (x; 0)
where IA(x; a) is the Atkinson’s inequality measure for parameter a.

For the range a < 1, the entropy measures behave the same way as the
Atkinson’s measures. Over the range a > 1, the general mean places
greater weight on higher incomes and yields a representative income that
is typically higher than the mean income. An example is the squared coef-
ficient of variation.
All measures in the generalized entropy class satisfy the invariance
properties: symmetry, normalization, population invariance, and scale invari-
ance. Furthermore, they all satisfy the transfer principle and subgroup con-
sistency. However, transfer sensitivity is satisfied only by the measures in
this class with a < 2. Measure IGE(x; 2) is, in fact, transfer neutral like
the Gini coefficient. It turns out that the generalized entropy measures
are the only inequality measures that satisfy the usual form of additive
decomposability (see Shorrocks 1980). If distribution x is divided into
two population subgroups, x' with population size N' and x" with popula-
tion size N", then the decomposition formula of the generalized entropy
measure for a ≠ 0,1 is
IGE(x; a) = w'IGE(x'; a) + w"IGE(x"; a) + IGE(x̄', x̄"; a), (2.32)
where the weights are w' = (N'/N)(x̄'/x̄)a and w" = (N"/N)(x̄"/x̄)a. For
example, when a = 2, the weights are w' = (N'/N)(x̄'/x̄)2 and w" = (N"/N)
(x̄"/x̄)2.
Note that the weights may not always sum to one. However, for the two
Theil measures, the weights do sum to one. The first Theil measure can be
decomposed as
IT1(x) = w'IT1(x') + w"IT1(x") + IT1(x̄', x̄"), (2.33)
where the weights are w' = x̄'/x̄ and w" = x̄"/x̄. Although it is difficult to
get an intuitive interpretation of the first Theil measure, the additive
99
decomposability property makes the first Theil measure useful in under-

standing within-group and between-group inequalities. The second Theil
measure can be decomposed as
IT2(x) = w'IT2(x') + w"IT2(x") + IT2(x̄', x̄"), (2.34)
where the weights are w' = (N'/N) and w" = (N"/N).
Inequality and Welfare
The Gini coefficient and the inequality measures in Atkinson’s family share
a social welfare interpretation. As we have already discussed, they can be
expressed as I = (x̄ − a)/x̄, where x̄ is the mean income of the distribution
x and a is an income standard that can be viewed as a welfare function
(satisfying the weak transfer principle). Note that the distribution in which
everyone has the mean income has the highest level of welfare among all
distributions with the same total income, and the distribution’s measured
welfare level is just the mean itself. This finding results from the normaliza-
tion property of income standards.
Thus, the mean WA(x) = x̄ is the maximum value that the welfare func-
tion can take over all income distributions of the same total income. When
incomes are all equal, a = WA(x) and inequality is zero. When the actual
welfare level a falls below the maximum welfare level WA(x), the percentage
welfare loss I = (WA(x) − a)/WA(x) is used as a measure of inequality. This
is the welfare interpretation of both the Gini coefficient and the Atkinson’s
class of measures.
The simple structure of these measures allows us to express the welfare
function in terms of the mean income and the inequality measure. A quick
rearrangement leads to a = WA(x)(1 – I), which can be reinterpreted as
saying that the welfare function a can be viewed as an inequality-adjusted
mean. If there is no inequality in the distribution, then (1 – I) = 1 and
a = WA(x). If the inequality level is I > 0, then the welfare level is obtained
by discounting the mean income by (1 – I) < 0.
For example, if we take I to be the Gini coefficient, IGini(x), then the Sen
mean (or Sen welfare function) can be obtained by multiplying the mean by
[1 – IGini(x)], that is, WS(x) = WA(x)[1 – IGini(x)]. Similarly, if we take I to
be the Atkinson’s measure with parameter a = 0, IA(x; 0), then the welfare
function is the geometric mean, and the geometric mean can be obtained by
multiplying the mean by [1 – IA(x; 0)], that is, WG(x) = WA(x)[1 – IA(x; 0)].
100
An inequality measure estimates, with a single number, the inequality level

in a society. A question may naturally arise: Do all inequality measures
compare two income distributions in the same way? In other words, if an
inequality measure evaluates income distribution x to be more equal than
distribution y, would another inequality measure evaluate distributions x
and y in the same way? The answer depends on the two inequality measures
we use for evaluation. Not all inequality measures evaluate various distribu-
tions in the same manner.
We can clarify this concern with an example. Consider the two income
vectors x = ($4k, $5k, $6k, $7k, $14k, $16k) and y = ($3.4k, $6.1k, $6k,
$6.5k, $14k, $16k). These two vectors have the same mean. The Gini coef-
ficient indicates that the inequality level in x is 0.282, which is higher than
the inequality in y (0.280). However, the Atkinson’s measure that is based
on the geometric mean shows that the inequality level in x is 0.127, which is
lower than the level of inequality in y (0.132). Therefore, different inequal-
ity measures may disagree in different situations.
Is there any condition in which different inequality measures agree with
each other? The answer is yes. Inequality measures that satisfy the four
basic properties—symmetry, population invariance, scale invariance, and
the weak transfer principle—agree with each other when Lorenz dominance
holds between two distributions. To understand Lorenz dominance, we need
to understand the Lorenz curve.
The Lorenz curve of an income distribution shows the proportion of total
income held by the poorest p percent of the population.10 We denote the
Lorenz curve of distribution x by Lx. Then Lx(p) is the share of total income
held by the poorest p percent of the population. Indeed, Lx(100) = 100
percent and Lx(0) = 0 percent. Suppose the total income of Nigeria is N25
trillion and only N1 trillion is received by the poorest 20 percent of the
population. Then LNig(20) = 4 percent. Suppose that income in Nigeria
is redistributed, keeping the total income unaltered, so that everyone has
identical income. Let us denote the equal income distribution by y. Then
the percentage of total income enjoyed by the poorest 20 percent of the
population is 20 percent, and Ly(20) = 20 percent.
In figure 2.13, the horizontal axis denotes the cumulative share of the
population (p), and the vertical axis shows the share of total income.
Note that the lowest and the highest values for both axes are 0 and 100,
101
Figure 2.13: Lorenz Curve
Lx(100) 100
Ly Lx ′
Income share
Lx
Ly(20) A 20
Lx ′(20) B 14
Lx(20) C 4
0
0 20 100
Population share
respectively. For income distribution x, Lx represents its Lorenz curve,

denoted by the dotted curve. Following the example of Nigeria, Lx(20) = 4
percent, which is the height of the curve Lx at point C.
If distribution y is obtained from distribution x by distributing income
equally across the population, then the Lorenz curve becomes a 45-degree
straight line, Ly (the solid line in figure 2.13). In this case, the share of the
population’s bottom 20 percent in distribution y is Ly(20) = 20 percent. This
is obtained at point A on Lorenz curve Ly.
Now, suppose the income distribution in Nigeria improves over time
and the new income distribution is denoted by x'. The Lorenz curve for x' is
denoted by the dashed curve Lx' in figure 2.13. The share of the bottom 20
percent in the total income increases from 4 percent to 14 percent. This is
shown at point B on the Lorenz curve Lx'.
Notice that every portion of Lorenz curve Lx' lies above that of Lorenz
curve Lx. This is what we mean by Lorenz dominance: the income share
of every cumulative population share in x' is higher than that in x. Thus,
distribution x' Lorenz dominates distribution x'. Similarly, distribution x
Lorenz dominates both distributions x and x'.
Any inequality measure satisfying the four basic properties—symmetry,
population invariance, scale invariance, and the weak transfer principle—
would evaluate distribution y as more equal than distributions x and x' and
distribution x' as more equal than distribution x. Thus, before comparing
distributions using different inequality measures, the distributions’ Lorenz
102
curves should be compared. If one distribution’s Lorenz curve dominates

that of another distribution, then all inequality measures satisfying these
four basic properties would consider the former distribution to be more equal
than the latter.
Well-known inequality measures satisfying these four basic properties are
the Gini coefficient, measures in the Atkinson’s family, measures in the gen-
eralized entropy family, and partial mean ratios. What happens when two
Lorenz curves cross? In this situation, Lorenz dominance does not hold, and
the inequality level needs to be judged using inequality measures when dif-
ferent inequality measures may agree or disagree with each other. However,
even in this case, the Lorenz curve can be helpful in identifying the winning
and losing portions of the distribution.
The Lorenz curve also has interesting relationships with income stan-
dards and inequality measures. First, consider its relationship with the
generalized Lorenz curve. A Lorenz curve may be obtained from a general-
ized Lorenz curve by dividing the latter by the mean. Thus, for distribu-
tion x, Lx(p) = GLx(p)/WA(x). The construction of a Lorenz curve can be
easily understood by following the construction of the generalized Lorenz
curve in figure 2.8. Next, recall that the height of the generalized Lorenz
curve at a certain percentile of population p is the lower partial mean
times p itself, that is, GLx(p) = p × WLPM(x; p). Therefore, the height
of the Lorenz curve at a certain percentile of population p is the ratio of
the lower partial mean to the overall mean times p itself, that is, Lx(p) = p ×
[WLPM(x; p)/WA(x)]. Note that the ratio of the lower partial mean to the
overall mean itself may be used to construct a partial mean ratio, denoted
by IPMR(x; 100/p).
Finally, an interesting relationship exists between the Lorenz curve and
the Gini coefficient. The Gini coefficient of distribution x is twice the area
between the Lorenz curves Lx and Ly in figure 2.13. Similarly, the Gini coeffi-
cient for distribution x" is twice the area between the Lorenz curves Lx' and Ly.
Inequality and Growth
The twin-standard view of inequality offers fresh insights into the relation-
ship between growth and inequality. Almost all inequality measures are
constructed in terms of a larger income standard b and a smaller income
standard a, and these income standards are expressed as 1 – a/b or b/a – 1.
Suppose income standard a changes to a' over time with growth rate g– a,
103
that is, a' = (1 + g– a)a, and income standard b changes to b' over time with
growth rate g– b, that is, b' = (1 + g– b)b. The inequality measure then changes
from I = 1 – a/b to I' = 1 – a'/b'. To have a fall in inequality, we require I' <
I or 1 – a'/b' < 1 – a/b, which occurs when g– a > g– b. Therefore, for a reduction
in inequality, the smaller income standard a needs to grow faster than the
larger income standard b.
Consider the example of the Gini coefficient, which is constructed from
two income standards. The larger income standard is the arithmetic mean
WA, and the smaller income standard is the Sen mean WS. Let us denote
the growth rate of the mean income by g– and the growth rate of the Sen
mean by g–S. The Gini coefficient will register a fall in inequality when the
growth rate of the Sen mean is larger than the growth rate of the arithmetic
mean, that is, g–S > g–. Similarly, inequality over time, in terms of the Gini
coefficient, increases when g–S < g–.
What about the Atkinson’s measures and the generalized entropy mea-
sures? Measures in these classes, including Theil’s second measure, are based
on the arithmetic mean and on any income standard from the class of gen-
eral means. For a < 1, the arithmetic mean is the larger income standard,
and the other general mean–based income standard is the smaller income
standard. In this case, if the growth rate of the smaller income standard of
order a is denoted by g–GM(a), then inequality decreases when g–GM(a) > g–.
If inequality is evaluated by Theil’s second index, then inequality falls when
the growth of geometric mean g–GM(0) is larger than that of the arithmetic
mean, that is, g–GM(0) > g–. For a > 1 in the generalized entropy measure,
the arithmetic mean is the smaller income standard, and the other general
mean–based income standard is the larger one. Inequality falls, according
to these indices, when the growth rate of the arithmetic mean g– is higher.
Is there any way to tell if all inequality measures in the Atkinson family
and the generalized entropy family have fallen? Yes, it is possible to do so
just by looking at the general mean growth curve, as described in figure 2.12.
A generalized mean growth curve is the loci of the growth rates of all
income standards in the class of general means. Comparing distributions x
and y for the general mean growth curve gGM(x,y) in figure 2.12 shows that
all inequality measures in Atkinson’s class and the generalized entropy class
agree that the inequality has fallen because the growth rates of the lower
income standards are higher than ḡ. The growth rates of the larger income
standards are lower than ḡ. However, for the general mean growth curve
gGM(x',y') in the same figure, all inequality measures in Atkinson’s class and
104
the generalized entropy class agree that the inequality has risen because the
growth rates of the lower income standards are lower than ḡ, whereas the
growth rates of the larger income standards are higher than ḡ.
In a similar manner, the growth incidence curve may be used to under-
stand the change in inequality using quantile ratios. If the growth rate of
the upper quantile income is larger than the growth rate of a lower quantile
income, then inequality has risen over time. In contrast, if the growth rate
of a lower quantile income is larger than the growth rate of the higher quan-
tile income, then inequality has fallen. For example, consider the growth
incidence curve gQI(x,y) in figure 2.10. If inequality is measured by the
90/10 measure IQR(x; 90/10), then inequality has fallen. Furthermore, for
growth incidence curve gQI(x',y'), the level of inequality has increased for
the same inequality measure.
Poverty Measures
The third aspect of a distribution is base, which is evaluated using a numeri-

cal poverty measure, assigning each distribution a number reflecting its
level of deprivation. In this section, before proceeding further, we introduce
additional notations that are more specific to poverty measures than income
standards and inequality measures. The income distribution of society X
with N people can be summarized by the vector x = (x1,x2, …, xN), where
xn is the income of person n. We also assume that the income distribution is
ordered, that is, x1 ≤ x2 ≤ … xN.
Any poverty measure is constructed in two steps. The first step is iden-
tification, where each person is identified as poor or nonpoor by using a
threshold called the poverty line, denoted by z. More specifically, a person
is identified as poor if his or her income falls below the poverty line z and
nonpoor if his or her income is greater than or equal to z. We denote the
number of poor in our reference society X by q. So the number of nonpoor
is N − q. Because elements in income distribution x are ordered, people 1,…,
q are poor and people q + 1, …, N are nonpoor.
Suppose society X consists of four people with the income vector x = ($1k,
$2k, $50k, $70k). If the poverty line is set at $10k, this means that a person
must have $10k to meet the minimum necessities to lead a healthy life. This
requirement would identify the first two people as poor with earnings $1k
and $2k, whereas the third person and the fourth person are identified as
105
nonpoor. In this example, society X has two poor people and two nonpoor
people. We summarize the incomes of the poor in vector x by the vector xq.
Poverty analysis is concerned only with the poor or the distribution’s
base, which should be the group targeted for public assistance. It naturally
ignores the incomes of nonpoor people in a society. In this way, the identifi-
cation step allows us to construct a censored distribution or censored vector
of incomes for society X, which we denote by x* = (x*1,x*2, …,x*N) such that
x*n = xn if income xn is less than the poverty line z and xn* = z if income xn is
greater than or equal to the poverty line z.
For the four-person income vector x = ($1k, $2k, $50k, $70k) in the
previous example, the censored vector is denoted by x* = ($1k, $2k, $10k,
$10k). Notice that incomes of the two nonpoor people are replaced by
the poverty line, and portions of their income above the poverty line are
ignored. A policy maker’s objective should be to include poor people at or
above the poverty line. Including all poor people at or above the poverty
line results in a nonpoverty censored distribution of income. We denote the
nonpoverty censored distribution of society X corresponding to poverty line
z by x– z* such that x– z* = (z,z,…,z).
The second step for constructing a poverty measure is aggregation. In this
step, incomes of individuals who are identified as poor using the poverty
line in the identification stage are aggregated to obtain a poverty measure.
Therefore, a poverty measure depends on both the incomes of the poor and
the criterion that is used for identifying the poor—that is, the poverty line.
In fact, it turns out that any poverty measure is obtained by aggregating ele-
ments in the censored distribution x∗.
In this section, we denote a poverty measure by P, where specific indi-
ces are denoted using corresponding subscripts. We denote the poverty
measure of distribution x for poverty line z by P(x; z). Alternatively, it may
be denoted by P(x∗). There are two different ways to understand a poverty
measure: one is based on the properties it satisfies and the other is through
its link with income standards. First, we discuss the properties that a poverty
measure should satisfy.
A useful poverty measure should satisfy some desirable properties. Like

income standards and inequality measures, poverty measure properties can
fall into two categories:
106
• Invariance properties leave poverty measures invariant to certain

changes in the dataset. Properties in the invariance category are sym-
metry, normalization, population invariance, scale invariance, and focus.
• Dominance properties cause a poverty measure to change in a particu-
lar direction. Properties in the dominance category are monotonicity,
transfer principle, transfer sensitivity, and subgroup consistency.
Six of these properties—symmetry, population invariance, scale invariance,

focus, monotonicity, and transfer principle—are called basic properties. Many
of these properties are analogous to the corresponding properties of income
standards and inequality measures.11
The first invariance property, symmetry, requires that switching the
income levels of two people while the poverty line remains the same leaves
poverty unchanged. In other words, a person should not be given priority on
the basis of his or her identity when evaluating the level of poverty within
a society. Formally, it requires that the poverty measure of distribution x be
equal to the poverty measure of another distribution x', if x' is obtained from
x by a permutation of incomes without changing the poverty line.
For example, recall the four-person income vector ($1k, $2k, $50k,
$70k). If the poverty line is z = $10k, then the first two people are poor and
the last two people are nonpoor. Now, if the income of the first and the
fourth individuals are switched, the new income vector becomes x' = ($70k,
$2k, $50k, $1k). This new vector x' is said to be obtained from vector x by
a permutation of incomes.
Symmetry: If distribution x' is obtained from distribution x by

permutation of incomes and the poverty line z remains the same,
then P(x'; z) = P(x; z).
The second invariance property, normalization, requires that the poverty

measure be zero if no one’s income in the society is less than the poverty
line. This is a natural property. For example, if the income vector of the
four-person society is ($1k, $2k, $50k, $70k), but the poverty line in this
case is $1k, then any poverty measure should be 0, reflecting that there are
no poor in the society.
Normalization: For any income distribution x and poverty line z, if

min{x} ≥ z, then P(x; z) = 0.
107
The third invariance property, population invariance, requires that pov-

erty be invariant to the population size, in the sense that a replication of
an income vector results in the same level of poverty as the original sample
vector if the poverty line does not change. The implication of this property
is as follows. Consider the income vector of society X, x = ($1k, $2k, $50k,
$70k). Suppose four more people with the same income distribution join the
society so that the new income vector is x' = ($1k, $1k, $2k, $2k, $50k, $50k,
$70k, $70k). The population invariance property requires that the poverty
level in society x remains unaltered, at least if the poverty line does not
change. This allows us to compare the extent of poverty across countries and
across time with varying population sizes. Furthermore, this property allows
any poverty measure to depend on a distribution function, which normalizes
the population size to one.
Population Invariance: If vector x' is obtained by replicating

vector x at least once and the poverty line remains unaltered, then
P(x'; z) = P(x; z).
The fourth invariance property, scale invariance, requires that if an

income distribution is obtained from another income distribution by
scaling all incomes and the poverty line by the same factor, then the pov-
erty level should remain unchanged. For example, if everyone’s income
and the poverty line in a society are tripled or halved, then the level of
deprivation of the society does not change. The scale invariance prop-
erty ensures that the measure is independent of the unit of measurement
for income. Consider the following example, where the income of each
person in vector x = ($1k, $2k, $50k, $70k) increases by three times and
becomes x' = ($3k, $6k, $150k, $210k) over time. If the poverty line also
increases from, say, $6k to $18k, then the level of deprivation should not
change over time.12
Scale Invariance: If distribution x' is obtained from distribution x

such that x' = cx and z' = cz where c > 0, then P(x'; cz) = P(x; z).
The fifth and final axiom in the invariance properties is focus, which
requires that if the income of a nonpoor person in a society changes but
does not fall below the poverty line, then the level of poverty should not
108
change. This property ensures that the measure focuses on the poor incomes
in evaluating poverty. In fact, focus ensures that the income distribution
is censored at the poverty line before evaluating a society’s poverty. For
example, suppose the initial income vector is x = ($1k, $2k, $50k, $70k)
and the poverty line income is $6k. Thus, the third person and the fourth
person are nonpoor. If the income of either the third or the fourth person
increases, but the poverty line remains unaltered at $6k, then the society’s
poverty level does not change.
Focus: If distribution x' is obtained from distribution x by increasing

the income of a nonpoor person while the poverty line remains the
same at z, then P(x'; z) = P(x; z).
The next group of properties are dominance properties. The first of these
properties requires that if the income of a poor person in a society increases,
then the poverty level should register a fall, or at least it should not increase.
There are two versions of this property. One is weak monotonicity, which
requires that poverty should not increase because of an increase in a poor
person’s income. The other is monotonicity, the stronger version, which
requires that poverty should fall if a poor person’s income in the society
increases.
These two properties are the same as the two corresponding properties
of income standards, except the ones introduced here are solely concerned
with incomes of the poor. For example, suppose the initial income vector
is x = ($1k, $2k, $50k, $70k) and the poverty line income is $6k so that
the first two people are identified as poor. If a new vector x' is obtained by
increasing the income of either the first or the second person, while the
poverty line remains unchanged, then according to the weak monotonicity
property, poverty should not be higher in x', and, according to the monoto-
nicity property, poverty should be lower in x'.
Weak Monotonicity: If distribution x' is obtained from distribution x

by increasing the income of a poor person while keeping the poverty
line unchanged at z, then P(x'; z) ≤ P(x; z).
Monotonicity: If distribution x' is obtained from distribution x by
increasing the income of a poor person while keeping the poverty
line unchanged at z, then P(x'; z) < (x; z).
109
The second dominance property is the transfer principle, which requires

that a regressive transfer between two poor people in a society increase pov-
erty and a progressive transfer between two poor people reduce poverty.13
(For definitions of regressive and progressive transfers, refer to the section
discussing the transfer principle for income standards.) Suppose the initial
income vector is x = ($1k, $2k, $50k, $70k) and the poverty line income
is $6k, so the first two people are poor. If a new vector x' is obtained by a
progressive transfer between the first and the second person such that x'=
($1.5k, $1.5k, $50k, $70k) and the poverty line is still fixed at $6k, then pov-
erty in x' should be lower. Note that the transfer principle property allows
the number of poor to change as a result of a regressive transfer because the
richer poor may become nonpoor because of a regressive transfer.14
Transfer Principle: If distribution x' is obtained from distribution x

by a regressive transfer between two poor people while the poverty
line is fixed at z, then P(x'; z) > P(x; z). If distribution x" is obtained
from another distribution x by a progressive transfer between two
poor people while the poverty line is fixed at z, then P(x"; z) < P(x; z).
As in inequality measurement, we also define a weaker version of trans-

fer principle in poverty measurement. It requires that a regressive transfer
between two people in a society not decrease poverty and a progressive
transfer between two people not increase poverty. Thus, the weaker prin-
ciple allows the possibility that the poverty level may remain unchanged
because of a progressive or a regressive transfer.
Weak Transfer Principle: If distribution x' is obtained from

distribution x by a regressive transfer between two poor people while
the poverty line is fixed at z, then P(x'; z) ≥ P(x; z). If distribution
x" is obtained from another distribution x by a progressive transfer
between two poor people while the poverty line is fixed at z, then
P(x"; z) ≤ P(x; z).
The transfer principle requires a poverty measure to decrease if the trans-

fer is progressive. However, it is not concerned with which part of the dis-
tribution the transfer is taking place. A same amount of transfer may take
place between two extremely poor people, who are further away from the
110
poverty line, or between two moderately poor people, who are much closer
to the poverty line.
Should the effect of transfer, no matter where it takes place, have
the same effect on the poverty level? We elaborate this situation with an
example. Consider the five-person income vector x = ($80, $100, $800,
$50, 000, $70,000). Let the poverty line be set at $1,050. Then the first four
people are identified as poor because their incomes are below the poverty
line. First, suppose $10 is transferred from the second person to the first per-
son. Then the post-transfer income vector is x' = ($90, $90, $800, $1,000,
$50,000, $70,000). Transferring 10 percent of the second person’s income
has increased the first person’s income by 12.5 percent.
Suppose, instead, that the same $10 transfer takes place between the
third and the fourth persons, who are also poor. The post-transfer income
vector is x'' = ($80, $100, $810, $990, $50,000, $70,000), where transfer-
ring 1 percent of the fourth person’s income increases the third person’s
income by 1.25 percent. This transfer makes hardly any difference in the
large pool of income of the two richer poor people. Therefore, one might
feel that a transfer of the same amount between two extreme poor and
two richer poor should not have the same effect on the society’s overall
poverty.
The third dominance property, transfer sensitivity, requires a poverty
measure to be more sensitive to a transfer between poor people at the lower
end of the income distribution of the poor. In other words, this property
requires that a poverty measure should change more when a transfer takes
place between two extremely poor people than between two richer poor
people. In terms of the example above, the level of deprivation should be
lower in x' than in x''.
Suppose the initial income distribution is x and distribution x" is obtained
from distribution x by a progressive (or regressive) transfer between two
extremely poor people. Suppose further that distribution x" is obtained from dis-
tribution x by a progressive (or regressive) transfer of the same amount between
two richer poor people. The following is the transfer sensitivity property:
Transfer Sensitivity: A poverty measure that satisfies transfer

sensitivity places greater emphasis on progressive (or regressive)
transfers at the lower end of the distribution of the poor than at the
upper end of the distribution of the poor; so P(x'; z) < (>) P(x"; z).
111
The final dominance property is subgroup consistency, which is concep-

tually the same as the corresponding property for income standards and
inequality measures. This property requires that if subgroup population sizes
are fixed, then overall inequality must rise when poverty rises in one sub-
group and does not fall in the rest of the subgroups. For example, suppose
that income vector x with population size N is divided into two subgroup
vectors: x' with population size N' and x" with population size N" such
that N' + N" = N. Let a new vector, y, be obtained from x with the same
population size N, and let its two corresponding subgroups be denoted by
y' with population size N' and y" with population size N". The subgroup
consistency property can be stated as follows:
Subgroup Consistency: Given that subgroup population sizes remain

unchanged, if P(y';z) > P(x';z) and P(y";z) ≥ P(x";z), then P(y;z) > P(x;z).
There is a property closely related to subgroup consistency that is often

useful for understanding how much of the overall poverty is attributed to
the poverty of a particular group, given a collection of population subgroups.
For example, a country’s population may be divided into subgroups such as
rural and urban areas, states, provinces, and other geographic regions; ethnic
and religious groups; genders; or age groups. Often, one may want to evalu-
ate a particular group’s contribution. The additive decomposability property
requires that overall poverty is expressed as a population-weighted average
of subgroup poverty levels. This property is similar in spirit to the corre-
sponding properties of income standards and inequality measures. However,
it is more analogous to that of income standards in the sense that there are
no within-group and between-group terms as we see for a decomposable
inequality measure.
To formally outline the property, we will use two groups to simplify
the interpretation, but the definition can be extended to any number of
groups. Suppose income vector x with population size N is divided into two
subgroup vectors: x' with population size N' and x" with population size N"
such that N' + N" = N. The additive decomposability property can be stated
as follows (see example 2.9):
Additive Decomposability: If income distribution x is divided into two

′ ′
subgroup distributions x' and x", then P(x ′; z) = N P(x ′; z) + N P(x′′; z).
N N
112
Example 2.9: Consider the six-person income vector x = ($80, $100,

$800, $1,000, $50,000, $70,000), which is divided into two subgroups
x' = ($80, $100, $50,000) and x" = ($800, $1,000, $70,000). Suppose
the poverty line is z = $1,100, which is the same across both subgroups.
Note that N' = 3, N" = 3, and N = 6, and, thus, N'/N = N"/N = 3/6
= 0.5. Then any additively decomposable poverty index can be
expressed as P(x;$1,100) = 0.5P(x';$1,100) + 0.5P(x";$1,100).
Poverty and Income Standards
The second way of understanding poverty measures is through the income

standards discussed earlier. Like inequality measures, most poverty measures
are based on a comparison between two income standards: a higher income
standard b and a lower income standard a. However, there is a crucial dif-
ference between inequality measures and poverty measures. In inequality
measures, the higher and lower income standards are two different income
standards applied to the same income vector. In poverty measures, the
higher and lower income standards are the same income standards applied
to two different income vectors: one is the censored distribution and the other
is the nonpoverty censored distribution. Recall that a censored distribution is
obtained from an original income distribution by replacing the income of the
nonpoor by the poverty line. The nonpoverty censored distribution is that
income distribution where all incomes are equal to the poverty line income.
It turns out that the higher income standard for poverty measures is the
poverty line itself. Why is that so? This can be understood by the normaliza-
tion property of income standards, which requires that if all incomes are equal
in an income distribution, then an income standard of the distribution should
be equal to that commonly held income. Because in a nonpoverty censored
income distribution all incomes are equal to the poverty line, any income
standard of the nonpoverty censored distribution should be equal to the pov-
erty line itself, that is, b = z. Many well-known poverty measures take the form
P = (z − a)/z or the form P = a/z or a monotonic transformation of either form.
Commonly Used Poverty Measures
In this section, we introduce various poverty measures that are in com-

mon use. We classify them into two categories. The first category lists basic
113
poverty measures, and the second category lists advanced poverty measures.
There are two basic poverty measures in common use: headcount ratio and
poverty gap measure.
Headcount Ratio
The headcount ratio (PH) is a crude measure of poverty that simply counts
the number of people whose incomes are below the poverty line z and
divides that number by the total number of people in the society. In society
X with population size N, if there are q poor people, then the headcount
ratio is simply q/N. It is obvious that the headcount ratio lies between zero
and one. If all people are poor in a society, then the headcount ratio is one.
When there are no poor, it is zero.
The headcount ratio can also be understood using income standards
applied to the nonpoverty censored distribution and a doubly censored dis-
tribution. What is a doubly censored distribution, and how do we obtain
it? A doubly censored distribution x** is obtained from an original income
distribution x by replacing nonpoor incomes with the poverty line income z
and by replacing the poor incomes with zero. Therefore, income distribution
x is censored upward at poverty line z for nonpoor and again censored at zero
for the poor. The term doubly censored comes from the fact that distribution
x*z* is obtained by censoring distribution x twice.
The arithmetic mean is the income standard used to understand head-
count ratio. The arithmetic mean of the nonpoverty censored distribution is
poverty line z, and the arithmetic mean of the doubly censored distribution
is called the dichotomous mean. If there are q poor people, or N − q nonpoor
people, in society X, then the dichotomous mean of the society is
N− q N− q
WA (x ** ) = q × 0 + z= z. (2.35)
N N
The headcount ratio of distribution x is a normalized shortfall of the

dichotomous mean from the mean of the nonpoverty censored distribution
(see example 2.10). Thus, the headcount ratio can be expressed as
N−q
WA (xz* ) − WA (x** ) z− z
N q
PH (x; z) = *
z
= = . (2.36)
WD (xz ) z N
114
Example 2.10: How is the headcount ratio calculated by different

methods? Consider the four-person income vector x = ($800, $1,000,
$50,000, $70,000). If the poverty line is set at z = $1,100, then two of
the four people are poor. Thus, the headcount ratio is PH(x;z) = 2/4 = 0.5
or 50 percent.
How can the headcount ratio be calculated using the concept of
doubly censored distribution?
The doubly censored vector of x is x*z* = (0, 0, $1,100, $1,100)
and the nonpoverty censored distribution is x̄*z = ($1,100, $1,100,
$1,100, $1,100).
Then WA(x̄*z) = 4 × $1,100/4 = $1,100 and WA(x*z*) = 2 ×
$1,100/4 = $550.
Hence, PH(x;z) = ($1,100 − $550)/$1,100 = 0.5.
The headcount ratio is the most well-known and most widely used
poverty measure because its interpretation is highly intuitive and simple.
However, the effectiveness of the headcount ratio depends on which prop-
erties the headcount ratio satisfies. It satisfies all invariance properties:
symmetry, normalization, population invariance, scale invariance, and focus.
However, it does not satisfy any dominance property except subgroup consis-
tency. The headcount ratio is not sensitive to changes in the income level
of the poor as long as incomes do not cross the poverty line. This is why
the headcount ratio does not satisfy the other dominance properties and
monotonicity, which require poverty measures to change as the incomes of
the poor change. The headcount ratio satisfies subgroup consistency because
the headcount ratio is additively decomposable, as shown by example 2.11.
Poverty Gap Measure
The second basic poverty measure is the poverty gap measure. Like headcount
ratio, it is also widely used. The poverty gap measure (PG) is the average
normalized shortfall with respect to the poverty line across the poor. In
society X, the normalized income shortfall of a person, say, n, is calculated as
(z − x*n)/z, which means that the normalized income shortfall of a nonpoor
person is zero. The average normalized income shortfall is the average of all
normalized income shortfalls within a society. We denote the normalized gap
vector of x by g* = ((z − x*1)/z,…,(z − x*N)/z). Then the poverty gap measure is
115
Example 2.11: Consider the six-person income vector x = ($80, $100,

$800, $1,000, $50,000, $70,000), which is divided into two subgroups
x' = ($80, $100, $800) and x" = ($1,000, $50,000, $70,000).
Suppose the poverty line, z = $1,100, is the same across both
subgroups.
Note that N' = 3, N" = 3, and N = 6; thus, N'/N = N"/N = 3/6 = 0.5
is the population share of each group.
The headcount ratio of x is PH(x;z) = 4/6 = 2/3; the headcount
ratio of x' is PH(x';z) = 3/3 = 1; and the headcount ratio of x" is
PH(x";z) = 1/3.
Thus, the overall headcount ratio may be obtained from the sub-
group headcount ratios. The population-weighted average headcount
ratio of the subgroups is 0.5P(x';z) + 0.5P(x";z) = 0.5 × 1 + 0.5 × 1/3
= 2/3.
1 N z − x *n
PG (x; z) = WA (g * ) = ∑
N n =1 z
. (2.37)
The poverty gap measure may also be understood and interpreted by

using two income standards. The higher income standard is the poverty line
z itself, obtained by taking an arithmetic mean of the nonpoverty censored
distribution x̄*z . The lower income standard is obtained by applying the
arithmetic mean to the censored income distribution x*. Thus, the poverty
gap measure can be expressed as
WA (xz* ) − WA (x * ) z − WA (x * ) 1 N z − x *n
PG (x; z) = = = ∑ . (2.38)
WA (xz* ) z N n =1 z
There is a third way to interpret the poverty gap measure, which is as a

product of the headcount ratio and the average normalized income shortfall
among the poor. The average normalized income shortfall among the poor
1 q
is PIG (x; z) = ∑ n =1(z − x n )/z. The poverty gap measure can be expressed as
q
N−q q 1 q z − xn
PG (x; z) = ×0+ × ∑ = PH × PIG (x; z). (2.39)
N N q n =1 z
The poverty gap measure lies between zero and one. Zero is obtained
when there are no poor in the society. A value of one is obtained when
116
everyone in the society is poor and has zero income. When everyone in
a society is poor, then the poverty gap measure is the average normalized
income shortfall among the poor, PIG, because the headcount ratio is one in
this situation, that is, PH = 1 (see example 2.12).
Example 2.12: How is the poverty gap measure calculated by different

$50,000, $70,000). The poverty line is set at z = $1,100. The cen-
sored income vector is x* = ($800, $1,000, $1,100, $1,100).
• Use the method in equation (2.37) to calculate the pov-
erty gap measure. The poverty gap vector is g* = (300/1100,
100/1100,0,0). Then the poverty gap measure is PG(x;z) =
WA(g*) = 0.09.
• The method in equation (2.38) uses two income standards. The
mean of the censored distribution is WA(x*) = 1,000. The non-
poverty censored distribution is x̄*z = ($1,100, $1,100, $1,100,
$1,100). Thus, the mean of the nonpoverty censored distribu-
tion is WA(x*) = 1,100. Hence, the poverty gap measure is
PG(x;z) = (1,100 − 1,000) / 1,100 = 0.09.
• The method in equation (2.39) uses the headcount ratio and
the income gap ratio to calculate the poverty gap measure. We
already know that the headcount ratio of x is 0.5. The income
gap ratio of x may be obtained by taking the mean of the first
two elements of Gx and so PIG(x;z) = 2/11. Thus, the poverty gap
measure is PG(x;z) = 0.5 × 2/11 = 0.09.
What properties does the poverty gap measure satisfy? It satisfies all
invariance properties: symmetry, normalization, population invariance, scale
invariance, and focus. Among dominance properties, it satisfies only mono-
tonicity and subgroup consistency and does not satisfy the transfer principle
and transfer sensitivity. Although it does not satisfy the transfer principle, it
satisfies the weak transfer principle, which means that the poverty gap mea-
sure does not increase (or decrease) because of a regressive (or progressive)
transfer but also does not fall (or increase). The poverty gap measure satis-
fies the monotonicity property, meaning that if the income of a poor person
increases, then (unlike the headcount ratio) the poverty gap increases. The
poverty gap measure satisfies subgroup consistency because, like the head-
count ratio, it is additively decomposable.
117
There is a long list of advanced poverty measures. These measures may

not necessarily be as intuitive and as easy to understand as the two basic
measures, but they are capable of moderating the limitations of the two basic
measures. The advanced measures discussed in this book include the Watts
index, the Sen-Shorrocks-Thon index, the squared gap measure, the Foster-
Greer-Thorbecke indices, the mean gap measure, and the Clark-Hemming-
Ulph-Chakravarty indices.
Watts Index
The Watts index was proposed by Watts (1968), and it is the average dif-
ference between the logarithm of the poverty line and the logarithm of
incomes. For income distribution x with population size N and poverty line
z, the Watts index can be written as
1 N
PW (x; z) = ∑ (ln z − ln x *n).
N n =1
(2.40)
The lowest value the Watts index can take is zero, which is obtained
when no one is poor in the society. However, unlike the headcount ratio
and the poverty gap measure, the Watts index has no maximum value.
Like the two basic measures, the Watts index can also be expressed as a
difference between two income standards. The income standard used for the
headcount ratio and the poverty gap measure is the arithmetic mean, where-
as the income standard for the Watts index is the geometric mean. The
higher income standard is obtained by applying the geometric mean to the
nonpoverty censored distribution x̄*z. Because the geometric mean satisfies
normalization, the higher income standard is equal to the common ele-
ment in x*, which is the poverty line z itself. The lower income standard is
obtained by applying the geometric mean to the censored income distribu-
tion x*. The Watts index is the logarithm of the ratio of the higher and the
lower income standards.
The other way of interpreting the measure is by calculating the differ-
ence of their logarithms (see example 2.13). The formulation of the Watts
index in terms of income standards is
⎡ W (x * ) ⎤ ⎡ z ⎤
PW (x; z) = ln ⎢ G z ⎥ = ln ⎢ ⎥ = ln z − ln ⎡⎣WG (x * )⎤⎦ . (2.41)
⎣ WG (x* ) ⎦ ⎣ WG (x* ) ⎦
118
Example 2.13: How is the Watts index calculated by different meth-

ods? Consider the four-person income vector x = ($800, $1,000,
$50,000, $70,000), with the poverty line set at z = $1,100. The cen-
sored vector is x* = ($800, $1,000, $1,100, $1,100). The logarithm of
the poverty line is Inz = In1,000 = 7.
• Use the method in equation (2.40) to calculate the Watts index.
The logarithmic differences between the poverty line and the
censored incomes are (7 − In800, 7 − In1,000,0,0) = (0.3, 0.1, 0, 0),
the mean of which is 0.103. Thus, PW(x;z) = 0.1.
• Calculate the Watts index using the income standards. The
geometric mean of x* is WG(x*) = 991.9 and In[WG(x*)] = 6.9.
Therefore, by equation 2.41, PW(x;z) = 7 − 6.9 = 0.1. Thus, both
calculation and understanding of the Watts index are much easier
in terms of income standards.
The Watts index satisfies all invariance properties: symmetry, normaliza-

tion, population invariance, scale invariance, and focus, as well as all dominance
properties: monotonicity, transfer principle, transfer sensitivity, and subgroup
consistency. It satisfies the transfer principle because poverty falls when
income is transferred from a richer poor person to a poorer poor person. It
satisfies transfer sensitivity because it is more sensitive to a transfer at the
lower end of the distribution than at the upper end of the income distribu-
tion of the poor. It satisfies the subgroup consistency property because, like
the two basic measures, it is additively decomposable.
Sen-Shorrocks-Thon Index
The Sen-Shorrocks-Thon (SST) poverty index was originally formulated in

terms of a basic poverty measure and an inequality measure. The poverty
gap measure is the basic poverty measure used for constructing the SST, and
the Gini coefficient is the inequality measure. Thus, the SST index can be
expressed as
PSST(x;z) = PG(x;z) + [1− PG(x;z)]IGini(x∗). (2.42)
Note that the Gini coefficient is applied to the censored income distri-
bution x*.15 This measure is sensitive to inequality among the poor, which
is evident from its formulation in equation (2.42). If there is no inequality
119
among the poor, then PSST(x;z) reaches its minimum. As inequality increases,
the values of PSST(x;z) increase because 1 − PG(x;z) > 0, which results from
the fact that PG(x;z) lies between zero and one. The Gini coefficient lies
between zero and one. When there are no poor in a society, the SST index is
zero. The maximum value of one is obtained when everyone in the society is
poor and has zero income.
The SST index has an interesting relationship with the average normal-
ized income shortfall among the poor, PIG. When everyone is poor in a
society, but has equal income, then the SST index is equal to the average
normalized income shortfall among the poor, that is, PSST(x;z) = PIG(x;z).
This is because in this situation IGini(x*) is zero and PH = 1. When the
inequality level among the poor increases while the average normalized
income shortfall remains the same, the SST index becomes larger than the
average normalized income shortfall.
The SST index can also be interpreted by an income standard. The
income standard in this case would be the Sen mean. The SST index is the
normalized difference between the Sen mean of the nonpoverty censored
distribution and the Sen mean of the censored distribution. The Sen mean
satisfies the normalization property of income standards. Thus, the Sen
mean of the nonpoverty censored distribution is the poverty line itself, that
is, WS(x̄*) = z. The Sen mean of the censored distribution x* is denoted by
WS(x*). The SST index16 can be presented as
WS (xz* ) − WS (x* ) z − WS (x * )
PSST (x; z) = = . (2.43)
WS (xz* ) z
Given a censored distribution, once the Sen mean is calculated using the
procedure discussed in the income standard section, the SST index can eas-
ily be obtained by applying equation (2.43). How do equations (2.42) and
(2.43) give the same result? That question can easily be answered as
z − WS (x * ) z − WA (x * ) WA (x* )
= + IGini (x * ) = PG + (1 − PG )IGini (x * ). (2.44)
z z z
In the previous section, when discussing dominance and ambiguity
results for income standards, we mentioned that the Sen mean is related to
the generalized Lorenz curve. The SST index is based on the Sen mean and
thus is naturally related to the generalized Lorenz curve, which has been
graphically depicted in Zheng (2000). Example 2.14 shows how to calculate
the SST index.
120
Example 2.14: How is the Sen-Shorrocks-Thon index calculated

by different methods? Consider the four-person income vector
x = ($800, $1,000, $50,000, $70,000); the poverty line is set at
z = $1,100. The censored vector is x* = ($800, $1,000, $1,100, $1,100).
• Calculate the SST index using equation (2.42). The poverty gap
measure, as we already know, is 0.09. The Gini coefficient of x* is
0.062. Then PSST(x;z) = 0.09 + (1 − 0.09) × 0.062 = 0.15.
• Calculate the SST index using equation (2.43). The Sen
mean of x* is 937.5. Thus, the SST index is PSST(x;z) =
(1,100 − 937.5)/1,100 = 0.15.
What properties does the SST index satisfy? It satisfies all invariance
properties: symmetry, normalization, population invariance, scale invariance,
and focus. However, it does not satisfy all dominance properties because it
is based on the poverty gap measure and the Gini coefficient. It inherits
the monotonicity property from the poverty gap measure, and it inherits
the transfer principle from the Gini coefficient. However, neither the Gini
coefficient nor the poverty gap ratio satisfies transfer sensitivity; conse-
quently, the SST index does not satisfy transfer sensitivity. Furthermore,
the Gini coefficient is neither subgroup consistent nor additively decom-
posable in the usual way. This shortcoming is also inherited by the SST
index.
Despite these shortcomings, the SST index is useful because it can be
broken down into the poverty gap measure and the Gini coefficient. In fact,
the poverty gap measure can be further broken down into the headcount
ratio (PH) and the average income gap of the poor (PIG).
Squared Gap Measure
The next poverty measure in the advanced measures category is the squared
gap measure. This measure is calculated by averaging the square of the nor-
malized income shortfalls and is denoted by
2
1 N ⎛ z − x *n ⎞
PSG (x; z) = ∑
N n =1 ⎜⎝ z ⎟⎠
. (2.45)
One way of interpreting the squared gap measure is as the weighted aver-
age of normalized income shortfalls, where each normalized income shortfall
121
is weighted by itself. This method of weighting puts greater emphasis on

larger shortfalls during aggregation. Thus, a transfer of income from a richer
poor person to a poorer poor person should reduce poverty. Like the SST
index, the squared gap measure can also be expressed as a function of the
headcount ratio (PH), the average normalized income shortfall (PIG), and the
generalized entropy measure for a = 2 of the incomes of the poor (denoted
by the vector xq), such that
PSG(x;z) = PH[PIG
2 + 2(1 − P )2I (xq;2)].
IG GE (2.46)
The squared gap measure lies between zero and one (see example 2.15).
A zero value is obtained when there are no poor people in the society
because the headcount ratio is zero. The maximum value of one is reached
when everyone in the society is poor and has zero income.
Example 2.15: How is the squared gap measure calculated by different

$50,000, $70,000). The poverty line is set at z = $1,100. The cen-
sored vector is x* = ($800, $1,000, $1,100, $1,100).
• Use the method in equation (2.45) to calculate the squared gap
measure. The squared gap vector is sg* = ([300/1100]2, [100/1100]2,
0, 0). Then the squared gap measure is PSG(x;z) = WA(sg*) = 0.02.
• The method in equation (2.46) uses the headcount ratio, average
normalized income shortfall, and generalized entropy measure
to calculate the squared gap measure. We already know that the
headcount ratio of x* is 0.5 and that the poverty gap measure is
0.18. The inequality measure IGE(x q ; 2) among the poor is 0.006.
Then the squared gap measure is PSG(x;z) = 0.5[0.182 + 2 × (1 −
0.18)2 × 0.006] = 0.02.
What properties does the squared gap measure satisfy? It satisfies all
invariance properties: symmetry, normalization, population invariance, scale
invariance, and focus. However, among the dominance properties, it satisfies
monotonicity, the transfer principle, and subgroup consistency, but it does not
satisfy transfer sensitivity because the headcount ratio, the income gap ratio,
and the generalized entropy of order 2 do not satisfy this property. Hence,
like the basic poverty measures and the SST index, the squared income gap
122
measure is transfer neutral. However, unlike the SST index, it satisfies sub-
group consistency because it is additively decomposable.
Foster-Greer-Thorbecke (FGT) Family of Indices
This family of measures was proposed by Foster, Greer, and Thorbecke

(1984). The FGT family of measures has the following formulation:
a
N ⎛ z − xj ⎞
*
1
PFGT = (x; z,a ) = ∑ n =1 ⎜ ⎟ , (2.47)
N ⎝ z ⎠
where a ≥ 0. The parameter a can be interpreted as the inequality aver-

sion parameter among the poor, which is conceptually the same as that for
Atkinson’s class of inequality measures. As a increases, a society’s aversion
toward inequality among the poor increases.
Notice that there is a minor difference between parameter a in this
case and parameter a in Atkinson’s class of inequality measures, where a
lower value of a leads to greater aversion toward inequality. This differ-
ence exists because inequality is measured in the income space and poverty
is measured in the normalized gap space, where large gaps imply worse
situations.
Measures in the FGT family take the form of various well-known poverty
measures introduced earlier for different values of a. For example, for a = 0,
the formulation in equation (2.45) becomes the headcount ratio because
(z − x*n/z)0 =1 when xn < z and because (z − x*n/z)0 = 0 when xn ≥ z. Thus,
PFGT(x;z,0) = q/N = PH(x;z). For a = 1, the formula becomes the poverty
gap measure, which is the average of all normalized income shortfalls. For
a = 2, the formula is the squared gap measure, which is the average of the
square of all normalized income shortfalls.
As a increases and becomes very large, PFGT approaches a Rawlsian
measure17 placing more emphasis on the largest normalized income gap of
the poorest person. However, note that the value of PFGT for any distri-
bution decreases as a increases, and, for a very large a, the overall value
of PFGT may be infinitesimally small. This occurrence can be verified by
expressing the FGT formulation in equation (2.47) in general mean form
using equation (2.3) as follows:
PFGT(x;z,a) = [WGM(g∗; a)]a for a > 0. (2.48)
123
Recall that the general mean of a distribution converges toward the

maximum or largest element in a vector or distribution. The largest element
in the gap vector g* belongs to the poorest person in the society.
We have already discussed the properties that the headcount ratio, the
poverty gap measure, and the squared gap measure satisfy. Thus, we know
what properties the FGT family of indices satisfies when a = 0, 1, and 2. The
additional property that the measures in this family satisfy is transfer sensi-
tivity when a > 2, which implies that if a similar amount of transfer takes
place between two poorer poor people and two richer poor people, then this
measure is able to distinguish between these two situations.
An aspect that is not so intuitive in this family of measures is interpreta-
tion of the inequality aversion parameter. A larger value of a implies greater
aversion to inequality among the poor. However, when there is no inequal-
ity in the society, should the poverty measure alter because of a change in α?
For example, suppose that in a society of 100 people, everyone is poor and all
people have an equal income of $500. If the poverty line is z = $1,000, then
the normalized income gap of each person is one-half in this society. Given
that there is no inequality in the society, it should not matter how averse the
society is to inequality because there is no inequality.
However, the FGT family of measures may not remain the same for all α.
For the simple example considered above, PFGT(x;z,1) = PG(x;z) = 1/2 and
PFGT(x;z,2) = PSG(x;z) = 1/4. However, this problem can be easily solved
by calculating a monotonic transformation of the original FGT family of
measures as
P'FGT(x;z,a) = [PFGT(x;z,a)]1/a = WGM(g*; a) for a > 0. (2.49)
Note that this formula is not valid for the headcount ratio when a = 0.
For the example above, P'FGT(x;z,a) = 1/2 for all a > 0 because the general
mean satisfies the normalization property of income standards.
Mean Gap Measure
The mean gap measure of poverty can be obtained by taking the Euclidean
mean (WE) of the normalized income shortfalls. This is a monotonic trans-
formation of the squared gap measure. More specifically, the mean gap mea-
sure is the square root of the squared gap measure. The mean gap measure
can be expressed as
124
1
1 ⎛ 1 N ⎛ z − x* ⎞ 2 ⎞ 2
PMG (x; z) = WE (g ) = P = ⎜ ∑ ⎜
* 2
SG
n
⎟⎠ ⎟ . (2.50)
N
⎝ n =1 ⎝ z ⎠
There is another interpretation of the mean gap measure: P'FGT(x;z,2).

Because the mean poverty gap is a monotonic transformation of the squared
gap measure, it satisfies all the properties that are satisfied by the squared gap
measure except the additive decomposability. One advantage of the mean
gap measure compared with the squared gap measure is that the values of
the mean gap measure are commensurate with the values of the poverty
gap measure as discussed using equation (2.49). Values of the squared gap
measure tend to be much smaller than the poverty gap measure, and these
numbers are not comparable to each other.
Unlike the squared gap measure, values of the mean gap measure tend
to be higher than those of the poverty gap measure, because it uses the
Euclidean mean instead of the arithmetic mean. For example, for the four-
person income vector x = ($800, $1,000, $50,000, $70,000) and poverty line
z = $1,100, the poverty gap measure is 0.09, whereas the mean gap measure
is (0.02)1/2 = 0.14. However, had the income of the poor been equally dis-
tributed, the income vector would have been x' = ($800, $1,000, $50,000,
$70,000), and the poverty gap measure would remain the same as that of x
(that is, 0.09), but the mean gap measure would be 0.13.
Like the squared gap measure, the mean gap measure also lies between
zero and one. Moreover, this measure has an interesting relationship with
the average normalized income shortfall. When everyone in a society is
poor, but there is no inequality, then the squared gap measure is equal to
the average normalized income shortfall among the poor because CV = 0
and PH = 1. Thus,
PMG = PSG = PH ⎡⎣P12G + z (1 − P1G )2 IGE (xa; z)⎤⎦ = P12G = P1G . (2.51)
Clark-Hemming-Ulph-Chakravarty (CHUC) Family of Indices
The final measure in our discussion of poverty measures is the Clark-

Hemming-Ulph-Chakravarty (CHUC) family of indices (see Clark,
Hemming, and Ulph 1981; Chakravarty 1983). This family is an extension
of the Watts index. The CHUC index is based on the generalized mean
125
and is the normalized shortfall of the generalized mean of the observed cen-
sored income distribution x* from the generalized mean of the nonpoverty
censored income distribution x̄*. Again, the generalized mean satisfies the
normalization property of income standards; thus, the generalized mean of
the nonpoverty censored income distribution is the poverty line itself. The
CHUC index for a ≤ 1 can be expressed as
WGM (xz*;a ) − WGM (x *;a ) z − WGM (x *;a )

PCHUC (x; z) = = . (2.52)
WGM (xz*;a ) z
The CHUC index lies between zero and one. The minimum value of
zero is obtained when there are no poor people in a society. However, the
maximum value of the CHUC index cannot be larger than one. When
everyone in a society is poor, having equal income, this measure is equal
to the average normalized income shortfall. It satisfies all invariance and
dominance properties. However, not all measures in this class are addi-
tively decomposable. For a = 1, the CHUC index is the poverty gap mea-
sure, and for a = 0, the CHUC index is a monotonic transformation of the
Watts index.
Advantages and Disadvantages of Each Measure
We have shown that the two basic measures—the headcount ratio and the
poverty gap measure—do not satisfy transfer-related properties and so are
not sensitive to inequality across the poor. Besides not being sensitive to
inequality, the headcount ratio does not satisfy monotonicity, which, if it is
used as a target for public policy, may cause inefficiency in public spending.
All of the subsequent advanced poverty measures, in contrast, are sensitive
to inequality across the poor. The SST index and the mean gap measure are
both equal to the poverty gap measure when everyone in a society is poor
and no inequality exists among them. These two measures become larger
than the poverty gap measure when the income gap remains the same, but
inequality among the poor increases.
Each advanced measure, however, has its own pros and cons. Let us
begin with the SST measure. We know from our previous discussion that
this measure is not subgroup consistent, which means that it may lead to
inconsistent outcomes when group-level analysis is of interest. This measure
is also not transfer sensitive, which means that if a similar amount of transfer
126
takes place between two poorer poor people and two richer poor people,
then this measure cannot distinguish between the two situations.
What, then, are the SST index’s advantages? The first is that it can
be neatly broken down into the headcount ratio, the average normalized
income shortfall among the poor, and the well-known Gini coefficient. If
one is not interested in group-level analysis, then this measure can be bro-
ken down into these three components to understand the source of change
in poverty. In fact, the Gini coefficient can be broken down further into a
within-group and a between-group component using the Gini decomposi-
tion formula introduced earlier. The within-group component assesses
inequality among the poor, and the between-group component measures
inequality between the average income of the poor and the poverty line.
This decomposition reveals whether the change in the measure’s inequal-
ity component is caused by the change in inequality among the poor or due
to a change in the average income of the poor compared to the poverty line.
Note that there is no within-group inequality among the nonpoor because they
all have the same income equal to the poverty line. Furthermore, there is no
residual term, which is commonly seen in the Gini decomposition, because
there is no income overlap between the poor and the nonpoor.
Second, consider the squared gap measure. This measure has many posi-
tive features, such as it is additively decomposable and subgroup consistent.
Furthermore, like the SST index, it can be broken down into the head-
count ratio, the average normalized income shortfall among the poor, and
the generalized entropy measure order of 2 among the poor to understand
the poverty composition. However, like the SST index, this measure is not
transfer sensitive, which means that if a similar amount of transfer takes
place between two poorer poor people and two richer poor people, then this
measure cannot distinguish between these two situations.
Also, the generalized entropy measure order of 2 may be a bit unintuitive
in the sense that it may range from zero to infinity, unlike the Gini coefficient
that ranges from zero to one. The same pros and cons apply to the mean gap
measure, which is just a monotonic transformation of the squared gap measure.
Third, consider the Watts index. This measure appears to be a perfect
measure of poverty in the sense that it satisfies all the properties that we dis-
cussed earlier: it is additively decomposable, is transfer sensitive, and satisfies
the transfer principle and all other properties. However, this measure has two
shortcomings. One is that it is not applicable when there are zero incomes
because the logarithm of zero is undefined. The second shortcoming is that
127
it does not have an intuitive interpretation like the two basic measures, the
SST index and the squared gap measure and its monotonic transformation
(the mean gap measure). Also, like these other measures, it does not have an
upper bound of one. Finally, the CHUC class of indices is a generalization of
the Watts index. Like the Watts index, its members satisfy all the properties
discussed earlier and also lie between zero and one. However, measures in
this class are not defined for zero incomes when α ≤ 0.
Policy Relevance of Poverty Measures
Besides gauging the level of deprivation in a society, a poverty measure can

have crucial policy relevance. In fact, different measures may have different
policy implications. We discuss three policy implications below with cer-
tain examples. First is the influence of poverty measures as targeting tools.
Second is the relevance of poverty measures in guiding public policies. Third
is the use of the additive decomposability property for geographic targeting.
How Do Different Poverty Measures Influence the Targeting Exercise?
Besides gauging the level of deprivation in a society, a poverty measure is a

useful tool that can influence a policy maker’s targeting exercise. An impor-
tant question that is often asked is the following: if a policy maker has allot-
ted a certain amount of the budget that he or she can spend on the welfare
program for the poor, how should that budget be allocated among the poor?
For instance, consider the following six-person society with income vector
x = ($80, $100, $800, $1,000, $50,000, $70,000). The poverty line is set at
$1,100 so that four people are poor and two people are nonpoor.
It is evident that the society’s policy maker requires at least $2,420 so
that he or she can drive all four poor people out of poverty. Suppose that
the policy maker can allot only $1,000 toward the welfare program for the
poor. Then how should that budget of $1,000 be allocated among the poor?
The answer depends on which poverty measure is used to assess the society’s
deprivation. Different poverty measures provide different answers for this
targeting exercise.
We begin this analysis when the society’s poverty is assessed by the
headcount ratio. The easiest way for a policy maker to reduce the headcount
ratio is to bring as many poor people as possible up to the poverty line.
Therefore, the first $100 of the allotted budget would be spent on the richest
128
poor person (with an income of $1,000). The next $300 would be spent on
the second-richest poor person (with an income of $800).
After bringing these two poor people out of poverty, the policy maker
still has $600 in his or her budget that remains unused. How and whom
should this amount assist? Given that the headcount ratio does not satisfy
the monotonicity property, because even if this entire amount is transferred
to either of the two remaining poor people, the poorest people still remain
under the poverty line and do not add to the headcount ratio. The policy
maker in this situation would have no incentive to spend the remaining
budget. This lack of incentive creates inefficiency in public spending.
Although poverty is reduced by 50 percent, the poverty status of the two
severely deprived people remains unchanged.
What if the society’s poverty is assessed by the poverty gap measure? Recall
that, unlike the headcount ratio, the poverty gap measure satisfies monotonicity;
but, like the headcount ratio, it does not satisfy the transfer principle or transfer
sensitivity. Thus, it is not sensitive to inequality among the poor. What implica-
tion does it have on the targeting exercise? In this case, the policy maker will
be inclined to spend his or her entire budget because the poverty gap measure
satisfies monotonicity. An increase in a poor person’s income, even when he or
she is not driven out of poverty, reduces the poverty gap measure. Therefore,
unlike the headcount ratio, inefficiency in public spending does not arise.
Then how should the budget of $1,000 be allocated among the poor? The
straightforward way is to spend the budget on any of the four poor people as
long as they do not surpass the poverty line income. Given that the poverty
gap measure is not sensitive to inequality among the poor, it does not matter
who among the poor receives the assistance. For example, in one case, out
of the budget of $1,000, the richest poor person, with an income of $1,000,
may receive $100; the second-richest poor person may receive $300; and the
third-richest poor person may receive the rest, or, in another case, the poor-
est person, with an income of $80, may receive the entire amount. In both
cases, the improvements in the poverty gap measure are the same. Thus, the
poverty gap measure is insensitive to whoever receives the assistance. The
poorest section of a society may perpetually remain poor in spite of showing
decent progress in terms of the poverty gap measure.
How would this policy exercise be affected when the society’s poverty is
gauged by a distribution-sensitive poverty measure? A distribution-sensitive
measure requires that assistance should go to the poorest of the poor first.
Thus, out of the $1,000 budget allotted for the poor, the first $20 should go
129
to the poorest person whose income is $80 so that the incomes of the two
poorest poor people are made equal. Then the rest of the budget should be
equally divided between the two poorest people so that, after allocating
the entire budget, the income distribution becomes x' = ($590, $590, $800,
$1,000, $50,000, $70,000).
What if, instead of $1,000, there was $1,600 allotted to the welfare of
the poor? Then the first $20 would be transferred to the poorest person.
Next, out of $1,580, $1,400 would be divided equally between the two poor-
est people so that the incomes of all three of the poorest people would be
equalized at $800. Finally, the rest of the budget of $180 is equally divided
among the three poorest poor so that the post-allocation income vector
is x" = ($860, $860, $860, $1,000, $50,000, $70,000). All distribution-
sensitive poverty measures support this type of targeting. However, not all
measures reflect similar amounts of decrease in poverty, which depends on
how these measures weight different people.
Can Poverty Measures Influence Public Policy?
Like the targeting exercise, can a poverty measure influence public policy?
Consider an example of a developing country where the major staple food is
rice. As with other agricultural producers, rice producers are poor and their
incomes are scattered around the country’s poverty line income. Some rice
producers earn enough income to live just above the poverty line, but many
rice producers are unfortunate enough to live below the poverty line.
There are other poor people in the country, such as those whose major
occupation is agricultural labor, plantation labor, or other unskilled jobs.
These poor people are the poorest in the country, and their major source
of energy and nutrition is the staple food, rice. Rice is, in fact, a necessary
commodity in that country, and the government controls its price.
Being benevolent, the government wants to see a reduction in poverty
by adjusting the price of rice. Which of the following two policy options
would reduce poverty?
• Option 1: Reduce the price of rice.

• Option 2: Increase the price of rice.
Suppose poverty in the country is assessed by the headcount ratio. If

the government decides to choose option 1 and reduce the price, then rice
130
producers would be adversely affected because their income would fall, and
rice consumers would benefit because their real incomes would increase.
Given that most rice consumers are poorer than rice producers, one does
not know whether more or fewer people would become poor. Thus, the
impact on the headcount ratio is uncertain.
However, if the price of rice increases, then producers gain, but the
poorer consumers lose because their real incomes fall. Given that the
already poor consumers become poorer, this is not taken into account by
the headcount ratio because it does not satisfy monotonicity. Therefore,
the number of poor people would most likely fall, thereby leading to a fall
in the country’s headcount ratio. Thus, the potential assessment of poverty
using the headcount ratio would incline the government to choose option
2 and increase the price because poverty, according to the headcount ratio,
would fall.
Note, however, that the decrease in the headcount ratio has ignored the
change in inequality among the poor. The marginally poor producers would
become better off because of the price increase, but the severely poor people
would be worse off for the same reason. This occurrence is very similar to the
idea of regressive transfer. The higher price paid by the poorer consumers is
obtained by the lesser poor producers as profit.
Any inequality-sensitive poverty measure, such as the squared gap,
the Watts index, or the SST index, would be sensitive to such inequality
among the poor. Suppose the poverty level in that country is now assessed
with one such measure that is sensitive to inequality among the poor. If
the government now chooses option 1 and reduces the price of rice, then
the poorer consumers benefit at the cost of a reduction in the producers’
income. The result is uncertain. If some producers become poorer than
some consumers, then the poverty measure may increase. But if the pro-
ducers remain less poor than the consumers, then the poverty measure
may fall.
However, if option 2 is chosen and the rice price rises, then inequality
among the poor increases and, most certainly, the poverty measure would
increase. Hence, the potential assessment of poverty using any inequality-
sensitive poverty measure would incline the government to not raise the
price because poverty, according to any inequality-sensitive measure, would
increase. The conclusion is that different poverty measures would incline the
government to choose different policies.
131
Additive Decomposability and Geographic Targeting
A poverty measure of a population subgroup reflects the level of depriva-

tion for that subgroup. A higher value of a population subgroup’s poverty
measure reflects a higher level of deprivation. The poverty measures we
have discussed in this chapter satisfy population replication invariance to
be able to compare the poverty levels of different population sizes, so these
measures are invariant to population size. However, a population subgroup
with a higher level of poverty does not necessarily imply that the subgroup
has a larger contribution to overall poverty.
A subgroup’s contribution to overall poverty also depends on the popula-
tion distribution across subgroups. Therefore, targeting a region or a group
based on only a poverty measure may not be completely accurate. We also
need to take the population distribution into account. If P is an additively
decomposable poverty measure and the income distribution x with total popu-
lation size N is divided into M subgroups—x1 with population size N1, x2 with
population size N2, …, xM with population size NM—then the contribution
of group m to total poverty is NmP(xm;z)/NP(x;z), where z is the poverty line.
Consider the situation when poverty is assessed by the headcount ratio.
A population subgroup’s headcount ratio denotes the population percentage
identified as poor. Interpreting a population subgroup’s contribution to over-
all poverty in terms of the headcount ratio is intuitive. If the total number
of poor is q, and qm is the number of poor in subgroup m, then the overall
headcount ratio is q/N and that of subgroup m is qm/Nm for all m = 1,…, M.
Then subgroup m’s share of overall poverty is Nm[qm/Nm]/N[q/N] = qm/q.
Thus, the contribution of the subgroup’s poverty to overall poverty in terms
of the headcount ratio is just the share of overall poor in that subgroup.
For example, consider table 3.9 in chapter 3, which shows the distribu-
tion of the poor across Georgian subnational regions for years 2003 and
2006. Suppose that, in 2003, the headcount ratio of the subnational region
Kvemo Kartli is 44.4 percent, which is more than twice the headcount ratio
of 20.9 percent in Tbilisi. However, the share of total poor living in Tbilisi
is, in fact, slightly larger than that living in Kvemo Kartli, because the popu-
lation size of Tbilisi is more than twice that of Kvemo Kartli. In 2006, the
headcount ratio of Kvemo Kartli decreased to 35.1 percent, which is still
10 percent higher than the headcount ratio of Tbilisi, but the share of the
poor living in Tbilisi increased to 20.4 percent alongside only 12.2 percent
in Kvemo Kartli. Therefore, the Georgian government needs to understand
132
that, despite having a lower headcount ratio, a massive number of poor

people reside in Tbilisi.
The share of subgroup poverty in overall poverty also has an intuitive inter-
pretation that can be relevant for geographic targeting. Using the same nota-
tions as in the previous paragraph, we can express the poverty gap measure as
q
[∑i = 1(z − xi)]/Nz,
q
where [∑i = 1(z − xi)] is the total sum of financial assistance required to bring
all poor people just to the poverty line to eradicate poverty. If the distribu-
tion x is divided into M subgroups as earlier, then the poverty gap measure
of subgroup m is
q
[∑ i m= 1(z − xi)]/Nmz,
q
where [∑ i m= 1(z − xi)] is the total amount of financial assistance required to
eradicate poverty in subgroup m. The contribution of subgroup m’s poverty
gap measure to the overall poverty gap ratio is
q q q q
[Nm∑ i m= 1(z − xi)]/Nmz]/N[∑i = 1(z − xi)]/Nz = ∑ i m= 1(z − xi)/∑i = 1(z − xi). (2.53)
Therefore, a subgroup’s contribution is nothing but the share of total

financial assistance that should be received by that subgroup to eradicate pov-
erty. Thus, the contribution in terms of the poverty gap measure may be used
to understand the requirement for fund allocation across geographic regions.
The subgroup contribution of other additively decomposable poverty
measures that are sensitive to inequality, such as the squared gap or the
Watts index, may not have such an intuitive implication for targeting.
However, their additively decomposable property enables us to understand
the subgroup’s contribution to overall poverty and monitor the targeting
exercise. Although for these examples we have considered only the popula-
tion subgroups in terms of subnational regions, the population may well be
grouped alternatively by gender, occupation, or household head character-
istics, as depicted in chapter 3.
Poverty, Inequality, and Welfare
Poverty measures that satisfy the transfer principle are called distribution-
sensitive poverty measures. The distribution-sensitive poverty measures
133
introduced earlier were the Watts index, the SST index, the FGT family of
measures for α > 1, and the CHUC family of indices. Each of these distribu-
tion-sensitive poverty measures is built on a specific income or gap standard
that is closely linked to an inequality measure. For example, the Watts index
is closely linked with Theil’s second measure of inequality, the SST index
is closely linked with the Gini coefficient, the FGT family of indices for
α > 1 is linked with the generalized entropy measures, and the CHUC fam-
ily of indices is linked with Atkinson’s family of measures.
For the Watts index, SST index, and CHUC family of indices, the
inequality measure is applied to the censored distribution x*, with greater
censored inequality being reflected in a higher level of poverty for a given
poverty gap level. The FGT indices for α > 1, however, use generalized
entropy measures applied to the gap distribution g*, with greater gap inequal-
ity leading to a higher level of poverty for a given poverty gap level.
Recall from our earlier discussion in the income standard section that
certain income standards can be viewed as welfare functions, and this link
provides yet another lens for interpreting poverty measures. The Sen mean
used in the SST index and the general means for α ≤ 1 that are behind the
CHUC indices can be interpreted as welfare functions. In each poverty
measure, the welfare function is applied to the censored distribution to
obtain the censored income standard, which is now seen to be a censored
welfare function that takes into account poor incomes and only part of non-
poor incomes up to the poverty line. For these measures, poverty and cen-
sored welfare are inversely related—every increase in poverty can be seen as
a decrease in censored welfare.
A poverty measure assesses the level of poverty within a society by a single

number for a given poverty line. Two obvious questions arise: (a) Does a
single poverty measure evaluate two distributions in the same way for all
poverty lines? and (b) Do all poverty measures evaluate two income distri-
butions in the same way? More specifically, according to the first question, if
one distribution has more poverty than another distribution for a particular
poverty line, is there any certainty that the former distribution would have
more poverty than the latter for any other poverty line?
Consider the following example with two four-person income distribu-
tions x = ($800, $900, $5,000, $70,000) and x' = ($200, $1,200, $1,600,
134
$70,000). Let poverty be measured by the headcount ratio. If the poverty

line is $1,000, then distribution x has more poverty than distribution x'.
What happens if the policy maker decides that the correct poverty line
should be $800? Then distribution x has no poor people, but distribution
x" has one poor person. Similarly, if the poverty line is $2,000, then, again,
distribution x has less poverty than distribution x". Hence, the choice of
poverty line affects the poverty comparison.
According to the second question, if one poverty measure determines
income distribution x to have more poverty than distribution x', would
other poverty measures compare these two distributions in the same way?
This situation is analogous to our discussion of dominance and ambiguity
for inequality and income standards. The answer is not too optimistic and
depends on the poverty measure used—not all poverty measures evaluate
different distributions in the same manner.
Consider the same two four-person income vectors used above: x = ($800,
$900, $5,000, $70,000) and x' = ($200, $1,200, $1,600, $70,000). Let the
poverty line be z = $1,000. We have already seen that the headcount ratio
reflects more poverty in distribution x than in distribution x'. How does the
poverty gap measure PG compare these two distributions? It turns out that
PG(x; z) = 0.08 < PG(x'; z) = 0.18. Distribution x has less poverty than distribu-
tion x'. Thus, these two basic measures disagree with each other.
Is there any way we can devise situations where we have unanimous
results? To start, we try to answer the first question using a concept intro-
duced at the beginning of this chapter: the cumulative distribution function, or
cdf.18 Recall that the cdf of distribution x denotes the proportion of people in
the distribution whose income falls below a given income level. In the pov-
erty analysis context, if that income level is the poverty line z, then the
value of the cdf at z is nothing but the headcount ratio at poverty line z (see
figure 2.14 below).
Poverty Incidence Curve
The horizontal axis of figure 2.14 denotes income, and the vertical axis
denotes the values of a cumulative distribution function. If the poverty line
is set at z, then the headcount ratio is PH(x; z), which is the percentage of
people in distribution x who have incomes less than z. Similarly, PH(x; z')
and PH(x; z") are the headcount ratios of distribution x corresponding to
poverty lines z' and z'', respectively.
135
Figure 2.14: Poverty Incidence Curve and Headcount Ratio
Cumulative distribution (%)

100
Fx ′
PH(x ′; z ″) Fx
PH(x ′; z ′)
PH(x; z ″)
PH(x; z ′)
PH(x ′; z)
PH(x; z)
z z′ z″ xN
Income
Suppose there is another distribution x'. One can see in figure 2.14 that
the headcount ratios corresponding to poverty lines z, z', and z" lie above
the respective headcount ratios for distribution x. Is there any other poverty
line that reflects a higher headcount ratio in x than in x'? The answer is no.
The cdf of x lies to the right of the cdf of x', which means that the headcount
ratio for x' for no poverty line can be lower than the headcount ratio for x.
When a cdf lies to the right of another cdf, first-order stochastic dominance
(introduced earlier) occurs. When such dominance relation holds between
two cdfs, not only do the headcount ratios agree for all poverty lines, but the
poverty gap measure, the squared gap measure, the mean gap measure, the
Watts index, and the CHUC indices also agree for all poverty lines.
This approach also answers the second question, which asks when all
poverty measures agree. Therefore, if the first-order stochastic dominance
holds, then there is no need to compare any two distributions by any poverty
measure introduced earlier with respect to varying the poverty line. The
choice of poverty measure and the choice of poverty line simply do not mat-
ter when the first-order dominance condition holds. The cdf in the context
of poverty measurement is also known as the poverty incidence curve.
Poverty Deficit Curve
What if two poverty incidence curves cross? Then a unanimous relationship

in terms of the headcount ratio does not hold. However, there are two other
136
poverty-value curves that lead to a unanimous relationship in terms of the

poverty gap measure and the squared gap measure. These two curves are
known as the poverty deficit curve and the poverty severity curve.
When the poverty deficit curve of one distribution lies above the poverty
deficit curve of another distribution, then the former distribution has higher
poverty—in terms of the poverty gap measure for all poverty lines—than
the latter distribution. Similarly, if the poverty severity curve of a distribu-
tion lies above the poverty severity curve of another distribution, then the
former distribution has higher poverty in terms of the squared gap measure
for all poverty lines. We now elaborate these two concepts.
Figure 2.15 outlines the poverty deficit curve concept. We use the pov-
erty incidence curve (panel a) to construct the poverty deficit curve (panel
b). The poverty incidence curve of distribution x is denoted by Fx. The
height of a poverty deficit curve at a poverty line is the area underneath
the poverty incidence curve to the left of the poverty line. In figure 2.15,
the height of the poverty incidence curve at poverty line z is denoted by
height B, which is the shaded area below the poverty incidence curve Fx
to the left of z. For instance, for the poverty line z, if q people are identi-
fied as poor, then Fx(z) = q/N percent, which is the percentage of the poor
population.
What does the area underneath the incidence curve denoted by B
mean? To understand, first note that the lightly shaded area denoted by
A is the average income of the q poor people times the share of the poor.
Figure 2.15: Poverty Deficit Curve and the Poverty Gap Measure
a. Poverty incidence curve b. Poverty deficit curve

C
Deficit
Fx
Dx
Fx(z)
Dx ′
A B
B
z xN z xN
Income Income
137
This can be easily verified from the quantile function as described earlier
in figure 2.5.
Recall that an income distribution’s cdf is just the inverse of the relevant
distribution’s quantile function. Thus, A is WA(xA)(xq)q/N = (x1 + … + xq )/N.
Another interpretation of area A is that it is the per capita income of an aver-
age poor person in the society. The combined area A + B denotes the society’s
per capita income, which, if held by each poor person, means that the poor will
not be poor anymore.
This per capita income is qz/N. Thus, area B, which is also the height of
the poverty deficit curve Dx at poverty line z, is the difference between the
area A + B and the area A, or the average income shortfall or the deficit,
that is, [z − WA(xq)]q/N. This deficit is the minimum per capita income of
the society, which, if transferred to the poor, will lift the poor out of poverty.
Area B is also zPG(x; z). The maximum height of the poverty deficit curve
is denoted by C, which is xN − WA(x).
Example 2.16: Suppose in a country of 100 million (m) people with a

per capita income of $20,000, 30 million people are poor. The aver-
age income of these poor people is $400. So the per capita income
held by an average poor person is ($1,000 − $400) × 30m ÷ 180m.
If the poverty line is $1,000, then the deficit is ($1,000 − $400) ×
30m ÷ 100m = $180.
Thus, $180 per capita, which is only 0.9 percent of the per capita
income of the country, is the minimum amount required to bring all
30 million poor people out of poverty.
Note that the larger height of the poverty deficit curve Dx compared
to the poverty deficit curve Dx' at z reflects a larger poverty gap measure
in distribution x than in distribution x' at poverty line z. It is evident from
figure 2.15 that the poverty deficit curve Dx lies above the poverty deficit
curve Dx' for all poverty lines. Hence, distribution x has higher poverty than
distribution x' for all poverty lines in terms of the poverty gap measure.
This type of unanimity result, however, fails to hold when two poverty
deficit curves cross each other. We should then check the poverty severity
curve of these two distributions. If the poverty severity curve of a distribu-
tion lies above the poverty severity curve of another distribution, then the
former distribution has higher poverty than the latter in terms of the squared
gap measure or the mean gap measure for all poverty lines.
138
Poverty Severity Curve
Panel a of figure 2.16 displays the poverty deficit curve that we will use to
show how a poverty severity curve is constructed. As explained earlier, the
height B of a poverty deficit curve is proportional to the poverty gap measure
and is the poverty gap measure times the poverty line. As shown in panel b,
the height of the poverty severity curve Sx at poverty line z is D, which is the
area underneath the poverty deficit curve Dx. Area D is proportional to the
squared gap measure. Therefore, the larger the height of the poverty sever-
ity curve Sx than the poverty severity curve Sx at z, the larger the squared
gap measure in distribution x than in distribution x' at poverty line z. It
turns out that the poverty severity curve Sx lies above the poverty severity
curve Sx' for all poverty lines. Hence, distribution x has higher poverty than
distribution x' for all poverty lines.
Note that the dominance by the poverty deficit curve is equivalent to
the second-order stochastic dominance, and the dominance by the poverty
severity curve is equivalent to the third-order stochastic dominance.19
When there is dominance in terms of poverty incidence curves, all pov-
erty measures satisfying the invariance properties and monotonicity agree
with each other when ordering distributions according to the level of pov-
erty for any poverty line. Such dominance relationships do not always hold.
When two poverty incidence curves cross, one distribution has higher or
lower poverty only for a part of the entire range of incomes. In fact, different
poverty measures may order two distributions differently.
Figure 2.16: Poverty Severity Curve and the Squared Gap Measure
a. Poverty deficit curve b. Poverty severity curve
E
C
Severity
Deficit
Dx Sx
Sx ′
B D
D
z xN z xN
Income Income
139
One way of examining the robustness of poverty comparisons is by cal-

culating the vector of poverty levels of different measures for a fixed pov-
erty line. For instance, the headcount ratio, the poverty gap measure, the
squared gap measure, the Watts index, and the SST index can be depicted
in a five-dimensional vector. If there are two distributions x and x', then the
five-dimensional vector of x for poverty line z is
(PH(x; z), PG(x; z), PSG(x ;z), PW(x ;z), PSST(x ;z)),
and the five-dimensional vector of x' for poverty line z is
(PH(x'; z), PG(x'; z), PSG(x'; z), PW(x'; z), PSST(x'; z)).
Vector dominance between these two vectors would then be interpreted
as a variable measure poverty ordering that ranks distributions when all five
measures unanimously agree. If each element in the vector x is greater than
each corresponding element in the vector x', then distribution x has unani-
mously more poverty than distribution x' for poverty line z.
Sensitivity Analysis with Respect to the Poverty Line
The dominance analysis discussed earlier helps us understand whether one dis-
tribution has more or less poverty than another distribution. It is not concerned
about the level of poverty, which is often of particular policy interest. The num-
ber of poor people in a country or the fact that many poor people have been
moved out of poverty over a particular time period are always matters of great
concern. These data, of course, depend on the particular poverty line chosen.
As discussed in the introductory chapter, there are three different types
of poverty lines:
• An absolute poverty line may be adjusted with the rate of inflation over
time, but it is not adjusted with income growth over time.
• A relative poverty line is not fixed over time, and it changes with income
growth. For example, if a poverty line is set at 50 percent of the median
income, then the poverty line changes as the median income changes
over time. Or the poverty line may be set at 50 percent of mean
income. In this case, the growth rate of the poverty line over time is
the same as the growth rate of per capita income over time.
• A hybrid poverty line is created by taking a weighted average of an
absolute poverty line and a relative poverty line.
140
No matter how a poverty line is chosen, one can argue that it is arbitrary.
It is possible to propose a feasible alternative, which may change the
perspective of poverty significantly. Thus, one must examine the sensitiv-
ity of poverty with respect to the poverty line. One way of conducting the
sensitivity analysis is to change the poverty by certain percentages, then
estimate how much the poverty level has changed.
For example, suppose the headcount ratio of society x is 25 percent for
poverty line z = $10,000. Let this figure increase to 30 percent when the
poverty line is increased to $10,200. This means that a 2 percent increase in
the poverty line increases the headcount ratio by 5 percent. The lower the
change in the poverty estimate because of change in the poverty line, the
more reliable the point estimate based on a particular poverty line. If there is
too much variation, then the poverty estimate may not be considered reliable.
Growth and Poverty
When a country is rapidly growing, one must evaluate the quality of the
growth. By growth, we generally mean a country or society’s growth in mean
income, and, by merely looking at the growth, there is no way of knowing
who has benefited from this growth. This growth may result from a rise in
incomes of the richer part of the distribution or from a rise in incomes of the
poorer part of the distribution.
There are various ways of understanding if the growth is pro-poor or
anti-poor. First, we may be interested in knowing directly if poverty has
increased or decreased because of the growth. Second, we may want to
know if the growth has relatively benefited or hurt the population with lower
incomes. In this case, it is not enough just to understand if poverty has
increased or decreased; it is also important to understand whether the situ-
ation of the poor has changed in comparison to others in the distribution.
Third, we may be interested in knowing if the growth has lowered poverty
more than a counterfactual-balanced growth path would. In this case, one
may be interested in knowing how much of the change in poverty is due to
growth and how much is due to the redistribution.
Consider some examples to clarify these various ways of understand-
ing pro-poor growth. Suppose the society consists of four people and the
income vector is x = ($80, $100, $200, $260). The society’s mean income
is $160. First, if the poverty line income is $120, then two people are
poor. Suppose that, over time, incomes of these four people change to
141
x' = ($100, $125, $160, $575). The society’s mean income has grown by
50 percent to $240. If the poverty line remains unchanged at $120, then
the headcount ratio goes down. In fact, poverty goes down for any poverty
measure that satisfies the monotonicity property. Thus, if one is merely
interested in knowing if poverty has decreased because of growth, then the
growth has been pro-poor for a fixed poverty line. If, instead of $120, the
poverty line is set at $180, then the change in poverty may not appear to
be pro-poor by all measures. For example, despite growth of 50 percent, the
headcount ratio deteriorates. Thus, in terms of the headcount ratio, the
growth in the distribution appears to be anti-poor.20
Given that a fixed poverty line is difficult to defend, we must understand
the change in poverty for a variable poverty line. The approach is analogous
to the dominance analysis. If one poverty curve (incidence, deficit, or sever-
ity) dominates another poverty curve, then poverty has improved unambigu-
ously in the dominant distribution because of growth. Besides merely knowing
the direction of change in poverty, we may be interested in the magnitude of
the reduction in poverty relative to the growth in mean—the growth elastic-
ity of poverty. The growth elasticity of poverty is defined as the percentage
change in poverty resulting from a 1 percent change in the mean income.
If the elasticity is greater than one, then the percentage change in poverty
has been larger than the percentage change in mean income, or the growth
of mean income. For an application of the growth elasticity of poverty using
the headcount ratio, see Bourguignon (2003). To understand the change in
the growth or elasticity of poverty for a variable poverty line, various poverty
growth curves can be constructed (similar to the various growth curves dis-
cussed in the income standard section).
A second way of understanding a change in poverty as pro-poor is by look-
ing at the gain of the poor relative to the gain in the mean. Reconsider the two
income vectors in the previous example. The growth rate of the mean was
50 percent. Have the incomes of individuals at the bottom of the distribution
improved enough to catch up with the growth in mean? The answer is no. The
growth of the poorest person’s income was 25 percent. The income growth of
the two poorest people also totaled 25.0 percent, and the growth of the three
poorest people totaled 1.3 percent. Then how was the 50 percent growth
achieved? It was achieved because the richest person’s income grew by about
121 percent. Thus, this second way understands the relationship between
poverty and growth from an inequality perspective and may be referred to as
an inequality-based approach, as discussed in chapter 1.
142
The tools we used to understand the relationship between growth and

inequality can also be used here. Comparing the growth rates of two income
standards may provide some insight. If a lower income standard grows faster
than the mean, then incomes of the poorer section of the distribution must
have grown faster than the mean. In contrast, if an upper income standard
grows faster than the mean, then incomes of the richer section of the dis-
tribution must have grown faster than the mean. For example, one may
compare the growth rate of the Sen mean (emphasizing lower incomes) vis-
à-vis the growth of the average. Indeed, the growth in the Sen mean is only
24 percent compared to 50 percent growth in mean income.
One can also use other income standards, such as the general means,
for this exercise. For example, Foster and Székely (2008) computed the
growth in general means for different a to show that although the growth
rate mean incomes in Mexico and Costa Rica were the same, the growth
of general means was starkly different. In Mexico, the growth in mean
income was mostly driven by the increase in the income of the richer
section of the population. In Costa Rica, the growth in mean was driven
by the increase in the income of the poorer section. The same amount of
growth may have improved the situation of the poor in Costa Rica, but it
may have deteriorated the situation of the Mexican poor.
One may also be interested in understanding the composition of change
in poverty because of growth and because of change in inequality.21 As discussed
in chapter 1, pro-poor growth may be understood as a difference between
the growth rate of an original distribution and a counterfactual distribution
that has the same mean and relative distribution as the original distribution.
Then the overall change in poverty can be split into a change because of
growth and a change because of redistribution.
Consider the following simple example using the vectors above:
x = ($80, $100, $200, $260) and x' = ($100, $125, $160, $575). The mean
of x is $160, whereas the mean of x' is $240. We now rescale each element
of vector x' in such a way that it has the same mean as x, and we denote the
transformed vector by x". Thus, x" = (66.7, 83.3, 106.7, 383.3).
Let us simply measure poverty by the headcount ratio (this exercise can
be performed using any poverty measure). For the poverty line of $120, the
headcount ratio in x is 2/4, which decreases to 1/3 in x'.
How was this reduction obtained? Distribution x" is obtained from x by
redistribution while keeping the mean unchanged. The headcount ratio for
x", as a result, increases from two-fourths to three-fourths. Thus, poverty has
143
increased because of redistribution. However, distribution x' may be seen as

being obtained from distribution x" by merely increasing everyone’s income
by the same proportion with balanced growth. As a result, the headcount
ratio falls from three-fourths to one-fourth. Hence, the improvement in
poverty in this case has resulted from growth rather than redistribution.22
Exercises
1. Consider the following table that enables you to construct a cumula-

tive distribution function (cdf) from income data.
Number of
Category Income
people pi F(xi) (pi ë xi)
(i) ($ xi)
(ni)
1 12,000 10
2 13,000 15
3 14,000 40
4 15,000 20
5 16,000 15
There are five income categories (Xi) in the economy. Each category
contains a certain number of people (ni).
a. What is the total number of people (n) in the economy?
b. Let pi denote the proportion of people in each category. Fill in the
column corresponding to pi for each i. The probability mass function is
defined as a function that gives the probability of a discrete variable
taking the same value. Now draw the probability mass function.
Hint: Draw a diagram with x on the horizontal axis and p on the
vertical axis.
c. Let F(xi) denote the proportion of people who have an income no
higher than xi. Fill in the column corresponding to F(xi) for each i.
Now draw the cdf.
Hint: Draw a diagram with x on the horizontal axis and F(x) on
the vertical axis.
d. What is the relationship between pi and F(xi)?
e. Calculate the proportion of people having an income less than
$14,100. What is the proportion of people having an income more
than $14,900?
f. What is the average income for the economy?
g. Fill in the last column, and find the sum of all cells in that column.
What does the sum give you?
144
h. Use the cdf to calculate the area to the left of the cdf bounded by
x = 0 and F(x) = 1. What do you get?
i. Calculate the median, the 95th percentile, and the 20th percentile
using the cdf that you drew in 1c.
2. The Gini coefficient is probably the most commonly used index of
relative inequality. What are some of the advantages and disadvan-
tages of this measure?
3. The variance of logarithm (VL) is an inequality measure that is com-
puted as
1 N
VL (x) = ∑[ln x n − WL (x)]2,
N n =1
where WL(x) is the mean of the logarithm of elements in x as defined

in the chapter.
a. Verify that the variance of logarithms satisfies scale invariance. What
property of the variance of logarithms ensures scale invariance?
b. Graph the Lorenz curves for the two distributions x = (1,1,1,1,41)
and y = (1,1,1,21,21). Can the curves be ranked?
c. Find the variance of logarithms of the two distributions. What is
wrong here?
d. Find the mean log deviation (the second Theil measure) of the
two distributions. What is correct here?
4. Construct an inequality measure that violates replication invariance.
5. Are the following statements true, false, or uncertain? In each case,
support your answer with a brief but precise explanation.
a. The Kuznets ratios satisfy the Pigou-Dalton transfer principle.
b. Distribution y = (1,2,3,2,41) is more unequal than distribution
x = (1,8,4,1,36) in terms of the Lorenz criterion.
c. The four basic properties of inequality measurement are enough to
compare any two income distributions in terms of relative inequality.
d. If everyone’s income increases by a constant dollar amount,
inequality must fall.
6. Consider the distribution x = (1,3,6).
a. Draw the Lorenz curve, and calculate the area between the
45-degree line and the curve.
b. Calculate the Gini coefficient for x. What is the relationship
between the Gini coefficient and the calculated area?
145
7. Consider the distribution x = (3,6,9,12,24,36).

a. Divide the distribution into the following two subgroups: x1 = (3,6,9)
and x2 = (12,24,36). Calculate the Gini coefficient for x, x1, and x2.
Using the traditional additive decomposability formula, check if
the Gini coefficient is decomposable in this situation.
b. Divide distribution x into the following two subgroups:
x3 = (3,24,36) and x4 = (6,9,12). Again, using the traditional
additive decomposability formula, check if the Gini coefficient is
decomposable in this situation.
c. What is the difference between these two circumstances? Explain.
d. What is the residual for the Gini coefficient in these two
circumstances?
8. For the two distributions x = (2,100; 700; 1,100; 200) and y = (3,410;
620; 2,170; 6,510), do the following:
a. Calculate the WGM(.; a) and use it to calculate the Atkinson
measure IA(.; a) for a = 0, –1.
b. Do you have the same IA(.; a) for both distributions or not? What
is going on here?
9. For the income distributions x = (3,3,5,7) and y = (2,4,6,6), do the
following:
a. Calculate the generalized entropy measure and IGE(x; a) and
IGE(y; a) for a = 1,0,1,2,3,4.
b. Plot the values of a on the horizontal axis and the values of IGE(x; a)
and IGE(x; a) on the vertical axis.
c. Join the points, and check if they intersect. If they intersect, then
report at what value of a they intersect, and explain why.
10. Are the following statements true, false, or uncertain?
a. The second Theil measure is subgroup consistent.
b. The arithmetic mean is higher than the harmonic mean but less
than the geometric mean.
c. The sum of the decomposition weights of the generalized entropy
measure is always less than 1.
11. How is the generalized Lorenz curve GL(p) derived from a cdf? Draw
this process and explain. What value does the generalized Lorenz
curve take at p = 1?
12. Suppose an inequality measure is given by I(x) = (x̄—e(x))/x̄, where
e(x) is one of the equally distributed equivalent income functions used
by Atkinson (namely, a general mean with a parameter less than one).
146
a. Which equivalent income function is the lower income

standard?
b. Show that if the lower income standard grows at a faster rate than
the upper income standard, then inequality will fall.
c. Suppose the mean income grows at a rate of 3 percent. Under
what circumstances will the Atkinson index fall? When will the
Gini index fall?
13. Because of economic growth, the income distribution changes as fol-
lows over time: (1,1,1,1), (1,1,1,2), (1,1,2,2), (1,2,2,2), (2,2,2,2).
a. Explain the relevance of this example to the development literature.
b. Can unambiguous inequality comparisons be made between these
distributions?
c. How does the Gini coefficient change over time in this example?
14. Provide an example illustrating that the Gini coefficient violates
subgroup consistency. Explain why it does.
15. Country A has a more equal income distribution than Country B
such that Country A’s Lorenz curve dominates that of Country B.
a. What should be the relationship between these two countries in
terms of generalized Lorenz?
b. What does this finding say about welfare and inequality?
16. Why should a poverty measure be sensitive to the distribution of
income among the poor?
17. Suppose that the incomes in a population are given by x = (4,2,10)
and the poverty line is z = 6.
a. Find the number of people who are poor.
b. Find the headcount ratio PH.
c. Find the (normalized) poverty gap measure PG.
d. Find the squared poverty gap measure PSG.
e. If the income of person 2 falls by one unit so that the new distribu-
tion is y = (4,1,10), what happens to PH, PG, and PSG?
f. If person 2 gives person 1 a unit of income, resulting in distribution
u = (5,1,10), what happens to PH, PG, and PSG? Explain.
18. One of the big problems in evaluating poverty levels is arriving at a
single poverty line that represents the cutoff level between the poor
and the nonpoor. Many people believe that a poverty line must be
arbitrary to some extent. But if this is so, and if changing the pov-
erty line reverses poverty judgments, then all our conclusions about
poverty might be ambiguous. To solve this problem, we might make
147
comparisons not only for a single poverty line but also for a range
of poverty lines. Consider the three distributions from the previous
example: x = (4,2,10), y = (4,1,10), u = (5,1,10).
a. If z = 6 is the poverty line, does x or y have more poverty accord-
ing to the headcount ratio? Will this determination be reversed at
some other poverty line? Explain. Does x or y have more poverty
according to the poverty gap measure? Will this determination be
reversed at some other poverty line? Explain.
b. If z = 6 is the poverty line, does x or u have more poverty accord-
ing to the headcount ratio? Will this determination be reversed at
some other poverty line? Explain. Does x or u have more poverty
according to the poverty gap measure? Will this determination be
reversed at some other poverty line? Explain.
c. Do you think unambiguous comparisons with variable poverty
lines might be made in practice? If not, why not? If so, why?
19. Which inequality measure is the Sen-Shorrocks-Thon (SST) poverty
index based on?
a. Explain why the SST index is not subgroup consistent and provide
a counterexample to illustrate your point.
b. Which inequality measure is the Foster-Greer-Thorbecke (FGT)
index PSG(x; z) based on? Show that the measure is subgroup con-
sistent.
20. Why should a measure of poverty satisfy scale invariance (homoge-
neity of degree 0 in incomes and the poverty line)? Which poverty
measures satisfy scale invariance?
21. Suppose instead of the PSG(x; z) measure one were to use the
PMG(x; z) measure.
a. What is the main constructive difference between these two
measures?
b. What would be the advantages and disadvantages of using the
PMG(x; z) measure?
22. Why do inequality decompositions have a between-group term but
poverty decompositions do not?
23. Suppose inequality decreases without growth of mean income. What
may likely happen to poverty? Suppose growth of mean income
occurs without a change in inequality. What may likely happen to
poverty? Explain.
148
24. Suppose that the per capita poverty gap measure is used with a rela-
tive poverty line that sets z = αμ for some α > 0. When does one
distribution have a lower level of relative poverty for all α > 0? (Hint:
Think Lorenz.)
25. We have already shown that the poverty measures are different from
each other and differ in their sensitivity to a distribution. Please pro-
vide certain examples with illustrative distributions and poverty lines
such that
a. The SST index rises, but the three FGT indices fall.
b. The headcount ratio rises, but the SST index, poverty gap mea-
sure, and squared gap measure fall.
c. The poverty gap measure rises, but the headcount ratio, SST, and
squared gap measures fall.
d. The squared gap measure rises, but the headcount ratio, poverty
gap measure, and SST measure fall.
Notes
1. For further discussion on the use of consumption expenditure data ver-

sus income data, see Atkinson and Micklewright (1983) and Grosh and
Glewwe (2000).
2. For a more detailed discussion of some of these issues, see Deaton (1997).
3. For the concept and a more detailed discussion about the principle, see
Pigou (1912, 24–25); Dalton (1920); Atkinson (1970); Dasgupta, Sen,
and Starrett (1973); and Rothschild and Stiglitz (1973).
4. For further discussion of the concept, see Foster and Shorrocks (1991).
–
5. Going forward in this book, we will use the notation WA(x) and x inter-
changeably. They both denote the mean of distribution (x).
6. The measure was originally proposed by Sen (1976b) and thus we
named the income standard after him. See also Foster and Sen (1997).
7. A related property has been developed by Zheng (2007a). Called unit
consistency, it has a weaker requirement than the scale invariance
property. The unit consistency property requires that if one distribu-
tion is more unequal than another distribution, then just changing the
unit of measurement keeps the former distribution more unequal than
the latter. The property can be formally stated as follows: for any two
distributions x and x', if I(x) < I(x'), then I(cx) < I(cx') for any c > 0.
149
For example, if the elements of two distributions are converted from

Indian rupees to U.S. dollars, then the direction of inequality between
any two distributions should not change if the inequality measure satis-
fies unit consistency. An inequality measure that satisfies scale invari-
ance also satisfies unit consistency, but the converse is not necessarily
true. A class of decomposable inequality measures satisfying unit con-
sistency has been developed by Zheng (2007a). In this book, however,
we focus on relative inequality measures satisfying the scale invariance.
8. For a more in-depth theoretical discussion of the transfer sensitivity
property, see Shorrocks and Foster (1987).
9. A geographical interpretation of the residual term can be found in
Lambert and Aronson (1993), where the residual term is shown to be
an effect of the re-ranking effect. The inequality of a distribution is
computed in three steps: (a) within-group inequalities are computed
in each subgroup; (b) the groups are ranked by their mean incomes
and a concentration curve representing between-group inequalities is
constructed; and (c) the Lorenz curve is constructed. The difference
between the Lorenz curve of the distribution in the third step and the
concentration curve from the second step is known as the residual term.
10. The Lorenz curve was developed by Max Lorenz (1905).
11. Interested readers, who may desire to have further theoretical under-
standing of the properties and their interrelationship, should see Zheng
(1997) and Chakravarty (2009).
12. A related but weaker property has been developed by Zheng (2007b).
See note 8.
13. This axiom is also known in the literature as strong transfer (see Zheng
2000). However, to keep the terminologies comparable across sections,
we prefer to use the term transfer principle.
14. A weaker version of this property exists that is known in the literature
as weak transfer (see Chakravarty 1983), which can be stated as follows:
if distribution x' is obtained from distribution x by a regressive transfer
between two poor people while the poverty line is fixed at z and the
number of poor does not change, then P(x'; z) > P(x; z). If distribu-
tion x" is obtained from another distribution x by a progressive transfer
between two poor people while the poverty line is fixed at z and the
number of poor does not change, then PS(x"; z) < P(x; z). Note that this
property is different from the weak transfer principle that we define in
this book.
150
15. Previously, Sen (1976b) proposed the index PS(x; z) = PH[PIG + (1 − PIG)-
IGini(xq)], where xq is the income distribution of the poor only. This mea-
sure was modified later by Thon (1979) and Shorrocks (1995).
16. For a more elaborated discussion on various formulations of the SST
index, see Xu and Osberg (2003).
17. Rawls’s welfare function maximizes the welfare of society’s worse-off
member. “Social and economic inequalities are to be arranged ... to the
greatest benefit of the least advantaged...” (Rawls 1971, 302).
18. For an in-depth discussion on poverty ordering, see Atkinson (1987),
Foster and Shorrocks (1988), and Ravallion (1994).
19. Note that the poverty deficit curve and the generalized Lorenz curve
have an interesting relationship. They are based on the area under-
neath the cdf and the quantile function, where a quantile function is
an inverse of a cdf. See figure 2.7.
20. For various approaches to measuring pro-poor growth for a fixed poverty
line, see Kakwani and Son (2008).
21. For a discussion on the poverty-growth-inequality triangle, see
Bourguignon (2003).
22. The growth-redistribution decomposition becomes a bit more compli-
cated when there is interregional migration. For such decomposition
with change in population, see Huppi and Ravallion (1991). An appli-
cation of their method can be found in table 30 of chapter 3.
References
Atkinson, A. B. 1970. “On the Measurement of Inequality.” Journal of

Economic Theory 2 (1970): 244–63.
———. 1987. “On the Measurement of Poverty.” Econometrica 55 (4): 749–64.
Atkinson, A. B., and J. Micklewright. 1983. “On the Reliability of Income
Data in the Family Expenditure Survey 1970–1977.” Journal of the Royal
Statistical Society, Series A (146): 33–61.
Bourguignon, F. 2003. “The Growth Elasticity of Poverty Reduction: Explaining
Heterogeneity across Countries and Time Periods.” In Inequality and Growth:
Theory and Policy Implications, edited by T. Eicher and S. Turnovsky, 3–26.
Cambridge, MA: Massachusetts Institute of Technology.
Chakravarty, S. R. 1983. “A New Index of Poverty.” Mathematical Social
Sciences 6: 307–13.
151
———. 2009. Inequality, Polarization and Poverty: Advances in Distributional

Analysis. New York: Springer.
Clark, S., R. Hemming, and D. Ulph. 1981. “On Indices for the Measurement
of Poverty.” The Economic Journal 91 (362): 515–26.
Dalton, H. 1920. “The Measurement of the Inequality of Incomes.” The
Economic Journal 30: 348–61.
Dasgupta, P., A. Sen, and D. Starrett. 1973. “Notes on the Measurement of
Inequality.” Journal of Economic Theory 6 (2): 180–87.
Deaton, A. 1997. The Analysis of Household Surveys: A Microeconometric
Approach to Development Policy. Baltimore: World Bank.
Deaton, A., and S. Zaidi. 2002. “Guidelines for Constructing Consumption
Aggregates for Welfare Analysis.” Living Standards Measurement Study
Working Paper 135, World Bank, Washington, DC.
Foster, J. E., J. Greer, and E. Thorbecke. 1984. “A Class of Decomposable
Poverty Measures.” Econometrica 52 (3): 761–66.
Foster, J. E., and A. Sen. 1997. On Economic Inequality. 2nd ed. Oxford,
U.K.: Oxford University Press.
Foster, J. E., and A. F. Shorrocks. 1988. “Poverty Orderings.” Econometrica
56 (1): 173–77.
———. 1991. “Subgroup Consistent Poverty Indices.” Econometrica 59 (3):
687–709.
Foster, J. E., and M. Székely. 2008. “Is Economic Growth Good for the Poor?
Tracking Low Incomes Using General Means.” International Economic
Review 49 (4): 1143–72.
Gini, C. 1912. “Variabilità e mutabilità.” Reprinted in Memorie di metodolog-
ica statistica, edited by E. Pizetti and T. Salvemini. Rome: Libreria Eredi
Virgilio Veschi (1955).
Grosh, M., and P. Glewwe, eds. 2000. Designing Household Survey
Questionnaires for Developing Countries: Lessons from 15 Years of the Living
Standards Measurement Study. Washington, DC: World Bank.
Huppi, M., and M. Ravallion. 1991. “The Sectoral Structure of Poverty dur-
ing an Adjustment Period: Evidence for Indonesia in the Mid-1980s.”
World Development 19 (12): 1653–78.
Kakwani N., and H. H. Son. 2008. “Poverty Equivalent Growth Rate.”
Review of Income and Wealth 54 (4): 643–55.
Kanbur, R. 2006. “The Policy Significance of Inequality Decompositions.”
Journal of Economic Inequality 4 (3): 367–74.
152
Lambert, P., and J. R. Aronson. 1993. “Inequality Decomposition Analysis

and the Gini Coefficient Revisited.” The Economic Journal 103 (420):
1221–27.
Lorenz, M. O. 1905. “Methods of Measuring the Concentration of Wealth.”
Publications of the American Statistical Association 9 (70): 209–19.
Pigou, A. C. 1912. Wealth and Welfare. London: Macmillan.
Ravallion, M. 1994. Poverty Comparisons. Chur, Switzerland: Harwood
Academic Press.
Rawls, J. 1971. A Theory of Justice. Cambridge, MA: Harvard University Press.
Rothschild, M., and J. E. Stiglitz. 1973. “Some Further Results on the
Measurement of Inequality.” Journal of Economic Theory 6 (2): 188–204.
Sen, A. K. 1976a. “Poverty: An Ordinal Approach to Measurement.”
Econometrica 44 (2): 219–31.
———. 1976b. “Real National Income.” Review of Economic Studies 43 (1):
19–39.
Shorrocks, A. F. 1980. “The Class of Additively Decomposable Inequality
Measures.” Econometrica 48 (3): 613–25.
———. 1983. “Ranking Income Distributions.” Economica 50 (197): 3–17.
———. 1995. “Revisiting the Sen Poverty Index.” Econometrica 63 (5):
1225–30.
Shorrocks, A. F., and J. E. Foster. 1987. “Transfer Sensitive Inequality
Measures.” Review of Economic Studies 54 (3): 485–97.
Thon, D. 1979. “On Measuring Poverty.” Review of Income and Wealth 25
(4): 429–39.
Watts, H. W. 1968. “An Economic Definition of Poverty.” Discussion paper,
Institute for Research on Poverty, University of Wisconsin, Madison.
Xu, K. and L. Osberg. 2003. “The Social Welfare Implications,
Decomposability, and Geometry of the Sen Family of Poverty Indices.”
Canadian Journal of Economics 35 (1): 138–52.
Zheng, B. 1997. “Aggregate Poverty Measures.” Journal of Economic Surveys
11 (2): 123–62.
———. 2000. “Poverty Orderings.” Journal of Economic Surveys 14 (4):
427–66.
———. 2007a. “Unit-Consistent Decomposable Inequality Measures.”
Economica 74 (293): 97–111.
———. 2007b. “Unit-Consistent Poverty Indices.” Economic Theory 31 (1):
113–42.
153
Chapter 3
How to Interpret ADePT Results
In this chapter, we discuss how to interpret tables and graphs generated by

the ADePT analysis program. The chapter is organized in six sections:
• In the first section, we discuss how to interpret results at the country

level, decomposing across rural and urban areas.
• In the second and third sections, we move into analyses at a more
disaggregated level: across subnational regions in the second section
and across various population subgroups—such as household charac-
teristics, employment situation, and so forth—in the third section.
• In the fourth and fifth sections, we perform sensitivity and domi-
nance analyses. These are useful for policy evaluation, because results
in the first two sections are based on many assumptions, such as
choice of poverty line and selection of methodologies for measuring
poverty and inequality.
• It is always important to check how robust these results are with
respect to the assumptions. For example, we may assume the poverty
line to be a certain level of income or per capita expenditure and find
poverty decreasing over time. Then how can we be sure that poverty
has not increased for other possible poverty lines?
• Insights revealed in the first five sections may be helpful when prepar-
ing any report on poverty and inequality.
• In the final section, we discuss some advanced analyses.
155
Tables and graphs in this chapter were generated by ADePT’s Poverty

and Inequality modules using the Integrated Household Survey of Georgia
dataset for 2003 and 2006. Calculations assumed the equivalence scale
parameter is 1, which implies that every household member is assumed to
be adult equivalent. Hence, per capita expenditure was calculated by divid-
ing the total expenditure by the number of household members regardless of
their age and gender. Calculations assumed the economy-of-scale parameter
is 1. This implies that no economies of scale exist when two or more indi-
viduals share a household. (Other scale choices are, of course, possible, and
these parameters can be changed in ADePT.)
Consumption expenditures are in lari (or GEL, the Georgian national
currency) per month. Many tables use one or two poverty lines of GEL 75.4
and GEL 45.2 per month. In the first case, if a household fails to meet a
monthly consumption expenditure of GEL 75.4 for each member in that
household, then the household (and each member in the household) is
identified as poor. In the second case, a household is identified as poor if the
household fails to meet a per capita expenditure of GEL 45.2 per month.
Tables may have an occasional small numerical inconsistency. To
improve readability, ADePT displays data with a limited number of decimal
places by rounding the underlying raw data. This process can result in values
that appear incorrect, such as 29.9 + 1.0 = 31.0 (as opposed to 29.9 + 1.0 =
30.9, or 29.9 + 1.1 = 31.0). Spreadsheets generated by ADePT (the sources
for tables in this chapter) include raw data, which are visible in the formula
bar when a cell is selected.
Rounding numbers also affects how we present some of the results.
Certain poverty and inequality measures are traditionally reported in
decimals. However, this presentation does not provide us enough power to
differentiate between numbers. For example, the Gini coefficient of 0.26
and the Gini coefficient of 0.34 both may read as 0.3. Similarly, the FGT2
poverty index, or the squared poverty gap index, may take reasonable low
values in decimals such as 0.019 or 0.024. Again, these numbers may be
significantly different. Therefore, to improve readability, we normalize all
poverty and inequality figures in a 0–100 scale.
The text in this chapter has numerous references to table cells. To help
you quickly find data in tables, numbers and letters in brackets reference
table cells by row and column. For example, [3,E] refers to the cell in row 3,
column E.
156
Chapter 3: How to Interpret ADePT Results
Analysis at the National Level and Rural/Urban

Decomposition
While preparing a report on poverty and inequality, one would first be inter-
ested in results at the national level. This part of the chapter contains seven
tables with results at the national level. We then decompose the results
across urban and rural areas.
Income Distribution across the Population
Initially, understanding income distribution across the population is impor-

tant. A distribution’s density function is the percentage of population that
falls within a range of per capita expenditure. Figure 3.1 graphs the per
capita expenditure density function for urban Georgia. The vertical axis
shows probability density function of consumption expenditures. The hori-
zontal axis is per capita expenditure or any other equivalent achievement.
Figure 3.1: Probability Density Function of Urban Georgia
Urban
0.008
0.006
0.004
0.002
Median
0
0 200 400 600 800
Welfare aggregate
2003 2006
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
157
In figure 3.1, the solid curve is urban Georgia’s density function for 2003,
and the dotted curve is the density function of urban consumption expendi-
ture distribution for 2006. The median is an important income standard that
can be found in the diagram. It is indicated by the corresponding vertical
lines: solid line for 2003 and dotted line for 2006.
A density function can also be useful for understanding a distribution’s
skewness. As can be seen from figure 3.1, the density functions for both years
are positively skewed. However, an important change from 2003 to 2006 is
that more people mass around the distribution’s median in 2006. We can
also see that the density functions for both years are unimodal. When more
than one mode exists, a society is considered to be polarized by consumption
expenditure or income.
Standard of Living and Inequality across the Population
Table 3.1 reports the mean and median per capita consumption expenditure
and their growth over time, and the inequality across the population using
the Gini coefficient. It also decomposes them across rural and urban areas
and across two years: 2003 and 2006. Table rows denote three geographical
regions: urban area, rural area, and all of Georgia (row 3). Per capita con-
sumption expenditure is measured in lari per month.
Columns A and B report the mean per capita consumption expenditure
for 2003 and 2006, respectively. Column C reports the percentage change
or growth in per capita expenditure over the course of these three years. The
average per capita expenditure of the urban area in 2003 is GEL 128.9 [1,A],
which is larger than the average rural per capita expenditure of GEL 123.5
[2,A]. The mean urban per capita expenditure in 2006 is GEL 127.3 [1,B],
Table 3.1: Mean and Median Per Capita Consumption Expenditure, Growth, and the Gini
Coefficient

2003 2006 Growth 2003 2006 Growth Change
(GEL) (GEL) (%) (GEL) (GEL) (%) 2003 2006 (%)
Region A B C D E F G H I
1 Urban 128.9 127.3 −1.2 108.4 101.1 −6.8 33.5 35.6 2.2
2 Rural 123.5 124.8 1.0 101.5 105.3 3.7 35.3 35.1 −0.3
3 Total 126.1 126.0 −0.1 104.7 103.3 −1.4 34.4 35.4 0.9
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
158
which fell by 1.2 percent [1,C]. The mean rural per capita expenditure, in
contrast, increased by 1.0 percent to GEL 124.8 in 2006 [2,B]. Georgia’s
overall per capita consumption expenditure in 2003 is GEL 126.1 [3,A],
which fell by 0.1 percent to GEL 126.0 in 2006 [3,B].
Columns D, E, and F report the median per capita expenditures for
2003 and 2006 and their growth rates. The percentage changes in medians
or median growths are much larger than the mean per capita expenditure
growth. The rural median growth is 3.7 percent [2,F], whereas the urban
median “growth” is –6.8 percent [1,F]. The overall change in median is
–1.4 percent [3,F].
Columns G, H, and I use the Gini coefficient to capture inequality in
the distribution. The rural Gini coefficient has marginally fallen from 35.3
[2,G] to 35.1 [2,H], while the urban Gini coefficient over these three years
increased from 33.5 in 2003 [1,G] to 35.6 in 2006 [1,H]. The overall Gini
coefficient changed by 0.9 from 34.4 [3,G] to 35.4 [3,H]. (Gini coefficient
is reported on a scale from 0 to 100 in this chapter, rather than from 0 to 1.)
Lessons for Policy Makers
Note that the mean and the median, two different measures of standard of
living, are differently sensitive to the distribution of per capita consumption
expenditure. Mean is more sensitive to extreme values, whereas median is
more robust to extreme values. For example, if the only change in the dis-
tribution of per capita expenditure is at the highest quintile or the lowest
quintile, the change would be reflected by the mean, but the median would
not change. In contrast, in certain situations, when changes occur in the
middle of the distribution, mean per capita expenditures may remain unal-
tered, but the median may reflect the change.
It is important to analyze and understand the growth in both these
measures of central tendency. However, changes in different measures of
central tendency do not provide enough information about the change
in the overall distribution. They do not tell us how the spread or inequal-
ity within the distribution changes over time, which can be captured by
an inequality measure. In the above exercise, rural mean and median per
capita expenditure increased, but rural inequality marginally fell. On the
contrary, the urban inequality has increased over these three years from
33.5 in 2003 [1,G] to 35.6 in 2006 [1,H], while the mean and median
have fallen.
159
Overall Poverty
Table 3.2 examines the performance of groups of people considered poor.

It analyzes poverty in Georgia by decomposing across rural and urban areas
using three different poverty measures: headcount ratio, poverty gap measure,
and squared gap measure. These three poverty measures belong to the FGT
(Foster-Greer-Thorbecke) family of poverty measures. Table rows denote
three geographic regions: urban, rural, and all of Georgia (rows 3 and 6). The
variable is monthly per capita consumption expenditure in lari. There are
two poverty lines: GEL 75.4 per month and GEL 45.2 per month.
Columns A and B report headcount ratios for 2003 and 2006, respec-
tively. A region’s headcount ratio is the proportion of the population that
is poor compared to that region’s total population. When the poverty line
is GEL 75.4 per month, then the urban headcount ratio in 2003 is 28.1
percent [1,A]. This means that 28.1 percent of the population in the urban
area belongs to households that cannot afford the per capita consumption
expenditure of GEL 75.4 per month. The urban headcount ratio for 2006 is
30.8 percent [1,B]. Column C reports the change in urban headcount ratios
over the course of these three years, which is an increase of 2.7 percentage
points [1,C].
In contrast, the rural headcount ratio decreased by 0.5 percentage point
from 31.6 percent [2,A] in 2003 to 31.1 percent [2,B] in 2006. Overall,
Georgia’s poverty headcount has increased by 1.0 percentage point from
29.9 percent [3,A] to 31.0 percent [3,B]. Similarly, for the poverty line of
Table 3.2: Overall Poverty

percent
Headcount ratio Poverty gap measure Squared gap measure

2003 2006 Change 2003 2006 Change 2003 2006 Change
Poverty line = GEL 75.4
1 Urban 28.1 30.8 2.7 8.6 9.3 0.7 3.9 4.0 0.1
2 Rural 31.6 31.1 −0.5 10.7 10.9 0.2 5.2 5.5 0.3
3 Total 29.9 31.0 1.0 9.7 10.1 0.4 4.6 4.8 0.2
4 Urban 8.9 9.3 0.4 2.4 2.4 0.0 1.0 1.0 −0.1
5 Rural 11.4 12.1 0.7 3.6 4.0 0.3 1.7 1.9 0.2
6 Total 10.2 10.7 0.5 3.0 3.2 0.2 1.4 1.4 0.1
160
GEL 45.2 per month, Georgia’s headcount ratio increased from 10.2 percent
in 2003 [6,A] to 10.7 percent in 2006 [6,B]. The rural headcount ratio in this
case increased from 11.4 percent [5,A] to 12.1 percent [5,B]. This change
implies that the proportion of extreme poor (per capita expenditure below
GEL 45.2) in the rural area increased, but the proportion of nonextreme
poor (per capita expenditure between GEL 45.2 and GEL 75.4) decreased.
Columns D, E, and F analyze the poverty gap measure in 2003 and 2006.
The poverty gap measure lies between a minimum of 0 and a maximum of
100, where the minimum is when no one in a region is poor and the maxi-
mum is when everyone has zero consumption expenditure and the poverty
line is positive. When the poverty line is GEL 75.4, the urban area’s poverty
gap measure is 8.6 in 2003 [1,D], which increases by 0.7 to 9.3 in 2006 [1,E].
Likewise, the rural area’s poverty gap measure increases by 0.2 from 10.7 in
2003 [2,D] to 10.9 in 2006 [2,E]. The total increase in poverty gap measure
is 0.4 from 9.7 [3,D] to 10.1 [3,E]. When the poverty line is GEL 45.2, the
overall poverty gap measure increases by 0.2 from 3.0 in 2003 [6,D] to 3.2
in 2006 [6,E].
Columns G, H, and I analyze the squared gap measure. The squared gap
measure also lies between a minimum of 0 and a maximum of 100, where
the minimum is when no one in a region is poor and the maximum is when
everyone has zero consumption expenditure and the poverty line is positive.
This measure is sensitive to inequality across the poor. Column I shows
that the rural area’s squared gap measure when the poverty line is GEL 75.4
increased by 0.3 from 5.2 in 2003 [2,G] to 5.5 in 2006 [2,H]. For the rural
area it increased by 0.1 point from 3.9 [1,G] to 4.0 [1,H]. A similar pattern
of changes is visible for the lower poverty line.
Consider the situation when the poverty line is GEL 75.4. From column C,
one can see that the headcount ratio increased in the urban area by 2.7 per-
centage points and it decreased in the rural area by 0.5 percentage point. In
other words, the rural area performed better than the urban area in reducing
the proportion of poor people.
However, when we look at the poverty gap numbers, we see a different
scenario. It turns out, in fact, from column F that the poverty gaps for both
regions have registered increases, with the urban area registering a larger
increase (0.7 point increase in the urban area compared with 0.2 point
161
increase in the rural area). Thus, although the number of poor in the rural
area decreased, the same is not true when deprivation is measured in terms
of the average relative shortfall. Column F still reflects that the increase in
the rural poverty gap is lower than that of its urban counterpart. But col-
umn I shows that the increase in the squared gap measure is larger in the
rural area (0.3) than in the urban area (0.2), which implies that inequality
among the rural poor has been sufficiently high that despite a fall in the
headcount ratio, the increase in the squared gap measure is larger than that
in the urban area.
The change in the rural area’s headcount ratio is quite different when
the poverty line is GEL 45.2 per month. The increase in rural poverty is
much higher than the increase in urban poverty by all three measures. In
fact, the squared gap measure slightly decreases for the urban area. We con-
clude from this result that the situation for the rural area’s extreme poor has
actually worsened in 2006 compared with 2003.
Distribution of Poor across Rural and Urban Areas
Table 3.3 analyzes the distribution of population and poor people across
rural and urban areas. Table rows denote three geographic regions: urban,
rural, and all of Georgia (rows 3 and 6). The variable is per capita consump-
tion expenditure in l per month. There are two poverty lines: GEL 75.4 per
month and GEL 45.2 per month.
Columns A, B, and C analyze the headcount ratio, that is, the popula-
tion percentage that is poor. Columns A and B report the headcount ratio
Table 3.3: Distribution of Poor in Urban and Rural Areas

percent
Headcount ratio Distribution of the poor Distribution of population

Poverty Line = GEL 75.4
1 Urban 28.1 30.8 2.7 45.6 48.6 3.0 48.5 48.9 0.3
2 Rural 31.6 31.1 −0.5 54.4 51.4 −3.0 51.5 51.1 −0.3
3 Total 29.9 31.0 1.0 100.0 100.0 0.0 100.0 100.0 0.0
Poverty Line = GEL 45.2
4 Urban 8.9 9.3 0.4 42.4 42.3 −0.1 48.5 48.9 0.3
5 Rural 11.4 12.1 0.7 57.6 57.7 0.1 51.5 51.1 −0.3
6 Total 10.2 10.7 0.5 100.0 100.0 0.0 100.0 100.0 0.0
162
for the years 2003 and 2006, respectively, while column C reports the differ-
ence across these two years. Columns D, E, and F report the distribution of
poor people across rural and urban areas, with the number in the cell being
the proportion of poor people located in that region. Another way of seeing
this is as the region’s percentage contribution to poverty, or the headcount
ratio times the share of the region’s overall population divided by the overall
headcount ratio. Columns G, H, and I provide the population distribution
across rural and urban areas, or the percentage of the overall population
residing in that region.
The headcount ratio for the urban area’s population in 2003 is 28.1
percent [1,A]. In other words, 28.1 percent of the urban area popula-
tion is poor. The headcount ratio increased for urban Georgia in 2006 to
30.8 percent [1,B].
Of all poor people in Georgia in 2003, 45.6 percent [1,D] reside in
urban areas. The share of all poor people living in urban areas increases to
48.6 percent in 2006 [1,E]. This represents an increase of 3.0 percentage
points [1,F]. The shares of rural and urban area population do not change
much over the course of the three years. But when the poverty line is GEL
75.4 per month, the share of poor in urban areas increases in 2006 because
of the increase in headcount ratio.
This exercise has a very useful policy implication because the headcount
ratio does not provide any information about where most poor people live.
A region may have a lower headcount ratio, but if that region is highly
populated, then the number of poor may be high. Thus, policies should focus
on regions with high headcount ratios as well as regions with larger shares
of poor.
Composition of the FGT Family of Indices
Table 3.4 analyzes the composition of poverty figures reported in table 3.2.
Table rows denote three geographic regions: urban, rural, and all of Georgia
(rows 3 and 6). The variable is per capita consumption expenditure in lari
per month. There are two poverty lines: GEL 75.4 Lari per month and GEL
45.2 Lari per month.
163
Table 3.4: Composition of FGT Family of Indices by Geography
Headcount ratio Income gap Poverty gap GE(2) among Squared gap
(%) ratio measure the poor measure
Region A B C D E
2003
1 Urban 28.1 30.5 8.6 4.6 3.9
2 Rural 31.6 33.7 10.7 5.9 5.2
3 Total 29.9 32.3 9.7 5.3 4.6
2006
4 Urban 30.8 30.1 9.3 4.1 4.0
5 Rural 31.1 34.9 10.9 6.4 5.5
6 Total 31.0 32.6 10.1 5.3 4.8
2003
7 Urban 8.9 26.8 2.4 4.0 1.0
8 Rural 11.4 31.8 3.6 5.3 1.7
9 Total 10.2 29.7 3.0 4.7 1.4
2006
10 Urban 9.3 25.7 2.4 3.3 1.0
11 Rural 12.1 32.7 4.0 5.7 1.9
12 Total 10.7 29.7 3.2 4.7 1.4
The headcount ratio reports the proportion of people within a region

who are poor. The poverty gap measure and the squared gap measure can be
broken down as follows:
• The poverty gap measure is the headcount ratio multiplied by the

income gap ratio divided by 100.
• The income gap ratio is the average per capita expenditure shortfall
from the poverty line divided by the poverty line.
The squared gap (PSG) can be decomposed into three factors: headcount
ratio (PH), income gap ratio (PIG), and generalized entropy measure (GE)
for α = 2 among the poor, such that PSG = PH [P2IG + 2(1 − PIG)2 IGE (x; 2)].
These measures make possible a richer set of information for policy
analysis. An improvement in the poverty gap measure may result from a
reduction in the number of poor or a reduction in the average normalized
gap among the poor. Similarly, an improvement in the squared coefficient of
variation may result from a decrease in the number of poor, a decrease in the
164
average normalized gap among the poor, or a decrease in inequality among

the poor in terms of the generalized entropy measure.
For the GEL 75.4 per month poverty line, the poverty gap measure
for Georgia increased from 9.7 in 2003 [3,C] to 10.1 in 2006 [6,C]. This
increase comes from both a headcount ratio increase from 29.9 percent
[3,A] to 31.0 percent [6,A] and an income gap ratio increase from 32.3 [3,B]
to 32.6 [6,B]. However, the urban poverty gap measure increase derives
from an increase in the headcount ratio and a reduction in the income gap
ratio. In contrast, the rural poverty gap measure increase was a result of
an increase in the income gap ratio because the rural headcount ratio fell
slightly between 2003 and 2006.
Some interesting results are also evident when the poverty line is set
at GEL 45.2 per month. The urban poverty gap measure does not change
because an increase in the number of poor has been offset by an income
gap ratio decrease. In fact, the total poverty gap measure increase from 3.0
in 2003 [9,C] to 3.2 in 2006 [12,C] was caused solely by an increase in the
headcount ratio from 10.2 percent [9,A] to 10.7 percent [12,A], because the
income gap ratio remained unchanged at 29.7 [9,B] and [12,B].
The squared gap measure depends on another component: inequality among

the poor. Surprisingly, inequality among the poor does not change between
2003 and 2006 for both the higher and the lower poverty lines. For both
poverty lines and both years, inequality among the poor is higher in the
rural area. Thus, not only does the number of rural poor increase when the
poverty line is GEL 45.2, but also the average normalized shortfalls and
inequality across the poor go up.
Quantile Incomes and Quantile Ratios
Besides analyzing poverty, one must understand the situation of the rela-
tively poor population compared to the rest of the population. Table 3.5
reports five quantile per capita expenditures (PCEs) and certain quantile
ratios of per capita consumption expenditure for Georgia and its rural
and urban areas. It compares two different periods: 2003 and 2006. Table
rows denote three geographic regions: urban, rural, and all of Georgia
165
Table 3.5: Quantile PCEs and Quantile Ratios of Per Capita Consumption Expenditure
Percentile
Quantile ratio
10th 20th 50th (median, 80th 90th
(GEL) (GEL) GEL) (GEL) (GEL) 90-10 80-20 90-50 50-10
2003
1 Urban 47.4 64.1 108.4 182.1 229.6 79.3 64.8 52.8 56.3
2 Rural 42.2 58.8 101.5 173.1 230.0 81.6 66.0 55.9 58.4
3 Total 44.8 61.4 104.7 177.0 229.8 80.5 65.3 54.4 57.3
2006
4 Urban 46.7 61.2 101.1 174.0 231.3 79.8 64.8 56.3 53.8
5 Rural 41.0 58.5 105.3 175.9 229.1 82.1 66.8 54.0 61.1
6 Total 43.8 59.8 103.3 175.0 230.5 81.0 65.8 55.2 57.6
Note: PCE = per capita expenditure.
(rows 3 and 6). Per capita consumption expenditure is measured in lari

per month.
Columns A through E denote quantile PCE for five percentiles. Column
A denotes the quantile PCE at the 10th percentile, column B denotes the
quantile PCE at the 20th percentile, and so forth. Columns F through I
report the quantile ratios based on the quantile PCE reported in the first
five columns. Column F, for example, reports the 90/10 ratio, computed as
(quantile PCE at the 90th percentile – quantile PCE at the 10th percentile) /
quantile PCE at the 90th percentile. The larger the 90/10 ratio, the larger is
the gap between these two percentiles.
In 2003, the quantile PCE at the 10th percentile of Georgia is GEL 44.8
[3,A], implying that 10 percent of the Georgian population lives with per
capita consumption expenditure less than 44.8. Similarly, 20 percent of the
Georgian population lives with per capita consumption expenditure less than
61.4 [3,B]. In contrast, 10 percent of the Georgian population lives with per
capita expenditure more than GEL 229.8 [3,E], which is the 90th percentile.
The corresponding 90/10 quantile ratio using these two quantile PCEs
is 80.5 [3,F], which means that the gap between the two percentiles is
80.5 percent of the quantile PCE at the 90th percentile. Stated another
way, the quantile PCE at the 90th percentile is 100 / (100 – 80.5) = 5.1
times larger than the 10th percentile. Likewise, the quantile PCE at the
80th percentile of Georgia is GEL 177.0 [3,D], which is nearly three times
larger than the quantile PCE at the 20th percentile [3,B]. The correspond-
ing 80/20 measure is 65.3 [3,G]. Inequality between the quantile PCE at
166
the 90th percentile per capita expenditure and the quantile PCE at the
10th percentile is larger in the rural area (81.6 [2,F]) than in the urban area
(79.3 [1,F]) in 2003. The 90/10 measure increases for Georgia and both its
urban and rural areas in 2006 [4,F] and [5,A].
This table is helpful in holistically understanding inequality across the per

capita consumption expenditure distribution. The mean and median are
measures of a distribution’s central tendency and the distribution’s size, while
the Gini coefficient is a single measure of the overall distribution that does
not provide any information about which part of the distribution changed.
The four additional quantile PCEs reported in table 3.5 provide infor-
mation about different parts of the distribution. For example, the Gini
coefficient analysis in table 3.1 shows that inequality in the rural area has
decreased, whereas inequality in the urban area has increased. Which part
of the distribution is responsible for such changes? The Gini coefficient does
not provide an answer to this question. A decrease in inequality in the rural
area has not been obtained by increasing the income of the poorest because
the quantile PCE at the 10th percentile in the rural area fell to GEL 41.0 in
2006 [5,A] compared to GEL 42.2 in 2003 [2,A]. The quantile PCE at the
80th percentile increased from GEL 173.1 in 2003 [2,D] to GEL 175.9 in
2006 [5,D]. In other words, even though the Gini coefficient fell, inequality
between the quantile PCEs at the 80th percentile and the 20th percentile
increased in the rural area: from 66.0 in 2003 [2,G] to 66.8 in 2006 [5,G],
according to the 80/20 measure.
Table 3.6 reports two lower partial means, two upper partial means, and two
partial mean ratios, based on the partial means between two periods: 2003
and 2006. Table rows denote three geographic regions: urban, rural, and all
of Georgia (rows 3 and 6). Per capita consumption expenditure is measured
in lari per month.
Columns A and B report two lower partial means (LPM), columns C and
D report two upper partial means (UPM), and columns E and F report partial
mean ratios. The first partial mean ratio, for example, reports the 90/10
partial mean ratio, computed as (90th percentile UPM – 10th percentile
167
Table 3.6: Partial Means and Partial Mean Ratios
Lower partial mean Upper partial mean

Partial mean ratio
10th percentile 20th percentile 90th percentile 80th percentile
(GEL) (GEL) (GEL) (GEL) 90-10 80-20
Region A B C D E F
2003
1 Urban 34.5 45.2 319.5 261.8 89.2 82.7
2 Rural 29.0 39.9 321.3 259.1 91.0 84.6
3 Total 31.5 42.3 320.4 260.5 90.2 83.8
2006
4 Urban 34.5 44.3 347.7 273.5 90.1 83.8
5 Rural 27.8 39.0 317.0 258.2 91.2 84.9
6 Total 30.8 41.6 332.0 265.7 90.7 84.4
LPM) / 90th percentile UPM). The larger the 90/10 ratio, the larger is the
gap between these two partial means.
A lower partial mean is the average per capita expenditure of all people
below a specific percentile cutoff. An upper partial mean is the mean per
capita expenditure above a specific percentile. A partial mean ratio captures
inequality between a lower partial mean and an upper partial mean.
It is evident from the table that the average per capita expenditure of the
urban Georgian population’s poorest 20 percent is only GEL 45.2 in 2003
[1,B], whereas the average income of the population’s richest 20 percent is
GEL 261.8 [1,D]. The corresponding 80/20 partial mean ratio is 82.7 [1,F],
which means that the gap between the two partial means is 82.7 percent
of the 80th upper partial mean. Stated another way, the mean per capita
expenditure of the population’s richest 20 percent is 100 / (100 – 82.7) =
5.8 times larger than the mean per capita expenditure of the population’s
poorest 20 percent. Likewise, in rural areas, the mean per capita expendi-
ture of the population’s richest 20 percent (GEL 259.1 [2,D]) is 6.5 times
larger than the mean per capita expenditure of the population’s poorest
20 percent (GEL 39.9 [2,B]) in 2003. The corresponding 80/20 partial mean
ratio is 84.6 [2,F].
In table 3.5, we reported different percentiles of a distribution. For example,

the 10th percentile for Georgia in 2003 is GEL 44.8 [3,A], meaning that
168
10 percent of the Georgian population lives with a per capita expendi-

ture less than GEL 44.8. But what is the average income of these people?
Similarly in table 3.5, 10 percent of the Georgian population has a per
capita expenditure more than GEL 229.8 [3,E], which is the 90th percen-
tile for Georgia, but we do not know exactly how rich this group is. Partial
means are useful for answering this question, and the partial mean ratios tell
us the difference in the average per capita expenditures between a poorer
and a richer group.
Distribution of Population across Quintiles
Table 3.7 analyzes the population distribution in Georgia and its rural and
urban areas across five quintiles of per capita consumption expenditure.
It compares two time periods: 2003 and 2006. Table rows denote three
geographic regions: urban, rural, and all of Georgia (row 1). Per capita
consumption expenditure is measured in lari per month. Each of the five
columns denotes a quintile. Column A denotes the lowest, or first, quintile,
column B denotes the second quintile, and so forth.
All cells in row 1 have a value of 20, obtained by dividing Georgia’s
entire population into five equal groups in terms of per capita expenditure.
Each group contains 20 percent of the population. The fifth quintile con-
tains the richest 20 percent of the population, the fourth quintile consists
of the second-richest 20 percent of the population, and so on, and the first
quintile consists of the poorest 20 percent of the population.
Table 3.7: Distribution of Population across Quintiles

percent
Quintile
First Second Third Fourth Fifth
Region A B C D E
1 Total 20.0 20.0 20.0 20.0 20.0
2003
2 Urban 18.1 19.6 20.4 20.8 21.1
3 Rural 21.8 20.4 19.6 19.2 19.0
2006
4 Urban 19.0 21.6 20.6 19.2 19.7
5 Rural 21.0 18.5 19.4 20.8 20.3
169
Rows 2 and 3 report the population distribution in urban and rural areas
for 2003 using the national quintiles. Consider the value 18.1 [2,A] in the
urban row. This value implies that 18.1 percent of the total urban popula-
tion falls in the first quintile. The next cell is 19.6 [2,B], meaning that
19.6 percent of the total urban population falls in the second quintile. Similarly,
21.1 percent [2,E] of the total urban population falls in the fifth quintile.
The picture is slightly different for the rural area, where 19.0 percent
[3,E] of the total rural population falls in the fifth quintile and 21.8 per-
cent [3,A] falls in the lowest quintile. In 2006, the urban population share
in the first two quintiles increased to 19.0 percent [4,A] and 21.6 percent
[4,B], respectively, but the rural population share in the same two quintiles
decreased to 21.0 percent [5,A] and 18.5 percent [5,B], respectively. In
contrast, the rural population share in the two highest quintiles increased,
[3,D] and [3,E] compared with [5,D] and [5,E], but the urban population
share in the two highest quintiles decreased, [2,D] and [2,E] compared with
[4,D] and [4,E].
This table is helpful in understanding the population’s mobility across dif-

ferent consumption expenditure levels in different regions. A single welfare
measure—inequality or poverty—cannot reflect this mobility.
Analysis at the Subnational Level
Analyses in the previous section concentrate at the national level and

across rural and urban areas. For better policy implementation, we need to
understand the results at a more disaggregated level, such as across subna-
tional or geographic regions, or across population groups having different
characteristics.
In this section, we conduct subnational analysis, and in the next section,
we conduct analysis across other population subgroups. Some tables here are
similar to tables discussed in the previous section, and we occasionally refer
to those tables.
During the analysis across population subgroups, we assume the poverty
line to be the same across all subgroups. However, in the ADePT program
different poverty lines can be used for different subgroups in the analyses.
170
Standard of Living and Inequality
Table 3.8 results from calculating the mean and median per capita consump-
tion expenditure, and the Gini coefficient, for Georgia’s subnational regions.
Columns A and B report the mean per capita consumption expenditure for
years 2003 and 2006, respectively. Column C reports the percentage change
or growth in per capita expenditure over the course of these three years.
The mean per capita expenditure decreases for some regions (such as
Kakheti [1,C], Tbilisi [2,C], and Imereti [9,C]) and increases for others
(such as Shida Kartli [3,C], Kvemo Kartli [4,C], and Samtskhe-Javakheti
[5,C]). Imereti registers the steepest fall (7.0 percent [9,C]) in mean per
capita consumption expenditure, from GEL 150.3 in 2003 [9,A] to GEL
139.9 in 2006 [9,B]. In contrast, Kvemo Kartli reflects the highest increase
in mean per capita expenditure, 16.1 percent [4,C]. It increased from GEL
93.5 in 2003 [4,A] to GEL 108.5 in 2006 [4,B].
Columns D, E, and F report median per capita expenditures and their
growth. Although the change in overall median is −1.4 percent [11,F] (much
larger than the change in overall mean), changes in subnational regions are
mixed. For Kvemo Kartli, the growths of mean and median are almost the
same [4,C] and [4,F]. For Samtskhe-Javakheti, the growth in mean [5,C] is
three times larger than the growth of median [5,F]. In contrast, the growth
Table 3.8: Mean and Median Per Capita Income, Growth, and the Gini Coefficient across
Subnational Regions

2003 2006 Growth 2003 2006 Growth Change
(GEL) (GEL) (%) (GEL) (GEL) (%) 2003 2006 (%)
1 Kakheti 107.9 102.2 −5.2 92.7 80.4 −13.2 34.4 38.5 4.0
2 Tbilisi 144.5 143.1 −0.9 122.2 111.4 −8.8 32.1 36.4 4.3
3 Shida Kartli 122.9 125.6 2.3 98.7 101.7 3.0 36.6 35.9 −0.7
4 Kvemo Kartli 93.5 108.5 16.1 81.0 94.1 16.2 32.6 32.7 0.1
5 Samtskhe-Javakheti 116.5 121.5 4.3 98.8 100.3 1.5 32.9 31.1 −1.8
6 Ajara 107.8 101.8 −5.6 91.6 83.3 −9.0 33.9 34.4 0.4
7 Guria 134.3 125.6 −6.5 113.9 101.3 −11.1 33.9 35.0 1.1
8 Samegrelo 117.2 125.1 6.7 97.0 109.5 12.8 34.1 32.3 −1.9
9 Imereti 150.3 139.9 −7.0 128.6 122.4 −4.8 33.0 32.9 −0.1
10 Mtskheta-Mtianeti 113.0 123.6 9.3 103.7 96.7 −6.7 33.5 37.4 3.9
11 Total 126.1 126.0 −0.1 104.7 103.3 −1.4 34.4 35.4 0.9
171
of median in Samegrelo [8,F] is twice as large as the growth of mean per

capita consumption expenditure [8,C]. The most interesting pattern can be
seen for Mtskheta-Mtianeti, where the mean grows by 9.3 percent [10,C],
but the median falls by 6.7 percent [10,F].
Columns G, H, and I analyze inequality within subnational regions
using the Gini coefficient, which lies between 0 and 100. Although the
overall Gini coefficient has increased by 0.9 [11,I], a mixed picture is
found across subnational regions. In Tbilisi and Kakheti, inequality rises by
4.3 percent [2,I] and 4.0 percent [1,I], respectively. In Samtskhe-Javakheti,
inequality falls by 1.8 percent [5,I], while in Kvemo Kartli and Imereti, the
Gini coefficient changes by a meager 0.1 [5,I] and [9,I], going up and down,
respectively.
Headcount Ratio and the Distribution of Poor
Table 3.9 analyzes the headcount ratio of Georgia by population subgroup,

where each subgroup is classified by subnational regions—such as Kakheti,
Ajara, and Imereti—which could be states or provinces. The poverty line
for this table is GEL 75.4 per month (we use only one poverty line here, but
the analysis could be conducted for any number of poverty lines).
Table 3.9: Headcount Ratio by Subnational Regions, 2003 and 2006

percent
Headcount ratio Distribution of the poor Distribution of population

1 Kakheti 38.9 46.2 7.3 12.6 13.8 1.3 9.7 9.3 −0.4
2 Tbilisi 20.9 25.2 4.3 17.1 20.4 3.3 24.6 25.2 0.6
3 Shida Kartli 35.2 30.8 −4.5 8.3 7.2 −1.1 7.0 7.2 0.2
4 Kvemo Kartli 44.4 35.1 −9.3 16.8 12.2 −4.6 11.3 10.8 −0.5
5 Samtskhe-Javakheti 30.0 24.4 −5.7 4.6 3.8 −0.8 4.6 4.8 0.2
6 Ajara 37.1 44.6 7.5 10.7 13.7 2.9 8.7 9.5 0.8
7 Guria 25.3 34.4 9.2 2.7 3.5 0.7 3.2 3.1 −0.1
8 Samegrelo 33.5 29.4 −4.1 11.8 9.0 −2.8 10.5 9.5 −1.1
9 Imereti 20.6 23.0 2.3 12.1 13.4 1.3 17.5 18.0 0.5
10 Mtskheta-Mtianeti 34.3 35.2 0.9 3.3 3.1 −0.2 2.9 2.7 −0.2
11 Total 29.9 31.0 1.0 100.0 100.0 n.a. 100.0 100.0 n.a.
Note: n.a. = not applicable.
172
Table rows list subnational regions. Columns A, B, and C analyze

headcount ratios. Columns D, E, and F outline the distribution of poor
people across the subgroups, with the number in the cell being the pro-
portion of all poor people in the country that are included in that sub-
group. Another way of seeing this is the percentage contribution of the
subgroup to overall poverty, or the headcount ratio times the population
share in that group, divided by the overall headcount ratio. Columns G,
H, and I depict the population distribution in subnational regions, or the
percentage of the population that resides in that region. Row 11 shows
that the overall headcount ratio increases from 29.9 percent in 2003
[11,A] to 31.0 percent in 2006 [11,B], reflecting a 1.0 percentage point
(rounded) increase.
In cell [1,A], we find that in 2003, 38.9 percent of the population in
Kakheti is poor. In other words, the headcount ratio for this population
subgroup is 38.9 percent. Cell [1,B] is 46.2, the headcount ratio for the
same population subgroup in 2006. Thus, the headcount increased by
7.3 percentage points [1,C] over the course of these three years. In row 4, we
see that Kvemo Kartli’s headcount ratio decreased by 9.3 percentage points,
from 44.4 percent [4,A] to 35.1 percent [4,B]. The headcount ratio also fell
between 2003 and 2006 in other regions, such as Shida Kartli [3,C] and
Samtskhe-Javakheti [5,C].
Cell [1,D] is 12.6, meaning that of all poor people in Georgia in 2003,
12.6 percent can be found in Kakheti. The share of all poor living in
Kakheti increases to 13.8 percent in 2006 [1,E], an increase of 1.3 percentage
points.
Now compare Kvemo Kartli and Imereti. Clearly, Kvemo Kartli’s pov-
erty headcount ratio (44.4 percent [4,A]) is more than twice as large as
Imereti’s poverty headcount ratio (20.6 percent [9,A]) in 2003. However,
the share of all poor people is only around 40 percent larger in Kvemo
Kartli (16.8 percent in Kvemo Kartli [4,D], compared with 12.1 percent
in Imereti [9,D]). This is due to the different population shares of the two
regions as given in the table’s final columns. The population share living
in Imereti in 2003 is 17.5 percent [9,G], while the Kvemo Kartli share
is only 11.3 percent [4,G]. Therefore, a policy maker should take into
account a region’s population share in addition to the headcount ratio,
because a region may have a lower headcount ratio because of a higher
number of poor.
173
Poverty Gap Measure and Subnational Contribution

to Overall Poverty
Table 3.10 analyzes Georgia’s poverty gap measure across subnational regions.
The poverty line is GEL 75.4 per month. Table rows list subnational regions.
Columns A, B, and C analyze poverty gap measures for 2003, 2006, and the
changes over time. Columns D, E, and F report the percentage contribution
of the subnational regions to the overall poverty gap measure. Columns G,
H, and I depict the population distribution of the subnational regions, or the
percentage of the overall population that resides in each region.
The overall poverty gap measure increases from 9.7 in 2003 [11,A] to
10.1 in 2006 [11,B], reflecting a 0.4 point increase [11,C]. For Kakheti,
the poverty gap measure in 2003 is 13.4 [1,A]. The poverty gap measure
for the same population subgroup in 2006 is 17.8 [1,B]. Thus, the poverty
gap measure increased by 4.4 points [1,C] over three years. The poverty gap
measure in Kvemo Kartli decreased by 3.5 points, from 15.4 in 2003 [4,A] to
11.9 in 2006 [4,B]. The poverty gap measure also fell between 2003 and 2006
in other regions, such as Samegrelo [8,C] and Mtskheta-Mtianeti [10,C].
Kakheti’s contribution to the overall poverty gap measure is 13.4 percent
[1,D]. Its contribution increased to 16.3 percent in 2006 [1,E], an increase
of 2.9 percentage points [1,F].
Table 3.10: Poverty Gap Measure by Subnational Regions
Contribution to Distribution of
Poverty gap measure overall poverty (%) population (%)
1 Kakheti 13.4 17.8 4.4 13.4 16.3 2.9 9.7 9.3 −0.4
2 Tbilisi 5.5 7.3 1.8 14.0 18.2 4.2 24.6 25.2 0.6
3 Shida Kartli 11.7 10.9 −0.8 8.5 7.8 −0.7 7.0 7.2 0.2
4 Kvemo Kartli 15.4 11.9 −3.5 18.1 12.8 −5.3 11.3 10.8 −0.5
6 Ajara 12.8 14.6 1.8 11.5 13.7 2.2 8.7 9.5 0.8
7 Guria 8.3 10.6 2.3 2.8 3.3 0.5 3.2 3.1 −0.1
8 Samegrelo 11.0 8.8 −2.2 12.0 8.2 −3.8 10.5 9.5 −1.1
9 Imereti 6.1 7.5 1.4 11.1 13.4 2.4 17.5 18.0 0.5
10 Mtskheta-Mtianeti 13.1 11.7 −1.4 3.9 3.1 −0.8 2.9 2.7 −0.2
11 Total 9.7 10.1 0.4 100.0 100.0 n.a. 100.0 100.0 n.a.
174
Now compare Guria and Imereti. Clearly, Guria’s poverty gap measure
(8.3 [7,A]) is larger than Imereti’s poverty gap measure (6.1 [9,A]) in 2003.
But Guria’s contribution is only 2.8 percent [7,D], whereas Imereti’s contri-
bution is 11.1 percent [9,D]. The contribution of subnational regions to the
overall poverty gap and the share of poor in each region are quite different.
The share of poor in each of Kakheti and Ajara is almost identical in 2006
(9.3 percent for Kakheti [1,H], compared with 9.5 percent in Ajara [6,H]),
but their contributions to the total poverty gap measure are quite different
(16.3 percent in Kakheti [1,E], compared with 13.7 percent in Ajara [6,E]).
Thus, the average normalized shortfall of per capita expenditure from the
poverty line is much higher in Kakheti, and that is not captured by the
headcount ratio analysis.
Squared Gap Measure and Subnational Contribution

to Overall Poverty
Table 3.11 analyzes Georgia’s squared gap measure across subnational

regions. The poverty line is GEL 75.4 per month. Table rows list subna-
tional regions. Columns A, B, and C analyze the squared gap measure for
2003, 2006, and the difference over time. Columns D, E, and F report the
Table 3.11: Squared Gap Measure by Subnational Regions
Contribution to overall Distribution of

Squared gap measure poverty (%) population (%)
2003 2006 Change (%) 2003 2006 Change 2003 2006 Change
1 Kakheti 6.6 9.4 2.7 14.0 18.2 4.2 9.7 9.3 −0.4
2 Tbilisi 2.1 3.0 0.9 11.4 15.9 4.6 24.6 25.2 0.6
3 Shida Kartli 6.0 5.5 −0.6 9.3 8.2 −1.1 7.0 7.2 0.2
4 Kvemo Kartli 7.8 6.2 −1.7 19.4 13.9 −5.5 11.3 10.8 −0.5
6 Ajara 6.4 6.8 0.5 12.1 13.6 1.5 8.7 9.5 0.8
7 Guria 3.7 4.6 0.9 2.6 3.0 0.4 3.2 3.1 −0.1
8 Samegrelo 5.2 3.7 −1.4 11.9 7.4 −4.5 10.5 9.5 −1.1
9 Imereti 2.7 3.6 0.9 10.3 13.6 3.3 17.5 18.0 0.5
10 Mtskheta-Mtianeti 6.8 5.9 −0.8 4.2 3.3 −0.9 2.9 2.7 −0.2
11 Total 4.6 4.8 0.2 100.0 100.0 n.a. 100.0 100.0 n.a.
175
percentage contribution of the subnational regions to the overall squared

gap measure. Columns G, H, and I depict the population distribution of
the subnational regions, or the percentage of the overall population that
resides in each region. Row 11 shows that the overall squared gap mea-
sure increased from 4.6 in 2003 [11,A] to 4.8 in 2006 [11,B], reflecting a
0.2 point increase [11,C].
The squared gap measure for Kakheti is 6.6 in 2003 [1,A]. The squared
gap measure for the same population subgroup is 9.4 in 2006 [1,B]. Thus,
the squared gap measure increased by 2.7 points in three years [1,C]. The
squared gap measure in Kvemo Kartli decreased by 1.7 points, from 7.8 in
2003 [4,A] to 6.2 in 2006 [4,B]. The squared gap measure also fell between
2003 and 2006 in other regions, such as Samegrelo [8,C] and Mtskheta-
Mtianeti [10,C]. Kakheti’s contribution to the overall squared gap measure
in 2003 is 14.0 percent [1,D]. The contribution increased to 18.2 percent in
2006 [1,E], an increase of 4.2 percentage points [1,F].
Comparing the contribution of subnational regions to the overall squared

gap measure to the contribution to the overall squared gap measure and
the share of poor in each region, we see they are not necessarily the same.
Tbilisi’s contribution to overall poverty in 2006 is larger than Kakheti’s
contribution when poverty is measured by headcount ratio and poverty
gap measure, but Tbilisi’s contribution is lower in 2006 (3.0 [2,B]) than
that of Kakheti (9.4 [1,B]) when poverty is measured using the squared gap
measure. This finding may reflect that inequality across the poor, captured
by the squared normalized shortfalls, is much higher in Kakheti, and that
is not captured by the analysis based on headcount ratio or poverty gap
measure.
Quantile Incomes and Quantile Ratios
In addition to analyzing poverty, understanding how a population’s poor

segment compares to the rest of the population is important. Table 3.12
reports quantile per capita expenditure for five percentiles and certain
quantile ratios of per capita consumption expenditure for Georgia’s sub-
national regions in 2003. Each of the first five columns denotes a quantile
PCE. Column A denotes the quantile PCE at the 10th percentile, column
176
Table 3.12: Quantile PCE and Quantile Ratio of Per Capita Consumption Expenditure, 2003
Quantile PCE
50th
Quantile ratio
10th 20th percentile 80th 90th
percentile percentile (median, percentile percentile 90-10 80-20 90-50 50-10
(GEL) (GEL) GEL) (GEL) (GEL) (%) (%) (%) (%)
1 Kakheti 37.8 52.6 92.7 150.4 191.1 80.2 65.0 51.5 59.2
2 Tbilisi 56.0 74.3 122.2 202.8 252.9 77.9 63.3 51.7 54.2
3 Shida Kartli 38.6 55.9 98.7 169.8 228.4 83.1 67.1 56.8 60.9
4 Kvemo Kartli 34.3 48.3 81.0 126.5 165.1 79.2 61.8 51.0 57.7
5 Samtskhe-Javakheti 43.0 61.2 98.8 160.5 190.2 77.4 61.9 48.0 56.5
6 Ajara 37.8 53.1 91.6 146.5 203.3 81.4 63.7 54.9 58.7
7 Guria 47.7 64.0 113.9 189.1 241.9 80.3 66.1 52.9 58.1
8 Samegrelo 41.2 56.2 97.0 160.7 208.5 80.2 65.0 53.5 57.5
9 Imereti 54.0 74.1 128.6 211.6 267.0 79.8 65.0 51.8 58.0
10 Mtskheta-Mtianeti 33.9 52.5 103.7 162.0 200.1 83.1 67.6 48.2 67.3
11 Total 44.8 61.4 104.7 177.0 229.8 80.5 65.3 54.4 57.3
Note: PCE = per capita expenditure.
B denotes the quantile PCE at the 20th percentile, column C denotes the
median, column D denotes the quantile PCE at the 80th percentile, and
column E denotes the quantile PCE at the 90th percentile.
Columns F through I report the quantile ratios based on the quantiles
reported in the first five columns. Column G, for example, reports the 80/20
ratio, computed as (quantile PCE at the 80th percentile – quantile PCE at
the 20th percentile) / quantile PCE at the 80th percentile. The larger the
80/20 ratio, the larger is the gap between these two percentiles.
In 2003, the quantile PCE at the 10th percentile of Kakheti is 37.8 [1,A],
which implies that 10 percent of the population in Kakheti lives with per
capita consumption expenditure less than GEL 37.8. Similarly, 20 percent
of Kakheti’s population lives with per capita consumption expenditure less
than GEL 52.6 [1,B]. In contrast, 10 percent of people in Kakheti live with
per capita expenditure more than GEL 191.1 [1,E], the quantile PCE at the
90th percentile. The corresponding 90/10 measure using these two quantile
PCEs is 80.2 [1,F], meaning that the gap between the two quantile PCEs is
80.2 percent of the quantile PCE at the 90th percentile. Described another
way, the quantile PCE at the 90th percentile is 100 / (100 – 80.2) = 5.1
times larger than the quantile PCE at the 10th percentile.
Likewise, the quantile PCE at the 80th percentile of Kakheti is
GEL 150.4 [1,D], nearly three times larger than the quantile PCE at the
177
20th percentile per capita expenditure [1,B]. It is evident that Shida Kartli
has a lower quantile PCE at the 10th percentile than Samegrelo but a larger
quantile PCE at the 90th percentile. As a result, the 90/10 quantile ratio of
Shida Kartli [3,F] is higher than the 90/10 quantile ratio of Samegrelo [8,F].
This table is helpful in holistically understanding inequality across the per

capita consumption expenditure distribution. The mean and median measure
a distribution’s central tendency and measure. The Gini coefficient is a single
measure of the overall distribution, but it does not provide any information
about which part of the distribution has changed. The four additional quan-
tile PCEs reported in the table provide further information about different
parts of the distribution.
Table 3.13 reports two lower partial means, two upper partial means, and
two partial mean ratios for Georgia’s subnational regions in 2003. Columns
A and B report the two lower partial means, columns C and D report the two
upper partial means, and columns E and F report the partial mean ratios.
The first of the partial mean ratios, for example, reports the 90/10 partial
Table 3.13: Partial Means and Partial Mean Ratios for Subnational Regions,
2003
Lower partial mean Upper partial mean Partial mean ratio (%)
p10 p20 p20 p10 90-10 80-20
Region A B C D E F
1 Kakheti 25.6 35.9 222.3 276.0 90.7 83.8
2 Tbilisi 44.1 54.9 286.7 348.8 87.3 80.9
3 Shida Kartli 26.2 37.1 263.3 331.0 92.1 85.9
4 Kvemo Kartli 23.9 32.3 186.7 230.9 89.6 82.7
5 Samtskhe-Javakheti 30.5 41.5 234.4 294.8 89.6 82.3
6 Ajara 26.2 36.5 222.4 273.4 90.4 83.6
7 Guria 35.9 45.8 275.3 337.2 89.4 83.4
8 Samegrelo 30.8 40.1 241.1 302.9 89.8 83.4
9 Imereti 39.8 52.4 299.1 362.1 89.0 82.5
10 Mtskheta-Mtianeti 25.0 34.7 222.7 265.7 90.6 84.4
11 Total 31.5 42.3 260.5 320.4 90.2 83.8
178
mean ratio, computed as (90th percentile UPM – 10th percentile LPM) /

90th percentile UPM. The larger the 90/10 partial mean ratio, the larger is
the gap between these two partial means.
A lower partial mean is the average per capita expenditure of all people
below a specific percentile cutoff. An upper partial mean is the mean per
capita expenditure above a specific percentile. A partial mean ratio captures
inequality between a lower partial mean and an upper partial mean. In
table 3.5, we reported a distribution’s different quantile PCEs. For example,
the quantile PCE at the 10th percentile of Georgia in 2003 was GEL 44.8,
meaning that 10 percent of the Georgian population lives with per capita
expenditure less than GEL 44.8. However, that does not tell us the average
income of these people. Similarly, 10 percent of the Georgian population
has per capita expenditure more than GEL 229.8, Georgia’s quantile PCE
at the 90th percentile, but we do not know exactly how rich this group
is. Partial means are useful for determining these values, and partial mean
ratios tell us the difference in the average per capita expenditures between
a poorer and a richer group.
It is evident from table 3.13 that the average per capita expenditure of
the poorest 20 percent of people in Ajara is only GEL 36.5 in 2003 [6,B],
whereas the average income of the richest 20 percent of the population is
GEL 222.4 [6,C]. The corresponding 80/20 partial mean ratio is 83.6 [6,F],
meaning that the gap between the two partial means is 83.6 percent of the
80th upper partial mean. Stated another way, the mean per capita expendi-
ture of the population’s richest 20 percent is 100 / (100 – 83.6) = 6.1 times
larger than the mean per capita expenditure of the population’s poorest
20 percent. Likewise, in Shida Kartli, the mean per capita expenditure of
the population’s richest 20 percent (GEL 263.3 [3,C]) is 7.1 times larger
than the mean per capita expenditure of the population’s poorest 20 percent
(GEL 37.1 [3,B]) in 2003. The corresponding 80/20 partial mean ratio is
85.9 [3,F].
A larger inequality in terms of the quantile ratio does not necessarily trans-
late into higher inequality in terms of the partial mean ratio. In table 3.12,
we found that the 80/20 quantile ratio for Imereti (65.0) was larger than
that of Ajara (63.7), but in table 3.13 Ajara’s 80/20 partial mean ratio (83.6
[3,F]) is slightly larger than Imereti’s (82.5 [9,F]).
179
Distribution of Population across Quintiles by Subnational

Region
Table 3.14 analyzes the population distribution in subnational regions across

five quintiles of per capita consumption expenditure. Column 1 denotes the
lowest or the first quintile, column 2 denotes the second quintile, and so
forth.
entire population into five equal-sized groups in terms of per capita expen-
diture. Each group contains 20 percent of the population. The fifth quintile
contains the richest 20 percent of the population; the fourth quintile con-
sists of the second-richest 20 percent of the population, and so on, and the
first quintile consists of the poorest 20 percent of the population.
For the subnational regions, table cells report population percentage in
each quintile. Consider the value 27.6 for Kakheti [2,A]. This value implies
that 27.6 percent of Kakheti’s population lives with per capita expenditure
less than the first quintile. The next cell to the right is 20.9 [2,B], imply-
ing that 20.9 percent of Kakheti’s population falls in the second quintile.
Similarly, only 12.5 percent [2,E] of Kakheti’s population falls in the fifth
quintile.
The picture is slightly different for Imereti, where only 13.3 percent
[10,A] and 15.3 percent [10,B] of its population fall in the first and second
Table 3.14: Distribution of Population across Quintiles by Subnational

Region, 2003
percentage of population
Quintile
Region A B C D E
1 Total 20.0 20.0 20.0 20.0 20.0
2 Kakheti 27.6 20.9 20.8 18.3 12.5
3 Tbilisi 12.4 17.9 19.9 22.5 27.2
4 Shida Kartli 23.0 21.7 17.0 19.7 18.5
5 Kvemo Kartli 30.0 27.5 21.2 13.5 7.9
6 Samtskhe-Javakheti 20.1 24.0 21.6 20.2 14.1
7 Ajara 25.9 22.9 21.7 15.6 13.8
8 Guria 17.4 17.7 20.7 20.4 23.8
9 Samegrelo 23.5 19.8 20.6 21.3 14.7
10 Imereti 13.3 15.3 18.2 22.9 30.3
11 Mtskheta-Mtianeti 25.5 17.5 20.2 21.2 15.6
180
quintiles, respectively, and 30.3 percent [10,E] of its population falls in the
fifth or richest quintile. Kvemo Kartli appears to be the poorest among all
subnational regions because 30.0 percent [5,A] of its population falls in the
poorest quintile and only 7.9 percent [5,E] of its population falls in the rich-
est quintile.
This table is helpful in understanding population mobility across different

consumption expenditure levels in different regions, which a single measure
of welfare, inequality, or poverty cannot reflect.
Subnational Decomposition of Headcount Ratio
Table 3.15 decomposes poverty to explore the factors that caused a change
in headcount ratio. Table rows are divided into two categories. Rows 1
through 4 report the change in the overall poverty and three factors affect-
ing this change: total intrasectoral effect, population-shift effect, and inter-
action effect. Rows 5 through 14 report the intrasectoral effects for various
regions in Georgia.1 Column A reports the absolute change in headcount
Table 3.15: Subnational Decomposition of Headcount Ratio, Changes

between 2003 and 2006
Absolute change Percentage change

A B
1 Change in headcount ratio 1.04 100.00
2 Total intrasectoral effect 1.09 104.98
3 Population-shift effect −0.18 −17.38
4 Interaction effect 0.13 12.40
Intrasectoral effects by region
5 Kakheti 0.70 67.93
6 Tbilisi 1.06 102.38
7 Shida Kartli −0.31 −30.37
8 Kvemo Kartli −1.05 −101.76
9 Samtskhe-Javakheti −0.26 −25.06
10 Ajara 0.65 62.79
11 Guria 0.30 28.70
12 Samegrelo −0.43 −41.25
13 Imereti 0.41 39.18
14 Mtskheta-Mtianeti 0.03 2.44
181
poverty and the size of the factors contributing to this change. Column B
shows how these factors change the headcount ratio.
The change in overall headcount ratio between 2003 and 2006 is 1.04
[1,A]. This overall change of 1.04 percentage points is divided into three
different effects. The first is the total intrasectoral effect, 1.09 [2,A]. The
total population-shift effect is negative and amounts to –0.18 [3,A]. The
interaction between the intrasectoral factor and the population shift fac-
tor is 0.13 [4,A]. If we sum these three effects, we get the overall absolute
change in headcount ratio poverty: (1.09 – 0.18 + 0.13) = 1.04 [1,A].
The next column reports the proportion these effects have relative to the
overall change. The proportion of the total intrasectoral effect on the over-
all change in poverty is 104.98 percent [2,B]. This number is calculated by
dividing the total intrasectoral effect by the change in poverty: (100 × 1.09)
/ 1.04 = 104.98. The corresponding entries for the population-shift effect
and the interaction effect are calculated by the same method. For example,
(100 × –0.18) / 1.04 = –17.38 and (100 × 0.13) / 1.04 = 12.40 [4,B].
The next set of results decomposes the total intrasectoral effect across
Georgia’s regions. Column A reports the size of the intrasectoral effect, and
column B reports the intrasectoral effect as a proportion of the total change
in the overall headcount ratio. For example, the intrasectoral effect for
Kakheti is 0.70 [5,A], and its proportion of the overall poverty change is
67.93 percent [5,B], calculated as (100 × 0.70) / 1.04 = 67.93.
The intrasectoral effect of Kakheti is calculated as the change in
headcount ratio between 2003 and 2006, which is 7.3 percentage points
(reported in column C of table 3.9), multiplied by its population share in
2003 (reported in column G of table 3.9). The intrasectoral effects are nega-
tive for regions such as Shida Kartli, Kvemo Kartli, Samtskhe-Javakheti, and
Samegrelo, because the poverty headcount ratio fell in these regions. For the
rest of the regions, the intrasectoral effects are positive. The contribution of
this effect is highest for Tbilisi and lowest for Kvemo Kartli.
The total intrasectoral effect is even higher than the total change in the
overall headcount ratio. Thus, if the region-wise population shares are
kept constant, then the change in poverty is 1.09 percentage points [2,A].
However, if we keep the regional headcount ratios constant and consider
182
only the changes in regional population shares, then the poverty rate would
have fallen by 0.18 percentage point [3,A]. Thus, the intrasectoral effect
dominates and the overall headcount ratio increases. Finally, the second set
of results gives us an idea about the headcount ratio’s regional contribution
in terms of intrasectoral effect.
Poverty Analysis across Other Population Subgroups
In this section, we discuss the results when the population is divided in various
ways: household head’s characteristics, household member’s employment sta-
tus, education level, age group, demographic composition, and landownership.
Standard of Living and Inequality by Household Head’s

Characteristics
Table 3.16 reports the mean and median per capita consumption expendi-
ture and their growth over time and inequality across the population using
the Gini coefficient across various household characteristics. Table rows
denote various household characteristics. Columns A and B report the
mean per capita consumption expenditure for 2003 and 2006, respectively.
Column C reports the percentage change in per capita expenditure over
these three years. It is evident from rows 1 and 2 that the mean per capita
expenditure goes up by 1.1 percent [2,C] for female household heads but
decreases by 0.5 percent [1,C] for male household heads.
Columns D, E, and F report the median per capita expenditures for 2003
and 2006 and the growth rates between these years. Although the overall
change in median is –1.4 percent [20,F] (much larger than the change in
overall mean of –0.1 percent [20,C]), the changes in the groups with vari-
ous household characteristics are mixed. For female household heads, the
median increases by 1.5 percent [2,F], but it falls by 2.2 percent [1,F] for
male household heads. We find a mixed picture for the other household
characteristics.
Columns G, H, and I report inequality by household head’s characteris-
tics using the Gini coefficient, which lies between 0 and 100. Although the
overall Gini coefficient increases by 0.9 [20,I] in 2006, changes for different
household characteristics vary over a broad range.
183
Table 3.16: Mean and Median Per Capita Consumption Expenditure, Growth, and Gini
Coefficient, by Household Characteristics
Mean per capita Median per capita

consumption expenditure consumption expenditure Gini coefficient
2003 2006 Change 2003 2006 Change
(GEL) (GEL) (%) (GEL) (GEL) (%) 2003 2006 Change
Characteristic of
household head A B C D E F G H I
Gender
1 Male 127.2 126.6 −0.5 106.7 104.3 −2.2 33.7 34.8 1.1
2 Female 122.9 124.3 1.1 98.9 100.4 1.5 36.5 37.0 0.5
Age
3 15–19 110.8 217.8 96.6 90.0 150.7 67.3 16.2 31.7 15.5
4 20–24 188.0 223.5 18.9 131.7 188.3 43.0 40.6 35.0 −5.6
5 25–29 121.1 153.9 27.1 114.8 121.9 6.2 32.1 33.8 1.7
6 30–34 130.1 121.7 −6.5 111.4 98.1 −12.0 33.2 38.1 4.8
7 35–39 121.3 124.2 2.4 103.9 105.1 1.2 32.7 34.3 1.6
8 40–44 127.9 128.5 0.5 109.7 105.1 −4.2 33.8 35.3 1.5
9 45–49 127.6 132.7 4.0 102.9 104.4 1.4 35.7 36.2 0.5
10 50–54 121.5 120.7 −0.6 100.9 105.0 4.1 34.4 32.6 −1.8
11 55–59 134.7 132.8 −1.4 117.0 104.2 −10.9 33.5 38.0 4.5
12 60–64 130.5 123.0 −5.7 109.4 102.5 −6.4 32.3 34.3 2.0
13 65+ 122.8 121.8 −0.8 100.9 99.9 −1.0 35.1 34.8 −0.3
Education
14 Elementary or less 101.3 101.6 0.4 80.9 84.6 4.5 36.5 37.5 1.0
15 Incomplete secondary 109.5 106.7 −2.6 90.8 90.3 −0.5 34.5 33.4 −1.0
16 Secondary 116.2 118.6 2.1 97.3 99.6 2.3 33.7 34.1 0.4
17 Vocational-technical 127.7 116.3 −8.9 107.1 97.5 −9.0 34.6 34.6 0.0
18 Special secondary 134.4 127.5 −5.2 113.1 106.1 −6.2 33.9 33.0 −1.0
19 Higher education 153.7 155.1 0.9 129.7 123.7 −4.7 31.9 36.0 4.1
20 Total 126.1 126.0 −0.1 104.7 103.3 −1.4 34.4 35.4 0.9
Headcount Ratio by Household Head’s Characteristics
Table 3.17 analyzes poverty by population subgroup according to various

household head characteristics. The poverty line is set at GEL 75.4 per month.
Table rows report categories for three household head characteristics:
gender, age, and education level. Columns A, B, and C analyze the pov-
erty headcount ratios for 2003, 2006, and the change between those years.
Columns D, E, and F outline the distribution of poor people across the
subgroups, with the number in the cell being the proportion of all poor
people in the country contained in each subgroup. We can also call this
the subgroup’s percentage contribution to overall poverty, or the headcount
ratio times the population share included in that group. Columns G, H, and
184
Table 3.17: Headcount Ratio by Household Head’s Characteristics

percent
Poverty headcount ratio Distribution of the poor Distribution of population

Characteristic of
household head A B C D E F G H I
Gender
1 Male 28.4 30.0 1.6 69.6 71.5 1.9 73.3 73.6 0.3
2 Female 34.1 33.5 −0.5 30.4 28.5 −1.9 26.7 26.4 −0.3
Age
3 15–19 0 0 0 0 0 0 0 0.1 0.1
4 20–24 18.5 8.2 −10.3 0.3 0.2 −0.2 0.5 0.6 0
5 25–29 33.4 18.4 −15.0 1.3 0.7 −0.7 1.2 1.1 −0.1
6 30–34 26.9 36.2 9.3 3.3 3.1 −0.2 3.7 2.7 −1.0
7 35–39 31.6 31 −0.6 5.7 4.8 −0.9 5.4 4.7 −0.6
8 40–44 28.5 29.9 1.4 9 8.2 −0.8 9.5 8.5 −1.0
9 45–49 30.1 28.2 −1.9 11.9 10.7 −1.3 11.9 11.7 −0.2
10 50–54 32.8 31.1 −1.7 12.7 12.2 −0.5 11.6 12.2 0.6
11 55–59 26.0 30.0 4.0 7.7 11.2 3.5 8.9 11.6 2.7
12 60–64 24.2 32.4 8.2 8.7 7.6 −1.1 10.8 7.3 −3.5
13 65+ 32.1 32.4 0.2 39.2 41.4 2.2 36.5 39.6 3.1
Education
14 Elementary or less 44.2 43.1 −1.0 12.4 10.2 −2.2 8.4 7.3 −1.1
15 Incomplete secondary 38.4 38.7 0.3 12.7 10.0 −2.6 9.9 8.0 −1.9
16 Secondary 33.3 32.5 −0.7 42.9 40.1 −2.9 38.6 38.1 −0.5
17 Vocational-technical 30.2 36.5 6.3 8.4 11.7 3.4 8.3 9.9 1.7
18 Special secondary 26 26.9 0.9 9.9 11.2 1.3 11.4 12.8 1.5
19 Higher education 17.5 21.9 4.4 13.7 16.9 3.1 23.4 23.8 0.3
20 Total 29.9 31.0 1.0 100 100 n.a. 100 100 n.a.
I depict the subgroup population distributions, or the population percent-

age contained in each subgroup. Row 20 shows that overall headcount ratio
increases from 29.9 percent in 2003 [20,A] to 31.0 percent in 2006 [20,B],
reflecting a 1.0 percentage point increase [20,C] in the headcount ratio.
We see that 28.4 percent of male household heads [1,A] are poor. In
other words, the headcount ratio for this population subgroup is 28.4 percent.
The headcount ratio of the same group in 2006 is 30.0 percent [1,B]. So the
headcount ratio for the population in the male-headed household increased
by 1.6 percentage points [1,C] from 2003 to 2006.
In row 4, we find that 18.5 percent [4,A] of the population from house-
holds headed by someone in the 20–24 age group is poor. The headcount
ratio for the same population subgroup in 2006 is 8.2 percent [4,B], a
185
change of –10.3 percentage points [4,C]. In fact, headcount ratios have

also decreased for households with the head in the 25–29 age group [5,C],
35–39 [7,C], 45–49 [9,C], and 50–54 [10,C]. When subgroups are divided
according to household head’s education, we find that the headcount ratio
for the population living in the households where the head’s education is
elementary or less is 44.2 percent [14,A]. In both years, the population in
this subgroup had the highest headcount ratio.
Of all people who were poor in Georgia in 2003, 69.6 percent [1,D]
were from male-headed households. The share of all poor living in male-
headed households increased to 71.5 percent in 2006 [1,E], an increase of
1.9 percentage points [1,F]. In contrast, the share of poor in female-headed
households fell by 1.9 percentage points from 30.4 percent [2,D] in 2003 to
28.5 percent [2,E] in 2006.
There was not a large change in the population share in either male- or
female-headed households. For male-headed households, the proportion
of population increased by 0.3 percentage point from 73.3 percent [1,G]
to 73.6 percent [1,H]. For the female-headed households, the propor-
tion of population decreased by 0.3 percentage point from 26.7 percent
[2,G] to 26.4 percent [2,H]. Similarly, headcount ratios increased from
38.4 percent [15,A] to 38.7 percent [15,B] for the subgroup having house-
hold heads with incomplete secondary education. But the headcount ratio
for the subgroup having household heads with secondary education fell
from 33.3 percent [16,A] to 32.5 percent [16,B]. The shares of poor in both
groups decreased over the course of these three years: from 12.7 percent
[15,D] to 10.0 percent [15,E] for heads with incomplete secondary and
from 42.9 percent [16,D] to 40.1 percent [16,E] for heads with secondary
education.
One might wonder why the share of poor in households with heads
having incomplete secondary education decreased despite the increase in
the headcount ratio. The answer can be found if we look at columns G
and H. Notice that the population share with heads having incomplete
secondary or less decreased from 9.9 percent in 2003 [15,G] to 8.0 percent
in 2006 [15,H]. At the same time, headcount ratios for other subgroups
increased. For example, headcount ratios for the subgroups with household
heads in vocational-technical education and higher education increased by
6.3 [17,C] and 4.4 [19,C] percentage points, respectively. Thus, despite an
increase in headcount ratio, the shares of the poor population decreased for
the subgroup with heads having incomplete secondary education.
186
Population Distribution across Quintiles by Household

Head’s Characteristics
Table 3.18 analyzes the distribution of population across five quintiles of

per capita consumption expenditure by household head’s characteristics.
Column 1 denotes the lowest or first quintile, column 2 denotes the second
quintile, and so forth.
population into five equal-sized groups in terms of per capita expenditure.
Each group consists of 20 percent of the population. The fifth quintile con-
tains the richest 20 percent of the population, the fourth quintile consists
of the second-richest 20 percent of the population, and so on, and the first
quintile consists of the poorest 20 percent of the population.
Table 3.18: Distribution of Population across Quintiles by Household Head’s

Characteristics, 2003
percentage of per capita expenditure
Quintile
Characteristic of household head A B C D E
1 Total 20.0 20.0 20.0 20.0 20.0
Gender
2 Male 18.6 20.2 20.1 20.7 20.3
3 Female 23.8 19.3 19.7 18.0 19.1
Age (years)
4 15–19 0.0 27.1 51.1 17.1 4.8
5 20–24 10.5 19.9 12.1 19.1 38.4
6 25–29 23.4 15.5 13.4 26.1 21.6
7 30–34 16.5 21.5 16.8 23.4 21.8
8 35–39 21.6 19.9 20.0 17.7 20.8
9 40–44 19.5 17.5 21.1 20.9 21.0
10 45–49 21.7 19.9 19.2 19.2 19.9
11 50–54 21.4 20.5 19.9 20.0 18.2
12 55–59 16.2 18.6 18.2 24.1 22.9
13 60–64 15.5 19.5 22.4 20.9 21.8
14 65+ 21.5 21.1 20.3 18.6 18.5
Education
15 Elementary or less 32.5 23.8 16.5 15.9 11.3
16 Incomplete secondary 25.5 23.7 20.1 16.9 13.9
17 Secondary 22.3 22.4 20.3 18.5 16.5
18 Vocational-technical 20.1 18.9 20.1 21.0 19.9
19 Special secondary 17.4 18.1 19.9 22.7 21.8
20 Higher education 10.7 14.4 20.6 23.6 30.7
187
The rows below row 1 report the distribution of population by vari-

ous household head characteristics for 2003 using the national quintiles.
Consider the value 18.6 [2,A] for male household heads. This value implies
that 18.6 percent of the total population in male-headed households lives
with per capita expenditure less than the first quintile. The population
living in male-headed households is 20.2 percent in the second quintile
[2,B]. Similarly, 20.3 percent [2,E] of the population from the male-headed
households falls in the fifth quintile. The population distribution is almost
the same across all five quintiles.
The largest proportion of population living in the lowest quintile belongs
to households headed by someone who has not acquired education beyond
elementary level [15,A]. At the other extreme, the largest proportion of
population living in the highest quintile belongs to the households headed
by someone in the 20–24 age group [5,E].
This table is helpful in understanding population mobility across different

levels of consumption expenditure across different regions that a single wel-
fare, inequality, or poverty measure cannot reflect.
Headcount Ratio by Employment Category
Table 3.19 analyzes Georgia’s headcount ratio by population subgroups

according to household members’ employment category. The poverty line
is set at GEL 75.4 per month. Table rows list employment sectors (agricul-
ture, industry, government, and so on) as well as unemployed and inactive
categories to account for those not working.
Columns A, B, and C analyze poverty headcount ratios for 2003, 2006,
and the change over time. Columns D, E, and F outline the distribution of
poor people across the subgroups, with the number in the cell being the per-
centage of all poor people in the country that are located in that subgroup.
Stated another way, this is the percentage contribution of the subgroup to
overall poverty, or the headcount ratio times the population share in that
group. Columns G, H, and I depict subgroup population distribution, or the
population percentage found in that subgroup. The last row shows that overall
headcount ratio increases from 29.9 percent in 2003 [15,A] to 31.0 percent in
2006 [15,B], reflecting a 1.0 percentage point increase in the headcount ratio.
188
Table 3.19: Headcount Ratio by Employment Category

percent

Employment A B C D E F G H I
Self-employed
1 Agriculture 29.4 28.2 −1.3 23.2 20.2 −3.0 23.6 22.2 −1.4
2 Industry 20.5 32.2 11.7 0.4 0.5 0.1 0.5 0.5 −0.1
3 Trade 23.8 22.1 −1.6 2.5 1.8 −0.7 3.2 2.5 −0.7
4 Transport 19.2 28.9 9.7 0.4 0.7 0.3 0.7 0.7 0.1
5 Other services 20.7 27.8 7.2 0.7 0.9 0.2 1.0 1.0 −0.0
Employed
6 Industry 21.3 24.7 3.4 1.5 1.6 0.0 2.1 2.0 −0.2
7 Trade 19.5 24.1 4.6 1.1 1.1 0.1 1.6 1.5 −0.2
8 Transport 21.1 28.2 7.1 0.7 0.8 0.1 0.9 0.9 −0.1
9 Government 18.9 17.8 −1.1 1.4 1.1 −0.3 2.2 1.8 −0.4
10 Education 19.1 17.4 −1.7 2.1 1.7 −0.3 3.3 3.1 −0.2
11 Health care 16.7 19.1 2.5 0.6 0.7 0.1 1.1 1.2 0.0
12 Other 23.1 24.9 1.8 2.9 3.1 0.2 3.7 3.8 0.1
13 Unemployed 37.3 40.3 3.1 8.9 10.8 1.9 7.2 8.3 1.1
14 Inactive 32.9 33.7 0.8 53.6 55.1 1.5 48.8 50.7 1.8
15 Total 29.9 31.0 1.0 100.0 100.0 n.a. 100.0 100.0 n.a.
We find that 29.4 percent [1,A] of people engaged in the agricultural

sector are poor in 2003. In other words, the headcount ratio for this popula-
tion subgroup (with a household head employed in the agricultural sector) is
29.4 percent. The headcount ratio for the same population subgroup (with a
household head in the agricultural sector) fell to 28.2 percent in 2006 [1,B].
Thus, a 1.3 percentage point decrease [1,C] occurred in the headcount ratio
between the two years. We see that the headcount ratio among members
in the other services sector increased by 7.2 percentage points [5,C], from
20.7 percent [5,A] to 27.8 percent [5,B]. This headcount ratio increase from
2003 to 2006 is found in other sectors, such as employed industry [6,C],
trade [7,C], and transport [8,C].
Of all people who are poor in Georgia in 2003, 23.2 percent [1,D] are
employed in agriculture. We find that the share of all poor employed in
agriculture fell to 20.2 percent in 2006 [1,E]. This represents a decrease of
3.0 percentage points [1,F].
Contrast those results with the figures for the unemployed population
subgroup. Clearly, the poverty headcount ratio among this group in 2003
189
[13,A] is larger than the poverty headcount ratio in 2003 among the sub-
group employed in the agricultural sector [1,A]. However, if we consider
the share of all poor people who are found in these two subgroups in 2003,
this number is nearly twice as large in the agricultural sector as that among
the unemployed group. This is because of the different population shares of
the two subgroups as given in the final columns. The population share in the
agriculture subgroup in 2003 is 23.6 percent [1,G], while the share in the
unemployed subgroup is only 7.2 percent [13,G].
In row 1, the agricultural subgroup’s poverty headcount ratio falls 1.3
percentage points [1,C], while the share of poor in this subgroup falls by
3.0 percentage points [1,F]. For the other services subgroup, the headcount
ratio increased 7.2 percentage points [5,C] between 2003 and 2006, while
the share of poor in this subgroup increased by only 0.2 percentage point,
from 0.7 percent [5,D] to 0.9 percent [5,E].
One might wonder why these two ways of evaluating changes are so
different. Look at columns G and H. Notice that the population share
employed in the agricultural sector is more than 20 percent of the total
population in both 2003 [1,G] and 2006 [1,H]. In comparison, the popu-
lation share engaged in other services is only 1.0 percent in 2003 [5,G]
and 2006 [5,H]. Consequently, a change of smaller magnitude in the
headcount ratio in the agricultural sector has a larger impact on its share
of the poor and vice versa.
Headcount Ratio by Education Level
Table 3.20 analyzes poverty by education levels. The poverty line is set at
GEL 75.4 per month. Columns A, B, and C analyze poverty headcount
ratios for 2003, 2006, and the difference over time. Columns D, E, and F
outline the distribution of poor people across the subgroups, with the num-
ber in the cell being the proportion of all poor people in the country located
in that subgroup. This is the subgroup’s percentage contribution to overall
poverty, or the headcount ratio times the population share in that group.
Columns G, H, and I depict subgroup population distribution, or the popula-
tion percentage in that subgroup. Row 7 shows that the overall headcount
190
Table 3.20: Headcount Ratio by Education Level

percent

Education level A B C D E F G H I
1 Elementary or less 40.4 35.9 −4.6 6.5 5.7 −0.7 4.6 4.1 −0.5
2 Incomplete secondary 36.1 38.2 2.1 14.3 13.9 −0.5 11.5 10.9 −0.6
3 Secondary 33.2 31.9 −1.3 46.8 44.1 −2.6 40.8 41.7 0.9
4 Vocational-technical 30.0 35.0 5.0 7.7 8.5 0.7 7.5 7.3 −0.2
5 Special secondary 25.2 27.7 2.5 10.1 11.2 1.2 11.6 12.2 0.6
6 Higher education 17.6 20.9 3.4 14.6 16.6 1.9 24.1 23.8 −0.3
7 Total 29.9 31.0 1.0 100.0 100.0 n.a. 100.0 100.0 n.a.
ratio increases from 29.9 percent in 2003 [7,A] to 31.0 percent in 2006 [7,B],
reflecting a 1.0 percentage point (rounded) increase in the headcount ratio.
We find that 40.4 percent [1,A] of the population who have elementary-
level education or less are poor. In other words, the headcount ratio for this
population subgroup is 40.4 percent. The headcount ratio for the same popu-
lation subgroup fell to 35.9 percent in 2006 [1,B]. Thus, the headcount ratio
fell by 4.6 percentage points [1,C] between these three years. At the other
extreme, the headcount ratio for the subgroup with higher education increased
by 3.4 percentage points, from 17.6 percent [6,A] to 20.9 percent [6,B].
Of all people who are poor in Georgia in 2003, 6.5 percent [1,D]
have elementary education or less. The share of all poor with elementary
education or less decreased to 5.7 percent in 2006 [1,E], a decrease of 0.7
percentage point [1,F].
Clearly, the poverty headcount ratio among the population with incom-
plete secondary education in 2003 [2,A] is larger than the poverty head-
count ratio in 2003 among the higher education subgroup [6,A]. However,
if we consider the share of all poor people who are found in these two
subgroups in 2003, the number is larger for the population with higher edu-
cation because of the two subgroups’ different population shares, as given
in the table’s final columns. The population share with higher education
in 2003 is 24.1 percent [6,G], whereas the population share with incom-
plete secondary education is only 11.5 percent [2,G]. The headcount ratios
increased for the population with incomplete secondary education from
191
36.1 percent [2,A] to 38.2 percent [2,B], for vocational-technical education

from 30 percent [4,A] to 35 percent [4,B], for special secondary education
from 25.2 percent [5,A] to 27.7 percent [5,B], and for higher education from
17.6 percent [6,A] to 20.9 percent [6,B].
Headcount Ratio by Demographic Composition
Table 3.21 analyzes poverty by population subgroup, where each subgroup

is based first on the number of children 0–6 years of age in the household,
then on the household’s size. The poverty line is set at GEL 75.4 per month.
Columns A, B, and C analyze poverty headcount ratios for 2003, 2006, and
the difference over time. Columns D, E, and F outline the distribution of
poor people across the subgroups, with the number in the cell being the
proportion of poor people in the country contained in that subgroup. This is
the subgroup’s percentage contribution to overall poverty, or the headcount
ratio times the population share that falls in that group. Columns G, H, and
I depict subgroup population distribution, or the percentage of the popula-
tion in that subgroup. Row 12 shows that overall headcount ratio increased
Table 3.21: Headcount Ratio by Demographic Composition

percent

Demographic characteristic A B C D E F G H I
Number of children 0–6 years
1 None 28.8 28.5 −0.4 69.6 66.1 −3.5 72.2 72.0 −0.2
2 1 31.2 36.2 5.0 20.5 22.2 1.7 19.7 19.0 −0.7
3 2 35.5 39.9 4.5 8.3 10.3 2.0 7.0 8.0 1.0
4 3 or more 43.7 40.6 −3.1 1.5 1.3 −0.2 1.0 1.0 −0.0
Household size
5 1 25.8 24.1 −1.7 2.6 2.6 −0.0 3.1 3.4 0.3
6 2 23.1 21.0 −2.1 6.7 5.9 −0.8 8.7 8.7 −0.0
7 3 25.0 23.2 −1.8 11.1 9.9 −1.2 13.3 13.2 −0.1
8 4 24.4 26.2 1.7 19.5 18.5 −1.1 23.9 21.8 −2.1
9 5 31.9 33.8 1.8 23.0 23.0 0.0 21.6 21.1 −0.5
10 6 36.2 35.4 −0.9 19.6 19.1 −0.4 16.2 16.7 0.6
11 7 or more 39.3 43.2 4.0 17.3 20.9 3.6 13.2 15.0 1.8
12 Total 29.9 31.0 1.0 100.0 100.0 n.a. 100.0 100.0 n.a.
192
from 29.9 percent in 2003 [12,A] to 31.0 percent in 2006 [12,B], reflecting
a 1.0 percentage point (rounded) increase in the headcount ratio.
First, consider the results based on the number of children in households.
We find that 28.8 percent [1,A] of the population with no child in the
household is poor in 2003. In other words, the headcount ratio for this popu-
lation subgroup is 28.8 percent. The headcount ratio for the same population
subgroup decreased to 28.5 percent in 2006 [1,B]. Thus, the headcount ratio
decreased by 0.4 percentage point [1,C] over the course of these three years.
Headcount ratios also decreased for the population with three or more
children in the household by 3.1 percentage points from 43.7 percent
[4,A] in 2003 to 40.6 percent [4,B] in 2006. Similarly, consider the set of
results corresponding to the household size. The headcount ratio among the
population with only one member in the household in 2003 is 25.8 percent
[5,A], which falls by 1.7 percentage points to 24.1 percent in 2006 [5,B]. At
the other extreme, the headcount ratio among the people living in house-
holds with seven or more members increased by 4.0 percentage points from
39.3 percent [11,A] to 43.2 percent [11,B].
The next cell in row 1 is 69.6 [1,D], meaning that of all people who are
poor in Georgia in 2003, 69.6 percent of the population live in households
with no child. In the next column, we find that the share of poor with no
child decreased to 66.1 percent in 2006 [1,E], a decrease of 3.5 percentage
points [1,F].
Compare those results with the subgroup having three or more children.
It is evident that the headcount ratio among the subgroup with no child in
both years (28.8 percent in 2003 [1,A] and 28.5 percent in 2006 [1,B]) is
lower than the headcount ratio for the subgroup with three or more children
(43.7 percent in 2003 [4,A] and 40.6 percent for 2006 [4,B]). Note that the
share of the former subgroup to total poverty is 69.6 percent in 2003 [1,D],
which fell by 3.5 percentage points to 66.1 percent in 2006 [1,E]. The share
of the latter to total poverty is 1.5 percent in 2003 [4,D], which fell by
0.2 percentage point to 1.3 percent in 2006 [4,E]. However, in both years,
the share of poor in the former subgroup is more than 40 times higher than
that in the latter subgroup.
Note that the poverty rate among the subgroup with three or more children
is higher than the subgroup with no child. However, the population share
193
in the subgroup with no child is so large (72.2 percent in 2003 [1,G] and
72 percent in 2006 [1,H]), compared to the subgroup with three or more
children (only 1.0 percent in both years [4,G] and [4,H]), that the share of
the subgroup with no child in total poverty is high. The analysis in table
3.21 enables a policy maker to understand the origin of poverty at a more
disaggregated level. A policy maker should also focus on households with no
child, even though the headcount ratio is lowest in this subgroup. Similar
intuition should hold for the next set of results where the subgroups are
based on household size.
Headcount Ratio by Landownership
Table 3.22 analyzes poverty by population household landownership sub-

groups for 2003, 2006, and the change across those years. The poverty line
is set at GEL 75.4 per month. Columns A, B, and C analyze the poverty
headcount ratios. Columns A and B report the headcount ratio for 2003
and 2006, respectively, while column C reports the difference over time.
Columns D, E, and F outline the distribution of poor people across the
subgroups, with the number in the cell being the proportion of poor people
in the country located in that subgroup. This is the subgroup’s percent-
age contribution to overall poverty, or the headcount ratio times the
population share that lies in that group. Columns G, H, and I depict the
subgroups’ population distribution, or the population percentage found in
Table 3.22: Headcount Ratio by Landownership

percent

Size of landholding
(hectares) A B C D E F G H I
1 0 29.4 32.7 3.3 39.0 46.4 7.3 39.7 43.9 4.2
2 Less than 0.2 39.4 36.2 −3.1 12.7 11.9 −0.7 9.6 10.2 0.6
3 0.2–0.5 33.9 36.9 2.9 17.2 18.4 1.1 15.2 15.4 0.2
4 0.5–1.0 25.1 24.3 −0.8 19.5 15.4 −4.1 23.2 19.6 −3.6
5 More than 1.0 28.2 22.4 −5.8 11.5 7.9 −3.6 12.2 10.9 −1.3
6 Total 29.9 31.0 1.0 100.0 100.0 n.a. 100.0 100.0 n.a.
194
each subgroup. Row 6 shows that the overall headcount ratio increases from
29.9 percent in 2003 [6,A] to 31.0 percent in 2006 [6,B], reflecting a
1.0 percentage point (rounded) increase in the headcount ratio.
We find that 29.4 percent [1,A] of people who belong to households
with no landownership are poor in 2003. In other words, the headcount
ratio for this population subgroup is 29.4 percent. The headcount ratio for
the same population subgroup increases to 32.7 percent in 2006 [1,B]. Thus,
the headcount ratio increased by 3.3 percentage points [1,C] over these
three years. We see that the headcount ratio for the population in house-
holds with landownership of 0.5–1.0 hectare decreased by 0.8 percentage
point, from 25.1 percent [4,A] to 24.3 percent [4,B].
Of all poor people in Georgia in 2003, 39 percent [1,D] lived in house-
holds with no landownership. The share of poor with no landownership
increased to 46.4 percent in 2006 [1,E]. The headcount ratio among the
subgroup with landownership of 0.5–1.0 hectare (25.1 percent in 2003
[4,A] and 24.3 percent in 2006 [4,B]) is lower than the headcount for the
subgroup with a landownership of less than 0.2 hectare (39.4 percent in
2003 [2,A] and 36.2 percent for 2006 [2,B]). Note that the share of the
former subgroup to total poverty is 19.5 percent in 2003 [4,D], which fell
by 4.1 percentage points to 15.4 percent in 2006 [4,E]. The share of the
latter to total poverty is 12.7 percent in 2003 [2,D], which fell by only 0.7
percentage point to 11.9 percent in 2006 [2,E]. Note that despite a larger
fall in the poverty rate of 3.1 percentage points [2,C] for the subgroup with
landownership of less than 0.2 hectare, the share of poor in that subgroup
fell by only 0.7 percentage point [2,F]. One might wonder about the reason
behind this phenomenon.
The answer can be found if we look at columns G and H. Notice that the
population share with landownership of less than 0.2 hectare is 9.6 percent
in 2003 [2,G], and it increased by 0.6 percentage point to 10.2 percent in
2006 [2,H]. In contrast, the population share with landownership of 0.5–1.0
hectare fell by 3.6 percentage points, from 23.2 percent [4,G] in 2003 to
19.6 percent [4,H] in 2006. Moreover, the population share in the latter
subgroup is almost twice as high as that in the former subgroup in both
years. Thus, despite a larger fall in headcount ratio for the subgroup with
landownership of less than 0.2 hectare, its share in total number of poor did
not decrease significantly compared to the subgroup with landownership of
0.5–1 hectare.
195
Headcount Ratio by Age Groups
Table 3.23 analyzes poverty by population subgroup according to individuals’

ages. The poverty line is set at GEL 75.4 per month. Columns A, B, and C
analyze poverty headcount ratios for 2003, 2006, and the difference over
time, respectively. Columns D, E, and F outline the distribution of poor
people across the subgroups, with the number in the cell being the propor-
tion of poor people located in that subgroup. This is the subgroup’s percent-
age contribution to overall poverty, or the headcount ratio times the overall
population share that lies in that group. Columns G, H, and I depict the sub-
groups’ population distribution, or the percentage of the population that can
be found in that subgroup. Row 14 shows that the overall headcount ratio
increased from 29.9 percent in 2003 [14,A] to 31.0 percent in 2006 [14,B],
reflecting a 1.0 percentage point (rounded) increase in headcount ratio.
We see that 32.8 percent of the population in age group 0–5 years [1,A]
is poor. In other words, the headcount ratio for this population subgroup
is 32.8 percent. The headcount ratio for the same population subgroup
increased to 34.9 percent in 2006 [1,B]. Thus, the headcount ratio increased
by 2.1 percentage points [1,C] during these three years. In fact, the head-
count ratio increased among all age groups except 50–54 and 65+ years.
Table 3.23: Headcount Ratio by Age Groups

percent

Age group (years) A B C D E F G H I
1 0–5 32.8 34.9 2.1 5.9 6.2 0.2 5.4 5.5 0.1
2 6–14 33.3 34.5 1.2 14.4 12.6 −1.8 12.9 11.3 −1.7
3 15–19 33.3 33.7 0.4 9.6 9.5 −0.1 8.6 8.7 0.1
4 20–24 30.7 31.6 0.9 8.0 8.7 0.7 7.8 8.5 0.8
5 25–29 30.9 31.5 0.7 7.3 7.4 0.1 7.1 7.3 0.2
6 30–34 30.2 32.6 2.4 6.9 6.8 −0.2 6.9 6.4 −0.4
7 35–39 30.2 32.1 1.9 6.8 6.8 −0.0 6.7 6.5 −0.2
8 40–44 27.9 31.4 3.5 7.2 7.0 −0.2 7.7 7.0 −0.8
9 45–49 28.6 29.1 0.5 6.9 6.8 −0.1 7.2 7.2 −0.0
10 50–54 28.3 27.1 −1.1 5.6 5.4 −0.2 6.0 6.2 0.2
11 55–59 23.0 25.8 2.8 3.2 4.5 1.2 4.2 5.4 1.2
12 60–64 23.0 26.7 3.8 3.5 3.0 −0.5 4.5 3.4 −1.1
13 65+ 29.3 28.8 −0.6 14.7 15.5 0.8 15.0 16.6 1.6
14 Total 29.9 31.0 1.0 100.0 100.0 n.a. 100.0 100.0 n.a.
196
Between 2003 and 2006, the headcount ratios decreased for age group
50–54 years by 1.1 percentage points, from 28.3 percent [10,A] to 27.1 per-
cent [10,B], and for the age group 65+ years by 0.6 percentage point from
29.3 percent [13,A] to 28.8 percent [13,B]. In contrast, headcount ratios
increased for all other groups by 0.4 to 3.8 percentage points. For example,
the headcount ratio for age group 30–34 years increased by 2.4 percentage
points from 30.2 percent [6,A] in 2003 to 32.6 percent [6,B] in 2006.
Of all poor people in Georgia in 2003, 5.9 percent are in the age group
of 0–5 years [1,D]. The share of all poor in age group 0–5 years increased to
6.2 percent in 2006 [1,E], an increase of 0.2 percentage point. Now consider
age groups 6–14 and 65+ years. The headcount ratio among the population
in age group 6–14 years increased by 1.2 percentage points from 33.3 percent
in 2003 [2,A] to 34.5 percent in 2006 [2,B], but the headcount fell by 0.6
percentage point for age group 65+ years [13,C]. However, if we consider
the change in share of all poor people found in these two subgroups in 2003
(column F), this number went up for age group 65+ (0.8 [13,F]) and fell for
age group 6–14 years (–1.8 [2,F]).
One might ask why the share of the poor has fallen in spite of an increase in
headcount ratios. The answer can be found in columns G and H. Note that
the share of people in the age group 6–14 years decreased by 1.7 percent-
age points from 12.9 percent in 2003 [2,G] to 11.3 percent in 2006 [2,H].
In contrast, the population share in age group 65+ years increased by 1.6
percentage points from 15.0 percent in 2003 [13,G] to 16.6 percent in 2006
[13,H]. Thus, despite a decrease in headcount ratio for age group 65+ years,
its share of poor increased. A policy maker, therefore, should notice that a
decrease in headcount among the 65+ years age group did not necessarily
decrease the number of total poor in that age group.
Headcount Ratio and Age-Gender Pyramid
Until now, we have analyzed headcount ratios across individual population

subgroups. We have not analyzed the headcount ratio across two different
population subgroups simultaneously. Figure 3.2 presents one such example
using a graph known as an age-gender pyramid. The age-gender pyramid
analyzes the headcount ratios across gender and across different age groups
197
Figure 3.2: Age-Gender Poverty Pyramid
2003
95 33 30 95
90 31 38 90
85 36 31 85
80 31 27 80
75 31 29 75
30 25
Female poverty rate

70 70
Male poverty rate

Age in years 65 23 22 65
Age in years
60 23 21 60
55 25 29 55
50 28 28 50
45 26 28 45
40 31 29 40
35 29 30 35
30 29 30 30
25 31 29 25
20 34 30 20
15 35 34 15
10 32 31 10
5 31 34 5
6 5 4 2 1 0 1 2 4 5 6
Share in total population, %
Total population Poor population
simultaneously. However, it can be used to analyze other subgroups with

proper justification. As before, the variable for our analysis is per capita
consumption expenditure in lari, and the poverty line is set at GEL 75.4 per
month. The outside vertical axes denote the age of the members in years,
and the horizontal axis presents the share of the population.
The figure is divided vertically by gender: the right-hand side repre-
sents males and the left-hand side represents females. The distance from
the middle to each side in dark gray denotes the total population share in
that age group. The distance in light gray is the proportion of poor people
in that age group of the total number of poor, again for each gender. Data
are aggregated in five-year increments, and each increment is displayed as
a bar centered on the highest age in the increment. The data for ages 25 to
30 years, for example, are represented by the bar at 30 years. For those zero
to five years of age, the shares of both males and females are 2.2 percent,
and nearly 0.7 percent of both males and females in that age group reside in
198
poor households. The headcount ratio among females in that age group is
32 percent and among males it is 31 percent. The headcount ratio is highest
among male members in the 85–90 years age group: 38 percent of the males
in that age group reside in poor households. The largest headcount ratio
among females is seen in the 80–85 years age group.
Sensitivity Analyses
In this section, we perform sensitivity analysis of poverty line choice, pov-

erty measures, and inequality measures, mostly at the national level and
across urban and rural areas. In certain cases, the results are reported at the
subnational levels or across geographic regions. However, all sensitivity
analysis can be replicated at any disaggregated level.
Elasticity of FGT Poverty Indices to Per Capita Consumption
Table 3.24 provides a tool for checking the sensitivity of the three poverty
measures to consumption expenditure. The table shows the result of increas-
ing everyone’s consumption expenditure by 1.0 percent and compares those
values across two years, 2003 and 2006. There are two poverty lines: GEL
75.4 and GEL 45.2 per month.
The percentage change in poverty caused by a 1.0 percent change in the
mean or average per capita consumption expenditure is referred to as the
elasticity of poverty with respect to per capita consumption. The particular way
Table 3.24: Elasticity of FGT Poverty Indices to Per Capita Consumption Expenditure
Headcount ratio Poverty gap measure Squared gap measure

1 Urban −1.89 −1.72 0.17 −1.95 −2.03 −0.09 −2.09 −2.23 −0.14
2 Rural −1.66 −1.53 0.13 −1.71 −1.64 0.07 −1.82 −1.72 0.09
3 Total −1.77 −1.62 0.15 −1.81 −1.82 0.00 −1.93 −1.93 0.00
4 Urban −2.06 −2.35 −0.29 −2.36 −2.47 −0.12 −2.24 −2.50 −0.26
5 Rural −1.87 −1.64 0.23 −1.86 −1.78 0.07 −1.94 −1.86 0.08
6 Total −1.95 −1.94 0.01 −2.05 −2.04 0.01 −2.05 −2.06 −0.02
Note: FGT = Foster-Greer-Thorbecke.
199
in which we consider an increase in the average per capita consumption

expenditure is by increasing everyone’s consumption expenditure by the
same percentage. This type of change is distribution neutral, because the
relative inequality does not change.
The main columns denote three different sets of poverty measures:
headcount ratio, poverty gap measure, and squared gap measure. The first
two columns within each set report the elasticities for 2003 and 2006,
respectively, while the third column reports the difference between these
two years.
Let us start with the GEL 75.4 per month poverty line. The elasticity of
poverty with respect to the mean consumption expenditure for the urban
area in 2003 is –1.89 [1,A]. In other words, if the consumption expenditure
increases by 1.0 percent for everyone, then the mean per capita consump-
tion expenditure increases by 1.0 percent and the urban headcount ratio
falls by –1.89 percent, or stated differently, 1.89 percent of the population
who were living under the poverty line of GEL 75.4 will be out of poverty.
If the mean consumption expenditure is increased by 1.0 percent, then
the headcount ratio for the urban area falls by 1.72 percent in 2006 [1,B]. A
higher value implies higher sensitivity. The urban headcount ratio elasticity
is less sensitive to consumption expenditure in 2006 than in 2003 by 0.17
[1,C]. Similarly, the elasticity of poverty gap to the per capita consumption
expenditure for the urban area in 2003 is –1.95 [1,D], which increases by
–0.09 (rounded) to –2.03 in 2006 [1,E]. The elasticity of squared gap mea-
sure in 2003 is –2.09 [1,G], which increases by –0.14 to –2.23 in 2006 [1,H].
Negative elasticities mean a fall in poverty caused by an increase in con-
sumption expenditure. The higher magnitude implies higher elasticity even
though the sign is negative. Note that both the poverty gap measure and the
squared gap measure, unlike the headcount ratio, are more sensitive to con-
sumption expenditure in 2006 than in 2003. A similar pattern is seen for the
GEL 45.2 per month poverty line: the poverty gap measure and the squared
gap measure are more sensitive to the per capita consumption expenditure.
All elasticities in the rural area are lower in magnitude compared to
what we see in the urban area for both poverty lines. In other words, all
rural poverty measures are less sensitive to the per capita consumption
expenditure. The overall headcount ratio elasticity decreases slightly from
–1.95 in 2003 [6,A] to –1.94 in 2006 [6,B] for the GEL 45.2 poverty line,
but it decreases by 0.15 from –1.77 in 2003 [3,A] to –1.62 in 2006 [3,B] for
the GEL 75.4 poverty line. The elasticities of the overall poverty gap and
200
the squared gap measure did not change much between these two years for
either poverty line.
Note that poverty lines are set normatively, which is difficult to justify
exclusively. A slight change in per capita consumption expenditure may or
may not change the poverty rates by significant margins. If the distribution is
highly polarized or, in other words, if the society has two groups of people—
one group consisting of rich people and the other group consisting of extreme
poor—then a slight change in everyone’s income by the same proportion
may not affect the headcount ratio.
In contrast, if marginal poor are concentrated around the poverty line,
then a slight change in everyone’s income by the same proportion would
have a huge impact on the poverty measures. For example, in the table
the poverty measures are more sensitive to the lower GEL 45.2 per month
poverty line than the higher GEL 75.4 per month poverty line. This is
because the concentration of poor around the lower poverty line is much
larger than that around the higher poverty line. Hence, this type of analysis
may tell us about the impact of any policy on the poverty rate used by the
policy maker.
Sensitivity of Poverty Measures to the Choice of Poverty Line
Table 3.25 presents a tool for checking the sensitivity of the headcount ratio
with respect to the chosen poverty line. This exercise is similar to the exer-
cise for checking the elasticity of poverty measures to per capita consump-
tion expenditure, but it is more rigorous. It is always possible to find a certain
percentage of decrease in the poverty line that matches the increase in the
consumption expenditure for everyone by 1.0 percent. In this exercise,
we check the sensitivity of the poverty measure by changing the poverty
line in more than one direction. Thus, in the table, we ask how the actual
headcount ratio changes as the poverty line changes from its initial value,
whether it is GEL 75.4 per month or GEL 45.2 per month.
Rows denote the change in poverty line, both upward and downward.
Columns report the change in three poverty measures: the headcount ratio,
the poverty gap measure, and the squared gap measure, and their change
from actual. The variable is per capita consumption expenditure, measured
201
Table 3.25: Sensitivity of Poverty Measures to the Choice of Poverty Line, 2003
Headcount Change from Poverty gap Change from Squared gap Change from
ratio actual (%) measure actual (%) measure actual (%)
A B C D E F
1 Actual 29.9 0.0 9.7 0.0 4.6 0.0
2 +5 percent 32.6 9.0 10.7 10.7 5.1 11.4
3 +10 percent 35.3 18.0 11.7 21.7 5.6 23.3
4 +20 percent 40.5 35.2 13.9 44.3 6.8 48.5
5 −5 percent 26.9 −10.0 8.7 −10.2 4.1 −10.8
6 −10 percent 24.2 −19.1 7.7 −19.9 3.6 −21.1
7 −20 percent 19.4 −35.3 6.0 −38.1 2.7 −40.0
8 Actual 10.2 0.0 3.0 0.0 1.4 0.0
9 +5 percent 11.4 11.8 3.4 12.2 1.5 12.4
10 +10 percent 12.7 24.1 3.8 25.2 1.7 25.6
11 +20 percent 15.8 54.5 4.7 53.8 2.1 54.6
12 −5 percent 9.2 −9.9 2.7 −11.6 1.2 −11.6
13 −10 percent 8.0 −21.4 2.4 −22.4 1.1 −22.4
14 −20 percent 6.0 −40.9 1.8 −41.6 0.8 −41.4
in lari. In this table, we report the results only for 2003, but this analysis can
be conducted for any year.
Column A reports the headcount ratios for different poverty lines, and
column B reports the change in the headcount ratios from the actual pov-
erty line, which can be either GEL 75.4 per month or GEL 45.2 per month.
Rows 2 and 9, corresponding to +5 percent, denote the increase in poverty
line by 5 percent. Thus, when the poverty line is GEL 75.4, then a 5 percent
increase means the poverty line becomes GEL 79.2 and the headcount ratio
increases by 3.7 percentage points from 29.9 percent [1,A] to 32.6 percent
[2,A], or the headcount ratio increases by 9.0 percent [2,B] from its actual
level of 29.9 percent.
Similarly, if the poverty line is decreased by 10 percent (–10 percent)
from GEL 75.4, then the poverty headcount rate falls by 5.7 percentage
points from 29.9 percent [1,A] to 24.2 percent [6,A], or the headcount ratio
decreases by 19.1 percent from the actual level of 29.9 percent. The head-
count ratio is more sensitive to the change in poverty line when the actual
poverty line is GEL 45.2 than when the poverty line is GEL 75.4. In fact,
the poverty gap measure and the squared gap measure are also more sensitive
to change in poverty line when the actual poverty line is GEL 45.2 rather
than GEL 75.4.
202
This table helps us understand the robustness of a particular poverty esti-

mate. Selection of any poverty line is debatable, because it is set with nor-
mative judgment. If a change in the poverty line causes a drastic change in
a poverty measure, then a cautious policy conclusion should be drawn from
the analysis based on that particular poverty line. In contrast, if a poverty
measure does not vary much because of a change in the poverty line, then a
more robust conclusion can be drawn.
Other Poverty Measures
Table 3.26 analyzes the overall poverty for Georgia and decomposes it across
rural and urban areas using three other poverty measures not in the FGT class.
The table reports three different sets of poverty measures: the Watts index,
Sen-Shorrocks-Thon (SST) index, and Clark-Hemming-Ulph-Chakravarty
(CHUC) index (these measures are defined in chapter 2). This is a type of
sensitivity analysis, but of the poverty measurement methodology. There are
two poverty lines: GEL 75.4 per month and GEL 45.2 per month.
Columns A and B report the Watts index for both years. The Watts
index is the mean log deviation relative to the poverty line. It is evident
from row 1 that the urban Watts index increases from 12.0 in 2003 [1,A]
to 12.7 in 2006 [1,B] when the poverty line is GEL 75.4 but falls slightly
between the same years when the poverty line is GEL 45.2 [4,A] and [4,B].
Table 3.26: Other Poverty Measures
Watts index Sen-Shorrocks-Thon index CHUC index

A B C D E F G H I
1 Urban 12.0 12.7 0.7 15.7 16.8 1.1 16.6 16.5 0.0
2 Rural 15.6 16.2 0.5 19.2 19.6 0.4 22.2 22.8 0.6
3 Total 13.9 14.5 0.6 17.5 18.3 0.7 19.6 19.8 0.3
4 Urban 3.3 3.2 −0.1 4.7 4.6 0.0 5.1 4.5 −0.6
5 Rural 5.2 5.7 0.5 7.0 7.6 0.6 8.5 9.0 0.4
6 Total 4.3 4.5 0.2 5.9 6.2 0.3 6.9 6.8 −0.1
Note: CHUC = Clark-Hemming-Ulph-Chakravarty.
203
Columns D and E report the SST index, which is also based on the
headcount ratio, the income gap ratio, and the Gini coefficient across the
censored distribution of consumption expenditure. The last is obtained
by replacing consumption expenditure of all nonpoor people by the pov-
erty line. We see that when the poverty line is GEL 75.4, the SST index
for the urban region in 2003 is 15.7 [1,D], and it increases by 1.1 to 16.8
in 2006 [1,E]. Likewise, the rural region’s SST index increased by 0.4,
from 19.2 in 2003 [2,D] to 19.6 in 2006 [2,E], for the same poverty line.
The total increase in SST index is 0.7, from 17.5 in 2003 [3,D] to 18.3
in 2006 [3,E].
The final three columns report the CHUC index and its changes across
time. Unlike the SST index, the CHUC index does not reflect an increase
in poverty across all regions. In fact, urban poverty falls marginally between
2003 [1,G] and 2006 [1,H] when the poverty line is GEL 75.4. Furthermore,
when the poverty line is set at GEL 45.2, the CHUC index shows a fall in
Georgia’s overall poverty [6,I].
If these three measures, capturing different aspects of poverty and inequal-

ity among the poor, agree with the results from the measures in the FGT
class, then the poverty analysis is robust. In contrast, if these measures do
not agree with each other, the policy conclusion should be drawn with more
care. Comparing table 3.2 with table 3.26, we see that the three measures
reported in table 3.2 do not always agree with the results based on the
poverty gap measure and squared gap measure. Thus, any conclusion about
whether poverty has increased or decreased should be made cautiously.
Other Inequality Measures
Table 3.27 reports the Atkinson inequality measures and generalized entropy
measures for 2003, then decomposes them across different regions. This is
a type of sensitivity analysis for inequality measurement methodology. We
report the Gini coefficient only in the last two sections of this chapter.
However, the Gini coefficient may not be subgroup consistent (subgroup
consistency is defined in chapter 2). Rows denote results for urban and rural
population subgroups and for different geographic regions, such as Kakheti,
Tbilisi, and Shida Kartli.
204
Table 3.27: Atkinson Measures and Generalized Entropy Measures by

Geographic Regions, 2003
Atkinson measure Generalized entropy measure

A(1/2) A(0) A(−1) GE(0) GE(1) GE(2)
A B C D E F
1 Urban 9.1 17.7 34.3 19.4 18.8 22.8
2 Rural 10.1 19.8 38.9 22.0 21.0 25.6
Subnational regions
3 Kakheti 9.8 19.2 39.1 21.3 20.1 24.4
4 Tbilisi 8.3 15.9 29.8 17.3 17.3 20.8
5 Shida Kartli 11.0 21.6 44.8 24.4 22.8 28.2
6 Kvemo Kartli 8.9 17.3 33.9 19.0 18.6 24.0
7 Samtskhe-Javakheti 9.0 17.4 34.1 19.1 19.0 24.6
8 Ajara 9.4 18.5 36.5 20.4 19.2 22.6
9 Guria 9.4 18.2 35.7 20.1 19.9 27.2
10 Samegrelo 9.5 18.3 35.3 20.2 19.7 23.7
11 Imereti 8.8 17.3 33.8 19.0 18.0 20.8
12 Mtskheta-Mtianeti 9.3 18.6 36.7 20.6 18.5 20.2
13 Total 9.6 18.8 36.8 20.8 20.0 24.2
Note: GE = generalized entropy.
Columns A, B, and C report the Atkinson measures for a = 1/2, 0, and –1,
respectively, and columns D, E, and F report the generalized entropy measures
for a = 0, 1, and 2, respectively. (For a theoretical discussion on the Atkinson
inequality measure and generalized entropy measures, please refer to chapter
2.) Intuitively, an Atkinson inequality measure of order a is the gap between
the mean achievement and the generalized mean of achievements of order
a divided by the mean achievement. Generalized mean is sensitive to inequal-
ity across the distribution, where a lower value of a reflects higher sensitivity
to inequality across the distribution. In other words, a lower value of a reflects
higher aversion toward inequality and, thus, it is also known as the inequality
aversion parameter. When everyone has identical achievement, then it does
not matter how sensitive one is toward inequality, so the generalized mean
is equal to the arithmetic mean for all a. For the analysis in table 3.27, the
inequality measures put more emphasis on the lower end of the distribution
and thus assume a < 1. The Atkinson measure lies between 0 and 1. Similarly,
if a household has equal per capita expenditure in a region, then the general-
ized entropy measure is also 0 for all a.
The Atkinson measure for a = 1/2, or A(1/2), for the urban area is
9.1 [1,A]. Intuitively, the number implies that the generalized mean of
205
order 0.5 for urban Georgia is 9.1 percent lower than Georgia’s mean per
capita expenditure in 2003. The next two cells to the right report A(0) and
A(–1) for urban Georgia, where A(0) = 17.7 [1,B] and A(–1) = 34.3 [1,C].
Therefore, A(0) is 17.7 percent lower than the mean per capita expenditure
and A(–1) is 34.3 percent lower than the mean per capita expenditure.
Columns D, E, and F report three generalized entropy measures for a = 0, 1,
and 2, denoted by GE(0), GE(1), and GE(2), respectively.
Row 2 reports the three Atkinson measures and three generalized
entropy measures for rural Georgia. Each of these six measures shows that
rural Georgia is more unequal than urban Georgia. For example, the A(1/2)
for the rural area is 10.1 [2,A], compared with 9.1 in the urban area [1,A],
and A(0) for the rural area is 19.8 [2,B], compared with 17.7 for the urban
area [1,B]. However, the difference is much larger when the two regions are
compared with respect to A(–1): 38.9 for the rural area [2,C] and 34.3 for
the urban area [1,C].
Next, we consider the results across regions. The level of inequality
of Ajara according to A(1/2) is 9.4 [8,A], which is higher than that of
Samtskhe-Javakheti at 9.0 [7,A]. This means that Ajara has larger income
inequality than Samtskhe-Javakheti. Even according to A(0), A(–1),
GE(0), and GE(–1), Ajara has higher income inequality than Samtskhe-
Javakheti. However, in terms of GE(2), which gives more weight to larger
incomes across the population, Samtskhe-Javakheti [7,F] has higher income
inequality than Ajara [8,F].
A region’s income standards reflect that region’s welfare level. However,

higher welfare does not necessarily mean more equal distribution. A high
level of inequality may be detrimental to a region’s welfare. We already
reported the Gini coefficient for that purpose. However, given that the Gini
coefficient has certain limitations, we report three Atkinson inequality mea-
sures and three generalized entropy measures to check the inequality ranking
for regions. These six inequality measures are commonly used separately
from the Gini coefficient.
Also unlike the Gini coefficient, Atkinson and generalized entropy
class inequality measures are normative measures, in which we may choose
varying degrees of inequality aversion. If these six measures agree with the
Gini coefficient, then a conclusion based on the Gini coefficient can be
206
considered robust. However, if these six measures provide different rankings

than the Gini coefficient, then a more cautious policy conclusion should be
drawn based only on Gini.
Dominance Analyses
In the previous section, we conducted some dominance analysis with respect

to the choice of poverty lines and measurement methodologies. In this sec-
tion, we perform additional dominance analyses. Note that when we analyze
sensitivity with respect to the poverty line, we do not compare the results
for all poverty lines. Similarly, when we check the sensitivity of inequal-
ity using different Atkinson and generalized entropy measures, we do not
conduct the analysis for all parameter values. The dominance tests in this
part of the chapter go beyond the sensitivity analyses. For example, accord-
ing to the dominance analyses in this section, we can say that poverty has
unambiguously risen for all poverty lines, or inequality has risen, no matter
which inequality measure is used to assess it.
Poverty Incidence Curve
A poverty incidence curve is the distribution function of the welfare indi-

cator across the population. The poverty incidence curve is useful while
performing a dominance analysis of the headcount ratio with respect to the
poverty line. In this dominance exercise, the welfare indicator is per capita
consumption expenditure, assessed by lari. The horizontal axis of figure 3.3
denotes per capita consumption expenditure. The height of the poverty
incidence curve at any per capita consumption expenditure denotes the
proportion of people having less than that per capita expenditure.
Therefore, the link between the poverty incidence curve and the head-
count ratio is clear. The height of the poverty incidence curve is the head-
count ratio when the poverty line is set at a particular per capita consumption
expenditure. For a poverty line, a larger height denotes a larger headcount
ratio or a larger share of the population having per capita expenditure below
the poverty line. If the poverty incidence curve of a distribution lies to the
right of the poverty incidence curve of another distribution, then the former
distribution is understood to have an unambiguously lower headcount ratio
or the former distribution has lower headcount ratios for all poverty lines.
207
Figure 3.3: Poverty Incidence Curves in Urban Georgia, 2003 and 2006
Urban
1.0
0.8
0.6
0.4
0.2
0
0 160 320 480 640 800
Welfare indicator
2003 2006
Note: The red vertical line is the poverty line.
Figure 3.3 graphs the poverty incidence curves for urban Georgia in
2003 and 2006. The vertical axis reports the headcount ratio. The solid
line denotes the poverty incidence curve for 2003, while the dashed line
denotes the poverty incidence curve for 2006. We saw earlier that the urban
headcount ratio is higher in 2006 for both poverty lines: GEL 75.4 and GEL
45.2. What about other poverty lines? Can we say that poverty has unam-
biguously fallen for any poverty line? Figure 3.3 suggests that we may not
be able to. If we set the hypothetical poverty line somewhere between GEL
320 and GEL 480, then the headcount ratio would have been lower in 2006
than that in 2003.
Although such a poverty line seems very high and unlikely to be set at that
value, the main point of the exercise is clear. When two poverty incidence
curves cross, then an unambiguous judgment cannot be made. The crossing
208
may take place at a much lower level, as happened in the rural area. We
have already seen that the headcount ratio showed an increase in 2006
when the poverty line is set at GEL 75.4 but showed a decrease when the
poverty line is set at GEL 45.2. Given the infinite number of possible pov-
erty lines, it would be cumbersome to check them all one by one. Instead,
the poverty incidence curve is a convenient way of checking for dominance
(if two poverty incidence curves never cross). If dominance does not hold,
then the graph can tell us which part is responsible for the ambiguity.
Poverty Deficit Curve
A poverty deficit curve is useful while performing a dominance analysis of the

poverty gap measure with respect to the poverty line. In this dominance
exercise, the welfare indicator is per capita consumption expenditure,
assessed by lari. In figure 3.4, the horizontal axis denotes the welfare indica-
tor or per capita consumption expenditure. The height of the poverty den-
sity curve is proportional to the poverty gap measure, so that a larger height
Figure 3.4: Poverty Deficit Curves in Urban Georgia, 2003 and 2006
Urban
300
200
Total deficit
100
0
0 160 320 480 640 800
Welfare indicator
2003 2006
209
for a poverty line denotes a larger poverty gap measure. If a distribution’s

poverty deficit curve lies to the right of another distribution’s poverty deficit
curve, then the former distribution is understood to have an unambiguously
lower poverty gap measure, or the former distribution has lower poverty gap
measures for all poverty lines.
Figure 3.4 graphs the poverty deficit curves of urban Georgia for 2003
and 2006. The vertical axis reports total deficit, which is directly propor-
tional to the poverty gap measure for the corresponding poverty line. The
solid line denotes the poverty deficit curve for 2003, while the dashed line
denotes the poverty deficit curve for 2006. We saw earlier that the urban
poverty gap measure is higher in 2006 for both poverty lines: GEL 75.4 and
GEL 45.2. What about other poverty lines? Can we say that poverty has
unambiguously fallen for any poverty line? The graph suggests that we may
not be able to. If we set the hypothetical poverty line to about GEL 320,
then the poverty gap measure would have been lower in 2006 than in 2003.
Although such a poverty line seems very high and unlikely to be set at that
value, the main point of the exercise is clear. When two poverty deficit
curves cross, then an unambiguous judgment cannot be made based on the
poverty gap measure. Given the infinite number of possible poverty lines, it
would be cumbersome to check them all one by one. Instead, the poverty
deficit curve is a convenient way of checking for dominance (if two poverty
deficit curves never cross). If dominance does not hold, then the graph can
tell us which part is responsible for the ambiguity.
Poverty Severity Curve
A poverty severity curve is useful when performing a dominance analysis of

the squared gap measure with respect to the poverty line. In this dominance
exercise, the welfare indicator is the per capita consumption expenditure,
assessed by lari. In figure 3.5, the horizontal axis denotes the welfare indica-
tor or the per capita consumption expenditure. The height of the poverty
severity curve is proportional to the squared gap measure, so that a larger
height for a poverty line denotes a larger squared gap. If a distribution’s pov-
erty severity curve lies to the right of another distribution’s poverty severity
curve, then the former distribution is understood to have an unambiguously
210
Figure 3.5: Poverty Severity Curves in Rural Georgia, 2003 and 2006
Rural
100
80
Total severity, '000
60
40
20
0
0 160 320 480 640 800
Welfare indicator
2003 2006
lower squared gap, or the former distribution has a lower squared gap for all
poverty lines.
Figure 3.5 graphs the poverty severity curves of rural Georgia for 2003 and
2006. The figure’s vertical axis reports total severity, which is directly propor-
tional to the squared gap measure of the corresponding poverty line. The solid
line denotes the poverty severity curve for 2003, while the dashed line denotes
the poverty severity curve for 2006. We saw earlier that the rural squared gap
measure is higher in 2006 for both poverty lines: GEL 75.4 and GEL 45.2.
What about the other poverty lines? Can we say that poverty has unambigu-
ously fallen for any poverty line? The figure suggests that we may not be able
to. One of the poverty severity curves does not lie below another poverty
severity curve for all poverty lines. When two poverty severity curves cross,
then an unambiguous judgment cannot be made based on the squared gap
211
measure. Given the infinite number of possible poverty lines, it would be

cumbersome to check them all one by one. Instead, the poverty severity
curve is a convenient way of checking for dominance (if two poverty sever-
ity curves never cross). If dominance does not hold, then the graph can tell
us which part is responsible for the ambiguity.
Growth Incidence Curve
Figure 3.6 graphs the growth incidence curve of Georgia’s per capita con-
sumption expenditure. The vertical axis reports the growth rate of consump-
tion expenditure between 2003 and 2006, and the horizontal axis reports
the per capita consumption expenditure percentiles. We earlier reported the
growth rate of mean per capita consumption expenditure and found that the
overall growth rate was slightly negative. We also compared the median and
four other quantile incomes.
Figure 3.6: Growth Incidence Curve of Georgia between 2003 and 2006
Urban
3
2
Annual growth rate %
–1
–2
–3
1 10 20 30 40 50 60 70 80 90 100
Expenditure percentiles
Growth-incidence 95% confidence bounds
Growth at median Growth in mean
Mean growth rate
212
However, that analysis does not give us the entire picture, so we perform
this dominance analysis through a growth incidence curve that graphs the
growth rate of per capita consumption expenditure for each percentile of the
population. The height of a growth incidence curve for a particular percentile
of population is the per capita consumption expenditure growth of that per-
centile. In fact, a growth incidence curve assesses how the quantile incomes
change over time. If the growth rates of the lower quintiles are larger than
the growth rates of the upper quintiles, then the growth is said to be pro-poor.
The dotted-dashed straight line denotes the growth in mean per capita
expenditure, which is negative in this case. It is not necessary that the entire
population received an equal share of this growth. It is evident from the fig-
ure that the per capita expenditure growth rate for the population’s higher
percentiles between 2003 and 2006 is much larger and more positive than
that for their lower percentile counterparts. Even though growth has been
mixed throughout the quantile incomes, the lowest quantile income has a
large negative growth. Given that the growth rate was negative, this means
that the population’s poorer section had a proportionally larger decrease in
its per capita expenditure.
What we can state by looking at the figure is that the quantile ratios—
such as 90/10, 80/20, or 90/50—increased between 2003 and 2006. The
shaded area around the growth incidence curve reports the 95 percent con-
fidence bounds. Can we say something about poverty? Yes, we can. For an
absolute poverty line, the headcount ratio between 2003 and 2006 should
not fall because the per capita expenditures of the population’s lower per-
centile decreased. Thus, growth in Georgia between 2003 and 2006 was not
pro-poor.
Lorenz Curve
Figure 3.7 graphs the Lorenz curve of urban Georgia’s per capita expenditure
for 2003 and 2006. The vertical axis reports the share of total consumption
expenditure, and the horizontal axis reports the percentile of per capita
expenditure. A Lorenz curve graphs the share of total consumption expendi-
ture spent by each percentile of the population. Thus, the height of a Lorenz
curve for a particular percentile is the share of total consumption expenditure
spent by that percentile out of the region’s total consumption expenditure.
The Lorenz curve’s height is 1 when the percentile is 1. In other words, the
share of the total consumption expenditure spent by the entire population is
213
Figure 3.7: Lorenz Curves of Urban Georgia, 2003 and 2006
Urban
1.0
0.8
Lorenz curve
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1.0
2003, Gini=33.49 2006, Gini=35.65

Line of equality
100 percent. The diagonal straight line denotes the situation of perfect
equality: each person has the same per capita expenditure.
As inequality increases, the Lorenz curve bows out, and the area between
the Lorenz curve and the line of perfect equality increases. The area between
a Lorenz curve and the line of perfect equality is proportionally related to
the Gini coefficient: it is twice the Gini coefficient. If a distribution’s Lorenz
curve lies completely to the right of another Lorenz curve, then the former
distribution has unambiguously lower inequality, and any Lorenz-consistent
measure—such as the Gini coefficient, the Atkinson class of indices, and
the generalized entropy measures—ranks the former distribution as less
unequal. If the Lorenz curves of two distributions cross, we cannot unam-
biguously rank those two distributions, even when one is ranked as more
unequal than another by all the Lorenz-consistent measures we discussed
earlier. Therefore, the Lorenz curve provides an opportunity to conduct a
sensitivity analysis for the reported inequality measures.
214
The solid line represents the Lorenz curve for 2003, while the dotted
line corresponds to 2006. It is evident that the dotted curve lies nowhere to
the left of the solid curve. This implies that the inequality in urban Georgia
unambiguously increased in 2006 compared with 2003. If these two curves
had crossed, then the reported inequality measures would not have neces-
sarily agreed with each other.
Standardized General Mean Curve
Dominance in terms of the Lorenz curves is not very common. Therefore,

for inequality comparisons, we need to rely on various measures we cov-
ered earlier. We reported the Atkinson measures and generalized entropy
measures in addition to the Gini coefficient. The Gini coefficient is not
subgroup consistent, which means that if inequality in one region increases
but remains the same in another region, the overall inequality may fall. We
also showed in chapter 2 that each generalized entropy for a < 1 is a mono-
tonic transformation of the Atkinson inequality measures, and for a ≠ 1 it
is a monotonic transformation of the general means. However, we report the
Atkinson measures and the generalized entropy measures for only certain
values of parameter a. This exercise should be understood as a dominance
analysis of the Atkinson measures and the generalized entropy measures.
Figure 3.8 graphs the standardized general mean curve of Georgia’s
per capita expenditure for 2003 and 2006. The vertical axis reports the
standardized general mean of per capita expenditure, where standardiza-
tion is done by dividing the general mean of per capita expenditures by
their mean. The horizontal axis reports parameter a, which is the degree
of generalized mean and also known as the degree of a society’s aversion
toward inequality.
The general mean of a distribution tends toward the maximum and the
minimum per capita expenditure in the distribution when a tends to ∞
and – ∞, respectively. Given that the largest per capita expenditure in any
distribution is usually several times larger than the mean per capita expendi-
ture, allowing a to be very large prevents meaningful analysis. Therefore, we
restrict a to between – 5 and 5, which we consider large enough. The height
of a standardized general mean curve for a particular value of parameter a
is the general mean per capita expenditure divided by the mean per capita
expenditure. The height of any standardized general mean curve is 1 at a = 1.
215
Figure 3.8: Standardized General Mean Curves of Georgia, 2003 and 2006
Total
3.0
2.4
Generalized mean
1.8
1.2
0.6
0
–5 –4 –3 –2 –1 0 1 2 3 4 5
Alpha
2003 2006
The solid line represents Georgia’s standardized general mean curve in

2003, while the dashed line represents Georgia’s standardized general mean
curve in 2006. If a standardized general mean curve lies completely above
another standardized general mean curve to the left of a = 1 and completely
below to the right of a = 1, then every Atkinson inequality measure and
generalized entropy measure for a ≠ 1 agree that the former distribution has
lower inequality than the latter. It is evident from the figure that for larger
values of parameter a, inequality in 2006 has worsened. However, for a
less than 1, inequality has not significantly deteriorated. The standardized
general mean curve is a convenient way of verifying the robustness of the
Atkinson inequality measures and the generalized entropy measures.
Advanced Analysis
In this chapter’s final section, we discuss certain advanced analysis methods.

These techniques require knowledge of regression analysis. We assume read-
ers have the required background.
216
Consumption Regression
Table 3.28 analyzes determinants of the variable used for measuring welfare
(the per capita consumption expenditure in this case). Rows denote the set
of regressors (such as logarithm of household size, share of children in the
age group of 0–6 years, share of male adults, share of elderly) and a set of
Table 3.28: Consumption Regressions
2003 2006
Urban Rural Urban Rural
Coef SE Coef SE Coef SE Coef SE
Factors A B C D E F G H
Household characteristics
1 Log of household size −0.093 0.06 −0.010 0.05 −0.001 0.06 0.051 0.05
2 Log of household size squared −0.020 0.03 −0.078*** 0.02 −0.102*** 0.03 −0.114*** 0.02
3 Share of children age 0–6 years (dropped) (dropped) (dropped) (dropped)
4 Share of children age 7–16 years −0.252*** 0.09 0.223** 0.09 0.249** 0.10 0.076 0.09
5 Share of male adults −0.064 0.10 0.254*** 0.09 0.477*** 0.11 0.251*** 0.10
6 Share of female adults −0.004 0.10 0.453*** 0.10 0.592*** 0.11 0.435*** 0.10
7 Share of elderly (age ≥60 years) −0.124 0.11 0.462*** 0.10 0.488*** 0.12 0.355*** 0.10
Characteristics of household head
8 Log of household head’s age −0.063 0.05 0.076 0.05 −0.318*** 0.05 0.210*** 0.05
Regions
9 Kakheti (dropped) (dropped) (dropped) (dropped)
10 Tbilisi 0.446*** 0.05 (dropped) 0.258*** 0.05 (dropped)
11 Shida Kartli 0.182*** 0.06 0.147*** 0.03 −0.050 0.06 0.182*** 0.03
12 Kvemo Kartli 0.061 0.06 0.075** 0.03 −0.023 0.06 0.183*** 0.03
13 Samtskhe-Javakheti −0.115* 0.06 0.185*** 0.03 0.231*** 0.07 0.163*** 0.04
14 Ajara 0.226*** 0.06 −0.035 0.04 0.103* 0.06 0.067* 0.04
15 Guria −0.077 0.08 0.250*** 0.03 0.030 0.08 0.131*** 0.04
16 Samegrelo 0.112** 0.06 0.194*** 0.03 −0.007 0.06 0.238*** 0.03
17 Imereti 0.270*** 0.05 0.529*** 0.03 0.208*** 0.05 0.381*** 0.03
18 Mtskheta-Mtianeti −0.060 0.07 0.164*** 0.03 0.020 0.08 0.144*** 0.04
sland
19 0 ha (dropped) (dropped) (dropped) (dropped)
20 Less than 0.2 ha 0.121*** 0.03 0.162*** 0.05 0.104*** 0.03 0.166*** 0.04
21 0.2–0.5 ha 0.180*** 0.04 0.356*** 0.04 0.138*** 0.05 0.193*** 0.03
22 0.5–1.0 ha 0.255*** 0.05 0.478*** 0.04 0.125* 0.07 0.365*** 0.03
23 More than 1.0 ha 0.021 0.09 0.565*** 0.05 0.192** 0.08 0.484*** 0.04
Gender of household head
24 Male (dropped) (dropped) (dropped) (dropped)
25 Female −0.073*** 0.02 −0.002 0.02 −0.101*** 0.02 −0.027 0.02
Education of household head
26 Elementary or less (dropped) (dropped) (dropped) (dropped)
27 Incomplete secondary −0.067 0.07 0.034 0.03 0.226*** 0.07 0.086*** 0.03
28 Secondary 0.021 0.06 0.105*** 0.03 0.179*** 0.06 0.196*** 0.03
29 Vocational-technical 0.118* 0.06 0.147*** 0.04 0.225*** 0.07 0.255*** 0.04
30 Special secondary 0.156*** 0.06 0.217*** 0.03 0.269*** 0.06 0.322*** 0.04
31 Higher education 0.289*** 0.06 0.274*** 0.03 0.441*** 0.06 0.477*** 0.04
(continued)
217
Table 3.28: Consumption Regressions (continued)
2003 2006
Coef SE Coef SE Coef SE Coef SE
Factors A B C D E F G H
Employment status of household head
Self-employed
32 Agriculture (dropped) (dropped) (dropped) (dropped)
33 Industry −0.028 0.09 0.430*** 0.09 −0.122 0.11 0.208* 0.11
34 Trade 0.082 0.05 0.275*** 0.05 0.056 0.06 0.193*** 0.06
35 Transport 0.026 0.08 0.311*** 0.07 −0.039 0.08 0.311*** 0.07
36 Other services 0.072 0.07 0.340*** 0.08 −0.099 0.07 0.033 0.09
Employed
37 Industry −0.043 0.06 0.127** 0.06 −0.036 0.06 0.140** 0.06
38 Trade −0.094 0.06 0.144 0.09 −0.024 0.07 0.115 0.10
39 Transport −0.021 0.06 0.212*** 0.08 −0.174** 0.07 0.282*** 0.08
40 Government −0.041 0.06 0.227*** 0.06 0.012 0.07 0.277*** 0.08
41 Education −0.037 0.06 0.054 0.06 −0.029 0.07 0.045 0.07
42 Health care −0.041 0.08 0.085 0.15 −0.039 0.08 0.279* 0.15
43 Other −0.120** 0.05 0.150*** 0.04 −0.022 0.06 0.005 0.05
44 Unemployed −0.376*** 0.05 −0.138** 0.06 −0.325*** 0.05 −0.066 0.06
45 Inactive −0.219*** 0.04 −0.117*** 0.02 −0.169*** 0.05 −0.067*** 0.02
Other
46 Constant 4.851*** 0.21 3.425*** 0.19 5.328*** 0.20 2.976*** 0.20
47 Number of observations 4,525 7,106 4,112 6,773
48 Adjusted R2 0.18 0.20 0.17 0.16
Note: Coef = coefficient. ha = hectare. SE = standard error, sland = area of land owned.
*** p < 0.01, ** p < 0.05, * p < 0.1.
dummy variables (such as regional dummies, gender dummies, dummies for

household head education, and dummies for household head employment
status). Columns report regression coefficients (Coef) and standard errors
(SE) of four different ordinary least square regressions, where the depen-
dent variable, or the regressand, is the logarithm of per capita consumption
expenditure. The four regression results correspond to the urban and rural
areas for 2003 and 2006.
Each regression result has two columns. The first column reports regres-
sion coefficients and the second reports standard errors of the coefficients.
A regression coefficient of any regressor indicates the change in the regres-
sand caused by a one-unit increase in that regressor. The standard error
of a regressor indicates the reliability of its coefficient. Standard errors are
always positive, and a higher standard error indicates lower reliability of the
coefficient.
218
Rows 46, 47, and 48 report the intercept term, number of observations,
and adjusted R-squares (R2), respectively. The intercept term, or constant
term, denotes the level of the consumption expenditure logarithm not
explained or determined by any regressors, or adjusted R-square denoted
power of prediction of all regressors, or the model’s goodness-of-fit. If the
adjusted R-square is 1, then the regressors predict the regressand with com-
plete accuracy. If a regressor’s p-value is less than 0.01, then *** is added
to the coefficient. If the p-value is less than 0.05, then ** is added to the
coefficient. Finally, if the p-value is less than 0.1, then * is added to the
coefficient. P-values denote regressors’ significance level.
Note that all variables in the regions, sland, gender of household head, edu-
cation of household head, and employment status of household head categories
are dummy variables or binary variables. They take a value of only 0 or 1.
A binary variable coefficient denotes the change in regressand when
the dummy variable’s value changes from 0 to 1. For example, consider the
coefficient of the regressor Female in the household head gender category for
urban regression in 2003. The coefficient is –0.073 [25,A], implying that the
per capita expenditure logarithm for a member in a female-headed household
is 0.073 units lower than that of a male-headed household. The regressor’s
standard error is 0.02 [25,B] with a p-value less than 0.01 (indicated by ***
after the regressor), and thus the coefficient is highly significant. The coef-
ficient of the same regressor for urban regression in 2006 is –0.101 [25,E]
with a p-value of less than 0.01, implying that the per capita consumption
expenditure gap between female- and male-headed households increased
over the three-year period.
The table provides a detailed analysis of the determinants of per capita

consumption expenditure. If we focus on column A, it is evident that vari-
ables such as the share of children age 7–16 years [row 43], female-headed
households [row 25], and household head unemployed [row 44] and inactive
employment [row 45] status have significant negative effects on per capita
consumption expenditure for the urban area in 2003.
In contrast, the variables such as 0.5–1.0 hectare of landholding [row 22],
household head higher education [row 31], and living in Imereti [row 17]
have a significant positive impact on per capita expenditure for both urban
and rural areas in both years. Hence, the analysis summarized in table 3.28
219
provides a tool to understand per capita consumption expenditure determi-

nants and to develop appropriate poverty eradication policies.
Changes in the Probability of Being in Poverty
Table 3.29 analyzes changes in the probability of being in poverty using a

probit regression model based on the consumption regression presented in
table 3.28. Rows denote changes in values for various variables—such as
change from having no children 0–6 years old to having one child, change
Table 3.29: Changes in the Probability of Being in Poverty

percent
2003 2006
Variables A B C D
Demographic event, child born in the family
1 Change from having no children 0–6 years old to having 1 child 2.0 18.0 31.5 17.8
2 Change from having no children 0–6 years old to having 2 children 4.7 33.1 57.9 33.5
Land acquisition event
3 Change from “0 ha” to “less than 0.2 ha” −18.9 −15.3 −16.6 −18.0
4 Change from “0 ha” to “0.2–0.5 ha” −27.4 −33.6 −21.8 −20.9
5 Change from “0 ha” to “0.5–1.0 ha” −37.4 −44.5 −19.8 −38.5
6 Change from “0 ha” to “over 1.0 ha” −3.5 −51.9 −29.5 −49.6
Change of household head (following divorce, migration, and so forth)
7 Change from “Male” to “Female” 13.0 0.2 18.4 3.9
Education event: change in household head’s education
8 Change from “Elementary or less” to “Incomplete 10.5 −4.5 −28.0 −10.3
secondary”
9 Change from “Elementary or less” to “Secondary” −3.3 −13.6 −22.4 −23.0
10 Change from “Elementary or less” to “Vocational-technical” −17.6 −18.8 −27.8 −29.5
11 Change from “Elementary or less” to “Special secondary” −23.0 −27.2 −33.0 −36.7
12 Change from “Elementary or less” to “Higher education” −40.3 −33.7 −51.4 −51.8
Sector of employment event: household head’s sector of employment
13 Change from “Agriculture” to “Industry” 5.7 −53.1 25.3 −28.1
14 Change from “Agriculture” to “Trade” −15.6 −36.5 −10.6 −26.2
15 Change from “Agriculture” to “Transport” −5.2 −40.5 7.7 −40.3
16 Change from “Agriculture” to “Other Services” −13.8 −43.8 20.3 −4.7
17 Change from “Agriculture” to “Industry” 8.8 −17.7 7.1 −19.4
18 Change from “Agriculture” to “Trade” 19.6 −20.0 4.8 −16.0
19 Change from “Agriculture” to “Transport” 4.2 −28.7 36.8 −37.0
20 Change from “Agriculture” to “Government” 8.4 −30.7 −2.3 −36.4
21 Change from “Agriculture” to “Education” 7.5 −7.8 5.7 −6.4
22 Change from “Agriculture” to “Health Care” 8.3 −12.0 7.8 −36.7
23 Change from “Agriculture” to “Other” 25.3 −20.9 4.3 −0.7
24 Change from “Agriculture” to “Unemployed” 87.4 20.7 72.1 9.7
25 Change from “Agriculture” to “Inactive” 48.2 17.4 35.6 9.8
Source: Based on consumption regression presented in table 3.28.
220
from owning 0 hectare of land to > 1 hectare of land, and change from
male-headed household to female-headed household. Columns report the
percentage changes in the probability of being in poverty for rural and urban
areas and across 2003 and 2006.
Recall from our discussion about table 3.28 that the interpretation of
dummy or binary variables is different from that of continuous variables.
A dummy variable, unlike a continuous variable, may take only a value of
either 0 or 1. Table 3.28 described how the probability of being in poverty
changes as values of certain variables change.
The probability of being in poverty in 2003 increased by 2.0 percent
[1,A] if an individual moved from an urban household with no children in
the 0–6 years age group to an urban household with one child in the same
age group, all other factors being identical. The probability of being in pov-
erty in 2003 is increased by 18.0 percent [1,B] if an individual moved from
a rural household with no children in the 0–6 years age group to a rural
household with one child in the same age group, all else being identical. In
the urban area, the increase in the probability of being in poverty in 2006
for the same reason is 31.5 percent [1,C].
Similarly, in 2003 if an individual moved from a male-headed urban
household to a female-headed urban household, all else being identical,
then the probability of being in poverty increased by 13.0 percent [7,A].
If an individual moved from a male-headed rural household to a female-
headed rural household, all else being identical, then the probability of
being in poverty increased by only 0.2 percent [7,B]. The largest increase in
the probability of being in poverty in 2003 in the urban area occurred when
an individual moved from a household where the head is employed in the
agricultural sector to a household where the head is unemployed [24,A], all
else being identical.

The table provides a detailed analysis of how the probability of being in pov-
erty changes when some of the crucial determinants of poverty are adjusted.
Note that if the household head’s education in the urban area in 2006
increased from elementary education or less to secondary education, all else
remaining equal, then the probability of a member in that household being
in poverty fell by 22.4 percent [9,C]. Similarly, in rural Georgia for both
years, if the household head transferred from the agricultural sector to any
221
other employment sector, all else being equal, then the probability of being
in poverty fell. Hence, this analysis provides a tool to better understand the
source of poverty and what type of policy would be more efficient in terms
of eradicating poverty.
Growth and Redistribution Decomposition of Poverty Changes
Table 3.30 decomposes the change in poverty into a change in the mean per
capita consumption expenditure and a change in distribution of consump-
tion expenditure around the mean, following Huppi and Ravallion (1991).
Table rows denote three regions—urban, rural, and all of Georgia—for two
different poverty lines. The per capita consumption expenditure is measured
in lari per month. Poverty lines are set at GEL 75.4 (poor) and GEL 45.2
(extremely poor) for each household and household member. For simplicity
in this table, we present the decomposition for headcount ratio only, but the
technique is equally applicable to other poverty measures in the FGT class.
Columns A and B report the headcount ratio of the three regions for
years 2003 and 2006, respectively, and column C reports the changes over
time. Columns D, E, and F decompose the change in the headcount ratio
between 2003 and 2006 into three different terms. Column D reports the
effect of growth on poverty, referred to as the growth effect. Column E reports
the effect of redistribution on poverty and is called the redistribution effect.
Column F reports the interaction term and is referred to as the interaction
effect, following Huppi and Ravallion (1991).
Table 3.30: Growth and Redistribution Decomposition of Poverty Changes,

Headcount Ratio
percent
2003 2006 Actual change Growth Redistribution Interaction

Region A B C D E F
1 Urban 28.1 30.8 2.7 0.6 1.9 0.1
2 Rural 31.6 31.1 −0.5 −0.7 −0.1 0.3
3 Total 29.9 31.0 1.0 0.0 1.0 0.0
4 Urban 8.9 9.3 0.4 0.3 0.0 0.1
5 Rural 11.4 12.1 0.7 −0.2 1.0 0.0
6 Total 10.2 10.7 0.5 0.0 0.5 0.0
222
It is evident from the table that the overall headcount ratio in 2003 is
29.9 percent [3,A], which increased to 31.0 percent in 2006 [3,B]. These
numbers can be verified from table 3.2. The actual change in the overall
headcount ratio is 1.0 percentage point (rounded) [3,C]. The actual change
is broken down into three components: growth effect, redistribution effect,
and interaction effect. By looking at the corresponding figures in columns D,
E, and F, we see that the change is caused mainly by redistribution rather
than growth. We can verify from table 3.1 that growth in mean is negligible
compared to change in inequality in terms of the Gini coefficient.
The picture is slightly different for the urban and rural areas. The urban
headcount ratio rose by 2.7 percentage points from 28.1 percent [1,A] to
30.8 percent [1,B], with both growth effect and redistribution effect being
positive. The urban redistribution effect [1,E] is more than three times
larger than the urban growth effect [1,D]. For the rural area, the headcount
ratio fell from 31.6 percent [2,A] to 31.1 percent [2,B]. In this case, both the
growth effect [2,D] and the redistribution effect [2,E] are negative.
The appendix contains additional tables and figures that may be helpful
in understanding concepts and results in terms of the data for Georgia in
2003 and 2006.
Note
1. For technical details, see Huppi and Ravallion (1991).
Reference
Huppi, M., and M. Ravallion. 1991. “The Sectoral Structure of Poverty dur-
ing an Adjustment Period: Evidence for Indonesia in the Mid-1980s.”
World Development 19 (12): 1653–78.
223
ESTIMATING WEALTH EFFECTS 115
ESTIMATING WEALTH EFFECTS WITHOUT EXPENDITURE DATA—OR TEARS:

AN APPLICATION TO EDUCATIONAL ENROLLMENTS IN STATES OF INDIA*
DEON FILMER AND LANT H. PRITCHETT
Using data from India, we estimate the relationship between tures. The econometric evidence suggests that the asset index,
household wealth and children’s school enrollment. We proxy wealth as a proxy of economic status for use in predicting enrollments,
by constructing a linear index from asset ownership indicators, us- is at least as reliable as conventionally measured consumption
ing principal-components analysis to derive weights. In Indian data expenditures, and sometimes more so.
this index is robust to the assets included, and produces internally
This straightforward and pragmatic method of construct-
coherent results. State-level results correspond well to independent
data on per capita output and poverty. To validate the method and ing a proxy for household economic status produces results
to show that the asset index predicts enrollments as accurately as that are reassuringly consistent with other approaches (see
expenditures, or more so, we use data sets from Indonesia, Paki- Montgomery et al. 2000 and references therein) and poten-
stan, and Nepal that contain information on both expenditures and tially is broadly applicable. Demographic and Health Sur-
assets. The results show large, variable wealth gaps in children’s veys (DHS) have been conducted with nearly identical sur-
enrollment across Indian states. On average a “rich” child is 31 vey instruments in more than 40 countries. (In India this sur-
percentage points more likely to be enrolled than a “poor” child, vey is called the National Family Health Survey, or NFHS.)
but this gap varies from only 4.6 percentage points in Kerala to 38.2 These surveys include assets and housing conditions but not
in Uttar Pradesh and 42.6 in Bihar. consumption expenditures (except for an experimental con-
sumption module included in Indonesia in 1994, a feature
that we exploit later in this paper).
T his paper has both an empirical and a methodological goal.
The empirical goal is to investigate the effect of household eco-
Our proposed method for estimating household wealth in
the DHS/NFHS allows estimates of the association of wealth
nomic status on children’s educational attainment across the with education across households and permits a comparison of
states of India. To accomplish this aim, we propose and defend wealth gaps across countries (and in India across states). In
a method for estimating the effect of household wealth on edu- separate papers we use the asset index to examine wealth and
cational outcomes even in the absence of survey questions on gender gaps in enrollment and educational attainment in more
income or expenditures. We use data on asset ownership (e.g., than 35 countries (Filmer 2000; Filmer and Pritchett 1999a).
owning a bicycle or radio) and housing characteristics (e.g., We limit ourselves here to demonstrating the validity
number of rooms, type of toilet facilities), henceforth called and usefulness of an asset index in analyzing education out-
“asset indicators” or “asset variables,” to construct an “asset incomes. The same method might be equally useful in examin-
dex.” We handle the vexing problem of choosing the appropri- ing wealth differences in other socioeconomic outcomes in
ate weights by using the statistical procedure of principal com- the DHS data, such as mortality, morbidity, utilization of
ponents. We demonstrate the empirical validity of this approach health facilities, fertility, and contraceptive use (Bonilla-
for India by comparing the state-level averages of the asset in- Chacin and Hammer 1999; Gwatkin et al. 2000; Stecklov,
dex with data on poverty rates and gross state product per Bommier, and Boerma 1999). Sahn and Stifel (1999) have
capita. Going further, we use data sets from Indonesia, Nepal, taken a similar “index” approach to making comparisons of
and Pakistan, which contain both expenditures and asset vari- relative poverty over time and across countries, using DHS
ables for the same households, to show a reasonable correspon- data for nine countries in Africa.
dence between the classification of households based on the as- A proxy for wealth not only is useful in examining ef-
set index and a classification based on consumption expendi- fects of wealth, but also is needed as a “control” variable in
estimating effects of variables potentially correlated with
*
Deon Filmer, Development Research Group, The World Bank, 1818 household wealth, such as maternal education. The method
H Street NW, Washington, DC 20433; E-mail: dfilmer@worldbank.org. proposed here provides a simple technique for creating a
Lant H. Pritchett, John F. Kennedy School of Government and The World wealth proxy in the absence of either income or expenditure
Bank. This research was funded in part through World Bank research sup-
port grant (RPO 682-11). We would like to thank Harold Alderman, Zoubida data (an issue discussed further in Montgomery et al. 2000).
Allaoua, Gunnar Eskeland, Jeffrey Hammer, Keith Hinchliffe, Valerie
Kozel, Alan Krueger, Peter Lanjouw, Marlaine Lockheed, Berk Ozler, and OVERCOMING THE PROBLEM OF RANKING
Martin Ravallion for valuable discussions and comments, including some HOUSEHOLDS IN THE ABSENCE OF
on an earlier version of this paper (Filmer and Pritchett 1998). The findings, CONSUMPTION OR EXPENDITURE DATA
interpretations, and conclusions expressed here are entirely those of the au-
thors. They do not necessarily represent the views of the World Bank, its The DHS and NFHS data present both an opportunity and a
executive directors, or the countries they represent. challenge. The opportunity is a rich set of large, representative
Demography, Volume 38-Number 1, February 2001: 115–132 115
116 DEMOGRAPHY, VOLUME 38-NUMBER 1, FEBRUARY 2001
surveys with nearly identical questionnaires covering a large Construction of an Asset Index
number of countries or Indian states. The challenge is that the
DHS/NFHS surveys contain no data on income or on house- How does one aggregate various asset ownership indicators
hold consumption expenditures. The latter is unfortunate be- into one variable to proxy for household “wealth”? Even if
cause a large literature establishes the theoretical underpin- the question is simplified by limiting the aggregation to a
nings of consumption expenditures as a measure of current linear index, how should the weights be chosen? Equal
and long-run household (and implicitly individual) welfare. weights have the appeal of simplicity and apparent objectiv-
(For example, see Deaton 1997; Deaton and Muellbauer ity, but these qualities only mask the fact that the imposition
1980; Deaton and Zaidi 1999.) As a result, expenditures are of numeric equality is completely arbitrary.
routinely used in measuring poverty. We overcome the ab- A second option would be to estimate the current value
sence of expenditure data by using the information collected of household assets using explicit and implicit “prices” as
on assets owned by household members and on housing char- weights. If the data did not include current values, then the
acteristics. These data are used to generate an asset index that purchase price, the date of purchase, and a suitable deprecia-
proxies for wealth and hence for long-run economic status. tion rate could be used to estimate them. The DHS and NFHS
Three clarifications will avoid confusion at the outset. data, however, are typical in that they include binary indica-
First, we are not proposing this asset index as a measure tors on asset ownership and housing characteristics but not
either of current welfare or of poverty.1 We use the asset current value, purchase price, or vintage of assets.
index as a determinant of current enrollment, which de- A third possible solution is not to build an index at all but
pends on long-run as well as current expenditures: house- simply to enter all of the asset variables separately in a linear
holds will smooth schooling expenditures over time and are multivariate regression equation. This procedure implicitly
unlikely to respond to temporary shocks by withdrawing creates weights on the variables. Such an approach handles
children from school. the problem of controlling for wealth in estimating the im-
Second, unlike Montgomery et al. (2000), we are not cre- pact of other, nonwealth, variables. Yet the linear index of the
ating an asset index as a proxy for current consumption ex- assets using regression weights does not estimate the wealth
penditures. We view the asset index and expenditures as a effect because many assets exert both a direct and an indirect
proxy for something unobserved: a household’s long-run eco- effect on outcomes. For instance, the household’s use of elec-
nomic status. Therefore, although it is reassuring that these tricity for lighting may serve as a proxy for wealth, but also
two proxies are related, discrepancies in the classification of may make study easier and hence may lower the opportunity
households cannot be assumed to be “mistakes” of the asset costs of schooling. The availability of piped water not only
index. The two, as proxies, have conceptually and empirically may indicate greater wealth but also may reduce the time
distinct limitations; therefore discrepancies could just as eas- needed for water collection and thus may reduce the opportu-
ily indicate limitations of current consumption expenditures. nity costs of schooling. This argument is even clearer with
The major problem with the asset index is that the weights on health outcomes because water and sanitation variables have
individual indicators are not grounded theoretically. The ma- strong independent effects on children’s health (Bonilla-
jor problem with current expenditures as a proxy for long-run Chacin and Hammer 1999). Therefore, although linear regres-
wealth is the presence of short-term fluctuations.2 sion coefficients implicitly produce weights for the linear in-
Third, we emphasize that the principal-components ap- dex of the asset variables that predicts the dependent variable
proach is a pragmatic response to a data constraint problem. most closely, there is no way to infer from these uncon-
In this paper we set the modest goal of establishing the valid- strained coefficients the impact of an increase in wealth.
ity of this particular approach in this application; we do not
establish that our approach is “optimal” because there may be Using Principal Components
other methods that possess superior statistical properties. We implement a different approach: we use the statistical
procedure of principal components to determine the weights
1. The limitation in poverty analysis is twofold. First, the conventional for an index of the asset variables. Principal components is a
notion of poverty is based on the flow of consumption relative to some pre- technique for extracting from a set of variables those few
determined poverty threshold, whereas we, by aggregating assets, are estab-
lishing only a measure of a stock. Second, the categorization used is based
orthogonal linear combinations of the variables that capture
on a relative measure (that is, the household’s ranking within the distribu- the common information most successfully. Intuitively the
tion), whereas poverty thresholds typically are based on the expenditures first principal component of a set of variables is the linear
necessary for the consumption of a determined bundle of goods. index of all the variables that captures the largest amount of
2. Current expenditures would be a perfect measure only under the unre- information that is common to all of the variables. (For read-
alistically strong assumptions of perfect foresight and perfect capital markets.
Current expenditures are a popular proxy for long-run wealth for two main able and intuitive descriptions of principal components, see
reasons: the theoretical justification that expenditures are superior to current Lindeman, Merenda, and Gold 1980; StataCorp 1999.)
income as a proxy for long-run income because of consumption smoothing, Suppose we have a set of N variables, a*1j to a*N j, repre-
and, perhaps even more important, the pragmatic justification that expendi- senting the ownership of N assets by each household j. Prin-
tures are easier to measure than income in most rural settings. Current expen-
ditures, however, are not preferred unambiguously to asset ownership on ei-
cipal components starts by specifying each variable normal-
ther of these dimensions: asset ownership also reflects smoothing and is (if ized by its mean and standard deviation: for example, a1j =
anything) easier to measure than either income or expenditures. (a* 1j – a* 1) / (s* 1), where a*1 is the mean of a* 1j across
households and s*1 is its standard deviation. These selected holds surveyed in each Indian state varies from 9,963 in
variables are expressed as linear combinations of a set of un- Uttar Pradesh to around 1,000 in the small northeastern
derlying components for each household j: states. The survey includes data on 21 asset indicators that
can be grouped into three types: household ownership of
a1j = v11 × A1j + v12 × A2j +...+ v1N × ANj
consumer durables, with eight questions (clock/watch, bi-
... j = 1,...J
cycle, radio, television, bicycle, sewing machine, refrigera-
aNj = vN1 × A1j + vN2 × A2j +...+ vNN × ANj , (1)
tor, car); characteristics of the household’s dwelling, with
where the As are the components and the vs are the coeffi- 12 indicators (three about toilet facilities, three about the
cients on each component for each variable (and do not vary source of drinking water, two about rooms in the dwelling,
across households). Because only the left-hand side of each two about the building materials used, and one each about
line is observed, the solution to the problem is indeterminate. the main source of lighting and cooking); and household
Principal components overcomes this indeterminacy by landownership.
finding the linear combination of the variables with maxi- Table 1 reports the scoring factors from the principal-
mum variance—the first principal component A1j— and then components analysis of the 21 variables (see Eq. (3)). The
finding a second linear combination of the variables, orthogo- mean value of the index is 0; the standard deviation is 2.3.
nal to the first, with maximal remaining variance, and so on. Because all the asset variables (except “number of rooms”)
Technically the procedure solves the equations (R – nI)vn = take only the values 0 or 1, the weights have an easy inter-
0 for n and vn, where R is the matrix of correlations between pretation: a move from 0 to 1 changes the index by f1i / s*i
the scaled variables (the as) and vn is the vector of coeffi- (reported in column 4). A household that owns a clock has
cients on the nth component for each variable. Solving the an asset index higher by 0.54 than one that does not; own-
equation yields the characteristic roots of R, n (also known ing a car raises a household’s asset index by 1.21 units; us-
as eigenvalues) and their associated eigenvectors, vn. The fi- ing biomass as the main cooking fuel lowers the asset index
nal set of estimates is produced by scaling the vns so the sum by 0.67.
of their squares sums to the total variance, another restriction We sort individuals by the asset index and establish cut-
imposed to achieve determinacy of the problem. off values for percentiles of the population. We then assign
The “scoring factors” from the model are recovered by households to a group on the basis of their value on the index.
inverting the system implied by Eq. (1), and yield a set of For expository convenience, we refer to the bottom 40% as
estimates for each of the N principal components: “poor,” the next 40% as “middle,” and the top 20% as “rich,”
asking the reader to remember that this classification does
A1j = f11 × a1j + f12 × a2j +...+ f1N × aNj
not follow any of the usual definitions of poverty.
... j = 1,...J
The difference in the average index between the poorest
ANj = fN1 × a1j + fN2 × a2j +...+ fNN × aNj. (2)
and the middle group is 2.07 units. One example of a combi-
The first principal component, expressed in terms of the nation of assets that would produce this difference is owning
original (unnormalized) variables, is therefore an index for a radio (0.54), having a kitchen as a separate room (0.37),
each household based on the expression having electricity for lighting (0.57), and not having a dwell-
ing of all low-quality materials (0.55). The average asset in-
A1j = f11 × (a*1j – a*1)/(s*1)+...+ f1N × (a*Nj – a*N) / (s*N). (3)
dex is 3.78 units higher for the richest than for the middle
The crucial assumption for our analysis (and it is just that—an group. This difference is equivalent to owning a motor
assumption) is that household long-run wealth explains the scooter (0.91) and a television (0.83), having a flush toilet
maximum variance (and covariance) in the asset variables. (0.75) and a house of all high-quality materials (0.73), and
There is no way to test this assumption directly, but in the next not using biomass as the main cooking fuel (0.67).
two sections we provide evidence that this method produces
reasonable results. Below we illustrate the method using the The Reliability of the Asset Index
states of India, and show the internal and external coherence of For India the asset index performs well on three dimensions.
the asset index. Then we compare this the method with the use First, it is internally coherent because average asset owner-
of consumption expenditures in three countries; we find that ship differs markedly across the poor, middle, and rich house-
this simple asset index has reasonable agreement with con- holds for each asset; second, it is robust to the assets included;
sumption expenditures and performs as well, or better, in pre- third, it produces reasonable comparisons with measures of
dicting education outcomes in two of the countries. state-level poverty and output. The index has drawbacks, how-
ever, especially possible problems with urban/rural compari-
CONSTRUCTION AND INTERNAL VALIDITY OF sons.
THE ASSET INDEX IN INDIA Internal coherence. The last three columns of Table 1
Construction of the Asset Index Using the NFHS compare the average ownership of each asset across the
Data poor, middle, and rich households. We find large differences
across groups for almost all assets. Clock ownership is 16%
The NFHS survey covers about 88,000 households and for the poor versus 98% for the rich. Also, the poor cook
about one-half million individuals. The number of house- with biomass fuel almost exclusively (96%), whereas only
TABLE 1. SCORING FACTORS AND SUMMARY STATISTICS FOR VARIABLES ENTERING THE COMPUTATION OF THE
FIRST PRINCIPAL COMPONENT
All India
_______________________________________________ Means
_________________________________
Scoring Scoring Poorest Middle Richest
Factors Mean SD Factor × SD 40% 40% 20%
Own Clock/Watch 0.270 0.533 0.499 0.54 0.164 0.739 0.985
Own Bicycle 0.130 0.423 0.494 0.26 0.264 0.510 0.621
Own Radio 0.248 0.396 0.489 0.51 0.101 0.522 0.838
Own Television 0.339 0.209 0.407 0.83 0.000 0.127 0.866
Own Sewing Machine 0.253 0.182 0.385 0.66 0.015 0.179 0.580
Own Motorcycle/Scooter 0.249 0.082 0.274 0.91 0.001 0.031 0.375
Own Refrigerator 0.261 0.068 0.252 1.04 0.000 0.006 0.353
Own Car 0.129 0.012 0.107 1.21 0.000 0.001 0.059
Drinking Water From Pump/Well –0.192 0.609 0.488 –0.39 0.800 0.569 0.242
Drinking Water From Open Source –0.041 0.040 0.195 –0.21 0.057 0.036 0.005
Drinking Water From Other Source –0.002 0.019 0.138 –0.01 0.016 0.027 0.012
Flush Toilet 0.308 0.217 0.412 0.75 0.005 0.175 0.797
Pit Toilet/Latrine 0.040 0.086 0.280 0.14 0.040 0.127 0.111
None/Other Toilet 0.001 0.001 0.029 0.03 0.001 0.001 0.001
Main Source of Lighting Electric 0.284 0.510 0.500 0.57 0.143 0.700 0.989
Number of Rooms in Dwelling 0.159 2.676 1.957 0.08 1.975 2.965 3.739
Kitchen a Separate Room 0.183 0.536 0.499 0.37 0.312 0.643 0.848
Main Cooking Fuel Biomass (Wood/Dung/Coal) –0.281 0.776 0.417 –0.67 0.956 0.841 0.224
Dwelling All High-Quality Materials 0.309 0.237 0.425 0.73 0.005 0.218 0.821
Dwelling All Low-Quality Materials –0.273 0.483 0.500 –0.55 0.832 0.308 0.017
Own > 6 Acres Land 0.031 0.115 0.319 0.10 0.075 0.155 0.126
Economic Status Index 0.000 2.32 –2.00 0.071 3.857
Notes: Each variable beside number of rooms takes the value 1 if true, 0 otherwise. Scoring factor is the “weight” assigned to each variable (normalized by its
mean and standard deviation) in the linear combination of the variables that constitute the first principal component. The percentage of the covariance explained
by the first principal component is 26%. The first eigenvalue is 5.37; the second eigenvalue is 1.66.
Source: Authors’ calculation from NFHS 1992–1993.
22% of the rich do so. One might ask whether the asset in- tion between poor and rich on housing variables not related
dex tends too much to reflect community variables (espe- to infrastructure, such as “all high-quality materials in the
cially locally available infrastructure such as electricity for dwelling” (less than 1% of the poor versus 82% of the rich)
lighting or piped water) rather than household-specific vari- and “having a kitchen as a separate room” (31% of the poor
ables. On this score, we are reassured by the clear separa- versus 85% of the rich).
TABLE 2. CLASSIFICATION DIFFERENCES OF THE POOREST 40%: ALL INDIA

All Variables Except Asset Ownership,
Base Case: Drinking Water and Housing, and Only 8 Asset
All Variables Toilet Facilities Land Ownership Ownership Variables
Poorest 40% 100.0 95.1 87.7 80.2
Middle 40% 0.0 4.9 12.3 19.7
Richest 20% 0.0 0.0 0.0 0.1
Total 100.0 100.0 100.0 100.0
Spearman Rank
Correlation Coefficient,
Ranking of Households 1.0 .94 .87 .79
Robustness. The asset index produces very similar clas- TABLE 3. DISTRIBUTION OF INDIVIDUALS ACROSS
sifications when different subsets of variables are used in its GROUPS, STATE-LEVEL POVERTY, AND NET
construction. Table 2 reports the percentage of households DOMESTIC PRODUCT
classified in the poorest 40% when all assets are used, com- State Per Capita
pared with indices based on (1) all the variables except those Percentage in Poverty Rate Net State
related to drinking water and toilet facilities, (2) ownership Nationwide (Headcount Domestic
of consumer durables, housing quality, number of rooms, and Poorest 40% Index) Product
land ownership, and (3) only the ownership of consumer
Delhi 1.3
durables (e.g., watch, radio). Almost no households classi-
fied in the poorest group by the index using all variables Goa 5.6 10,128
would be classified as rich by any of the more limited mea- Himachal Pradesh 6.8 28.58
sures. The robustness of the classification is similar for the Punjab 8.4 11.46 10,857
middle and the rich groups. Haryana 10.5 25.22 9,609
A more general measure of the differences in rankings Jammu 14.5
can be derived from the rank correlation coefficient, which Kerala 15.1 25.12 5,065
compares the degree to which two methods produce the same
Mizoram 18.1
ranking of households. Even when the index is constructed
with only consumer durables, the correlation with the index Nagaland 20.3
that uses all the variables is .79 (all correlation coefficients Gujarat 26.8 24.15 7,586
in Table 2 are significant; p < .001, N = 87,175). Adding Maharashtra 26.9 36.82 9,270
more variables in constructing the index only increases the Karnataka 27.6 32.91 6,313
similarity of the rankings. Manipur 27.6
An additional check for robustness is made by using a Tamil Nadu 32.5 35.40 6,205
different methodology for deriving the weights. Although
Meghalaya 37.9 5,769
the theoretical underpinnings and the algorithms used in
factor analysis are close to those for principal components, Arunachal Pradesh 38.1 6,359
the two methodologies differ sufficiently to make factor Andhra Pradesh 39.0 21.87 5,802
analysis a possible alternative approach. The first factor de- Rajasthan 39.7 27.46 5,035
rived from a model analogous to that described above yields Tripura 41.8
a household ranking that has a .988 Spearman rank correla- West Bengal 44.3 36.94 5,901
tion with a ranking derived from principal components. Uttar Pradesh 48.6 41.55 4,280
Clearly, the results drawn from the asset index approach are
Madhya Pradesh 49.4 42.46 4,725
robust to whether one picks one or the other of these meth-
ods. Orissa 54.4 48.64 3,963
Comparisons across states. Because the poor, middle, Assam 58.3 41.09 5,056
and rich groups are defined on an “all-India” basis, states Bihar 61.5 55.15 3,280
differ as to the percentage of households in each group. All India 40.0 36.16 6,380
Thus we can compare state-by-state rankings with more
conventional measures. The national poverty rate based on Notes: The rank correlation coefficient between the percentage asset poor
consumption expenditures is 36%, roughly comparable to and the poverty rate is .794 (p < .001, N = 16); the rank correlation between
the percentage asset poor and per capita state product is –.864 (p < .001, N =
defining the “asset poor” as the poorest 40% nationally by 18).
the asset index. The rank correlation of the poverty rate
Sources: Authors’ calculation from NFHS 1992–1993, World Bank
with the proportion asset poor across states is .794 (p < (1998), and Agrawal and Varma (1996). Data on the headcount index are for
.001, N = 16). In a comparison of the first and the second 1993–1994.
columns of Table 3, both classifications show that Punjab,
Haryana, and Kerala have better than average economic sta-
tus and that Bihar, Orissa, and Uttar Pradesh are below av- whereas Assam looks poorer, with 58% asset poor but a per
erage. Differences exist, however: Maharashtra appears capita SDP of Rs 5,056.
richer by the asset index, while Andhra Pradesh looks Alternative interpretations. The first principal compo-
poorer. The rank correlation of the proportion of a state’s nent explains 25.6% of the variation in the 21 asset vari-
population asset poor with SDP per capita is –.864 (p < ables; this percentage is substantial but not overwhelming.
.001, N = 18). (By comparison, the rank correlation between Although the first principal component might well serve as
the poverty rate and per capita state domestic product is – a reasonable overall index, it is uncertain whether this com-
.729, p = .002, N = 15). Again, although the rankings agree ponent alone contains all of the relevant information. The
overall, certain states look different according to the two factor scores derived for the second principal component
rankings. For example, Kerala appears richer by the index, are more difficult to interpret. The procedure produces a
with only 15% asset poor and a per capita SDP of Rs 5,065, pattern of factor scores, which appears to assign positive
weights to assets that one might think richer rural house- see Dreze and Srinivasan 1997; Lanjouw and Ravallion
holds would own, and negative weights to assets that one 1995). The asset index is not adjusted for household size be-
might expect other types of households to own. (See Appen- cause the benefits of many of the assets, such as quality of
dix A, available from the authors.) This is somewhat worri- housing materials, source of fuel, or lighting, are present at
some because the asset index rankings show rural house- the household level. The Spearman rank correlations across
holds to be less “wealthy” than do conventional expenditure households are .64 for Nepal (p < .001, N = 3,372), .56 for
measures. Indonesia (p < .001, N = 16,242), and .43 for Pakistan (p <
One explanation for this discrepancy is that many of the .001, N = 1,192). Clearly, the degree of agreement among
asset variables depend on the availability of infrastructure the different rankings varies across the countries. Generally,
(electricity, piped water, sewerage); therefore urban house- the smaller the , the better the fit between assets and ex-
holds are more likely than rural households to appear well- penditure classifications; thus the asset index classification
off. On the other hand, this point may imply that standard fits total household expenditures better than is reported and
expenditure measures underestimate the difference between fits per capita expenditures worse than is reported. At the
rural and urban households’ welfare by not adequately ad- suggestion of a referee, we repeated the analysis using C/A:
justing real incomes for the implicit price differences for the that is, adjusting for the number of adults in the household.
services provided by infrastructure. For the analysis of en- The results differ neither qualitatively nor, to any substantial
rollment decisions, however, we want an index that captures degree, quantitatively.3
the dimensions of wealth relevant to education, which is not We assigned households to the poorest 40, middle 40,
necessarily net welfare. and richest 20 percentiles, using either the asset index or the
Finally, in this particular application, nothing in our expenditure measure. Table 4 shows the results of compar-
analyses depends on urban/rural comparisons. The bivariate ing the two classifications. The results in Indonesia and
analysis uses the pooled urban/rural data sets only to ana- Nepal are quite similar: roughly two-thirds of those classi-
lyze the stability of rankings within the entire samples. The fied into the poorest 40% by expenditures are also classified
subsequent multivariate analysis either uses only the rural into the poorest 40% by assets, and only 5% of those in the
data or, when data are pooled, controls for rural/urban sta- poorest 40% by expenditures appear in the richest 20% by
tus. Thus the analysis should not be affected by any level assets. The classification of the richest 20% shows less
difference due to systematic over- or understatement of the agreement: only 49 to 56% of households rich by expendi-
differences. (We refrain from interpreting the “urban” tures are also in the richest 20% by assets. Reassuringly,
coefficient because it will combine the “true” effect of ur- however, only 10 to 13% of those ranked in the richest 20%
ban residence with any effect of the construction of the in- by expenditures are in the poorest 40% by assets.
dex.) The results for Pakistan show less coherence between
Even though the procedure produces higher-order com- the two rankings. Although it is still the case that only 4%
ponents, they are, by construction, orthogonal to the first and of those who are poor by expenditures are rich by assets,
therefore will not create direct “excluded variable” bias in bi- only 60% of the expenditure poor are also asset poor. More-
variate comparisons of wealth and education. In the multi- over, only 43% of those in the richest 20% by expenditures
variate analysis discussed below, the results are robust to in- are also in the richest 20% by assets, and 22% of the richest
cluding higher-order components as linear indices in the 20% of households by expenditures are in the poorest 40%
specification. by assets.
ASSET INDEX VERSUS CONSUMPTION Comparison of Enrollment Rate Regressions

EXPENDITURES AS PROXIES FOR LONG-RUN Using the Two Measures
WEALTH Table 5 reports the results of probit regressions of “currently
Comparisons of Consumption Expenditures With enrolled” and “ever attended” for children age 6 to 14, and
Asset Index Classifications “completed grade 5” for those age 15 to 19. The regressions
control for child characteristics, residence, and characteris-
The Nepal Living Standards Survey (NLSS) conducted in tics of the household head—a specification that mimics the
1996 and the Pakistan Integrated Household Survey (PIHS) one used in the India analysis conducted below. The mar-
conducted in 1991 are “standard” surveys from the Living ginal wealth effect reported is the increase in the probability
Standards Measurement Study (LSMS) (Grosh and Glewwe that the education variable will equal 1 for a child in the rich-
1998). The Indonesian DHS (IDHS) conducted in 1994 in-
cluded an experimental module on consumption expenditures 3. We compare the asset index with consumption expenditures and not
(based closely on Indonesia’s National Socio-Economic Sur- with predicted consumption expenditures, where assets and other house-
vey—SUSENAS) for about half of the households. hold variables are used as instruments. Although some of the results would
In these three countries we compare an asset index with have appeared more similar if we had taken this “best practice” approach
(recently used and explored in Behrman and Knowles 1999), the conven-
total consumption expenditures (C) adjusted for household tional approach—particularly for bivariate/tabular analysis—is to use ac-
size (N), C/N, where the economies of scale parameter is tual, not predicted, expenditures. Therefore we use this as our baseline for
set to 0.6 (For discussions of the choice of this parameter, comparison.
TABLE 4. CLASSIFICATION DIFFERENCES: NEPAL, INDO- est versus the poorest quintile when all other variables are at
NESIA, AND PAKISTAN, URBAN AND RURAL their mean.
AREAS The results for current attendance in Nepal are almost
identical for the two proxies: the marginal effect of being in
Groups Based on Household the richest relative to the poorest quintile by assets is 34.0
a
Consumption Expenditures per Adjusted Size
____________________________________________ percentage points, versus 33.8 percentage points by expendi-
Groups Based tures. In contrast, the marginal effect in Pakistan is 36.7 per-
on Asset Index Poorest 40% Middle 40% Richest 20% centage points when quintiles are defined according to assets,
Nepal
but only 27.5 percentage points when the groups are defined
according to expenditures. The Indonesian results fall be-
Poorest 40% 65.2 37.8 12.6
tween the other two: the wealth gap in current enrollment is
Middle 40% 29.9 46.8 31.4 10.5 percentage points when quintiles are based on assets and
Richest 20% 4.9 15.4 56.0 8.7 percentage points when quintiles are based on expendi-
Total 100.0 100.0 100.0 tures.
Indonesia In Indonesia and Pakistan the asset index produces a
Poorest 40% 63.9 35.3 10.5 larger predicted gap between the richest and the poorest quin-
Middle 40% 31.7 49.1 41.0 tiles than do expenditures for all three education outcomes.
Richest 20% 4.4 15.6 48.5 In Nepal the differences are very close to 0 for current en-
Total 100.0 100.0 100.0 rollment and ever attended. For completion of grade 5, they
Pakistan even indicate a slightly smaller gap when the asset index is
Poorest 40% 61.2 40.0 20.2
used.
As discussed earlier, higher-order principal components
Middle 40% 34.9 42.5 37.1
are orthogonal to the asset index by construction. Neverthe-
Richest 20% 3.9 17.5 42.7 less, the results may be affected by introducing the higher-
Total 100.0 100.0 100.0 order terms into a nonlinear multivariate model. When we
Sources: Authors’ calculations from NLSS 1996, IDHS 1994, and PIHS include up to the fifth principal component in the set of re-
1991. gressors in the models underlying Table 5, the changes in the
a
Adjusted household size is equal to household size to the power 0.6. estimated marginal wealth effects are very small for all three
TABLE 5. MARGINAL EFFECTS OF BEING IN THE RICHEST QUINTILE (RELATIVE TO THE POOREST QUINTILE): NEPAL,
INDONESIA, AND PAKISTAN, URBAN AND RURAL AREAS
Based on Household Difference in “Wealth Gap”
Consumption Between Asset Index
Expenditures and Expenditures
Based on Asset Index (Adjusted for Size) Classification
Nepal
Currently attending school (ages 6 to 14) 34.0 33.8 0.2
Ever went to school (ages 6 to 14) 31.9 30.7 1.2
Completed at least grade 5 (ages 15 to 19) 44.9 48.5 –3.6
Indonesia
Completed at least grade 6 (ages 15 to 19) 8.5 4.3 4.2
Pakistan
Completed at least grade 5 (ages 15 to 19) 51.0 27.4 23.6
Notes: Values are derived from a probit regression of the outcome on a set of four dummy variables for quintile (the poorest category is the reference category)
and controls for gender, urban residence, age, gender of the head of the household, age of the head of the household, and education of the head of the household.
(Control variables are available from the authors.) Marginal effects are evaluated at the means of all the variables: that is, a specification that mimics that used in the
subsequent analysis for India.
Sources: Authors’ calculations from IDHS 1994, NLSS 1996, PIHS 1991.
outcome variables.4 For example, the marginal effect of be- TABLE 6. DIFFERENCE BETWEEN THE RICHEST QUIN-
longing to the richest versus the poorest quintile for current TILE AND THE POOREST QUINTILE, AVERAGE
attendance declines from 34.0 to 32.2 percentage points in ENROLLMENT RATES OF RURAL CHILDREN
Nepal, remains at 10.5 points for Indonesia, and increases AGE 6 TO 14
from 36.7 to 39.2 points in Pakistan.
Because the NFHS does not include both assets and con- Quintiles Constructed by
____________________________
sumption expenditures, we cannot conduct the same exercise Per Capita
for India as for Nepal, Indonesia, and Pakistan. Yet we can Consumption
compare wealth gaps in enrollment across quintiles calcu- Asset Index Expenditures Difference
lated with the asset index in the NFHS to wealth gaps based All India 52 33 19
on per capita consumption expenditures ( = 1) from an
analysis of Indian National Sample Survey (NSS) data Andhra Pradesh 55 37 19
(Haque, Lanjouw, and Ravallion 1998). We use a simple bi- Assam 36 21 15
variate analysis because we do not have the multivariate pre- Bihar 67 43 25
dicted probabilities for the NSS data. (In the case of Nepal, Gujarat 46 27 19
Indonesia, and Pakistan, the bivariate results produce exactly
Haryana 49 39 10
the same pattern as that from the multivariate specification.)
The first row of Table 6 shows the difference in the en- Karnataka 51 38 13
rollment rate of rural children age 6 to 14 between the rich- Kerala 12 3 9
est and the poorest quintiles, based on either the asset index Madhya Pradesh 55 33 22
(first column) or on consumption expenditures (second col- Maharashtra 34 25 9
umn) for the national sample. The enrollment of the poorest Orissa 47 38 10
asset quintile is 7 percentage points lower than for the poor-
est expenditure quintile (42 versus 49); the enrollment of the Punjab 56 52 5
richest asset quintile is 12 percentage points higher than for Rajasthan 52 41 11
the richest expenditure quintile (94 versus 82). As a result, Tamil Nadu 25 15 9
the wealth gap in enrollment rates between the richest and Uttar Pradesh 52 30 21
the poorest quintiles is only 33 percentage points when con- West Bengal 51 40 11
sumption expenditures are used; when we use the asset in-
dex, however, the gap is 52 percentage points, or 19 percent- Sources: Authors’ calculation from NFHS 1992–1993. Enrollment for
age points higher. consumption quintiles from Haque, Lanjouw, and Ravallion (1998).
The subsequent rows of Table 6 show that the wealth-
education profile is “flatter” (that is, the rich-poor differen-
tial is smaller) when rankings are based on consumption ex- plication of the method described in Ashenfelter and Krueger
penditures for every state of India. One possible interpreta- 1994.) We refer to these as “pseudo”-IV estimates because
tion of this regularity is attenuation bias due to greater “mea- we are not making the (strong) assumption that measurement
surement error” in consumption expenditures than in the as- errors are uncorrelated. We do not make such an assumption
set index. “Error” here is defined in relation to its use as a because some of the asset variables may be used both in the
proxy for the relevant indicator of economic status in the asset index and in estimating expenditures. The resulting in-
analysis of education outcomes. struments therefore are not “valid” in the sense of producing
We explore this issue of relative measurement error fur- consistent parameter estimates.
ther, using two multivariate regression approaches: Nevertheless, the ratio of OLS to pseudo-IV estimates is
“pseudo”-instrumental variables (IV) and reverse regres- an estimate of the relative signal to signal-plus-noise ratio
sion. for the two variables. Because the degree of inconsistency in
Pseudo-IV. Under the hypotheses that (1) expenditures the pseudo-IV estimates depends only on the measurement
and the asset index are both proxies for long-run economic error common to both measures, the pseudo-IV estimates for
status and (2) the measurement errors in each are not corre- expenditures and assets will converge in probability to the
lated perfectly, each proxy can be used as an instrument for same (inconsistent) estimate. In contrast, the degree of in-
the other to mitigate the attenuation bias due to measurement consistency in the two OLS estimates depends on both the
error. This is true even if the measurement errors in expendi- common measurement error and the error specific to either
tures and in assets are correlated and hence neither of the the asset index or consumption expenditures. Hence the ratio
resulting IV estimates are consistent. (This is a different ap- of the ratios of OLS to pseudo-IV for each measure is a valid
indicator of the relative degree of measurement error in the
two proxies. In the multivariate context, the inconsistency in
4. To maintain consistency with the rest of the paper, we enter the first
principal component into the model as dummy variables for each quintile. both the OLS and the pseudo-IV estimate will depend in ad-
We enter the higher-order principal components into the model simply as dition on the correlation with the other variables in the re-
linear indices as given by the factor scores from the analysis. gression. Because the measurement error is assumed to be
TABLE 7. PROBABILITY LIMITS USING TWO ALTERNATIVE NOISY MEASURES OF X

OLS Pseudo-IV Ratio of OLS to Pseudo-IV
X* × ( x2 / ( x2 + 2 + 2)) × ( x2 / ( x2 + 2)) (x2 + 2) / ( x2 + 2 + 2)
X+ × ( x2 / ( x2 + 2 + +2)) × ( x2 / ( x2 + 2) (x2 + 2) / ( x2 + 2 + +2)
Notes: y2 is the variance of variable y. Additional discussion is available from the authors.
TABLE 8. SCHOOL ENROLLMENT AS A FUNCTION OF THE ASSET INDEX OR CONSUMPTION EXPENDITURES: ALTERNA-
TIVE ESTIMATES OF RELATIVE MEASUREMENT ERROR OF EXPENDITURES VERSUS ASSET INDEX
Pseudo-IV Method
__________________________________________________
Reverse Regression Method
______________________________________________________________
OLS to IV Ratio: OLS to IV Ratio: Reverse to Direct Reverse to Direct
Asset Indexa Cons. Expend.b Ratio of Ratios Ratio: Asset Indexc Ratio: Cons. Expend.d Ratio of Ratios
Nepal 0.37 0.71 0.52 27.1 14.2 0.52
Indonesia 0.52 0.29 1.79 60.2 108.8 1.81
Pakistan 0.65 0.35 1.86 20.8 38.5 1.85
Notes: Where is the coefficient on the asset index (*) or consumption (+) in a regression of enrollment on the asset index or consumption (and other variables),
IV is the IV estimate of ; r is the coefficient on enrollment in the (reverse) regression of the asset index or consumption on enrollment (and other variables).
Further discussion and full regression results are available from the authors.
Sources: Authors’ calculations from IDHS 1994, NLSS 1996, and PIHS 1991.
a
(* / *IV)
b
(+ / +IV)
c
(r*–1 / *)
d
(r+–1 / +)
uncorrelated with those variables, only the correlation be- tively similar, and even more striking than those shown
tween the error-free underlying variable and other variables here.
in the regression will matter. Consequently an additional In Nepal, when we regress current enrollment on the as-
term will be introduced into the probability limits, but this set index using OLS and with consumption as an instrument,
will not qualitatively affect the analysis because it will be the ratio of the OLS to IV estimates is 0.37 (0.046/0.124).
the same for the two variables. When we regress enrollment on the consumption measure
In a simple case where there are two noisy measures of using OLS and with the asset index as an instrument, the ra-
a variable X, namely X* and X +, the common measurement tio of the OLS to IV estimates is 0.71 (0.211/0.298), yield-
error component is , and the measure-specific errors are * ing a ratio of the two ratios of 0.52 (0.37/0.71). In Nepal it
and +, we summarize the probability limits and their ratios appears that there is more measurement error in the asset in-
(see Table 7). The ratio of the OLS to pseudo-IV for the two dex than in consumption expenditures.
proxies will depend only on the relative measurement error. With data from Indonesia and Pakistan, the finding is
The discussion here revolves around probability limits, the opposite. When current enrollment is regressed on the
which are valid asymptotically but may not necessarily hold asset index, the ratio of the OLS to IV estimate is 0.52 in
in a given sample. Indonesia and 0.65 in Pakistan. When we regress enrollment
In the first two columns of Table 8 we report the ratios on the consumption measure using the asset index as an in-
of OLS to pseudo-IV estimates for Nepal, Indonesia, and strument, the ratio of the OLS to IV estimates is 0.29 in In-
Pakistan. (The full multivariate regressions are reported donesia and 0.35 in Pakistan.5 This yields respective ratios
in Appendix B, which is available from the authors.) To
keep the setup simple for this exploratory work, we use the 5. Behrman and Knowles (1999) find that their estimates of the elas-
index itself here rather than quintiles based on the index. ticities of various education outcome measures in Vietnam with respect to
We compare this with the log of consumption expenditures income, or consumption expenditures, increase by 50 to 60% when they use
household assets and other household characteristics as instruments for con-
per adjusted household size. In an alternative specification sumption. In the case of India we do not have expenditures, but we have 21
(not shown here) we used the level of consumption expendi- assets. We divided the asset variables into two groups and constructed an
tures per adjusted household size; the results were qualita- asset index from each set to form repeat measurements on long-run
of the ratios equaling 1.79 and 1.86. In these two countries TABLE 9. CLASSIFICATION DIFFERENCES USING ASSET
the results indicate much larger measurement error in con- INDEX DERIVED FROM TWO SAMPLES (WITH
sumption expenditures as a proxy in predicting enrollments OVERLAP) IN MOROCCO
than in the asset index. For instance, if the true signal-plus- Quintiles
measurement error common to both (which is unobservable) Based Quintiles Based on 1995 Ranking
is roughly equal to the idiosyncratic error in assets, then a ________________________________________
on 1992
ratio of ratios equaling about 1.8 would imply a measure- Ranking Poorest 2 3 4 Richest Total
ment error three times larger in consumption expenditures
Poorest 78.4 20.3 1.1 0.2 0.0 100.0
than in the asset index. (For this derivation, see Appendix A,
available from the authors.) 2 26.6 53.9 19.5 0.0 0.0 100.0
Reverse regression. An alternative approach to mea- 3 2.8 24.5 54.7 13.6 4.5 100.0
surement error is to use reverse regression: that is, to re- 4 0.0 2.5 15.7 58.4 23.4 100.0
gress enrollment on the wealth measure and estimate the co- Richest 0.0 0.0 3.1 32.6 64.2 99.9a
efficient on wealth (), and then to regress the wealth mea-
sure on enrollment and estimate the coefficient on Source: Authors’ calculations from Morocco DHS, 1992 and 1995.
enrollment (r). If enrollment and wealth are measured with a
Row does not sum to 100.0 because of rounding.
error, then, under certain conditions, the true regression pa-
rameter is bounded by and 1/ r (see, for example,
Maddala 1988). Using reverse regression in a multivariate
context, as we do, introduces several complications that are Stability of Household Rankings Over Time
absent in the bivariate case, but under reasonable conditions
they will not substantially affect our conclusions. (For fur- These results from Indonesia and Pakistan are consistent
ther discussion see Appendix A, which is available from the with an asset index that is less sensitive to transitory fluc-
authors.) tuations than are consumption expenditures. Therefore one
A comparison of the ratio of to 1/r, using the asset explanation of the “superior” performance of the asset index
index versus using expenditures, indicates the relative mea- is that household rankings based on the asset index are more
surement error in the two variables (as whatever measure- stable than those based on a consumption expenditure mea-
ment error in enrollment rates is the same for the two analy- sure.6
ses). In Nepal the reverse regression yields an estimate (1/ A panel survey of households in Morocco conducted in
r), which is 27.1 times higher than the direct regression 1992 and 1995 provides the basis for exploring this issue. A
when the asset index is used and 14.2 times higher when DHS survey conducted in 1992 covered 6,407 households; a
expenditures are used, yielding a ratio of 0.52 (Table 8). In 1995 survey covered 2,751, of which 2,489 can be matched
Indonesia the reverse regression estimate is 60.2 times across surveys. Table 9 presents the classification differences
higher for the asset index and 108.8 times higher for con- across the two periods for the subsample of overlapping
sumption; for Pakistan the numbers are 20.8 and 38.5, households. (Households are classified in each period ac-
yielding respective ratios of 1.81 and 1.85. For all three cording to their position with respect to the entire sample in
countries these figures are strikingly consistent with the that period, not merely the subsample that can be matched
pseudo-IV results. over time.) For example, 78.4% of the households that are
To make more precise statements about the relative classified as belonging to the poorest quintile in 1992 are
magnitude of measurement error, one would need to make also in the poorest quintile in 1995, and virtually none (1.3%)
more assumptions. If one is willing to assume that the mea- move out of the poorest 40%.
surement error is uncorrelated with the included control In a recent survey, Fields (1998) reports a similar
variables and that the true variability and noise for assets analysis of stability of classifications based on expenditures
are roughly equal (as in the illustration of pseudo-IV (or incomes) in a secondary analysis of results from four
above), then one can use the ratio of ratios to work out the countries (China, Peru, Malaysia, and Chile). In addition,
relationship between the measurement error variance in ex- Skoufias (1999) calculates similar numbers for Indonesia
penditures and in the asset index. Under these assumptions, between 1997 and 1998 based on a panel survey of about
if the ratio is 1.8, approximately the ratio in Indonesia and 12,000 households. Table 10 summarizes the results on
Pakistan, then the measurement error variance in consump- changes in household rankings from these studies. Particu-
tion expenditures is approximately 2.5 times larger than in
the asset index. 6. On the basis of a recent six-year panel of households in China, Jalan
and Ravallion (1998) find that annual consumption expenditures are highly
wealth. Although both of these will be imperfect proxies for long-run wealth, variable. In particular, they find that the average standard deviation of con-
the measurement errors will not be correlated perfectly; hence each can be sumption per person across households is 384 yuan (the mean is 342 yuan
used as an instrument for the other. In this case the ratio of OLS to IV esti- per person per year at 1985 prices over the period 1985–1990) and that the
mates, approximately one-half, is an estimate of the ratio of the “true” vari- mean of the intertemporal standard deviation for any given household, over
ance to the total variance. The wealth index appears to include a sizable the entire period, is 189 yuan. Thus the standard deviation of a household’s
measurement error component. measured expenditures over time is about half that in the cross-section.
TABLE 10. STABILITY OVER TIME IN RANKINGS: COMPARISON FROM PANEL DATA SETS
% Staying in Same Quintile

___________________________
Variable Used to Poorest Richest
Country Start Year End Year Difference Rank Households (Individuals) Quintile Quintile
Indonesia 1997 1998 1 Per capita household expenditures 54 51
Morocco 1992 1995 3 Household asset index 78 64
China (Rural) 1978 1983 5 Household income 54 61
Lima, Peru 1985 1990 5 Per capita household expenditures 40 50
China (Rural) 1983 1989 6 Household income 41 49
Malaysia 1967 1976 9 Income of males 55 62
Chile 1968 1986 18 Per capita household income 8 58
Sources: Adapted from Fields (1998); Skoufias (1999) for Indonesia; authors’ calculations for Morocco (DHS 1992, 1995).
larly for the poorest quintile, these findings clearly show (Indonesia and Pakistan), the results are consistent with
more variability over time for the classifications based on the asset index containing less measurement error than tra-
income or consumption expenditure than do the results for ditional consumption expenditures as a proxy for long-run
Morocco, which use the measure based on the asset index. economic status in predicting enrollment differences. In
This is true even for the consumption-based data from Indo- Nepal the results are ambiguous.
nesia, which trace changes over only one year. Although Ultimately the question in each case is empirical and de-
these results may not surprise some readers, they may not pends on many factors, such as the quality of underlying data
be obvious to others, who argue that consumption expendi- and the degree of expenditure variability. In cases where the
tures are smoothed relative to income. As Table 10 makes data contain both expenditures and assets, we would not use
clear, consumption expenditures vary substantially across these results to argue that an asset index is the most appro-
time (possibly because of measurement error, transitory priate variable to employ. In such cases, using assets as in-
shocks, or unexpected shocks to permanent income). struments for household per capita expenditures is most
likely the more effective way of extracting the maximum
Methodological Summary amount of information from the data while reducing the im-
The conventional wisdom is that survey-based household pact of measurement error. Yet the apparent success of this
consumption expenditures are not only the best estimates of simple asset index in addressing the problem at hand intro-
current expenditures but also the most reliable proxy for a duces the possibility of applying the DHS and NFHS data on
household’s long-run wealth. This view implies that surveys household wealth to a broader array of socioeconomic out-
lacking consumption expenditures have limited value becomes. Here we have limited ourselves to investigating and
cause they cannot control for, or estimate, wealth effects. discussing its validity and usefulness in the study of educa-
Current consumption expenditures are more firmly grounded tion enrollments in states of India.
theoretically and have a much wider range of uses (e.g., esti-
mation of demand functions, absolute poverty analysis, esti- MULTIVARIATE ANALYSIS OF WEALTH GAPS IN
mates of current welfare); yet there is no a priori argument EDUCATIONAL ENROLLMENT IN INDIAN STATES
explaining why current consumption expenditures are a more Armed with data on educational outcomes, on the one hand,
reliable proxy for long-run household economic status than and the asset index, on the other, we now examine how
an index of assets. Because the two have different sources of children’s school enrollments differ within Indian states ac-
potential “error,” this is an open empirical question. cording to the household’s economic status.
Our results suggest a methodologically simple solu- In India only 68% of children age 6 to 10 and 66% of
tion to the vexing problem of creating a weighted asset in- those age 11 to 14 are reported as being in school. Enroll-
dex. In the absence of data on expenditures, the straight- ment rates vary dramatically across the states: the percent-
forward solution of applying the principal-components age of 6- to 10-year-olds in school ranges from only 50% in
technique to a collection of asset indicators works quite Bihar to 96% in Kerala, and the percentage of those age 11
well. The asset index is not difficult to defend empirically; to 14 in school ranges from 54% in Bihar to 94% in Kerala
it appears to be an internally and externally coherent and and Mizoram (see Appendix Table A1).
stable measure. Although each of the methods we use to To disentangle the determinants of school enrollment,
assess measurement error has its limitations, together they we estimate probit regressions with the school enrollment of
tell a consistent story. In two out of three countries studied the ith child age 6 to 14 as the (latent) dependent variable:
Ei* = q = 2,5 q × Qiq + × Xi + k = 2,25 k × ik +

i. (4) The wealth effects for rural areas only, where we can in-
clude numerous additional village-level factors in the model,
We use a probit model because we observe only whether a
are very similar. This is important because the rural sample
child is in school (corresponding to Ei = 1 if Ei* 0) or not
includes information on school availability; therefore these
in school (corresponding to Ei = 0 if Ei* < 0). Wealth is speci-
effects represent the relationship, controlling for the fact that
fied by including the Qiqs; these are dummy variables equal
the poor are more likely to live in villages without schools.
to 1 if the child’s household is in quintile q. (The poorest
Even with these additional controls, the marginal effects asso-
group is the reference quintile.)
ciated with the wealth quintiles are nearly identical to those in
In all of the samples the variables included in addition
the all-India sample (11.1, 18.5, 26.9, and 31.5).
to wealth quintiles are the child variables, namely a dummy
We estimate the same regressions separately for India as
variable for gender and the child’s age and age squared, and
a whole (Table 11) and for each state. (The all-India regres-
the household variables: age of the head of the household,
sions include state dummy variables.) Table 12 presents the
whether the household head ever attended school, the high-
marginal effects of higher wealth on the probability that a
est grade completed by the household head, whether the
child age 6 to 14 is in school when the effects are estimated
household is Hindu, and whether the household belongs to a
in state-by-state regressions. The state-level models include
scheduled caste or tribe (Patrinos 1997). (If the information
all the same control variables as the national models; here,
on the education of the household head is missing, we set
however, we report only the wealth effects. In a separate pa-
the “head ever attended school” and “head’s highest grade”
per we delve more deeply into the interpretation of the other
variables equal to 0 and set an indicator dummy variable
variables in the Indian context, including an examination of
equal to 1 in the regression.) Finally, the specification in-
gender effects and the state-specific effects (Filmer and
cludes a set of state dummy variables ik equal to 1 if child i
Pritchett 1999b).
lives in state k.
Although the effects are large, on average, the states
The other variables used (in Xi) depend on the sample
vary substantially in the magnitude of the wealth effects.
because data on school availability and other village charac-
For example, a child from the highest quintile in Kerala is
teristics are limited to rural areas. In the pooled urban/rural
only 4.6 percentage points more likely to be in school than
samples, we include an urban dummy variable. In the rural
a child from the poorest quintile in that state, whereas in
samples, the variables include three dummy variables for the
Bihar the difference is 42.6 percentage points, nearly 10
presence of (1) a primary school, (2) a primary and a
times larger. The differences are exacerbated in rural areas:
“middle” school, and (3) a primary, “middle,” and secondary
the rich-poor difference is 4.2 percentage points in Kerala
school. In addition, we include a number of variables captur-
and 52.6 in Bihar. These enormous differences in the wealth
ing village infrastructure and “development” (e.g., post of-
gap, even within the same country, certainly deserve further
fice, bank, cinema).
analysis.
Table 11 reports the estimation of Eq. (4) for all of In-
The magnitudes of the wealth gaps found here are con-
dia with both urban and rural samples, and with only the
sistent with those from other studies, both in the all-India
rural samples. (Summary statistics are displayed in Appen-
findings and in the state-by-state results. For example,
dix Tables A2 and A3.) The first rows of Table 11 report
NCAER (1994) finds that among children age 6 to 14 who
the marginal effect, or (for a dummy variable) the change
had ever attended school, there was an average difference of
in a variable from 0 to 1, on the probability of enrollment
25 percentage points, over 14 major states, between children
when all of the other variables in the regression are set to
from households with per capita incomes of less than Rs
their sample means. The probability of enrollment rises
3,000 and children from households with per capita incomes
sharply with household wealth. All else being equal, a child
of more than Rs 10,000. The difference ranged from virtu-
from a household in the highest quintile is about 30.7 per-
ally none in Kerala to 55 percentage points in Punjab. Haque
centage points more likely to be in school than a child from
et al. (1998), in bivariate tabulations, find similar differences
the poorest quintile. Moreover, the effects are ordered
across the quintiles in the raw enrollment rates (see the dis-
strictly across the quintiles: belonging to the second quin-
cussion of Table 6).
tile is associated with a 10-percentage-point increase in the
Behrman and Knowles (1999) review estimates on the
probability of being in school, and each subsequent quintile
income elasticity of educational attainment from many dif-
is associated with an increase of roughly 7 percentage
ferent countries. These are not strictly comparable with our
points in the probability of enrollment (10.3 to 16.9 to 24.1
results because they report elasticities of attainment rather
to 30.7).7
than enrollment probabilities. Even so, the evidence in the
7. The results shown in Table 11 are robust to the inclusion of higher- gradients implied by the two specifications are also very similar. In all speci-
order principal components in the specification. For example, including in- fications, the children from the richest quintile are 30 percentage points
dices up to the fifth component changes the estimated percentage-point ef- more likely to be enrolled than those in the poorest quintile. As above, to
fect on quintiles to 8.4, 13.5, 20.5, and 29.4 in the all-India model and to maintain simplicity in this robustness check, we include higher-order com-
8.0, 12.5, 20.3, and 29.2 in the rural model. Although these point estimates ponents as indices, whereas the first principal component enters in “quin-
are slightly different, the largest difference is no more than 3.6 percentage tile” form. (The results are not shown here but are available from the au-
points in the all-India model and 6.6 points in the rural-only sample. The thors.)
TABLE 11. MARGINAL EFFECTS ON THE PROBABILITY OF BEING “IN SCHOOL,” AGES 6–14 (PROBIT REGRESSION RE-
SULTS)
All India (Urban and Rural)

_______________________________ Rural Only
_______________________________
Zero / One Variable Marginal Effect T-Ratio Marginal Effect T-Ratio
a
Quintile 2 * 0.103 12.32 0.111 9.87
Quintile 3 * 0.169 16.94 0.185 17.92
Quintile 4 * 0.241 22.55 0.269 20.77
Quintile 5 * 0.307 23.53 0.315 18.69
Male * 0.237 8.42
Rural Maleb * 0.070 3.85
Urban Female * –0.107 –6.19
Rural Female * –0.149 –6.70
Scheduled Caste/Tribe * –0.047 –3.87 –0.053 –4.37
Age 0.206 13.37 0.232 13.20
Age Squared –0.011 –16.89 –0.012 –16.47
Head Is Male * –0.092 –5.64 –0.119 –5.90
Head’s Age 0.001 4.29 0.002 5.41
Head Ever Attended School * 0.072 6.73 0.071 6.88
Head’s Highest Grade Completed 0.019 16.27 0.023 19.31
Head Information Missing * 0.094 4.42 0.112 4.75
Hindu * 0.109 5.11 0.119 5.38
Primary School in Village * 0.037 2.10
Primary and Middle School in Village * 0.073 3.05
Primary, Middle, and Secondary in Village * 0.083 6.43
Nearest Town Within 5 km * 0.018 1.31
Nearest Railroad Within 5 km * –0.001 –0.11
Nearest Bus Within 5 km * 0.014 1.71
Paved Road in Village * 0.006 0.42
Electricity in Village * 0.019 1.10
PHC Clinic in Village * –0.006 –0.27
Health Subcenter in Village * –0.011 –1.09
Hospital in Village * –0.015 –1.00
Dispensary in Village * 0.001 0.11
Health Guide in Village * 0.001 0.05
Bank in Village * 0.009 0.92
Co-op in Village * 0.007 0.55
Post Office in Village * –0.009 –0.60
Market in Village * –0.021 –2.95
Cinema in Village * 0.003 0.31
Pharmacy in Village * 0.016 1.15
Mahila Mandal (Women’s Group) in Village * –0.022 –1.01
Flood Within Last 2 Years * –0.003 –0.22
Drought in Last 2 Years * –0.007 –0.56
Notes: The marginal effect for a zero/one variable is the effect of a change in the variable from 0 to 1 on the probability of a child being in school, evaluated at
the means of the other variables. The specification includes dummy variables for each state. T-ratios refer to the underlying probit coefficient.
a
Reference group is Quintile 1 (poorest).
b
Reference group is urban male.
TABLE 12. MARGINAL EFFECTS OF WEALTH ON THE PROBABILITY OF BEING IN SCHOOL, AGES 6 TO 14, URBAN AND
RURAL: PROBIT REGRESSION RESULTS FOR SELECTED VARIABLES
Pooled Urban and Rural Samples

_____________________________________________________ Rural Sample Only
_____________________________________________________
Quint. 2 Quint. 3 Quint. 4 Quint. 5 Quint. 2 Quint. 3 Quint. 4 Quint. 5
a a a
Mizoram 0.030 0.073 0.112 0.083 –0.012 0.026 0.018 –0.096a
Himachal Pradesh –0.035a 0.031a 0.045a 0.062a –0.086a 0.005a 0.013a 0.026a
Kerala 0.017a 0.038 0.059 0.046 0.014a 0.037 0.058 0.042
Goa 0.019a 0.042 0.064 0.098 0.024a 0.038 0.063 0.054
Nagaland –0.004a 0.027a 0.017a 0.064 0.001a 0.037a 0.007a 0.065a
Manipur 0.032a 0.055 0.085 0.073 0.037 0.049 0.095 0.095
Jammu 0.039a 0.079 0.146 0.160 0.028a 0.066 0.118 0.119
Tamil Nadu 0.006a 0.061 0.106 0.143 –0.001a 0.078 0.119 0.142
Tripura 0.080 0.115 0.138 0.079a 0.066 0.136 0.137 0.155
Delhi 0.055a 0.072a 0.115 0.446 0.087 0.160
Maharashtra 0.048 0.084 0.124 0.199 0.049 0.093 0.163 0.164
Assam 0.131 0.202 0.212 0.133 0.139 0.212 0.187 0.172
Haryana 0.072 0.093 0.186 0.234 0.084a 0.107 0.229 0.196
Arunachal Pradesh 0.137 0.215 0.239 0.242 0.121 0.217 0.226 0.212
Orissa 0.082 0.206 0.231 0.263 0.095 0.229 0.250 0.251
Meghalaya 0.011a 0.081 0.188 0.197 0.011a 0.083a 0.209 0.257
Gujarat 0.057 0.106 0.179 0.294 0.066 0.145 0.210 0.273
West Bengal 0.152 0.242 0.290 0.271 0.124 0.226 0.287 0.284
Punjab 0.035a 0.104 0.207 0.336 0.022a 0.110 0.246 0.286
Karnataka 0.088 0.185 0.253 0.296 0.074 0.191 0.267 0.303
Madhya Pradesh 0.121 0.198 0.268 0.348 0.135 0.220 0.297 0.371
Uttar Pradesh 0.135 0.188 0.271 0.382 0.152 0.196 0.282 0.372
Andhra Pradesh 0.077 0.151 0.261 0.322 0.083 0.126 0.270 0.387
Rajasthan 0.082 0.158 0.296 0.388 0.065 0.180 0.339 0.406
Bihar 0.150 0.248 0.400 0.426 0.167 0.255 0.425 0.526
All India 0.103 0.169 0.241 0.307 0.111 0.185 0.269 0.315
Notes: States are sorted by the “wealth gap” as measured by the Quintile 5 coefficient in the rural sample. Marginal effects are evaluated at the means of the
other variables. In addition to the displayed variables, the probit regression includes age, age squared; gender, age, and schooling of the head of the household; and a
dummy variable for Hindu. The regression for the rural sample includes dummy variables for village infrastructure (for example, for the presence of a paved road, a
PHC clinic, a post office, and a market shop). The all-India regression includes dummy variables for state.
a
Not significant at .05 level. For all other coefficients, p < .05.
poorer countries is consistent with an income elasticity that Indonesia, and Pakistan are consistent with the presence of
would produce wealth gaps similar to those we estimate less measurement error in the asset index than in consump-
here. tion expenditures as a proxy for long-run wealth in predict-
ing educational outcomes.
CONCLUSION When the asset index is applied to the Indian data, the
In this paper we show that the relationship between wealth results show large school enrollment differences by wealth
and enrollment can be estimated without income or expendi- that vary widely across states of India. On average a rich In-
ture data, and largely without apologies or tears, by using dian child is 31 percentage points more likely to be enrolled
household asset variables. Principal-components analysis than a poor child, but the wealth gap varies from only 4.6
provides plausible and defensible weights for an index of as- percentage points in Kerala to 38.2 in Uttar Pradesh and 42.6
sets to serve as a proxy for wealth. In the four countries ex- in Bihar.
amined—India, Indonesia, Nepal, and Pakistan—this ap- Many research possibilities are suggested by the ability
proach produces reasonable results. The results from India, to generate a proxy for household wealth from DHS-like
data. For example, Filmer and Pritchett (1999a) and Filmer created, and the output from the principal-components pro-
(2000) explore how educational attainment profiles differ by cedure (factor scores for higher-order components, eigen-
wealth and gender in more than 35 countries with a recent values, and proportion of the variance captured by each
DHS. Similar country-specific (or comparative) analyses po- component); discussion and illustration of the correspon-
tentially could be conducted for a wide array of socioeco- dence between the OLS-to-IV ratios and differential mea-
nomic indicators included in the DHS/NFHS, such as health surement error; discussion of the complications arising from
outcomes (mortality, morbidity, immunization, utilization of the multivariate case of reverse regression and demonstra-
health facilities), fertility, and use of family planning. tion that under reasonable conditions they will not substan-
tially affect our conclusions; discussion of additional as-
APPENDIX. AVERAGE ENROLLMENT IN INDIAN sumptions required and the implications of relaxing those
STATES AND ALL-INDIA REGRESSIONS assumptions, in the derivation of relative magnitudes of
The following additional appendix material is available measurement error in reverse regression; and full multivari-
from the authors: description of a simple model of the struc- ate regressions results for OLS, pseudo-IV, and reverse re-
tural relationship between individual assets and the index gressions.
APPENDIX TABLE A1. EDUCATION STATUS, BY WEALTH GROUP
Proportion of 6- to 14-Year-Olds Proportion of 15- to 19-Year-Olds

Currently in School
__________________________________________________ Who Have Completed at Least Grade 8
_________________________________________________
Poorest Richest Wealth Gap Poorest Richest Wealth Gap
State All 40% 20% (Rich – Poor) All 40% 20% (Rich – Poor)
Kerala 0.949 0.887 0.975 0.088 0.749 0.531 0.923 0.392
Goa 0.937 0.774 0.973 0.200 0.703 0.344 0.848 0.504
Himachal Pradesh 0.908 0.724 0.970 0.246 0.565 0.233 0.818 0.585
Mizoram 0.907 0.768 0.974 0.205 0.567 0.190 0.844 0.654
Manipur 0.902 0.804 0.991 0.186 0.610 0.359 0.927 0.568
Nagaland 0.896 0.824 0.980 0.157 0.572 0.354 0.865 0.511
Delhi 0.872 0.477 0.924 0.448 0.685 N/A 0.766 N/A
Jammu 0.857 0.666 0.979 0.313 0.541 0.195 0.833 0.638
Tamil Nadu 0.825 0.717 0.950 0.232 0.518 0.269 0.838 0.570
Maharashtra 0.820 0.671 0.962 0.290 0.579 0.279 0.832 0.554
Haryana 0.813 0.605 0.957 0.352 0.480 0.189 0.728 0.539
Punjab 0.808 0.427 0.957 0.531 0.571 0.153 0.777 0.624
Tripura 0.795 0.710 0.873 0.163 0.395 0.187 0.789 0.603
Gujarat 0.757 0.552 0.962 0.410 0.504 0.212 0.845 0.633
Meghalaya 0.749 0.601 0.959 0.358 0.326 0.150 0.667 0.516
Karnataka 0.708 0.507 0.943 0.437 0.447 0.205 0.816 0.611
Assam 0.703 0.615 0.846 0.231 0.422 0.229 0.866 0.637
Orissa 0.697 0.552 0.969 0.416 0.395 0.189 0.908 0.719
West Bengal 0.678 0.527 0.902 0.375 0.338 0.137 0.734 0.597
Andhra Pradesh 0.639 0.457 0.917 0.460 0.419 0.160 0.859 0.698
Madhya Pradesh 0.626 0.461 0.937 0.476 0.367 0.172 0.832 0.661
Uttar Pradesh 0.614 0.484 0.939 0.455 0.424 0.239 0.836 0.598
Rajasthan 0.593 0.414 0.910 0.496 0.345 0.141 0.773 0.632
Bihar 0.514 0.378 0.942 0.564 0.381 0.183 0.864 0.681
All India 0.677 0.500 0.942 0.442 0.447 0.204 0.824 0.620
APPENDIX TABLE A2. ESTIMATES AND SUMMARY STATISTICS OF MULTIVARIATE MODELS FOR INDIA (SEE TABLE 11)
Urban and Rural

________________________________________________ Rural Only
_______________________________________________
Standard Standard
Coefficient T-Statistic Mean Deviation Coefficient T-Statistic Mean Deviation
1 = Quintile 2 0.335 12.32 0.187 0.390 0.322 9.87 0.239 0.426
1 = Quintile 3 0.588 16.94 0.203 0.402 0.559 17.92 0.241 0.428
1 = Quintile 4 0.930 22.55 0.218 0.413 0.912 20.77 0.204 0.403
1 = Quintile 5 (Richest) 1.385 23.53 0.211 0.408 1.385 18.69 0.077 0.267
1 = Male 0.663 8.42 0.520 0.500
1 = Rural Male 0.218 3.85 0.367 0.482
1 = Urban Female –0.307 –6.19 0.141 0.349
1 = Rural Female –0.442 –6.70 0.339 0.473
1 = Scheduled Caste/Tribe –0.140 –3.87 0.254 0.435 –0.144 –4.37 0.291 0.454
Age 0.629 13.37 9.841 2.553 0.642 13.20 9.797 2.548
Age Squared –0.033 –16.89 103.369 51.017 –0.034 –16.47 102.464 50.783
1 = Head is Male –0.282 –5.64 0.795 0.404 –0.330 –5.90 0.784 0.412
Head’s Age 0.004 4.29 36.386 16.793 0.006 5.41 36.005 17.331
1 = Head Ever Went to School 0.222 6.73 0.523 0.499 0.199 6.88 0.453 0.498
Highest Grade Head Completed 0.059 16.27 4.091 4.802 0.065 19.31 3.125 4.143
1 = Head’s Information Missing 0.312 4.43 0.137 0.343 0.331 4.75 0.148 0.355
1 = Hindu 0.315 5.11 0.761 0.426 0.317 5.38 0.781 0.414
Andhra Pradesh 0.204 27.15 0.041 0.198 0.198 4.00 0.042 0.201
Assam 0.599 31.37 0.039 0.194 0.706 25.27 0.038 0.192
Delhi 0.190 4.86 0.035 0.183 0.113 1.57 0.004 0.061
Goa 1.082 43.87 0.031 0.173 1.213 20.40 0.024 0.153
Gujarat 0.403 31.47 0.039 0.194 0.462 14.36 0.038 0.190
Himachal Pradesh 0.925 40.84 0.031 0.172 1.010 27.99 0.033 0.178
Haryana 0.462 25.83 0.034 0.182 0.488 11.03 0.033 0.178
Jammu 0.860 39.16 0.033 0.179 0.926 20.04 0.034 0.180
Karnataka 0.309 24.39 0.049 0.216 0.335 8.38 0.050 0.218
Kerala 1.405 41.29 0.040 0.195 1.478 16.39 0.043 0.203
Meghalaya 0.872 13.00 0.013 0.114 0.930 11.19 0.015 0.122
Maharashtra 0.661 43.42 0.043 0.202 0.735 16.52 0.038 0.190
Manipur 1.221 30.56 0.014 0.116 1.254 20.61 0.010 0.100
Madhya Pradesh 0.179 21.42 0.072 0.259 0.204 4.92 0.079 0.270
Mizoram 1.301 16.09 0.013 0.114 1.295 11.57 0.010 0.099
Nagaland 1.359 17.60 0.013 0.114 1.387 17.21 0.015 0.122
Orissa 0.473 83.46 0.047 0.211 0.493 16.27 0.049 0.217
Punjab 0.570 17.02 0.036 0.187 0.612 11.98 0.037 0.190
Rajasthan 0.057 5.52 0.068 0.251 0.130 2.99 0.069 0.253
Tamil Nadu 0.728 50.32 0.035 0.184 0.834 16.76 0.033 0.179
Tripura 0.734 59.16 0.013 0.113 0.762 20.10 0.014 0.118
Uttar Pradesh 0.166 29.01 0.136 0.343 0.243 5.22 0.155 0.362
West Bengal 0.417 29.52 0.046 0.210 0.517 11.49 0.046 0.210
Constant –3.494 –11.29 –4.245 –12.42
APPENDIX TABLE A3. ESTIMATES AND SUMMARY STATISTICS OF MULTIVARIATE MODELS FOR INDIA, RURAL AREAS
ONLY (SEE TABLE 11)
Coefficient T-Statistic Mean Standard Deviation
Primary School in Village 0.104 2.10 0.372 0.483
Primary and Middle School in Village 0.208 3.05 0.244 0.429
Primary, Middle, and Secondary School in Village 0.236 6.44 0.284 0.451
Nearest Town Within 5 km 0.050 1.31 0.194 0.395
Nearest Railroad Within 5 km –0.004 –0.11 0.196 0.397
Nearest Bus Within 5 km 0.037 1.71 0.663 0.473
Paved Road in Village 0.017 0.42 0.506 0.500
Electricity in Village 0.052 1.10 0.783 0.413
PHC Clinic in Village –0.016 –0.27 0.116 0.321
Health Subcenter in Village –0.032 –1.09 0.309 0.462
Hospital in Village –0.041 –1.01 0.146 0.353
Dispensary in Village 0.004 0.11 0.357 0.479
Health Guide in Village 0.002 0.05 0.401 0.490
Bank in Village 0.024 0.92 0.237 0.425
Co-op in Village 0.018 0.55 0.386 0.487
Post Office in Village –0.026 –0.60 0.459 0.498
Market in Village –0.059 –2.95 0.478 0.500
Cinema in Village 0.008 0.31 0.147 0.354
Pharmacy in Village 0.044 1.16 0.261 0.439
Mahila Mandal (Women’s Group) –0.062 –1.01 0.335 0.472
Flood in Last 2 Years –0.008 –0.22 0.152 0.359
Drought in Last 2 Years –0.020 –0.56 0.199 0.399
REFERENCES Fields, G.S. 1998. “Income Mobility: Meaning, Measurement, and

Some Evidence for the Developing World.” Department of In-
Agrawal, A.N., H.O. Varma, and R.C. Gupta. 1996. India Economic dustrial and Labor Relations, Cornell University. Photocopied
Information Year-Book. New Delhi: National. document.
Ashenfelter, O. and A. Krueger. 1994. “Estimates of the Economic Filmer, D. 2000. “The Structure of Social Disparities in Education:
Returns to Schooling From a New Sample of Twins.” American Gender and Wealth.” World Bank Policy Research Working Pa-
Economic Review 84:1157–73. per 2268, Development Economics Research Group, World
Behrman, J.R. and J.C. Knowles. 1999. “Household Income and Bank, Washington, DC.
Child Schooling in Vietnam.” World Bank Economic Review Filmer, D. and L. Pritchett. 1998. “Estimating Wealth Effects With-
13(2):211–56. out Expenditure Data—Or Tears: With an Application to Edu-
Bonilla-Chacin, M. and J.S. Hammer. 1999. “Life and Death cational Enrollments in States of India.” World Bank Policy Re-
Among the Poorest.” Johns Hopkins University and Develop- search Working Paper 1994, Development Economics Research
ment Economics Research Group, World Bank, Washington, Group, World Bank, Washington, DC.
DC. Photocopied document. ———. 1999a. “The Effect of Household Wealth on Educational
Deaton, A. 1997. The Analysis of Household Surveys. Washington, Attainment: Evidence From 35 Countries.” Population and De-
DC: Johns Hopkins University Press. velopment Review 25(1):85–120.
Deaton, A. and J. Muellbauer. 1980. Economics and Consumer Be- ———. 1999b. “Determinants of Education Enrollment in India:
havior. Cambridge, UK: Cambridge University Press. Child, Household, Village and State Effects.” Journal of Edu-
Deaton, A. and S. Zaidi. 1999. “Guidelines for Constructing Con- cational Planning and Administration 13(2):135–64.
sumption Aggregates for Welfare Analysis.” Working Paper 192, Grosh, M. and P. Glewwe. 1998. “The World Bank’s Living Stan-
Princeton University Research Program in Development Stud- dards Measurement Study Household Surveys.” Journal of Eco-
ies, Princeton, NJ. nomic Perspectives 12:187–96.
Dreze, J. and P.V. Srinivasan. 1997. “Widowhood and Poverty in Gwatkin, D.R., S. Rutstein, K. Johnson, R. Pande, and A. Wagstaff.
Rural India: Some Inferences From Household Survey Data.” 2000. “Socio-Economic Differences in Health, Nutrition, and
Journal of Development Economics 54:217–34. Population.” HNP/Poverty Thematic Group, World Bank, Wash-
ington, DC. Patrinos, H.A. 1997. “Differences in Education and Earnings

Haque, T., P. Lanjouw, and M. Ravallion. 1998. “A Poverty Profile Across Ethnic Groups in Guatemala.” Quarterly Review of Eco-
for India 1993–94.” Development Research Group, The World nomics and Finance 37:809–21.
Bank, Washington, DC. Photocopied document. Sahn, D. and D.C. Stifel. 1999. “Poverty Comparisons Over Time
Jalan, J. and M. Ravallion. 1998. “Transient Poverty in Postreform and Across Countries in Africa.” Cornell Food and Nutrition
Rural China.” Journal of Comparative Economics 26:338–57. Policy Program, Cornell University. Photocopied document.
———. 1995. “Poverty and Household Size.” Economic Journal Skoufias, E. 1999. “Krismon and Its Impact on Household Wel-
105(433):1415–34. fare: Preliminary Evidence From Household Panel Data From
Lindeman, R.H., P.F. Merenda, and R.Z. Gold. 1980. Introduction the 100 Village Survey in Indonesia.” International Food Policy
to Bivariate and Multivariate Analysis. New York: Scott, Research Institute, Washington, DC. Photocopied document
Foresman. StataCorp. 1999. Stata Statistical Software: Release 6.0. College
Maddala, G.S. 1988. Introduction to Econometrics. New York: Station, TX: Stata Corporation.
Macmillan. Stecklov, G., A. Bommier, and T. Boerma. 1999. “Trends in Equity
Montgomery, M.R., M. Gragnolati, K.A. Burke, and E. Paredes. in Child Survival in Developing Countries: A Illustrative Ex-
2000. “Measuring Living Standards With Proxy Variables.” De- ample Using Ugandan Data.” MEASURE Evaluation Project/
mography 27:155–74. Carolina Population Center, INED, Chapel Hill, NC. Photocop-
National Council of Applied Economic Research (NCAER). 1994. ied document.
“Non-Enrollment, Drop-Out, and Private Expenditures on El- World Bank. 1998. “Reducing Poverty in India: Options for More
ementary Education: A Comparison Across States and Popula- Effective Public Services.” Report 17881-IN, The World Bank,
tion Groups.” New Delhi. Photocopied document. Washington, DC.
larger volumes of data to be collected and pro-
REVIEW cessed than were accessible to previous gener-
ations of scholars. The second reason for this
time gap in using tax data is that most modern
Inequality in the long run research on inequality has focused on micro-
survey data that became available in the 1960s
and 1970s in many countries. Survey data, how-
Thomas Piketty1* and Emmanuel Saez2 ever, cannot measure top percentile incomes
accurately because of the small sample size and
This Review presents basic facts regarding the long-run evolution of income and wealth top coding. The top percentile plays a very large
inequality in Europe and the United States. Income and wealth inequality was very high a role in the evolution of inequality that we will
century ago, particularly in Europe, but dropped dramatically in the first half of the 20th discuss. Survey data also have a much shorter
century. Income inequality has surged back in the United States since the 1970s so that time span—typically a few decades—than tax
the United States is much more unequal than Europe today. We discuss possible data that often cover a century or more.
interpretations and lessons for the future. Kuznets-type methods to construct top in-
T
come shares were first extended and updated to
he distribution of income and wealth is a for the top decile and percentile of the U.S. the cases of France (8, 9), the United Kingdom
widely discussed and controversial topic. population. By dividing by national income, (10), and the United States (11). By combining
Do the dynamics of private capital ac- Kuznets obtained series of U.S. top income shares the efforts of an international team of over 30
cumulation inevitably lead to the con- for 1913 to 1948. scholars, similar series covering most of the
centration of income and wealth in ever In the 1960s and 1970s, similar methods 20th century were constructed for more than
fewer hands, as Karl Marx believed in the 19th using inheritance tax records were developed to 25 countries (12–15). The resulting “World Top
century? Or do the balancing forces of growth, construct top wealth shares (2, 3). Inheritance Incomes Database” (WTID) is the most ex-
competition, and technological progress lead declarations and probate records dating back tensive data set available on the historical
in later stages of development to reduced in- to the 18th and 19th centuries were also ex- evolution of income inequality. The series is
equality and greater harmony among the classes, ploited by a growing number of scholars in constantly being extended and updated and is
as Simon Kuznets thought in the available online (http://topincomes.
20th century? What do we know parisschoolofeconomics.eu/) as
about how income and wealth a research resource for further
have evolved since the 18th cen- Income inequality in Europe and the United States, analysis.
tury, and what lessons can we de- 1900–2010 Historical top wealth shares se-
rive from that knowledge for the Share of top income decile in total pretax income ries have also been constructed with
century now under way? For a long similar methods, albeit for a smaller
time, social science research on the number of countries so far, but with
distribution of income and wealth Top 10% income Top 10% income a longer time frame (16–21). Draw-
50 percent
was based on a relatively limited share: Europe share: U.S. ing on previous attempts to collect
set of firmly established facts to- historical national balance sheets
gether with a wide variety of pure- 45 (22), long-run series on the evolu-
ly theoretical speculations. In this tion of aggregate wealth-income
Review, we take stock of recent ratios in the eighth largest devel-
progress that has been made in 40 oped economies were established,
this area. We present a number some of them going back to the
of basic facts regarding the long- 18th century (23).
35
run evolution of income and wealth This Review draws extensively
inequality in advanced countries. on this body of historical research
We then discuss possible inter- 30 on income and wealth, as well as
pretations and lessons for the on a recently published interpre-
future. tive synthesis (24). We start by
25 presenting three basic facts that
Data and Methods 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 emerge from this research pro-
Modern data collection on the dis- gram (Figs. 1 to 3), and then turn
tribution of income begins in the to interpretations.
1950s with the work of Kuznets (1). Fig. 1. Income inequality in Europe and the United States, 1900 to 2010.
Shortly after having established The share of total income accruing to top decile income holders was higher in Three Facts About Inequality
the first national income time series Europe than in the United States from 1900 to 1910; it was substantially in the Long Run
for the United States, Kuznets set higher in the United States than in Europe from 2000 to 2010. The series We find large changes in the lev-
himself to construct time series of report decennial averages (1900 = 1900 to 1909, etc.) constructed using els of inequality, both over time
income distribution. He used tab- income tax returns and national accounts. See (24), chapter 9, Fig. 9.8. Series and across countries. This re-
ulated income data coming from available online at piketty.pse.ens.fr/capital21c. flects the fact that economic trends
income tax returns—available since are not acts of God, and that
the creation of the U.S. federal income tax in France, the United States, and the United King- country-specific institutions and historical cir-
1913—and statistical interpolation techniques based dom (4–7). cumstances can lead to very different inequality
upon Pareto laws (power laws) to estimate incomes Such data collection efforts on income and outcomes.
wealth dynamics have started to become more
1
Department of Economics, Paris School of Economics, Paris,
systematic and broader in scope and time only Income Inequality
France. 2Department of Economics, University of California
at Berkeley, Berkeley, CA, USA. since the 2000s. This is due first to the advent First, we find that whereas income inequality
*Corresponding author. E-mail: thomas.piketty@psemail.eu of information technologies, which allow much was larger in Europe than in the United States a
838 23 MAY 2014 • VOL 344 ISSUE 6186 sciencemag.org SCIENCE

century ago, it is currently much larger in the by the top 10% of wealth holders was notably was about 70 to 80% in from 1870 to 1910, fell to
United States. This is true for every inequality larger in Europe than in the United States one 60 to 70% from 1950 to 1980, and has been
metric. The simplest and most powerful measure, century ago, while the opposite is true today rising above 70% in recent decades. Naturally,
on which we focus in this article, is the share of (Fig. 2). this means that wealth concentration has been
total income going to the top decile (Fig. 1). There are important differences between in- high throughout U.S. history. But this also implies
On the eve of World War I (WWI), in the come and wealth inequality dynamics, how- that there has always been a large fraction of U.S.
early 1910s, the top decile income share was ever. First, we stress that wealth concentration aggregate wealth—about 20 to 30% —that did not
between 45 and 50% of total income in most is always much higher than income concen- belong to the top 10%. As the bottom 50% wealth
European countries. This applies in particular tration. The top decile wealth share typically share has always been negligible, this remaining
to the United Kingdom, France, Germany, and falls in the 60 to 90% range, whereas the top 20 to 30% fraction corresponds to the share
Sweden, which are the four countries that we decile income share is in the 30 to 50% range. owned by the “middle 40%” (i.e., the intermediate
use to compute the European average series Even more striking, the bottom 50% wealth group between the bottom 50% and the top 10%),
reported in this article. At the same time, the share is always less than 5%, whereas the bot- a social group that one might want to call the
top decile income share was slightly above 40% tom 50% income share generally falls in the 20 “wealth middle class.” The important point is
in the United States. to 30% range. The bottom half of the popula- that, to a large extent, there has always been a
One century later, in the early 2010s, the in- tion hardly owns any wealth, but it does earn wealth middle class in the United States.
equality ordering between Europe and the United appreciable income: On average, members of In contrast, wealth concentration was so ex-
States is reversed. In Europe, the top decile in- the bottom half of the population (wealth-wise) treme in pre-WWI Europe that there was ba-
come share fell sharply, from 45 to 50% to about own less than one-tenth of the average wealth, sically no wealth middle class. That is, the top
30%, between 1914 and the 1950s–1960s. It has while members of the bottom half of the pop- decile wealth share was close to 90% (or even
been rising somewhat since the 1970s–1980s, and ulation (income-wise) earn about half the aver- somewhat higher than 90%, as in the UK), so
it is now close to 35% (somewhat less in con- age income. that the middle 40% wealth holders were almost
tinental Europe and somewhat more in the United In sum, the concentration of capital ownership as poor as the bottom 50% wealth holders (the
Kingdom, which has experienced an evolution is always extreme, so that the very notion of wealth share of both groups was close to or less
closer to that of the United States). That is, the capital is fairly abstract for large segments—if than 5%). Between 1914 and the 1950s–1960s, the
top decile share in Europe is currently almost not the majority—of the population. The inequal- top decile wealth share fell dramatically in Eu-
one-third smaller than what it used to be one ity of labor income can be high, but it is usually rope, from about 90% to less than 60%. It has
century ago. The secular decline in inequality much less extreme. It is also less controversial, been rising since the 1970s–80s, and is now close
would be even larger if we took into account the partly because it is viewed as more merit-based. to 65% (somewhat more in the United Kingdom,
rise of taxes and transfers, and measure instead Whether this is justified is a highly complex and and somewhat less in Continental Europe). In
income after taxes and transfers. Total tax rev- debated issue to which we later return. other words, the wealth middle class now com-
enues and public spending were less than 10% Next, in contrast to income inequality, U.S. mands a larger share of total wealth in Europe
of national income in every country before WWI, wealth inequality levels have still not regained than in the United States—although this share
and they are now on the order of 30 to 50% of the record levels observed in Europe before has been shrinking lately on both sides of the
national income in every developed country. Prop- World War I. The U.S. top decile wealth share Atlantic.
erly attributing taxes, transfers, and Given that wealth inequality is
public spending to each income dec- lower in the United States today
ile raises important measurement than in 1913 Europe, why is U.S.
issues, however, particularly regard- Wealth inequality in Europe and the United States, income inequality now as large as
ing in-kind transfers (such as health, 1870–2010 (or even slightly larger than) that
education, or public good spend- in 1913 Europe? The reason is that
Share of top wealth decile in total net wealth
ing). In this Review, we therefore modern U.S. inequality is based
focus on the long-run evolution of more on a very large rise of top
the inequality of primary income 100 percent Top 10% wealth labor incomes than upon the ex-
(pretax, pretransfer). share: Europe treme levels of wealth concentration
In the United States, the top dec- 90 that characterized the “patrimonial”
ile income share in 1910 was lower (wealth-based) societies of the past.
than in Europe, then rose in the Top 10% wealth In 1913 Europe, top incomes were
1920s, fell in the 1930s–1940s, and 80 share: U.S. predominantly top capital incomes
stabilized around 30 to 35% in the (rent, interest, and dividends) com-
1950s–1960s, slightly above Euro- ing from the very large concen-
70
pean levels of the time. It then rose tration of capital ownership. Top
at an unprecedented pace since the U.S. incomes today are composed
1970s–1980s, and is now close to 60 about equally of labor income and
50%. According to this measure, pri- capital income. This generates ap-
mary income concentration is cur- proximately the same level of total
50
rently higher than it has ever been income inequality, but it is not the
1870 1890 1910 1930 1950 1970 1990 2010
in U.S. history. It is also slightly same form of inequality.
higher than in pre-WWI Europe.
Wealth-to-Income Ratios
Wealth Inequality Fig. 2. Wealth inequality in Europe and the United States, 1870 to 2010. Before further discussing the dif-
Second, we observe the same “great The share of total net wealth belonging to top decile wealth holders became ferent possible interpretations for
inequality reversal” between Europe higher in the United States than in Europe over the course of the 20th these important transformations,
and the United States when we century. But it is still smaller than what it was in Europe before World War I. we introduce a third basic fact: If
look at wealth inequality rather than The series report decennial averages constructed using inheritance tax we look at the evolution of the ag-
income inequality. That is, the share returns and national accounts. See (24), chapter 10, Fig. 10.6. Series available gregate value of wealth relative to
of total net private wealth owned online at piketty.pse.ens.fr/capital21c. income, we also find large historical
SCIENCE sciencemag.org 23 MAY 2014 • VOL 344 ISSUE 6186 839

variations, again with striking differences be- experiments: We cannot replay the 20th-century simplest way to understand why capital accu-
tween Europe and the United States (Fig. 3). income and wealth dynamics as if the world wars, mulation is a slow process is to consider the
This ratio is of critical importance for the anal- the rise of progressive taxation, or the Bolshevik following elementary arithmetic: With a saving
ysis of inequality, as it measures the overall im- revolution did not happen. Still, we can try to rate of 10% per year, it takes 50 years to accu-
portance of wealth in a given society, as well as make some progress. mulate the equivalent of 5 years of income.
the capital intensity of production. How is the long-run equilibrium wealth-income
In every European country for which we have Wealth-to-Income Ratios ratio determined, and why does it seem to vary across
data, and in particular France, the United King- The relatively easier part of the story is the long- countries and over time? A simple yet powerful
dom, and Germany, the aggregate wealth-income run evolution of aggregate wealth-to-income ra- way to think about this issue is the so-called
ratio has followed a pronounced U-shaped pat- tios (Fig. 3). The fall of European wealth-income Harrod-Domar-Solow formula (23). In the long-
tern over the past century. On the eve of WWI, ratios following the 1914–1945 capital shocks can run, assuming no systematic divergence between
net private wealth was about equal to 6 to 7 years be well accounted for by three main factors: direct the relative price of capital assets and consumption
of national income in Europe. It then fell to war-related physical destruction of domestic capital goods, one can show that the wealth-to-income
about 2 to 3 years of national income in the assets (real estate, factories, machinery, equipment); (or capital-to-income) ratio bt = Kt/Yt converges
1950s. It has risen regularly since then, and it is lack of investment (a large fraction of 1914–1945 toward b = s/g, where s is the long-run annual
now back to about 5 to 6 years of national in- private-saving flows was absorbed by the enor- saving rate and g is the long-run annual total
come. Interestingly, we also find a similar pat- mous public deficits induced by war financing; growth rate. The growth rate g is the sum of the
tern for Japan (23). there was also massive dissaving in some cases, population growth rate (including immigration)
In contrast, the U.S. pattern is flatter: Net pri- e.g., foreign assets were sold to purchase gov- and the productivity growth rate (real income
vate wealth has generally equalled about 4 to ernment bonds; the resulting public debt was growth rate per person). This formula holds
5 years of national income in the United States, eventually wiped away by inflation); and a fall in whether savings are invested in domestic or
with much less variation than in Europe or Japan. relative asset prices (real estate and stock mar- foreign assets (it also holds at the global level).
The U.S. pattern is also slightly U-shaped—with ket prices were both historically very low in the That is, with a saving rate s = 10% and a growth
aggregate wealth-income ratios standing at a immediate postwar period, partly due to rent rate g = 3%, then b ≈ 300%. But if the growth
relatively lower level in the mid-20th century control, nationalization, capital controls, and rate drops to g = 1.5%, then b ≈ 600%. In short:
than at both ends of the century. But it is clearly various forms of financial repression policies). Capital is back because low growth is back.
much less marked than in Europe. In France and Germany, each of these three Intuitively, in a low-growth society, the to-
The comparison between Figs. 1 and 3 is factors seems to account for about one-third of tal stock of capital accumulated in the past
particularly striking. Both figures have two the total decrease in wealth-income ratios. In can become very important. In the extreme
U-shaped curves, but these are clearly differ- the United Kingdom, where domestic capital case of a society with zero population and pro-
ent. The United States displays a U-shaped destruction was of limited importance (less than ductivity growth, income Y is fixed. As long as
pattern for income inequality (mostly driven 10% of the total), the other two factors each there is a positive net saving rate s > 0, the
by the large rise of top labor incomes in recent account for about half of the decline in the ag- quantity of accumulated capital K will go to
decades). Europe (and Japan) shows a U-shaped gregate wealth-income ratio (23, 24). infinity. Therefore, the wealth-income ratio
pattern for aggregate wealth-income ratios. Why did the postwar recovery of European b = K/Y would rise indefinitely (at some point,
The United States is the land of booming top wealth-income ratios take so much time? The people in such a society would probably
labor incomes; Europe is the land stop saving, as additional capi-
of booming wealth (albeit with a tal units become almost useless).
lower wealth concentration than With positive but small growth,
in the United States). These are Wealth-to-income ratios in Europe and the the process is not as extreme:
two distinct phenomena, involv- United States, 1900–2010 The rise of b stops at some finite
ing different economic mechanisms level. But this finite level can be
Market value of net private wealth (% national income)
and different parts of the devel- very high.
oped world. One can show that this simple
700 percent logic can account relatively well
Interpreting the for why the United States accumu-
Long-Run Evidence Europe
lates structurally less capital relative
600
We now turn to the discussion of to its annual income than Europe
possible interpretations and les- U.S. and Japan. U.S. population growth
sons for the future. We stress at the 500 rates exceed 1% per year, thanks
outset that what we have to offer is to large immigration flows, so total
little more than an informed dis- U.S. growth rates—including pro-
400
cussion. Although we have at our ductivity growth of around 1 to
disposal much more extensive his- 1.5%—are at least 2 to 2.5% per
torical and comparative data than 300 year, if not 2.5 to 3%. By contrast,
were available to previous research- population growth in Europe and
ers, existing evidence is still far too Japan is now close to zero, so that
200
incomplete and imperfect for a rig- 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 total growth is close to produc-
orous quantitative assessment of tivity growth, i.e., about 1 to 1.5%
the various causes at play. Several per year. This is further reinforced
different mechanisms have clear- Fig. 3. Wealth-to-income ratios in Europe and the United States, 1900 to by the fact that U.S. saving rates
ly played an important role in the 2010. Total net private wealth was worth about 6 to 7 years of national income tend to be lower than in Europe
evolution of income and wealth in Europe before World War I, fell to 2 to 3 years in 1950–1960, and increased and Japan. To the extent that pop-
depicted in Figs. 1 to 3, but it is back to 5 to 6 years in 2000–2010. In the United States, the U-shape pattern ulation growth will eventually de-
extremely difficult to disentangle was much less marked. The series report decennial averages (1900 = 1900 to cline almost everywhere, and that
the individual processes. We are 1909, etc.) constructed using national accounts. See (24), chapter 5, Fig. 5.1. saving rates will stabilize, this al-
not in the domain of controlled Series available online at piketty.pse.ens.fr/capital21c. so implies that the return of high

capital-to-income ratios will apply at the global concentration, with a top decile wealth share of supplemented by high labor income); and so on.
level in the very long run (23, 24). around 90%. Importantly, for a given structure of shocks,
The share of capital income in national in- How can we account for the very high level of the variance of the multiplicative term wit is an
come is defined as a = rK/Y = rb, where r is the wealth concentration that we observe in histor- increasing function of r – g, where r is the (net-of-
average annual real rate of return on wealth. ical series, and what does this tell us about the tax) rate of return and g is the economy’s growth
For instance, if r = 5% and b = K/Y = 600%, then future? The most powerful model to analyze struc- rate. Intuitively, a higher r – g tends to amplify
a = 30%. Whether the rise in the capital income tural changes in wealth inequality is a dynamic initial wealth inequalities: It implies that past
ratio b will also lead to a rise in a is a compli- model with multiplicative random shocks. That is, wealth is capitalized at a faster pace, and that it is
cated issue. assume that the individual-level wealth process less likely to be overtaken by the general growth
In the standard economic model with per- has the following general form: zit+1 = wit zit + eit , of the economy. Under fairly general conditions,
fectly competitive markets, r is equal to the where zit is the position of individual i in the one can show that the top tail of the distribution
marginal product of capital (that is, the addi- wealth distribution prevailing at time t (i.e., zit = of wealth converges toward a Pareto distribution,
tional output produced by one additional cap- kit/kt where kit is net wealth owned by individ- and that the inverted Pareto coefficient (measur-
ital unit, all other things being equal). As the ual i at time t, and kt = average net wealth of ing the thickness of the upper tail and hence the
volume of capital b rises, the marginal product the entire population at time t), wit is a multi- inequality of the distribution) increases with r – g
r tends to decline. The important question is plicative random shock, and eit is an additive (3, 14, 24–26).
whether r falls more or less rapidly than the random shock. The dynamic wealth accumulation model with
rise in b. This depends on what economists de- The shocks wit and eit can be interpreted as multiplicative shocks can explain the extreme
fine as the elasticity of substitution s between reflecting different types of events that often levels of wealth concentration that we observe in
capital and labor in the production function occur in individual wealth histories, including the data much better than alternative models. In
Y = F(K, L). shocks to rates of return (some individuals may particular, if wealth accumulation were predomi-
A standard hypothesis in economics has been get returns that are far above average returns; nantly driven by lifecycle or precautionary mo-
to assume a unitary elasticity, in which case the investment strategies may fail and lead to fam- tives, then wealth inequality would not be as large
fall in r exactly offsets the rise in b, so that the ily bankruptcy); shocks to demographic param- as what we observe (it would be comparable in
capital share a = rb is a technological constant. eters (some families have many children; some magnitude to income inequality, or even lower).
However, historical variations in capital shares individuals die young); shocks to preferences The dynamic multiplicative model can also
are far from negligible: a typically varies in the parameters (some individuals like to save, some help to explain some of the important historical
20 to 40% range (and the labor share 1 – a in the prefer to consume their wealth); shocks to pro- variations that we observe in wealth concentra-
60 to 80% range). In recent decades, rich coun- ductivity parameters (capital income is sometimes tion series.
tries have experienced both a rise in b and a rise
in a, which suggests that s is somewhat larger
than 1. Intuitively, it makes sense to assume that
s tends to rise over the development process, as Rate of return vs. growth rate at the world level, Antiquity–2100
there are more diverse uses and forms for capital Annual rate of return or rate of growth
and more possibilities to substitute capital for
labor (e.g., replacing delivery workers by drones
6 percent Pure rate of return
or self-driving trucks). to capital r (after
Whether the capital share a will keep rising tax and capital
in future decades is an open question. It de- 5 losses)
pends both on technological forces and on the
bargaining power of capital and labor and the 4–5
collective institutions regulating the capital- 4
labor relationship (the simple economic mod-
el with perfectly competitive markets is likely
excessively naïve). But from a logical standpoint, 3
this is a plausible possibility, especially if the
population and productivity growth slowdown Growth rate of
2 world output g
pushes the global capital income ratio b toward
1.5
higher levels.
1
Wealth Inequality: r > g
We now move to an even more complicated—and
arguably more important—issue: the long-run dy- 0
namics of wealth inequality (Fig. 2). High capital 0-1000 1000-1500 1500-1700 1700-1820 1820-1913 1913-1950 1950-2012 2012-2050 2050-2100
intensity, as measured by high b and a, is not bad
in itself. After all, it would be good to have an
infinite quantity of robots producing most of the Fig. 4. Rate of return versus growth rate at the global level, from Antiquity until 2100. The
output, so that we can devote more time to leisure average rate of return to capital (after tax and capital losses) fell below the growth rate in the 20th
activities. The problem is twofold: Can we all find century. It may again surpass it in the 21st century, as it did throughout human history except in the
jobs as a robot designer (or in leisure-related ac- 20th century. The series was constructed using national accounts for 1700 and after and historical
tivities), and who owns the robots? In practice, sources on growth and rent to land values for the period before 1700. See (24), chapter 10, Fig. 10.10.
the concentration of capital ownership always Series available online at piketty.pse.ens.fr/capital21c. The future values for g are based upon UN
seems to be very high—much more than the con- demographic projections (median scenario) for population growth and on the assumption that
centration of labor income (Figs. 1 and 2). The between-country convergence in productivity growth rates will continue at its current pace. The
“patrimonial” (wealth-based) societies of Europe future values for r are simply based upon the continuation of current pretax values and the assump-
one century ago were characterized not only by tion that tax competition will continue. See (24), chapter 10, Fig. 10.10. Series available online at
very high b and a, but also by extreme capital piketty.pse.ens.fr/capital21c.

In particular, it is critical to realize that r – g which could lead to a structural rise in wealth equality declined in developed countries during
was very large during most of human history concentration. the first half of the 20th century. The compression
(Fig. 4). Growth was very low until the industrial This model seems to capture relatively well of incomes occurred primarily because of the
revolution (much less than 1% per year), whereas some of the evolutions that we are currently fall of top capital incomes induced by the world
average rates of return were typically on the order observing at the global level. For instance, if we wars, the Great Depression, and the regulatory
of 4 to 5% per year (historically, in preindus- use the global billionaires rankings published and fiscal policies developed in response to these
trial agrarian societies, annual rent on land, by Forbes magazine since 1987, we find that the shocks. In particular, there was no structural de-
the main capital asset, was about 4 to 5% of the very top fractiles of the global wealth distribu- cline in the inequality of labor income (8–13, 24).
land value) and taxes were minimal. Growth tion have been rising on average at about 6 to Kuznets’ overly optimistic theory of a natural
rates rose substantially during the 18th and 7% per year in real terms over the 1987–2013 pe- decline in income inequality in market econo-
19th centuries, but they remained relatively riod, i.e., more than three times as fast as average mies largely owed its popularity to the Cold
small (1 to 1.5%) compared to rates of return. global wealth (about 2% per year over the same War context of the 1950s as a weapon in the ideo-
This large gap between r and g explains why period) (24). logical fight between the market economy and
wealth concentration was so large until World We stress, however, that our ability to prop- socialism (24).
War I and why wealth concentration was smaller erly measure and monitor the dynamics of the What are the main forces that determine the
in the United States, where population growth global distribution of wealth is far from being level of labor income inequality in the long-run?
was faster. satisfactory. National statistical institutes as The most widely used economic model is based
During the 20th century, growth rates were well as international organizations are facing on the idea of a race between education and
exceptionally high (in particular due to very major difficulties in tracking down cross-border technology (30). That is, the expansion of educa-
high population growth, which even today rep- wealth, and magazines are ill-equipped to produce tion leads to a rise in the supply of skills, while
resents about half of global gross domestic rigorous statistics. Despite some recent progress technological change leads to a rise in the de-
product growth), and rates of return were se- in this area (28), our ability to measure global mand for skills. Depending on which process
verely reduced by capital shocks (destructions) wealth is also severely limited by the rise of tax occurs faster, the inequality of labor income will
and the rise of taxation. Simple simulations havens (29). either fall or rise.
show that this effect is quantitatively sufficiently One proposed explanation for the increase
important to explain why wealth concentration The Dynamics of Income Inequality of inequality in recent decades has been the
did not return to pre-WWI levels in the postwar We finally return to the most difficult and un- rise in the global competition for skills, itself
period. certain part: the long-run dynamics of income driven by globalization, skill-biased technical
Other factors might also have played a role. For inequality (Fig. 1). This is the most difficult part change and the rise of information technologies.
instance, the rise of the wealth middle class might because income inequality combines forces aris- Such skill-biased technological progress is not
partly come from the fact that the growth of in- ing from the inequality of capital ownership sufficient to explain important variations between
comes and living standards eventually induced the and capital income (which, as we have just countries: The rise of labor income inequality was
rise of middle class saving. However, this process seen, are relatively complex) and forces re- relatively limited in Europe (and Japan) com-
does not seem to have taken place in pre-WWI lated to the inequality of labor income (which pared to the United States, despite similar tech-
Europe, because of the powerful unequalizing involve a different set of economic and social nological changes. In the very long run, European
impact of the r – g factor (17, 21, 24, 27). processes). labor income inequality appears to be relatively
To the extent that population growth (and Kuznets posited that income inequality first rises stable (there is no major downward or upward
possibly productivity growth) will slow down in with economic development when new, higher- trend in the wage shares received by the various
the 21st century, and that after-tax rates of re- productivity sectors emerge (e.g., manufacturing deciles and percentiles of the wage distribution).
turn to capital will rise (due to rising interna- industry during the industrial revolution) but then This suggests that the supply and demand for
tional tax competition to attract capital, and decreases as more and more workers join the skills have increased approximately at the same
maybe also to changing technology), it is likely high-paying sectors of the economy. Our data pace in Europe
that r – g will increase again in the 21st century, show that this is not the reason that income in- Could the particularly large increase in U.S.
labor income inequality in recent decades be ex-
plained by insufficient educational investment
for large segments of the U.S. labor force? In that
Box 1. Income and wealth: definitions case, massive investment in higher education
would be the right policy to curb rising income
Income is a flow. It corresponds to the quantity of goods and services produced and dis- inequality (30). Although this view is very ap-
tributed each year. Income can be decomposed as the sum of labor income (wages, salaries, pealing, it cannot account for all of the facts. In
bonuses, earnings from nonwage labor, and other remuneration for labor services) and capital particular, the race between education and tech-
income (rent, dividends, interest, business profits, capital gains, royalties, and other income nology fails to explain the unprecedented rise of
derived from owning capital assets). In this Review, we focus on the long-run evolution of the very top labor incomes that has occurred in the
inequality of primary income, defined as income before taxes and government transfers. In United States over the past few decades. A very
contrast, disposable income is defined as income after taxes and government transfers. large part of the rise in the top 10% income share
Although we do not analyze disposable income in this article, comparing inequality of primary comes from the top 1% (or even the top 0.1%).
income and inequality of disposable income is useful to assess the role of the government in This is largely due to the rise of top executive com-
reducing income inequality. pensation in large U.S. corporations (both fi-
Wealth (or capital) is a stock. It corresponds to the total wealth owned at a given point in nancial and nonfinancial). We discuss in the
time. This stock comes from the wealth appropriated or accumulated in the past. In the context supplementary online material how changes
of this article, wealth is defined as nonhuman net worth, i.e., the sum of nonfinancial and in tax policy, as well as social norms regarding
financial assets, net of financial liabilities (debt). National wealth is the sum of private wealth pay equality, likely play a key role in shaping
(net worth owned by private individuals) and public wealth (net worth owned by the gov- labor income inequality.
ernment and other public agencies). In this article, we focus on the level and distribution of To summarize: Inequality does not follow
private wealth. More details on these definitions, concepts, and corresponding series are pro- a deterministic process. In a sense, both Marx
vided in (23, 24). and Kuznets were wrong. There are power-
ful forces pushing alternately in the direction

of rising or shrinking inequality. Which one
dominates depends on the institutions and pol-
icies that societies choose to adopt. REVIEW
RE FE RENCES
1. S. Kuznets, Shares of Upper Income Groups in Income and
Skills, education, and the rise of
Savings (National Bureau of Economic Research, Cambridge,
MA, 1953).
2. R. J. Lampman, The Share of Top Wealth holders in National
earnings inequality among
the “other 99 percent”
Wealth, 1922-1956 (Princeton Univ. Press, Princeton, NJ,
1962).
3. A. B. Atkinson, A. J. Harrison, Distribution of Personal Wealth
in Britain, 1923-1972 (Cambridge Univ. Press, Cambridge,
1978). David H. Autor
4. A. Daumard, Les fortunes françaises au 19e siècle.
Enquête sur la répartition et la composition des capitaux
privés à Paris, Lyon, Lille, Bordeaux et Toulouse d'après The singular focus of public debate on the “top 1 percent” of households overlooks the
l'enregistrement des déclarations de successions (Mouton, component of earnings inequality that is arguably most consequential for the “other
Paris, 1973). 99 percent” of citizens: the dramatic growth in the wage premium associated with higher
5. A. H. Jones, American Colonial Wealth: Documents and
Methods (Arno Press, New York, 1977).
education and cognitive ability. This Review documents the central role of both the supply
6. P. Lindert, J. Polit. Econ. 94, 1127–1162 (1986). and demand for skills in shaping inequality, discusses why skill demands have persistently
7. L. Soltow, Distribution of Wealth and Income in the United risen in industrialized countries, and considers the economic value of inequality alongside
States in 1798 (Univ. of Pittsburgh Press, Pittsburgh, PA, its potential social costs. I conclude by highlighting the constructive role for public policy in
1989).
8. T. Piketty, Les hauts revenus en France au 20e
fostering skills formation and preserving economic mobility.
P
siècle—Inégalités et redistributions, 1901–1998
(Grasset, Paris, 2001). ublic debate has recently focused on a gap between college and high school graduates
9. T. Piketty, J. Polit. Econ. 111, 1004–1042 (2003). subject that economists have been ana- has more than doubled in the United States over
10. A. B. Atkinson, J. R. Stat. Soc. Ser. A Stat. Soc. 168, 325–343
(2005).
lyzing for at least two decades: the steep, the past three decades. A third reason for focus-
11. T. Piketty, E. Saez, Q. J. Econ. 118, 1–41(2003). persistent rise of earnings inequality in ing on the skill premium is that it offers broad
12. A. B. Atkinson, T. Piketty, Eds., Top Incomes over the the U.S. labor market and in developed insight into the evolution of inequality within a
20th Century—A Contrast Between Continental European countries more broadly. Much popular dis- market economy, highlighting the social value of
and English Speaking Countries (Oxford Univ. Press, New York,
2007).
cussion of inequality concerns the “top 1 percent,” inequality alongside its potential social costs and
13. A. B. Atkinson, T. Piketty, Eds., Top Incomes—A Global referring to the increasing share of national in- illuminating the constructive role for public policy
Perspective (Oxford Univ. Press, New York, 2010). come accruing to the top percentile of house- in maximizing the benefits and minimizing the
14. A. B. Atkinson, T. Piketty, E. Saez, J. Econ. Lit. 49, 3–71 holds. Although this phenomenon is undeniably costs of inequality.
(2011).
15. F. Alvaredo, A. B. Atkinson, T. Piketty, E. Saez, J. Econ.
important, an exclusive focus on the concen- The rising skill premium is not, of course, the
Perspect. 27, 3–21 (2013). tration of top incomes ignores the component sole cause of growing inequality. The decades-
16. W. Kopczuk, E. Saez, Natl. Tax J. 57, 445–487 (2004). of rising inequality that is arguably even more long decline in the real value of the U.S. min-
17. T. Piketty, G. Postel-Vinay, J. L. Rosenthal, Am. Econ. Rev. consequential for the “other 99 percent” of imum wage (7), the sharp drops in non-college
96, 236–256 (2006).
18. J. Roine, D. Waldenstrom, Scand. J. Econ. 111, 151–187
citizens: the dramatic growth in the wage pre- employment opportunities in production, clerical,
(2009). mium associated with higher education and, and administrative support positions stemming
19. H. Ohlson, J. Roine, D. Waldenstrom, in J. B. Davies, Ed., more broadly, cognitive ability. This paper con- from automation, the steep rise in interna-
Personal Wealth from a Global Perspective (Oxford Univ. Press, siders the role of the rising skill premium in tional competition from the developing world,
Oxford, 2008), pp. 42–63.
20. D. Waldenstrom, Lifting all Boats? The Evolution of Income and
the evolution of earnings inequality. the secularly declining membership and bar-
Wealth Inequality Over the Path of Development (Lund There are three reasons to focus a discus- gaining power of U.S. labor unions, and the
University, Sweden, 2009) sion of rising inequality on the economic pay- successive enactment of multiple reductions in
21. T. Piketty, Q. J. Econ. 126, 1071–1131 (2011). off to skills and education. First, the earnings top federal marginal tax rates, have all served to
22. R. Goldsmith, Comparative National Balance Sheets: A Study of
Twenty Countries, 1688-1978 (Univ. of Chicago Press, Chicago,
premium for education has risen across a large magnify inequality and erode real wages among
IL, 1985) number of advanced countries in recent dec- less educated workers. As I discuss below, the
23. T. Piketty, G. Zucman, Q. J. Econ. 129, in press ades, and this rise contributes substantially to foremost concern raised by these multiple forces
(2014); http://piketty.pse.ens.fr/files/ the net growth of earnings inequality. In the is not their impact on inequality per se, but
PikettyZucman2013WP.pdf.
24. T. Piketty, Capital in the Twenty-first Century (Harvard Univ.
United States, for example, about two-thirds rather their adverse effect on the real earnings
Press, Cambridge, MA, 2014). of the overall rise of earnings dispersion be- and employment of less educated workers.
25. J. Stiglitz, Econometrica 37, 382–397 (1969). tween 1980 and 2005 is proximately accounted I begin by documenting the centrality of the
26. M. Nirei, “Pareto Distributions in Economics Growth Models,” for by the increased premium associated with rising skill premium to the overall growth of
Institute of Innovation Research Working Paper No. 09-05,
Hitotsubashi University, Tokyo (2009)
schooling in general and postsecondary edu- earnings inequality. I next consider why skills
27. T. Piketty, G. Postel-Vinay, J. L. Rosenthal, Explor. Econ. Hist. cation in particular (1, 2). Second, despite a are heavily rewarded in advanced economies
51, 21–40 (2014). lack of consensus among economists regard- and why the demand for them has risen over
28. J. Davies, S. Sandstrom, T. Shorrocks, E. Wolff, Econ. J. 121, ing the primary causes of the rise of very top time. I then demonstrate the substantial ex-
223–254 (2011).
29. G. Zucman, Q. J. Econ. 128, 1321–1364 (2013).
incomes (3–6), an influential literature finds planatory power of a simple framework that
30. C. Goldin, L. Katz, The Race Between Education and Technology that the interplay between the supply and embeds both the demand and supply for skills
(Harvard Univ. Press, Cambridge, MA, 2008). demand for skills provides substantial insight in interpreting the evolution of the inequality
into why the skill premium has risen and fallen over five decades. The final section considers
SUPPLEMENTARY MATERIALS
over time—and, specifically, why the earnings the productive role that inequality plays in a
www.sciencemag.org/content/344/6186/838/suppl/DC1
market economy and the potential risks attend-
Supplementary Text ing very high and rising inequality; evidence on
Department of Economics and National Bureau of Economic
Figs. S1 and S2 whether those risks have been realized; and
Research, Massachusetts Institute of Technology, 40 Ames
References (31, 32) the role of policy and governance in encour-
Street, E17-216, Cambridge, MA 02142, USA. E-mail: dautor@
10.1126/science.1251936 mit.edu aging skills formation, fostering opportunity,

Labour Economics 11 (2004) 355 – 371
www.elsevier.com/locate/econbase
Does education reduce wage inequality? Quantile

regression evidence from 16 countries
Pedro S. Martins a,*, Pedro T. Pereira b
a
Department of Economics, University of Warwick, Coventry CV47AL, UK
b
Universidade da Madeira, CEPR and IZA, UK
Received 31 July 2001; received in revised form 29 January 2003; accepted 30 May 2003
Abstract
Quantile regression estimates of returns to education are used to address the relation between
schooling and wage inequality. Empirical evidence for male workers from 16 countries for the mid-
1990s suggests a robust stylised fact: Returns to schooling are higher for the more skilled
individuals, conditional on their observable characteristics. This suggests that schooling has a
positive impact upon within-levels wage inequality. Factors such as over-education, ability –
schooling interactions and school quality or different fields of study may be driving this result.
D 2003 Elsevier B.V. All rights reserved.
JEL classification: C29; J31; I21

Keywords: Returns to education; Wage inequality; Quantile regression
1. Introduction
Returns to education have been thoroughly analysed in the labour economics

literature. This is understandable as the pay-off to schooling is important information
for both public and private decisions on how much to invest in the sector (see Card, 1999
for a survey).
A related issue concerns wage inequality. After witnessing major increases in the spread
of wages since the early 1980s,1 some Western decision-makers have portrayed schooling
* Corresponding author. Tel.: +44-2476-523-935; fax: +44-2476-523-032.

E-mail address: p.martins@warwick.ac.uk (P.S. Martins).
1
See Katz and Murphy (1992) and Juhn et al. (1993) for a description of the US case and some tentative
explanations.
0927-5371/$ - see front matter D 2003 Elsevier B.V. All rights reserved.
doi:10.1016/j.labeco.2003.05.003
356 P.S. Martins, P.T. Pereira / Labour Economics 11 (2004) 355–371
as the best tool to erode the supposedly globalisation-related forces that increase wage
inequality. As Ashenfelter and Rouse (2000, p. 111) put it, ‘‘The school is a promising
place to increase the skills and incomes of individuals. As a result, educational policies
have the potential to decrease existing, and growing, inequalities in income’’.
This line of thought carries with it the presumption that new highly educated
cohorts will benefit from such levels’ traditionally high returns. However, this
approach disregards whether such levels are characterised by reasonably concentrated
or disperse returns. If the latter situation turns out to be the most representative, then
one should acknowledge the potential problems concerning within-levels inequality of
educational policies designed to erode wage dispersion. Moreover, the scarce evidence
available suggests that ‘‘differences in the extent of earnings inequality among high-
income countries are heavily influenced by the rewards for educational attainment’’
(Sullivan and Smeeding, 1997).
Our aim with this paper is to fill in this gap by drawing on quantile regression estimates
of returns to education. This approach allows us to assess the differences in the schooling-
related pay increment across the wage distribution. We therefore compare the returns to
education for the ‘‘skilled’’ and the ‘‘unskilled’’ workers (conditional on their schooling
and experience) in order to shed light on the contribution of schooling upon within-levels
wage inequality.
Furthermore, we provide evidence on this matter for a large number of countries. This
was achieved under the framework of a research project, ‘Public Funding and Private
Returns to Education’ (PuRE), where each country team analysed their country data sets.
Special care was taken to assure that these data sources were as similar and thus
comparable as possible.
The paper is structured as follows. Section 2 outlines the quantile regression
methodology. The following section describes the data sets used and provides
comparable descriptive statistics for the 16 countries analysed. Section 4 describes
the results obtained and the following section discusses them. Finally, Section 6
concludes.
2. Quantile regression
An ordinary least squares (OLS) regression is based on the mean of the conditional
distribution of the regression’s dependent variable. This approach is used because one
implicitly assumes that possible differences in terms of the impact of the exogenous
variables along the conditional distribution are unimportant.
However, this may prove inadequate in some research agendas. If exogenous
variables influence parameters of the conditional distribution of the dependent variable
other than the mean, then an analysis that disregards this possibility will be severely
weakened (see Koenker and Bassett, 1978). Unlike OLS, quantile regression models
allow for a full characterisation of the conditional distribution of the dependent
variable.2
2
See Abadie et al. (2002) for a recent extension of quantile regressions, considering instrumental variables.
P.S. Martins, P.T. Pereira / Labour Economics 11 (2004) 355–371 357
In a wage equation setting, the quantile regression model can be written as:
lnwi ¼ xi bh þ uhi with Quanth ðlnwi j xi Þ ¼ xi bh ð1Þ
where xi is the vector of exogenous variables and bh is the vector of parameters. Quanth
(ln wAx) denotes the hth conditional quantile of ln w given x. The hth regression
quantile, 0 < h < 1, is defined as a solution to the problem:
( )
X X
min hAlnwi xi bh A þ ð1 hÞAlnwi xi bh A ð2Þ
baRk
i:lnwi zxi b i:lnwi <xi b
This is normally written as:

X
min qh ðlnwi xi bh Þ; ð3Þ
baRk i
where qh(e) is the check function defined as qh(e) = he if e z 0 or qh(e) =(u 1)e if e < 0.
This problem does not have an explicit form but can be solved by linear programming
methods. Standard errors are obtainable by bootstrap methods.
The least absolute deviation (LAD) estimator of b is a particular case within this
framework. This is obtained by setting h = 0.5 (the median regression). The first quartile is
obtained by setting h = 0.25 and so on. As one increases h from 0 to 1, one traces the entire
distribution of y, conditional on x.
Summing up, quantile regressions provide snapshots of different points of a conditional
distribution. They therefore constitute a parsimonious way of describing the whole
distribution and should bring much value-added if the relationship between the regressors
and the independent variable evolves across its conditional distribution.
This flexibility has so far been precluded in the returns-to-education literature. In so
doing, it has left unaddressed the possible impact of schooling upon inequality, through its
within-levels inequality component. If the schooling-related earnings increment were the
same across the wage distribution, then schooling would not impact upon within-levels
wage inequality as distributions of wages conditional on different levels of schooling
would differ only on their locations and not on their dispersions.
However, it may be the case that these dispersions do indeed vary across educational
levels, thus resulting in an impact of schooling upon the wage distribution, through its
within-levels channel. This is the possibility we test, by using quantile regression
estimates, in the next two sections.
3. Data-sets description
The results for each country considered here were derived from a specific cross-section
data set used by each country’s team, within the above-mentioned ‘‘PuRE’’ research
project. Table 1 describes such data sets, referring the year for which the information
applies and also the number of observations used. In Appendix A, we provide a more
thorough description of these data sources.
Most data sets are household surveys. The exceptions are administrative registers (the
case of Denmark), labour-market surveys (France) and employer-based data sets (Nether-
358
Table 1
Data-sets description, descriptive statistics and inequality measures
P.S. Martins, P.T. Pereira / Labour Economics 11 (2004) 355–371

Country Data set Year No. of Educ. Exp. Log Wage Wage Wage ratios (1)
observations Mean C.V. 10th 50th 90th 9/1 9/5 5/1
percentile percentile percentile
Austria Mikrozensus 1993 7175 10.1 21.3 4.57 0.077 65.8 93.8 150 2.28 1.6 1.43
Denmark Long. Lab. Market Reg. 1995 4416 12 19.4 4.97 0.072 96.5 138.4 230.4 2.39 1.67 1.43
Finland Labour Force Survey 1993 1175 11.4 19.5 4.16 0.091 41.9 62.1 106.1 2.53 1.71 1.48
France Training Qualif. + 1993 4606 11.4 21.9 10.92 0.036 19.8 29.8 54.1 2.73 1.81 1.5
Employment Survey
Germany Socio-Economic Panel 1995 1070 11.9 24.7 3.4 0.103 2.64 2.92 3.01 1.45 1.09 1.33
Greece Household Budget Survey 1994 2096 10.1 21.9 6.93 0.092 527 1103 1907 3.62 1.73 2.09
Ireland ESRI Household Survey 1994 1903 12.4 23.8 1.74 0.351 2.5 5.9 11.9 4.74 2.01 2.36
Italy Survey of Household Income 1995 3441 10.1 22.9 2.52 0.163 7.8 12.5 20.8 2.67 1.67 1.6
and Wealth
Netherlands Structure of Earnings Survey 1996 49805 12.5 20 3.23 0.142 15.5 24.9 43.8 2.83 1.75 1.61
Norway Level of Living Survey 1995 870 12.2 20.9 4.65 0.071 71.4 101.1 158 2.21 1.56 1.42
Portugal Personnel Records 1995 28055 6.5 24.5 6.42 0.095 318 531 1456 4.58 2.74 1.67
Spain Wage Structure Survey 1995 118005 8.8 26 7.3 0.071 761 1410 2999 3.94 2.13 1.85
Sweden Level of Living Surveys 1991 1508 11.8 21.5 4.45 0.070 61 81 127 2.08 1.57 1.33
Switzerland Labour Force Survey 1995 6334 13.2 19.8 3.6 0.111 23.9 35.9 60.3 2.53 1.68 1.51
UK Family Expenditures Survey 1995 2183 12.3 22.6 2 0.245 4.1 7.3 13.5 3.33 1.85 1.8
USA Current Population Survey 1995 42347 12.6 18.5 2.33 0.202 5.5 10 19 3.45 1.82 1.9
See Appendix A for a more detailed characterisation of the data sets.
Results for France and Spain refer to yearly earnings. Hourly wages for France and Spain were computing assuming 1760 h/year. Inequality figures (1, 5, 9) refer to 10th,
50th and 90th percentiles.
Fig. 1. Returns to education, OLS.
lands, Portugal and Spain). The number of observations used varies reasonably, ranging
from fewer than 2000 (Finland, Ireland, Norway and Sweden) to more than 20,000
(Netherlands, Portugal, Spain and the US). All observations refer to full-time male workers
only.
All countries draw on gross wages as the measure of earnings. The exceptions are
Austria, Greece and Italy, which use net wages, as this was the single type of wage data
available. This difference in types of wages may be important, on account of the
progressivity of tax systems, and thus trouble the comparability of the results. This will
be bore in mind in the following stages of the paper. Hourly wages, the dependent
variable, is computed by dividing total wages by total hours in all countries, except for
France and Spain. In these two countries, only yearly wages were available, which were
then divided by 1760 h.3 Another issue is top coding, which affects the US data set.
However, our analysis with a different data set suggests that this is likely not to distort the
qualitative results.4
3
The sensitivity of the results to the procedure adopted to deal with the lack of hours data was examined by
imposing the same number of total hours for all workers in the Portuguese and US data sets. The results, available
from the authors, are robust to different procedures.
4
We examine the role of top coding by drawing on Portuguese data and artificially censoring the 1% or 5%
top of the wage distribution. Again, this matter does not influence the qualitative results. Another issue that may
be particularly important in the US is the role of race. However, we extended the main specification with racial
dummies (blacks, Hispanics, Asians and native Americans) and found no substantial differences. Both sets of
results are also available from the authors upon request.
Table 2
Summary of results
Country OLS 1st decile 9th decile Difference
Austria 9.7% 7.2% 12.8% 5.6%
Denmark 6.6% 6.3% 7.1% 0.8%
Finland 8.9% 6.8% 10.1% 3.3%
France 7.6% 5.9% 9.3% 3.4%
Germany 8.0% 7.8% 8.1% 0.3%
Greece 6.5% 7.5% 5.6% 1.9%
Italy 6.4% 6.7% 7.1% 0.4%
Ireland 8.9% 7.8% 10.4% 2.6%
Netherlands 7.0% 5.3% 8.3% 3.0%
Norway 6.0% 5.5% 7.5% 2.1%
Portugal 12.6% 6.7% 15.6% 8.9%
Spain 8.6% 6.7% 9.1% 2.4%
Sweden 4.1% 2.4% 6.2% 3.8%
Switzerland 9.5% 8.7% 10.6% 1.9%
UK 8.6% 4.9% 9.7% 4.8%
USA 6.3% 3.9% 7.9% 4.0%
Means 7.9% 6.5% 9.1% 2.7%
St. Dev. 2.0% 1.6% 2.6% 2.7%
Coeff. Var. 0.25 0.24 0.29 1.00
Table 1 also presents descriptive statistics from each country’s data set. Most
countries exhibit levels of average schooling above 10 years, the highest value being
that of Switzerland (13.2). The lowest are those of Portugal and Spain, with 6.5 and 8.8
years of average schooling, respectively. Average experience (which corresponds, in all
countries, to Mincer experience, age-schooling-6) is generally above 19 and below 22
years. There are a few exceptions, namely Portugal and Spain, with much higher levels
(24.5 and 26, respectively), which is not surprising given their low levels of average
schooling.
The means and the coefficients of variation of the logarithm of hourly wages are also
reported. However, they cannot be compared in a straightforward manner as they are based
on different currencies (and slightly different years). The same applies to the hourly wages
at the first, fifth and ninth deciles. We used this evidence in the three last columns of Table
1, where we report simple inequality measures, such as the ratios between wages at
different deciles.
In the columns headed ‘‘Wage ratios’’, the ratio of the wages at the ninth and
first deciles’ are generally between 2 and 3. The exceptions are the UK (3.33), US
(3.45), Greece (3.62), Spain (3.94), Portugal (4.58) and Ireland (4.74).5 Comparing
the second and third columns, it can also be seen that in most countries the largest
share of the ninth –first deciles inequality is obtained from the top half of the
5
With respect to the cases of Austria, Greece and Italy, their measures of inequality should be understood as
lower bounds of the true measure, as their wage figures are net of taxes, unlike those of the other countries, and
therefore influenced by the progressive nature of the tax systems.
Table 3
Quantile regressions results (coefficients and SEs)
Austria (1993) Denmark (1995) Finland (1993) France (1993)
0.1 0.070 0.0034 0.061 0.0026 0.066 0.0067 0.0571 0.00203
0.2 0.075 0.0030 0.062 0.0020 0.083 0.0048 0.0652 0.00197
0.3 0.082 0.0020 0.061 0.0016 0.080 0.0052 0.0682 0.00183
0.4 0.087 0.0028 0.061 0.0017 0.081 0.0042 0.0728 0.00188
0.5 0.091 0.0031 0.061 0.0021 0.088 0.0045 0.0755 0.00174
0.6 0.098 0.0034 0.065 0.0019 0.087 0.0050 0.0809 0.00210
0.7 0.106 0.0042 0.067 0.0024 0.092 0.0048 0.0825 0.00216
0.8 0.113 0.0031 0.069 0.0024 0.092 0.0062 0.0840 0.00290
0.9 0.120 0.0046 0.069 0.0041 0.096 0.0096 0.0890 0.00345
OLS 0.093 0.0021 0.064 0.0018 0.086 0.0042 0.0733 0.00156
Germany (1995) Greece (1994) Ireland (1994) Italy (1995)

0.1 0.0748 0.00447 0.073 0.0072 0.075 0.0102 0.065 0.0034
0.2 0.0753 0.00528 0.063 0.0043 0.085 0.0056 0.063 0.0024
0.3 0.0788 0.00324 0.060 0.0041 0.087 0.0048 0.057 0.0021
0.4 0.0785 0.00328 0.059 0.0028 0.089 0.0040 0.057 0.0017
0.5 0.0820 0.00329 0.056 0.0027 0.099 0.0052 0.056 0.0015
0.6 0.0837 0.00318 0.056 0.0028 0.098 0.0055 0.057 0.0019
0.7 0.0865 0.00303 0.055 0.0029 0.100 0.0044 0.061 0.0020
0.8 0.0851 0.00367 0.053 0.0034 0.102 0.0032 0.065 0.0026
0.9 0.0780 0.00437 0.055 0.0047 0.099 0.0049 0.068 0.0033
OLS 0.0803 0.00372 0.063 0.0033 0.086 0.0047 0.062 0.0017
Norway (1995) Netherlands (1996) Portugal (1995) Spain (1995)

0.1 0.053 0.0071 0.051 0.0014 0.065 0.0010 0.065 0.0004
0.2 0.048 0.0043 0.054 0.0008 0.083 0.0010 0.076 0.0004
0.3 0.051 0.0042 0.059 0.0008 0.099 0.0009 0.083 0.0004
0.4 0.049 0.0025 0.061 0.0007 0.112 0.0009 0.086 0.0004
0.5 0.056 0.0039 0.063 0.0007 0.122 0.0009 0.087 0.0004
0.6 0.065 0.0044 0.066 0.0008 0.131 0.0011 0.087 0.0004
0.7 0.069 0.0060 0.070 0.0008 0.136 0.0012 0.087 0.0004
0.8 0.070 0.0049 0.074 0.0010 0.140 0.0013 0.087 0.0005
0.9 0.073 0.0080 0.079 0.0013 0.145 0.0017 0.087 0.0006
OLS 0.059 0.0039 0.068 0.0006 0.119 0.0009 0.082 0.0003
Sweden (1991) Switzerland (1995) United kingdom (1995) USA (1995)

0.1 0.024 0.0027 0.084 0.0036 0.048 0.0070 0.039 0.0012
0.2 0.028 0.0021 0.084 0.0024 0.056 0.0056 0.050 0.0012
0.3 0.031 0.0022 0.086 0.0022 0.066 0.0053 0.057 0.0011
0.4 0.036 0.0023 0.090 0.0016 0.071 0.0047 0.065 0.0012
0.5 0.043 0.0026 0.092 0.0014 0.070 0.0042 0.068 0.0010
0.6 0.045 0.0025 0.094 0.0018 0.070 0.0036 0.072 0.0009
0.7 0.050 0.0029 0.096 0.0016 0.069 0.0040 0.074 0.0010
0.8 0.055 0.0036 0.100 0.0020 0.075 0.0049 0.075 0.0010
0.9 0.060 0.0044 0.101 0.0026 0.092 0.0060 0.076 0.0015
OLS 0.041 0.0022 0.090 0.0019 0.083 0.0041 0.061 0.0008
Fig. 2. Returns to education, QR and OLS.

Fig. 2. (continued).
distribution (ninth– fifth deciles), the exceptions being Germany, Greece, Ireland and
the US.6
4. Empirical results
The empirical results were obtained by regressing the following version of the Mincer
(1974) equation, under Becker’s (1975) framework:
logyi ¼ ah þ bh educi þ dh1 expi þ dh2 exp2i þ ui ;
where i = 1,. . .,N (N being the number of observations for each year), h = 0.1,0.2,. . .,0.9
is the quantile being analysed, y is the hourly wage, educ is the number of schooling
6
These results are generally in accordance with those presented at Gottschalk and Smeeding (1997). However,
a thorough comparison is impossible as both the time period and the earnings measure covered there are different.
Fig. 3. Returns to education, quantile regression.
years7 and exp corresponds to Mincer experience. Only men working full time (35 h or
more per week) were considered. The case of women was disregarded on account of the
extra complication of potential selectivity biases. 8
As a benchmark, we first present the results obtained with the traditional OLS method.9
In Fig. 1, we rank countries by their OLS return to education. As one can see in Table 2,
the mean return is 7.9%, with a standard deviation of 2%. Sweden exhibits the lowest
value (4%) whereas Portugal displays the highest (12.6%).
In Table 3, we present the coefficients and associated standard errors for both
OLS and quantile regression estimates, for each country. These results are pictured
in Fig. 2A – E, where we exhibit both the returns at the mean (OLS) and at
different points of the wage distribution (namely the quantiles 0.1,0.2,. . .,0.9). The
stylised fact that comes out from this analysis, which is also the key result of the
7
We use information on the highest level achieved. Extra school attainment above the number of school
years associated with the degree is thus disregarded.
8
For an application of quantile regressions accounting for selectivity issues, see Buchinsky (1998).
9
See Asplund and Pereira (1999) and Harmon et al. (2001) for surveys of OLS returns to education across
Europe, which were also prepared by the same network involved in the ‘‘PuRE’’ project.
paper, is that returns to education are higher at higher points of the (conditional)
wage distribution.10
The Swedish case is a good example of this general result. Whereas the average return
to education is of 4%, the return at the first decile is no greater than 2% and the return at
the ninth decile reaches 6%. The more extreme case of this pattern is the Portuguese case.
Here an average return of 13% masks a return of only 6% at the first decile and more than
15% at the last decile.11
In the set of 16 countries considered here, only Greece does not follow this pattern.
However, the data for this country are based in net wages, which troubles a full
comparison with the remaining countries. In particular, progressive taxes are likely to
have a strong impact in eroding the returns to education at the top of the distribution than
at its bottom. This may explain the Greek results.
A summary graphical description of these results can be found in Fig. 3. Here we only
consider the returns to education at the first and ninth deciles for each country. One can see
that most countries are lying on the top left part of the figure, which means that their
returns at the top of the distribution (proxied by the ninth decile) are higher than those at
the bottom (first decile). The clear exception is Greece while Denmark, Germany and Italy
are relatively close to the 45j line that separates each situation.12
5. Discussion
This paper presents an empirical finding concerning returns to education across the
wage distribution. More skilled workers (individuals who receive higher hourly wages
conditional on their characteristics) are associated with a stronger education-related
earnings increment. We put forward three possible explanations for this result.
A first one lies on over-education. In fact, situations where highly schooled workers
take jobs with a low-skill requirement and consequent low pay would be consistent with
these results. In this case, the lower earnings of the over-educated will increase the within-
skill dispersion of pay by extending the lower tails of the highly educated wage
distributions. In the extreme case that the lower tail of the highly skilled workers earnings
distribution is fully taken by the over-educated (that is, by highly skilled workers with
low-skills jobs), returns to education at the bottom quantiles would be particularly low.
One reason why one would expect these returns to be positive is that the over-education
10
This result is in accordance with previous attempts at estimating wage equations with quantile regressions,
such as Machado and Mata (2001), Hartog et al. (2001) and Fersterer and Winter-Ebmer (2003). We have also
experimented with other specifications for a subset of countries (Portugal and the US), in particular considering a
quartic in experience, following Murphy and Welch (1990). The qualitative results—not shown but available
upon request—were unaltered, although we found higher estimates under the more extended specification.
11
Tests on whether these coefficients are significantly different have been performed for most countries and
are available upon request from the authors. Except for some countries whose sample sizes are small, these tests
support our result of significantly higher coefficients at the upper tail of the distribution.
12
These data are examined from the point of view of the risk involved in education in Pereira and Martins
(2002). In this paper, a significantly positive cross-country correlation between the spread of returns at the top and
bottom of the wage distribution and the mean return to education is presented and discussed.
literature (see Hartog (2000) for a survey) finds that, although returns to over-education are
lower than the returns to ‘‘appropriate’’ education, they are in any case still positive.
A second alternative or complementary explanation is about ability. In the line of the
much-debated results of Herrnstein and Murray (1995),13 ability may be the most relevant
force in explaining socio-economic success. We take this argument one step further and
hypothise that higher schooling levels may compound the role of ability if there is an
interaction between the two variables. This would mean that the role of ability differences
within a given schooling level would become increasingly amplified in terms of pay as one
considers high schooling levels. While in lower education levels, differences in ability
would not be too relevant (and thus the dispersion of earnings would be low), in higher
education levels, differences in ability would translate into substantial pay gaps between
the high- and low-ability workers.
Finally, differences in school quality or fields of study may be another driving force for
the pattern uncovered. Although the Mincer framework considers only differences in school
quantity, it may be the case that those individuals that fall into the bottom of their conditional
earnings distributions are precisely those who benefited from poorer school quality or who
have chosen fields of study with (ex-post) poor returns. These events are likely to jeopardise
the earnings potential from these individuals’ human capital, placing them at an increasingly
worse position (in relative terms) the higher their educational level. The bottom of the wage
distributions would then be over-represented with workers with low-level school quality or
who engaged on fields of study that attract scarce interest in the labour market. Moreover,
these differences are likely to be more prevalent at higher schooling levels, because those are
the stages that exhibit more variety in schooling paths—and possibly also in schooling
quality (conditional on the school-leavers at each given school level).
More generally, these explanations can be nested into an interaction between schooling
and some factor or set of factors that also impacts upon pay differentials and which are
heterogeneously distributed across workers within any given skills level. To the extent that
these factors have a stronger influence upon earnings for workers with higher schooling
levels, then the empirical stylised fact documented here would be obtained.
A final, important issue concerns the possible impact of endogeneity on the results
presented. If high-ability individuals are over-represented in higher-schooling levels
(because they face lower marginal costs to schooling, for instance), then not considering
ability differences would wrongly lead to an overestimate (underestimate) of the return to
schooling at the upper (lower) part of the conditional distributions. This would indeed
correspond to the pattern uncovered here.
However, an important recent contribution in this topic, Arias et al. (2001), suggests
this is not the case. In this paper, the authors draw on the data from the twins study by
Ashenfelter and Rouse (1998) in order to control for ability differences.14 They find that,
although this attenuates the extent of differences in returns to education across the
distribution, these returns do still remain significantly different.
13
See Heckman (1995) for a critique of the methodology and conclusions of ’The Bell Curve’.
14
This procedure—using twins data to identify the returns to education—is not, however, completely
satisfactory. For instance, one may be concerned about how exogenous are the differences in schooling between
twins which are implicitly used to overcome the endogeneity of the returns to education.
6. Conclusions
In this paper we analyse the dispersion of the returns to education at different Western
countries during the mid-1990s, with a view to addressing the link between schooling and
within-levels inequality. These issues have not received much attention so far in the
literature, which has implicitly assumed that the schooling-related earnings increment is
constant across the wage distribution. Notable exceptions are Card (1999) and Buchinsky
(1994) who examine these phenomena, both theoretically and empirically—the latter
focusing on the US. In particular, Card (1994) asked, ‘Is the labour force reasonably well
described by a constant return to education for all workers?’ The evidence we provide
suggests that this question should be answered negatively.
In fact, we found a stylised fact over the 16 developed countries we cover: returns to
schooling increase over the wage distribution. Or, to put it differently, the earnings
increment associated to schooling is higher for those individuals whose unobservable
characteristics place them at the top of the conditional wage distribution.
These findings imply that schooling may have a positive impact upon within-group
wage inequality, as the spread of returns increases for higher educational levels. One
possible explanation for this is over-education (when individuals with higher schooling
attainment take jobs requiring lower skills). Another possibility is that there is an
interaction between schooling and ability, in which the most able can benefit more from
their schooling and the pay gap between the more and less able deepens for higher
educational levels. A final explanation may lie on differences in school quality or fields of
study. Here, the bottom of the wage distributions would be over-represented with workers
with low-level school quality or who engaged on fields of study that attract scarce interest
in the labour market. These differences are likely to be more prevalent at higher schooling
levels, when there is more variety in schooling paths and schooling quality.
More generally, these explanations can be nested to an interaction between schooling
and some factor or set of factors that also impacts upon pay differentials and which is
heterogeneously distributed across workers within any given education level. To the extent
that these factors have a stronger influence upon earnings for workers with higher
schooling levels, then the empirical stylised fact documented here would be obtained.
In any case, regardless of the true source of this result—and this is certainly an area
deserving further research—we believe this finding rings some alarm bells at policies
designed to cut wage inequality by simply investing in the attainment of higher schooling
levels. Even if a given population were made of only highly skilled individuals—which
would, on average, command high wages—such population, according to our evidence,
would still exhibit considerable levels of pay inequality.
Acknowledgements
This paper was written under the scope of the 15-country ‘PuRE—Public Funding and
Private Returns to Education’ European Commission TSER project. The results presented
here are based on work done by country teams headed by Mahmood Arai, Rita Asplund,
Erling Barth, Giorgio Brunello, Colm Harmon, Joop Odink, Pedro Pereira, José Luis
Raymond, Ali Skalli, Viktor Steiner, Panagiotis Tsakloglou, Ian Walker, Niels West-
ergaard-Nielsen, Rudolf Winter-Ebmer and Stefan Wolter. We also thank Robin Naylor,
participants at several seminars and a referee and the associate editor (D. Black) for
comments and suggestions. Financial support from Fundacß ão para a Ciência e a
Tecnologia (SFRH/BD/934/2000 and POCTI/ECO/33089/99) and the British Council and
research support from the Bank of Portugal are also gratefully accredited. The usual
disclaimer applies.
Appendix A . Description of data sets used
A.1 . Common features
This section presents the common features across all data sets. Exceptions to this
pattern are documented in the following section, which also describes in greater and
somewhat varying detail other relevant information about each country’s data set.
The information used in all data sets refers to men, aged 15– 65, working at least 35 h/
week. The dependent variable is the logarithm of gross hourly wages, obtained by dividing
weekly, monthly or annual wages by the correspondent number of hours worked. This
variable is not censored. Schooling years are obtained from considering the highest grade
obtained by the worker and then the number of schooling years typically required for a
student to achieve such grade. Experience is Mincer experience (age-education-6).
A.2 . Specific Features
Austria: Mikrozensus. A representative 1% household survey, including detailed

information about human capital variables. Information on net monthly earnings. All
employees (white-collar, blue-collar and civil servants) aged between 15 and 65 years are
included in the sample. Apprentices have been eliminated from the analysed population.
Denmark: Longitudinal Labour Market Register. A random 0.5% sample of the
adult population, covering the years 1976 –1995. All information in LLMR is drawn from
administrative registers and is merged by Statistics Denmark. Estimations are based on
people aged 16– 64.
Finland: Finnish Labour Force Survey. Compiled by Statistics Finland. Represen-
tative sample of the whole Finnish population. The sample has traditionally contained
some 9000 individuals aged 15– 64. Apart from standard individual characteristics, also
the information on and income is register based. The rest of the information is self-
reported through questionnaires and interviews undertaken by Statistics Finland. The
earnings concept refers to the individual’s average gross hourly wage as calculated from
tax record information on taxable annual earnings and self-reported numbers of months
and normal hours worked. The annual earnings comprise all types of compensation, such
as overtime and vacation pay, except for fringe benefits.
France: Training and Professional Qualifications. Survey conducted by INSEE, the
French national statistics institute. Richest French data set in terms of initial as well as
post-school education and their professional outcomes. The household survey called
Enquête Emploi (Employment Survey) is also used. Only FQP gives detailed information
on individuals’ educational record and family background. In FQP, only gross annual
earnings are available.
Germany: German Socio-Economic Panel. Longitudinal household survey con-
ducted on a yearly basis. Detailed information about income, labour market status,
education and various other socio-economic variables is collected. The sample was
restricted to West German citizens. The self-employed, pensioners, military personnel,
people still engaged in education or training were also excluded, as well as foreigners,
whose educational background may fundamentally differ from that of native Germans.
Greece: Household Budget Surveys (1993/1994). The only data set available that
covers in a consistent way the entire labour force over the last 25 years and contains
income information. Carried out by the national statistical service of Greece. It covers
the entire non-institutional population of the country and their sampling fraction is
2x. It contains detailed information about consumption expenditures, incomes and
socio-economic characteristics of the households and their members. The income
component used in the paper is ‘‘earnings net of income taxes and social insurance
contributions’’. It includes wages, salaries, overtime payments, bonuses, holiday
payments, and related benefits received from the main and secondary employer,
normalised on a monthly basis. Further, the surveys report the number of hours
normally worked per week. Division of monthly income adjusted on a weekly basis
by this figure yields ‘‘net hourly earnings’’. The samples used consist of employees
outside the agricultural sector aged 14 – 64. Thus, self-employed, employers, unpaid
family members and apprentices are excluded.
Ireland: Household Survey. Carried out by the Economic and Social Research
Institute (ESRI), in 1994. First wave of the Irish part of the European Community
Household Panel. Rich survey with regards to labour market experience, including a very
comprehensive list of current gross earnings, deductions and net earnings. Includes a
measure of the number of schooling years and the highest level of education attained.
There is a high level of confidence in the reliability of the data, in terms of how
representative it is of the population.
Italy: Survey of Household Income and Wealth. Conducted by the Bank of Italy.
Available from 1977 annually and at odd years after 1987. It contains information both on
households (family composition) and on individuals. This information includes net yearly
earnings, average weekly hours of work and number of months of employment per year. It
also contains information on family background (the education, age, occupation and sector
of parents). There are no other nationally representative surveys in Italy that cover the
same range of information. The sample is restricted to non-agricultural employees aged
from 14 to 65.
Netherlands: Structure of Earnings Survey. Conducted by the Dutch Central Bureau
of Statistics. Until 1979, the Structure of Earnings Surveys were large cross-sectional
employer surveys in which information on gross earnings, educational level, sex, age, and
industry of employees was gathered. The 1995 version of the Structure of Earnings Survey
was created by combining information at the individual level of three different data
sources: the 1995 ‘‘Employment and Wages Survey’’, the 1995 ‘‘Insured Persons
Register’’, and the 1994– 1996 ‘‘Labour Force Surveys’’.
Norway: Level of Living Surveys. Ongoing project at Statistics Norway, surveying

a sample of the Norwegian adult population. Panel survey, adding young persons in
every wave. Every LLS contains about 5000 individuals, comprising around 2500
wage observations. The analysis is limited to wage earners, 16 – 67 years of age. The
wage variable is calculated as reported monthly/weekly/hourly gross wage, divided by
the number reported weekly hours. The individuals are asked to report their usual level
of wages and hours—including usual level of overtime. In the surveys from 1989 and
onwards, the educational variable is merged from administrative registers (5-digit
code).
Portugal: Personnel Records. Every year since the late 1970s, all firms, either private
or public, provide data concerning every employee and also detailed firm-specific
information. This data set covers the entire Portuguese labour market (employees) and
has a panel structure (at both the employee and employer levels) since the late 1980s.
Spain: Wage Structure Survey 1995. Employer survey of 175,000 wage earners,
which contains an important amount of characteristics related to each worker (qualifica-
tion, tenure, type of contract, type of job, sector, firm size, and so on). All surveys were
purged dropping those observations with wages below minimum wage and age below 18
or above 65.
Sweden: Swedish Level of Living Survey, 1991. SLL surveys are the most widely
used Swedish data sets for wage equations. SLLS contains about 6000 randomly sampled
people between ages 16 and 75 (18 –75 for 1991), where the primary sample from 1968
constitutes the base for these surveys: 1974, 1981 and 1991. The non-response rate has
increased over time, from 9% in 1968 to roughly 20% in 1991.
Switzerland: Swiss Labour Force Survey. Main data source for labour-market related
questions in Switzerland. Produced annually by the Swiss Federal Statistical Office since
1991. The sample of 32,000 in 1995 is representative for the adult population (older than
15 years) permanently living in Switzerland. The data are collected by telephone
interviews.
United Kingdom: Family Expenditure Survey. Random sample of approximately
7000 households each year, available since the 1960s. In addition to education and
earnings FES contains some information relating to union status and has smoking and
other expenditures. Refers to Great Britain (i.e., UK excluding Northern Ireland).
United States: Current Population Survey. Monthly household survey used to
address labour-market related issues, e.g., unemployment rates.
References
Abadie, A., Angrist, J., Imbens, G., 2002. Instrumental variables estimates of the effect of subsidized training on
the quantiles of trainee earnings. Econometrica 70, 91 – 117.
Arias, O., Hallock, K., Sosa-Escudero, W., 2001. Individual heterogeneity in the returns to schooling: instru-
mental variables quantile regression using twins data. Empirical Economics 26, 7 – 40.
Ashenfelter, O., Rouse, C., 1998. Income, schooling and ability: evidence from a new sample of identical twins.
Quarterly Journal of Economics 113, 253 – 284.
Ashenfelter, O., Rouse, C., 2000. Schooling, intelligence and income in America. In: Arrow, K., Bowles, S.,
Durlauf, S. (Eds.), Meritocracy and Economic Inequality. Princeton Univ. Press, Princeton.
Asplund, R., Pereira, P., 1999. Returns to Human Capital in Europe: A Literature Review. ETLA, Helsinki.
Becker, G., 1975. Human Capital: A Theoretical and Empirical Analysis, with Special Reference to Education.
University of Chicago Press, Chicago.
Buchinsky, M., 1994. Changes in the US wage structure, 1963 – 1987: application of quantile regression. Econ-
ometrica 62, 405 – 458.
Buchinsky, M., 1998. The dynamics of changes in the female wage distribution in the USA: a quantile regression
approach. Journal of Econometrics 13, 1 – 30.
Card, D., 1994. Earnings, schooling and ability revisited. NBER Working 4832.
Card, D., 1999. The causal effect of education on earnings. In: Ashenfelter, O., Card, D. (Eds.), Handbook of
Labour Economics, vol. 3. North-Holland, Amsterdam, pp. 1801 – 1863.
Fersterer, J., Winter-Ebmer, R., 2003. Are Austrian returns to education falling over time? Labour Economics 10,
73 – 89.
Gottschalk, P., Smeeding, T., 1997. Cross-national comparisons of earnings and income inequality. Journal of
Economic Literature 35, 633 – 687.
Harmon, C., Walker, I., Westergaard-Nielsen, N., 2001. Education and Earnings in Europe: A Cross-Country
Analysis of Returns to Education. Edward Elgar, Cheltenham.
Hartog, J., 2000. Over-education and earnings: where are we, where should we go? Economics of Education
Review 19, 131 – 147.
Hartog, J., Pereira, P., Vieira, J.C., 2001. Changing returns to education in Portugal during the 1980s and early
1990s: OLS and quantile regression estimators. Applied Economics 33, 1021 – 1037.
Heckman, J., 1995. Lessons from the bell curve. Journal of Political Economy 103, 1091 – 1120.
Herrnstein, R., Murray, C., 1995. The Bell Curve: Intelligence and Class Structure in American Life. Free Press,
New York.
Juhn, C., Murphy, K., Pierce, B., 1993. Wage inequality and the rise in returns to skill. Journal of Political
Economy 101, 3.
Katz, L., Murphy, K., 1992. Changes in relative wages, 1963 – 1987: supply and demand factors. Quarterly
Journal of Economics 107, 35 – 78.
Koenker, R., Bassett, G., 1978. Regression quantiles. Econometrica 46, 33 – 50.
Machado, J., Mata, J., 2001. Earnings functions in Portugal 1982 – 1994: evidence from quantile regressions.
Empirical Economics 26, 115 – 134.
Mincer, J., 1974. Schooling, experience and earnings. NBER, New York.
Murphy, K.M., Welch, F., 1990. Empirical age-earnings profiles. Journal of Labor Economics 8, 202 – 229.
Pereira, P., Martins, P., 2002. Is there a return-risk link in education? Economics Letters 75, 31 – 37.
Sullivan, D., Smeeding, T., 1997. Educational attainment and earnings inequality in eight nations. International
Journal of Educational Research 27, 513 – 525.
of rising or shrinking inequality. Which one
dominates depends on the institutions and pol-
icies that societies choose to adopt. REVIEW
RE FE RENCES
1. S. Kuznets, Shares of Upper Income Groups in Income and
Skills, education, and the rise of
Savings (National Bureau of Economic Research, Cambridge,
MA, 1953).
2. R. J. Lampman, The Share of Top Wealth holders in National
earnings inequality among
the “other 99 percent”
Wealth, 1922-1956 (Princeton Univ. Press, Princeton, NJ,
1962).
3. A. B. Atkinson, A. J. Harrison, Distribution of Personal Wealth
in Britain, 1923-1972 (Cambridge Univ. Press, Cambridge,
1978). David H. Autor
4. A. Daumard, Les fortunes françaises au 19e siècle.
Enquête sur la répartition et la composition des capitaux
privés à Paris, Lyon, Lille, Bordeaux et Toulouse d'après The singular focus of public debate on the “top 1 percent” of households overlooks the
l'enregistrement des déclarations de successions (Mouton, component of earnings inequality that is arguably most consequential for the “other
Paris, 1973). 99 percent” of citizens: the dramatic growth in the wage premium associated with higher
5. A. H. Jones, American Colonial Wealth: Documents and
education and cognitive ability. This Review documents the central role of both the supply
Downloaded from www.sciencemag.org on September 8, 2015

Methods (Arno Press, New York, 1977).
6. P. Lindert, J. Polit. Econ. 94, 1127–1162 (1986). and demand for skills in shaping inequality, discusses why skill demands have persistently
7. L. Soltow, Distribution of Wealth and Income in the United risen in industrialized countries, and considers the economic value of inequality alongside
States in 1798 (Univ. of Pittsburgh Press, Pittsburgh, PA, its potential social costs. I conclude by highlighting the constructive role for public policy in
1989).
8. T. Piketty, Les hauts revenus en France au 20e
fostering skills formation and preserving economic mobility.
P
siècle—Inégalités et redistributions, 1901–1998
(Grasset, Paris, 2001). ublic debate has recently focused on a gap between college and high school graduates
9. T. Piketty, J. Polit. Econ. 111, 1004–1042 (2003). subject that economists have been ana- has more than doubled in the United States over
10. A. B. Atkinson, J. R. Stat. Soc. Ser. A Stat. Soc. 168, 325–343
(2005).
lyzing for at least two decades: the steep, the past three decades. A third reason for focus-
11. T. Piketty, E. Saez, Q. J. Econ. 118, 1–41(2003). persistent rise of earnings inequality in ing on the skill premium is that it offers broad
12. A. B. Atkinson, T. Piketty, Eds., Top Incomes over the the U.S. labor market and in developed insight into the evolution of inequality within a
20th Century—A Contrast Between Continental European countries more broadly. Much popular dis- market economy, highlighting the social value of
and English Speaking Countries (Oxford Univ. Press, New York,
2007).
cussion of inequality concerns the “top 1 percent,” inequality alongside its potential social costs and
13. A. B. Atkinson, T. Piketty, Eds., Top Incomes—A Global referring to the increasing share of national in- illuminating the constructive role for public policy
Perspective (Oxford Univ. Press, New York, 2010). come accruing to the top percentile of house- in maximizing the benefits and minimizing the
14. A. B. Atkinson, T. Piketty, E. Saez, J. Econ. Lit. 49, 3–71 holds. Although this phenomenon is undeniably costs of inequality.
(2011).
15. F. Alvaredo, A. B. Atkinson, T. Piketty, E. Saez, J. Econ.
important, an exclusive focus on the concen- The rising skill premium is not, of course, the
Perspect. 27, 3–21 (2013). tration of top incomes ignores the component sole cause of growing inequality. The decades-
16. W. Kopczuk, E. Saez, Natl. Tax J. 57, 445–487 (2004). of rising inequality that is arguably even more long decline in the real value of the U.S. min-
17. T. Piketty, G. Postel-Vinay, J. L. Rosenthal, Am. Econ. Rev. consequential for the “other 99 percent” of imum wage (7), the sharp drops in non-college
96, 236–256 (2006).
18. J. Roine, D. Waldenstrom, Scand. J. Econ. 111, 151–187
citizens: the dramatic growth in the wage pre- employment opportunities in production, clerical,
(2009). mium associated with higher education and, and administrative support positions stemming
19. H. Ohlson, J. Roine, D. Waldenstrom, in J. B. Davies, Ed., more broadly, cognitive ability. This paper con- from automation, the steep rise in interna-
Personal Wealth from a Global Perspective (Oxford Univ. Press, siders the role of the rising skill premium in tional competition from the developing world,
Oxford, 2008), pp. 42–63.
20. D. Waldenstrom, Lifting all Boats? The Evolution of Income and
the evolution of earnings inequality. the secularly declining membership and bar-
Wealth Inequality Over the Path of Development (Lund There are three reasons to focus a discus- gaining power of U.S. labor unions, and the
University, Sweden, 2009) sion of rising inequality on the economic pay- successive enactment of multiple reductions in
21. T. Piketty, Q. J. Econ. 126, 1071–1131 (2011). off to skills and education. First, the earnings top federal marginal tax rates, have all served to
22. R. Goldsmith, Comparative National Balance Sheets: A Study of
Twenty Countries, 1688-1978 (Univ. of Chicago Press, Chicago,
premium for education has risen across a large magnify inequality and erode real wages among
IL, 1985) number of advanced countries in recent dec- less educated workers. As I discuss below, the
23. T. Piketty, G. Zucman, Q. J. Econ. 129, in press ades, and this rise contributes substantially to foremost concern raised by these multiple forces
(2014); http://piketty.pse.ens.fr/files/ the net growth of earnings inequality. In the is not their impact on inequality per se, but
PikettyZucman2013WP.pdf.
24. T. Piketty, Capital in the Twenty-first Century (Harvard Univ.
United States, for example, about two-thirds rather their adverse effect on the real earnings
Press, Cambridge, MA, 2014). of the overall rise of earnings dispersion be- and employment of less educated workers.
25. J. Stiglitz, Econometrica 37, 382–397 (1969). tween 1980 and 2005 is proximately accounted I begin by documenting the centrality of the
26. M. Nirei, “Pareto Distributions in Economics Growth Models,” for by the increased premium associated with rising skill premium to the overall growth of
Institute of Innovation Research Working Paper No. 09-05,
Hitotsubashi University, Tokyo (2009)
schooling in general and postsecondary edu- earnings inequality. I next consider why skills
27. T. Piketty, G. Postel-Vinay, J. L. Rosenthal, Explor. Econ. Hist. cation in particular (1, 2). Second, despite a are heavily rewarded in advanced economies
51, 21–40 (2014). lack of consensus among economists regard- and why the demand for them has risen over
28. J. Davies, S. Sandstrom, T. Shorrocks, E. Wolff, Econ. J. 121, ing the primary causes of the rise of very top time. I then demonstrate the substantial ex-
223–254 (2011).
29. G. Zucman, Q. J. Econ. 128, 1321–1364 (2013).
incomes (3–6), an influential literature finds planatory power of a simple framework that
30. C. Goldin, L. Katz, The Race Between Education and Technology that the interplay between the supply and embeds both the demand and supply for skills
(Harvard Univ. Press, Cambridge, MA, 2008). demand for skills provides substantial insight in interpreting the evolution of the inequality
into why the skill premium has risen and fallen over five decades. The final section considers
SUPPLEMENTARY MATERIALS
over time—and, specifically, why the earnings the productive role that inequality plays in a
www.sciencemag.org/content/344/6186/838/suppl/DC1
market economy and the potential risks attend-
Supplementary Text ing very high and rising inequality; evidence on
Department of Economics and National Bureau of Economic
Figs. S1 and S2 whether those risks have been realized; and
Research, Massachusetts Institute of Technology, 40 Ames
References (31, 32) the role of policy and governance in encour-
Street, E17-216, Cambridge, MA 02142, USA. E-mail: dautor@
10.1126/science.1251936 mit.edu aging skills formation, fostering opportunity,

and countering the possibility that extremes of The median earnings comparisons in Fig. 1 also education can account for 55% of the rise in
inequality erode economic mobility and reduce convey a key feature of rising inequality that male hourly wage variance from 1973–1975 to
economic dynamism. cannot be inferred from trends in top incomes: 2003–2005. Firpo et al. (13) found that rising
Wage inequality has risen throughout the earn- returns to education can explain just over 95% of
The Critical Role of Skills in the ings distribution, not merely at the top percent- the rise of the U.S. male 90/10 earnings ratio be-
Labor Market iles. Figure S1 documents this pattern by plotting, tween 1984 and 2004. That is, holding the ex-
There is no denying the extraordinary rise in for 12 Organization for Economic Cooperation panding education premium constant over this
the incomes of the top 1% of American house- and Development (OECD) member countries over period, there would have been essentially no in-
holds over the past three decades. Between three decades (1980 to 2011), the change in the crease in the relative wages of the 90th-percentile
1979 and 2012, the share of all household in- ratio of full-time earnings of males at the 90th worker versus the 10th-percentile worker.
come accruing to the top percentile of U.S. percentile relative to males at the 10th percent- I have so far used the terms education and
households rose from 10.0% to 22.5% (8, 9). To ile of the wage distribution. Although the 90/10 skill interchangeably. What evidence do we have
get a sense of how much money that is, con- earnings ratio differed greatly across countries that it is skills that are rewarded per se, rather
sider the conceptual experiment of redistri- at the earliest date of the sample—from a low than simply educational credentials? The Pro-
buting the gains of the top 1% between 1979 and of 2.0 in Sweden to a high of 3.6 in the United gram for the International Assessment of Adult
2012 to the bottom 99% of households (10). States—this earnings ratio increased substan- Competencies (PIAAC) provides a compelling
How much would this redistribution raise house- tially in all but one of them (France) over the data source for gauging the importance of
hold incomes of the bottom 99%? The answer next 30 years, growing by at least 25 percentage skills in wage determination. The PIAAC is an
is $7107 per household—a substantial gain, equal points in 10 countries, by at least 50 percentage internationally harmonized test of adult cog-
to 14% of the income of the median U.S. house- points in 8 countries, and by more than 100 per- nitive and workplace skills (literacy, numeracy,
hold in 2012. (I focus on the median because it centage points in three countries (New Zealand, and problem-solving) that was administered
reflects the earnings of the typical worker and the United Kingdom, and the United States). by the OECD to large, representative samples
thus excludes the earnings of the top 1%.) How much does the rising education premium of adults in 22 countries between 2011 and
Now consider a different dimension of in- contribute to the increase of earnings inequality? 2013 (14). Figure 2, sourced from (15), plots the
equality: the earnings gap between U.S. work- Although data limitations make it difficult to relationship between adults’ earnings and their
ers with a 4-year college degree and those with answer this question for most countries, we do PIAAC numeracy scores across these 22 coun-
only a high school diploma (11). Economists fre- know the answer for the United States. Goldin tries. The length of each bar reflects the av-
quently use this college/high school earnings and Katz (1) found that the increase in the edu- erage percentage earnings differential between
gap as a summary measure of the “return to cation wage premium explains about 60 to 70% full-time workers ages 35 to 54 who differ by
skill”—that is, the gain in earnings a worker of the rise in the dispersion of U.S. wages be- one standard deviation in the PIAAC score.
can expect to receive from investing in a col- tween 1980 and 2005 and, similarly, Lemieux (12) The whiskers on each bar provide the 95%
lege education. As illustrated in Fig. 1, the earn- calculated that higher returns to postsecondary confidence intervals for the estimates.
ings gap between the median college-educated
and median high school–educated among U.S.
males working full-time in year-round jobs was
$17,411 in 1979, measured in constant 2012 dol- College/high school median annual earnings gap, 1979–2012
lars. Thirty-three years later, in 2012, this gap In constant 2012 dollars
had risen to $34,969, almost exactly double its
1979 level. Also seen is a comparable trend among
U.S. female workers, with the full-time, full- Household gap
year college/high school median earnings gap 70,000 dollars
$30,298 to $58,249
nearly doubling from $12,887 to $23,280 be-
tween 1979 and 2012. As Fig. 1 underscores, the 60,000
economic payoff to college education rose stead-
ily throughout the 1980s and 1990s and was 50,000
barely affected by the Great Recession starting Male gap
$17,411 to $34,969
in 2007. 40,000
Because the earnings calculations in Fig. 1 re-
flect individual incomes while the top 1% cal- 30,000
culations reflect household incomes, the two
calculations are not directly comparable. To
20,000
put the numbers on the same footing, consider
the earnings gap between a college-educated Female gap
10,000 $12,887 to $23,280
two-earner husband-wife family and a high school–
educated two-earner husband-wife family, which
0
rose by $27,951 between 1979 and 2012 (from
1979 1982 1985 1988 1991 1994 1997 2000 2003 2006 2009 2012
$30,298 to $58,249). This increase in the earn-
ings gap between the typical college-educated
and high school–educated household earn-
ings levels is four times as large as the redis- Fig. 1. College/high school median annual earnings gap, 1979–2012. Figure is constructed using
tribution that has notionally occurred from Census Bureau P-60 (1979–1991) and P-25 (1992–2012) tabulations of median earnings of full-time,
the bottom 99% to the top 1% of households. full-year workers by educational level and converted to constant 2012 dollars (to account for
What this simple calculation suggests is that inflation) using the CPI-U-RS price series. Prior to 1992, college-educated workers are defined as
the growth of skill differentials among the “other those with 16 or more years of completed schooling, and high school–educated workers are those
99 percent” is arguably even more consequen- with exactly 12 years of completed schooling. After 1991, college-educated workers are those who
tial than the rise of the 1% for the welfare of report completing at least 4 years of college, and high school–educated workers are those who
most citizens. report having completed a high school diploma or GED credential.

This figure conveys three points. First, cog- for skill in the labor market—specifically, why infrastructure. This was not always the case. In
nitive skills are substantially rewarded in the labor they fluctuate over time and how their inter- 1900, 4 in 10 U.S. jobs were in agriculture, 11%
market across all 22 economies. The average action helps to determine the skill premium. I of the population was illiterate, a substantial
wage premium corresponding to one “unit” (i.e., focus on the United States in this section to al- fraction of economic activity required hard phys-
one standard deviation) increase in measured low a deeper exploration of the data. ical labor, and workers’ strength and physical
cognitive skills is 18%. In addition, cognitive earn- stamina were key job skills (17, 18). Few citizens
ings premiums differ substantially across coun- Education and Inequality would have predicted at the time that a cen-
tries. The premium is below 13% in Sweden, Workers’ earnings in a market economy de- tury later, health care, finance, information tech-
the Czech Republic, and Norway. It is above pend fundamentally (some economists would nology, consumer electronics, hospitality, leisure,
20% in six countries. The United States stands say entirely) on their productivity—that is, the and entertainment would employ far more work-
out as having the highest measured return to value they produce through their labor. And in ers than agriculture—which employed only 2%
skill, with a premium of 28% per unit increment turn, workers’ productivity depends on two fac- of U.S. workers in 2010. As physical labor has
to cognitive ability. Concretely, comparing two tors. One is their capabilities, concretely, the given way to cognitive labor, the labor market’s
U.S. workers who are one standard deviation tasks they can accomplish (i.e., their skills). A demand for formal analytical skills, written com-
above and one standard deviation below the second is their scarcity: The fewer workers that munications, and specific technical knowledge—
population average of cognitive ability, we would are available to accomplish a task, and the more what economists often loosely term cognitive
expect their full-time weekly earnings to dif- employers need that task accomplished, the skills—has risen spectacularly.
fer by 50 to 60%. Notably, the high return to higher is workers’ economic value in that The central determinant of the supply of
cognitive ability in the United States does not task. In conventional terms, the skill premium skills available to an advanced economy is its
follow automatically from high levels of U.S. depends upon what skills employers require (skill education system. In 1900, the typical young,
earnings inequality. If U.S. wages were deter- demand) and what skills workers have acquired native-born American had only a common school
mined mainly by luck, beauty, or family con- (skill supply). To interpret the evolution of this education, about the equivalent of six to eight
nections, we would expect little connection premium, we need to account for both forces. grades (19). By the late 19th century, however,
between workers’ cognitive ability and their la- many Americans recognized that farm employ-
bor market rewards (16). Figure 2 demonstrates Skill Demands: The Long View ment was declining, industry was rising, and
that this is not the case. A technologically advanced economy requires their children would need additional education
Of course, these data do not explain why a literate, numerate, and technically and scien- to earn a living. Over the first four decades of the
the skill premium has risen over time, nor tifically trained workforce to develop ideas, man- 20th century, the United States became the first
why the United States has a higher skill pre- age complex organizations, deliver healthcare nation in the world to deliver universal high
mium than so many other advanced nations. services, provide financing and insurance, ad- school education to its citizens. Tellingly, the high
The next section considers the supply and demand minister government services, and operate critical school movement was led by the farm states.
As the high school movement reached its
Fig. 2. Cross-national differences conclusion, postsecondary education became
in wage returns to skills, increasingly indispensable to the growing oc-
2011–2013. Reproduced with Cross-national differences in wage returns cupations of medicine, law, engineering, sci-
ence, and management. In 1940, only 6% of
permission from Hanushek et al. to skills, 2011–2013 Americans had completed a 4-year college
[(15), table 2]. Estimates are
Percentage increase for a one standard deviation degree. From the end of the Second World
obtained by regressing the increase in skill
natural logarithm of workers’ War to the early 1980s, however, the ranks of
weekly full-time earnings on test college-educated workers rose robustly and
Sweden Earnings steadily, with each cohort of workers enter-
scores while controlling for sex
Czech R. gain ing the labor market boasting a proportion-
and labor market experience
Norway ately higher rate of college education than
(both a linear and a quadratic 95% confidence
term). Regression estimates are Italy interval the cohort that preceded it. This intercohort
performed separately for each Denmark pattern, which was abetted by the Second
country and test scores are World War and Korean War GI Bills (20) and
Cyprus
normalized with mean zero and by huge state and federal investments in pub-
Finland lic college and university systems, is depicted in
unit standard deviation within
Belgium Fig. 3A. From 1963 through 1982, the fraction
each country. Estimates that
France of all U.S. hours worked that were supplied
normalize test scores on a
common basis across countries, Estonia by college graduates rose by almost 1 percentage
or that use literacy or Slovak R. point per year, a remarkably rapid gain.
problem-solving scores rather After 1982, however, the rate of intercohort
Austria
than numeracy scores, increase fell by almost half—from 0.87 percentage
Netherlands points to 0.47 percentage points per year—and
yield qualitatively similar patterns.
Japan did not begin to rebound until 2004, nearly
Poland two decades later. As shown in fig. S2, this de-
Canada celeration in the supply of college graduates is
Korea particularly stark when one focuses on young
adults with fewer than 10 years of experience—
U.K.
that is, the cohorts of recent labor market
Spain entrants at each point in time. Although the
Germany supply of young college-educated males rela-
Ireland tive to young high school–educated males in-
U.S. creased rapidly in the 1960s and early 1970s
(and indeed throughout the postwar period), this
0 5 10 15 20 25 30 percent rising tide reached an apex in 1974 from which

it barely budged for the better part of the next graduates in 1982 but were double those of high effect of these forces is to further raise the de-
30 years. Among young females, the deceleration school graduates by 2005. mand for formal education, technical expertise,
in supply was also unmistakable, although not as Why is this deceleration in supply relevant and cognitive ability (23–27).
abrupt or as complete as for males. to the college premium? After all, although the
The counterpart to this deceleration in the growth of supply slowed in 1982, it was still Bringing the Supply-Demand
growth of supply of college-educated workers rising. A likely answer is that the demand for Framework to the Data
is the steep rise in the college premium com- college workers rose in the interim. Through- The persistently rising demand for educated
mencing in the early 1980s and continuing for out much of the 20th century, successive waves labor in advanced economies was first noted
25 years. Concretely, when the influx of new of innovation—electrification, mass production, by the Nobel Prize–winning economist Jan
college graduates slowed, the premium that a motorized transportation, telecommunications— Tinbergen (28) and is often referred to as the
college education commanded in the labor mar- have reduced the demand for physical labor “education race” model (19). Its primary im-
ket increased. The critical role played by the and raised the centrality of cognitive labor in plication is that if the supply of educated labor
fluctuating supply of college education in the practically every walk of life. The past three does not keep pace with persistent outward
rise of U.S. inequality is documented in Fig. 3B, decades of computerization, in particular, have shifts in demand for skills, the skill premium
which plots the college wage premium from extended the reach of this process by displac- will rise. In the words of the Red Queen in
1963 through 2012 (blue line). This premium ing workers from performing routine, codifiable Lewis Carroll’s Alice in Wonderland, “…it takes
fluctuated in a comparatively narrow band dur- cognitive tasks (e.g., bookkeeping, clerical work, all the running you can do, to keep in the same
ing the 1960s and 1970s, as rising demand for and repetitive production tasks) that are now place.” Thus, when the rising supply of edu-
educated workers was met with rapidly rising readily scripted with computer software and cated labor began to slacken in the early 1980s,
year-over-year increases in supply. In 1981, the performed by inexpensive digital machines. This a logical economic consequence was an increase
average college graduate earned 48% more per ongoing process of machine substitution for rou- in the college skill premium.
week than the average high school graduate—a tine human labor complements educated work- To more formally account for the impact of
significant earnings gap but not an earnings ers who excel in abstract tasks that harness the fluctuating growth rate of supply of college-
gulf. When the supply deceleration began in problem-solving ability, intuition, creativity, and educated workers on the college wage differen-
1982, however, the college premium hit an in- persuasion—tasks that are at present difficult tial, Fig. 3B depicts the fit of a simple regression
flection point. This premium notched remark- to automate but essential to perform. Simulta- model that predicts the college wage premium
ably rapid year-over-year gains from 1982 forward, neously, it devalues the skills of workers, typ- in each year as a function of two factors: (i) the
reaching 72% in 1990, 90% in 2000, and 97% in ically those without postsecondary education, contemporaneous supply of college graduates,
2005 (21, 22). Thus, the average earnings of college who compete most directly with machinery in and (ii) a time trend, which serves as a proxy for
graduates were 1.5 times those of high school performing routine-intensive activities. The net the secularly rising demand for college-educated
The supply of college graduates and the U.S. college/high school premium, 1963–2012
College share of hours worked (%), 1963–2012: College versus high school
All working-age adults wage gap (%)
A B
50 percent 105 percent
Measured Gap
45 95
40 85
35 75 Predicted by Supply-
Demand Model
30 65
25 55
20 45
35
15
1964 1970 1976 1982 1988 1994 2000 2006 2012 1964 1970 1976 1982 1988 1994 2000 2006 2012
Fig. 3. The supply of college graduates and the U.S. college/high school are aggregated using CPS sampling weights. (B) College versus high school
premium, 1963–2012. (A) College share of hours worked in the United wage gap. Figure uses March CPS data for earnings years 1963 to 2012. The
States, 1963–2012: All working-age adults. Figure uses March CPS data for series labeled “Measured Gap” is constructed by calculating the mean of
earnings years 1963 to 2012. The sample consists of all persons aged 16 to the natural logarithm of weekly wages for college graduates and non–
64 who reported having worked at least 1 week in the earnings years, college graduates, and plotting the (exponentiated) ratio of these means for
excluding those in the military. Following an extensive literature, college- each year. This calculation holds constant the labor market experience and
educated workers are defined as all of those with four or more completed gender composition within each education group. The series labeled
years of college plus half of those with at least 1 year of completed college. “Predicted by Supply-Demand Model” plots the (exponentiated) predicted
Non-college workers are defined as all workers with high school or less values from a regression of the log college/noncollege wage gap on a
education, plus half of those with some completed college education. For quadratic polynomial in calendar years and the natural log of college/
each individual, hours worked are the product of usual hours worked per noncollege relative supply. See text and supplementary material for further
week and the number of weeks worked last year. Individual hours worked details.

workers (29). Comparing the fitted values (red boosted college attendance. This created some- One should not, of course, take this model as
series) from this simple supply-demand model thing of a glut of college enrollments in the late irrefutable. A puzzling pattern evident in the
alongside the actual data (blue series) reveals an 1960s and early 1970s, which in turn depressed data is that the rising demand for skilled workers
extremely tight correspondence over the course the college earnings premium in the 1970s (see appears to have slowed in the early 1990s, a
of five decades and three distinct eras: a declining Fig. 3) and likely reduced the attractiveness of phenomenon that is not anticipated by the “edu-
skill premium in the 1970s; an explosive rise in college-going absent the military draft. Thus, when cation race” model (37). This discrepancy un-
the premium during the 1980s, 1990s, and early the war ended in the early 1970s, college enroll- derscores that the supply-demand model is
2000s; and, most recently, a plateau commencing ment rates dropped sharply, particularly among necessarily incomplete—in part for the sake of
after 2005. A key implication of this figure is that males. The fall in enrollment produced a corre- expositional clarity and, in larger part, because
a central causal factor behind rising inequality sponding decline in college completions half a our understanding of macroeconomic phenom-
in the United States has been the slowdown in decade later, and a surge of inequality followed. ena is typically imperfect. Nevertheless, the data
the accumulation of skills by young adults almost This supply-demand explanation for the rise of speak sufficiently clearly to warrant two eco-
30 years ago. Had the supply of college graduates U.S. inequality may appear almost too simple to nomic inferences. The first is that although pop-
risen as rapidly in the decades after 1980 as it did be credible. After all, we are comparing just two ular accounts frequently assert that the United
in the decades immediately before, it is quite plau- economic variables: the college wage premium States is in the midst of a “college bubble”—too
sible that there would have been no sustained and the supply of college graduates in the U.S. many students going to college at too high a
rise in the skill premium in the U.S. labor market. workforce. But a host of rigorous studies com- cost—abundant economic evidence strongly sug-
Of course, this set of facts raises another puz- mencing with Katz and Murphy (31) confirm the gests otherwise. Yes, college tuitions have risen
zle: If slackening college supply sparked rising remarkable explanatory power of this simple far faster than inflation, and indeed, student
inequality, what caused rising U.S. postsecondary supply-demand framework for explaining trends debt has risen rapidly, with more than $100
achievement to grind to a sudden halt in 1982? in the college versus high school earnings gap billion in federal student aid dollars loaned in
Work by Card and Lemieux (30) highlights that over the course of nine decades of U.S. history, as 2012–2013 alone (38). But the doubling of the
one critically important factor was the United well as across other industrialized economies (most college weekly wage differential over the past
States’ involvement in the Vietnam War. Because notably, the United Kingdom and Canada) and 30 years also implies that there have been sizable
draft-eligible males in the Vietnam era were often among age and education groups within countries increases in the lifetime earnings of college grad-
able to defer their military service by enrolling (19, 31–36). The United States was far from the uates relative to high school graduates. How large
in postsecondary schooling, the war artificially only Western country to experience this surge. are these gains? Figure 4, reproduced from (39),
reports the estimated lifetime college earnings
differential net of tuition for cohorts of students
entering the labor market between 1965 and
Present discounted value of college relative to high school degree 2008. For both males and females, the expected
net of tuition, 1965–2008 net present value of a college degree relative to a
high school diploma roughly tripled in this pe-
College/high school difference, 2009 dollars
590K riod, with the fastest gains accruing during the
582K
1980s and 1990s. Note that this growing college/
600,000 dollars Men high school gap reflects the rising payoff to the
4-year college degree, the even steeper rise in the
439K premium associated with graduate and profession-
500,000
385K al degrees (see below), and the growing fraction of
368K
college graduates who obtain higher degrees; thus,
400,000 an additional payoff to the college degree is that
261K Women it opens the door to further specialization. This
300,000 213K lifetime earnings differential would, of course, have
387K
370K risen further still if college tuitions had held steady
rather than rising. But the inevitable sticker shock
200,000 284K that households feel when confronting the cost of
college should not obscure the fact that the real
225K
100,000 198K lifetime earnings premium to college education
has likely never been higher (40).
129K 138K The second positive economic news im-
0 plied by Fig. 3 above is that the ongoing rise
1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 of skill differentials is not inevitable. Prior co-
horts of U.S. students, particularly males, were
slow to react to the rising return to education
Fig. 4. Present discounted value of college relative to high school degree net of tuition, during the 1980s and 1990s, but the message
1965–2008. Reproduced from Avery and Turner with permission of the American Economic As- appears to have finally gotten through. Dur-
sociation (39). Expected earnings are calculated from the March Current Population Survey files ing the first decade of the 21st century, the U.S.
for full-time, full-year workers using sample weights. The estimates equal what a man or woman high school graduation rate rose sharply after
would expect to earn working full-time, full-year over a career of 42 years, with a discount rate of having been essentially stagnant since the late
3%, assuming that college graduates delay the start of earnings for 4 years while in school. 1960s (41). This unanticipated rise was followed
Earnings expectations are formed in each year by assuming that future high school and college just a few years later by a surge in college com-
graduates will have future earnings at each age equal to the average earnings of high school and pletions. Between 2004 and 2012, the supply of
college graduates (respectively) currently observed at each age; for example, expected earnings in new college graduates to the U.S. labor market
1980 are based on data across ages for 1980. Results for college-educated workers are net of 4 years rose at a rate not seen in several decades (Fig.
of tuition and fees associated with appropriate year-specific values for public universities. Plotted 3A). As this influx of supply took hold, the col-
points show the difference between expected earnings for college graduates and for high school lege wage premium halted its enduring rise (Fig.
graduates. 3B). What these observations and our simple

supply-demand model suggest is that the flat- concern is that high inequality at a point in time Indeed, two of the strongest predictors of child-
tening of the college premium after 2005 is in may serve to reduce mobility over time. If, for ren’s ultimate educational attainment are pa-
large part a consequence of the quickening pace example, adults who became wealthy through rental education and parental earnings (45, 46).
of educational attainment. hard work are able to “buy” success for their Hence, when the return to education is high,
children through outsized investments and per- children of better-educated parents are doubly
Inequality: Causes for Concern? sonal connections, while adults who are unpro- advantaged—by their parents’ higher education
A market economy needs some inequality to ductive or unlucky in their careers are unable and higher earnings—in attaining greater edu-
create incentives. If, for example, students were to muster the resources to foster their children’s cation while young and greater earnings in
not ultimately rewarded for spending their early potential, then inequality of incomes could be- adulthood. Figure 5 therefore lends credence
adulthoods pursuing undergraduate, graduate, come self-perpetuating even if it originally ema- to the concern that rising inequality may erode
and professional degrees, or if the hardest-working nates from high market returns to skill (43). economic mobility.
and most productive workers were paid the To understand the importance of high and Has this erosion occurred? Surprisingly, the
same as the median worker, then citizens would rising U.S. inequality, it is therefore useful to best evidence to date suggests that it has not.
have little incentive to develop expertise, to exert ask how U.S. economic mobility compares to Evidence from Chetty et al. (46), documented
effort, or to excel in their work (42). Having ac- that of other developed countries, and whether in the supplementary material, underscores the
knowledged that some inequality is necessary, U.S. mobility has fallen as inequality has risen. message from Fig. 5 that there is substantial
however, how can we gauge whether there is The answers to both questions will surprise economic immobility in the United States. Chil-
too much of it? I offer two analytical perspec- many. Contrary to conventional civic mythol- dren born three deciles apart in the household
tives on this question. ogy, U.S. intergenerational mobility is relative- income distribution are on average one decile
ly low. The left panel of Fig. 5, reproduced from apart in the earnings distribution at age 29 or
Earnings Mobility (44), which plots the relationship between cross- 30. Similarly, children born three deciles apart
One metric by which to evaluate the conse- sectional inequality (x axis) and earnings mobil- in the household income distribution differ by
quences of inequality is via its relationship ity (y axis) among a set of 13 OECD member 20 percentage points in their probability of at-
with economic mobility—that is, the degree to countries for which consistent data are available, tending college at age 19 (relative to a mean of
which individual economic fortunes change documents that the United States has both the approximately 55%). Yet these data offer no
over time. Of particular interest is the degree of lowest mobility and highest inequality among evidence that mobility has appreciably changed
intergenerational mobility, meaning the likeli- all wealthy democratic countries. The right panel among children born prior to the historic rise
hood that children born to low-income fami- of Fig. 5, also sourced from (44), suggests one of U.S. inequality (1971–1974) and those born
lies become high-income adults and vice versa. proximate explanation for this pattern: Coun- afterward (1991–1993). As far as we can measure,
High levels of economic inequality at a given tries with high returns to education tend to rising U.S. income inequality has not reduced
point in time are not intrinsically inimical to have relatively low mobility. Why, if education is intergenerational mobility so far. These find-
economic mobility; a society with high inequal- “the great equalizer” in the words of Horace ings, which also appear to hold over a longer
ity may be dynamic, with lots of movement up Mann, do high educational returns predict low historical time frame (47), suggest that U.S.
and down the economic ladder, and one with mobility? A key reason is that educational at- mobility has not trended downward as many
low inequality may be dynastic. But a natural tainment is highly persistent within families. social scientists would have anticipated, and as
Earnings inequality and economic mobility: cross-national relationships

Generational earnings elasticity Generational earnings elasticity
(higher values imply lower mobility) (higher values imply lower mobility)
United
A B
0.5 Kingdom 0.5
Italy Italy United Kingdom
United States United States
France France
Spain
0.4 0.4
Japan
Germany
New Zealand Germany
0.3 0.3
New Zealand Australia
Sweden Sweden
Australia Canada
Canada
0.2 Finland 0.2
Norway Norway Finland
Denmark Denmark
0.1 0.1
20 25 30 35 100 120 140 160 180
Income inequality (more inequality ) College earnings premium (men 25 to 34)
Fig. 5. Earnings inequality and economic mobility: Cross-national rela- mobility. In the left panel, cross-sectional income inequality is measured
tionships. Reproduced from Corak [(44), figs. 1 and 4] with permission of using a “Gini” index that ranges from 0 to 100, where 0 indicates complete
the American Economic Association. In both panels, the mobility measure is equality of household incomes and 100 indicates maximal inequality (all
equal to the intergenerational earnings “elasticity,” meaning the average income to one household). In the right panel, the college earnings premium
proportional increase in a son’s adult earnings predicted by his father’s refers to the ratio of average earnings of men 25 to 34 years of age with a
adult earnings measured approximately three decades earlier. A higher in- college degree to the average earnings of those with a high school diploma,
tergenerational earnings elasticity therefore implies lower intergenerational computed by the OECD using 2009 data. See (44) for further details.

many policymakers and popular accounts fre- conveys the positive economic news that educa- collar office, clerical, and administrative support
quently assume. tional investments offer large returns, this wage positions, and has reduced the set of middle-
It is important to interpret these results in premium also masks a discouraging truth: The skill career jobs available to non–college-educated
context. The most recent birth cohorts whose rising relative earnings of workers with post- workers more generally (25). A second factor
adult outcomes can be observed at present secondary education is not simply due to rising is the globalization of labor markets, seen par-
were born no later than the early 1990s, which real earnings among college-educated workers ticularly in the greatly increased U.S. trade
is still relatively early in the rise of U.S. in- but is also due to falling real earnings among non– integration with developing countries. Global-
equality. Another 10 years of data, focusing college-educated workers. Between 1980 and ization has become particularly important for
on children born since 2000, may suggest a 2012, real hourly earnings of full-time college- U.S. labor markets since the early 1990s, when
different conclusion. Moreover, the fact that educated U.S. males rose anywhere from 20% to China began its extremely rapid integration
mobility has stayed constant while inequality 56%, with the greatest gains among those with into the world trading system. The influx of
has risen means that the lifetime relative dis- a postbaccalaureate degree (Fig. 6A). During the Chinese goods lowered consumer prices but
advantage of children born to low- versus high- same period, real earnings of males with high also fomented a substantial decline in U.S. man-
income families has increased substantially; school or lower educational levels declined substan- ufacturing employment, contributing directly
concretely, the rungs of the economic ladder tially, falling by 22% among high school dropouts to the decline in production worker employment
have pulled farther apart but the chance of and 11% among high school graduates. Although (50). A third factor impinging on the earnings
ascending the ladder has not improved. Fi- the picture is generally brighter for females (Fig. of non–college-educated males is the decline in the
nally, it is possible to interpret the fact that 6B), real earnings growth among females with- penetration and bargaining power of labor unions
mobility has remained unchanged as evidence out at least some college education over this three- in the United States, which have historically
that U.S. mobility would have declined had it decade interval was extremely modest. obtained relatively generous wage and benefit
not been for the other compensatory steps Accompanying the fall in real wages among packages for blue-collar workers. Over the past
taken by the federal government during this less educated workers has been a pronounced three decades, however, U.S. private-sector union
period, including, for example, expanding the drop in their labor force participation rates, density—that is, the fraction of private-sector
Earned Income Tax Credit for low-income work- particularly among less educated males. Be- workers who belong to labor unions—has fallen
ers in the 1980s, enlarging the early childhood tween 1979 and 2007, prior to the onset of the by approximately 70%, from 24% in 1973 to 7% in
education Head Start program in the 1990s, Great Recession, the fraction of working-age 2011 (51, 52).
and increasing federal student grant and loan males in paid employment fell by 12 percentage Notably, these three forces—technological
programs to support college-going (48). Declines points among high school dropouts and 10 per- change, deunionization, and globalization—
in racial and gender discrimination during this centage points among those with exactly a high work in tandem. Advances in information and
period likely also complemented these policies school diploma. Conversely, employment rates were communications technologies have directly
(49). A cautious read of the evidence is that al- generally stable for males with postsecondary changed job demands in U.S. workplaces while
though the United States is not a “land of oppor- education and rose for females of all education simultaneously facilitating the globalization of
tunity” by conventional economic mobility metrics, levels except for high school dropouts. production by making it increasingly feasible
it has not become less so in recent decades. The causes for the sharp falls in real earnings and cost-effective for firms to source, monitor,
among non–college-educated workers are mul- and coordinate complex production processes
Real Earnings tiple. One likely force, as noted above, is the at disparate locations worldwide. In turn, the
A second gauge of economic health is the tra- ongoing substitution of computer-intensive ma- globalization of production has increased com-
jectory of earnings and employment. Here, the chinery for workers performing routine task- petitive conditions for U.S. manufacturers and
data present substantial cause for concern. Al- intensive jobs. This has depressed demand for U.S. workers, eroding employment at unionized
though the substantial college wage premium workers in both blue-collar production and white- establishments and decreasing the capability
Changes in real wage levels of full-time U.S. workers by sex and education, 1963–2012
Real weekly earnings relative to 1963 (men) Real weekly earnings relative to 1963 (women)
A B
2.0 2.0 Some college
Bachelor's
1.8 degree 1.8 High school
> Bachelor's graduate
degree High school
1.6 1.6 dropout
1.4 1.4
1.2
1.2
1.0
1.0
1964 1968 1972 1976 1980 1984 1988 1992 1996 2000 2004 2008 2012 1964 1968 1972 1976 1980 1984 1988 1992 1996 2000 2004 2008 2012
Fig. 6. Change in real wage levels of full-time workers by education, 1963–2012. (A) Male workers, (B) female workers. Data and sample construction are
as in Fig. 3.

of unions to negotiate favorable contracts, attract policy levers for shaping inequality’s trajectory 14. See www.oecd.org/site/piaac/surveyofadultskills.htm for more
new members, and penetrate new establishments. and socioeconomic consequences. Policies that information. The PIAAC program will encompass 33 countries, but
data for only 22 were available at this writing.
In all cases, the foremost concern raised by appear most effective over the long haul in rais-
15. E. A. Hanushek, G. Schwerdt, S.Wiederhold,
these multiple forces impinging on the earnings ing prosperity and reducing inequality are those L.Woessmann, Returns to Skills Around the World:
of workers at different skill levels is not their that cultivate the skills of successive generations: Evidence from PIAAC (NBER Working Paper 19762,
impact on inequality per se, but rather their ad- excellent preschool through high school education; Cambridge, MA, 2013).
16. Hanushek et al. (15) also found that the correlation between
verse effect on the real earnings and employment broad access to postsecondary education; and
numeracy skills and years of schooling is 0.45. When including
of less educated workers. These declines in both good nutrition, good public health, and high- both numeracy skills and years of schooling in an earnings
earnings and employment bode ill for the welfare quality home environments. Such policies address regression, they found that both are substantial and significant
of non–college-educated U.S. adults and are likely inequality from two directions: (i) enabling a larger predictors of earnings, although each is attenuated relative
to a model where only one factor is included at a time. This
to have broader detrimental social consequences fraction of adults to attain high productivity,
pattern of results suggests, logically, that neither test scores
that frequently accompany non-employment: rewarding jobs, and a reasonable standard of nor years of schooling is a complete measure of labor
greater criminality, increased social dependency, living; and (ii) raising the total supply of skills market skills.
and (more mundanely) reduced tax receipts. available to the economy, which in turn moderates 17. T. D. Snyder, 120 Years of American Education: A Statistical
the skill premium and reduces inequality (54). Portrait (National Center for Education Statistics, U.S.
Do Supply and Demand Make Department of Education, 1993).
Of course, building skills is a multigenera- 18. L. D. Johnston, “History lessons: Understanding
Policy Irrelevant? tional process and thus has little impact on in- the decline in manufacturing.” MinnPost, 22 February
One potential interpretation of the evidence equality in the short term. There are, however, 2012; www.minnpost.com/macro-micro-minnesota/
above is that, because rising inequality is sub- numerous nearer-term levers that moderate 2012/02/history-lessons-understanding-decline-
manufacturing.
stantially a consequence of the impersonal forces inequality directly without imposing substan- 19. C. Goldin, L. F. Katz, The Race Between Education and
of supply and demand, public policy has no role tial economic costs: applying progressive tax and Technology (Harvard Univ. Press, Cambridge, MA, 2008).
to play in shaping the trajectory of inequality or transfer policies that fund public investments 20. M. Stanley, Q. J. Econ. 118, 671–708 (2003).
its social impact. This conclusion is incorrect for and foster opportunities for children of all socio- 21. B. Pierce, in Labor in the New Economy, K. G. Abraham,
J. R. Spletzer, M. Harper, Eds. (Univ. of Chicago Press, Chicago,
two reasons. First, there are multiple channels economic backgrounds; applying well-crafted 2010), pp. 63–98.
by which policy has contributed to the rise of labor regulations that ensure safe and non- 22. These comparisons hold labor market experience and gender
U.S. inequality, many of which are not fully exploitive working conditions; providing wage constant. This doubling of the college premium very likely
evident in the education earnings premium. subsidies such as the Earned Income Tax Credit understates the magnitude of the increase in inequality
between college-educated and non–college-educated workers.
These include the fall over several decades in that increase the payoff to employment for those
Alongside higher hourly earnings, college-educated workers
the real value of the U.S. minimum wage (7); the with limited skills; setting modest but nonzero enjoy greater job stability, lower rates of unemployment, more
declining prevalence and bargaining power minimum wage rules; and offering numerous generous fringe benefits, and better working conditions;
of U.S. labor unions; mounting international social insurance policies (health and disability Pierce (21) found that these differentials have generally increased
in the same time period.
competition that places particular pressure on insurance, flood insurance, disaster assistance,
23. D. H. Autor et al., Q. J. Econ. 118, 1279–1333 (2003).
the wages and employment of less educated food assistance) that buffer misfortune for the 24. D. Acemoglu, D. H. Autor, Skills, tasks and technologies:
workers; and sharp reductions in top federal unfortunate. Although it is outside the scope of Implications for employment and earnings. In Handbook of
marginal tax rates that have raised after-tax this article to evaluate these policies, it is crit- Labor Economics, D. Card, O. Ashenfelter, Eds. (Elsevier-North
inequality and increased the incentive of highly ical to underscore that policy and governance Holland, Amsterdam, 2011), vol. 4, pp. 1043–1171.
25. D. H. Autor, D. Dorn, Am. Econ. Rev. 103, 1553–1597 (2013).
paid workers to seek still higher compensa- has played and should continue to play a central 26. M. Goos et al., www.aeaweb.org/forthcoming/output/
tion. As discussed in the companion paper by role in shaping inequality—even when a central accepted_AER.php
Piketty and Saez, there is also disagreement cause of rising inequality is the changing supply 27. Extensive recent literature, commencing with Autor et al. (23)
among economists about whether the rising and demand for skills. and summarized in Acemoglu and Autor (24), considers the
role of technological change in displacing workers performing
share of household incomes accruing to the top routine tasks and complementing workers performing
few percentiles of households in numerous nonroutine tasks. An additional implication of this framework is
developed countries over the past several dec- RE FERENCES AND NOTES that an increasing share of employment will be found in
ades is also primarily a market phenomenon, or 1. C. D. Goldin, L. F. Katz, Brookings Pap. Econ. Act. (fall), 135 (2007). comparatively low-skill nonroutine manual tasks that require
2. Goldin and Katz (1) found that the increase in the education situational adaptability, visual and language recognition,
instead reflects changing social norms, growing wage premium, particularly the college premium, explains and in-person interactions but limited formal education
corporate misgovernance, slackening regula- about 60 to 70% of the rise in wage inequality (variance) (e.g., janitors and cleaners, home health aides, construction
tory oversight, or increasing political capture of between 1980 and 2005. laborers, and security personnel). See Autor and Dorn (25) and
the policymaking process by elites (3–6). It would 3. F. Alvaredo et al., J. Econ. Perspect. 27, 3–20 (2013). Goos et al. (26) for evidence that employment in the
4. J. Bivens, L. Mishel, J. Econ. Perspect. 27, 57–78 (2013). U.S. and among OECD member countries has increasingly
therefore be a vast overstatement to conclude that 5. A. Bonica et al., J. Econ. Perspect. 27, 103–124 (2013). polarized into high-paid, abstract-intensive occupations and
the rise of U.S. inequality is exclusively due to 6. S. N. Kaplan, J. Rauh, J. Econ. Perspect. 27, 35–56 (2013). low-paid, manual-intensive occupations.
conventional market forces, or that public policy 7. D. Autor et al., The Contribution of the Minimum Wage to U.S. 28. J. Tinbergen, Kyklos 27, 217–226 (1974).
has not played a role. Wage Inequality over Three Decades: A Reassessment (NBER 29. Details of this model are given in the online supplement.
Working Paper 16533, Cambridge, MA, 2010). 30. D. Card, T. Lemieux, Am. Econ. Rev. 91, 97–102
But let us assume for the sake of argument 8. T. Piketty, E. Saez, Q. J. Econ. 118, 1–41 (2003). (2001).
that the rise of income inequality is entirely a 9. These calculations use data from (8), with data updated to 31. L. F. Katz, K. M. Murphy, Q. J. Econ. 107, 35–78
market phenomenon. Would this imply that 2012 available at http://elsa.berkeley.edu/saez/~Tab- (1992).
there is no role for public policy? A moment’s Fig.2012prel.xls. Average U.S. household incomes, including 32. L. Katz, D. H. Autor, Changes in the wage structure and
the top 1%, rose by 20.2%, while the average household earnings inequality. In Handbook of Labor Economics,
reflection suggests otherwise. As the economist income of the bottom 99% of households rose by only 3.5%. D. Card, O. Ashenfelter, Eds. (Elsevier-North Holland, Amsterdam,
Arthur Goldberger once famously observed, the 10. Thus, the top 1% maintains its share of household income at a 1999), vol. 3, pp. 1463–1555.
fact that nearsightedness is substantially a genetic constant 10.0% while average household incomes rise by 33. D. Card, T. Lemieux, Q. J. Econ. 116, 705–746 (2001).
disorder has no bearing on whether doctors 20.2%, as actually occurred. 34. D. H. Autor et al., Rev. Econ. Stat. 90, 300–323 (2008).
11. This point is due to Lawrence Katz of Harvard University, who 35. E. Crivellaro, “College wage premium over time: Trends in
should prescribe eyeglasses (53). What is rele- offers these calculations in his graduate labor economics Europe in the last 15 years.” University Ca’ Foscari
vant is whether the benefits of addressing myopia lecture notes. of Venice, Department of Economics Research Paper
exceed the costs. In the case of myopia, the avail- 12. T. Lemieux, Post-Secondary Education and Increasing Wage Series no. 03/WP/2014 (2014); http://dx.doi.org/10.2139/
ability of eyeglasses make this an easy call. Inequality (Working Paper 12077, National Bureau of Economic ssrn.2383795.
Research, 2006). 36. Summarizing evidence on the college premium in 12 European
Although there is no “remedy” for inequality 13. S. Firpo et al., Decomposition methods in economics. In countries between 1994 and 1999, Crivellaro (35) found a
that is as swift or cheap as eyeglasses, prosperous Handbook of Labor Economics, D. Card, O. Ashenfelter, Eds. pattern of increasing skill differentials except in countries that
democratic countries have numerous effective (Elsevier-North Holland, Amsterdam, 2011), vol. 4, pp. 1–102. have had a large increase in the relative supply of college

graduates, a pattern consistent with the conceptual model laid
out below.
37. Although this deceleration is not evident from Fig. 3, it is REVIEW
detected by the regression equation, as discussed in the online
supplement.
38.
39.
College Board, Trends in Student Aid: 2013 (College Board,
New York, 2013).
C. Avery, S. Turner, J. Econ. Perspect. 26, 165–192 (2012).
Income inequality in the
40. Three sources of uncertainty should be kept in mind when
interpreting these estimates. First, they encompass
substantial heterogeneity. Although the average college
developing world
graduate earns substantially more than the average high
school graduate, the least successful college graduates may Martin Ravallion
earn substantially less than the median among high school
graduates, and the most successful high school graduates Should income inequality be of concern in developing countries? New data reveal less
may earn substantially more than the median among college
graduates. Second, for students who acquire substantial
income inequality in the developing world than 30 years ago. However, this is due to falling
student debt but do not complete the college degree, it is inequality between countries. Average inequality within developing countries has been
far from certain that college will prove a good investment. slowly rising, though staying fairly flat since 2000. As a rule, higher rates of growth in
Finally, these calculations assume that the lifetime profile of average incomes have not put upward pressure on inequality within countries. Growth
earnings observed in the year of college graduation will
persist throughout the career. As Fig. 3 indicates, this
has generally helped reduce the incidence of absolute poverty, but less so in more
premium has changed substantially over time, so this unequal countries. High inequality also threatens to stall future progress against
assumption is only a rough approximation. However, the poverty by attenuating growth prospects. Perceptions of rising absolute gaps in living
college premium is so high at present that even with a standards between the rich and the poor in growing economies are also consistent
substantial decline, college would remain an attractive
financial proposition on average from a lifetime earnings
with the evidence.
D
perspective.
41. R. J. Murnane, J. Econ. Lit. 51, 370–422 (2013).
evelopment economics emerged as a sub- to John Rawls’s “difference principle” that (subject
42. D. Acemoglu, J. Robinson, Why Nations Fail (Crown, New York,
2012). discipline of economics in the 1950s, and to assuring liberty and equal opportunity) higher
43. As with cross-sectional inequality, there is no economically its initial focus was on economic growth, inequality can be justified as long as it benefits the
“ideal” level of intergenerational mobility. Even in a society with inequality as a secondary concern. worst-off group in society (1).
with perfect equality of opportunity, one would expect children
The prevailing orthodoxy for many dec- The period since 2000 has seen a deeper and
of successful parents to have above-average success as
adults, simply because many attributes that contribute to ades was that a period of rising inequality was to more widespread questioning of this long-standing
success (appearance, intellect, athleticism) are partly be expected in growing poor countries. Rising view of pro-poor inequality. New concerns have
heritable. inequality was seen to be more or less inevitable emerged about the instrumental importance of
44. M. Corak, J. Econ. Perspect. 27, 79–102 (2013). and not something to worry about, particularly equity to other valued goals, including poverty
45. S. F. Reardon, The widening academic achievement gap
between the rich and the poor: New evidence and possible
if the incidence of poverty was falling. Another reduction and human development more broad-
explanations. In Whither Opportunity? Rising Inequality, commonly held view was that policy efforts to ly. It appears more likely today that high inequal-
Schools, and Children’s Life Chances, G. J. Duncan, reduce inequality were likely to impede growth ity will be seen as a threat to future development
R. J. Murnane, Eds. (Russell Sage Foundation, New York, 2011), and (hence) poverty reduction. than as an inevitable and unimportant conse-
pp. 91–115.
46. R. Chetty et al., Is the United States Still a Land of Opportunity?
The existence of high inequality within many quence of past progress. The long-standing idea
Recent Trends in Intergenerational Mobility (NBER Working developing countries, side by side with persistent of a substantial growth-equity trade-off has come
Paper No. 19844, Cambridge, MA, 2014). poverty, started to attract attention in the early to be seriously questioned.
47. C.-I. Lee, G. Solon, Rev. Econ. Stat. 91, 766–772 (2009). 1970s. Nonetheless, through the 1980s and This paper reports new estimates of the levels
48. Between 2002–2003 and 2012–2013, the sum of federal
Pell Grants and loans for higher education increased by 105%,
well into the 1990s, the mainstream view in and changes in income inequality measures for
from $83 billion to $170 billion in constant 2012 dollars development economics was still that high the developing world. The new estimates take
[(38), table 1]. and/or rising inequality in poor countries was us up to 2010, embracing the period of higher
49. C.-T. Hsieh et al., The Allocation of Talent and U.S. Economic a far less important concern than assuring suf- growth rates in the developing world since the
Growth (NBER Working Paper No. 18693, Cambridge, MA, 2013).
50. D. H. Autor et al., Am. Econ. Rev. 103, 2121–2168 (2013).
ficient growth, which was the key to poverty turn of the millennium. In the light of these new
51. D. Card et al., J. Labor Res. 25, 519–559 (2004). reduction. The policy message for the develop- data, I revisit past and ongoing debates on in-
52. B. T. Hirsch, J. Econ. Perspect. 22, 153–176 (2008). ing world was clear: You cannot expect to have equality in developing countries and the trade-
53. A. S. Goldberger, Economica 46, 327 (1979). both lower poverty and less inequality while offs with growth and poverty reduction.
54. The extensive involvement of state and federal government in
education at all levels also underscores the fact that the
you remain poor, and, if you choose to give pov-
erty reduction highest priority, then focus on Income Inequality Measures
distribution of education and skills today is in no sense a
“free market” outcome; it is a consequence of both individual growth. To measure inequality in the developing world
and public choices. Other objections could still be raised to as a whole, one ignores country borders—pooling
ACKN OW LEDG MEN TS high income inequality. The classical utilitarian all residents and measuring inequality among
I thank D. Acemoglu, L. Katz, J. Van Reenen, M. Tatsutani, and formulation—whereby social welfare is judged them. This overall measure will naturally depend
two anonymous referees for valuable comments and advice, and by the sum of utilities, assuming diminishing on the inequality between countries as well as
C. Patterson and B. Price for expert research assistance. marginal utility of income—pointed to social wel- that within them. Thus, its evolution over time
Supported by NSF grant SES-1227334, Russell Sage Foundation
grant 85-12-07, and Alfred P. Sloan Foundation grant 2011-10-12.
fare losses from high inequality at a given mean. will depend on whether poorer countries are
All data and code that are unique to this article (Figs. 1, 3, But that did not persuade those who believed seeing lower growth rates as well as the things
and 6; fig. S2) are available from the author. All other figures that there was a trade-off between equity and happening within countries—economic changes
(Figs. 2, 4, and 5; figs. S1 and S3) are reproduced from growth. A moral defense could also be mounted and policies—that affect inequality.
other publications, as noted, with permission of the authors.
for the view that inequality is not an important If we are comparing country or regional per-
SUPPLEMENTARY MATERIALS issue for a growing developing country by appeal formance, then we want to isolate the within-
www.sciencemag.org/content/344/6186/843/suppl/DC1 country component of inequality as distinct from
Supplementary Text that between countries. Although there are many
Figs. S1 to S3 Department of Economics, Georgetown University,
Washington, DC 20057, and National Bureau of Economic
inequality measures, not all of them allow a clean
References (55–61)
Research, Cambridge, MA 02138, USA. E-mail: mr1185@ separation of the between and within compo-
10.1126/science.1251868 georgetown.edu nents. For example, such a decomposition is

SERIES
IZA DP No. 7155

PAPER
Inequality of Opportunity, Income Inequality and

Economic Mobility: Some International Comparisons
Paolo Brunori
DISCUSSION
Francisco H. G. Ferreira
Vito Peragine
January 2013
Forschungsinstitut
zur Zukunft der Arbeit
Institute for the Study
of Labor
Inequality of Opportunity, Income
Inequality and Economic Mobility:
Some International Comparisons
Paolo Brunori
University of Bari
World Bank and IZA
Vito Peragine
University of Bari
Discussion Paper No. 7155

January 2013
IZA
P.O. Box 7240

53072 Bonn
Germany
Phone: +49-228-3894-0
Fax: +49-228-3894-180
E-mail: iza@iza.org
Any opinions expressed here are those of the author(s) and not those of IZA. Research published in
this series may include views on policy, but the institute itself takes no institutional policy positions.
The IZA research network is committed to the IZA Guiding Principles of Research Integrity.
The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center
and a place of communication between science, politics and business. IZA is an independent nonprofit
organization supported by Deutsche Post Foundation. The center is associated with the University of
Bonn and offers a stimulating research environment through its international network, workshops and
conferences, data service, project support, research visits and doctoral program. IZA engages in (i)
original and internationally competitive research in all fields of labor economics, (ii) development of
policy concepts, and (iii) dissemination of research results and concepts to the interested public.
IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion.
Citation of such a paper should account for its provisional character. A revised version may be
available directly from the author.
IZA Discussion Paper No. 7155
January 2013
ABSTRACT
Inequality of Opportunity, Income Inequality and Economic Mobility:
Some International Comparisons 1
Despite a recent surge in the number of studies attempting to measure inequality of
opportunity in various countries, methodological differences have so far prevented
meaningful international comparisons. This paper presents a comparison of ex-ante
measures of inequality of economic opportunity (IEO) across 41 countries, and of the Human
Opportunity Index (HOI) for 39 countries. It also examines international correlations between
these indices and output per capita, income inequality, and intergenerational mobility. The
analysis finds evidence of a “Kuznets curve” for inequality of opportunity, and finds that the
IEO index is positively correlated with overall income inequality, and negatively with
measures of intergenerational mobility, both in incomes and in years of schooling. The HOI is
highly correlated with the Human Development Index, and its internal measure of inequality
of opportunity yields very different country rankings from the IEO measure.
JEL Classification: D71, D91, I32
Keywords: equality of opportunity, income inequality, social mobility, mobility
NON-TECHNICAL SUMMARY
Many different indices to measure inequality of opportunity have been proposed, but only two
have been applied to enough countries to permit reasonably meaningful international
comparisons. The inequality of economic opportunity index (IEO) estimates the (lower bound)
share of income inequality that can be attributed to differences in people’s pre-determined
circumstances (such as race, gender and family background). It has been applied to 41
countries, and ranges from 2% in Norway to 34% in Guatemala. The second approach is
known as the Human Opportunity Index (HOI): an index of children’s access to basic
services, penalized by unequal opportunities in that access. Like the IEO, the HOI must lie
between 0 and 100%. In the 39 countries where it has been computed, it ranges from 10% in
Niger to 91% in Chile. The IEO is positively correlated with income inequality, and negatively
with intergenerational mobility – both in incomes and in years of schooling. The HOI is highly
correlated with the Human Development Index. Its internal measure of inequality of
opportunity – the dissimilarity index – yields very different country rankings from the IEO,
highlighting the differences between the two methods. The IEO and the HOI may well be
complementary, but users should be cautious to understand what each one really captures.
Corresponding author:
The World Bank
1818 H Street, NW
Washington, DC 20433
USA
E-mail: fferreira@worldbank.org
1
This paper was prepared for a volume on “The Triple Challenge of Development: Changing the rules
in a global world”, which draws on a conference held at Mount Holyoke College in March 2012. We are
grateful to Michael Grimm, Peter Lanjouw, Branko Milanovic and Eva Paus (the editor) for various
helpful comments and suggestions. We are also grateful to Ambar Narayan and Alejandro Hoyos
Suarez for help with the data on human opportunity indices and to Tor Eriksson and Yingqiang Zhang
for providing us additional estimates for China. Ferreira would like to acknowledge support from the
Knowledge for Change Program, under project grant P132865. All errors are exclusively our own. The
views expressed in this paper are those of the authors, and they should not be attributed to the World
Bank, its Executive Directors, or the countries they represent.
1. Introduction
The relationship between inequality and the development process has long been of interest, and
both directions of causality have been extensively investigated. The idea that the structural
transformation that takes place as an economy develops may lead first to rising and then to falling
inequality – known as the Kuznets (1955) hypothesis – was once hugely influential. The view that
inequality may, conversely, affect the rate and nature of economic growth has an equally distinguished
pedigree, dating back at least to Kaldor (1956). In the 1990s, a burgeoning theoretical literature
suggested a number of mechanisms through which wealth inequality might be detrimental to economic
growth: when combined with credit constraints and increasing returns; through political channels;
fertility effects; etc. See Voitchovsky (2009) for a recent survey of that literature.
But popular concern about inequality in developing (and developed) countries does not originate
exclusively – or even primarily – from its possible instrumental effects - on growth, on the growth
elasticity of poverty, on health status, on crime, or on any number of other factors that are possibly
influenced by the distribution of economic well-being. Many of those who worry about inequality do so
because they consider it – or at least some of it – “unjust”. Most development economists, however,
share the broader profession’s discomfort with normative concepts such as justice and, until recently
and with some distinguished exceptions, have had little to say about it.
That is a pity. Behavioral economics has taught us that notions of fairness and justice affect
individual behavior – in the precise and well-documented sense that they induce sizable deviations from
the behaviors predicted by models based on the assumption of purely self-regarding preferences (e.g.
Fehr and Schmidt, 1999; Fehr and Gachter, 2000; Fehr and Fischbacher, 2003). Some recent
experimental evidence suggests that, when assessing outcome distributions, people do distinguish
between factors for which players can be held responsible, and those which are beyond their control
(Cappelen et al., 2010). If fairness matters to economic agents and alters their behavior, then
understanding fairness ought to matter even to the purest positive economist. If people assess
distributional outcomes differently depending on how much of the inequality they observe is thought to
be “fair” or “unfair”, then it may be useful to measure the extent to which inequality is unfair.
Efforts in this direction have already taken place. Drawing primarily on the welfare economics
literature on “inequality of opportunity” (I. Op.), researchers have started to measure unfair inequality
in both poor and rich countries. In that literature, there is now widespread agreement on the basic
principle of what equality of opportunity refers to: inequalities due to circumstances beyond individual
control are unfair, and should be compensated for, while inequalities due to factors for which people
can be held responsible (sometimes called “efforts”), may be considered acceptable. But this broad
concept can be interpreted in a number of different ways, some of which have been shown to be
mutually inconsistent. And there is an array of actual indices that have been proposed to implement
these concepts, and used to measure inequality of opportunity in different countries or at different
times. The relatively high ratio of different (and incomparable) approaches to actual empirical
applications means that it has so far been difficult to make a reasonably broad comparison of inequality
of opportunity levels across countries.
2
This paper takes a first step towards making such a comparison, by drawing on two specific
approaches that have been relatively widely used. The first is the measurement of ex ante inequality of
economic opportunity. The second is the measurement of (children’s) access to basic services adjusted
for differences associated with circumstances – commonly known as the Human Opportunity Index
(HOI). The latter is not a measure of inequality of opportunity per se; it is better seen as a development
index that is designed to be sensitive to inequality of opportunity. Our objective is a modest one: we
collect and summarize the results of empirical applications of these two measures to as many countries
as possible, and describe the correlations between these measures and a number of other indicators of
interest, including GDP per capita, overall income inequality, and two measures of intergenerational
mobility.
We hope that the collected evidence on the degree of inequality of opportunity in different
countries, and its pattern of association with other variables, might help to shed light on the nature of
the (often increasing) inequalities observed today in many areas of the world. The paper is organized as
follows. Section 2 contains a brief overview of the concepts and approaches to the measurement of
inequality of opportunity. This provides essential background not only for an understanding of where
the inequality of opportunity measures come from and what they do, but also of what they do not do,
and the concepts they do not capture. Section 3 contains our review of inequality of opportunity
measures for 41 countries, and examines how they correlate with other indicators. Section 4 presents a
comparison of HOI applications across 39 developing countries, and how it correlates with other
relevant indices, including the United Nations’ Human Development Index (HDI). Section 5 contains a
discussion of the results and some concluding remarks.
2. Concepts and measurement
The economics literature on inequality of opportunity builds explicitly on a few key contributions
from philosophy, including Dworkin (1981a, b), Arneson (1989) and Cohen (1989). The basic idea, as
noted above, is that outcomes that are valued by all or most members of society (such as income,
wealth, health status, etc.), and which are often termed “advantages”, are determined by two types of
factors: those for which the individual can be held responsible, and those for which she cannot.2
Inequalities due to the former - which we will call “efforts” - are normatively acceptable, whereas those
due to the latter - which we call “circumstances” - are unfair, and should in principle be eliminated.3
However, as economists sought to formalize this idea so as to make it more precise, they quickly
faced some fundamental choices, both conceptual and methodological. Some of these are actually
choices between mutually inconsistent principles or approaches. Following Fleurbaey (1998, 2008) and
Fleurbaey and Peragine (2012) we focus on two such fundamental dichotomies: the distinction between
2
Which factors belong to which category is a subject of considerable debate in the philosophical literature.
3
The terminology of advantages, circumstances and efforts follows Roemer (1998). Other authors prefer the term
“responsibility factors” to efforts, for example.
3
the compensation and reward principles, and the distinction between the ex-ante and ex-post
approaches.4
In order to understand these distinctions, it is helpful to introduce the concepts of types and
tranches, using some simple notation. For simplicity, consider the basic set up in which there is a single
advantage y and a vector of discrete circumstance variables, C. Let effort be measured by a continuous
scalar variable e. Then suppose that all determinants of y, including various different forms of luck, can
be classified into either the vector C or the scalar index e. The theory of inequality of opportunity is built
upon the idea that these circumstances and efforts determine advantage, as follows:
(1)
Because C is a vector with a finite number of elements, each of which is discrete, we can partition
the population into a set of groups that are fully homogeneous in terms of circumstances. Formally, this
is the partition   T1 , T2 ,...,TK  such that Ci  C j , i, j i  Tk , j  Tk , k. Each of these subgroups,
indexed by k, is called a type Tk , and clearly individuals within each type can differ only in their effort
level . Let denote the advantage distribution in type k and denote its population share. The
overall distribution for the population as a whole is .
Effort variables have been treated in a number of different ways in the literature. In this
exposition, we follow the influential approach due to Roemer (1993, 1998), in which effort is treated as
unobserved. Roemer argues that the absolute level of effort is not actually an appropriate basis for
comparison across individuals, because the average level of effort expended in each type may vary. The
children of well-educated parents may on average dedicate greater effort to their studies than those of
less educated parents, for example. Roemer argues that such average differences in effort levels should
be treated as characteristics of the types, rather than of the individuals – effectively as unobserved
circumstances. He proposes that effort comparisons be based instead on relative effort, which he
equates with the percentile of the distribution of advantage within each type: . This is known
in the literature as the Roemer Identification Assumption. It naturally gives rise to an alternative
partition of the population, by grouping in separate tranches all those who are at identical percentiles of
the advantage distribution, across types:   R1 , R2 ,..., RP  .
So we have a population of individuals, each of whom is fully characterized by the triple (y, C, e).
This population can be partitioned in two ways: into types (within which everyone shares the same
circumstances), and into tranches (within which everyone shares the same degree of effort). Figure 1
provides a simple illustration, in which there are three types, T1, T2 and T3. The (inverse) cumulative
advantage distribution of each type is given by , and their means are indicated on the vertical axis,
where advantages (or incomes) are mapped. Tranches are not shown in the figure but, under the
Roemer Identification Assumption, they would correspond to ‘vertical’ sections across the three type
distributions, at each percentile pk on the horizontal axis. With this very basic toolkit, we are ready to
4
This section is intended as a brief non-technical overview. It cannot – and is not intended to – do justice to the
recent literature. Two excellent full-length reviews of the literature on the measurement of I. Op. are Pignataro
(2011) and Ramos and van de Gaer (2012).
4
understand the distinction between the compensation and reward principles, and between ex-ante and
ex-post approaches.
The compensation principle states the first basic idea of inequality of opportunity as follows:
"inequalities due to circumstances should be eliminated". There are two main versions of this principle
in the literature. The ex-ante approach to compensation (associated with van de Gaer, 1993) seeks to
evaluate – i.e. attribute a numerical value vi to – the opportunity set faced by individual i. Inequality of
opportunity would then be eliminated when all types faced opportunity sets with the same value:
. If that did not hold, inequality of opportunity could be measured by computing an
appropriate inequality measure I(.) over the counterfactual distribution where each person’s advantage
is replaced by the value of his or her opportunity set, vi:
, where (2)
Under this ex-ante compensation approach, then, there are two questions left before a precise
measure can be proposed. First, how should opportunity sets be valued, i.e. how should be chosen?
And second, what inequality index I(.) should be applied to the counterfactual distribution? Most
attempts to evaluate the opportunity set faced by individuals in a given type k are based on information
on the type’s advantage distribution . The advantage prospect of individuals in the same type is
interpreted as the set of opportunities open to each individual in that type. A specific version of this
model, extensively used in empirical analyses, further assumes that the value of the opportunity set
can be summarized by a single statistic such as its mean, .5 In that case,
Hence, starting from a multivariate distribution of income and circumstances, a smoothed

distribution is obtained, which is interpreted as the distribution of the values of the individual
opportunity sets. In this model, measuring opportunity inequality with Equation (2) simply amounts to
measuring inequality in the smoothed distribution6. Clearly, focusing on the mean imposes full neutrality
with respect to inequality within types.
There are also alternatives with respect to the inequality index: van de Gaer (1993) argues for a
measure with infinite inequality aversion, effectively . Other authors have suggested alternative
inequality measures, such as a transformation of the Gini coefficient (Lefranc et al., 2008), a rank
dependent mean (Aaberge et al., 2011), or the mean logarithmic deviation (Checchi and Peragine, 2010;
Ferreira and Gignoux, 2011).
The ex-post approach to compensation, on the other hand, argues that inequalities should be
eliminated among any individuals who exert the same degree of effort. Under this approach there is no
need to evaluate opportunity sets, but one must observe (or agree on a measure of) effort. Under
Roemer’s identification assumption, eliminating ex-post inequality of opportunity would require
eliminating all income differences among individuals at a given percentile of their type’s advantage
5
Alternative approaches propose to use the equally distributed equivalent income (EDEI), see Atkinson (1970), or
other welfare indicators (see Lefranc et al. 2008)
6
The concept of smoothed (and standardized) distributions is introduced by Foster and Shneyerov (2000). In the
present context, a smoothed distribution is one where individual incomes are replaced by their subgroups means.
5
distribution, across types: Inequality of opportunity can be measured by applying
an inequality measure I(.) to the distribution of advantages within each tranche, and then aggregating
across tranches.
In terms of our illustration in Figure 1, eliminating ex-ante inequality of opportunity (when

) would be achieved by shifting those inverse distribution curves up or down (i.e. transferring
incomes between individuals of different types) until they had the same mean. Eliminating ex-post
inequality of opportunity, on the other hand, would require making those distributions identical to one
another. The latter requirement clearly demands a more complex set of transfers, so that inequality is
eliminated within each and every tranche. Indeed, ex-post equality of opportunity implies ex-ante
equality of opportunity, but not the reverse. In this example:
(3)
Let us now briefly turn to the reward principle, which maintains that "inequalities due to unequal
effort should be considered acceptable". This is, in some sense, the other side of the coin (from the
compensation principle) of the basic idea of inequality of opportunity expressed in the first paragraph of
this section. This principle too can be formalized in various ways, the two most prominent ones being
the liberal reward principle that "inequalities due to unequal effort should be left untouched" ---
prohibiting redistribution between individuals with identical circumstances --- and the utilitarian reward
principle that "inequalities due to unequal effort do not matter" --- advocating a sum-maximizing policy
among subgroups with identical circumstances7.
An interesting recent result from the theoretical literature (see Fleurbaey, 2008, and Fleurbaey
and Peragine (2012), is that both of these reward principles are incompatible with the ex-post
compensation principle: full respect for the differences in reward to effort within each type is not
consistent with full equality within tranches. Although the result is proved for a more general set up, its
essence is easily understood from Figure 1 again, focusing on types 1 and 2. The liberal reward principle
requires that policy makers do nothing about the differential rewards between high and low percentiles
within each of those types. The ex-post compensation principle requires that the two distributions
become identical – with the functions lying on top of each other. Those two things cannot both be
achieved.
Figure 1 is also suggestive of another result in Fleurbaey and Peragine (2012): there is no such
clash between the ex-ante compensation principle and the reward principles. One could “re-scale” the
advantage distributions across types so that they would all have the same mean (or some other value),
without changing the absolute advantage differences (the rewards to effort) across percentiles within
each type. The ex-post approach to the compensation principle is more demanding, but a conceptual
7
These various distinctions are discussed in detail in Fleurbaey (2008).
6
price must be paid for its stringency, namely consistency with the reward principles that also underpin
the theory of equality of opportunity.8
Most measures of inequality of opportunity computed in practice have followed an ex-ante

approach. A notable exception is Checchi and Peragine’s (2010) work on inequality of opportunity in
Italy, which reports both ex-ante and ex-post measures. There is also a related literature that
acknowledges the incompatibility between ex-post compensation and reward, and proposes fair
allocation rules that satisfy somewhat weakened versions of those principles. If one treats these fair
allocation rules as income norms (that individuals would have received under that particular definition
of fairness) then unfair inequality can be defined as some aggregate of the differences between actual
and norm incomes across the population. See Ramos and van de Gaer (2012) for an excellent discussion
of these measures, and Almas et al. (2011) and Devooght (2008) for examples of the approach.9 But
neither ex-post compensation nor norm-based measures have been computed in similar ways across
many countries.
In contrast, the particular version of the ex-ante approach where equation (2) is computed with
, has been applied to at least some forty countries, by a number of authors. The measure I(.)
used does vary across some of the papers but most use the mean logarithmic deviation, following
Checchi and Peragine (2010) and Ferreira and Gignoux (2011). In a few cases, as detailed below, the
Theil (T) index and even the variance are employed. Despite these differences, as well as a variety of
caveats on data comparability across – or even within – studies, the eight papers reviewed in Section 3
comprise the most closely comparable sources on actual I. Op. measures across countries that we are
aware of.
In closing this section, we turn to another approach that has been applied to a number of
countries in recent years, namely the Human Opportunity Index of Barros et al. (2009, 2011). This index
is defined over a different set of advantages (which, confusingly, are sometimes referred to as ‘basic
opportunities’), namely access to certain basic services, such as piped water, electricity or sanitation. In
a discrete population of size n, let denote the probability that person i has access to service j.
then denotes the expected coverage of service j in the population. In practice, probabilities
are often estimated econometrically from binary data on access, and can be interpreted as the
average coverage of service j. Let this population also be partitioned into K types, by   T1 ,T2 ,...,TK as
before. Denote the population share of type k by wk, and the average coverage of service j in type k as
. Then the human opportunity index for service j is defined as:
where (4)
8
There is also a potential practical price to be paid in empirical exercises of measuring inequality of opportunity.
Because the ex-post approach requires a partition into types and tranches, it is more demanding on the data.
When many circumstance variables are observed, precision is harder to achieve for ex-post measures. See Ferreira,
Gignoux and Aran (2011) for a discussion.
9
Brunori and Peragine (2011) compare the norm-based measures with the ex-ante and ex-post measures.
7
In equation (4), is a version of the dissimilarity index commonly used in sociology. In this
application, it simply computes an appropriately normalized (and population-weighted) average
deviation in service coverage from the mean, across types. The HOI (for service j) itself, denoted by Hj, is
simply the average access rate in the population, penalized by the degree of dissimilarity in that
coverage across types. It is clearly analogous to the Sen welfare function, where mean outcomes are
adjusted by one minus a measure of inequality. Sometimes an aggregate index is calculated as an
average of Hj across a number of different services, .10 Various versions of the HOI have now
been computed for at least 39 countries, and basic results are compared in Section 4 below.
3. Ex-ante inequality of opportunity in 41 countries
As noted above, the ex-ante approach to the measurement of inequality of opportunity essentially
consists of computing an inequality measure over a counterfactual distribution, where individual
advantages are replaced with some valuation of the opportunity set of the type to which the individual
belongs. In this section, we review eight papers that have adopted this approach and applied it, in total,
to 41 countries, ranging from Guinea and Madagascar (with annual per capita GNIs of PPP$980, to
Luxembourg, with a per capita GNI of almost PPP$ 64,000). The eight papers are Checchi et al. (2010);
Ferreira and Gignoux (2011); Ferreira et al. (2011); Pistolesi (2009); Singh (2011); Belhaj-Hassine (2012),
Cogneau and Mesple-Somps (2008) and Piraino (2012).
All of these papers use a measure of economic well-being as the advantage indicator: household
per capita income, household per capita consumption, or individual labor earnings. All use the mean
value of this indicator for each type as the value of the type’s opportunity set. We refer to the measure
generated by this specific version of the ex-ante approach as an index of inequality of economic
opportunity (IEO). There are, in fact, two closely related versions of the index: the absolute or level
estimate of inequality of opportunity (IEO-L) is given simply by the inequality measure computed over
the smoothed distribution, where each person is given the mean income of their types: . The ratio
of IEO-L to overall inequality in the relevant advantage variable (e.g. household per capita income) yields
the relative measure, IEO-R11:
(5)
The partition of types varies across studies, ranging from six types to 7,680 (although in four of
the eight studies, the range is a more comfortable 72-108 types). Because in some cases the data sets
are not large enough to yield precise estimates of for all types, some authors compute IEO-L using a
parametric shortcut. After estimating the reduced-form regression of income on circumstances:
(6)
10
However, see Ravallion (2011) on the potential pitfalls of such arbitrary aggregate indices or, as he calls them,
“mashup indices” of development.
11
Ferreira and Gignoux (2011) refer to the corresponding measures that are obtained when the mean log
deviation is used as the inequality measure I(.) as IOL and IOR. They also note that IEO-R is an application of a
standard between-group inequality decomposition, which has long been familiar. See e.g. Bourguignon (1979).
8
and obtaining coefficient estimates , these authors use predicted incomes as a parametric
approximation to the smoothed distribution:
, where (7)
Parametric estimates are also presented either as levels (IEO-L) or ratios (IEO-R), analogously. This
approach follows Ferreira and Gignoux (2011), which in turn draws on Bourguignon et al. (2007).
Empirically, parametric estimates of inequality of opportunity tend to be a little lower than their non-
parametric counterparts but, at least in the case of Latin America, the differences are not great:
proportional differences between the two average 6.6% in Ferreira and Gignoux (2011).
The fact that the parametric estimates are conservative – i.e. generally lower than the non-
parametric ones – is consistent with another important property of these estimates of IEO-R and IEO-L.
They are, in each and every case, lower-bound estimates of inequality of opportunity. A formal proof of
the lower-bound result is contained in Ferreira and Gignoux (2011), but the intuition is straight forward.
The set of circumstances which is observed empirically - and used for partitioning the population into
types - is a strict subset of the theoretical vector of all circumstance variables. The existence of
unobserved circumstances – virtually a certainty in all practical applications – guarantees that these
estimates of I.Op. – whether parametric or non-parametric – could only be higher if more circumstance
variables were observed.
As discussed in Ferreira and Gignoux (2011), the existence of effort variables, observed or
unobserved, is entirely immaterial to this result, since (6) is written as a reduced-form equation, where
any effect of circumstances on incomes through their effects on effort (such as years of schooling or
hours worked) is captured by the regression coefficients, and hence influence the smoothed
distribution. In a setting where some variables are treated as observed efforts (as in Bourguignon et al.
2007), Equations (6) and (7) capture the reduced-form influence of circumstances on advantages, both
directly and indirectly through efforts. By construction, therefore, the only omitted variables that matter
for IEO are omitted circumstances.12
Table 1 presents the estimates of IEO-L and IEO-R for each of the forty-one countries studied by
the eight aforementioned papers. The table also lists their gross national income (GNI) per capita;
overall inequality and, when available, a measure of intergenerational earnings elasticity (IGE) reported
in the literature; a measure of the intergenerational correlation of education from Hertz at al. (2007);
and the Human Opportunity Index. Overall inequality is measured by whatever index was used in the
construction of the IEO indices for each country. Except where indicated, this measure was the mean
logarithmic deviation, also known as the Theil-L index, and a member of the generalized entropy class of
inequality measures. Whereas overall inequality, IEO-L and IEO-R come from the eight studies
mentioned above, the other variables come from other sources. GNI per capita comes from the World
Bank’s World Development Indicators database. Our measure of intergenerational correlation of
12
Of course, this does not hold for the estimates of the individual coefficients . First, these coefficients are
reduced-form, rather than structural, estimates. In addition, they are likely to be biased (upwards or downwards)
even as reduced-form estimates, by the omission of unobserved circumstances. The lower-bound result applies
only to the overall measures of inequality of opportunity, IEO-L and IEO-R.
9
education is simply the correlation coefficient between the parents’ education and the child’s education,
where both are measured by years of completed schooling, as reported by Hertz et al. (2007). Parental
education is the average of mother’s and father’s attainment “wherever possible” (Hertz et al, 2007,
p.11). The correlation we report is what the authors call a measure of “standardized persistence”.
The measures of intergenerational earnings elasticity reported in Table 1 come from eleven
different studies published over the last ten years, namely Azevedo and Bouillon (2010); Cervini Pla
(2009); Christofides et al. (2009); Corak (2006); D’Addio (2007); Dunn (2007); Ferreira and Veloso (2006);
Grawe (2004); Hnatkovskay et al. (2012); Hugalde (2004); Nuñez and Miranda (2006); and Piraino
(2007). Denoting parental earnings (or income) by , and the adult child’s earnings by , these
elasticity estimates generally come from an equation of the form:
(8)
An elasticity (β) of 0.4, for example, would mean that income differences of 100% between two
fathers (say), would lead to a 40% gap between their sons (on average). As in the case of the IEO
measures, the datasets and econometric methods used for estimating this elasticity are not
homogeneous across studies. This comparative exercise is very much in the same spirit as Corak (2012),
and the same caveats he discusses are applicable here. The values for the Human Opportunity Index
reported in Table 1 come from Molinas et al. (2011) for Latin America, and World Bank (2012a, b) for
Africa.
Table 1 should be read in close conjunction with Table 2, which provides some basic information
on each of the eight studies used to construct the inequality of opportunity estimates in Table 1. Table 2
describes which countries are studied in each paper; the specific data sets (including survey year); the
precise income and circumstance variables used; whether the estimation was parametric or otherwise,
and the number of types included in each calculation. The table highlights a number of problems for
comparability across these studies. First is the nature of the advantage variable (y) itself: whereas
Checchi et al. (2010), Pistolesi (2009), Singh (2011) and Belhaj-Hassine use labor earnings, Ferreira and
Gignoux (2011) and Piraino (2012) use incomes, Cogneau and Mesple-Somps (2008) use consumption,
and Ferreira et al. (2011) use imputed consumption. And the definitions of earnings and incomes are not
exactly the same across each of these papers either.
These distinctions are not immaterial: in a comparison of six Latin American countries, Ferreira
and Gignoux (2011) found substantially higher estimates of IEO-R for consumption expenditure than for
income distributions, in the same countries.13 They attributed this finding to the fact that income
inequality measures are thought to contain greater amounts of measurement error, as well as transitory
income components, which are less closely correlated with circumstances than permanent income or
consumption might be. Bourguignon et al. (2007) also noted differences between estimates for
individual earnings and for household per capita incomes, which they attributed to the fact that unequal
opportunities affect the latter not only through earnings, but also through assortative mating, fertility
decisions, and non-labor income sources.
13
Similarly, Singh (2010) finds a higher IEO-L for consumption than for earnings in India.
10
Second, the studies differ in the number of types used for the decomposition and, naturally, in the
exact set of circumstances used in each case. On one extreme, the Cogneau and Mesple-Somps study
has a mere three types for Uganda, based on father’s occupation and education levels, while on the
other Pistolesi has 7,680 types, constructed on the basis of information on age (20 levels), parental
education (4 levels for the mother and 4 for the father), occupational group of the father (6 categories),
individual ethnic group (2 categories), individual region of birth (2 categories). There is, fortunately, a
middle range of studies which account for most countries in the sample, with 72 to 108 types each.
Nevertheless, results for Africa and the US should certainly be interpreted with caution, in light of the
number of types used in each case. Finally, a third comparability caveat, on which we have already
dwelled, is the fact that some studies use non-parametric estimates while others use parametric ones.
Bearing these caveats in mind, Table 1 nevertheless illustrates the substantial variation in
inequality levels across countries – both in advantages and in opportunities. The mean log deviation for
incomes (or the corresponding advantage indicator) ranges from 0.083 in Denmark to 0.675 in South
Africa. Norway, Slovenia and Sweden also have comparatively low levels of overall inequality, while
Brazil and Guatemala stand out at the upper end. Inequality of opportunity levels (IEO-L) range from
0.003 in Norway and 0.005 in Slovenia to 0.199 in Guatemala and 0.223 in Brazil. In other words, the
level of inequality in the distribution of values of opportunity sets across types (the smoothed
distribution described in Section 2) in Brazil is almost three times as large as the inequality (measured by
the same index) in the distribution of actual incomes in Denmark. One can also observe substantial
differences in IEO-L among countries at closer levels of development, and more methodologically
comparable: Madagascar’s level of inequality of opportunity is twice that of Ghana; those of the US and
the UK are ten times those of Norway and almost four times higher than Denmark’s.
The ratio of these two inequality measures, i.e. the (lower bound) share of the overall inequality
due to inequality of opportunity (IEO-R), also varies substantially, from 0.02 in Norway to 0.34 in
Guatemala. Slovenia also has a remarkably low inequality of opportunity ratio, at 0.05, while Brazil
closely follows Guatemala in the upper tail, at around 0.32. Figure 2 shows the range of relative
measures of inequality of opportunity graphically, for the entire sample, highlighting those countries
where consumption (actual or predicted) was used instead of earnings or incomes.
It may be of interest to look at how these measures of inequality of opportunity correlate with
some other important variables. Output per capita, overall income inequality, and measures of
intergenerational mobility – a concept closely related to I.Op. – are natural candidates. Figures 3, 4, 5
and 6 depict the associations between the relative measure of inequality of opportunity (IEO-R) and four
other variables – log per capita GNI, total inequality, the intergenerational elasticity of income, and the
intergenerational correlation of education. Figure 3 reveals a non-linear relationship between inequality
of opportunity and the level of development, as measured by log per capita income levels. In fact, the
association appears to have an inverted-U shape, much as the “Kuznets curve” that used to be
hypothesized for the relation between income inequality and the “level of development”. The
regression of IOR on a quadratic of log GNI is shown in the figure; the coefficient on the linear term is
0.32 (p-value: 0.05), and that on the quadratic term is -0.017 (p-value: 0.05).
11
A very similar relationship (not shown) is found between IEO-L and log per capita GNI (with a
coefficient of 0.37 on the linear term, and on the square term of -0.02, both significant at the one
percent level). While the poorest countries in this figure are all located in Africa, the middle income
countries near the turning point of the inverted-U include a number of Latin American countries, as well
as Egypt, South Africa and Turkey. The richer part of the sample is dominated by European countries and
the United States. Although these tend to be more I. Op. egalitarian, there is still a considerable spread
among them.
It is, of course, impossible to interpret this inverted-U pattern solely on the basis of the
information available in our data. One can weave hypotheses: the non-linearity might reflect two
opposite effects at play, the relative strengths of which change as incomes grow. Perhaps at very low
levels of development, new income opportunities are initially captured by a narrow privileged group – a
few well-educated families, or a small ruling ethnic group. During that phase, disparities across types
may grow even faster than overall income inequality. At some point, however, the grip of the elite on
economic opportunities must weaken if growth is to continue. Such mechanisms have been modeled
formally: the transition can occur when, at a certain point, the elite decides that the costs of expanding
education to “the masses” (in terms of their own share of political power) is outweighed by the likely
economic gains from a more skilled labor force (Bourguignon and Verdier, 2000) Alternatively, the
threat of revolution may impose the franchise and a broader sharing of political influence, even upon a
less enlightened elite (Acemoglu and Robinson, 2000). There is also some evidence that lower inequality
of opportunity may be associated with faster growth, at least in richer countries (see, e.g., Marrero and
Rodriguez, 2010, for a sample of US states).
But these are only hypotheses consistent with the pattern in Figure 3. It is equally possible, of
course, that the pattern is spurious: other variables may cause inequality of opportunity first to rise, and
then decline with GNI. As we have learned from work on the (income) Kuznets hypothesis, it would also
be foolhardy to infer much about the time-series pattern in any given country from a simple cross-
sectional association. At some level, in fact, it is probably fruitless to look for evidence of causal
relationships between two variables at such a high order of aggregation. Both overall output levels (GNI)
and inequality of opportunity are summary statistics, jointly determined by the full general equilibrium
of the economy, including all of the key political economy processes that determine policy variables
such as tax rates and spending allocations. It is likely that one can more easily find causality at the
microeconomic level. From that vantage point, disentangling causality in the relationship depicted in
Figure 3 may well be pointless, even if the correlation between the two aggregate variables reflects
genuine economic processes, which are both real and important.
Another question that naturally arises is whether there is any observable empirical relationship
between inequality of opportunity and income inequality. Since the former is measured as a component
of the latter there is a mechanical aspect to the relationship in levels, but it is not obvious that there is
any mechanical reason to expect a correlation between income inequality levels and the relative extent
of inequality of opportunity. Figure 4 shows the association between overall inequality (in economic
advantage) and the share of that inequality associated with inequality of opportunity (IEO-R). The
correlation coefficient is 0.523 (p-value: 0.0004). A number of possible mechanisms might drive this
12
correlation as well. One that appears eminently plausible is the notion that today’s outcomes shape
tomorrow’s opportunities: large income gaps between today’s parents are likely to imply bigger gaps in
the quality of education, or access to labor market opportunities, among tomorrow’s children (Ferreira,
2001). Naturally, the reverse causality probably holds too: if opportunity sets differ a great deal among
people, then individual outcomes are also likely to be unequal. Inequalities in income and opportunities
are both endogenously determined: once again, the quest for causality at the aggregate level may be
futile, even if the correlation reflects real underlying political and economic processes.14
The use of the links between parents’ and children’s incomes to describe an important
manifestation of inequality of opportunity suggests that the concept should be closely related to
intergenerational mobility. Indeed, if we wrote and , equations (6) and (8) would
be identical suggesting that, if the set of observed circumstances becomes restricted to parental income,
then our lower-bound measure of inequality of opportunity is very closely related to the commonest
measure of intergenerational mobility, namely the IGE. It can easily be checked that the R2 of (8) is
identical to the IEO-R measure defined by (5) and (7) when the variance of logarithms is used as the
inequality index.
Figure 5 documents the association between IEO-R and (inverse) economic mobility, as measured
by the intergenerational elasticity of earnings (or incomes). The correlation across the 23 countries for
which we have both variables in Table 1 is 0.5853 (p-value: 0.0172). Of course, the two measures are not
exactly the same, in part because the vector of circumstances C used to partition types and generate
IEO-R is not the same as a measure of parental income or earnings. In fact, C does not contain that
variable for any of the 41 countries in Table 1. It does, however, usually contain parental education (and
in some cases parental occupation), which are themselves determinants of log parental incomes. And it
often contains additional information, such as race or the region of the person’s birth.
For these reasons, we expected the correlation in Figure 5 to be strong, but not perfect. Given the
likely correlation between most circumstances and parental economic status, it would be surprising if
this association turned out to be weak. Given the isomorphism between the ex-ante measurement of
inequality of opportunity and the measurement of intergenerational mobility, we find it intriguing that
these comparisons do not appear to have been made before.
It should also be noted that Figure 5 is close in spirit to Figure 2 in Corak (2012), which plots the
intergenerational earnings elasticity against income inequality (measured by the Gini coefficient) across
countries.15 Instead of plotting the estimates of IGE against overall inequality, we plot the
intergenerational elasticity of income against a broader measure of inequality of opportunity.
14
If an inverted U-shaped relationship is observed between income inequality and per capita GNI levels across
countries – i.e. if a cross-sectional “Kuznets curve” holds empirically - then the positive association between
income inequality and IEO-R shown in Figure 4 actually implies the inverted U shape in Figure 3. We are grateful to
Branko Milanovic for pointing this out.
15
Corak’s figure has rapidly become well-known, in part because Alan Krueger, Chairman of President Obama’s
Council of Economic Advisers, referred to it in a speech as “the Great Gatsby curve”, relating the distance between
the rungs of the economic ladder, and the ease with which it is climbed.
13
Reassuringly, a very similar correlation is found between the same measure of inequality of
opportunity (IEO-R) and a different gauge for intergenerational (im)mobility, namely the correlation
between parental and child schooling attainment. As noted earlier, the intergenerational correlations of
education reported in Table 1 come from Hertz et al. (2007), and use the average years of schooling
completed by a person’s mother and father as the measure of parental education. Figure 6 shows the
scatter-plot for the 23 countries for which data on both variables is available. The correlation coefficient
is 0.5965 (p-value: 0.0021). So, inequality of economic opportunity, as measured by IEO-R, is clearly
negatively associated with two independent measures of intergenerational mobility (as opposed to
persistence), one based on incomes and the other on educational attainment.
4. Measuring development with a penalty for unequal opportunities
The country composition of Table 1 was determined by the availability of information on ex-ante
measures of inequality of opportunity, IEO-L and IEO-R, and drew on the eight papers listed in Table 2.
The last column of Table 1 contains estimates of the aggregate Human Opportunity Index, defined as a
weighted average of the dimension-specific HOI.16 This information was only available for ten of the 41
countries in Table 1, largely because the index has not been calculated in rich countries.
In Table 3, however, we list the component (or dimension-specific) human opportunity indices for
a larger set of countries, and for the following advantages (or “basic opportunities”, or “services”):
school attendance (10-14 year olds); access to water; access to electricity; access to sanitation; and
whether or not the child finished primary school on time (i.e. with zero grade-age delay). The indices are
multiplied by 100, so the possible range is 0-100. The 39 countries included - all of them in either Africa
or Latin America - is the full set available at the time of writing. As noted earlier, they come from
Molinas Vega et al. (2011) for Latin America, and World Bank (2012a, b) for Africa. Following the
authors, the table also reports the simple average of the school attendance and primary school
completion indices, as the HOI for education, and the simple average of the other three indices as the
HOI for housing conditions. The simple average of these two numbers in turn yields the overall HOI
reported in the last column of the table.
The motivation behind the HOI, as initially proposed by Barros et al. (2009), was to measure the
extent to which children in various developing countries have access to basic opportunities. Although
the authors do not motivate it this way, one could view the index as an example of the ex-ante approach
applied to a multidimensional advantage space, with each dimension corresponding to access to a
particular service – such as water or schooling – and the valuation of the opportunity set of each type
being given by the coverage of the service in that type. The particular inequality index applied to that
smoothed distribution of probabilities is the dissimilarity index (see equation 4).
16
The averaging procedure is the same suggested by Barros et al. (2011) for the HOI summary index: first calculate
a HOI for education obtained as the mean of the two education components and a HOI for housing conditions (the
mean of the other three components). Then obtain a summary HOI as a simple average of the two.
14
Although the dissimilarity index might therefore be seen as a measure of inequality of
opportunity, the HOI itself clearly cannot.17 It is intended – and defined – as a measure of average
access, adjusted (or penalized) by inequality of opportunity. Unsurprisingly, therefore, it is closely
correlated with other indicators of “level of development”. This association is already clear in Figure 7,
which ranks the average HOI for all countries in Table 3, ranging from 9.6 in Niger, to 91.6 in Chile. There
is almost no overlap in HOI between the African and the Latin American sub-samples, and the
correlation between the HOI and GNI per capita for these countries is 0.89 (p-value: 0.0005).
Perhaps more striking is the correlation with the UNDP’s Human Development Index which is even
higher (at 0.94) and highly statistically significant. Figure 8 presents the scatter plot. This is remarkable
because the two indices are constructed on the basis of completely different data. Until 2010 (the year
used in Figure 8), the Human Development Index was calculated as a simple average of three normalized
indices in the dimensions of health, income and education. 18 The income index used GNP per capita,
and the health index was based on life expectancy at birth, while the education index combined
information on literacy and the gross school enrolment ratio. Of these four basic components, only one
is close to the indicators used to construct the HOI, namely gross enrolment ratio, which is related to the
“school attendance” data used in the first column of Table 3. The other four components of the HOI,
listed above, do not enter directly into the computation of the HDI, and neither does the latter explicitly
adjust for dissimilarity across types in any way. Conversely, life expectancy at birth, GDP per capita and
literacy do not enter the HOI explicitly.
A correlation of 0.94 between these two indices, albeit calculated only over a non-representative
sample of 39 countries in two of the world’s regions, suggests two things. First, it suggests that the
average coverage rates of services like access to water, electricity, etc. are highly correlated with the
constituent elements of the HDI. Second, it suggests that the HOI is determined, to a very large extent,
by the first term in the product . In fact, the correlations between average coverage and the
component-specific HOI in this sample are extremely high: they are greater than 0.99 for school
attendance; access to water; access to electricity; and having finished primary school on time. It is 0.987
for access to sanitation. This implies, of course, that the penalty for inequality of opportunity, ,
accounts for a much smaller share of the variance in the HOI than mean coverage.
A final international comparison issue our data can shed light on is the association between the
dissimilarity index (the measure of inequality of opportunity contained within the HOI) and the index of
inequality of economic opportunity (IEO-R). The dissimilarity index can be interpreted as the proportion
17
A possible caveat with viewing the dissimilarity index within the HOI as a measure of inequality of opportunity is
that the index is typically calculated “for children”. This justifies the use of certain variables - like geographic
location or education of the adults in the household - as circumstances, which are clearly in the realm of choices
for the adults. The argument is that the index applies to children, and these are circumstances from their
perspective. But this then raises the issue of age of responsibility, and whether or not all inequalities in access to
services for children below a certain age should not be considered inequality of opportunity. Under that view,
unequal access to water or sanitation among five-year olds within the same type (i.e. sharing identical observed
circumstances) should also be counted as inequality of opportunity.
18
The correlation with the inequality-adjusted Human Development Index introduced for the first time in 2011 is
almost the same: 0.95.
15
of “basic opportunities” that is improperly allocated, relative to equal access across all types (Barros et
al. 2011). In other words, it is a measure of how much re-distribution in access to a particular service
would be required to move from the observed allocation to one in which average access was the same
across types. Subject to the caveat in footnote 17, this is a perfectly plausible measure of between-type
inequality in a particular dimension (that of service j). IEO-R, on the other hand, measures inequality of
opportunity as the between-type share of income (or consumption) inequality. How do these two
measures correlate? Do they yield essentially the same country ranking, even though their information
bases are quite different, as appears to be the case with the HDI and the HOI?
It is probably too early to answer this question in cross-country terms. The overlap between the
country samples in Table 1 (for which we have estimates of IEO-R) and in Table 3 (for which we have
estimates of the dissimilarity index) is only ten countries, six in Latin America and four in Africa. Very
little can be said, even about descriptive correlations, on the basis of such a small and unrepresentative
sample. Nevertheless, for what it is worth, Figure 9 plots the IEO-R index against the dissimilarity index,
averaged across its five dimensions. The correlation is -0.6989 (p-value: 0.0245), suggesting that the two
alternative approaches to measuring inequality of opportunity can yield very different country rankings.
It is true, of course, that in this sample the negative correlation is driven primarily by a dichotomy
between Africa and Latin America, where the latter has lower dissimilarity in access to services, but a
higher share of income inequality driven by unequal opportunities. Given that the IEO-R data for Africa
in our sample is based on coarser partitions than in most other cases, one really should not read too
much into this correlation. Nevertheless, it equally cannot be taken for granted that the IEO-R and the
part of the HOI which seeks to capture inequality of opportunity are measuring the same things.
5. Concluding remarks
Inequality of opportunity is a complex concept that can be measured in a number of different

ways. A number of measures have recently been proposed, both under the ex-ante and the ex-post
approaches, or indeed seeking a compromise between them. But most of these approaches have been
applied to a single country or a very small group of countries, making cross-country comparisons
impossible. Two exceptions are ex-ante measures of inequality of economic opportunity (IEO), and the
Human Opportunity Index (HOI). Our review of this empirical literature yielded (roughly) comparable
measures of the IEO for forty-one countries, and of the HOI for thirty-nine. Most countries in the first set
are in Europe and Latin America, but there are examples from North America, Asia, Africa and the
Middle-East. The second set covers countries in Africa and Latin America exclusively, and the overlap
between the two samples is ten countries.
The evidence reviewed suggests that an important portion of income inequality observed in the
world today cannot be attributed to differences in individual efforts or responsibility. On the contrary, it
can be directly ascribed to exogenous factors such as family background, gender, race, place of birth,
etc. There was considerable cross-country variation in the (lower-bound) relative measure of inequality
of economic opportunity: Brazil’s share (0.32) is sixteen times as large as Norway’s. Although there
certainly is noise in these measures, and various comparability caveats, there appears to be some signal
as well.
16
In addition, the data reveal a positive correlation between inequality of opportunities and income
inequality. Countries with a higher degree of income inequality are also characterized by greater
inequality of opportunity. This result is consistent with the empirical literature on social mobility, which
considers only one exogenous circumstance (family background measured on the basis of income or
social status of the parents) and finds a negative correlation between inequality and mobility (see the
“Great Gatsby Curve” of Corak, 2012): less unequal countries are also those that have a higher degree
intergenerational mobility.
In fact, the IEO-R measure is strongly positively correlated with two different measures of
intergenerational persistence (the converse of mobility): the intergenerational elasticity of income, and
the correlation coefficient of parental and child schooling attainment. It bears emphasis that these
measures of intergenerational transmission refer to different variables, collected in different data sets,
and reported by different studies. This suggests that the cross-country association between inequality of
economic opportunity and intergenerational mobility is rather robust.
In a sense, this is not surprising: inequality of opportunity is the missing link between the concepts
of income inequality and social mobility: if higher inequality makes intergenerational mobility more
difficult, it is likely because opportunities for economic advancement are more unequally distributed
among children. Conversely, the way lower mobility may contribute to the persistence of income
inequality is through making opportunity sets very different among the children of the rich and the
children of the poor.
We also found an inverted-U relationship between per capita GNI and inequality of economic
opportunity, reminiscent of the old Kuznets curve for income inequality. We argued that it is impossible
to treat that relationship as causal (in either direction), but that this is due primarily to the order of
aggregation of the two variables. It is quite possible that the relationship is underpinned by real
economic processes, although it is likely that disentangling them requires looking for specific
relationships among well-defined microeconomic variables.
Our international comparison exercise also revealed some interesting differences between the
IEO-R index and the Human Opportunity Index, even though both can be thought of as belonging to the
ex-ante family of I.Op. measures. These differences fall into at least three categories. First, the
advantage space for the IEO index is unidimensional, and usually refers to a measure of economic well-
being, such as income or consumption, while the HOI focuses on binary indicators of access to services.
If it is constructed as an average of the measure for different services, it can be thought of as having a
multidimensional advantage space (although aggregation across them is fairly ad-hoc).
Second, the HOI is deliberately constructed as a development index, with a functional form
analogous to Sen’s welfare index: a mean penalized by an inequality measure. The HOI is not a measure
of inequality of opportunity; it contains a measure of inequality of opportunities (in the space of access
to services), which is the dissimilarity index. As we have seen, however, most of the cross-country
variation in the HOI is driven by the mean coverage term, with correlations above 0.98 for each of the
five main dimensions usually included. Partly as a result, the HOI is very highly correlated with the HDI,
another famous aggregate development index, at least over the currently available sample of countries.
17
It is not obvious that the extent of this correlation is well-understood by the analysts working on either
approach.
Third, over the (small and unrepresentative) sample of countries for which both measures are
available, the dissimilarity index and the IEO-R – each an ex-ante measure of inequality of opportunity,
albeit with respect to different advantage spaces – are actually negatively correlated. While sample size
and comparability issues preclude taking this correlation too seriously, it may nevertheless serve as a
cautionary tale that different ways of measuring inequality of opportunity can measure (very) different
things, and yield widely disparate country rankings.
We argued in the introduction that fairness matters to people, and affects individual behavior.
There is also (anecdotal) evidence that measures of fair or unfair inequality matter to governments, and
international institutions like the World Bank increasingly use measures of inequality of opportunity in
country dialogue. We hope that this simple description of how the two most commonly-used measures
vary across countries, and co-vary with related indicators, may both contribute to greater clarity in those
discussions and help spur further analytical work.
18
References
Aaberge, Rolf, Magnus Mogstad, & Vito Peragine (2011): “Measuring Long-term Inequality of
Opportunity”, Journal of Public Economics, 95 (3-4), 193-204.
Acemoglu, Daron and James Robinson (2000): "Why Did the West Extend the Franchise? Growth,
Inequality and Democracy in Historical Perspective." Quarterly Journal of Economics 115(4): 1167–
99.
Almas, I., A.W. Cappelen, J.T. Lind, E. O. Sorensen and B. Tungodden (2011): “Measuring unfair
(in)equality”, Journal of Public Economics 95: 488-499.
Arneson, Richard (1989): “Equality of Opportunity for Welfare”, Philosophical Studies, 56, 77--93.
Azevedo V.M.R. and Bouillon C.P. (2010): “Intergenerational Social Mobility In Latin America: A Review
of Existing Evidence”, Revista de Analisis Economico, 25 (2): 7-42.
Barros, Ricardo, Francisco Ferreira, Jose Molinas and Jaime Saavedra (2009): Measuring Inequality of
Opportunity in Latin America and the Caribbean. Washington, DC: The World Bank.
Barros, Ricardo, Jose Molinas and Jaime Saavedra (2011): “Measuring Progress toward Basic
Opportunities for All”, Brazilian Review of Econometrics 30 (2): 335-367.
Belhaj-Hassine, Nadia (2012): “Inequality of Opportunity in Egypt”, World Bank Economic Review 26 (2):
265-295.
Bourguignon, François (1979): “Decomposable Income Inequality Measures”, Econometrica 47 (4): 901-
920
Bourguignon, François, Francisco Ferreira and Marta Menendez (2007): “Inequality of Opportunity in
Brazil”, Review of Income and Wealth, 53 (4): 585-618.
Bourguignon, Francois, and Thierry Verdier (2000): "Oligarchy, Democracy, Inequality and Growth."
Journal of Development Economics 62(2):285–313.
Cappelen, A. W., E. O. Sorenson and B. Tungodden (2010): “Responsibility for What? Fairness and
Individual Reponsibility”, European Economic Review (54): 429-441.
Cervini Pla M. (2009), Measuring intergenerational earnings mobility in Spain: A selection-bias-free,

Department of Applied Economics at Universitat Autonoma of Barcelona in its series Working Papers
n. wpdea0904.
Checchi, Daniele, & Vito Peragine (2010): “Inequality of Opportunity in Italy”, Journal of Economic
Inequality 8 (4), 429-450.
19
Checchi, Daniele, Vito Peragine and Laura Serlenga (2010): “Fair and unfair income inequalities in
Europe”. ECINEQ working paper 174-2010.
Christofides L. N., Kourtellos A., Theologou A., Vrachimis K. (2009): “Intergenerational Income Mobility
in Cyprus”, University of Cyprus, Economic Policy Research, Economic Policy Papers.
Cogneau D. and S. Mesple-Somps (2008): “Inequality of Opportunity for Income in Five Countries of
Africa”, DIAL Document de travail DT/2008-04.
Cohen, Gerry A. (1989: “On the Currency of Egalitarian Justice”, Ethics, 99, 906--944.
Corak Miles (2006): “Do Poor Children Become Poor Adults? Lessons from a Cross Country Comparison
of Generational Earnings Mobility”, IZA DP No. 1993.
Corak, Miles (2012): “Inequality from Generation to Generation: The United States in Comparison”, in R.
Robert Rycroft (ed.): The Economics of Inequality, Poverty and Discrimination in the 21st Century,
ABC-CLIO.
D'Addio A. C., (2007), Intergenerational Transmission of Disadvantage: Mobility or Immobility Across

Generations?, OECD Social, Employment and Migration Working Papers 52, OECD Publishing.
Dunn, C. (2007): “The Intergenerational Transmission of Lifetime Earnings: Evidence from Brazil”, The
B.E. Journal of Economic Analysis & Policy, 7 (2).
Dworkin, Ronald (1981a): “What is equality? Part 1: Equality of welfare”, Philos. Public Affairs, 10, 185--
246.
Dworkin, Ronald (1981b): “What is equality? Part 2: Equality of resources”, Philos. Public Affairs, 10,
283--345.
Fehr, Ernst, and Urs Fischbacher (2003): "The Nature of Human Altruism." Nature 425(October):785–91.
Fehr, Ernst, and Simon Gachter. 2000. "Cooperation and Punishment in Public Goods Experiments."
American Economic Review 90:980–994.
Fehr, Ernst, and Klaus M Schmidt. 1999. "A Theory of Fairness, Competition and Cooperation." Quarterly
Journal of Economics 114 (3):817–68.
Ferreira, Francisco H. G. (2001): “Education for the Masses? The interaction between wealth,
educational and political inequalities”, Economics of Transition 9 (2): 533-552.
Ferreira, Francisco H. G. and Jérémie Gignoux, (2011): “The Measurement of Inequality of Opportunity:
Theory and an Application to Latin America”, Review of Income and Wealth, 57 (4): 622-657.
Ferreira, Francisco H. G, Jérémie Gignoux and Meltem Aran (2011): “Measuring Inequality of
Opportunity with Imperfect Data: The case of Turkey”, Journal of Economic Inequality 9 (4): 651-680
20
Ferreira Sérgio G. and Fernando Veloso (2006): “Intergenerational Mobility of Wages in Brazil”, Brazilian
Review of Econometrics, 26 (2):.
Fleurbaey, Marc (1998): “Equality among Responsible Individuals” in J.–F. Laslier, M. Fleurbaey, N.
Gravel and A. Trannoy (eds.) Freedom in Economics: New Perspectives in Normative Economics.
London: Routledge.
Fleurbaey, Marc (2008): Fairness, Responsibility and Welfare, 1st Edition. Oxford: Oxford University
Press.
Fleurbaey Marc and Vito Peragine (2012): “Ex ante versus ex post equality of opportunity”, Economica
Foster, James and Artyom Shneyerov (2000): "Path Independent Inequality Measures," Journal of
Economic Theory, 91 (2): 199-222.
Grawe, N.D. (2004): "Intergenerational mobility for whom? The experience of high- and low-earning
sons in international perspective", Chapter 4 in M. Corak (ed.), Generational Income Mobility in
North America and Europe, Cambridge, Cambridge University Press, pp. 58-89
Hertz, T. Jayasunderay, T., Piraino P., Selcuk S., Smithyy N., Verashchagina A., (2007) “The Inheritance of
Educational Inequality: International Comparisons and Fifty-Year Trends”, B.E. Journal of Economic
Analysis and Policy (Advances), 7 (2): 1-46.
Hnatkovskay V., Lahiriy A., Pauly S. B. (2012): “Breaking the Caste Barrier: Intergenerational Mobility in
India” Department of Economics, University of British Columbia, Canada, mimeo.
Hugalde, A. S. (2004): “Movilidad intergeneracional de ingresos y educativa en España (1980-90)”,

Document de treball 2004/1, Institut d'Economia de Barcelona.
Kaldor, Nicholas (1956): “Alternative Theories of Distribution”, Review of Economic Studies, 23(2): 94-
100.
Kuznets, Simon (1955): “Economic Growth and Income Inequality”, American Economic Review 65 (1): 1-
29.
Marrero, Gustavo A. & Juan G. Rodríguez, (2010): “Inequality of opportunity and growth”, Working
Papers 154, ECINEQ, Society for the Study of Economic Inequality.
Molinas Vega, Jose, Ricardo Paes de Barros, Jaime Saavedra and Marcelo Giugale (2011): Do our children
have a chance? Washington, DC: World Bank
Nunez, J. and Miranda L. (2006): “Recent findings on intergenerational income and educational mobility
in Chile”, Universidad de Chile. Mimeo.
21
Pignataro, Giuseppe (2011): “Equality of Opportunity: Policy and Measurement Paradigms”, Journal of
Economic Surveys.
Piraino P. (2007): “Comparable Estimates of Intergenerational Income Mobility in Italy”, The B.E. Journal
of Economic Analysis & Policy, vol 7 n. 2.
Piraino P. (2012): “Inequality of opportunity and intergenerational mobility in South Africa”, paper
presented at the 2nd World Bank Conference on Equity. June 27th, 2012; Washington DC, USA.
Pistolesi, Nicolas (2009): “Inequality of opportunity in the land of opportunities, 1968-2001”, Journal of

Economic Inequality 7: 411-433.
Ramos, Xavi and Dirk van de Gaer (2012): “Empirical Approaches to Inequality of Opportunity: Principles,
Measures and Evidence”
Ravallion, Martin (2011): “Mashup Indices of Development”, Policy Research Working Paper 5432,
Washington, DC: World Bank.
Roemer, John (1993): “A Pragmatic Theory of Responsibility for the Egalitarian Planner”, Philosophy &
Public Affairs, 10, 146-166.
Roemer, John (1998). Equality of Opportunity. Cambridge, MA: Harvard University Press.
Singh A. (2011): “Inequality of opportunity in earnings and consumption expenditure: The case of Indian
men” Review of Income and Wealth 58 (1)
Van de Gaer, Dirk (1993): “Equality of opportunity and investment in human capital” Ph.D. Dissertation,
Katholieke Universiteit Leuven.
Voitchovsky, Sarah (2009): “Inequality and Economic Growth”, Chapter 22 in W. Salverda, B. Nolan and
T. Smeeding (eds.) Oxford Handbook of Economic Inequality. London: Oxford University Press.
World Bank (2006): World Development Report: Equity and Development. Washington, DC: World Bank.
World Bank (2012a): Do African Children Have a Chance? A Human Opportunity Report for Twenty
Countries in sub-Saharan Africa. Draft version June 2012
World Bank (2012b): South Africa Economic Update. Issue 3. July 2012.
Zhang Y. and Eriksson T. (2010), "Inequality of opportunity and income inequality in nine Chinese
provinces, 1989–2006". China Economic Review, 21, pp. 607-616.
22
Table 1: Inequality of opportunity, income inequality and economic mobility in 41 countries
Intergenerational
GNI per Total Intergenerational
Country IEO-L IEO-R Method correlation of HOI
capita PPP inequality income elasticity
education
Austria (1) 39,410 0.1800 0.0390 0.2167 parametric

Belgium (1) 37,840 0.1450 0.0250 0.1724 parametric 0.400
Brazil (3) 10,920 0.6920 0.2230 0.3223 parametric 0.5733 0.590 75.90
Colombia (3) 9,000 0.5720 0.1330 0.2325 parametric 0.590 79.25
Cyprous (1) 30,160 0.1700 0.0510 0.3000 parametric 0.3430
Czec Rep. (1) 23,620 0.1760 0.0190 0.1080 parametric 0.370
Denmark (1) 40,140 0.0830 0.0120 0.1446 parametric 0.0710 0.300
Ecuador (3) 9,270 0.5800 0.1500 0.2586 parametric 0.610 76.25
Egypt (5) 5,910 0.4230 0.0491 0.1160 non parametric 0.500
Estonia (1) 19,500 0.2430 0.0260 0.1070 parametric 0.400
Finland (1) 37,180 0.1360 0.0130 0.0956 parametric 0.1353 0.330
France (1) 34,440 0.1630 0.0210 0.1288 parametric 0.4100
Germany (1) 38,170 0.1910 0.0350 0.1832 parametric 0.2130
Ghana (2) 1,600 0.4000 0.0450 0.1125 non parametric 0.390 39.30
Greece (1) 27,360 0.2000 0.0340 0.1700 parametric
Guatemala (3) 4,610 0.5930 0.1990 0.3356 parametric 51.73
Guinea (2) 980 0.4200 0.0560 0.1333 non parametric
Hungary (1) 19,280 0.2080 0.0210 0.1010 parametric 0.490
India (8) 3,560 0.4218 0.0822 0.1949 parametric 0.5500
Ireland (1) 32,740 0.1880 0.0420 0.2234 parametric 0.460
Italy (1) 31,090 0.1960 0.0280 0.1429 parametric 0.4095 0.540
Ivory Coast (2) 1,650 0.3700 0.0500 0.1351 non parametric
Latvia (1) 16,360 0.2290 0.0280 0.1223 parametric
Lithuania (1) 17,880 0.2280 0.0350 0.1535 parametric
Luxemburg (1) 63,850 0.1480 0.0350 0.2365 parametric
Madagascar (2) 980 0.4400 0.0920 0.2091 non parametric 22.62
Netherlands (1) 42,580 0.1920 0.0360 0.1875 parametric 0.2200 0.360
Norway (1) 57,130 0.1300 0.0030 0.0231 parametric 0.2050 0.350
Panama (3) 12,980 0.6300 0.1900 0.3016 parametric 0.610 63.98
Peru (3) 8,940 0.5570 0.1560 0.2801 parametric 0.6000 0.660 69.18
Poland (1) 19,020 0.2710 0.0250 0.0923 parametric 0.430
Portugal (1) 24,710 0.2470 0.0300 0.1215 parametric
Slovakia (1) 23,140 0.1320 0.0180 0.1364 parametric 0.370
Slovenia (1) 26,970 0.1040 0.0050 0.0481 parametric 0.520
South Africa (6) 10,280 0.6750 0.1690 0.2504 parametric 0.7055 0.440 58.09
Spain (1) 31,550 0.2160 0.0420 0.1944 parametric 0.4533
Sweden (1) 39,600 0.1060 0.0120 0.1132 parametric 0.2125 0.400
Turkey (4) 14,580 0.3620 0.0948 0.2620 parametric
Uganda (2) 1,230 0.4300 0.0400 0.0930 non parametric 27.00
UK (1) 36,580 0.2040 0.0420 0.2059 parametric 0.4760 0.310
US (7) 47,020 0.2200 0.0409 0.1860 semiparametric 0.4800 0.460
Notes: The source for inequality and IEO measures for each country is given in parentheses after the country's name, and refers to the studies
below. GNI per capita is from the World Bank's World Development Indicators, for the year 2010, using PPP exchange rates for 2005. Total
inequality is measured by the mean logarithmic deviation in all cases except those from source (2), which use the Theil-T index. IEO indices are
always based on the same inequality measure used for total inequality in that country. Sources for the numbers in the last three columns are given
in the text.
(1) Checchi et al. (2010)
(2) Cogneau and and Mesple-Somps (2008)
(3) Ferreira and Gignoux (2011)
(4) Ferreira et al. (2011)
(5) Belhaj-Hassine (2012)
(6) Piraino (2012)
(7) Pistolesi (2009)
(8) Singh (2011)
23
Table 2: Comparing eight studies of ex-ante inequality of opportunity across 41 countries.
Number of
References Countries Data sources Outcome Method Circumstances
types
Austria, Belgium, Czech Republic, Germany,

Denmark, Estonia, Greece, Spain, Finland, parental education,
post-tax
Checchi et al. France, Hungary, Ireland, Italy, Lithuania, parental occupation,
1 EU-Silc 2005 individual parametric 72
(2010) Latvia, Netherlands, Norway, Poland, gender, nationality,
earnings
Portugal, Sweden, Slovenia, Slovakia, geographical location
United Kingdom.
Ivory Coast, EPAMCI, 1985-88
Cogneau and Ghana, 1998, GLSS per capita 3 groups based on
Ivory Coast, Ghana, Guinea, Madagascar, 6
2 Mesple-Somps Guinea, 1994, EICVM household non parametric father’s occupation and
Uganda. (3 Uganda)
(2008) Madagascar, 1993, EPAM consumption education, region of birth
Uganda, 1992, NIHS
Brazil, PNAD 1996;
Colombia, ECV 2003; gender, ethnicity,
Ferreira and Brazil, Colombia, Ecuador, Guatemala, Ecuador ECV 2006; household per parental education, 108
3 parametric
Gignoux (2011) Panama, Peru Guatemala, ENCOVI 2000; capita income father’s occupation, (54 Peru)
Panama, ENV 2003; region of birth.
Peru, ENAHO 2001
urban/rural, region of
Ferreira, imputed per
birth, parental education,
4 Gignoux, Aran Turkey TDHS 2003-2004 and HBS 2003 capita parametric 768
mother tongue, number
(2011) consumption
of sibling
gender, father’s
education, mother’s
Belhaj-Hassine total monthly
5 Egypt ELMPS 2006 non parametric education, father’s 72
(2012) eraning
occupation, region of
birth.
Individual gross
6 Piraino (2012) South Africa NIDS 2008-2010 parametric race, father's education 24
income
age, parental education,
individual annual
7 Pistolesi (2009) US PSID 2001 semiparametric father's occupation, 7,680
earnings
ethnicity, region of birth
father’s education,
household per father’s occupation, caste,
8 Singh (2011) India IHDS 2004–2005 parametric 108
capita earnings religion, geographical
area of residence.
Table 3: The Human Opportunity Index for five service indicators and 39 countries
HOI HOI
HOI HOI HOI HOI
School Finished HOI
Country Period Access to Access to Access to Housing HOI
Attendance primary on Education
Water Electricity Sanitation conditions
(10-14 yrs) time
Argentina 2008 96.80 97.30 100.00 64.40 82.60 89.70 87.23 88.47
Brazil 2008 97.30 82.50 96.40 78.20 34.90 66.10 85.70 75.90
Cameroon 2004 79.11 4.91 24.38 1.89 24.50 51.80 10.40 31.10
Chile 2006 98.40 93.90 99.20 86.10 82.00 90.20 93.07 91.63
Colombia 2008 93.00 54.00 100.00 77.00 70.00 81.50 77.00 79.25
Costa Rica 2009 95.50 95.40 98.80 92.80 66.40 80.95 95.67 88.31
Dem. Rep. Congo 2007 72.92 2.73 5.33 1.65 18.64 45.78 3.24 24.51
Dominican Republic 2008 96.50 70.10 95.40 48.80 53.40 74.95 71.43 73.19
Ecuador 2006 85.90 67.60 90.90 50.90 79.50 82.70 69.80 76.25
El Salvador 2007 89.40 18.30 83.00 18.60 42.50 65.95 39.97 52.96
Ethiopia 2011 69.09 0.93 5.61 0.14 15.75 42.42 2.23 22.32
Ghana 2008 84.59 4.90 36.70 3.91 42.26 63.42 15.17 39.30
Guatemala 2006 80.40 63.90 68.20 21.10 24.40 52.40 51.07 51.73
Honduras 2006 82.00 19.70 53.20 25.60 45.10 63.55 32.83 48.19
Jamaica 2002 95.00 23.40 85.40 35.70 93.00 94.00 48.17 71.08
Kenya 2008-09 93.34 8.36 4.92 1.53 47.31 70.32 4.93 37.63
Liberia 2007 59.10 1.03 1.04 4.70 8.45 33.78 2.26 18.02
Madagascar 2008-09 72.49 0.83 3.84 0.44 14.59 43.54 1.70 22.62
Malawi 2010 90.24 1.67 2.51 0.26 24.10 57.17 1.48 29.32
Mali 2006 39.32 3.17 6.14 1.08 10.85 25.09 3.47 14.28
Mexico 2008 92.50 80.30 98.30 72.00 86.70 89.60 83.53 86.57
Mozambique 2003 69.91 1.45 3.00 0.47 5.81 37.86 1.64 19.75
Namibia 2006-07 92.66 25.70 15.48 11.58 53.46 73.06 17.59 45.32
Nicaragua 2005 84.60 14.80 52.50 36.50 33.50 59.05 34.60 46.83
Niger 2006 29.98 1.03 2.54 0.17 5.88 17.93 1.25 9.59
Nigeria 2008 63.00 1.80 29.31 4.20 42.35 52.68 11.77 32.22
Panama 2003 90.80 50.20 60.20 31.40 70.60 80.70 47.27 63.98
Paraguay 2008 92.00 67.20 94.70 48.40 56.30 74.15 70.10 72.13
Peru 2008 95.00 42.60 64.40 54.40 74.10 84.55 53.80 69.18
Rwanda 2010 93.33 0.95 2.90 0.06 8.73 51.03 1.30 26.17
Senegal 2010-11 55.33 36.52 32.28 13.89 24.68 40.00 27.57 33.78
Sierra Leone 2008 65.73 2.37 3.24 0.61 24.41 45.07 2.07 23.57
South Africa 2010 98.72 20.57 78.82 24.95 50.74 74.73 41.44 58.09
Tanzania 2010 81.52 2.84 2.89 0.33 45.72 63.62 2.02 32.82
Uganda 2006 90.64 0.56 1.62 0.10 15.95 53.30 0.76 27.03
Uruguay 2008 94.80 89.30 98.20 96.60 78.40 86.60 94.70 90.65
Venezuela, R. B. de 2005 94.60 88.10 98.50 83.70 73.40 84.00 90.10 87.05
Zambia 2007 87.97 4.69 6.44 3.56 29.81 58.89 4.90 31.89
Zimbabwe 2010-11 92.05 8.48 12.63 7.58 78.00 85.03 9.56 47.30
Note: HOI Education is the simple average of HOI for school attendance and HOI for finishing primary school on time. HOI
Housing Conditions is the simple average of the other three individual HOIs. The last column is the simple average of the two
preceding sub-aggregates. This follows the authors in the sources below.
Source: Molinas Vega et al. (2011) and World Bank (2012a)
Figure 1: An illustration: inverse advantage distribution for three types
y k  pk  F31
3
F21
2
F11
 1
1 pk  F k  y 
Figure 2: inequality of economic opportunity: lower-bound estimates
Inequality of economic opportunity index (IEO-R)
26
Figure 3: Inequality of economic opportunity and the level of development
Figure 4: Inequality of opportunity and income inequality
Income Inequality (mean logarithmic deviation)
27
Figure 5: Inequality of opportunity and intergenerational mobility
Figure 6: Inequality of opportunity and the intergenerational correlation of education
28
Figure 7: The Human Opportunity Index in Africa and Latin America
Figure 8: The Human Opportunity and Development Indices
29
Figure 9: IEO-R and the Dissimilarity Index in the common subsample
30
MEASURING INEQUALITY OF OPPORTUNITIES FOR CHILDREN
Ricardo Paes de Barros, Jose R. Molinas Vega, and Jaime Saavedra

(with collaboration of Mirela Carvallo, and Samuel Franco and Samuel Freije).
1. INEQUALITY OF OPPORTUNITY: CONCEPTS AND ANALYTICAL FRAMEWORK

1.1. Concepts
Types of inequality 1
Even though absolute poverty and inequality are closely related concepts, the aims to
reduce poverty and inequality have received clearly different supports. While the call for
reductions in absolute poverty has always received universal and unconditional support,
being the first of the Millennium Development Goals, the call for reduction in inequality
has never received unconditional support.
The support for inequality reduction seems to vary among groups, as well as according to
the level and nature of inequality. When inequality is extremely high, as in most Latin
American countries, and closely related to family origin, gender and ethnicity, consensus
that inequality must be reduced can, in general, be more easily achieved. However, when
inequality is the result of differences in effort, choices or even talents among persons who
had access to the same opportunities and received the same treatment, then the support
for compensatory policies to reduce inequality is usually much weaker.
Thus, in the debate about public policy and inequality reduction we must recognize that
inequality is made of heterogeneous components, some much more unjust, undesirable
and unnecessary than others. Even though we could easily reach consensus toward
policies devoted to reduce or eliminate some of these components, consensus for fighting
other sources might never be achieved. Some potential sources of inequality may be
necessary to give everyone proper incentives (inequality caused by differential efforts).
Some may be tolerated (like the inequality due to differences in talents), in particular,
when attempts to reduce it could interfere with other ethical objectives, like privacy.
Hence, for the debate about the relation between public policy and inequality reduction to
advance, it is necessary to decompose overall inequality and evaluate the magnitude of its
components. Diagram 1 presents a basic framework to decompose inequality by source
and to relate the effort to reduce it to public policy. Next, we use this diagram to
elaborate on this breakdown of the overall inequality level.
1
Diagram 1 was constructed with the objective of providing a synthesis of inequality decomposition,
presented in this section, and its relation to public policy.
Diagram 1: Analytical Framework
Inequality of outcomes
(the quantity-quality debate)
Inequality of opportunity
Inequality due to choice and luck (inequality of outcomes beyond
(justifiable, but not desirable ) children control)
(unjust )
Inequality due to differences in Inequality of access to social

Second chance and social policy
talent and motivation (may be services among equally talented
(safety nets and minimum income
necessary as an incentive, may and motivated children
and access to social services)
be meritocratic ) (unproductive and unnecessary )
Social inclusion and social policy Inequality due to differences in Inequality due to differences in
(making everyone equally social treatment (discrimination family resources and location
productive) or inequality of treatment) (inequality of conditions)
Minimal resources guaranteed to

Anti-discriminatory laws (formal Targeting social programs to the
everyone (safety nets and
equality of opportunities) poor
conditional cash transfers)
Choices and circumstances
To a larger extent, there are two types of determinants on a person’s welfare. On one
hand, we have those factors that could have been different if the person had chosen
different paths, which were actually available to her. In this case, individuals with
identical choice sets reach different outcomes only due to differences in their choices.
Hence, one may argue that, in these situations, they must, themselves, be held
accountable or responsible for the resulting inequality.
On the other hand, we have differences in welfare among individuals who would have
made the same choices if they had had the same opportunities. In this case, all the
differences are due to exogenous factors that lead individuals to face different
opportunity sets. Contrary to the previous situation, now one must argue that they cannot
possibly be accountable or responsible for the resulting inequality. We usually refer to
this component as the inequality of opportunity.
Although many would argue that this source of inequality is intrinsically unjust, not
everyone would unconditionally agree that public policy must be used to reduce it. The
inequality of opportunity, as well as the overall inequality, is made of heterogeneous
components that must be disentangled before we can reach a consensus on its
justifiability, desirability and on how much we think public policy should be used to
reduced it.
Outcomes and Circumstances

The very concepts of outcomes and circumstances depend on context. If we define
inequality of opportunity as the inequality of outcomes due to factors beyond individual
control, then it would be natural to define circumstances as the factors (beyond individual
control) and outcomes as advantages that are influenced by them, but we wish they were
not.
In principle, outcomes themselves and their relation to circumstances, in particular, do

not need to be influenced by social choice. However, if the measurement of inequality of
opportunity is to have any policy relevance, outcome level and its relation to
circumstances must at least conceptually be influenced by social choice.
Hence, we define circumstances as factors beyond individual control (like race) that
should not but that actually do affect outcomes of interest (like wage). Outcomes of
interest are advantages (like wages) that can be modified by social choice (like subsidized
education). Moreover, outcomes (like wage) are in general dependent on circumstances
(like race) with this dependency, however, being, at least conceptually, modifiable by
social choice (like affirmative action programs). It is important to emphasize that
outcomes should depend on social choice through two channels. Social choice must
conceptually be able to modify both the level of the outcome and its dependence on
circumstances, see Figure 1.
Figure 1: The relationship between Outcomes, Circumstances and the role of Social
Choice
Improving the
level of the
outcome
outcome value
Reducing the
impact of
circumstances
Original
Situation
circumstances
Second chance and preference for equity
The fact that inequality caused by choice may be justifiable does not necessarily prevent
it from being undesirable. Even though we should, due to other competing principles,
agree that everyone has the right to make their own choices, nothing prevent us to cheer
for better choices. The freedom of choice does not prevent the society for preferring that
everyone makes the choices that are best for them. Hence, the resulting inequality in
welfare, although justifiable, may not be considered desirable or even necessary.
Support for public policies devoted to give a second chance to those who recognize
having not chosen properly the first time are evidence that, although justifiable, society is
still interested in devoting resources to reduced inequality caused by differences in
choices. Society does not seem indifferent to inequality, even when it is entirely caused
by differences in choices between equally well informed individuals with identical
opportunity sets.
Types of inequality of opportunity
Even when we constrain ourselves to exogenous sources of inequality, not all of them are
perceived as being equally undesirable. Hence, to better design public policy it is also
fundamental to isolate and measure each one of the components of the inequality of
opportunity. In this case it may be important to differentiate three types of factors:
intrinsic and personal characteristics, discriminatory treatment and the access to social
services.
First, we have the factors that, although still exogenous, are intrinsic to the person. These
are mainly biological factors (genetic inheritance and genetic lottery) and post-born daily
luck. To the extent that these factors lead to differences in productivity, achievement,
results or virtue (i.e., some being preferred to others as a function of what they can
perform or provide), in a meritocratic environment in which we give to everyone what
they deserve, the inequality so generated may be considered, if not just, at least
acceptable.
A recurrent problem with this source of inequality is how much advantage or recognition
should be assigned to productivity, achievement or virtue. A meritocratic society may be
perceived by some as valuing same talents too much, or it may be criticized by valuing
just a very narrow set of talents. Above all, for the inequality among unequally talented
individual to be considered fair, we need first to demonstrate that the values we are
attributing to talents are just.
Ideally, we would prefer to live in a society in which, despite all human diversity, every
person is equally deserving. In other words, we would prefer societies where a wide
variety of talents and virtues are valued. In this case, even though individuals may be
very different regarding their talents, talents’ values are such that, at the end, everyone is
equally productive and deserving. Even those strongly in favor of a meritocratic
environment would certainly dislike a society where only one type of virtue has value and
only a small fraction of the population has this virtue.
Although usually public policy has not been extensively used to modify the value of
talents and virtues, it can, and probably should, be used with this role. It may be an
important role for public policy to broaden the use and value of all talents (integrated and
inclusive schools may be an important policy in this group). It is certainly extremely
important to have a public policy that celebrates human diversity and values all types of
talents. Broadening the set of talents with value might lead to important reductions in
inequality of welfare.
In the next two topics we elaborate on the other two components of the inequality of
opportunity
Inequality of treatment and discrimination
But the inequality of opportunity does not result only from the inequality due to
deserving traits in a pure meritocratic environment. Quite often, equally talented and
productive individuals are treated differently, receiving different access to the best jobs or
receiving different wages even when performing the same tasks. In this case, inequality is
generated by the unequal treatment of equally deserving individuals. It constitutes a
violation of both meritocratic principles and equal opportunity values. We usually refer to
this unequal treatment of equals as discrimination.
Consensus on the unfairness of this source of inequality and on the fact that it is
avoidable and unnecessary seems easy to be achieved. Nevertheless, the amount of
resources societies are willing to allocate to eliminate this source of inequality could still
vary considerably and be open to debate.
There are at least two worth mentioning channels through which discrimination could be
generated. In the first one, it operates through the unequal access to social services (like
education or health). Due to this unequal access, inequality in valuable acquired
characteristics, such as formal education, might be produced even among equally talented
persons. This inequality in valuable traits will lead to unequal outcomes even in societies
where everyone receives what they deserve (i.e., outcome-meritocratic societies).
Naturally, the problem with these societies is the lack of equal opportunity to acquire the
valuable characteristics that are being so meritocratically translated in higher desirable
outcomes (income and welfare, for instance).
Through the second channel, discrimination operates after the access to services is
granted. In this case, even when everyone is treated equally when accessing relevant
social services, differential treatment is received during the service, leading to differential
benefits. Different ethnic groups may be equally treated for admission to a school, but
discriminated while studying. Likewise, different genders may be equally treated when
considered for a job, but they may be paid differently for performing the same task.
People dressing differently might get equal access to a restaurant, but they might not
receive the same treatment by the waiter.
Through either of the two channels formal equality of opportunity is violated, since
equally talented and motivated individuals are being treated differently, leading to
differential access to opportunities and to better positions. Anti-discriminatory policies
would be the appropriate and necessary measure in these situations.
Inequality of access to social services
A third component of the inequality of opportunity results from the unequal access to
social services, due to differences in family resources. In this case, children from poor
parents are not being discriminated, they just do not have equal access to services,
considered important for them, to develop and fully utilize their talents, just because their
families lack the necessary resources. In fact, one may be hungry just because one does
not have the necessary resources to buy food.
Lack of resources may impair not just the access to social services but also the ability to
benefit from these services. Children from poor parents may have both: more limited
access to schools and learning disadvantages. They may not have books at home or they
may have parents with greater difficulty to read for them.
Whenever the access or the ability to benefit from social services depends on family
resources, the ideal of equal opportunity is violated and social immobility is generated. In
this case, equally talented children from different social backgrounds are not going to
have the same opportunities and outcomes for reasons outside their control.
The role of public policy in this case is unanimously recognized. Equal access to social
services must be provided to all. Equal conditions for equally benefiting from these
services must also be guaranteed to all. The natural implications are (a) subsidized access
to social services for all, or at least for the poor, and (b) a guaranteed minimum income to
ensure that all have the necessary conditions for benefiting from available services.
The central role of the inequality of access to social services and the equal opportunity to
benefit from them
As we have emphasized in the previous section, for getting support and for designing
adequate public policies aimed at reducing inequality it is important to disentangle the
overall inequality into its components. Among them, the unequal access to social services
and the unequal opportunity to benefit from these services, due to discrimination or
differences in family resources, are perhaps the components whose reductions command
the greatest consensus. They are also, probably, the components more suitable for direct
public policy interventions.
To the extent unequal access to social services and unequal opportunity to benefit from
them are due to exogenous factors (like gender, ethnicity, place of origin or family
background), they are not just part of the overall inequality, but certainly also a
constituent part of the inequality of opportunity. Even though the concept and
measurement of inequality of opportunity may remain very debatable, there are no doubts
that all the unequal access to social services and the unequal opportunity to benefit from
them, due to differences in family resources and discrimination, are integral parts of it.
This is particularly true when we concentrate the attention on children. For them, family
resources and composition are certainly exogenous.
Moreover, even among the components of the inequality of opportunity, the unequal
access to social services and the unequal opportunity to benefit from them, due to
discrimination and lack of family resources, play a major and central role. In fact,
although the complete endogeneity of choice (are people really free to choose? All of
them are equally informed and have the same opportunity set?) and the fairness of
rewarding inherited talents are always under questioning, nobody would question the
unfairness of children’s unequal access and opportunity to fully benefit from social
services due to discrimination or differences in family resources.
1.2. FROM CONCEPTS TO MEASUREMENT
Inequality of opportunity has always received great attention in the policy debate as well
as by philosophers 2 . Nevertheless, empirically it does not seem to have a universal
definition. It has traditionally been measured as the association between family
background and children’s outcomes by sociologists. 3 Recently, economists have been
given growing attention to the measurement of inequality of opportunity for continuous
outcomes like income 4 . In this section we seek to review the main notions of inequality
of opportunity important for empirical and applied studies.
Defining an empirically tractable notion of inequality of opportunity
Let y be an individual outcome of general interest: education, income or access to an

important social service, for instance. Let us divide its determinants in two broad groups.
One group made of all factors beyond individual control. Let c denote a vector with all
these determinants. We refer to them as circumstances. The other group is formed by all
factors under individual control. Let e be a vector with all these determinants. We refer to
them as choices or efforts. Individuals are considered fully responsible for choices and
effort.
2
For recent reviews of this literature in economics, see Paragine (1999) and Fleurbaey and Maniquet
(2004). For classical treatments in the philosophical literature, see Dworkin (1981a,b), Cohen(1989) and
Arneson (1990). For excellent recent discussions on the concept of equality of opportunity see Hild and
Voorhoeve (2004) and Fleurbaey (2005).
3
See, for instance, the classical work of Boudon (1974).
4
See, for instance, Betts and Roemer (1999), Checchi, Ichino and Rustichini (1999), Benabou and Ok
(2001), Goux and Maurin (2003), Bouguignon, Ferreira and Menéndez (2007), Checchi and Peragine
(2005), Waltenberg and Vandenberghe (2007), Schuetz, Ursprung and Woessmann (2005), Brunello and
Checchi (2007), Lefranc, Pistolesi and Trannoy (2006),
Even though this dichotomist decomposition is widely used, it is not without major
difficulties. Surely, factors outside individual control are easy to define and to recognize,
choices however are much more difficult to classify. Suppose the choice set for two
persons are completely distinct. Their choices, as a consequence, ought to be quite
different too. Would we consider the differences in the choices they made a result of
personal decisions or a result of exogenous differences in the choice sets? If, due to
factors beyond individual control, the choice sets are quite different, then it will be
awkward to consider the individuals fully responsible for the choices they made.
Let ψ be the function relating the outcome to its determinants. Hence, y = ψ (c, e ) and the
outcome distribution and its corresponding degree of inequality are a function of ψ and
the joint distribution of its determinants (c, e ) .
Equality of opportunity as the stochastic independence of outcome and circumstances
Under perfect equality of opportunity, circumstances should not have any impact on
outcome. More precisely, the outcome, y, and the circumstances, c, would ideally be
stochastically independent, i.e., the distribution of y conditional on c should be the same
for all possible values of c. If y were income and c, race, equal opportunity would require
the distribution of income among one race to be identical to the distribution among other
races. Under this definition, a measure of inequality of opportunity would be a measure
of the differences between the outcome conditional distributions. If race were the unique
circumstance, inequality in opportunity would be a measure of the distance from the
outcome distribution for blacks to the corresponding distribution for whites, for instance.
It should be clearly understood, however, that the distance between these conditional
distributions and, hence the magnitude of the inequality of opportunity, is determined by
the shape of the function ψ as well as by the stochastic dependence between effort and
circumstances. In fact, for the ideal of equal opportunity to be achieved, it is necessary (a)
that circumstances are not really among the outcome determinants and (b) that effort and
circumstances are not stochastically dependent. For instance, if income is a function of
race and effort, then the distance between the distribution of income among white and
blacks will depend on race, in part, to the extent that race has a direct impact on income
and in part to the extent that the distribution of effort is different in the two groups.
To see why the stochastic independence of effort and circumstances is not sufficient,
let y = c + u . In this case, Fy (t c) = Fu (t − c) . The outcome conditional distribution
depends on c. To see why ψ being invariant with c is not sufficient, let y = u . In this case,
Fy (t c) = Fu (t c) and the outcome conditional distribution depends on c, as long as effort
and circumstances are not stochastically independent.
Equality of opportunity as equal outcome for equal effort

There is, however, an alternative definition of equal opportunity were the stochastic
dependence of effort on circumstances is not included. In this alternative we reverse the
conditioning and consider the distribution of outcomes conditional on effort. Equality of
opportunity would imply perfect equality of outcomes given effort, i.e., equal outcome
for equal effort. In this case the inequality of opportunity should be measured not as a
contrast between outcome distributions conditional on circumstances, as in the previous
case, but directly as the inequality in the outcome distributions conditional on effort.
Naturally, there will be many measures of inequality in opportunity, one for each choice
of the level of effort.
The two approaches are not equivalent at all, and the choice between them poses serious
ethical questions. After all, the dependence of effort on circumstances is a departure from
the equal opportunity ideal. Only when effort is stochastically independent of
circumstances the two approaches would be identical.
Before going into measurement issues it should be highlighted that effort being a choice
does not imply that all differences in outcome due to effort are fair or compatible with the
equality of opportunity. For instance, people may be discriminated based on their choices
and nobody would consider this situation as fair. If due to pure discrimination, the wage
of a worker is lower as a result of his choice of religion, sexual orientation, local of
residence (favela, for instance) or the race of his/her spouse, then we have a situation
where an important outcome depends on choices in a clearly unfair way. Hence, for all
outcome differences due to efforts to be compatible with equality of opportunity, we have
to be sure that the relation between effort and outcome is perfectly meritocratic.
Clearly, these two concepts of equal opportunity have quite different informational
requirements to be measured. In the first approach, measures could be obtained without
effort being observed and the function ψ being known or estimated. In the second
approach, however, either effort must be observed or the function ψ needs to be
estimated. But, how ψ could be estimated when effort is unobserved and stochastically
dependent on circumstances?
1.3. UNOBSERVED CIRCUMSTANCES AND STOCHASTICALLY DEPENDENT EFFORT
In order to construct a practical index of inequality of opportunity we are going to

proceed as if (a) the relation between log-outcome, circumstance and effort were
additively separable, (b) effort were stochastically independent of circumstances, (c) the
relation between effort and outcome were meritocratic, and (d) all circumstances were
observed. Surely, none of these conditions are strictly valid in the real world.
Unfortunately, we do not have information necessary to adjust the index to these

departures. All we can actually do is to clarify the consequences of these departures and
stress the care that one should have when using this index for policy recommendations.
The next two sections are dedicated to identify the problems and highlight their potential
consequences.
As before, let y be an observed outcome and c a vector with all circumstances. Since we
do not observed all of the circumstances and some are more prone to policy intervention
than others, we are going to concentrate on a set of observed socially-determined
circumstances. Let x denote a vector with all these observed socially-determined
circumstances. We are going to treat as socially determined all the circumstances that are
not intrinsic to the person. They result from differences on how a person is socially
treated. Examples are differences related to family background, differential treatment due
to gender and race, and different access to public services due to place of birth.
To complete the list of determinants, let z denote the additional set of socially-determined
but unobserved circumstances and a, the set of non-socially determined circumstances 5 .
Finally, let w denote measures of efforts. Hence, y = ψ ( x, z , a, w) . Additionally, assume
that either the factors ( z , a, w) are not observed or we have decided not to control for
them. Hence, from a practical point of view they can be treated as unobservable. To
simplify, assume the relation determining the outcome is log-additive. Hence,
Ln( y ) = g ( x ) + h1 ( z ) + h2 (a) + h3 ( w)
Given that y and x are observed we could estimate, for instance, the log-mean regression
of y on x, i.e., we could estimate μ ( x ) ≡ E (Ln( y ) x ) . Unless the excluded variables
( z , a, w )
are mean-independent of the observed socially-determined circumstances, x, it
will not be, in general, true that g ( x) = α +μ ( x ) . If they were mean-independent, then, of
course, this relation would be valid with
α = −(E (h1 ( z ) ) + E (h2 (a ) ) + E (h3 ( w) ))
In general, however,
μ ( x) = g ( x)+ E (h1 ( z ) x ) + E (h2 (a ) x ) + E (h3 ( w) x ) (1)
5
Even though it would be nice to treat all choices as effort and all exogenous variables as circumstances,
this separation is not always possible. Consider a worker who chooses to live in a nearby favela and to be
self-employed, as opposed to live in a distant community and work as an employee in a factory. Let us
assume this choice is rational and improve his welfare. In addition, suppose that teachers discriminate
children from favela and employer discriminate workers from favela. As a result of the discrimination this
worker may prefer to be an informal self-employed and his children may evade school earlier. In this case,
the advantages of proximity with a better labor market dominate the discrimination effects. For this family
it is better to be discriminated but to live nearby the labor market, than to avoid being discriminated by
working in a distant factory as a low-paying employee. Nevertheless, it remains true that this family
standard of living would be much better if there were no discrimination against favela. As it is always the
case, discrimination is a factor leading to equally hard-working people to achieve different outcomes. It
should, therefore, be considered a source of inequality of opportunity, even when families choose the
situation where they are discriminated.
In our practical applications we are going to use the inequality in μ (x ) to measure the
dependence of the outcome on socially-determined circumstances and to construct an
index of inequality in socially-determined opportunities. As expression (1) reveals, this
index will be also capturing a host of other effects in addition to the impact of observed
socially-determined circumstances on outcome. In the next section we aim to organize
and illustrate these potential departures. The reader only interested in the computation of
the inequality of opportunity index could skip the next section, since its contents matter
only for the interpretation of the measure we propose.
1.4. WHY WOULD OUTCOMES AND OBSERVED SOCIALLY DETERMINED CIRCUMSTANCES

BE CORRELATED?
As expression (1) aims to clarify, in addition to their direct causal link, there are at least
three reasons for an outcome to be correlated with observed socially-determined
circumstances. Their formal origins are well described in expression (1). In this section
we aim to illustrate their substantive content.
The association between observed and unobserved socially-determined circumstances

E (h1 ( z ) x ) ≠ E (h1 ( z ) )
To the extent that observed and unobserved socially-determined circumstances are

correlated, μ (x ) will also capture at least part of the impact of unobserved socially-
determined circumstances. For instance, to the extent that living in a favela is correlated
with race, the fraction of the inequality of opportunity we are going to assign to race will
bias its own real effect. In fact, even when teachers do not discriminate by race, but do
discriminate residents from favela, we are going to capture some inequality in
educational opportunity by race. Hence, our indicator has difficulties in isolating the
contribution of each observed circumstance.
The association between observed socially-determined circumstances and intrinsic

circumstances E (h2 (a ) x ) ≠ E (h2 (a ) )
To the extent that intrinsic circumstances and observed socially-determined

circumstances are correlated, μ (x ) will also capture at least part of the impact of intrinsic
circumstances. For instance, suppose that parents´ talents are correlated with their
children talents and own talent is correlated with own education. In this case, some of the
inequality of opportunity we assign to differences in parents´ education may overestimate
their contribution. In fact, even when parents´ education has no effect on their children
education, we are still going to capture some inequality of opportunity associated to
parents´ education due to the inheritance of talents. Again, our indicator has difficulties in
isolating the real contribution of a specific circumstance.
The association between effort and observed socially-determined circumstances
E (h3 ( w) x ) ≠ E (h3 ( w) )
Suppose that differences by race in educational effort are a reaction to discrimination in

school. In this case, the inequality of opportunity we assign to race includes not only the
direct effect of discrimination in school, but also its indirect effect on effort. Our
indicator will not be comparing outcomes between students who are putting the same
amount of effort. Our indicator will consider all differences by race including those
related to effort.
Even more worrisome is the possibility that differences in effort – which we are including
in our measure of inequality of opportunity – may be due to the anticipation of future
discrimination in the labor market. Hence, we could estimate a sizeable degree in
educational opportunity related to race, even when schools are a perfect non-
discriminatory institution. This example illustrates that our measure may have severe
difficulties in identifying where the inequality of opportunity really occurs.
Furthermore, since differences in public policy (for instance, any form of positive
discrimination) may influence effort, gender differences in work effort could be, for
instance, related to the generosity of maternity benefits. In this case, we are going to
consider as inequality of opportunity differences in effort generated by discriminatory
social policy.
The association between observed socially-determined circumstances, risks and other

exogenous random events
Some groups may choose greater risks or they may be forced to bear greater risks. In both
cases, outcomes may be related to group membership, even though there are no other
forms of discrimination. For instance, black children may have worse performance at
school because they have poorer health due to uneven access to basic preventive health
care. Moreover, due to their greater poverty, blacks may choose more hazardous working
activities and as a consequence spent a larger fraction of their life without working, due
to poor heath conditions. In the first case, we are going to estimate a degree of inequality
in educational opportunity that is unrelated to schools. In the second case, we are going to
estimate a degree of inequality in working opportunity that is totally related to differences
in endogenous choice.
Summary
As all these situations clearly emphasize, we have to be very careful in interpreting the
measure of inequality of opportunity we are going to estimate. It may be capturing
differences in effort and luck stochastically associated to circumstances. It can also be
ascribing differential treatment in the labor market or in the access to heath to inequality
in educational opportunity. Finally, it may be attributing differences to socially-
determined inequality of opportunity due to intrinsic circumstances like talents.
1.5. THE DIRECT AND THE RESIDUAL APPROACHES
If the inequality of opportunity is defined as the inequality in average outcome between

groups defined in terms of their common circumstances, we could estimate it based on
two types of counterfactuals. On one hand, we could define it as the outcome inequality
that would remain if all the inequality within groups with the same circumstances were
eliminated. On the other hand, we could define it as the difference between the overall
outcome inequality and the level that would remain if all the inequality in means between
groups with the same circumstances were eliminated. In the sequence, we define
precisely each of these two approaches and identify when they are equal.
Inequality of opportunities as the between component of the inequality in outcomes
To isolate the outcome inequality between groups it is enough to eliminate the inequality
within groups and re-compute the overall inequality. More precisely, we change the
outcome for every person to equal the average of her circumstance group. More
precisely, we make
yb (ω ) = E ( y x = x(ω ) ) = λ ( x(ω ))
where x = x(ω ) represents the set of circumstances of a specific group. In this case, if
I ( y ) is a proper measure of inequality in the distribution of y, then the inequality of
opportunity IOb ( y x) will be given by IOb ( y x)= I ( yb ) . Notice that this measure of
inequality of opportunity does not require y to be observed or the overall inequality
I ( y ) to be measure, as long as the function λ is known and the socially-constructed
circumstances observed.
Let’s work the precise expression of this indicator for the Theil-L inequality measure. To
begin with, notice that by its definition
L( y )= Ln(E ( y )) − E (Ln( y ))
Hence,
IOb ( y x) = L( yb )= Ln(λ ( x) ) − E (Ln(λ ( x) ))
The relation between the overall inequality and the inequality in opportunities can be
perceived by the expressions
L( y )= E (Ln(E ( y x )) − E (Ln( y ) x )) + Ln(E (E ( y x) )) − E (Ln(E ( y x )))
or
L( y )= E (Ln(λ ( x) ) − E (Ln( y ) x )) + Ln(E (λ ( x) )) − E (Ln(λ ( x) ))
Since,
L( yb )= Ln(λ ( x) ) − E (Ln(λ ( x) ))
Hence,
L( y )= E (Ln(λ ( x) ) − E (Ln( y ) x )) + L( yb )
Inequality of opportunities as the difference between the overall outcome inequality and
the within component
In this approach we estimate the inequality in opportunity by how much the outcome
inequality would decline if the inequality between circumstance groups were eliminated.
To estimate the inequality of opportunity in this case, it is therefore necessary to isolate
the within component of the outcome inequality, i.e., the inequality that would remain if
all inequality between circumstance groups were eliminated. To isolate the inequality
within groups, we should change the outcome for every person proportionally to the gap
between the average for the group and the overall average. More precisely, we make
μ . y (ω )
yd (ω ) =
λ ( x(ω ))
In this case if I ( y ) is a proper measure of inequality in the distribution of y, then the

inequality in opportunity IOd ( y x) will be given by IOd ( y x)= I ( y ) − I ( yd ) .
Next, we obtain the precise expression of this indicator for the Theil-L. To begin with,
notice that by its definition
L( y )= Ln(E ( y )) − E (Ln( y ))
Hence,
⎛ ⎛ μ. y ⎞ ⎞ ⎛ ⎛ μ. y ⎞ ⎞
L( yd )= Ln⎜⎜ E ⎜⎜ ⎟⎟ ⎟⎟ − E ⎜⎜ Ln⎜⎜ ⎟⎟ ⎟⎟ = E (Ln(λ ( x) )) − E (Ln( y ))
⎝ ⎝ λ ( x ) ⎠⎠ ⎝ ⎝ λ ( x ) ⎠⎠
The relation between the overall inequality and the within inequality can be expressed as
L( y )= (E (Ln(λ ( x) )) − E (Ln( y ))) + L( yb )
Hence,
L( y )= L( yd ) + L( yb )
Therefore
IOd ( y x)= L( y ) − L( yd ) = L( yb ) = IOb ( y x)
2. MEASURING INEQUALITY OF OPPORTUNITY FOR DISCRETE OUTCOMES
If the inequality of opportunity is defined as the inequality in average outcome between

groups defined in terms of their common circumstances, we could estimate it based on
two types of counterfactuals. As explained in the previous section, we could define it as
the outcome inequality that would remain if all the inequality within groups with the
same circumstances were eliminated: the direct approach. On the other hand, we could
define it as the difference between the overall outcome inequality and the level that
would remain if all the inequality in means between groups with the same circumstances
were eliminated: the indirect approach. In the case of discrete outcomes the direct
approach has important advantages. For this reason we based our development of a
measure of inequality of opportunity for discrete outcomes on this approach.
2.1. THE ADVANTAGES OF THE DIRECT APPROACH FOR MEASURING INEQUALITY OF

OPPORTUNITIES FOR DISCRETE OUTCOMES
As already mentioned, the direct approach requires only the knowledge of the regression
function, whereas the residual approach would need information on outcomes and
regression residuals. When outcomes are completed observed, the two approaches convey
equivalent information. However, when outcomes are only partially observed the direct
approach may have great practical advantages. In this section we clarify these
advantages.
As before, assume that, y = λ (x) + ε . But now assume that we only observe if the
outcome is smaller or greater that η, let d be an indicator of this event. Hence,
d = 1 ⇔ y > η and d = 0 ⇔ y ≤ η . The only observed variables are the indicator d and
the socially-determined circumstances x.
Since, μ ( x) ≡ E ( y x ) it follows that E (ε x ) = 0 . Let strength this condition, assuming that

ε is stochastically independent of x and have a known distribution, Fε . Hence,
E (d x ) = P (d = 1 x ) = P( y > η x ) = P(ε > η − μ ( x) ) = 1 − Fε (η − μ ( x) )

Since the pair (d , x) is observed we can estimate the regression function E (d x ) . To the
extent that Fε is known, we could obtain μ (x) , up to a constant, without fully observing
y, via
μ ( x) = η − Fε−1 (1 − E (d x ))
and proceed to construct a measure of inequality of opportunity as described in the

previous section.
It must, however, be clarified that given all these hypotheses the residual approach could
also be pursued. In fact, the distribution of the residuals we know by hypothesis.
Therefore, once we have estimated the regression function, we could obtain the outcome
distribution, since by hypothesis ε is stochastically independent of x. It is still true that we
can not recover the outcome values for each person, but its distribution is completely
determined combining the observed data with the hypotheses we have made.
To assume the distribution of ε is known may be a very high price to be paid to construct
an indicator of equality of opportunities. Mainly, because we can avoid this hypothesis if
we construct an inequality of opportunity measure entirely based on E (d x ) . Let us
denote E (d x ) by p ( x ) , i.e., p(x ) ≡ E (d x ) . If, instead of insisting in measuring the
inequality of opportunity from the inequality in μ ( x ) we accept to measure it from the
inequality in p( x ) , we could obtain a measure of inequality in opportunity without
having to assume a distribution for effort.
We have, nevertheless to keep the hypothesis that effort is stochastic independent of

circumstances, otherwise E (d x ) would not be a function of x only through μ ( x ) , but
also due to the stochastic relation between effort and circumstances. In other words, for
E (d x ) = 1 − Fε (η − μ ( x) ) we require ε and x to be stochastically independent, otherwise
we would have E (d x ) = 1 − Fε (η − μ ( x) x ) , and there would be two sources of
dependence on x built in this regression: one related to the direct impact of circumstance
on outcome, the source of inequality of opportunity we want to measure, and another due
to the nuisance stochastic dependence of effort on circumstance.
Measuring differences between distributions
Given estimates for the regression function p ( x ) and for the distribution of x, many
measures of inequality could be easily computed. All of them would be measures of the
inequality of opportunity. However, the binary nature of d and the fact that p ( x ) is a
conditional probability and so bounded between 0 and 1 permit and may also require
specialized treatment. Let p ≡ E (d ) = P (d = 1) .
We begin noticing that, as a consequence of the binary nature of d, mean-independence is
equivalent to full stochastic independence. Hence,
p ( x ) = p ⇔ Fx d =1 = Fx d = 0
In other words, the probability of a low outcome is independent on the circumstances, if

and only if, the distribution of circumstances among those with low outcomes is identical
to the distribution of circumstances among those with high outcome. Therefore, we could
measure the degree of inequality of opportunity by computing the distance from the
distribution of circumstances for those with low outcomes to the distribution of
circumstances for those with high outcomes. One advantage of this procedure is that we
could select any measure from the large literature on measuring distances between two
distributions.
Two measures are widely used. In sociology and demography, the Dissimilarity Index is
the most commonly used. In statistics, the Kullback-Leibler measure is probably
preferred. We are going to concentrate our attention on the dissimilarity index due to its
immediate, intuitive and simple interpretation. 6
Let us consider a situation where circumstances have a discrete distribution and

{x1,..., xm } denotes the set of all possible values. Then, the dissimilarity between the
distribution of circumstances among those with high outcomes, f1 ( xk ) = P (x = xk y > η )
and the distribution of circumstances among those with low outcomes,
f 0 ( xk ) = P (x = xk y ≤ η ) is defined as
m
D = Λ.∑ f1 ( xk ) − f 0 ( xk )
k =1
1− p
where, Λ = and as defined before p = P( y > η ) = P(d = 1) . The Dissimilarity Index
2
is, therefore, proportional to the absolute distance between the two distributions. The
proportional factor Λ plays an important role in providing a simple interpretation for this
index. Before considering its interpretation, notice that it could be alternatively expressed
as the absolute distance between the distribution of circumstances among those with high
outcomes, f1 ( xk ) = P (x = xk y > η ) and the overall distribution of circumstances,
f ( xk ) = P( x = xk ) . In fact,
6
An alternative to the dissimilarity index is the Kullback-Leibler measure. As a measure of the distance
between the distribution of circumstance among the high outcome persons and the overall distribution of
circumstances this measure would be given by
m
⎛ f (x ) ⎞
KL = ∑ f1 ( xk ).Ln⎜⎜ 1 k ⎟⎟
k =1 ⎝ f ( xk ) ⎠
1 m
D = .∑ f1 ( xk ) − f ( xk )
2 k =1
Moreover, since f ( xk ). p( xk ) = f1 ( xk ). p , it follows that the Dissimilarity Index could also

be expressed as
m
D = Γ.∑ p(xk ) − p f (xk )
k =1
1
where, Γ = . Hence, the Dissimilarity Index is proportional to the area between the
2p
mean-regression function, p(x) , and the overall average p (see Figure 2).
Figure 2: Graphical representation for Dissimilarity Index
70
probability of com pleting sixth grade on tim e
65
60
55
50
45
40
35
Inequality of
30 Opportuntity
1 m
∑
25
20
D= β i pi − p
15
2 p i =1
10
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
per capita income (US$ (PPP)/day)
To see the usefulness of the proportional factors, Λ and Γ , consider a finite population
with N members. Hence, L = N . p would be the total number of persons with high
outcome, and L( xk ) = N . f ( xk ). p( xk ) the corresponding number among those with
circumstances equal to xk . Moreover, notice that L ( xk ) = N . f ( xk ). p would be the
expected number of persons with high outcome if equal opportunity prevails. Hence, a
natural measure of inequality in opportunity would be
1 m
IO = .∑ L( xk ) − L ( xk )
2 L k =1
Notice that IO = D . Moreover, IO represents the minimum fraction of the total number
of persons with high outcome, L, that needs to be redistributed across circumstance
groups in order to achieve equal opportunity, i.e., to ensure equal proportion to low
outcome persons in all circumstance groups, p( xk ) = p .
2.2. COMPUTING THE DISSIMILARITY INDEX
Assume that one has access to a random sample of the population with information on
whether person i had or not access to a given opportunity ( I i = 1 if that person had access
and I i = 0 otherwise) and a vector of variables indicating his/her circumstances,
xi = ( x1i ,...., xmi ) .
Given this information, one need to follow three steps to estimate the inequality of
opportunity index
E P (I = 1 x ) − P (I = 1)
D=
2 P (I = 1)
Before we proceed, however, it is worth noticing that since,
P (I = 1) = E (P (I = 1 x ))
we can rewrite D as
E P(I = 1 x ) − E (P (I = 1 x ))
D=
2 E (P(I = 1 x ))
which is actually the expression we mimic in order to estimate D. This expression also
indicates the central role of group specific coverage rates, P (I = 1 x ) , in estimating D.
These conditional probabilities could be estimated through a variety of parametric,

nonparametric or semi-parametric procedures. One could impose separability restrictions
or consider interactions. In all case the three step procedure we describe in this note
would apply.
Since we were planning to apply this procedure to all Latin America and Caribbean
countries with available data and for several points in time we privileged a standard
specification that could be feasibly applied to all countries at all times. Our choice was a
separable logistic model. Hence, our first step in estimating D was to fit the following
separable logistic regression
⎛ P(I = 1 x1 ,..., xm ) ⎞ m
⎟ = hk ( x k )
⎜ 1 − P(I = 1 x ,..., x ) ⎟ ∑
Ln⎜
⎝ 1 m ⎠ k =1
where xk denotes a vector of variables representing the k-dimension of circumstances,

hence, x = ( x1 ,..., xm ) . The dimensions consider include parents education, family per
capita income, gender, number of siblings, family structure (number of siblings, single-
parent household) and area of residence (urban versus rural) 7 . The functions {hk } were
chosen accordingly to the needs of each dimension: quadratic on education, logarithmic
on income, nonparametric (dummies) on age and the other dimensions. In all cases all
functions end up being linear in the parameters, so that, hk (x k ) = xk β k . A complete
specification of this logistic regression is presented in Table 1. From the estimation of
this logistic regression one obtains estimates of the parameters {β k } that will be denoted
{ }
by β̂ k . 8
Table 1: Specification of separable logistic regression function
Dimensions of circumstances Specification
Gender free (dummy)
Parents education quadratic
Per capita income logarithimic
Number of siblings linear
Presence of parents free (dummy)
Area of residence (urban versus

free (dummy)
rural)
7
In the case of education age was also a variable use to predict the probability of completing each grade.
8
The results for these logistic regressions are shown in Tables A.3.1 to A.3.6 at the end of this paper.
As our second step, given this coefficient estimates we obtain for each individual in our
universe his/her predicted probability of access to the opportunity in consideration, i.e.,
for each individual i we compute
⎛ m
⎞
Exp⎜ βô + ∑ x ki βˆk ⎟
pˆ i = ⎝ k =1 ⎠
⎛ m
⎞
1 + Exp⎜ βô + ∑ x ki βˆk ⎟
⎝ k =1 ⎠
Next, as a third and final step we compute
n
p = ∑ wi pˆ i
1
and
1 n
Dˆ = ∑ wi pˆ i − p
2 p i =1
1
where wi = or some sampling weights. 9
n
Since, almost surely, Lim( p ) = P(I = 1) and, under the assumptions that (i) the regression
n →∞
has been correctly specified and (ii) its coefficients consistently estimated, also
⎛ n ⎞
Lim⎜ ∑ wi pˆ i − p ⎟ = E P(I = 1 x ) − P(I = 1)
n→∞
⎝ i =1 ⎠
( )
almost surely. It follows then that Lim Dˆ = D almost surely. In other words D̂ is a
n →∞
consistent estimator of D.
Since D̂ is a ratio of two linear functions of p̂i , it can be shown that the asymptotic
variance of D̂ , σ D2 , is given by σ D2 = ΓΩ β Γ′ , where
1 ⎛⎛ ⎞⎛ ⎞ ⎛ ⎞⎛ ⎞⎞
Γ= ⎜ ⎜ w pˆ ⎟.⎜ ∑ wi pˆ i (1 − pˆ i )xi ⎟ − ⎜ ∑ wi pˆ i ⎟.⎜ ∑ wi pˆ i (1 − pˆ i )xi ⎟ ⎟⎟
2 ⎜ ∑ i i
pˆ ⎝ ⎝ i∈L ⎠ ⎝ i∈U ⎠ ⎝ i∈U ⎠ ⎝ i∈L ⎠⎠
9
The Dissimilarity index for the 19 selected Latin American countries are shown in Table A.3.7 at the end
of this paper.
Ω β is the asymptotic variance matrix for βˆ and L denotes the set of all individuals with
predicted probability of access below average and U the complement. So, L = {i : pˆ i ≤ pˆ }
and U = {i : pˆ i > pˆ }
We conclude considering two practical problems. First, it must be emphasized that the
entire procedure is consistent if we keep or if we drop the non-significant parameters.
Asymptotically the choice is irrelevant. We, however, have decided for keeping the non-
significant coefficients. Our choice was based on a couple of considerations. Firstly, we
have to recognize that dropping a coefficient will always reduce the measured level of
inequality. Since the presence of non-significant coefficients depends as much on their
magnitude as on sample size, dropping them would favor smaller sample size countries.
Secondly, by the same reason that we do not revert to 2 all coefficients that are not
significantly different from 2, we have no reason to give zero a special treatment.
A second practical question is the treatment to be given to coefficients with sign contrary
our expectations. For instance, what should we do when rural areas are estimated to have
greater access than the corresponding urban areas? We opted for keeping all the
coefficients at their estimated values. Again our decision was based on a couple of
reasons. First, because respecting the data could prove informative. Today, throughout
Latin America girls are performing better than boys in school, so the male dummy
receives a consistent negative sign. We are all now quite familiar with these findings, so
we would not consider their sign incorrect. But the first econometricians encountering
this finding a decade ago could consider it unlikely and discard the information.
Secondly, most of our intuition and reasons for discarding unexpected signs may be
based on univariate comparisons. We all expect the access to opportunities to be worse in
rural areas, but are we absolute sure about this contrast after we have extensively control
for parents´ education, family income and composition?
2.3. PROPERTIES OF THE DISSIMILARITY INDEX
Range of the Inequality of Opportunity Index
Consider a situation in which we have m groups. Let pi be the proportion of children in

group i with access to a basic social service and let β i be the proportion of children in
this group. In this case our new dissimilarity index will be given by
1 m
D= ∑ βi pi − p
2 p i =1
where
m
p = ∑ β i pi
i =1
Proposition 1: D ≥ 0
m
In fact, since pi − p ≥ 0 , it follows that ∑β
i =1
i pi − p ≥ 0 . Hence, D ≥ 0 .
Proposition 2: D ≤ (1 − p )
Let us, without any loss of generality, order the groups increasingly according to their
values of pi . Let k be such that pk < p and pk +1 ≥ p . Hence, k is the number of groups
with probability of access below average. In this case, the numerator of the dissimilarity
index can be rewritten as;
m k m
⎛ k m
⎞ ⎛ k m
⎞
∑ βi pi − p = ∑ βi ( p − pi ) +
i =1 i =1
∑ βi ( pi − p ) = p⎜ ∑ βi −
i = k +1 ⎝ i =1
∑ βi ⎟ − ⎜ ∑ βi pi −
i = k +1 ⎠ ⎝ i =1
∑ β p ⎟⎠
i = k +1
i i
hence,
m
⎛ m m
⎞ ⎛ m m
⎞ ⎛ m m
⎞
∑β
i =1
i pi − p = p⎜ ∑ βi − 2 ∑ βi ⎟ − ⎜ ∑ βi pi − 2 ∑ βi pi ⎟ = 2⎜ ∑ βi pi − p ∑ βi ⎟
⎝ i =1 i = k +1 ⎠ ⎝ i =1 i = k +1 ⎠ ⎝ i = k +1 i = k +1 ⎠
therefore,
m
⎛ m m
⎞ ⎛ ⎛ k m
⎞⎞
∑ βi pi − p = 2⎜ ∑ βi pi − p ∑ βi ⎟ = 2⎜⎜ p − ⎜ ∑ βi pi + p ∑ βi ⎟ ⎟⎟
⎝ i = k +1 ⎠ ⎝ i =1 ⎠⎠
i =1 i = k +1 ⎝ i = k +1
and
k m k m
∑ βi pi + p ∑ βi > p∑ βi pi + p ∑ βi pi = p 2
i =1 i = k +1 i =1 i = k +1
Therefore,
∑β i pi − p ≤ 2 p (1 − p )
i =1
implying that
1 m
D= ∑ βi pi − p ≤ 1 − p
2 p i =1
Next let us now consider two extreme situations:
Case A: pi = p for all i =1,…,m.

Case B: pi = 0 for all i =1,…,m-1 and pm = 1
Proposition 3: In Case A, D = 0
In fact, in this case, pi − p = 0 for all i and consequently, D = 0
Proposition 4: In Case B, D = 1 − β m = 1 − p . Therefore, D ↑ 1 as β m ↓ 0

Since in thus case pi = 0 for all i = 1,…m-1, p = β m pm . Consequently, for all i
=1,…, m-1 , pi − p = p = β m pm . Hence,
m −1 m −1
∑β i pi − p = ∑ β i β m pm = (1 − β m )β m pm
i =1 i =1
In addition, since
pm − p = (1 − β m ) pm
m m −1
∑ βi pi − p = ∑ βi pi − p + β m pm − p = (1 − β m )β m pm + (1 − β m )β m pm = 2(1 − β m ) p
i =1 i =1
Hence,
1 m
2(1 − β m ) p = (1 − β m )
1
D= ∑
2 p i =1
β i pi − p =
2p
In sum, D = 1 − β m . As a consequence, D ↑ 1 as β m ↓ 0 .
Sensitivity of Dissimilarity Index to variations in the coverage rate
In principle, any inequality of opportunity index to deserve this denomination must be

insensitive to a balanced increase in the availability of opportunities. In fact, new
opportunities that are allocated using the same criteria as the pre-existing ones should not
have an effect on the level of inequality. That a proper measure of inequality of
opportunity must have this property is clear, the difficulty is how to define the concept of
balanced increase in availability of opportunities.
In measuring income inequality, balanced growth is easy to define. To obtain balanced

growth one has just to distribute the new income in the same way previous income was
distributed. If a person holds δ% of the original total income, he/she must receive δ% of
the additional income. As a consequence, he/she will remain with δ% of the new total
income. In this case, every and each person will receive some extra income 10 .
Some specific opportunities, either one has or does not have, as opposed to income where
there may be large variations among those who have. So, an increase in the availability of
opportunities necessarily means to provide access for those who did not have access
previously, leaving those who already had access unchanged. In fact, when we are
dealing with a specific opportunity, we can not increase the opportunity of those who
already had access to it. In this case the notion of balanced increase must naturally be
more complex.
To construct a balanced increase in this situation, we have to change the unity of analysis
from individuals to circumstance groups. This is not a limitation, since measures of the
inequality of opportunity should be insensitive to the distribution of opportunities within
circumstance groups. So, we restrict our attention to the average access of each group.
Accordingly, a balanced increase would be a situation in which new opportunities are
assigned to circumstances groups in the same way as the pre-existing ones were in the
past.
It seems, however, that we have just changed labels. We have defined balanced increase
based on the assignment of opportunities to circumstances groups in the same way as the
pre-existing opportunities were originally assigned. But what we mean by in the same
way? A natural interpretation is to assume that for a balanced increase to occur one need
the distribution of the new opportunities to be distributed among circumstance groups in
the same way as the pre-existing distributions are. In this case, the distribution of
opportunities across circumstance groups would remain unchanged.
In this case, due to the increase in the availability of opportunities both the overall
coverage, p , and the group specific access probabilities, {p j }, would increase. Given the
way the new opportunities are distributed, all specific probabilities increase
proportionally to the overall coverage rate. Hence, if p *j denotes the new group i specific
pj p *j
access probability, then p = (1 + λ ) p j for all j and also p = (1 + λ ) p . So,
*
j = * since
p p
neither the proportion of the population in each group, β j , neither the proportion of
opportunities allocated to each group, γ j , change as a result of this balanced increase in
pj γj
opportunities (remember that = ).
p βj
As long as we measure the inequality of opportunity using D, defined as before
10
An exception will be those that originally had no income. To keep the balance, they will receive nothing.
1 m 1 m
D= ∑ j j
2 p j =1
β p − p = ∑γ j −βj
2 j =1
and the inequality of opportunity will be insensitive to this type of balanced increase in
opportunities.
Traditionally, however, the dissimilarity index has been defined slightly differently. In
most cases is defined in terms of the lack of access, i.e.,
( )
m m
D1 =
1
∑ j β 1 − p − (1 − p ) =
1
∑β j pj − p
2(1 − p ) j =1 2(1 − p ) j =1
j
In this case, the interpretation changes slightly. As mentioned before, D is the proportion
of all opportunities that need to be rearranged to ensure equal access for all groups. In
other words, is the amount of opportunities that need to be rearranged as a proportion of
the number of individuals who already have access. The more traditional measure, D1 ,
keeps the same numerator, the amount of opportunities that need to be rearranged for
equal opportunity to prevail, but changes the denominator to become the number of
individuals without access. So, it becomes the amount of opportunities that need to be
rearranged as a proportion of individuals without access to this opportunity.
In short, D centers the attention on access and D1 on the lack of access. In fact,
1 m
D = ∑ γ j − β j . Hence, it measures the distance between the distribution of the overall
2 j =1
population, {β j }, and the distribution of individuals with access to the opportunity of
interest, {γ j }, across groups. On the other hand, D1 =

1 m
∑ η j − β j , where η j denotes
2 j =1
the proportion of individuals without access to the opportunity in consideration who
belong to circumstance group j. So it measures the distance between the distribution of
the overall population, {β j }, and the distribution of individuals without access across
opportunity groups, {η j }.
In fact, a more symmetric measure, D2 , would be the distance between the distribution of
the individuals with access, {γ j }, and the distribution of individuals without access across
opportunity groups, {η j }. In this case,
1 m
D2 = ∑ γ j −η j
2 j =1
It can be shown that in this case D2 can also be written as

m
1
D2 = ∑β j pj − p
2 p (1 − p ) j =1
Above we demonstrate that D is insensitive to a balanced increase in the availability of

opportunities, at least as we have defined it above. D1 and D2 , however, do not share this
property 11 . Both measures would increase, indicating that a balanced increase in the
availability of opportunities (one that would keep the distribution of opportunities across
groups unchanged) would increase the inequality of opportunity as measured by these
two indices. This undesirable property is one of the main reasons why we opted to
measure the inequality of opportunity using D.
Although our concept of balanced increase in opportunity is quite natural, it has at least
one important practical limitation worth mentioning. In fact, it does not respect the
restriction that each individual can have at most one of the specific opportunity we are
considering (in fact, either one has or does not have access to school), so that we always
must have p j ≤ 1 . If the coverage is already very high in certain groups, it might not exist
a way of allocating the new opportunities as the old ones were allocated. Suppose we
have two groups and all the original opportunities were allocated just to one of the groups
ensuring full coverage for this group. In this case, new opportunities cannot be allocated
to the advantage group because their coverage rate is already 100%. Moreover, the new
opportunities cannot be allocated either to the disadvantage group without disturbing the
distribution of opportunities across groups. So, using this concept, no balanced increase
in opportunity is possible in this situation. Given this drawback, one might be compel to
review the concept. However, it does not seem to exist a simple alternative without its
own limitations. Let us consider a couple of them for further analysis.
One possibility would be to consider as balanced an increase in opportunities in which

the new opportunities are distributed at random. This would be equivalent to distribute
opportunities across groups according to their population size. In this case the access
probability of all circumstance groups will increase by the same amount. So, while
according to our previous concept of balanced increase in opportunities
( )
Ln pi* = Ln( pi ) + Ln(1 + λ ) , under this alternative concept pi* = pi + λ .
In this case, the absolute number of opportunities that need to be reallocated,

N m
∑ β j p j − p , would not be influenced by the any balanced increase. As a
2 j =1
consequence, following a balanced increase in opportunity of this type D would decline,
D1 will increase and D2 will decrease if p < 1 / 2 and increase if p > 1 / 2 . So, when the
11
The proof is simple. First, notice that D2 = D + D1 . Since D is insensitive to a balanced increase in
1
opportunities, the sensitivity of D1 and D2 must be the same. Next notice that D2 = D . Hence,
1− p
any increase in p not affecting D would increase D2 and so D1 .
majority of the population already has access to the opportunity in consideration, out of
the three measures only D will decline as a consequence of such balance increase in
opportunity.
It should be noticed that this concept share the same limitation as the previous one. The
restriction that no individual can have more than one opportunity implies that no
balanced increase may be possible in some situations.
A second and more elaborate alternative would be to consider balanced proportional

increase in the odds ratio of all circumstance groups. In this case, a balanced increase in
opportunities occur when
p *j pj
= (1 + λ )
1− p *
j 1− p j
which is equivalent to
1+ λ
p *j = pj
1 + λp j
So, in this case the proportional increase in specific coverage will vary accordingly the
baseline situation, with the proportional increase being greater for the more disadvantage
groups. For groups with original coverage closer to zero the coverage will increase by
λ %, whereas for groups with original coverage closer to universal coverage will not be
affected.
It can be shown that also in this case the inequality as measure by D would always
decline as a consequence of a balanced increase in opportunity of this type. The other
alternative measures, D1 and D2 , may decline or even increase. In any case, out of the
three measures D would be the one experiencing the greatest decline.
In sum, we consider three alternative measures for the inequality of opportunity, D, D1

and D2 , and their response to three notions of a balanced increase in the availability of
opportunities. Table 1 summarizes the main results. Ideally, we would like these
measures to be insensitive to a balanced increase in opportunities. The fact that D1 would
always increase and D2 would increase if p > 1 / 2 (the most common case) if one
distributes the new opportunities at random, lead us to strongly prefer the measure D. D
would always decline if the newly available opportunities were distributed at random,
The preference for the measure D also comes from the fact that it is insensitive to an
increase in opportunities whenever the new opportunities are distributed across groups
accordingly the distribution of the pre-exiting opportunities. In this case, both D1 and D2
would increase (see Table 2). If the concept of a balanced increase in opportunities is
taken to represent a proportional shift in all group specific odds ratio, D would always
decline as a result of such balanced increase in opportunity. Hence, out of the three
concepts of balanced increase in opportunity we consider in these notes, D is insensitive
to one and inverse related to the other two. One may conclude therefore that D is an
inequality measure exhibiting some pro-growth bias.
Table 2: Sensitivity of Alternative measures of inequality of opportunity
to a balanced increase in available opportunities
Inequality Measure
Concept of balanced D D1 D2
increase in
opportunity
p j * = (1 + λ ) p j remains unchanged increases increases
pj* = pj + λ decreases increases increases if p>1/2
1+ λ decreases may increase may increase
pj* = pj
1 + λp j
Decomposability of the inequality of opportunity index
The degree of inequality of opportunity in this report is measured by
1 m
D= ∑ β i pi − p
2 p i =1
If we denote by ν i the group i absolute gap in specific coverage relative to the overall
mean, then ν i = pi − p and,
1 m
D= ∑ β iν i
2 p i =1
This expression for D can be extremely useful to disentangle the immediate reasons for
variations over time or across countries in the inequality of opportunity. It reveals that D
has only three immediate determinants: ( p, β ,ν ) , where β = (β 1 ,.., β m ) and similarly
ν = (ν 1 ,..,ν m ) . As this expression indicates, D can change if and only if at least one of
these tree determinants changes. For this reason we refer to them as the immediate
determinants of the inequality of opportunity.
Usually we associate inequality of opportunity with coverage gaps between circumstance

groups. The component of D capturing these gaps is ν . The greater the gaps, the greater
will tend to be the elements of ν and, as consequence, the greater will be D, the
inequality of opportunity. The effect of changes in ν on changes on D will be called the
“gap effect”.
The factor p appears in the expression to translate an absolute index of inequality into a
relative index. It is essential to isolate changes in scale. Without the division by p , a
balanced increase in the available opportunities (i.e., and increase in which the new
opportunities were distributed precisely as previously available) would increase the level
of inequality of opportunity. Moreover, without the division by p , an increase in
available opportunities proportionally assigned would not reduce inequality. It would
actually leave inequality unchanged. The effect of changes in p on changes on D will be
called the “scale effect”.
Finally, the component β captures the distribution of individuals across circumstance

groups. In the making of D, it is the component that allows a vector of coverage gaps ν
to be translated in a scalar measure of inequality. In this construction, β serves as the
weights of a weighted average. Precisely for this reason, variations in D can be caused to
some extent by changes in these weights. It may be useful to isolate them because they
refer to changes in the relative size of each circumstance group, not to changes in gaps
between groups nor to changes in available opportunities. Hence, the effect of changes in
β on changes on D will be called the “composition effect”.
Consider two situations: A and B. It could be a country at two points in time or two
countries at the same point in time. Let Δ p be the changes on D if the only difference
between situations A and B were a difference in their p s. The β s and ν s are exactly de
same. So Δ β are the changes on D if the only difference between situations A and B
were a difference in their β s, and Δν the changes on D if the only difference were a
difference in their ν s.
Next, we consider a procedure to decompose variations in the inequality of opportunity to

concomitant variations in each of its three immediate determinants. Let Δ p , Δ β and Δν
be defined as
m m
1 1
Δp =
2pB
∑ β jBν Bj −
j =1 2pA
∑β
j =1
B
jν Bj
m m
1 1
Δβ =
2pA
∑ β jBν Bj −
j =1 2pA
∑β
j =1
A
jν Bj
m m
1 1
Δν =
2pA
∑ β jAν Bj −
j =1 2pA
∑β
j =1
A
jν jA
Notice that these differentials add to the total change on the degree of inequality of
opportunity, Δ = D A − D B . In fact,
m m
1 1
Δ p + Δ β + Δν =
2pB
∑ β jBν Bj −
j =1 2pA
∑β
j =1
A
j ν jA = D A − D B = Δ
so, Δ p + Δ β + Δν = Δ . Moreover, notice that (a) if p A = p B then Δ p = 0 , (b) if α A = α B
then Δ α = 0 and finally, (c) if ν A = ν B then Δν = 0 .
Properties (a)-(c) indicates that each differential is in fact capturing only the impact of the
corresponding immediate determinant. In addition, Δ p + Δ β + Δν = Δ . These differentials
add to the total change on the degree of inequality of opportunity, Δ = D A − D B . Hence
the impact of the three immediate determinants adds up to the total change, leading,
therefore, to a proper decomposition of the overall change
3. FROM COVERAGE RATE AND INEQUALITY OF OPPORTUNITIES TO AN OPPORTUNITY

INDEX
Everyone recognizes that for equality of opportunity to prevail and poverty to be

eradicated, access to basic opportunities should be universal. Let I be an indicator of
access to a basic opportunity (access to primary education of quality, for instance), so that
I=1 for those with access and I=0 for those without access to this opportunity. Since
equality of opportunity would require access for all, a natural indicator to monitor is the
coverage rate (i.e., the percentage of subjects with access to this opportunity),
p = P[I = 1] = E [I ] .
Figure 1 presents estimates of the current access to primary education of quality in a

variety of Latin American and Caribbean countries. We measure the access to this
opportunity by the completion of sixth grade on time. According to Figure 3, the access
to primary education of quality varies from less than 50% in Guatemala and Nicaragua to
near 90% in Jamaica and Chile.
Figure 3: Average probability of completing the sixth grade on time – National,
circa 2005
Jamaica
Chile
Mexico
Ecuador
Panama
Uruguay
Venezuela
Argentina
Costa Rica
Bolivia
Paraguay
Dominican Republic
Colombia
Honduras
Peru
Brazil
El Salvador
Nicaragua
Guatemala
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
average probability
Despite the central importance of this indicator (coverage rate), its total lack of sensitivity
to how available opportunities are distributed may be considered an important drawback.
It could be argued that the way the available opportunities are distributed should be
reflected in any meaningful index of opportunity. Countries with the same coverage
would have different opportunities depending on whether the available opportunities are
distributed according race, other family origin and socio-economic conditions
(circumstances) instead of randomly.
Probably most of the interest and motivation for incidence analysis comes from concerns
of this nature. Figure 4 presents how our measure of access to primary education of
quality varies for children in poor and rich families. This figure reveals that in addition of
lower coverage, in Nicaragua and Guatemala the gap in access between the rich and poor
is much greater than in Chile and Jamaica.
Figure 4: Probability of completing the sixth grade on time for the poorest and
richest quintile
Uruguay
Mexico
Panama poorest quintile

Ecuador
Jamaica
richest quintile
Chile
Bolivia
Costa Rica
Colombia
Venezuela
Dominican Republic
Paraguay
Argentina
Peru
Brazil
Honduras
Nicaragua
El Salvador
Guatemala
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
If z is a scalar indicator of one of the dimensions of personal origin, gender or family

socio-economic background (per capita income in Figure 2), traditional incidence
analysis seek to estimate p ( z ) = P[I = 1 z ] = E [D z ] and evaluate how sensitive to z
is p(z ) . Although no general measure of the dependence of p(z ) on z is proposed, the ratio
between the specific coverage rate for the top and bottom quintiles are widely used.
Figure 5 presents estimates for this ratio.
Figure 5: Ratio of the probability of completing the sixth grade on time for the
richest and poorest quintile
Guatemala
Nicaragua
Brazil
El Salvador
Peru
Colombia
Bolivia
Paraguay
Dominican Republic
Costa Rica
Honduras
Uruguay
Panama
Ecuador
Mexico
Venezuela
Chile
Jamaica
Argentina
1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 3.8 4.0
To the extent these gaps in access due to differences in race, gender, family origin and
socio-economic condition are important, they should also be part of a general measure of
the availability of opportunities. What we need then is a distributive sensitive coverage
indicator. Before we attempt to construct such index, let us solve first certain limitations
of the most basic incidence analysis.
A sizeable fraction of incidence analysis studies investigate the relation between access
and each circumstance separately, without considering the impact of others circumstances
or even holding them constant. These studies seek correlations not causal relations.
Figures 4 and 5 certainly demonstrate that access to primary education of quality remains
quite correlated with income in almost all countries of the region. The statistical analysis
behind this table, however, cannot identify whether the gaps are really caused by
differences in income or to other associated variables, like parents education.
One alternative is to estimate multivariate models relating the probability of access to all
measurable circumstances. One hopes that proceeding this way would lead closer to
identifying the actual causal impact of each circumstance on the probability of access.
Needless to say that due to either (a) our limited ability of measuring all circumstances,
(b) the existence of other factors, beyond circumstances, affecting the probability of
access or (c) the measurement err associated to circumstances the estimated relation
between the probability of access and observed circumstances may end up being quite
apart from their true causal relation. Nevertheless, expecting to get closer to a causal
relationship, we proceed to a multivariate estimation. More specifically, we estimate
p ( x) = P[I = 1 x ] = E [I x ] where now x denotes a comprehensive vector of circumstances.
Based on this estimated relation a large variety of differentials could be computed. Figure
6 and 7 illustrate this possibility. In this figure we contrast the access to primary
education of quality for children in families with favorable and unfavorable
circumstances. In this comparison a variety of circumstances are varying at the same
time. It may be worthwhile to isolate the impact of each dimension of the circumstances,
holding all others constant. This can be easily done. Figures 8a to 8c illustrate it.
Figure 6: Differential probability of completing sixth grade on time, circa 2005
El Salvador
Mexico
Bolivia
Uruguay
Panama
Colombia
Jamaica
Paraguay
Venezuela
Costa Rica
Dominican Republic
Chile
Nicaragua
Peru
Guatemala
Ecuador
Argentina
Brazil
Honduras
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
average probability
Figure 7: Ratio of the probability of completing the sixth grade on time for the
advantage and disadvantage groups
Brazil
Nicaragua
Guatemala
Peru
Colombia
Dominican Republic
Costa Rica
Uruguay
Ecuador
Paraguay
Bolivia
Venezuela
Panama
El Salvador
Honduras
Chile
Mexico
Argentina
Jamaica
1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0 11.0 12.0 13.0 14.0 15.0
There is not, however, a unique way for doing it, since the relationships are non-linear
and the other circumstances could be held constant at different levels. In Figures 6a-c we
opt for taking the average of all these possibilities. More precisely, we estimate
q( x k ) = ∫ p( x)dF ( x − k ) , where k denotes the dimension of immediate interest and –k all the
others 12 . If the dimension of interest were gender, then all we have to do would be to
estimate for every child two access probabilities. In one case assuming that he or she
were a male; in the other assuming that he or she were a female. The average of the first
probability would be q for males and the average of the second probability would be q for
females.
12
[ ]
Note that p( x k ) = E D x k , the univariate relation between the probability of access and the dimension k
of the circumstance, is given by p( x k ) = ∫ p( x)dF ( x −k x k ) . Therefore, in general q( x k ) ≠ p( x k ) , unless
F ( x − k x k ) = F ( x − k ) , i.e., circumstance x k is stochastically independent of all others.

Figure 8.a: Probability of completing the sixth grade on time by parents education
Mexico
Ecuador
Illiterate
Panama
College education
Peru
Dominican Republic
Chile
Costa Rica
Bolivia
Venezuela
Paraguay
Colombia
Jamaica
Nicaragua
El Salvador
Brazil
Guatemala
Honduras
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Figure 8.b: Probability of completing the sixth grade on time by per capita family
income
Ecuador
Mexico
1 US$ per day
Jamaica
Peru
50 US$ per day
Panama
Chile
Venezuela
Bolivia
Colombia
Costa Rica
Dominican Republic
Paraguay
Brazil
Honduras
El Salvador
Nicaragua
Guatemala
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Figure 8.c: Probability of completing the sixth grade on time accordingly the
number of siblings
Jamaica
Ecuador
5 children 1 child
Mexico
Chile
Venezuela
Panama
Peru
Costa Rica
Paraguay
Bolivia
Colombia
Dominican Republic
Honduras
Nicaragua
Brazil
El Salvador
Guatemala
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
The movement to a multidimensional relation has hopefully led us closer to causal

relationship, but it has a major practical drawback. We now have a very large number of
probability gaps related to differences in circumstances. How should them be
summarized? And once having found a summary statistics, how should we use this
statistics to construct a distributive sensitive coverage measure?
Let us consider the first problem first. Surely, there are several alternatives to summarize
the strength of the relationship between circumstances and the probability of access. Few
procedures, however, share a simple and clear interpretation. Among these few, we
consider the minimum proportion of all available opportunities that one must reallocated
to ensure equal access to all circumstance groups, i.e., a situation in which p(x) would be
equal to p .
To construct this index consider a population with N subjects divided among m disjoint
circumstance groups. Within each of these groups all subjects have the same value of the
circumstances, i.e, within group j all subjects have x = a j . Moreover, if a subject has x = a j
then he/she should belong to circumstance group j. Let M j denote the number of subjects
in group j with access to the opportunity in study (i.e., for whom I = 1 ) and let N j be the
M Nj
total number of subjects in this group. Then, p(a j ) ≡ p j = j
and let β j =
denote
Nj N
the proportion of subjects in circumstance group j. Finally, let M j denote the number of
subjects that would have access to the opportunity in study if the probability of access
were identical for all groups. From its very definition, it follows that M j = p.N j
Relative to a situation in which the access is identical for all circumstance groups, the
number of opportunities in excess or lacking in group j is given by M j − M j . Therefore,
the minimum number of opportunities that need to be rearranged to reach equal access
m
∑M
1
across groups is given by j −M j . Since the total number of opportunities
2 j =1
m
available is M = ∑ M j = pN , it follows that as a percentage of the total number of
j =1
available opportunities the minimum amount that need to rearranged to ensure equal
m
∑M
1
access across circumstance groups, D, is given by D = j −M j
2 pN j =1
Notice that this is exactly the same Dissimilatity Index developed in section 2.1 above.
This index can be expressed in many different forms. In fact, notice:
1 m
1 m Nj Mj −Mj 1 m
D=
2 pN
∑ Mj −Mj =
j =1
∑
2 p j =1 N Nj
= ∑β j pj − p
2 p j =1
As this expression indicates, the index is proportion to the average absolute distance of
the group specific access probabilities and the overall probability of access. In this sense,
it is a measure of the inequality of opportunity.
There is yet another alternative expression for the same index from which a new
interpretation is worthwhile to register. Let γ j denote the proportion of the overall
Mj
available opportunities that are being currently allocated to group j. Hence, γ j = and
M
we can rewrite D as
m m
1 1 1 m
D=
2 pN
∑ Mj −Mj =
j =1 2M
∑ M j − pN j =
j =1
∑γ j −βj
2 j =1
So, D has the alternative interpretation of being the distance between the distribution of
the opportunities, {γ j }, and the distribution of the population , {β j }, across circumstance
groups. Since ideally each member of the population would have an opportunity, the
distribution of the population across groups is identical to the distribution that
opportunities would have if equality of opportunity prevails. Hence, the proposed
measured of inequality of opportunity is identical to the distance between the actual
distribution of opportunities across circumstance groups and the distribution that would
occur if equality of opportunity prevails. In this sense is a clear measure of the inequality
of opportunity.
Figure 9 presents estimates of the inequality of access to primary education of quality for
a variety of Latin American and Caribbean countries, using this inequality measure.
Consistently with our previous findings, the inequality of opportunity is particularly large
in Nicaragua and Guatemala and particularly small in Jamaica and Chile. Quantitatively,
this figure reveals that almost 30% of the educational opportunities in Guatemala and
Nicaragua would need to be rearranged for equality of opportunity to prevail. In Chile,
Jamaica and Argentina only around 3% of the opportunities would need to be rearranged
to reach equality of opportunity.
Figure 9: Inequality of educational opportunities measured by the dissimilarity

index – National, circa 2005
Nicaragua
Guatemala
El Salvador
Brazil
Peru
Colombia
Dominican Republic
Bolivia
Paraguay
Costa Rica
Honduras
Uruguay
Venezuela
Panama
Ecuador
Mexico
Argentina
Chile
Jamaica
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20 0.22 0.24 0.26 0.28 0.30
Dissimilarity Index
The estimates presented in this figure are just a scalar synthetic measure of the gaps in
access to primary education of quality presented in Figures 2 to 6. It is certainly just one
possible measure. However, as we try to demonstrate it is a measure with a simple and
clear interpretation that can be computed in a very transparent way.
Having proposed a solution to our first problem let us consider the second: How to
construct a distributive sensitive opportunity index related to this measure of inequality of
opportunity.
M
Let us begin analyzing further the meaning of p . Since p = and N is the number of
N
opportunities needed to ensure access for all, we can reinterpret p as the percentage of
the total number of opportunities required for universal access that are actually available.
This form of interpreting p clarifies both its strengths and its weaknesses. It
demonstrates that p is certainly a measure of the stock of available opportunities, but it
is completed insensitive to how these opportunities are allocated.
These observations give us clear directions for improving upon p . A simple and intuitive
idea would be to modify the numerator so that are considered as valid only those
allocated accordingly to the principle of equal opportunity for all. Hence, if we let O
denote the available opportunities allocated respecting the principle of equal opportunity,
O
then the desired index, r, can be expressed as r = .
N
It remains, however, to be more specific about O. One alternative we already have. Since
our measure of inequality of opportunity, D, is the proportion of opportunities that must
be reallocated for equality of opportunity to prevail, then 1-D is the proportion properly
allocated and M .(1 − D ) the total number allocated accordingly the principle of equal
opportunity for all. Hence, one possibility is to let O = M (1 − D ) . In this case, the overall
measure of opportunity, r, will be given by
O M
r= = (1 − D ) = p(1 − D )
N N
Using this measure, Figure 10 presents estimates of the access to primary education of
quality for a variety of Latin American and Caribbean countries. This figure reveals that
in Chile and Jamaica around 85% of the educational opportunities required for universal
access are available and have been allocated consistently with the principle of equal
opportunity. On the other hand, in Guatemala and Nicaragua less than 30% of the
educational opportunities required for universal access are available and have been
allocated consistently with the principle of equal opportunity. 13
13
The Opportunity index for the 19 selected Latin American countries are shown in Table A.3.8 at the end
of this paper.
Figure 10: Index of Opportunity
Jamaica
Chile
Mexico
Ecuador
Panama
Uruguay
Venezuela
Argentina
Costa Rica
Bolivia
Paraguay
Dominican Republic
Colombia
Honduras
Peru
Brazil
El Salvador
Nicaragua
Guatemala
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Although a clear definition must be an essential attribute of any opportunity index, its
usefulness will also necessarily depend on its properties. Next, we highlight a few
important and useful properties of the proposed index.
3.2. PROPERTIES OF THE OPPORTUNITY INDEX
Sensitivity to redistribution
We begin rewriting the index. In order to simplify the exposition and without any loss of
generality we assume that all circumstance groups are ordered increasingly by their
probability of access. Hence,
1⎛ l ⎞
⎜ ∑ β j (p − p j )+ ∑ β (p − p )⎟⎟
1 m m
r = p (1 − D ) = p − ∑ j j β p − p = p −
2 ⎜⎝ j =1
j j
2 j =1 j = l +1 ⎠
where l denote the number of groups with access probability below average. Let λ denote
the proportion of subjects in circumstance groups with access probability below average,
l
then λ = ∑ β j . Notice that if p , in addition to be the average of the group specific
j =1
probability of access, were also the median, then λ = 0.5 . In any case,
l m
r = (2 − λ )∑ β j p j + (1 − λ ) ∑ β j p j
j =1 j =l +1
or alternatively
1⎛ l m ⎞
r= ⎜ (2 − λ )∑ M j + (1 − λ ) ∑ M j ⎟
N ⎜⎝ ⎟
j =1 j =l +1 ⎠
This expression for the Opportunity Index r reveals several of its properties. First, it
demonstrates that despite its distributive sensitivity the index is Pareto consistent. In fact,
an increase in the number of opportunities available to any group would always increase
the index. Its distributive sensitiveness, however, implies that the impact would be
greater if the increase benefits the groups with below average access to opportunities.
In fact, if an extra opportunity is given to any group with access probability above
1− λ
average, the index would increase by 100 percentage points. Notice that if λ = 0.5
N
and N = 100 , then a one percentage point increase in the availability of opportunities to
above average groups would increase the Opportunity Index in one-half of a percentage
point.
However if the additional opportunities were given to any group with access probability
2−λ
below average, the index would increase by 100 percentage points. Notice, again,
N
that if λ = 0.5 and N = 100 , then an one percentage point increase in the availability of
opportunity to above average groups would increase in the Opportunity Index in 1.5
percentage point.
Finally, it follows from these observations that a balanced increase in available

opportunities that give λ to below average groups and 1 − λ for the above average groups
1
would increase the Opportunity Index precisely by 100 percentage points. Hence if
N
λ = 0.5 and N = 100 , then a balanced increase of one-half percentage point in the
availability of opportunity to above average groups couple with the same increase for the
bellow average group would increase the Opportunity Index by precisely one percentage
point.
Decomposability of changes in the opportunity index
Consider two situations: A and B. It could be a country at two points in time or two
countries at the same point in time. Any change in the opportunity index, O, could be
decomposed in a scale, Δ p , and a distributional, Δ D , effect via
( ) ( )
Δ = O B − O A = p B 1− DB − p A 1− D A = Δ p + Δ D
where the scale effect, Δ p , and the distributional effect, Δ D , are defined as follows
( ) (
Δ p = p B 1− D A − p A 1− D A )
and
( )
ΔD = p B 1− DB − p B 1− D A ( )
Moreover, notice that (a) if p A = p B then Δ p = 0 and (b) if D A = D B then Δ D = 0 .
These properties indicate that each differential is in fact capturing only the impact of the
corresponding factor. Since, by construction, Δ p + Δ D = Δ , the impact of the two
determinants of the opportunity index adds up to the total change, leading, therefore, to a
proper decomposition. 14
4. SUMMARY
This chapter has provided a broad discussion of the concept of inequality of opportunities
and the issues involved with measuring this concept. Measuring inequality of
opportunities requires special attention to understanding how circumstances, efforts,
innate talents, luck and other factors affect outcomes and how these elements can be
appropriately gauged and controlled for given the usually available data. Several
approaches to measuring inequality of opportunities are explained, and its relative
advantages are discussed in terms of both its relevance and applicability. After these
considerations, an index for measuring inequality of opportunities (the Dissimilarity
Index, D) is proposed. In addition, an Opportunity Index is proposed that takes into
consideration expansion in the coverage of basic opportunities as well as its distribution.
Both indexes have a series of properties, fully elaborated in the text, that make them
useful for analytical and policymaking purposes. Finally, the indexes are illustrated for a
sample of 19 Latin American countries with available data for recent years with respect
to access to selected basic opportunities (i.e., primary education, water and sanitation,
and electricity).
14
Other decomposition formulas are possible.
REFERENCES
Arneson R. J. 1990, “Liberalism, distributive subjectivism, and equal opportunity for welfare”,
Philosophy and Public Affairs 19: 158-194.
Betts, Julian and John Roemer, "Equalizing Opportunity through Educational Finance Reform,"
Mimeo, Public Policy Institute of California, San Francisco, CA, 1999.
Boudon R. (1974): Education, Opportunity and Social Inequality, New York: John Wiley Sons
Bourguignon, François; Francisco H. G. Ferreira, Marta Menéndez (2007) Inequality of
Opportunity in Brazil Review of Income and Wealth 53 (4) , 585–618
Brunello, Giorgio and Daniele Checchi, 2007. "Does school tracking affect equality of
opportunity? New international evidence," Economic Policy, CEPR, CES, MSH, vol. 22,
pages 781-861, October.
Checchi, Daniele & Vitorocco Peragine, 2005. "Regional Disparities and Inequality of
Opportunity: The Case of Italy," IZA Discussion Papers 1874, Institute for the Study of
Labor (IZA).
Cohen G. A. 1989, “On the currency of egalitarian justice”, Ethics 99: 906-944.
Dworkin R. 1981a, “What is equality? Part 1: Equality of welfare”, Philosophy & Public Affairs
10: 185-246.
Dworkin R. 1981b, “What is equality? Part 2: Equality of resources”, Philosophy & Public
Affairs 10: 283-345.
Fleurbaey, Marc (2005) “Health, Wealth, and Fairness” Journal of Public Economic Theory 7 (2)
, 253–284 doi:10.1111/j.1467-9779.2005.00203.x
Goux, Dominique & Maurin, Eric, 2005. "The effect of overcrowded housing on children's
performance at school," Journal of Public Economics, Elsevier, vol. 89(5-6), pages 797-
819, June
Hild M., A. Voorhoeve 2004, “Equality of opportunity and opportunity dominance”, Economics
and Philosophy 20: 117-146..
Lefranc, Arnaud & Nicolas Pistolesi & Alain Trannoy, 2006. "Inequality of opportunities vs.
inequality of outcomes: Are Western societies all alike?," Working Papers 54, ECINEQ,
Society for the Study of Economic Inequality.
Lefranc, Arnaud & Nicolas Pistolesi & Alain Trannoy, 2006. "Equality of Opportunity:
Definitions and testable conditions, with an application to income in France," THEMA
Working Papers 2006-13, THEMA (THéorie Economique, Modélisation et
Applications), Université de Cergy-Pontoise
Peragine V. 1999, “The distribution and redistribution of opportunity”, Journal of Economic
Surveys 13: 37-69.
Rustichini, A. & Ichino, A. & Checchi, D., 1998. "More Equal but Less Mobile? Education
Financing Intergenerational Mobilityin Italy and in the US," Economics Working Papers
eco98/8, European University Institute.
Schuetz Gabriela & Heinrich Ursprung & Ludger Woessmann, 2005. "Education Policy and
Equality of Opportunity," CESifo Working Paper Series CESifo Working Paper No. ,
CESifo GmbH.
Waltenberg, Fabio D. & Vandenberghe, Vincent, 2007. "What does it take to achieve equality of
opportunity in education?: An empirical investigation based on Brazilian data,"
Economics of Education Review, Elsevier, vol. 26(6), pages 709-723, December
ANNEX 3.1: REFERENCE TABLES
Table A.3.1:Logistic regression estimates

I ne q ua l ity o f op po rtu ni ty: P ro ba b il ity o f c om p le ti ng th e s i x th g ra d e - c i rc a 1 9 9 5
C o s ta Do m i n i c a n
B o l iv i a B ra zi l Ch il e C olom bia E c ua d o r E l S al v a d or G u a te m a l a H on d ur a s J a m ai c a M ex ic o N i c a r a gu a P ana m a P ara g
R ic a R e pu b l ic
In te rc e p t - 3 .1 2 - 8.1 3 -4 .6 8 - 3.5 3 - 3.8 7 - 2 .8 8 -3 . 0 5 - 4.14 -5 .8 7 - 4 .5 3 - 2.24 -2 .2 7 -4 . 0 3 - 2 .2 6 - 5 .9
A ge
8 ye a r s . . . . . . . . . . . . . . .
9 ye a r s . . . . . . . . . . . . . . .
1 0 y e ar s . . . . . . . . . . . . . . .
1 1 y e ar s . . . . . . . . . . . . . . .
1 2 y e ar s . . . . . . . . . . . . . . .
1 3 y e ar s 1 .2 3 2 .3 0 2 .1 5 1. 22 2 .0 2 0 .8 5 1 .4 2 1. 52 1.5 1 1 .6 7 2 .5 1 1. 08 0 .9 5 1 .4 5 1 .7 8
1 4 y e ar s 2 .1 6 3 .1 9 3 .0 9 1. 79 2 .7 0 1 .6 8 1 .8 6 2. 35 2.3 2 2 .6 7 3 .9 2 1. 52 1 .5 9 1 .9 9 2 .50
1 5 y e ar s 2 .4 9 3 .5 8 3 .6 9 2. 28 3 .3 4 2 .0 7 2 .3 1 2. 76 2.8 7 2 .9 7 2 .8 8 2. 14 2 .1 2 2 .5 2 3 .40
1 6 y e ar s 2 .8 2 3 .9 2 4 .0 8 2. 53 3 .1 8 2 .2 3 2 .4 2 3. 05 3.1 4 3 .2 7 1 .9 5 1. 93 2 .3 6 2 .9 4 3 .69
G en d er ( m a le ) 0 .2 1 - 0.5 5 -0 . 3 6 -0 .4 0 - 0.2 8 - 0 .5 0 -0 . 1 1 -0 .36 0.2 2 - 0 .2 5 0 .7 1 -0 .2 2 - 0. 43 - 0 .2 4 - 0 .1
Y e a rs o f s c h o o li ng o f th e
0 .1 6 0 .2 8 0 .1 7 0. 26 0 .2 6 0 .1 4 0 .3 2 0. 11 0.3 1 0 .2 0 0 .1 1 0. 23 0 .1 2 0 .2 2 0 .2 7
fa m i ly h e a d
Y e a rs o f s c h o o li ng o f th e
0 .0 0 - 0.0 1 0 .0 0 -0 .0 1 - 0.0 1 0 .0 0 -0 . 0 1 0. 00 -0 . 0 1 0 .0 0 0 .0 0 -0 .0 1 0 .0 0 - 0 .0 1 - 0 .0
fa m i ly h e a d to th e s q u a r e
L o g a r it h m o f t h e p e r c a p i ta
0 .2 7 0 .6 4 0 .2 9 0. 15 0 .1 5 0 .2 6 0 .1 3 0. 28 0.3 7 0 .2 7 0 .0 7 0. 32 0 .2 8 0 .3 1 0 .2 9
inc om e
P r e s en ce of fa m il y h e ad a n d
- 0 .2 6 0 .3 5 0 .1 4 0. 13 0 .1 6 0 .0 1 0 .4 1 -0 .09 -0 . 1 5 0 .1 2 0 .2 8 0. 29 - 0. 16 0 .2 0 0 .0 8
s pou s e
N u m b e r o f ch il d r e n ag e 0 t o
- 0 .1 1 - 0.1 5 -0 . 1 5 -0 .2 3 - 0.1 4 - 0 .2 3 -0 . 1 5 -0 .11 -0 . 1 0 - 0 .0 8 0 .1 6 -0 .1 3 - 0. 13 - 0 .1 4 - 0 .1
1 6 in th e re s id e n c e
A r ea ( ur ba n) 0 .6 4 0 .4 8 0 .0 9 0. 83 0 .1 0 0 .2 8 0 .3 7 0. 76 0.9 3 0 .4 1 - 0.54 0. 51 1 .2 6 0 .3 9 - 0 .0
Table A.3.2: Logistic regression estimates
Inequality of opportunity: Probability of completing the sixth grade - circa 2005

Domi nica n
Argen ti na Boli via Bra zil Ch ile Co lom bia Co sta Rica Ecuad or El Salva dor G ua te mal a Ho nd uras Jama ica Mexic o Nica ragu a Pa nama Parag ua y Pe ru U rugu ay
R epu bli c
Inte rcept 0.19 -2.75 -5.71 -2.54 -4.75 -4.71 -3.64 -0 .5 8 -3 .7 6 -4 .7 6 -3 .7 6 -0 .2 3 - 2.12 -3.80 -1.25 -4.21 -4.10 -4.60
Ag e
8 ye ars . . . . . . . . . . . . . . . . . .
9 ye ars . . . . . . . . . . . . . . . . . .
1 0 years . . . . . . . . . . . . . . . . . .
1 1 years . . . . . . . . . . . . . . . . . .
1 2 years . . . . . . . . . . . . . . . . . .
1 3 years 0.65 1.32 2.39 2.14 0.94 2.42 1.02 0.77 1.4 7 2 .0 0 1 .5 6 2 .1 4 1 .1 9 1 .10 1.70 1.50 1.87 1.64
1 4 years -0.33 1.87 3.28 3.35 1.52 3.37 1.54 1.15 2.5 3 2 .8 1 2 .4 5 2 .8 9 2 .0 0 1 .70 2.31 2.35 2.85 2.73
1 5 years 0.71 2.23 3.74 4.01 1.58 4.07 2.03 1.30 2.7 2 3 .3 5 2 .7 7 2 .6 5 2 .1 3 2 .11 2.90 2.75 3.21 3.26
1 6 years 1.81 2.09 4.06 4.50 1.80 3.95 2.28 1.53 3.1 4 3 .5 3 2 .8 7 2 .6 1 2 .2 0 2 .47 2.76 2.97 3.29 3.51
G e nde r (m ale ) 0.11 0.01 -0.65 -0.24 -0.52 -0.42 -0.77 -0.15 -0 .1 7 0 .1 3 -0 .3 2 -0 .4 4 0 .0 1 - 0.54 -0.48 -0.41 -0.03 -0.55
Ye ars of scho oli ng of th e

0.17 0.25 0.21 0.15 0.17 0.24 0.18 0.22 0.1 7 0 .2 6 0 .2 0 -0 .0 2 0 .2 2 0 .18 0.17 0.23 0.28 0.22
fa mily he ad
-0.01 -0.01 -0.01 0.00 0.00 -0.01 0.00 0.00 0.0 0 -0 .0 1 0 .0 0 0 .0 0 -0 .0 1 0 .00 0.00 -0.01 -0.01 -0.01
fa mily he ad to th e squ are
L og arithm of the pe r ca pi ta
-0.01 0.23 0.46 0.13 0.36 0.24 0.36 0.25 0.2 4 0 .2 4 0 .2 8 0 .1 0 0 .3 3 0 .35 0.16 0.25 0.29 0.48
i nco me
Pr esen ce of fa mily he ad an d
0.08 0.06 0.29 0.17 0.02 0.29 0.16 -0.08 0.0 0 -0 .0 8 0 .0 9 0 .1 3 0 .1 2 0 .14 0.01 0.19 -0.07 0.29
s pou se
N umb er of chil dren a ge 0 to
-0.09 -0.11 -0.13 -0.12 -0.22 -0.22 -0.13 -0.17 -0 .1 0 -0 .1 2 -0 .1 4 -0 .0 9 -0 .1 5 - 0.21 -0.20 -0.19 -0.18 -0.20
1 6 in the re side nce
Ar ea (urb an) -0.17 0.69 0.41 -0.09 0.80 0.11 0.10 0.19 0.4 5 0 .5 2 0 .2 8 0 .3 1 -0 .3 4 0 .80 0.26 -0.11 0.55 -0.11
Inequality of opportunity: Access to eletricity - circa 1995
Domi nica n
R epu bli c
Inte rcept ....... -4.02 -3.85 -5.13 -0.88 ....... ....... -3 .2 5 -2 .4 0 -2 .1 0 -6 .8 5 0 .7 3 - 0.59 -3.13 -3.29 -4.32 -5.79 .......
G e nde r (m ale ) ....... -0.19 -0.68 -0.62 -0.35 ....... ....... -0.04 -0 .4 5 -0 .2 0 -0 .2 0 0 .2 3 0 .2 1 - 0.25 -0.17 -0.84 -0.23 .......

....... 0.18 0.31 0.09 0.17 ....... ....... 0.05 0.0 7 0 .1 6 0 .1 0 -0 .2 1 -0 .0 4 0 .17 0.20 0.26 0.15 .......
fa mily he ad
....... -0.01 0.00 0.00 0.00 ....... ....... 0.01 0.0 0 0 .0 0 0 .0 1 0 .0 1 0 .0 2 0 .00 0.00 -0.01 0.00 .......
i nco me ....... 0.67 0.82 0.61 0.29 ....... ....... 0.48 0.5 1 0 .4 3 0 .9 0 0 .0 0 0 .4 3 0 .59 0.72 0.40 0.88 .......
....... -0.15 0.84 0.36 -0.27 ....... ....... -1.42 0.3 4 -0 .5 0 -0 .2 3 0 .5 1 0 .0 1 - 0.24 -0.42 0.86 -0.20 .......
s pou se
....... -0.07 -0.02 -0.16 -0.14 ....... ....... -0.08 -0 .0 8 0 .0 0 0 .0 9 0 .1 1 -0 .0 2 - 0.02 -0.06 0.00 0.01 .......
Ar ea (urb an) ....... 2.97 2.81 3.49 3.01 ....... ....... 3.46 1.9 5 1 .6 9 2 .9 3 0 .9 2 2 .1 5 2 .05 3.43 1.90 2.50 .......

Inequality of opportunity: Access to eletricity - circa 2005
Domi nica n
R epu bli c
Inte rcept -3.53 -2.52 -1.01 -4.73 -1.24 -4.47 -5.62 0.5 6 -2 .8 6 -3 .0 2 -4 .9 5 1 .6 7 - 0.68 -4.07 -2.28 -9.29 -3.64 -8.68
G e nde r (m ale ) 0.36 -0.79 -0.22 -1.01 -0.64 -0.75 -1.16 -0.83 -0 .0 9 -0 .4 0 -0 .5 4 0 .4 1 -0 .9 2 - 0.58 -0.43 0.11 -0.17 -0.23

-0.76 0.06 0.20 0.10 0.08 0.13 0.17 0.09 0.1 2 0 .1 3 0 .0 5 -0 .1 3 -0 .0 9 0 .23 0.06 0.09 0.12 0.00
fa mily he ad
0.07 0.00 0.00 0.00 0.00 0.02 0.00 0.00 0.0 0 0 .0 0 0 .0 1 0 .0 1 0 .0 3 - 0.01 0.00 0.01 0.00 0.02
i nco me 1.53 0.33 0.50 0.64 0.30 0.76 0.87 0.48 0.6 0 0 .6 7 0 .7 4 0 .0 7 0 .6 5 0 .62 0.84 0.90 0.61 1.27
0.24 0.22 0.20 1.10 0.12 0.30 0.78 0.36 0.0 8 -0 .1 9 0 .1 9 0 .0 1 0 .9 7 - 0.07 -0.18 0.15 -0.02 0.50
s pou se
0.55 -0.02 -0.10 -0.06 0.02 -0.09 0.19 -0.14 -0 .0 1 -0 .0 2 0 .0 0 -0 .1 2 -0 .2 1 - 0.05 -0.23 -0.01 -0.13 -0.02
Ar ea (urb an) 1.34 2.44 3.07 2.63 0.10 4.45 1.05 2.91 1.4 2 1 .0 6 2 .9 2 0 .7 8 0 .9 6 2 .77 2.16 1.30 2.88 2.79
Inequality of opportunity: Adequate access to water and sanitation - circa 1995
Dominican
Argentina Bolivia Brazil Chile Colombia Costa Rica Ecuador El Salvador Guatemala Honduras Jamaica Mexico Nicaragua Panama Paraguay Peru Uruguay Venezuela
Republic
Intercept ....... -5.37 -5.88 -12.92 -3.40 ....... ....... -5.47 -9.21 -5.79 -4.50 -1.27 -6.95 -8.45 -4.01 -13.56 -8.03 ....... -4.90
Gender (male) ....... -0.12 -0.22 -0.39 -0.38 ....... ....... 0.02 -0.21 -0.66 -0.07 0.39 -0.40 -0.32 0.19 -0.40 0.48 ....... -0.38
Years of schooling of the

....... 0.02 0.19 -0.06 0.13 ....... ....... 0.20 0.06 0.18 0.07 -0.31 0.05 0.02 0.03 0.15 0.00 ....... 0.10
family head
Years of schooling of the
....... 0.01 -0.01 0.01 0.00 ....... ....... 0.00 0.00 0.00 0.00 0.03 0.01 0.01 0.01 0.00 0.01 ....... 0.00
family head to the square
Logarithm of the per capita
....... 0.58 0.77 1.02 0.32 ....... ....... 0.30 0.86 0.57 0.55 -0.12 0.93 0.61 0.52 0.94 0.89 ....... 0.66
income
Presence of family head and
....... -0.34 0.03 -0.25 -0.02 ....... ....... -0.21 -0.19 -0.12 -0.05 -0.57 0.14 -0.51 -0.78 0.08 -0.81 ....... -0.22
spouse
Number of children age 0 to
....... -0.08 -0.04 0.25 -0.14 ....... ....... -0.12 0.01 -0.10 0.01 -0.04 -0.05 0.03 -0.05 -0.06 0.03 ....... -0.03
16 in the residence
Area (urban) ....... 2.19 2.01 3.25 2.41 ....... ....... 1.48 3.44 2.49 0.89 2.28 2.02 4.27 2.06 1.80 3.28 ....... 1.84

Inequality of opportunity: Adequate access to water and sanitation - circa 2005
Domi nica n
R epu bli c
Inte rcept -1.80 -4.69 -4.90 -1 0.20 -4.86 -7.87 -8.42 -3 .0 4 -7 .3 9 -6 .8 1 -5 .7 6 -2 .8 7 - 7.17 -9.21 -4.30 -13 .8 4 -4.81 -1 4.78
G e nde r (m ale ) -0.53 -0.44 -0.20 0.01 -0.33 -0.26 0.02 0.09 -0 .0 7 -0 .6 1 -0 .1 0 0 .6 8 -0 .0 4 - 0.02 0.12 0.06 0.03 0.18

0.08 0.04 0.20 -0.12 0.11 0.15 0.09 0.09 0.0 5 0 .1 4 0 .1 8 -0 .1 3 0 .0 7 0 .21 0.00 0.19 -0.05 0.11
fa mily he ad
0.01 0.00 -0.01 0.01 -0.01 0.01 0.00 0.00 0.0 1 0 .0 0 0 .0 0 0 .0 1 0 .0 0 - 0.01 0.01 0.00 0.01 0.00
i nco me 0.54 0.44 0.61 0.88 0.47 0.94 0.88 0.52 0.7 4 0 .7 8 0 .5 4 0 .0 6 0 .9 0 0 .74 0.71 1.03 0.72 1.43
0.07 0.10 -0.10 -0.44 -0.79 -0.11 -0.45 -0.31 -0 .3 4 0 .2 9 -0 .1 3 -0 .2 9 -0 .0 2 - 0.45 -0.65 -0.49 -0.49 -0.14
s pou se
-0.12 -0.12 0.01 0.11 -0.07 -0.16 0.01 -0.11 0.0 0 -0 .0 8 0 .0 3 -0 .1 2 -0 .0 5 - 0.02 -0.07 -0.14 -0.07 -0.04
Ar ea (urb an) 0.28 2.05 2.14 3.00 2.08 1.03 1.51 1.73 2.5 4 2 .2 4 1 .8 2 1 .8 8 1 .0 5 3 .28 1.53 1.80 2.16 4.29
Table A.3.7: Index of Inequality of Opportunity (D) in 19 LAC Countries
Completion of 6th grade on time Electricity Water and sanitation
Country
circa 1995 circa 2005 reduction cir ca 1995 circa 2005 reduction circa 1995 circa 2005 reduction
Argentina ... .... 3% ...... . ......... 1% ......... ......... 8% ... ..... .

Bolivia 16% 12% 3% …….. 26% …….. 34% 37% -3%
Brazil 36% 20% 16% 8% 3% 5% 25% 18% 7%
Chile 6% 3% 4% 3% 1% 3% 14% 8% 6%
Colombia 19% 15% 5% ……. 2% ……. …….. 21% ………
Costa Rica 12% 9% 3% ........ 1% ........ ..... ... 5% .. ......
Dominican
Republic 16% 13% 3% ........ 4% ........ ..... ... 25% .. ......
Ecuador 10% 6% 4% 10% 4% 6% 33% 23% 10%
El Salvador 25% 20% 5% 14% 9% 5% 47% 37% 11%
Guatemala 37% 27% 10% 14% 11% 3% 52% 43% 10%
Honduras 22% 19% 3% 27% 22% 5% ……… 37% ……. .
Jamaica 3% 2% 1% 5% 3% 2% 39% 38% 1%
Mexico 10% 5% 4% 4% 1% 3% 33% 19% 14%
Nicaragua 34% 24% 10% 20% 24% -3% 50% 50% 0%
Panama 11% 8% 2% 23% 19% 4% 38% 32% 6%
Paraguay 15% 11% 4% 6% 3% 3% 34% 28% 6%
Peru 26% 17% 9% 29% 23% 6% 43% 31% 11%
Uruguay ... .... 7% ...... . ....... 1% ....... ....... 12% .. .....
Venezuela 10% 6% 4% 1% 1% 0% 8% 7% 1%
Average 18% 12% …… 13% 8% …. 35% 25% …..
Table A.3.8: Opportunity Index in 19 LAC Countries
Education Water and sanitation Electricity
Country circa circa circa circa circa circa
change change change
1995 2005 1995 2005 1995 2005
Argentina ....... 0.73 ....... ......... 0.75 ......... ......... 0.98 .........
Bolivia 0.50 0.61 0.11 0.23 0.19 -0.04 ........ 0.45 ........
Brazil 0.15 0.37 0.22 0.39 0.50 0.11 0.81 0.92 0.11
Chile 0.73 0.83 0.09 0.66 0.78 0.12 0.93 0.98 0.05
Colombia 0.43 0.56 0.13 ........ 0.48 ........ ........ 0.86 ........
Costa Rica 0.55 0.65 0.10 ........ 0.87 ........ ........ 0.98 ........
Dominican
Republic
0.45 0.57 0.12 ........ 0.36 ........ ........ 0.89 ........
Ecuador 0.64 0.78 0.13 0.23 0.42 0.19 0.75 0.90 0.15
El Salvador 0.28 0.34 0.07 0.16 0.24 0.08 0.65 0.75 0.10
Guatemala 0.16 0.24 0.09 0.10 0.18 0.08 0.57 0.66 0.10
Honduras 0.28 0.38 0.10 ........ 0.22 ........ 0.42 0.51 0.08
Jamaica 0.87 0.86 -0.01 0.11 0.10 -0.02 0.74 0.83 0.09
Mexico 0.68 0.80 0.12 0.28 0.46 0.18 0.90 0.97 0.06
Nicaragua 0.17 0.33 0.16 0.09 0.10 0.02 0.53 0.49 -0.03
Panama 0.67 0.70 0.03 0.25 0.29 0.04 0.53 0.59 0.06
Paraguay 0.45 0.59 0.14 0.29 0.38 0.10 0.83 0.91 0.08
Peru 0.25 0.45 0.19 0.20 0.32 0.12 0.42 0.52 0.10
Uruguay ....... 0.75 ....... ....... 0.66 ....... ....... 0.97 .......
Venezuela 0.62 0.73 0.11 0.73 0.77 0.04 0.98 0.98 0.00
Average 0.46 0.59 0.11 0.31 0.43 -0.05 0.70 0.80 -0.05
Chapter 8
Understanding the Determinants

of Poverty
Summary
A poverty profile describes the pattern of poverty, but is not principally concerned
with explaining the causes of poverty. Yet, a satisfactory explanation of why some
people are poor is essential if we are to be able to tackle the roots of poverty.
Among the key causes, or at least correlates, of poverty are
• Region-level characteristics, which include vulnerability to flooding or typhoons,

remoteness, quality of governance, and property rights and their enforcement
• Community-level characteristics, which include the availability of infrastructure

(roads, water, electricity) and services (health, education), proximity to markets,
and social relationships
• Household and individual characteristics, among the most important of which are
– Demographic, such as household size, age structure, dependency ratio, gender
of head
– Economic, such as employment status, hours worked, property owned
– Social, such as health and nutritional status, education, shelter.
Regression analysis is commonly undertaken to identify the effects of each of

these characteristics on income (or expenditure) per capita. Attention is needed to
choose the independent variables carefully, to be sure that they are indeed exoge-
nous. A number of more exotic techniques are now available for this purpose,
including classification and regression tree (CART) models and multiple-adaptive
regression splines (MARS models). 145
Haughton and Khandker
8
Regression techniques are good at identifying the immediate, proximate causes
of poverty, but are less successful at finding the deep causes; they can show that a
lack of education causes poverty, but cannot so easily explain why some people lack
education.
Learning Objectives
After completing the chapter on Understanding the Determinants of Poverty, you

should be able to
1. Identify the main immediate (“proximate”) causes of poverty.
2. Classify the main causes of poverty by characteristics related to the country or

region, the community, and the household and individual.
3. Explain how regression techniques may be used to identify the proximate causes
of poverty and their relative importance.
4. Explain why researchers generally prefer to use regressions to explain income (or
expenditure) per capita rather than whether an individual is poor.
5. Evaluate the assertion that the weakest part of poverty analysis is the under-
standing of poverty’s fundamental causes, and that this represents a “missing
middle” that makes it difficult to define a successful antipoverty strategy.
Introduction: What Causes Poverty?
A poverty profile describes the pattern of poverty, but is not principally concerned
with explaining its causes. Yet, a satisfactory explanation of why some people are
poor is essential if we are to be able to tackle the roots of poverty. This chapter
addresses the question of what causes poverty.
Poverty may be due to national, sector-specific, community, household, or
individual characteristics. This chapter summarizes some of the characteristics of
the poor by region, community, household, and individual characteristics and
then discusses how regression techniques can be used to determine the factors
“causing” poverty.
Two cautions are in order. First, it can be difficult to separate causation from
correlation. For instance, we know that poor people tend to have low levels of edu-
cation; but are they poor because they have little education, or do they have little
education because they are poor? A statistical association alone is not enough to
146 establish causality, and additional information is likely to be required.
CHAPTER 8: Understanding the Determinants of Poverty
8
Second, most of the “causes” of poverty that we identify in this chapter are imme-
diate (or “proximate”) causes, but not necessarily “deep” causes. For instance, sup-
pose that we can demonstrate that low levels of education do indeed increase the risk
of poverty. This is interesting, but now begs the question of why some people have
low levels of education in the first place: Were the school fees too high? Was there no
school nearby? Was the quality of the education abysmal? Were their parents unsup-
portive, or even hostile to education? Was there a concern that an educated woman
could not find a husband?
The weakest part of poverty analysis—what Howard White and David Booth
(2003) call the “missing middle”—is developing a clear understanding of the funda-
mental causes of poverty in a way that leads naturally to an effective strategy to com-
bat poverty. Because there is no reason to believe that the root causes of poverty are
the same everywhere, country-specific analysis is essential.
Region-Level Characteristics
At the regional (or countrywide) level, numerous characteristics might be associated

with poverty. The relationship of these characteristics with poverty is country spe-
cific. In general, however, poverty is high in areas characterized by geographical iso-
lation, a low resource base, low rainfall, and other inhospitable climatic conditions.
For example, many argue that economic development in Bangladesh is severely
retarded because of its susceptibility to annual floods; and Nghe An province in
north-central Vietnam is poor in part because it is regularly hit by typhoons, which
destroy a significant part of the accumulated stock of capital. In many parts of the
world the remoteness of rural areas—which lowers the prices farmers get for their
goods and raises the prices they pay for purchases because of high transport costs—
is responsible for generating food insecurity among the poor. Inadequate public
services, weak communications and infrastructure, as well as underdeveloped mar-
kets, are dominant features of life in rural Cambodia, as in many other parts of the
world, and clearly contribute to poverty.
Other important regional and national characteristics that affect poverty include
good governance; a sound environmental policy; economic, political, and market
stability; mass participation; global and regional security; intellectual expression;
and a fair, functional, and effective judiciary. Region-level market reforms can boost
growth and help poor people, but they can also be a source of dislocation. The effects
of market reforms are complex, deeply linked to institutions and to political and
social structures. The experience of transition, especially in countries of the former
Soviet Union, shows that market reforms in the absence of effective domestic insti-
tutions can fail to deliver growth and poverty reduction, at least initially.
Inequality is also relevant to the analysis of poverty; its measurement is the sub-
ject of chapter 6. Gender, ethnic, and racial inequality are both dimensions of—and 147
8
causes—of poverty. Social, economic, and ethnic divisions in regions are often
sources of weak or failed development. In the extreme, vicious cycles of social divi-
sion and failed development erupt into internal conflict (within or across regions),
as in the Balkans and Liberia, with devastating consequences for people.
Community-Level Characteristics
As with regional characteristics, a variety of community-level characteristics may be

associated with poverty for households in that community. At the community level,
infrastructure is a major determinant of poverty. Indicators of infrastructure devel-
opment often used in econometric exercises include proximity to paved roads,
availability of electricity, proximity to large markets, availability of schools and
medical clinics in the area, and distance to local administrative centers. Other
indicators of community-level characteristics include average human resource
development, access to employment, social mobility and representation, and
land distribution.
Recently, there has been more emphasis on the importance of social networks and
institutions, and “social capital,” which includes, for instance, the level of mutual
trust in the community (Putnam 1995). In addition to removing social barriers,
effective efforts to reduce poverty require complementary initiatives to build up and
extend the social institutions of the poor. Social institutions refer to the kinship sys-
tems, local organizations, and networks of the poor and can be thought of as differ-
ent dimensions of social capital. Research on the roles of different types of social
networks in poor communities confirms their importance. An analysis of poor vil-
lages in north India, for example, shows that social groups play an important role in
protecting the basic needs of poor people and in reducing risk (Kozel and Parker
2000). A study of agricultural traders in Madagascar shows that social relationships
are central; close relationships with other traders help lower transactions costs, while
longstanding ties to creditors are vital sources of security and insurance (Fafchamps
and Minten 1998).
How does social capital affect development? The narrowest view holds social cap-
ital to be the social skills of an individual—one’s propensity for cooperative behav-
ior, conflict resolution, tolerance, and the like. A more expansive “meso” view
associates social capital with families and local community associations and the
underlying norms (trust, reciprocity) that facilitate coordination and cooperation
for mutual benefit. A “macro” view of social capital focuses on the social and politi-
cal environment that shapes social structures and enables norms to develop. This
environment includes formalized institutional relationships and structures, such as
government, the political regime, the rule of law, the court system, and civil and
political liberties. Institutions have an important effect on the rate and pattern of
148 economic development.
8
As the World Bank (2000, 129) writes, “An integrating view of social capital rec-
ognizes that micro, meso, and macro institutions coexist and have the potential to
complement one another. Macro institutions can provide an enabling environment
in which micro institutions develop and flourish. In turn, local associations help sus-
tain regional and national institutions by giving them a measure of stability and
legitimacy—and by holding them accountable for their actions.” Social capital is
clearly a complicated characteristic and often researchers find it difficult to identify
appropriate variables that measure social capital quantitatively.
Household and Individual-Level Characteristics
Some important household and individual characteristics would include the age
structure of household members, education, gender of the household head, and the
extent of participation in the labor force. In recent times, other components under
this category have included domestic violence prevention and gender-based antidis-
crimination policies. The following discussion organizes these characteristics into
groups and discusses them in greater detail. These groups are demographic, eco-
nomic, and social characteristics.
Demographic Characteristics
Indicators of household size and structure are important in that they show a possible
correlation between the level of poverty and household composition. Household
composition—the size of the household and characteristics of its members (such as
age)—is often quite different for poor and nonpoor households. The Cambodia
Socio-Economic Survey (CSES) of 1993–94 shows that the poor tend to live in larger
households, with an average family size of 6.6 persons in the poorest quintile com-
pared with 4.9 in the richest quintile (Gibson 1999). Similar patterns are found in
most countries, although the effect is attenuated if welfare is measured on a per adult
equivalent rather than a per capita basis. The poor also tend to live in younger house-
holds, with the bottom quintile having twice as many children under age 15 per fam-
ily as the top quintile, and slightly fewer elderly people over age 60. Better-off
households also tend to be headed by people who are somewhat older.
The dependency ratio is the ratio of the number of family members not in the
labor force (whether young or old) to those in the labor force in the household. This
ratio allows one to measure the burden weighing on members of the labor force
within the household. One might expect that a high dependency ratio will be asso-
ciated with greater poverty.
It is widely believed that the gender of the household head significantly influences
household poverty, and more specifically, that households headed by women are 149
8
poorer than those headed by men. This might be expected to be of particular impor-
tance in Cambodia. Because of male casualties in past wars, women are often the
heads of households. Women play an important role in the labor force, both in the
financial management of the household and in the labor market, but appear to face a
large degree of discrimination. They are severely affected by both monetary and non-
monetary poverty; for example, they have low levels of literacy, are paid lower wages,
and have less access to land or equal employment. Thus, many observers are surprised
to learn that poverty rates are not higher among female-headed than male-headed
households in Cambodia. Likewise, female-headed households in neighboring
Vietnam are no more likely to be in poverty than their male-headed counterparts.
Economic Characteristics
Apart from income or consumption—which are typically used to define whether a

household is poor—there are a number of other economic characteristics that cor-
relate with poverty, most notably household employment and the property and
other assets owned by the household.
There are several indicators for determining household employment. Within this
array of indicators, economists focus on whether individuals are employed, how
many hours they work, whether they hold multiple jobs, and how often they change
employment.
The property of a household includes its tangible goods (land, cultivated areas,
livestock, agricultural equipment, machinery, buildings, household appliances, and
other durable goods) and its financial assets (liquid assets, savings, and other finan-
cial assets). These indicators are of interest because they represent the household’s
inventory of wealth and therefore affect its income flow. Furthermore, certain house-
holds, especially in rural areas, can be poor in income, but wealthy when their prop-
erty is taken into consideration. Despite its importance, property is difficult to value
in practice in any reliable way. First, one encounters the problem of underdeclaration.
Second, it is very difficult to measure certain elements of property, such as livestock.
Finally, the depreciation of assets may be difficult to determine for at least two rea-
sons: (a) the life span of any given asset is variable, and (b) the acquisition of these
assets occurs at different moments in each household. Therefore, property is more
difficult to use than certain other elements in the characterization of poverty.
Social Characteristics
Aside from the demographic and economic indicators, several social indicators are
correlated with poverty and household living standards. The most widely used are
150 measures of health, education, and shelter.
8
Four types of indicators are normally used to characterize health in analyzing a
household’s living standards. These indicators include
• Nutritional status, for example, anthropometric indicators such as weight for age,
height for age, and weight for height
• Disease status, for example, infant and juvenile mortality and morbidity rates as
related to certain diseases such as malaria, respiratory infections, diarrhea, and
sometimes poliomyelitis
• Availability of health care services such as primary health care centers, maternity
facilities, hospitals and pharmacies, basic health care workers, nurses, midwives,
doctors and traditional healers; and medical service such as vaccinations and
access to medicines and medical information
• The use of these services by poor and nonpoor households.
Three types of indicators are normally used to characterize education in an analy-

sis of household living standards. These include the level of education achieved by
household members (basic literacy, years of education completed); the availability of
educational services, such as proximity to primary and secondary schools; and the
use of these services by the members of poor and nonpoor households. For this last
item, commonly used measures include children’s registration in school, the dropout
rate of children by age and gender and reasons for dropping out, the percentage of
children who are older than the normal age for their level of education, and average
spending on education per child registered.
Literacy and schooling are important indicators of the quality of life in their own
right, as well as being key determinants of poor people’s ability to take advantage of
income-earning opportunities. Based on CSES data, Cambodia by 1993–94 had
achieved a self-reported basic literacy rate of 67 percent among adults (older than
age 15), implying a high degree of literacy among the poor. However, the literacy gap
remained quite large, with literacy ranging from just over half of adults (58 percent)
among the poorest quintile of the population to 77 percent among the richest quin-
tile. Much larger differentials appear in the distribution of schooling attainment:
adults in the poorest quintile averaged 3.1 years of schooling, compared with 5.3
years among the richest quintile. Men averaged 5.1 years of education, compared
with 3.2 years for women.
Shelter refers to the overall framework of personal life of the household. It is evalu-
ated, by poor and nonpoor household groups, according to three components (some
of which overlap with the indicators mentioned above): housing, services, and the
environment. Housing indicators include the type of building (size and type of mate-
rials), the means through which one has access to housing (renting or ownership), and
household equipment. The service indicators focus on the availability and the use of 151
8
drinking water, communications services, electricity, and other energy sources. Finally,
the environmental indicators concern the level of sanitation, the degree of isolation
(availability of roads and paths that are usable at all times, length of time and avail-
ability of transportation to get to work), and the degree of personal safety.
Example: It is generally established that poor households live in more precari-

ous, less sanitary environments, which contribute to the poorer health and
lower productivity of household members. To illustrate, the data from the
CSES of 1993–94 show that water and sanitation are especially important influ-
ences on health and nutritional status. The CSES showed that only 4 percent of
the poorest quintile had access to piped water, while more than 17 percent of
the richest quintile had the same. Similar differences are apparent in access to
sanitation. Just 9 percent of the poor had access to a toilet in the home, while
around half of the richest quintile did.
Another indicator of housing standards is access to electricity. Here again,
access of the poor lagged far behind. Access to electricity from a generator or
line connection rose sharply with income, from a mere 1 percent among peo-
ple in the bottom quintile to 37 percent of Cambodians in the richest
quintile. Other indicators of household wealth include ownership of trans-
portation. Access to bicycles is quite evenly distributed, with at least one-half
of households owning a bicycle in every quintile, even the poorest. However,
access to cars, jeeps, or motorbikes is very rare among the poor and rises
sharply with income.
A summary of the main influences on poverty is provided in table 8.1.
Analyzing the Determinants of Poverty: Regression Techniques
Tabulated or graphical information on the characteristics of the poor is immensely

helpful in painting a profile of poverty. However, it is not always enough when one
wants to tease out the relative contributions of different influences on poverty. For
example, tabulated data from the Vietnam Living Standards Survey of 1998 showed
per capita expenditure to be significantly higher in female-headed households than
in households headed by a man. However, after controlling for other influences—
where the household lived, the size of the household, and so on—the effect proved
to be statistically insignificant.
By far, the most widespread technique used to identify the contributions of dif-
ferent variables to poverty is regression analysis, a subject treated in some detail in
chapter 14. Here we simply summarize the essentials of regression, to allow this
152 chapter to be self-contained.
8
Table 8.1 Main Determinants of Poverty
Regional characteristics Isolation or remoteness, including less infrastructure and poorer access to
markets and services
Resource base, including land availability and quality
Weather (for example, whether typhoons or droughts are common) and
environmental conditions (for example, frequency of earthquakes)
Regional governance and management
Inequality
Community characteristics Infrastructure (for example, piped water, access to a tarred road)
Land distribution
Access to public goods and services (for example, proximity of schools, clinics)
Social structure and social capital
Household characteristics Size of household
Dependency ratio (that is, unemployed old and young relative to
working-age adults)
Gender of head, or of household adults on average
Assets (typically including land, tools, and other means of production;
housing; jewelry)
Employment and income structure (that is, proportion of adults employed;
type of work—wage labor or self-employment; remittances inflows)
Health and education of household members on average
Individual characteristics Age
Education
Employment status
Health status
Ethnicity
Source: Created by the authors.
There are two main types of analysis:

• Attempts to explain the level of expenditure (or income) per capita—the depend-
ent variable—as a function of a variety of variables (the “independent” or
“explanatory” variables). The independent variables are typically of the type dis-
cussed above in Household and Individual-Level Characteristics.
• Attempts to explain whether a household is poor, using a logit or probit regres-
sion. In this case the independent variables are as above, but the dependent vari-
able is binary, usually taking a value of 1 if the family is poor and 0 otherwise.
We now consider each of these in somewhat more detail.
A regression estimate shows how closely each independent variable is related to
the dependent variable (for example, consumption per capita), holding all other
influences constant. There is scope for a wide variety of regressions; for instance, the
dependent variable could measure child nutrition, or morbidity, or schooling, or
other measures of capabilities; the regressions could be used to examine the deter-
minants of employment or labor income; or regressions could be used to estimate
agricultural production functions (which relate production to information on type
of crops grown per area, harvest, inputs into agricultural production, and input and 153
8
output prices). For an accessible discussion and many examples, in the context of
Vietnam, see Health and Wealth in Vietnam (Haughton et al. 1999).
A typical multiple regression equation, as applied to poverty analysis, would look
something like this:
(8.1)
where z is the poverty line, yi is per capita income or consumption, the are the
“explanatory” variables, and the αj are the coefficients that are to be estimated. Note
that yi/z is in log form, which is a common way of allowing for the log normality of the
variable. Because we are interested in the determinants of individual poverty, but typ-
ically have information at the level of the household, it is standard (but in this context,
not universal) to estimate the regression using weights that reflect the size of the house-
hold. The “regress” command in Stata is flexible and allows the use of weights.
The independent (right-hand side) variables may be continuous variables, such
as the age of the individual. But often we want to represent a categorical variable—
the gender of the person, or the region in which he or she lives. In this case we need
to create a “dummy” variable; for instance, the variable might be set to 1 if the per-
son is a man and 0 for a woman. If there are, say, 10 regions in a country, each region
would need to have its own dummy variable, but one of the regions needs to be left
out of the regression, to serve as the point of reference.
Often we believe that the determinants of poverty differ from one area to the
next, which would mean that there are differences in “structure.” In this case we
could estimate separate regressions, for instance, for each region in a country. Some-
times it is sufficient to specify the regression equation in a way that is flexible enough
to allow for such differences, by allowing interactive effects. For example, one could
create a variable that multiplies educational level by age, instead of estimating sepa-
rate regressions for individuals in different age groups.
The fit of the equation is typically measured using (“adjusted R squared”),
which will vary between 0 (no fit) and 1 (perfect fit). There is no hard and fast rule
for determining whether an equation fits well, although with household survey data,
one is often pleased to get an of 0.5 or more.
We also need to know how much confidence to place in the accuracy of the coef-
ficients as guides to the truth; this is commonly done by reporting t-statistics, which
are obtained by dividing a coefficient by its standard error. The rule of thumb is that
if the t-statistic is, in absolute terms, less than 2, the coefficient is not statistically sig-
nificantly different from zero (at about the 95 percent confidence level); in other
words, we cannot be sure that we have picked up an effect, and it is possible that the
coefficient just reflects noise in the data. Many researchers prefer to report p-values,
154 which give the confidence level directly; a p-value of, say, 0.03 indicates that we are
8
Table 8.2 Determinants of Household Spending Levels in Côte d’Ivoire, about 1993
Urban Rural
Coefficient t-statistic Coefficient t-statistic
Dependent variable: ln(expenditure/capita)
Educational level of most educated
male
Elementary 0.38 5.3 0.04 0.6
Junior secondary 0.62 8.6 0.08 0.9
Senior secondary 0.80 9.6 0.05 0.4
University 0.93 9.4
Educational level of most educated
female
Elementary 0.11 1.7 0.07 1.0
Junior secondary 0.24 3.1 0.27 2.2
Senior secondary 0.34 4.1
University 0.52 4.1
Value of selected household assets
Home 0.06 5.3
Business assets 0.04 3.3 0.16 4.9
Savings 0.08 4.7
Hectares of agriculture land
Cocoa trees 0.17 4.3
Coffee trees 0.04 1.3
Distance to nearest paved road –0.04 –2.9
Distance to nearest market –0.09 –3.3
Unskilled wage 0.37 6.4
Source: Adapted from Grosh and Munoz (1996, 169), based on Glewwe (1990).
97 percent confident that the coefficient is not 0. So we hope to find low p-values
(and we usually do when working with large data sets). Arbitrarily, it is standard to
consider a coefficient to be statistically significant if the p-value is less than 0.05, but
this rule is not graven in stone.
Table 8.2 shows typical regression output from an example based on data from
Côte d’Ivoire. Here, the dependent variable is the log of per capita household expen-
diture. Separate regressions were estimated for households in urban and in rural
areas, on the thinking that the determinants of poverty might be quite different in
these two areas.
The results of the urban equation show that education is an important determinant
of expenditure per capita. The coefficients for most of the educational variables are sta-
tistically significant and quite large; having an elementary education boosts income by
approximately 38 percent relative to someone with no education; this comes from the
coefficient of 0.38, and the fact that the dependent variable is in log form.1
However, in rural areas education does not appear to explain expenditure per
capita levels very well, a not uncommon finding. Conversely, the infrastructure vari-
ables have substantial predictive power: households located in villages that are nearer 155
8
to both paved roads and public markets are better off, as are households living in areas
with higher wage levels. The results raise further questions about the quality of edu-
cation in rural areas (or its applicability in rural areas), and the importance of rural
infrastructure in helping families grow out of poverty, which could be addressed in
putting together a poverty reduction strategy.
It is vital to choose the independent variables carefully, and to be sure that they
are truly exogenous. For instance, in the example above, one could have included
income as an independent variable, along with education, assets, and the like. But
that does not advance us much, because income is in turn determined by such vari-
ables as educational levels and household assets. In our drive to find the underlying
causes of poverty, we need to dig deep to find variables that are indeed predeter-
mined. A good start is to work with the variables identified in table 8.1.
When multiple cross-sectional surveys are available, the same regression can be
repeated for different years to see how the association of certain correlates with
income or consumption varies over time. Variations over time will be reflected in
changes in coefficients or parameters. The results of repeated cross-section regres-
sions can also be used to decompose variations in poverty by changes in household
characteristics, and changes in the returns to (or impact of) these characteristics (for
example, Baulch et al. 2004; van de Walle and Gunewardena 2001; Wodon 2000).
Some researchers prefer to use, on the left-hand side, a binary variable that is set
equal to 1 if the household is poor, and to 0 otherwise. Some of the information is lost
by doing this, and the resulting logit or probit regression is relatively sensitive to spec-
ification errors, which is why this is rarely the preferred approach. However, such an
analysis is likely to be useful when designing targeted interventions (for example, edu-
cational vouchers for poor households) because it allows one to assess the predictive
power of various explanatory variables used for means testing. It is also possible to
undertake a multiple logit analysis, where the dependent variable could be in one of
several categories, for instance, expenditure quintiles. For further details see chapter 14.
Recent research has explored more exotic forms of analysis, including nonpara-
metric regression, classification and regression trees (CART models), and multiple-
adaptive regression splines (MARS models). The goal of all such efforts is to unearth
a parsimonious number of determinants of poverty, and quantify their effects, even
when those effects are highly nonlinear.
Review Questions
1. By the “missing middle,” White and Booth mean those households that
are too affluent to be counted as poor, but too poor to be considered
comfortably off.
True
° False
156 °
8
2. Region-level characteristics that are expected to influence poverty include
all of the following except
A. Geographic isolation.
° B. Insufficient rainfall.
° C. Low educational levels of households.
° D. An ineffective judiciary.
°
3. Which of the following is not generally considered to be a component of

social capital?
A. An individual’s social skills.

° B. An individual’s level of education.
° C. The level of mutual trust in a society.
° D. The extent of the rule of law in a society.
°
4. The dependency ratio measures the proportion of young and old to
working-age individuals in a household.
True
° False
°
5. Shelter includes:
A. The type of building in which one lives.

° B. Whether a household has piped drinking water.
° C. The sanitation level of housing.
° D. All of the above.
°
An analyst estimates the following regression equation, based on household survey
data:
Ln(expenditure/capita)
2.1 0.3 (has elementary education) 0.03 (distance to nearest paved road)
t 5.7 t 4.1 t 1.5
The value of R2 is 0.37. Expenditure per capita is measured in thousands of dollars per
year. The following three questions refer to this equation:
6. Are the signs of the coefficients plausible?
Yes
° No
°
7. Are all the coefficients significantly different from zero with at least 95%
probability?
Yes
° No
° 157
8
8. Achieving elementary education will raise expenditure per capita by $300
per year.
True
° False
°
Note
1. Strictly speaking, in this case it boosts income by e0.38 1 0.462 46.2%. For small
changes it is common to ignore this refinement.
References
Baulch, Bob, Truong Thi Kim Chuyen, Dominique Haughton, and Jonathan Haughton. 2004.
“Ethnic Minority Development in Vietnam: A Socio-Economic Perspective.” In Economic
Growth, Poverty and Household Welfare in Vietnam, ed. Paul Glewwe, Nisha Agrawal, and
David Dollar. Washington, DC: World Bank.
Fafchamps, Marcel, and Bart Minten. 1998. “Relationships and Traders in Madagascar.” MSSD
Discussion Paper No. 24, International Food Policy Research Institute, Washington, DC.
Gibson, John. 1999. “A Poverty Profile of Cambodia, 1999.” Report to the World Bank and the
Ministry of Planning, Phnom Penh.
Glewwe, Paul. “1990. Investigating the Determinants of Household Welfare in Côte d’Ivoire
in 1985.” Living Standards Measurement Study Working Paper No. 29, World Bank,
Washington, DC.
Grosh, Margaret E., and Juan Muñoz. 1996. “A Manual for Planning and Implementing the
Living Standards Measurement Study Surveys.” LSMS Working Paper Series No. 126, World
Bank, Washington, DC.
Haughton, Dominique, Jonathan Haughton, Sarah Bales, Truong Thi Kim Chuyen, and
Nguyen Nguyet Nga. 1999. Health and Wealth in Vietnam: An Analysis of Household Living
Standards. Singapore: Institute of Southeast Asian Studies.
Kozel, Valerie, and Barbara Parker. 2000. “Integrated Approaches to Poverty Assessment in
India.” In Integrating Quantitative and Qualitative Research in Development Projects, ed.
Michael Bamberger. Washington, DC: World Bank.
Putnam, Robert D. 1995. “Bowling Alone: America’s Declining Social Capital.” Journal of
Democracy 6 (1): 65–78.
Smith, Stephen C. 2005. Ending Global Poverty: A Guide to What Works. New York: Palgrave
Macmillan.
van de Walle, Dominique, and Dileni Gunewardena. 2001. “Sources of Ethnic Inequality in
Vietnam.” Journal of Development Economics 65 (1): 177–207.
Vietnam General Statistics Office. 2000. Viet Nam Living Standards Survey 1997–1998. Statistical
158 Publishing House, Hanoi.
8
White, Howard, and Richard Booth. 2003. “Using Development Goals to Design Country
Strategies.” In Targeting Development: Critical Perspectives on the Millennium Development
Goals, eds. Richard Black and Howard White. London and New York: Routledge Studies in
Development.
Wodon, Quentin T. 2000. “Poverty and Policy in Latin America and the Caribbean.” Technical
Paper No. 467, World Bank, Washington, DC.
World Bank. 2000. World Development Report 2000/2001: Attacking Poverty. Washington, DC:
World Bank.
159
Chapter 14
Using Regression
Summary
Regression is a useful technique for summarizing data and is widely used to test
hypotheses and to quantify the influence of independent variables on a dependent
variable. This chapter first reviews the vocabulary used in regression analysis, and
then uses an example to introduce the key notions of goodness of fit and statistical
significance (of the coefficient estimates).
Much of the chapter is taken up with a review of the main problems that arise in
regression analysis: there may be errors in the measurement of the variables, some
relevant variables may be omitted or unobservable, “dependent” and “independent”
variables may in fact be determined simultaneously, the sample on which the esti-
mation is based may be biased, independent variables may be correlated (“multico-
linearity”), the error term may not have a constant variance, and outliers may have
a strong influence on the results.
The chapter suggests solutions to these issues, including the use of more and bet-
ter data, fixed effects (if panel data are available), instrumental variables, and ran-
domized experiments.
Logistic regression has a binary dependent variable and is often used to explain
why some people are poor and others are not. The chapter explains how to interpret
the results of logistic regression.
273
14
Learning Objectives
After completing this chapter on Regression, you should be able to
1. Explain how regression may be used to summarize data, and why it is useful for
testing hypotheses and quantifying the effects of independent on dependent
variables.
2. Define the essential terms used in regression, including coefficient, slope, error,
residual, y hat, p-value, R2, and ordinary least squares (OLS).
3. Evaluate an estimated regression equation based on its fit and the signs and mag-
nitudes of the coefficients.
4. Assess the confidence we have in a coefficient estimate, based on the t-statistic or

p-value.
5. Describe and explain the main problems that arise in regression analysis, including
• Measurement error
• Omitted variable bias
• Simultaneity bias
• Sample selectivity bias
• Multicolinearity
• Heteroskedasticity
• Outliers
6. Summarize the most important solutions to the problems that arise in regression,
including using more or better data, fixed effects, instrumental variables, and ran-
domized experiments.
7. Explain how to interpret the results of a logistic regression, and determine when
such an approach to regression is useful.
Introduction
At its most basic, regression is a technique for summarizing and describing data pat-
terns. For instance, figure 14.1 graphs food consumption per capita (on the vertical
axis) against expenditure per capita (on the horizontal axis) for 9,122 Vietnamese
households in 2006. The axes show spending in thousands of Vietnamese dong
(VND) per year; VND 1 million is equivalent to about US$65. The data points are so
numerous and dense that it is difficult to get a good feel for the essential underlying
274 relationships.
CHAPTER 14: Using Regression
14
Figure 14.1 Scatter Plot and Regression Lines for Food Consumption per Capita
against Total Expenditure per Capita, Vietnam, 2006
Source: Vietnam Household Living Standards Survey 2006.
One solution is to estimate a regression line—for instance using the regress

command in Stata—that summarizes the information. There are actually two regres-
sion lines shown in figure 14.1—a straight line (the continuous line) and a quadratic
(the curve of x’s). Both of these lines capture the essential feature of the numbers,
which is that food expenditure rises when total spending rises. The straight line esti-
mated here (using Stata) may be written as follows:
(food spending per capita) = 1,078 + 0.24 (total spending per capita).
This shows that just under a quarter of any additional spending is devoted to
food; in economic jargon, the marginal propensity to spend on food is 0.24. The
quadratic curve shown in figure 14.1 may be written as follows1:
(food spending per capita) = 725 + 0.33 (total spending per capita)
– 0.0038 (spending/capita)2.
It follows from this equation that, as total spending rises, food spending rises but
less quickly—an example of Engel’s Law at work. For someone who is very poor,
almost a third of incremental spending goes to food, but as per capita spending levels
rise, the squared term becomes increasingly important and moderates the extent to
which extra spending is devoted to food.
But regression is used for much more than just summarizing voluminous data. It
is widely used to test theories and hypotheses; indeed, it is the workhorse of much of 275
14
social science. The use of regression for statistical inference is more difficult than its
use for description, but an appreciation of the issues involved is essential both to
assess the work of others and to do good research oneself. In what follows, we out-
line the essential features of regression, interpret a useful equation, list and discuss
the most important problems that arise when using regression, consider solutions to
these problems, and explain how to use and interpret logistic regression. There is an
enormous literature on regression: Anderson, Sweeney, and Williams (2008) provide
a basic treatment; Wooldridge (2008) gives a comprehensive introductory treatment
of econometrics; and Greene (2007) is the essential reference for academic
researchers and advanced graduate students.
The Vocabulary of Regression Analysis
Suppose that we are interested in explaining why some households have higher
levels of consumption per capita than others. This is the dependent variable, con-
ventionally labeled y. For each of the N households, we have a different value of
yi, i =1,…, N.
On the basis of theory, or prior practice, we have some variables that we
believe “explain” differences across households in consumption per capita—for
instance, the educational attainment of adult household members or whether a
household lives in an urban area. These are the independent variables, sometimes
referred to as the covariates, and usually are denoted by X1, X2, … , Xk, if there are
k variables.
We may write the true linear relationship between the independent and depend-
ent variables, for the case with two covariates, as follows
(14.1)
Here β0 is the intercept (or constant term), and the βj are the coefficients (or
slopes) of the X variables. We have information on the y and X variables, and have to
estimate the β coefficients, which typically is done using statistical software such as
Stata, SPSS, or SAS. In practice, the independent variables never “explain” the
dependent variable exactly; this is why equation (14.1) includes a random error εi,
which picks up measurement error as well as the effects of unobservable (and unob-
served) influences.
By the nature of its construction, we cannot measure εi directly. But for the pur-
poses of statistical inference, we need to know how the random errors are distrib-
uted. Ideally, we would like the error terms εi to be identically and independently
distributed following a normal distribution with mean 0 and constant variance σ2.
In practice, this may not always be reasonable, in which case we may need to adapt
276 the model in equation (14.1), as explained below.
14
Although the true εi errors cannot be observed directly, we can estimate them.
First estimate the coefficients in equation (14.1); by convention, these estimated
values are typically written as . Now apply these coefficient estimates to the
independent variables to get the predicted values of y, which are usually written
as (“y hat”), as follows:
(14.2)
Now we may compute the residual, defined as which is the difference

between the actual value of y and the value as predicted by the model that we are
using. If equation (14.1) fits well, these residuals will be very small relative to y.
The most common method for estimating linear regression equations such as
that in equation (14.1) is ordinary least squares (OLS). It essentially picks estimates
of the coefficients (that is, the values) that minimize the sum, over all the obser-
vations, of the squared residuals. This is not the only possible method of estimation,
but it is so standard that it is the default in all statistical packages.
In the linear case shown here, the coefficient estimates are relatively easy to inter-
pret. Take for example; it measures the effect that a one-unit increase in X1 will
have on y, holding all other effects constant (or, put differently, “controlling for the
effects of other covariates”).
This is an important point. The tables in a poverty profile can suggest links
between variables such as education and the incidence of poverty, but they do not
quantify strengths of such links. Nor do they control for the effects of other variables.
For instance, in Cambodia, female-headed households are less likely to be in poverty
than male-headed households. However, if female-headed households tend to be in
the cities, then the higher income levels of these households may not be a result of the
fact that they are headed by a woman, but rather because they are urban. One of the
most powerful features of regression analysis is that it allows one to separate out
effects of this nature—in this case, to control for the urban effect and measure the
importance of female-headedness separately from any confounding influences.
Examining a Regression Example
We are now ready to look in more detail at an actual example of a regression esti-
mate. Based on information from the Cambodian Socio-Economic Survey (SESC)
undertaken in 1993–94, Prescott and Pradhan (1997, 55) estimate the following
regression:
Ln(Consumption/capita) = 7.17 – 0.64 dependency ratio
215 –17.9
+ 0.15 femaleness + 0.86 Phnom Penh
3.97 45.7 277
14
+ 0.23 Other urban + 0.06 Average years of education of adults,
11.6 21.1
where R2 = 0.47.
Let us now interpret this equation:
• The value of R2 is 0.47. This measures the goodness of fit of the equation and lies
between 0 (no fit) and 1 (a perfect fit). Although one cannot say that a value of
0.47 is good or bad, it is certainly in line with what one would expect from a
cross-section regression of this nature. It may often be interpreted as saying that
47 percent of the variation in the log of consumption per capita is “explained” by
the independent (that is, right-hand side) variables, although here as always one
must be cautious about inferring causality rather than mere correlation.
• The coefficients (that is, –0.64, +0.15, and so on) have the expected signs. For
instance, more years of education are associated with higher consumption per
capita. In this particular case, a household in which the adult members have, on
average, one more year of education than an otherwise identical household, will
have 6 percent higher consumption per capita. This is because an extra year of
education is associated with an increase in ln(consumption/capita) of 0.06, or
about 6 percent.
• Consumption per capita is expressed in log terms. This is common. The untrans-
formed measure of consumption per capita is highly skewed to the right (that is,
the distribution has a long upper tail), which probably means that the errors in this
regression are not normally distributed. A log transformation usually solves this
problem. The broader question here is how to make the appropriate choice of
functional form for the equation.
• The numbers under the coefficients are t-statistics. The t-statistic is computed as
the estimated coefficient divided by its standard error. The rule of thumb is that if
the t-statistic is greater (in absolute value) than about 2, then we are roughly 95
percent confident that the coefficient is statistically significantly different from
zero—that is, we have not picked up a statistical effect by mistake. Most software
packages automatically generate the p-value, which is one minus the degree of sta-
tistical significance of the coefficient. Thus, a p-value of 0.02 means that there is a
98 percent probability that the coefficient is different from zero. Low p-values, typ-
ically under 0.05 or 0.1, give us confidence in the estimates of the coefficients.
• The Phnom Penh variable is binary. This variable is set equal to 1 if the house-
hold is in Phnom Penh, and to 0 otherwise. This is also true of the “other urban”
variable. The “rural” category has been left out (deliberately) and serves as the
reference category. Thus we may say, in the above equation, that consumption
278 per capita in other urban areas is about 23 percent higher than in rural areas.
14
• The dependency ratio. This ratio measures the number of young plus old house-
hold members to prime-age household members. When the dependency ratio is
high, we expect consumption per capita to be low, and indeed this is what the
regression shows.
• The “femaleness” measure. This measures the percentage of working-age house-

hold members (15–60 years old) who are women. This is probably a better meas-
ure than the gender of the head of household, because it is not always clear what
is meant by “head” of household. The U.S. census no longer uses the term at all.
The regression here shows that more-female households have higher levels of
consumption per capita in Cambodia, holding other variables constant.
Problems in Regression Analysis
It is not too difficult to gather data and estimate regression equations. That may be
the easy part, because there are plenty of pitfalls in the analysis of regression, and
dealing with these problems requires some practice and experience. In this section,
we outline the most important problems; some ways to deal with these problems are
set out briefly in the subsequent section.
Measurement Error
No variable is measured with complete precision, and many socioeconomic vari-

ables, including income and expenditure variables, are quite imprecise. In some
cases, even a variable that should be easy to quantify, such as a respondent’s age, may
not be correct; in many surveys, too many people report their age as, say, 70 or 75,
and too few report their age as 71 or 74.
Let S be the true measure of a variable, but assume that we only observe S*. Thus
S* = S + w.
observed true value random error
The effects of measurement error depend on whether it appears in the depend-

ent or in the independent variables.
Case 1. Measurement Error in Y. Suppose that we have the true equation

Y = a + bX + ε
but all we observe is Y*, so
Y* = a + bX + (ε+w).
In this case, the estimate of the coefficient b will not be biased, but the overall fit
of the equation will be poorer, because the error term is larger and hence noisier. 279
14
Case 2. Measurement Error in X. Again suppose that the true equation is
Y = a + bX + ε
but what we observe is now
Y = a + bX* + (ε – bw).
This time we have a noisy measure of the true X, and the estimate of b will be
biased toward 0, because the dependence of Y on X is masked. Formally,
where σ2 represents the true variance of the variable in question. As the error w
becomes more variable, the estimated value of b (that is, ) tends to zero. This
problem can easily arise. For instance, if we have
health = a + b schooling + ε
and if there is error in measuring the schooling variable, then the effect of schooling
on health will be understated.
Omitted Variable Bias
Suppose we leave a right-hand side (“independent”) variable out of the equation that
we estimate—perhaps because it is unobserved, or unavailable, or just overlooked.
Then the estimated coefficients on the remaining variables generally will be wrong if
the included variables are correlated with the omitted variables.
To see this, suppose the correct model is
HC = a + bSM + cAM + ε
Child’s Mother’s Mother’s
health schooling ability
In this case, the health of the child depends in part on the mother’s ability, but
this is a variable that we cannot observe. So when we regress HC on SM, we have, in
effect, a compound error term c.AM + ε. In this case, it can be shown that
where the last item is the covariance between schooling and (unobserved) ability. If
higher ability leads to more schooling (as is likely, so σSA > 0), and if higher ability
leads to better health (so c > 0), then the estimated value of b (that is, ) will be too
high. In effect, the estimated value is picking up ability as well as schooling effects,
280 and in doing so, it overstates the contribution of schooling to child health.
14
The problem of omitted variables is widespread. We are rarely able to specify a
model so completely that nothing important has been overlooked. And several
personal characteristics—ability, drive, motivation, flexibility—defy reliable quan-
tification. If bright, motivated farmers are more likely to use fertilizer and to use it
well, then the measured effect of fertilizer use on output is likely to be overstated,
because it is confounded with relevant unobservables. If clever, driven individuals
are more likely to get a good education, then the measured effect of education on
earnings will surely be overstated, because it reflects, in part, the contribution that
should properly be attributed to personal traits rather than to education per se.
Simultaneity Bias
Many applied research problems come up against the problem of simultaneity bias.
Although common, this can be a difficult problem to solve. We may illustrate it as
follows: Suppose
Child health = a + b Nutrients + ε.
Our goal is to explain child health, and we believe, quite reasonably, that better-fed
children will be healthier. But the problem here is that the child’s health may deter-
mine how much the household feeds her or him. For instance, parents might feed a
sick child more, in the hope that he or she will get better faster that way. But then we
would see a negative relationship, whereby more nutrients could be associated—in
the regression analysis—with poorer child health.
Formally, if there is simultaneity present, the estimate of the coefficient will gen-
erally be wrong. In our example, if parents provide better feeding to sick children,
then the estimate will be too low. One solution is to include lots of predetermined
variables in the estimated equation so that in the presence of simultaneity, its effect
will be attenuated.
Sample Selectivity Bias
Frequently, observations are available only for a subset of the sample that interests us.
For instance, we might want to measure the spending of the very poor, but we have
information only for people with a home. In this case, our sample would omit the
homeless, and our results are likely to underestimate the true extent of poverty. Or
again, we might want to know how much people would be willing to pay for piped
water, but we have information only about willingness-to-pay for households that cur-
rently have piped water—hardly a representative sample of the population at large. In
both of these cases, the problem is that our data may not come from a random sample
of the relevant population, and so our regression estimates risk being biased. 281
14
Again, an illustration from Behrman and Oliver (2000) is helpful. Consider the
common situation in which we are interested in determining the extent to which
additional schooling leads to higher wages. The immediate problem is that we observe
wages only for those who are skilled, dynamic, or educated enough to receive a wage
(rather than work on a farm or in self-employment). This means that wage earners
are not a random sample of the population.
The result is that we may have a situation as illustrated in figure 14.2. As shown,
each dot represents one person, and the estimated line linking wages to schooling is
too flat compared with the true relation. Thus, our estimate is biased.
The most common solution to the sample selection problem is to use Heckman’s
two-step procedure.
• First, estimate a probit model to determine who earns a wage. This is similar to a
logistic model (see the section, “Logistic Regression”), in which the dependent
variable is binary (1 if the person earns a wage, 0 otherwise).
• Second, estimate the wage equation—that is, wage as a function of schooling,

experience, location, and so on—in which one includes an additional term (the
inverse Mills ratio; also sometimes referred to as Heckman’s λ) that is derived
from the residuals of the probit model. Most statistical packages have commands
that do this quite easily.
Figure 14.2 A Hypothetical Example of the Link between Schooling and Wages
Source: Authors’ creation.

Note: The true relation here is relatively steep; the estimated line is based only on individuals who work
282 for a wage.
14
For this procedure to work, the initial probit model needs to include at least some
variables that do not appear in the wage equation, so that the wage effect may be
identified. This is not always easy, because the variables that affect what wage you get
are also likely to influence whether you work for a wage or not.
Multicolinearity
One of the most common problems in regression analysis is that the right-hand vari-
ables may be correlated with one another. In this case, we have multicolinearity, and
the problem is that the estimated coefficients can be quite imprecise and inaccurate,
even though the equation itself may fit well.
To see how this might occur, suppose that we are interested in modeling the
determinants of the number of years of schooling that girls get (y). We believe that a
girl will tend to get more schooling if her mother has more education (ME), or her
father has more education (FE), or she lives in an urban area (URB). A simple regres-
sion model would then look like this:
yi = β0 + β 1MEi + β2FEi + β3URBi + εi.
Many studies have found that the education of the mother has a strong influence
on whether a girl goes to school. However, it is typically the case that educated peo-
ple tend to marry each other (assortative mating); thus, a high level of ME will be
associated with a high level of FE. The problem is that this makes it particularly dif-
ficult to disentangle the effect of ME on y from the effect of FE on y. If our only inter-
est is in the value of β3 then this may not be troublesome, but frequently, we cannot
get out of the dilemma so easily. And because ME and FE really do affect y, the fit of
the equation is likely to be good.
When an equation fits well but the coefficients are not statistically significant, it
is appropriate to suspect that multicolinearity is at work. It is a good idea to look at
the simple correlation coefficients among the various independent variables; if any
of these are (absolutely) greater than about 0.5, then multicolinearity is likely to be
a problem.
In the extreme case in which MEi = γ.FEi exactly, we are unable to measure either
β1 or β2 correctly. Substituting in for MEi gives the following:
yi = β0 + β 1(γ FEi) + β2FEi + β3URBi + εi

= β0 + (β lγ + β 2) FEi + β3URBi + εi.
In this regression, we have left out ME, but the coefficient on the FE variable is
no longer correct. In other words, dropping a variable that is collinear with other
variables does not solve the problem. Indeed there is no easy solution to multicol-
inearity; the best hope is more, or perhaps more accurate, data, but finding such
data is easier said than done. 283
14
Heteroskedasticity
When working with cross-sectional data—the numbers that come from household
surveys—one frequently encounters heteroskedasticity. This is the term used when
the error in the regression model, εi, does not have a constant variance. The situation
is illustrated in figure 14.3; the true relation here is
y = 2 + 0.8 X,
but it is clear from panel A that the relationship fits well at low values of X—the
observations are close to the line—but is more imprecise at higher values of X.
Heteroskedasticity does not bias the estimates of the coefficients, so even if it is
present, the estimates of the slopes will be correct. However, it reduces the efficiency
with which the standard errors of the coefficients (and hence the t-statistics and
p-values) are estimated. Sometimes the problem can be solved with a simple trans-
formation: the picture illustrated in figure 14.3, panel B, shows the same relationship
except that the dependent variable is now the natural log of y, ln(y), rather than y. In
this case, there is no visible evidence of heteroskedasticity, and we can proceed satis-
factorily with testing hypotheses and creating confidence intervals.
Outliers
One last problem is worth noting, which is that of outliers. Quite frequently, a
small number of observations take on values that are far outside the range of what
one would expect. Figure 14.4, panel A, shows a hypothetical data set with a clean-
looking estimated regression line. Panel B shows the same data, except that in the
case of one observation the value of the y variable is 80.2 rather than 8.02. This is
an outlier, and it had a major effect on the fit and form of the estimated equation.
Figure 14.3 Heteroskedasticity Illustrated
284 Source: Authors’ creation.

14
Figure 14.4 Outliers Illustrated
Source: Authors’ creation.
In this case, it is quite likely that the data point was entered incorrectly and
is simply wrong. The process of data cleaning is largely one of tracking down
errors (and missing values) in the original data set. The situation is more prob-
lematic if it appears that the outlier does not represent an error. Some researchers
remove outliers, or apply more robust estimation methods that reduce the influ-
ence of outliers on the regression results. However, it is difficult to justify the
exclusion of data observations just because they happen to be inconvenient; after
all, they convey information, and indeed may be more informative than most of
the other observations.
Solving Estimation Problems
As a general rule, it is not easy to solve the estimation problems outlined above.
However, here are some possibilities:
Get more, or better, data. This is typically expensive or difficult, but is occasionally
possible. More data are useful in dealing with multicolinearity; better data are essen-
tial if measurement error is a problem.
Use fixed effects. Consider the following problem: We would like to estimate the
determinants of rice production (Q), and believe that output will be influenced by
fertilizer use (F), and the inherent ability of the farmer. Thus,
rice output = a + b fertilizer input + c ability + ε. (14.3)
The problem is that ability is unobserved, so we end up estimating
rice output = d + e fertilizer input + w. (14.4) 285

14
In this case the coefficient on fertilizer input will be overstated if the level of
fertilizer used is positively related to one’s ability, as is plausible. The coefficient
would be picking up the effects of both fertilizer use and ability, but attributing these
effects to fertilizer use alone.
But now suppose that we have two observations per individual—for example,
from panel data—so that we have equation (14.3) at two points in time. Using sub-
scripts to denote the two time periods (0 and 1), we may difference equation (14.3)
to get
Q1 – Q0 = b(F1 – F0) + (ε1 – ε0) (14.5)
which will now give us an unbiased estimate of the return to fertilizer inputs. The
“ability” term has been dropped, assuming that ability does not change over time.
More generally, we could include fixed effects, so that we are in effect estimating
Qti = ai + bFti + εi , (14.6)
where t refers to the year and i to the household. Here, each household has a sepa-
rate constant term (the ai), which controls for such factors as differences in ability or
other unobserved differences as dynamism, health, or the like.
Use instrumental variables. The idea here is to estimate equation (14.4) but,
instead of using fertilizer input, use an estimated value of fertilizer input that is
purged of any contamination by the ability variable. Typically there are three steps:
(1) estimate a preliminary equation where fertilizer input is the dependent variable,
and the independent variables constitute of a set of variables that are not correlated
with any of the other variables in the production equation; (2), use this first equa-
tion to generate a set of predicted (“purged”) values for fertilizer input; and (3), esti-
mate the production equation itself using these purged values as instruments for
fertilizer input, rather than the actual values. Most statistical packages have straight-
forward commands for this estimation.
It is important that at least some of the variables in the preliminary equation do
not appear in the main equation of interest, yet are closely correlated with the vari-
able of interest (here, fertilizer input). It is often hard to find good instruments,
which is the main weakness of this approach.
Experiments. Glewwe et al. (2000) wanted to know whether the use of flip charts
affected education achievement. They were able to arrange for charts to be given to
some schools in Kenya and to compare these schools with a control sample. It is not
always easy, however, to design or implement experiments of this nature. An effort
to give gifts of $100 to some, but not all, of the households surveyed in the Vietnam
Living Standards Survey of 1998 to measure the pure income effect on consumption
was not considered appropriate by the Vietnamese authorities; it was seen as invidi-
ous. These are examples of efforts to measure the impact of projects, a topic that was
286 addressed in more detail in chapter 13.
14
Logistic Regression
Often, the research issue of interest is whether a household is poor or not, owns a
car, has another child, or works in industry. These cases all have a point in com-
mon: the dependent variable is binary, taking on values of 1 if the event is true, or
0 otherwise.
Mechanically, one can apply OLS to this situation. But the result is not satisfac-
tory, since the predicted values of the equation might, in some cases, be higher than
1 or lower than 0, which makes little sense. This is illustrated in figure 14.5 for some
hypothetical data. The horizontal axis shows household income per capita, and each
dot is set equal to 1 if the household owns a motorbike, or 0 if it does not own a
motorbike. Our interest is in fitting a curve to these observations. The OLS line is
shown as the dashed line in figure 14.5; for income above 30, it predicts that the
probability of owning a motorbike is greater than 1. The thick line in figure 14.5 is
based on a logistic regression, which is of the form
and calls for some further explanation.

The easiest way to approach this is to look in more detail at a real example,
which comes from Haughton and Haughton (1999). Table 14.1 gives the results of
Figure 14.5 Logistic Regression Compared with OLS
Source: Authors’ creation. 287

14
a logistic model of migration to urban areas in Vietnam, using data from the
Vietnam Living Standards Survey of 1993. At first sight, the output looks similar
to that of a standard regression, and indeed the p-values can be interpreted in the
same way: variables with small p-values are significant, and it is unlikely that their
true coefficients are zero. The difference lies in the interpretation of the coeffi-
cients, as the following discussion will show.
Define the odds of migrating as follows:
The key point is that in a logistic regression, when an independent variable

increases by 1 (for instance, the age of the migrant, AGE), and all other variables are
held constant, the estimated odds of migrating change by the exponential of the
value of the coefficient. In this example, for AGE, we have e0.438 = 1.5496. This
means that one more year of age multiplies the estimated odds of migrating by
1.5496. Since P(migrating) = 1–P(not migrating), we have
For instance, consider a person whose estimated probability of migrating is 0.03

(3 percent). An otherwise identical person who is one year older would then have
Odds(migrating) = 1.5496 × (0.03/0.97) ≈ 0.0479.
So, for this second person, the estimated probability of migrating is

p(migrating) ≈ 0.0479/1.0479 ≈ 0.0457.
Thus, one more year of age, holding all other variables constant, raises an esti-
mated probability of migrating of 3 percent to about 4.57 percent.
The most important variable in the migration model is the one that measures the
differential between the log of expenditure per capita now and the log of expendi-
ture per capita that the person would have experienced had he or she not migrated.
The coefficient of 2.368 means that, when the ratio of current expenditure per capita
to expenditure per capita at place of birth is multiplied by e (= 2.72, a large change),
and all other variables are held constant, an estimated probability of migrating of
3 percent would rise to 24.8 percent.
The coefficients for the regional dummy variables can be interpreted as follows.
Consider a person residing in one of the reference regions (here the Red River
Delta, North Central Coast, Central Highlands, or Mekong Delta) with an esti-
mated probability of migrating of 7 percent. An otherwise similar person residing
in the Northern Uplands would have an estimated probability of migrating of
288 11.6 percent.
14
Table 14.1 Logistic Model of Rural-Urban Migration, Vietnam, 1993
Estimated probability of migrating when

independent variable changes by one
Coefficients P–values unit and initial probability is (percent):
3.0 7.0 11.0
Dependent variable
Does person migrate from
rural to urban area? (Y=1)
Independent variables:
Age of person (years) 0.438 0.000 4.6 10.4 16.1
Years of education 0.117 0.000 3.4 7.8 12.2
Δ ln(expenditure per capita) 2.368 0.000 24.8 44.5 56.9
Male household head (Y=1) –1.498 0.000 0.7 1.6 2.7
Size of household 0.105 0.000 3.3 7.7 12.1
Married (Y=1) 1.700 0.000 14.5 29.2 40.4
Divorced (Y=1) 1.618 0.000 13.5 27.5 38.4
Separated (Y=1) 1.176 0.026 9.1 19.6 28.6
Widowed (Y=1) 1.146 0.000 8.9 19.1 28.0
Regional effects
Northern Uplands 0.558 0.000 5.1 11.6 17.7
Central Coast 0.905 0.000 7.1 15.7 23.4
Southeast 1.404 0.000 11.2 23.5 33.5
Source: Based on data from Vietnam Living Standards Survey 1992–93.

Note: Based on 10,954 observations. Pseudo R2 = 0.34. Effect for expenditure per capita variable in last three columns assumes
a 2.72-fold rise in ratio of expenditure per capita at current residence to estimated expenditure per capita if person had stayed at
place of birth. Omitted regions are Red River Delta, North Central Coast, Central Highlands, and Mekong Delta. Nonmigrants
include those who were born in a rural area and are still living in the location where they were born.
Table 14.1 also gives a value for the pseudo-R2, which can be interpreted in
roughly the same way as the nonadjusted R2 for standard regression. As in the case
of standard regression, one must guard against overfitting, which is the addition of
too many independent variables in the hope of getting a “better” model. Adding
more variables will always increase the pseudo-R2, but sometimes by minute
amounts. The problem is that a model with too many variables might not fit well on
other similar data sets.
The use of logistic regression is not always appropriate. For instance, suppose we
are trying to model the determinants of poverty. Based on a poverty line and expen-
diture per capita data, we have classified every households as either poor (1) or not
poor (0). We certainly could apply logistic regression to this variable, but such a
model is unlikely to be efficient, because it is throwing away information. We not
only know whether a household is poor (the binary variable used in the logistic
regression), but also have information on consumption per capita, so we know how
poor the household is. This is more informative, and it might well be more useful to
estimate a model of the determinants of expenditure. 289
14
Review Questions
1. You have estimated an equation using ordinary least squares regression.
To determine whether the equation fits the data well, the most useful
statistic is:
A. The t-statistic.
° B. The p-value.
° C. R2.
° D. The standard error of the coefficient.
°
2. In logistic regression:
A. The dependent variable is binary.

° B. A coefficient shows the effect of a unit change in the independent variable on
° the log of the odds ratio.
C. There is a loss of information relative to using a continuous variable on the
° left-hand side of the equation.
D. All the other answers are correct.
°
3. Measurement error
A. Always biases the coefficients toward zero.

° B. Biases the coefficients toward zero if the independent variables are not
° measured correctly.
C. Biases the coefficients toward zero if the dependent variable is not measured
° correctly.
D. Is fortunately relatively rare when using micro data from household surveys.
°
4. An equation that seeks to explain whether a child attends school or not

includes, as a right-hand variable, the schooling level of the child’s mother,
but not her ability. This is a case of
A. Attrition bias.
° B. Simultaneity bias.
° C. Sample selectivity bias.
° D. Omitted variable bias.
°
5. In a regression model based on household data, fixed-effects estimation

essentially amounts to including a separate intercept for each household,
and requires panel data.
True
° False
290 °
14
6. In a regression model of the form yi = β 0 + β 1X1i + β 2X2i + εi, the
covariates are
A. The εi.
° B. The Xki.
° C. The β k.
° D. The yi.
°
7. In a regression model of the form yi = β 0 + β 1X1i + β 2X2i + εi, β 1 measures
A. The elasticity of output with respect to the input X1.

° B. The change in yi that occurs when X1 rises, holding other variables constant.
° C. The value of yi for each value of X1i.
° D. The correlation between yi and X1.
°
8. The main problem with multicolinearity is that
A. The equation will be a poor fit.

° B. The estimated coefficients will be biased toward zero.
° C. The error term will not be normally distributed.
° D. The coefficients will be estimated imprecisely.
°
9. Which of the following is true about heteroskedasticity?
A. The variance of the error of the equation is not constant.

° B. The coefficient estimates are unbiased.
° C. Hypothesis testing is inefficient.
° D. All the other answers are correct.
°
Note
1. The (spending/capita)2 term is obtained by squaring spending per capita (which is in

thousands of Vietnamese dong per year) and then dividing by 1,000.
References
Anderson, David R., Dennis J. Sweeney, and Thomas A. Williams. 2008. Essentials of Statistics
for Business and Economics, 5th edition, South-Western College Pub/Cengage Learning.
Behrman, Jere, and Raylynn Oliver. 2000. “Basic Economic Models and Econometric Tools.” In
Designing Household Survey Questionnaires for Developing Countries: Lessons from Ten Years
of LSMS Experience, ed. Margaret Grosh and Paul Glewwe. Washington, DC: World Bank.
Glewwe, Paul, Michael Kremer, Sylvie Moulin, and Eric Zitzewitz. 2000. “Flip Charts in Kenya.”
NBER Working Paper 8018, National Bureau of Economic Research, Cambridge, MA. 291
14
Greene, William H. 2007. Econometric Analysis, 6th edition. Upper Saddle River, NJ: Prentice
Hall.
Haughton, Dominique, and Jonathan Haughton. 1999. “Statistical Techniques for the Analysis
of Household Data.” In Health and Wealth in Vietnam, ed. Dominique Haughton, Jonathan
Haughton, Sarah Bales, Truong Thi Kim Chuyen, and Nguyen Nguyet Nga. Singapore:
ISEAS Publications.
Prescott, Nicholas, and Menno Pradhan. 1997. “A Poverty Profile of Cambodia.” Discussion
Paper No. 373, World Bank, Washington, DC.
Wooldridge, Jeffrey. 2008. Introductory Econometrics: A Modern Approach, 4th edition. South-
Western College Pub/Cengage Learning.
292
Journal of Development Economics 68 (2002) 381 – 400
www.elsevier.com/locate/econbase
Why has economic growth been more pro-poor in

some states of India than others?
Martin Ravallion*, Gaurav Datt
World Bank, 1818 H Street NW, Washington, DC 20433, USA
Received 1 August 1999; accepted 1 July 2001
Abstract
We use 20 household surveys for India’s 15 major states spanning 1960 – 1994 to study how the
sectoral composition of economic growth and initial conditions interact to influence how much
growth reduced consumption poverty. The elasticities of measured poverty to farm yields and
development spending did not differ significantly across states. But the elasticities of poverty to
(urban and rural) non-farm output varied appreciably, and the differences were quantitatively
important to the overall rate of poverty reduction. States with higher elasticities did not experience
higher rates of non-farm growth. The non-farm growth process was more pro-poor in states with
initially higher literacy, higher farm productivity, higher rural living standards (relative to urban
areas), lower landlessness and lower infant mortality. D 2002 Elsevier Science B.V. All rights
reserved.
JEL classification: I32; O15; O40

Keywords: Poverty; Inequality; Economic growth; Rural development; Human development
1. Introduction
While cross-country comparisons indicate that measures of absolute poverty in

developing countries tend to fall with economic growth, there is considerable variance
in the poverty-reducing impact of a given rate of growth. For example, on studying
successive household surveys for a set of developing countries, Ravallion and Chen
(1997, Table 6) obtain a point estimate of (minus) three for the elasticity of the
*
Corresponding author. Tel.: +1-202-473-685; fax: +1-202-522-1153.
E-mail addresses: mravallion@worldbank.org (M. Ravallion), gdatt@worldbank.org (G. Datt).
0304-3878/02/$ - see front matter D 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 3 0 4 - 3 8 7 8 ( 0 2 ) 0 0 0 1 8 - 4
382 M. Ravallion, G. Datt / Journal of Development Economics 68 (2002) 381–400
proportion of the population living below US$1/day (at 1985 Purchasing Power Parity)
to the survey mean income or consumption. Their 95% confidence interval, however,
goes from about 1 to 5. On further factoring in the variance in how much GDP growth
translates into growth in average household incomes (Ravallion, 2001), one sees why a
rate of growth that brings rapid poverty reduction in one country can largely leave the
poor behind in another.
Why is growth more pro-poor in some economies than others? Answering this question
with cross-country data raises many problems, including weak comparability of the
primary data used for poverty measurement and the explanatory variables. This paper
pursues an alternative route where we compare the evolution of poverty measures across
major states of India, for which it is possible to construct a long time series of 20
reasonably comparable surveys spanning a 35-year period. We test for inter-state differ-
ences in the poverty impact of various sources of growth in India, and we try to explain the
differences in impact.
There are clearly countrywide factors that have influenced growth and poverty
reduction in India, and comparisons of experiences over time in different states cannot
throw much light on those factors. However, the diverse experiences of these states in
poverty reduction can illuminate one important question: are these diverse experiences
mainly due to differences in the rate and sectoral pattern of economic growth, or are there
important differences in the poverty-reducing impact of that growth? Conditional on
finding the latter factor to be important, we further explore the extent to which differences
in initial conditions at state level account for differing poverty-reducing impacts of
economic growth.
We build on our past work on poverty in India. Consistent with cross-country evidence
for developing countries, measures of absolute consumption-poverty in India tend to fall
with economic growth (Ravallion and Datt, 1996). But one also finds that:
(i) the sectoral composition of growth matters to the aggregate (country-wide) rate of
poverty reduction: the aggregate time series data for India indicate that poverty measures
have responded more to rural economic growth than urban economic growth (Ravallion
and Datt, 1996),1 and
(ii) differences in initial conditions related to rural development and human resource
development accounted for a sizable share of the long-run differences between states in
rates of rural poverty reduction (Datt and Ravallion, 1998a).
However, initial conditions entered additively with growth in Datt and Ravallion
(1998a); favorable conditions meant a higher rate of poverty reduction (by a fixed mark-
up) at any rate of growth, but a given growth rate was deemed to have the same impact on
the rate of poverty reduction whatever the initial conditions. Clearly, this type of
1
‘‘Rural (urban) economic growth’’ refers to growth in mean consumption in rural (urban) areas; Ravallion
and Datt (1996) also find that ‘‘primary’’ and ‘‘tertiary’’ sector growth had greater impact on poverty than
‘‘secondary’’ sector growth. Thorbecke and Hong-Sang (1996) come to a similar conclusion for Indonesia, using a
different method based on simulations with a Social Accounting Matrix. There is a large literature, and much
debate, on the role of agricultural growth in poverty reduction (Lipton and Ravallion, 1995, Section 5.2; Datt and
Ravallion, 1998b).
M. Ravallion, G. Datt / Journal of Development Economics 68 (2002) 381–400 383
specification cannot tell us how initial conditions might influence the impact on poverty of
economic growth and how that impact might depend on the sectoral composition of
economic growth.
Here we test an empirical model of the evolution of the aggregate (urban plus rural)
poverty measures at state level that allows for multiplicative interactions of the sectoral
composition of growth with initial conditions in determining the evolution of state-level
poverty measures. In addition to helping us understand what makes growth in a given
sector of the economy more pro-poor, we will be able to test for possible trade offs; for
example, do certain conditions foster a more pro-poor agricultural growth process, but a
less pro-poor non-farm growth process?
The next section discusses the arguments that have been made in the past as to why
existing inequalities in various dimensions matter to the prospects for poverty-reducing
economic growth. We do not attempt to unify these arguments in a formal model, but we
do draw out some implications for the measurable factors that might be expected to
influence how much impact economic growth has on poverty. Turning to our data for
India, we then establish that there are significant differences between states of India in the
extent to which the poor have benefited from (urban and rural) non-farm economic
growth. We then test for interaction effects between non-farm growth and initial
conditions.
2. What makes growth pro-poor? Arguments from the literature
It is widely agreed that economic growth is not sufficient for poverty reduction. A
number of other factors influence whether the growth is more or less poverty reducing.
The issue here is whether the extra things needed enter additively or multiplicatively. Is it a
matter of doing as many things as possible from a list of poverty-reducing actions, with
extra impact as each one is ticked off? Or are there important interaction effects, such that
only certain combinations do the trick? And what are the key combinations?
It is evident that the distribution of consumption matters to the level of consumption
poverty at a given level of mean consumption. But does initial distribution also matter to
the subsequent rate of poverty reduction? One way it could matter is through the rate of
growth in the mean. A long-standing view — though for long periods a minority view
amongst economists — has held that inequality can be harmful to the pace of economic
growth in poor countries. For example, in the 1920s and 1930s, Gunnar Myrdal believed
that ‘‘. . .an equalization in favor of the low-income strata was also a productive investment
in the quality of people and their productivity’’ (Myrdal, 1988, p. 154). A number of
arguments have been made as to why high inequality can impede growth (Aghion et al.,
1999). A plausible argument (in this context) is that credit market failures mean that the
poor are unable to exploit growth-promoting opportunities for investment in (physical and
human) capital. The higher the proportion of poor (and hence credit-constrained) people in
the economy, the lower the rate of growth.2
2
Versions of this argument can be found in Stiglitz (1969), Loury (1981), Binswanger et al. (1995), Benabou
(1996) and Aghion et al. (1999), amongst others.
In addition to its implications for the rate of growth, high initial inequality is a plausible
candidate in explaining why the same rate of economic growth might be less effective in
reducing poverty in one setting than another. Quite generally, the elasticity of any measure
of poverty to change in the mean of the distribution on which poverty is measured will
depend on other properties of that distribution. These effects are hard to characterize in
general terms. However, in an economy where inequality is persistently low, one can
expect that the poor will tend to obtain a higher share of the gains from growth than in an
economy in which inequality is high. There is supportive evidence from cross-country
distributional data that higher initial income inequality entails a lower (absolute) elasticity
of poverty to average incomes (Ravallion, 1997; Timmer, 1997; World Bank, 2000). For
example, a country with a Gini index of 0.25 can expect an elasticity of the headcount
index to mean household income of around 3.3, while for a country with a Gini index of
0.60, the elasticity is 1.8 (Ravallion, 1997). But what specific aspects of ‘‘inequality’’
are likely to matter? The Gini index for incomes can be thought of as a product of various
dimensions of inequality. Some inequalities may matter more than others to how much the
poor share in economic growth.
Asset distribution is likely to influence the extent to which poor people participate in
economic growth. Indeed, the credit-market failure argument discussed above as to why
initial distribution matters to the rate of growth suggests that it will be the asset poor who
are most locked out of growth prospects. Greater initial asset poverty will then mean that
the growth that does occur is less (income) poverty reducing. Land is clearly an important
asset in this context, so we might expect greater landlessness to entail that the poor share
less in the gains from economic growth.
Human capital is possibly no less important. Low basic education attainments are often
identified as a source of income inequality.3 Education will also influence how much the
poor are equipped to participate in (relative to farming) skill-demanding non-farm growth.
This too is not a new observation. In the 1950s, Jacob Viner wrote that ‘‘The first
requirements of high labor productivity under modern conditions are that the masses of
the population shall be literate, healthy, and sufficiently well fed to be strong and energetic’’
(Viner, 1953, p. 100). More recently, the view that there are important synergies between
human resource development and growth-oriented policy reforms has been a prominent
theme in writings on development; examples include Drèze and Sen (1995) writing on India,
and Thorbecke and Hong-Sang (1996) on Indonesia. The World Bank’s approach to poverty
reduction has also emphasized the importance of combining human resource development
with policies promoting economic growth (World Bank, 1990, 2000; Bruno et al., 1998).
Another potentially important factor in developing countries is the extent of the income
disparities between urban and rural sectors. The existence of earnings and other income
disparities between urban and rural sectors is clearly an important dimension of overall
inequality in developing countries (Fields, 1980; Bourguignon and Morrison, 1998, who
provide supporting evidence from cross-country comparisons). Ravallion and Datt (1999)
outline a simple dual economy model with this feature. Poverty reduction in this model
3
See, for example, the discussions in Tinbergen (1975) and Atkinson (1997). Evidence from cross-country
comparisons of income-inequality reducing effects of average education attainments can be found in Li et al.
(1998).
takes the form of absorption of poor farm-sector workers into the non-poor non-farm
sector. The model assumes that any farm worker who wants to participate in the non-farm
sector incurs a cost in doing so. This cost determines the equilibrium earnings differential
between the farm sector and the non-farm sector. It is evident that such a cost lowers
overall output. But this cost also reduces labor absorption into the non-farm sector, thus
implying a higher poverty rate. Moreover, the intersectoral wage gap also makes output
gains less pro-poor. The higher the initial wage gap (in turn implying a higher initial
poverty rate), the lower the rate of poverty reduction from a given rate of non-farm
economic growth. Arguably, dualism also limits prospects for pro-poor growth. There is a
long-standing view (though not a dominant one it would seem) that rural underdevelop-
ment constrains prospects for industrialization; see, for example, Clarke (1940). Again
factor market distortions also entail that urban – rural inequality impedes poverty reduction
through non-farm economic growth.
Consider, for example, the classic model of Harris and Todaro (1970) in which wages
in the non-farm sector are fixed above market clearing levels. While there is mobility
between the urban and rural sectors, rural workers who move to the city will not all be able
to get the new jobs, and so they will face unemployment, or turn to relatively low-paid
urban informal sector activities. Greater labor market dualism (as measured by the
intersectoral wage differential) means that there will be less growth, and that less of the
growth that does occur will benefit the poor.
The initial population distribution between urban and rural sectors can also be expected
to matter to the impact on poverty of economic growth. In general, a change in the share of
the population living in urban areas will shift the Lorenz curve in a dual economy. The
direction of that effect is, however, theoretically ambiguous. For example, under the
Kuznets Hypothesis, inequality will be low at both low and high levels of urbanization
(Robinson, 1976; Fields, 1980; Anand and Kanbur, 1993).
Urbanization is often viewed as a positive factor in promoting rural non-farm economic
growth, by expanding markets. Schultz (1953) noticed that rural non-farm activities tend
to be more developed in the periphery of urban-industrial centers, and this has been
confirmed in many countries. Enterprises are probably attracted to urban areas because of
the larger local product markets, the availability of a skilled workforce, the wider variety
of production inputs, the possibility of technological spillovers, and better infrastructure
(Lanjouw and Lanjouw, 1997).
It can be argued that these same factors will also matter to the impact on poverty of
growth in mean output. It is plausible that the poor will tend to be more constrained in their
access to markets and infrastructure than the non-poor, and that the poor will tend to gain
more from relaxing those constraints than do the non-poor. Assuming that the level of
urbanization in an area reflects these differences in access to markets and infrastructure,
one can expect that (other things being equal) the poor will be able to benefit more from
non-farm growth when they live in a more urbanized area. On the other hand, it is also
arguable that greater initial level of urbanization may lead to further concentration of non-
farm growth in the urban centers that are already better-off, and hence may dampen the
poverty alleviation impact of that growth (relative to a situation where that growth is more
evenly spread). The overall effect of initial urbanization on the elasticity of poverty to
mean output could thus be positive or negative.
Another factor influencing the impact on poverty of non-farm economic growth is the
productivity of the main competing sector for workers, namely farming. For example, by
allowing multiple cropping, irrigation and the spread of high-yielding varieties will
probably increase aggregate demand for agricultural labor, thus bidding up wages for
new entrants into an expanding non-farm sector.
Motivated by these diverse arguments from the literature, the rest of this paper tries to
see what the experience of India’s states suggests about the factors influencing how much
impact economic growth has on poverty.
3. Econometric model and results
There is a large literature studying the evolution of India’s rural poverty measures over
time, following (and debating) the seminal contribution of Ahluwalia (1978); Datt and
Ravallion (1998b) survey this literature. In addition to updating and revising the data, our
main points of departure are that: (i) we model the aggregate (urban and rural) state-wide
poverty measures and (ii) we relax the almost universal assumption in past work that the
poverty-reducing impact of growth is the same across all states.4 Thus, we allow for state-
specific elasticities of poverty, which vary with initial conditions. We deliberately condition
out interstate differences in the levels of poverty, by including state fixed effects. We also
allow for state-specific time trends as well as state effects in other time-varying factors that
could well bias our results if they were omitted. With this specification, we can focus on
whether changes over time in aggregate farm and non-farm outputs had different impacts on
poverty in different states and (if so) whether observed differences in initial conditions can
account for the heterogeneity in impacts.
If there are no significant inter-state differences in the elasticities of poverty, then we
will not have much to explain. So we first test for such differences.
3.1. Testing for inter-state differences in the elasticities of poverty
We have estimated various measures of absolute consumption-poverty for each of

India’s 15 main states using 20 rounds of the National Sample Survey (NSS) spanning the
period 1960 – 1961 to 1993 – 1994 at intervals of 0.9– 5.5 years. We will be concerned with
measures of absolute poverty, by which we mean that the poverty line is kept fixed in real
terms (or in terms of the standard of living it commands) over the entire (spatial and
temporal) domain of poverty measurement (Ravallion, 1994). The standard of living is
measured by real per capita consumption. We construct three different measures of poverty
within the Foster et al. (1984) class of measures: the headcount index (H), the poverty gap
index (PG) and the squared poverty gap index (SPG). We have collated the survey data
with data on farm yields, non-farm output, government spending, and variables describing
initial conditions. Appendix A provides more details.
4
The only exception appears to be Van de Walle (1985) who relaxes the pooling restrictions in the Ahluwalia
(1978) model to allow the elasticities of rural poverty to agricultural output to vary between states. No attempt is
made to explain the revealed differences.
Table 1
Trend rates of poverty reduction by state, 1960 – 1994
Trend rates of poverty reduction (percentage per annum)
Headcount Poverty gap Squared poverty
index index gap index
Andhra Pradesh 2.179 3.371 4.295
Assam 0.060 0.054 0.025
Bihar 0.107 1.027 1.797
Gujarat 1.568 2.744 3.619
Karnataka 1.114 1.694 2.159
Kerala 2.733 4.447 5.675
Madhya Pradesh 0.633 1.412 2.070
Maharashtra 1.013 1.522 1.887
Orissa 1.586 2.712 3.697
Punjab and Haryana 2.547 3.746 4.679
Rajasthan 1.154 1.883 2.423
Tamil Nadu 1.508 2.315 2.930
Uttar Pradesh 0.876 1.531 2.115
West Bengal 1.965 3.073 4.015
Jammu and Kashmir 1.023 1.382 1.635
Trends calculated as the OLS regression coefficients of logarithms on time.
Table 1 gives the trend rates of poverty reduction by state. There is a wide range, from a
trend rate of reduction in the headcount index in Kerala of 2.7% per year (with Punjab –
Haryana close behind) to trends close to zero for Assam and Bihar.5 We will attempt to
explain these differences below.
In testing for inter-state differences in the elasticities of poverty, the natural starting
point is an econometric specification in which the log of the poverty measure is regressed
on the log of mean income, allowing the regression coefficient to vary across states. We
want to extend this specification to allow for differences in the elasticities with respect to
different sectoral components of aggregate income, and for inflationary shocks. Since we
are interested in describing (and later modeling) the elasticities of poverty with respect to
its determinants rather than levels of poverty, we also control for differences between
states in the initial level of poverty. (So when we later try to explain the inter-state
differences in elasticities, we will not be confusing this question with that of the effects of
initial conditions on the initial level of poverty.) This is done by including a complete set
of state dummy variables in all regressions.6 As usual in fixed effects regressions, this also
means that our results are robust to any correlation between the explanatory variables and
the time-invariant error component. For example, we need not be concerned with bias
arising from the endogeneity of rates of economic growth to any latent time-invariant
factors at state level that also influence subsequent progress in poverty reduction.
5
To help put these numbers in perspective, the results of Chen and Ravallion (in press) indicate that the
percentage of the population of the developing world living below US$1 per day (at 1993 Purchasing Power
Parity) fell at a compound annual rate of 1.7% over 1987 – 1998 (1.0% excluding China).
6
Alternatively, one can think of our regressions as models of the deviations from time mean or the rates of
poverty reduction (difference in logs) as functions of the similarly transformed (state- and time-specific)
explanatory variables.
However, bias will remain due to any correlations between the explanatory variables and
time-varying omitted variables. To allow for time-trended omitted variables, we include a
state-specific trend.
Combining these features, our test equation takes the form:
lnPit ¼ bNFP
i lnNFPit þ bYLD
i lnYLDit þ bGOV
i lnGOVit þ ci INFit þ pi t þ gi þ eit ð1Þ
where Pit is the measure of absolute consumption poverty (on a per capita basis) in state i
( = 1,. . .,N ) at date t ( = 1,. . .,T ), NFPit is real non-farm product per head of the population
in state i at date t, YLD is farm yield (output per hectare),7 GOV is real state development
expenditure per capita, and INF is the inflation rate. (All these variables are defined more
precisely in Appendix A.) Consistently with Datt and Ravallion (1998a), we found that the
fit was improved using the 2-year moving averages of lnYLD and lnNFP, and the lagged
value of lnGOV, which also addresses a possible endogeneity concern about current
spending. We initially assumed that eit was an AR1 error term, allowing for the uneven
spacing of the surveys (following Datt and Ravallion, 1998a). However, the autoregression
coefficient was not significantly different from zero so we set it to zero to simplify the
estimation method.
To be as flexible as possible, we initially write the bi’s as linear functions of a vector of
state dummy variables. Since we have state fixed effects and state-specific time trends as
well as differing effects of inflation, estimating Eq. (1) is equivalent to running a separate
regression for each state.
However, we found that a degree of pooling was consistent with the data. In particular,
we could not reject the null hypothesis of constant coefficients at the 10% level for all
variables except non-farm output per person and the state effects in the intercept.8 We could,
however, reject the null hypothesis that the coefficients on NFP are the same across states.
Thus, we impose a constant-coefficients restriction for YLD, GOV, INF and the time
trend, leaving the coefficient on NFP free to vary between states and retaining the state
fixed effects in the intercepts. An implication of our finding that only the non-farm
elasticities vary significantly across states is immediate: we can reject the idea of significant
trade-offs; it is not the case that states with higher elasticities with respect to non-farm
growth tended to have lower elasticities to agricultural growth or development spending.
Table 2 gives the estimated parameters of the restricted model. Higher farm yields and
higher development spending reduce all three poverty measures, and the coefficients are
highly significant. Higher non-farm output per person lowers poverty in all states. Inflation
is poverty increasing. The significance of inflation confirms the results of Datt and
Ravallion (1997, 1998a,b). In Datt and Ravallion (1998b), we argued that the main
7
In past work on these data (Datt and Ravallion, 1998a,b), we have found that farm output per hectare is a
better predictor of poverty than output per person; on decomposing log output per person into output per hectare
and hectares per person, the latter is insignificant. Output per hectare is probably the better measure of farm
productivity (for further discussion, see Datt and Ravallion, 1998b).
8
The failure to reject the null hypothesis of constant coefficients for all except NFP was statistically
convincing; probabilities for the tests were no lower than 0.15, and most were above 0.25.
Table 2
Regressions for the state poverty measures allowing for inter-state differences in elasticities to non-farm output
Variable Headcount index Poverty gap index Squared poverty gap index
Coefficient t-Ratio Coefficient t-Ratio Coefficient t-Ratio
Real agricultural output per 0.110 4.74 0.201 5.46 0.271 5.35
hectare of net sown area
(current + lagged) (YLD)
Real per capita state 0.140 2.57 0.241 2.79 0.338 2.86
development expenditure
(lagged) (GOV)
Real non-agricultural output
per person (current + lagged)
(NFP)
Andhra Pradesh 0.291 8.89 0.425 8.19 0.524 7.37
Assam 0.199 5.05 0.259 4.13 0.314 3.65
Bihar 0.130 2.59 0.335 4.21 0.501 4.58
Gujarat 0.285 6.93 0.444 6.81 0.550 6.14
Karnataka 0.249 7.06 0.360 6.42 0.444 5.77
Kerala 0.542 14.80 0.859 14.79 1.087 13.64
Madhya Pradesh 0.184 4.92 0.318 5.35 0.421 5.16
Maharashtra 0.191 5.04 0.248 4.13 0.270 3.27
Orissa 0.330 9.67 0.531 9.80 0.700 9.42
Punjab and Haryana 0.343 10.09 0.466 8.65 0.554 7.49
Rajasthan 0.336 7.39 0.493 6.84 0.605 6.11
Tamil Nadu 0.277 7.97 0.397 7.20 0.479 6.33
Uttar Pradesh 0.253 6.12 0.359 5.47 0.444 4.93
West Bengal 0.618 11.57 0.937 11.06 1.204 10.35
Jammu and Kashmir 0.176 5.12 0.230 4.21 0.273 3.65
Inflation rate (INF) 0.419 5.19 0.587 4.58 0.704 4.00
Time trend 0.017 6.46 0.027 6.51 0.036 6.21
Root mean square error 0.0940 0.1491 0.2047

R2 0.918 0.918 0.910
Test for equality of non-farm 14.39 14.28 12.94
output elasticities across all (0.00) (0.00) (0.00)
states: F(14, 238) ( p-value)
All variables are measured in natural logarithms. The dependent variables are log poverty measures. A positive
(negative) sign indicates that the variable contributes to an increase (decrease) in the poverty measure. The
estimated model also included state-specific intercept effects, not reported in the table. The number of
observations used in the estimation is 272.
channel through which inflation mattered to India’s poor was through its short-term
adverse effect on the real wage rate for unskilled labor. We also find a significant positive
independent time trend. One possible interpretation is that it reflects an adverse distribu-
tional effect of population on poverty (Van de Walle, 1985). This is consistent with the fact
that our time trend became insignificant when we introduced log population as an
additional variable in our model.
Fig. 1 gives the estimated (absolute) elasticities of poverty to non-farm output. (Notice
that the elasticities are twice the bi estimate because lnNFP enters as the sum of the current
and the lagged values.) For the headcount index, the absolute values of the implied non-
Fig. 1. Elasticities of poverty to non-farm output. Absolute values of the elasticities by states of India implied by
Table 2 for the headcount index (H), poverty gap index (PG) and squared poverty gap index (SPG). States ranked
by elasticity for SPG. 1 = West Bengal; 2 = Kerala; 3 = Orissa; 4 = Rajasthan; 5 = Gujarat; 6 = Punjab and Haryana;
7 = Andhra Pradesh; 8 = Bihar; 9 = Tamil Nadu; 10 = Uttar Pradesh; 11 = Karnataka; 12 = Madhya Pradesh;
13 = Assam; 14 = Jammu and Kashmir; 15 = Maharashtra.
farm growth elasticities vary from a low of 0.26 in Bihar to a high of 1.08 in Kerala and
1.24 in West Bengal. For all states, the elasticities are higher (in absolute value) for PG
than H, and higher still for SPG. This implies inequality-reducing gains below the poverty
line. For SPG, the lowest elasticity is for Jammu and Kashmir (0.46) and the highest is for
West Bengal (2.41, though with Kerala close behind at 2.18).
To quantify the importance of the elasticity differences to the overall rate of poverty
reduction, we simulated rates of poverty reduction, in which we artificially set the non-
farm output elasticities of all states to a reference value (denoted bNFP*) set alternately at
the lowest and highest elasticities across all states:
dlnPi dlnPi dlnNFPi

¼ þ 2ðbNFP bNFP
i Þ ð2Þ
dt dt dt
where (dlnPi/dt) and (dlnNFPi/dt) are the trend rates of poverty reduction and non-farm
output per capita, respectively. This calculation assumes that the changes in elasticities leave
other variables in the model unchanged. That is questionable. Consider the simulations
when all states are given the highest elasticity of any state. If the more favorable initial
conditions for the elasticities would have also led to higher (lower) growth rates, then these
simulations will have under (over) estimated the gains to rates of poverty reduction.
Table 3 compares the national average trend rate of poverty reduction with the
simulated rates of poverty reduction using Eq. (2). Recall from Table 2 that the lowest
(absolute) elasticities are for Bihar (for the headcount index) and Jammu and Kashmir (for
the other two poverty measures). If all states had Bihar’s low elasticity, then the mean rate
of poverty reduction would have been only 0.3% per year for the headcount index (versus
the observed mean of 1.3%). Under the Jammu and Kashmir elasticities, the average
annual rate of decline would have been 0.9% and 1.1% for PG and SPG, respectively
(versus actual means of 2.2% and 2.9%).
At the other extreme, if all states had West Bengal’s elasticity, then the headcount index
would have fallen at a trend rate of 3.5% instead of 1.3% per year. The trend rate of
reduction in SPG would have been 7.2% per year instead of 2.9%.
Was there more growth in the non-farm economy in states where that growth would
have greater impact on poverty nationally? To determine the impact on poverty in India
of the same rate of growth in two different states, one must weight the elasticities in Table
2 by the shares of national poverty in those states. These share-weighted elasticities give
the impact of non-farm growth in any state on the national rate of poverty reduction.
Table 4 gives these calculations, and also the trend rates of growth in non-farm output per
person.
It can be seen that the impact on national poverty of non-farm growth has varied greatly
across states. A 1% growth rate of non-farm output in West Bengal brings down the
national poverty rate by 0.078% (reflecting both that state’s high elasticity of poverty to
growth and its high population share). At the other extreme, the same growth rate in
Jammu and Kashmir (with both a low elasticity and low share of poverty) would have
negligible effect. The differences are even more pronounced using the squared poverty gap
(which is sensitive to the severity of poverty, as well as its depth and incidence). Then we
find that non-farm growth in Kerala has the highest impact on national poverty — many
times that for most other states. Thus, the geographic distribution of growth in the non-
farm economy matters to the overall rate of poverty reduction in India.
We can also see from Table 4 that the trend rates of non-farm growth have not tended to
be higher in states with higher (weighted) elasticities of poverty. Indeed, it is notable that
Table 3
Actual and simulated mean trend rates of poverty reduction across states
Trend rates of change Headcount Poverty gap Squared poverty
in poverty measure index index gap index
(percentage/year)
Actual mean across all states 1.330 2.187 2.865
Simulated mean with lowest 0.287 0.865 1.130
elasticity amongst all states
Simulated mean with highest 3.495 5.514 7.245
elasticity amongst all states
The first row shows the unweighted mean of the actual rates of poverty reduction across states. The second and
third rows show unweighted means across states of the simulated rates of poverty reduction, evaluated using state-
specific trend growth rates of non-farm output per person (based on regressions of log NFP on time) and the lowest
and highest elasticities of poverty (w.r.t. non-farm output growth) across all states.
Table 4
Impacts of non-farm economic growth by state on national poverty
Trend growth rate in non-farm Impact on the national rate of poverty reduction
output per person, 1960 – 1994 of a 1% annual growth rate in non-farm output per
(percentage per annum) person (percentage per annum)
Headcount Poverty gap Squared poverty
index index gap index
Andhra Pradesh 4.166 0.060 0.094 0.120
Assam 3.148 0.008 0.006 0.005
Bihar 1.961 0.030 0.078 0.115
Gujarat 3.350 0.029 0.044 0.054
Karnataka 3.657 0.026 0.037 0.044
Kerala 3.418 0.055 0.107 0.154
Madhya Pradesh 3.294 0.029 0.049 0.065
Maharashtra 3.512 0.039 0.053 0.059
Orissa 3.151 0.030 0.050 0.069
Punjab and Haryana 4.367 0.020 0.023 0.024
Rajasthan 2.414 0.036 0.057 0.075
Tamil Nadu 3.883 0.051 0.081 0.108
Uttar Pradesh 2.917 0.071 0.092 0.106
West Bengal 2.058 0.078 0.091 0.094
Jammu and Kashmir 3.971 0.002 0.002 0.002
Total 3.285 0.563 0.865 1.093

The trend rates of growth are the regression coefficients of log real non-agricultural state domestic product per
person on time. The right panel gives the non-farm output elasticity of poverty (from Table 2) times the initial
share of poverty. The totals give the rate of poverty reduction nationally if non-farm output per capita in all states
grew at 1% per annum.
two of the three states with the highest (most negative) share-weighted growth elasticities
of the headcount index (West Bengal, Uttar Pradesh and Kerala) had below average trend
rates of non-farm growth. And similarly, two of the three states with the highest growth
rates (Punjab and Haryana, and Jammu and Kashmir; the third is Andhra Pradesh) had
below average weighted elasticities. Overall, there is no significant correlation.
So it is clear that, in the longer term, growth in India’s non-farm economy has not been
concentrated in states where it would have the most impact on poverty nationally. In short,
the geographic pattern of non-farm economic growth has not been pro-poor.
Next we try to explain these differences in the poverty impact of non-farm growth.
3.2. Initial conditions and the non-farm output elasticities of poverty
We now postulate that the elasticities of poverty to non-farm output depend on initial
conditions. Motivated by the discussion in Section 2, we let biNFP depend on the values
around the beginning of the period of NFP, YLD, the urban population share (URB), the
ratio of urban to rural average consumption (CDIF), and the share of the rural population
that is landless (LLESS) in the state, the state’s infant mortality rate (IMR) and the literacy
rate; we use the female literacy rate (FLIT) following our previous work (Datt and
Ravallion, 1998a), though it makes little difference if one uses the male rate or the average.
Table 5
Initial conditions around 1960
Female Urbanization Urban – rural Landlessness Infant Agricultural Non-farm
literacy rate (%) mean (percentage mortality output per product
rate (%) consumption rural rate hectare per person
ratio households) (per ‘000)
Andhra 12.0 17.4 1.24 6.8 96.2 19.3 1.06
Pradesh
Assam 16.0 7.2 1.25 27.8 72.1 52.0 1.60
Bihar 6.9 8.4 1.09 8.6 91.7 30.3 0.91
Gujarat 19.1 25.8 1.10 14.7 70.4 6.3 2.60
Karnataka 14.2 22.3 1.01 18.6 91.0 12.8 1.32
Kerala 38.9 15.1 1.19 30.9 67.3 96.4 1.02
Madhya 6.8 14.3 1.14 9.1 130.1 8.0 1.10
Pradesh
Maharashtra 16.7 28.2 1.46 16.0 95.4 8.0 4.45
Orissa 8.6 6.3 1.01 7.8 95.3 18.3 1.10
Punjab and 14.1 20.7 0.97 12.3 87.3 16.7 1.50
Haryana
Rajasthan 5.9 16.3 0.96 11.8 119.4 3.6 0.97
Tamil Nadu 18.2 26.7 1.47 24.2 98.7 51.0 1.80
Uttar Pradesh 7.1 12.9 0.94 2.8 179.2 41.9 1.19
West Bengal 16.9 24.4 1.46 12.6 64.4 76.0 4.86
Jammu and 4.3 16.6 1.08 10.9 64.8 47.3 0.86
Kashmir
The units of initial farm yield are Rs. ‘000 per hectare at October 1973 – June 1974 all-India rural prices, and those
of initial non-farm product are Rs. ‘000 per person also at October 1973 – June 1974 all-India rural prices. See
Appendix A for further details on data and sources.
Table 5 gives the data we will use on these initial conditions.9 Table 6 gives the results
when we replace the state dummy variables in the sub-function for biNFP by the variables
described above. All of these variables are entered in log form.10
We find that non-farm growth is more pro-poor in states with higher initial farm yields,
higher female literacy rates, lower infant mortality, lower urban – rural disparities in
consumption levels and lower initial landlessness. Controlling for these variables, we
do not find the initial urbanization rates or the initial non-farm product to exert a
significant influence on the non-farm output elasticity of poverty. The restriction that
the effects of these two factors is jointly insignificant is easily accepted statistically. Table
6 also gives a restricted form of the model imposing this joint restriction.
On comparing the R2 values of Tables 2 and 6, it can also be seen that the variables we
have used in explaining the inter-state differences in the non-farm output elasticities of
poverty account for a large share of the variance. For example, with full state dummy
9
Note that we do not include the economic and human resource development indicators in time-varying form
as additional explanatory variables in our model for two reasons. First, there are gaps in the available time series
data on these variables over the period covered by our analysis. But, even if a complete time series were available,
these indicators in time-varying form would be arguably endogenous to the model.
10
Notice that there is also a (positive) intercept coefficient in the effect of non-farm output on poverty; this is
the elasticity when all initial conditions are set at zero.
394
Table 6
Explaining inter-state differences in the elasticity of poverty to non-farm output
Headcount index Poverty gap index Squared poverty gap index
M. Ravallion, G. Datt / Journal of Development Economics 68 (2002) 381–400

Coefficient t-Ratio Coefficient t-Ratio Coefficient t-Ratio Coeffient t-Ratio Coeffient t-Ratio Coeffient t-Ratio
Real agricultural output per hectare 0.103 3.92 0.105 4.04 0.192 4.57 0.186 4.47 0.258 4.52 0.246 4.34
of net sown area: current +
lagged (YLD)
Real per capita state development 0.046 0.78 0.056 0.97 0.156 1.64 0.137 1.48 0.261 2.02 0.215 1.70
expenditure: lagged (GOV)
Real non-agricultural output per person: 0.160 0.30 0.192 0.58 0.209 0.24 0.292 0.55 0.375 0.32 0.358 0.49
current + lagged (NFP)
NFP * initial female literacy rate (FLIT) 0.149 6.01 0.153 6.89 0.251 6.35 0.233 6.58 0.320 5.95 0.285 5.89
NFP * initial urban – rural population 0.020 0.77 – – 0.021 0.52 – – 0.067 1.19 – –
ratio (URB)
NFP * initial urban – rural mean 0.183 2.09 0.166 2.33 0.196 1.40 0.277 2.44 0.216 1.14 0.371 2.39
consumption disparity (CDIF)
NFP * initial percentage of rural 0.074 2.79 0.072 2.81 0.120 2.85 0.115 2.82 0.157 2.74 0.153 2.74
landless households (LLESS)
NFP * initial infant mortality rate (IMR) 0.101 2.14 0.101 2.17 0.173 2.30 0.162 2.20 0.233 2.28 0.216 2.14
NFP * initial yield per hectare (YPH) 0.029 2.27 0.027 2.37 0.032 1.54 0.042 2.37 0.038 1.36 0.060 2.44
NFP * initial per capita 0.001 0.02 – – 0.036 0.79 – – 0.055 0.89 – –
non-agricultural output (NFP)
Inflation rate (INF) 0.372 4.00 0.377 4.07 0.578 3.89 0.565 3.83 0.725 3.59 0.698 3.46
Time trend 0.010 3.49 0.011 3.72 0.018 3.78 0.017 3.70 0.024 3.76 0.022 3.52
Root mean squared error 0.1108 0.1104 0.1767 0.1768 0.2405 0.2409
R2 0.883 0.883 0.882 0.881 0.872 0.871
Test for joint significance 0.31 0.57 1.43
of omitted variables: (0.73) (0.57) (0.24)
F(2, 245) with p-value in ( )
Absolute t-ratios in parentheses; 272 observations. All variables are measured in natural logarithms. A negative (positive) sign on the interaction terms indicates that the
variable contributes to more (less) pro-poor growth. The regressions also included state-specific intercepts.
variables, the value of R2 in the regression for the headcount index is 0.918. Using our
explanatory variables, it drops to 0.882 indicating that only 3.6% of the variance in the
poverty measures is accounted for by omitted variables (including measurement errors)
influencing the measured non-farm output elasticity of poverty.
We know already that inter-state differences in elasticities with respect to non-farm
output were quantitatively important to the rates of poverty reduction. From Tables 5 and
6, we can also say something about the relative importance of the various initial conditions
we have identified in explaining the differences in elasticities. Consider the state with the
lowest elasticity for the headcount index, namely Bihar with an absolute elasticity of 0.25,
well below Kerala’s elasticity of 1.08. What non-farm output elasticity of poverty would
we have seen in Bihar if it had Kerala’s initial conditions? The female literacy rate in 1960
was in Bihar was 6.9%, while it was 38.9% in Kerala. If Bihar had Kerala’s literacy rate,
then the parameter estimates in Table 6 (restricted form) imply that the (absolute) elasticity
of the headcount index to non-farm output per person in Bihar would have risen roughly
three-fold, from 0.25 to 0.79.11
Performing the same calculation for the other initial conditions, we find that if Bihar had
Kerala’s infant mortality rate, then Bihar’s elasticity would have risen by 0.06, while with
Kerala’s higher initial farm yields, Bihar’s elasticity would have risen by a further 0.06.
However, the urban – rural consumption disparity and landlessness was higher in Kerala; so
Kerala’s values for these initial conditions would have reduced Bihar’s (absolute) elasticity
by 0.02 and 0.18, respectively. So maintaining the literacy difference between the two
states but equalizing the four other factors would have meant an even lower elasticity in
Bihar relative to Kerala. In other words, the difference in literacy was the overwhelming
factor in explaining the difference in the poverty impact of non-farm output growth.
4. Conclusions
Using state-level poverty measures for India spanning 1960 to 1994 and allowing for
state fixed effects, we find that higher farm yields, higher state development spending, higher
(urban and rural) non-farm output and lower inflation were all poverty reducing. Except for
non-farm output, we could not reject the null hypothesis that all these variables had the same
elasticity across states for a given poverty measure. Farm yield growth, for example, was
poverty reducing but with a similar elasticity in states with dissimilar initial conditions; thus
it was the differences in the rate of agricultural growth that mattered to the poor.
However, the elasticity of poverty to non-farm output growth varied significantly across
states; growth in this sector brought much larger proportionate reductions in consumption-
poverty measures in some states than others. Thus, we have been able to derive a state-
specific measure of how pro-poor economic growth has been in India over this period.
Differing non-farm output elasticities of poverty appear to have had a powerful longer
term impact on the prospects of escaping absolute poverty in India. We have simulated the
rates of poverty reduction if all states had the non-farm output elasticity of West Bengal,
11
The number of literate women per 1000 adult women is logged, so 0.79 = 0.26 + 0.306x(5.96 4.23)
(recalling that the regressor is the sum of current and lagged output).
which had the highest elasticity of any state (with Kerala a close second). Then the trend
rate of reduction in the headcount index over this 35-year period would have been more
than two percentage points higher. For a state with the national average poverty incidence
in 1960 of 48%, this difference in trend rates of poverty reduction would have meant a
difference in the poverty rate by the mid-1990s of almost 16 percentage points; the poverty
rate would have fallen to 14% instead of 30%.
Our results also indicate that the national rate of poverty reduction depended on the
geographic composition of growth as well as its overall rate. However, the geographic pattern
of growth in India has not been pro-poor, in that the rate of growth in the non-farm economy
has not been any higher in states where it would have greater impact on national poverty.
Consistent with arguments found in the literature, we find that the inter-state differences
in the impact of a given rate of non-farm economic growth on consumption poverty reflect
observed differences in initial conditions. Low farm productivity, low rural living stand-
ards relative to urban areas, greater landlessness in rural areas, and poor basic education
and health all inhibited the prospects of the poor participating in growth of the non-farm
sector. Taken as a whole, our results suggest that non-farm economic growth was less
effective in reducing poverty in states with poorer initial conditions in terms of rural
development (in both absolute terms and relative to urban areas), human resources and
land distribution. The sectoral composition of economic growth was clearly more
important to poverty reduction in states with poor initial conditions.
Rural and human resource development and a more egalitarian distribution of land
appear to be strongly synergistic with poverty reduction though an expanding non-farm
economy. Amongst the initial conditions we have found to matter significantly to
prospects for pro-poor growth, the role played by literacy is particularly notable. For
example, nearly two-thirds of the difference between the elasticity of the headcount index
of poverty to non-farm output for Bihar (the state with lowest absolute elasticity) and
Kerala is attributable to the latter’s substantially higher initial literacy rate.
Acknowledgements
The support of the World Bank’s South Asia Poverty Reduction and Economic
Management Unit and the Bank’s Research Committee (under RPO 681-39) are gratefully
acknowledged. Helpful comments were received from Jock Anderson, Christina
Malmberg Calvo, Aart Kraay, Gus Ranis, Dominique van de Walle, seminar participants
at the World Bank, the International Monetary Fund, a workshop co-sponsored by the
International Food Policy Research Institute and the Institute for Development and
Communication, India, and the journal’s anonymous referees. These are the views of the
authors, and should not be attributed to the World Bank.
Appendix A
This appendix describes the main features of our data. A complete description of the
data set assembled for this study (including sources of all variables) can be found in Özler
et al. (1996) and the data and manual can be found at the following web site: http://www.
worldbank.org/poverty/data/indiap aper.htm.
We have used a consistent series of absolute poverty measures based on distributions of
consumption per capita from 20 rounds of the National Sample Survey (NSS) spanning the
period 1960– 1961 (round 16) to 1993 – 1994 (round 50). All 20 rounds of the survey are
covered for all 15 major states with the exception of Jammu and Kashmir, for which
surveys were not held for the 48th and 50th rounds due to the prevailing political unrest.
Punjab and Haryana had to be treated as a composite state because Haryana emerged as a
separate state only in 1964. For NSS rounds since then, the poverty measures for the two
states have been aggregated using population weights derived from the decennial
censuses. Altogether, we use data from 298 consumption distributions to construct
state-level poverty measures.
There is considerable variation in the sample sizes. For all states, the samples range
from 6330 households for the 16th round (July 1959 – June 1960) to 157,928 households
for the 32nd round (July 1977 – June 1978), with a median sample size of 25,761
households for the 43rd round (July 1986 –June 1987). The smallest sample size for
any state is 172 households for Assam for the 16th round. Assuming a simple random
sample, this implies a standard error, for a headcount index of 50%, of 3.8 percentage
points.
The poverty lines we use are those proposed by the Planning Commission (GOI, 1979).
These lines were defined at the per capita monthly expenditure levels of Rs. 49 for rural
areas and Rs. 57 for urban areas (rounded to the nearest rupee) at October 1973 –June
1974 all-India prices. The Planning Commission followed the ‘‘food-energy method’’ in
deriving the rural and urban lines; these poverty lines thus corresponded to levels of per
capita total expenditure at which certain caloric norms were typically attained in the rural
and urban sectors. The norms correspond to a per capita food energy intake of 2400
calories per day in rural areas and 2100 calories per day in urban areas. Poverty lines
constructed this way have sometimes been found not to have the same purchasing power
in urban and rural areas (Ravallion, 1994). However, independent estimates of the urban –
rural cost of living differential for 1973 – 1974 (see Bhattacharya et al., 1980) came up with
a similar figure of about 16% higher urban cost of living implicit in the Planning
Commission poverty lines (see Datt, 1997, for further discussion).
A substantial effort was invested into the construction of a consistent set of price
indices across states and survey periods, using monthly data on consumer price indices
from the Labour Bureau (disaggregated to the center level for the urban index) over the
whole 35-year period. Our primary deflators were the Consumer Price Index for Industrial
Workers (CPIIW) for the urban sector and the adjusted all-India Consumer Price Index for
Agricultural Labourers (CPIAL) for the rural sector. The adjustment carried out to the
CPIAL was for the price of firewood that has been held constant in the official CPIAL
series since 1960– 1961. The nominal state-level distributions were further normalized for
inter-state cost of living differentials estimated separately for urban and rural areas,
anchored to the consumption pattern of households in the neighborhood of the poverty
line. However, since a single price index is used for a given state and sector, we do not
allow for differences between expenditure groups (as would arise from non-homothetic
preferences with changes in relative prices). For further details on the construction of the
price indices, see Özler et al. (1996), Datt (1997), and Datt and Ravallion (1998a). The
poverty measures are estimated from the published grouped distributions of per capita
expenditure using parameterized Lorenz curves; for details on the methodology, see Datt
and Ravallion (1992).
As discussed above, the poverty measures are hypothesized to depend on both a set of
time-dependent variables as well as a set of initial condition variables that determine how
poverty-reducing the time-dependent variables are. Building on the empirical approach
used in our earlier work (Datt and Ravallion, 1998a), we use time-dependent variables
related to agricultural and non-agricultural growth, public spending on economic and
social services and inflation. The specific variables used are as follows:
(i) mean farm yield, given by real agricultural state domestic product (SDP) per hectare
of net sown area in the state (denoted YLD),12
(ii) non-farm output, measured by real non-agricultural state domestic product per
person (NFP),
(iii) rate of inflation in the rural sector measured as the change per year in the natural
log of the (adjusted) CPIAL,13
(iv) real state development expenditure per capita (GOV); development expenditure
includes expenditure on economic and social services. The economic services include
agriculture, rural development, special area programs, irrigation and flood control, energy,
industry and minerals, transport and communications, science, technology and environ-
ment. The social services include education, medical and public health, family welfare,
water supply and sanitation, housing, urban development, labor and labor welfare, social
security and welfare, nutrition, and relief for natural calamities.
The data on SDP and state development expenditure are available on an annual basis,
while the NSS surveys are not only not annual but they also do not always cover a full 12-
month period. To match the annual data with the poverty data by NSS rounds, we have
log-linearly interpolated the annual data to the mid-point of the survey period of each NSS
round.
We also identify a number of social and economic variables to describe initial
conditions around 1960 (that we will later use in attempting to explain the growth
elasticities of the poverty measures by state).14 The following variables (all measured in
natural logs) describe these initial conditions:
(i) the female literacy rate in 1961 defined as the number of literate females per
thousand females in the total state population (FLIT),
(ii) the percentage of landless rural households in 1961 –1962 (LLESS), as a measure of
initial asset inequality in rural areas,
12
All real values were calculated using the (adjusted) state-specific CPIAL as the deflator. For further details
on the State Domestic Product (SDP) data, see Datt and Ravallion (1998a).
13
This is state specific. However, the bulk of the effect is clearly through inter-temporal variation in the rate
of inflation. We also tried adding the log of the ratio of the CPIIW to the (adjusted) CPIAL as an additional
regressor, but that turned out to be insignificant.
14
The initial condition variables were assembled from a number of diverse data sources including the 1961
Census, the Statistical Abstract (Central Statistical Organization) for various years, and reports from a number of
NSS surveys dealing with village statistics, land holdings and utilization, fertility, and infant mortality.
(iii) the proportion of urban population in 1961 (URB),

(iv) the ratio of the initial urban real mean consumption to that in the rural sector, where
the initial real mean consumption in each sector is formed as an average over the first
three NSS rounds available for that state (CDIF), as a measure of initial inter-sectoral
disparity,
(v) the percentage of operated area which was irrigated in 1957 –1960 (IRR),
(vi) the initial levels of YLD and NFP (for 1960 –1961).
References
Aghion, P., Caroli, E., Garcia-Penalosa, C., 1999. Inequality and economic growth: the perspectives of the new
growth theories. Journal of Economic Literature 37 (4), 1615 – 1660.
Ahluwalia, M.S., 1978. Rural poverty and agricultural performance in India. Journal of Development Studies 14
(3), 298 – 323.
Anand, S., Kanbur, R., 1993. The Kuznets process and the inequality-development relationship. Journal of
Development Economics 40, 25 – 52.
Atkinson, A.B., 1997. Bringing income distribution in from the cold. Economic Journal 107, 297 – 321 (Mar).
Benabou, R., 1996. Inequality and growth. In: Bernanke, B., Rotemberg, R. (Eds.), National Bureau of Economic
Research Macroeconomics Annual. MIT Press, Cambridge, pp. 11 – 74.
Bhattacharya, S.S., Choudhury, A.B.R., Joshi, P.D., 1980. Regional consumer price indices based on NSS
household expenditure data. Sarvekshana 3, 107 – 121.
Binswanger, H., Deininger, K., Feder, G., 1995. Power, distortions, revolt and reform in agricultural land
relations. In: Behrman, J., Srinivasan, T.N. (Eds.), Handbook of Development Economics, vol. III. North-
Holland, Amsterdam.
Bourguignon, F., Morrison, C., 1998. Inequality and development: the role of dualism. Journal of Development
Economics 57 (2), 233 – 257.
Bruno, M., Ravallion, M., Squire, L., 1998. Equity and growth in developing countries: old and new perspectives
on the policy issues. In: Tanzi, V., Chu, K. (Eds.), Income Distribution and High-Quality Growth. MIT Press,
Cambridge.
Chen, S., Ravallion, M., 2001. How did the world’s poorest fare in the 1990’s? Review of Income and Wealth
47 (3), 283 – 300.
Clarke, C., 1940. The Conditions of Economic Progress Macmillan, London.
Datt, G., 1997. Poverty in India, 1951 – 1994: trends and decompositions. Mimeo, World Bank and IFPRI,
Washington, DC.
Datt, G., Ravallion, M., 1992. Growth and redistribution components of changes in poverty measures: a decom-
position with applications to Brazil and India in the 1980s. Journal of Development Economics 38, 275 – 295.
Datt, G., Ravallion, M., 1997. Macroeconomic crises and poverty monitoring: a case study for India. Review of
Development Economics 1 (2), 135 – 152.
Datt, G., Ravallion, M., 1998a. Why have some Indian states done better than others at reducing rural poverty?
Economica 65, 17 – 38.
Datt, G., Ravallion, M., 1998b. Farm productivity and rural poverty in India. Journal of Development Studies 34,
62 – 85.
Drèze, J., Sen, A., 1995. India: Economic Development and Social Opportunity. Oxford Univ. Press, Delhi.
Fields, G., 1980. Poverty, Inequality and Development. Cambridge Univ. Press, New York.
Foster, J., Greer, J., Thorbecke, E., 1984. A class of decomposable poverty measures. Econometrica 52, 761 – 765.
Government of India, 1979. Report of the Task Force on Projections of Minimum Needs and Effective Con-
sumption. Planning Commission, New Delhi.
Harris, J.R., Todaro, M.P., 1970. Migration, unemployment and development: a two-sector analysis. American
Economic Review, 126 – 142.
Lanjouw, J.O., Lanjouw, P., 1997. The Rural Nonfarm Sector: An Update World Bank, Washington, DC.
Li, H., Squire, L., Zou, H., 1998. Explaining international and intertemporal variations in income inequality.
Economic Journal 108, 26 – 43.
Lipton, M., Ravallion, M., 1995. Poverty and policy. In: Behrman, J., Srinivasan, T.N. (Eds.), Handbook of
Development Economics, vol. III. North-Holland, Amsterdam.
Loury, G., 1981. Intergenerational transfers and the distribution of earnings. Econometrica 49, 843 – 867.
Myrdal, G., 1988. In: Meier, G., Seers, D. (Eds.), Pioneers in Development. Oxford Univ. Press, New York.
Özler, B., Datt, G., Ravallion, M., 1996. A database on poverty and growth in India. Mimeo. Development
Research Group, World Bank, Washington, DC.
Ravallion, M., 1994. Poverty Comparisons Harwood Academic Press, Chur, Switzerland.
Ravallion, M., 1997. Can high inequality developing countries escape absolute poverty? Economics Letters 56,
51 – 57.
Ravallion, M., 2001. Measuring aggregate welfare in developing countries: how well do national accounts and
surveys agree? Policy Research Working Paper, World Bank, Washington, DC.
Ravallion, M., Chen, S., 1997. What can new survey data tell us about recent changes in distribution and poverty?
World Bank Economic Review 11 (2), 357 – 382.
Ravallion, M., Datt, G., 1996. How important to India’s poor is the sectoral composition of economic growth?
World Bank Economic Review 10 (1), 1 – 25.
Ravallion, M., Datt, G., 1999. When is growth pro-poor? Evidence from the diverse experience of India’s states.
Policy Research Working Paper WPS 2263. World Bank, Washington, DC.
Robinson, S., 1976. A note on the u-hypothesis relating income inequality and economic development. American
Economic Review 66, 437 – 440.
Schultz, T.W., 1953. The Economic Organization of Agriculture McGraw Hill, New York.
Stiglitz, J.E., 1969. The distribution of income and wealth among individuals. Econometrica 37 (3), 382 – 397.
Thorbecke, E., Hong-Sang, J., 1996. A multiplier decomposition method to analyze poverty alleviation. Journal
of Development Economics 48 (2), 279 – 300.
Timmer, P., 1997. How well do the poor connect to the growth process? Mimeo. University of California, San
Diego.
Tinbergen, J., 1975. Income Distribution: Analysis and Policies. North-Holland, Amsterdam.
Van de Walle, D., 1985. Population, growth and poverty: another look at the Indian time series evidence. Journal
of Development Studies 34 (4), 62 – 85.
Viner, J., 1953. International Trade and Economic Development. Oxford Univ. Press, Oxford.
World Bank, 1990. World Development Report: Poverty Oxford Univ. Press, New York.
World Bank, 2000. World Development Report: Attacking Poverty Oxford Univ. Press, New York.
SANEI WORKING PAPER SERIES
Exploring the Links between Inequality, Polarization

and Poverty: Empirical Evidence from Pakistan
No. 11 - 04
Abid aman burki
SOUTH ASIA NETWORK OF ECONOMIC

RESEARCH INSTITUTES
www.saneinetwork.net
SANEI WORKING PAPER SERIES
No. 11 - 04
Exploring the Links between Inequality,

Polarization and Poverty: Empirical Evidence
from Pakistan
ABID AMAN BURKI
March 2011
South Asia Network of Economic Research Institutes

Bangladesh Institute of Development Studies (BIDS)
E-17, Agargaon, Sher-e-Bangla Nagar, GPO Box # 3854, Dhaka-1207, Bangladesh
T: +88-02-9118324, F: +88-02-8181237
E-mail: saneibd@gmail.com, Web: www.saneinetwork.net
First Published March 2011
The SANEI Working Papers present research studies completed by the Network
under its different Programs. These papers reflect research carried out by researchers from
South Asian countries who were selected by SANEI to conduct research on specific topics. It
is expected that these papers would eventually be published in learned journals or form
chapters of books after undergoing due review process and hence comments are most
welcome. Neither the Management of SANEI nor any other agency associated with SANEI
necessarily endorses any or all of the views expressed in these papers. The Working Papers
reflect views based on professional analysis of the authors and the usual caveat of research
reports applies.
Terms of Use
The materials contained in the Working Papers are free for publication in its entirety
or in part in newspapers, wire services, internet-based information networks and newsletters.
One may also use the information in radio-TV discussions or basis for discussion in different
fora. We would, however, appreciate it if you could let us know when and where the
publication was used.
Recommended Citation:
Burki, A. A. 2011, Exploring the Links between Inequality, Polarization and Poverty:
Empirical Evidence from Pakistan, SANEI Working Paper Series No. 11-04, South Asia
Network of Economic Research Institutes, Dhaka.
Biography:
Abid Aman Burki is Professor of Economics at the Department of Economics,

Lahore University of Management Sciences (LUMS). He received his PhD in economics
from Kansas State University in 1996. He has also taught economics at Bahauddin Zakariya
University (1981 – 1985), Quaid-i-Azam University (1986 – 2002) and Kansas State
University (1995 – 1996). He was Chairman, Department of Economics, Quaid-i-Azam
University, between 2000 and 2002, and Director, Centre for Management and Economic
Research, Lahore University of Management Sciences, between 2003 and 2010. Professor
Burki is the author of two books and more than 70 papers in professional journals and edited
volumes. He has authored or coauthored several professional reports for government,
corporate sector, NGOs, multilateral donors and international think tanks. He has served on
several high-level committees and taskforces of the federal and provincial governments in
Pakistan. In recognition of his services, he was conferred President of Pakistan’s academic
distinction award Izaz-i-Fazeelat in 2001.
Cover Design
Samiul Ahsan
Page layout
Ahshan Ullah Bahar
Contents
Abstract.......................................................................................................................................1
1. Introduction................................................................................................................1
2. Methodology ..............................................................................................................3
2.1 Introduction......................................................................................................................3
2.2 Empirical Specification ...................................................................................................4
2.3 Headcount Poverty Ratios ...............................................................................................5
2.4 Inequality and Polarization Indices .................................................................................5
3. Data Description and Preliminary Survey Evidence..............................................................6
3.1 Introduction .....................................................................................................................6
3.2 Six Rounds of Household Survey Data ...........................................................................6
3.3 Calculating Poverty, Inequality and Polarization Indices................................................8
3.4 Econometric Results for The Basic Model......................................................................8
4. Public Investments on E&H and Poverty ............................................................................10
4.1 Introduction ...................................................................................................................10
4.2 Modeling Education and Health Indices with Principal Components ...........................10
4.3 Education and Health Data Used for Principal Component Analysis ...........................11
4.4 Six Principal Components Retained ..............................................................................12
4.5 Spatial Deprivation: Most and Least Developed Districts.............................................12
4.6 How Spatial Deprivation in E&H Affects Poverty........................................................13
4.7 Post-Primary School Systems and Poverty: Does Polarization Matter?........................14
4.8 Conclusions ...................................................................................................................16
5. Impact of Road Density on Poverty.....................................................................................16
5.1 Introduction ...................................................................................................................16
5.2 Road Density and Spatial Deprivation in Punjab ..........................................................16
5.3 Empirical Relationship between Road Density and Poverty .........................................17
5.4 The Nexus between Road Density, Inequality and Poverty ..........................................18
5.5 Conclusions ...................................................................................................................19
6. Concluding Comments.........................................................................................................20
7. References............................................................................................................................21
List of Tables
Table 1: Distribution of rural sample by household survey...................................................... 23
Table 2: Inflation-adjusted poverty lines based on cut-off of 2350 calories............................ 23
Table 3: Changes in poverty, inequality and polarization in Pakistan...................................... 23
Table 4: Changes in poverty, inequality and polarization in Punjab........................................ 23
Table 5: Definition of variables................................................................................................ 24
Table 6: Descriptive statistics – full sample............................................................................. 27
Table 7: Effects of household and individual characteristics on rural poverty, basic results...31
Table 8: District level variables used in principal component analysis.................................... 32
Table 9: Principal component rotated factor loadings.............................................................. 33
Table 10: Most and least developed districts in Pakistan based on principal component post-
primary school system and hospital index..............................................................34
ii Contents
Table 11: Most developed districts in Punjab based on principal component post-primary
school system and hospital index............................................................................35
Table 12: Least developed districts in Punjab based on principal component post-primary
school system and hospital index............................................................................ 35
Table 13: Descriptive statistics of policy variables in full sample...........................................36
Table 14: Impact of education and health infrastructure on rural poverty in Pakistan.............37
Table 15: Post-primary school system and poverty under different inequality and polarization
regimes....................................................................................................................38
Table 16: Summary statistics of Punjab’s sample...................................................................39
Table 17: Descriptive statistics of policy variables in Punjab..................................................41
Table 18: Impact of road density on poverty in Punjab............................................................42
Table 19: Road density and poverty under different inequality and polarization regimes......43
List of Figures
Figure 1: Geographic location of 20 most developed districts based on post-primary school
system and hospital index......................................................................................... 44
Figure 2: Geographic location of 20 least developed districts based on post-primary school
system and hospital index......................................................................................... 45
Figure 3: Geographic concentration of road density in Punjab, 2005-06................................. 46
Figure 4: Relative road density of Punjab’s districts, 1992-93 vs. 2005-06............................. 46
Figure 5: Road density and poverty headcount in Punjab........................................................ 47
Acronyms
CWIQ Core Welfare Indicator Questionnaire
E&H Education and health
FATA Federally Administered Tribal Areas
HIES Household Income and Expenditure Survey
KP Khyber Pakhtunkhwa
NWFP North-West Frontier Province
PATA Provincially Administered Tribal Areas
PC Principal component
PIHS Pakistan Integrated Household Survey
PSLM Pakistan Social and Living Standards Measurement Survey
PSU Primary Sampling Unit
SD Standard deviation
SMEs Small and Medium Enterprises
TB Tuberculosis
Acknowledgements
I wish to acknowledge and express my sincere gratitude to various individuals
for valuable discussions, comments and support during the course of this study. I wish
to express my sincere gratitude to Syed Zahid Ali and Mushtaq Ahmad Khan for their
thoughtful criticism, suggestions and comments. My special thanks go to Kiran
Naseer for her splendid support in collecting district level data on education and
health indicators. Abubakar Memon provided excellent support in compiling
published data from household surveys. He also provided invaluable assistance in
computing and graphs. I am also pleased to acknowledge excellent research assistance
provided by Kiran Javaid on various stages of the study. She looked up relevant
literature, assembled and compiled data, provided support in computation and graphs.
Abid A. Burki February 8, 2011

Exploring the Links between Inequality,
Polarization and Poverty: Empirical Evidence from
Pakistan
ABID AMAN BURKI
Abstract
The main purpose of this study was to explore how spatial variation in public sector
infrastructural investments affects rural poverty under various inequality and polarization
regimes. Using data from six rounds of Pakistan’s Household Income and Expenditure
Survey (HIES) and 349295 observations we find that the percentage of rural people
changes over time. The movement of Gini index and Wolfson’s polarization index confirm
that the issues of distribution change that are missed out by the inequality index are better
described by the polarization index. We find that, among other things, poverty coexists with
illiteracy of household heads and lack of asset ownership. Mapping of districts clearly
identifies the most developed and least developed districts, measured both by post-primary
school and hospital index and road density. The evidence shows that dispersion between
least developed and most developed districts is increasing with the passage of time.
Increased public sector investments on education and health infrastructure are associated
with decline in rural poverty throughout all but highly unequal and polarized districts: it
suggests that public sector investments do not necessarily lead to welfare gains and that the
returns on investments significantly differ across regions. Likewise, we find a strong
negative relationship between road density and probability of poverty; however, the long
run decline in poverty due to investment on roads almost doubles when we move from high-
inequality/polarized and medium-inequality/polarized districts to low-inequality/polarized
districts. With investment on road network, the poverty reduction potential of less polarized
districts far exceeds the poverty reduction potential of less unequal districts: a further
confirmation that issues of distributional change are indeed missed out by the inequality
index. We conclude that public policies that seek more regional equality and less
polarization are desirable for pro poor growth policies in developing countries.
1. INTRODUCTION
Development practitioners and policy makers recognize that the objective of
development effort in developing countries is to increase the living standards of the
common man. It is also obvious to the policy makers that to achieve this goal
economic growth is not sufficient. Therefore, to make sensible poverty reduction
strategies the policy makers must know the correlates of poverty in respective
countries.
A vast empirical literature on poverty is available which provides insights into the
variables that have significantly affected poverty in the past. A number of recent
papers offer explanations on correlates of poverty in developing countries. Muller
(2002), Ravallion and Datt (2002), Ravallion and Chen (2007), for example, show
that how prices affect living standards in developing countries. Basu (2006), Winters
et al. (2004), Harrison et al. (2003), and Minot and Goletti (1998) describe
circumstances when globalization and trade liberalization can lead to alleviation of
poverty. The association between agricultural productivity and poverty is discussed,
among others, by Binswanger and Braun (1991), Graff et al. (2006) and Ravallion and
Chen (2007). Some papers focus on the effect of public policy and institutions on
poverty, e.g., Besley et al. (2005), Fan et al. (2000), Ravallion and Datt (2002).
2 SANEI Working Paper Series
Alternatively, a number of studies address the questions of pro-poor growth where

the argument is that economic growth alone is not sufficient to alleviate poverty and
that there are a number of other factors that would critically determine whether or not
the growth is pro-poor [see, Ravallion and Datt (2002), Ravallion and Chen (2003),
Kraay (2006), Klasen (2008), Suryahadi et al. (2009)]. Some empirical papers [see,
for example, Ravallion (1997), Ravallion and Datt (2002), Ravallion and Chen
(2007)] have established that given a positive rate of economic growth, regions facing
inequalities in income will face greater difficulties in reducing poverty. Wagle (2010)
employs cross-national data to investigate the association between inequality and
poverty and finds “the pivotal role that reduction in inequality can play in effectively
curbing poverty” in developing countries.
In contrast to the literature on the issues of poverty and inequality, a number of
theoretical and empirical papers to date have explored the phenomenon conveniently
termed as polarization [see, for example, Esteban and Ray (1994), Wolfson (1997),
Duclos et al. (2004)]. As noted by Wolfson (1997), a more polarized distribution of
income refers to “more spread out” distribution that includes “a hallowed out middle”
or “individuals moving out from the middle to the tails of the distribution.” Defined in
this way, the concept of polarization is not the same as formally defined in the
inequality literature [Atkinson (1970)]. That polarization and inequality orderings can
indeed differ in practice has been emphasized by many studies, including Duclos and
Echevin (2005), Duro (2005), Duclos et al. (2004), Seshanna and Decornez (2003),
Fedorov (2002), Zhang and Kanbur (2001), Gradin (2000), Wolfson (1997), Esteban
and Ray (1994). Therefore, it is important to look at income differentials across
regions within these countries, especially in regions where there is substantial degree
of heterogeneity in economic performance.
In the context of growth-inequality-poverty nexus, Ravallion and Chen (1997)
discuss that conventional inequality measures such as Gini index may not capture
distributional changes of concern to policy makers in developing countries. They
postulate that to determine political feasibility of a reform, we need to take into
account its impact on middle strata of the population. But the fact remains that the
Gini index may not capture the changes in the share of income held by the middle
stratum. While they underline the need to employ income polarization measures to
study distributions where rich are becoming richer and poor are becoming poorer,
they do not investigate the effect of public investments on infrastructure on incidence
of poverty when the society is polarized. How sensitive is the incidence of poverty to
different measures of income distribution and polarization has not heretofore, so far as
we could determine, been investigated in any country.
The present study builds on our understanding of the implications of polarization
on income distribution and incidence of poverty. We try to assess the sensitivity of
measures of inequality and polarization with poverty incidence by using Pakistan’s
pooled cross-section data from six rounds of household surveys for rural areas over
the 1992-93 to 2005-06 periods.1
1
There is a large literature on poverty in Pakistan. For some earlier estimates, see Naseem (1973, 1977),
Irfan and Amjad (1984), Ercelawn (1988, 1989, and 1990) and Malik (1993). For more recent studies,
Burki: Exploring the Links between Inequality, Polarization and Poverty 3
While Pakistan is a federation where significant economic power rests with the
federal government, its four provinces are free to make their independent choices for
public policies towards regional economic development. The heterogeneity in inter-
provincial and intra-provincial public policies over time and space allows us to
observe and study the links between public policy and its outcomes in the form of
public sector infrastructural investments, inequality/polarization and poverty. We are
interested to explore in depth the sensitivity of incidence of headcount poverty with
measures of income inequality and income polarization for given levels of public
infrastructural investments.
The goal of the study is two fold. First, it will document consistent time-series of
headcount poverty, measures of income inequality and degree and dimensions of
polarization from 1992-93 to 2005-06. Second, the study would let the data tell us that
given public sector infrastructural investments how increase in measures of inequality
and polarization affect incidence of poverty? The public sector infrastructural
investments considered in this study are investments on education and health (E&H)
infrastructure and investments on road infrastructure. For E&H we consider 20
indicators of such investments and select six indices by using the principal component
analysis. Toward this end, we develop a framework that allows us to study these
relationships.
Chapter 2 outlines the methodology adopted in the study. The data and data
sources are described in Chapter 3. Chapter 4 explains the principal component
analysis used to reduce 20 E&H variables into six factors and explains their
association with poverty. Moreover, this Chapter also discusses how infrastructure
investments in post-primary school system & hospitals affect poverty under various
inequality and polarization regimes. Chapter 5 maps road infrastructure and studies
the effects of unequal investments on road infrastructure on spatial deprivation in
Pakistan’s Punjab province. Then it goes on to study its effects on poverty under
various inequality and polarization regimes. Concluding comments are presented in
Chapter 6.
2. METHODOLOGY
2.1 Introduction
This chapter describes the empirical methodology for estimating how rural
poverty responds to individual and household characteristics and regional
infrastructural variables considered in this study. We begin this Chapter by specifying
the estimating equations for the determinants of rural poverty. This is followed by the
methods of estimating headcount poverty, inequality and polarization indices and
explanation on how we implement these methods in this study.
2.2 Empirical Specification
We begin by writing our basic model aimed at investigating how individual,
household and other control variables affect incidence of rural poverty in our sample.
see for instance, Haq et al. (2008), Burki and Khan (2008), Arif (2006), Malik (2005), Jamal (2003),
Anwar and Qureshi (2002) and Jafri (1999). For a review of papers on income inequality, see Anwar
(2005).
Because our dependent variable is binary in nature, we estimate binary probit

regressions where we assume that the dependent variable yijt is a latent variable,
which can be expressed as a linear function of other variables that affect the
probability that the individual is poor versus non-poor. The basic form of the
empirical model can be written as
yijt = α + γ X ijt + (η k × τ t ) + D j + ε ijt (1)
where yijt is an index for individual i living in district j in province k , and t is
an index for the year of the survey. Several household and individual level control
variables are represented in vector X ijt , the interaction term for period fixed-effects
and province fixed-effects (η k × τ t ) captures variation in individual poverty due to
changes in macroeconomic indicators over-time and province-specific heterogeneity,
while D j is for time-invariant district fixed-effects.
The dependent variable yijt is not observable because this is a latent variable that
equals 1 if individual falls below the official poverty line during the survey year and
zero otherwise. The probit model assumes that the error structure ε ijt has a cumulative
normal distribution function where the probability of an individual falling below the
poverty line pijt can be written as
1 α + γ X ijt + (ηk ×τ t ) + D j − 12 t 2
pijt = Pr ( yijt = 1) =
2π ∫−∞
e dt (2)
where t is a standardized normal variable. The coefficient estimates are obtained

from the maximum likelihood estimation procedure. To investigate the impact of
infrastructure investments, and the nexus between infrastructure investments,
inequality/polarization and poverty, we modify the basic model which takes the
following form
yijt = α + β ( I kt × g jt ) + γ X ijt + (ηk ×τ t ) + D j + ε ijt (3)
where I kt is a measure of income inequality/polarization in district k at time t ,
g jt measures infrastructural investments in district j at time t while all other
variables are defined as above. Our variable of interest ( I kt × g jt ) is the interaction of
a measure of income inequality (or polarization) with a measure of public sector
infrastructural investments. This will allow us to test whether these investments are
sensitive to variation in measures of income inequality and polarization in alleviating
poverty.
2.2 Headcount Poverty Ratios
The head count ratio of poverty gives us the relative incidence of the poor, or the
number of people below the poverty line. If we denote expenditure (income) by y and
subscripts i refers to individuals, and p refers to the poverty line, then head count
denoted by H is given by the number of individuals so that yi < p . The head count
ratio ( p0 ) is given by p0 = ( H / n) × 100 , where n is the total population in the
reference group. Distributional sensitivities in poverty measures are addressed in the
class of poverty measures proposed by Foster, Greer and Thorbecke (1984). To

illustrate, for any power α , the Foster-Greer-Thorbecke measure is defined as
α
1 ⎛ p − yi ⎞
pα = ∑ ⎜ ⎟
n yi < p ⎝ p ⎠
where the measure is just the head count ratio when α = 0 ; the measure is the
poverty gap ratio when α = 1 ; and the measure takes into account the degree of
inequality and severity of poverty among the poor households when α = 2 . For the
present application, we just take the assumption of α = 0 . We use the official poverty
line suggested by the cutoff of Rs.723.4 for basic needs poverty in 2000-01 indicating
the income/expenditure required to achieve the minimum calories of 2350 per adult
equivalent per month [GoP (2007)].2 This official poverty line is inflated or deflated
to get consistent series for other survey years.
2.3 Inequality and Polarization Indices
Regional income inequality is measured by the most commonly referred measure
known as the Gini coefficient, which provides some good benchmark values and
written as
K K
1
G=
μ
∑∑
i =1 j =1
f ( yi ) f ( y j ) yi − y j
where yi depicts the value of an indicator in the ith region, μ represents average
value of the indicator for the whole country, f ( yi ) is the population share of the
ith region in total population of the country and K indicates the number of regions.
The values of Gini-coefficient range from 1 to 0, where 1 indicates perfect inequality,
and 0 indicate perfect equality.
The polarization index used is this study is the Wolfson index [Wolfson (1994,
1997)] that is defined as
2 ( 2T − Gini )
W=
(m μ )
where T represents the area of the trapezoid = 0.5 – L(0.05) and L(0.5) denotes
income share of the bottom half of the population; m is for median income; μ
represents the mean income. The higher the index, the more polarization increases and
the more the middle class decreases.
3. DATA DESCRIPTION AND PRELIMINARY SURVEY EVIDENCE
3.1 Introduction
This study examines the impact of public sector infrastructure investments on
rural poverty across regions in Pakistan. The federative structure of Pakistan consists
of four provinces, e.g., Punjab, Sindh, Khyber Pakhtunkhwa (formerly North West
Frontier Province or NWFP), Balochistan, Federally Administered Tribal Regions
2
This approach assumes that the households whose food expenditures are equal to the food poverty
line would also be able to satisfy their basic needs.
(FATA) and Provincially Administered Tribal Areas (PATA). Islamabad is the

federal capital. The four provinces are further divided into administrative divisions
and administrative districts. Historically, the survey data in Pakistan is only collected
from the four provinces and its administrative districts.
Provinces have a direct control on their budgets and they have the rights to levy
different kinds of taxes and levies in their jurisdiction. Investments in local
infrastructure, e.g., schools, colleges, hospitals, dispensaries, rural health units and
road infrastructure, etc. come from both the federal government as well as from
provincial governments. However, for the most part, investments on regional
infrastructure are made by keeping in view the political and strategic considerations.
Historically, funds for infrastructure development have disproportionately been
allocated to the metropolitan cities and their surrounding districts. By 1990-91, there
were 56 administrative districts, which were subdivided for administrative reasons in
the later period. These districts mostly rely on provincial and federal government for
local infrastructure development. For spatial mapping of Pakistan, districts are often
used as the unit of analysis [see among others, Jamal et al. (2003), Wasti and Siddiqui
(2008) and references therein]. The subject of this study is also a district: the data is
collected in household surveys with district level codes while information on
infrastructure indicators is published in provincial development reports.
This chapter provides information on the six rounds of household survey used for
empirical analysis in this study; calculates and describes the trends on headcount
poverty, Gini inequality index and Wolfson polarization index, and gives summary
statistics of key variables. We conclude this Chapter by presenting econometric
results for the basic model by using pooled cross-section data from six rounds of the
household survey.
3.2 Six Rounds of Household Survey Data
The household and individual level data for this study comes from Pakistan’s
Household Income and Expenditure Surveys (HIES) administered by the Federal
Bureau of Statistics, Government of Pakistan.3 We use data drawn from six rounds of
HIES surveys from 1992-93 to 2005-06. The included survey rounds are HIES 1992-
93, HIES 1993-94, HIES 1996-97, PIHS-HIES 1998-99, PIHS-HIES 2001-02, and
PSLM-HIES 2005-06.4,5 The survey selects households by using stratified random
sampling methods by identifying primary sampling units (PSUs) in 8 metropolitan
3
Most of the studies on poverty and inequality in Pakistan have used Household Income and
Expenditure Survey (HIES), and Pakistan Integrated Household Survey (PIHS) data.
4
The HIES has been conducted, with some breaks, since 1963. In 1990, the HIES questionnaire was
revised in order to address the requirements of a new system of national accounts and incorporated in a
larger survey called Pakistan Integrated Household Survey (PIHS). The series of four rounds – 1990-91,
1992-93, 1993-94 and 1996-97 followed the design of these new questionnaires. In 1998, the HIES
data collection methods and questionnaire were again changed to reflect the integration of the HIES
with the PIHS. Same format was adopted during 2001-02. Since 2004-05, PIHS has been replaced with
Pakistan Living Standards Measurement Survey (PSLM) including HIES, and Core Welfare Indicators
(CWIQ). Fortunately, HIES was integral part of all PIHS and PSLM surveys. These surveys cover
urban and rural households and four provinces, but exclude tribal and military areas. These surveys are
representative at the province level as well as at the level of rural and urban areas.
5
For other papers that have used pooled cross-section data at the individual and household level, see
for instance Kruger (2007).
cities, other urban stratum and rural stratum. For this study we use household data
only from the rural districts, which were easily identifiable from the rural district
codes.
The combined data file consisted of 373991 individual level observations from six
rounds of the survey. Altogether, 24002 observations were discarded from the original
sample because they did not report food expenditure needed to measure poverty. We
were guided by the previous literature on poverty and child labor that suggests that
landholding variable may be endogenous. To purge this problem, we deleted 694
observations from our sample for households that acquired land through sale or
purchase thus leaving us with a sample where landholding variable was representative
of only inherited or rented-in land, which were exogenous. After these adjustments in
the total sample, we were left with a sample of 349295 observations. The distribution
of our working sample by survey years is presented in Table 1.
An important question before merging six rounds of the survey was the changing
district boundaries across survey years whereby several new districts were created
from the old ones due to increased population. For the sake of consistency, we
retained the 1990-91 district classification and created consistent pooled cross-section
data from the six rounds. While the survey identified households from Punjab, Sindh
and KP on the basis of their district codes, households from Balochistan province
were identified on the basis of administrative divisions instead of districts. This leaves
us with a total of 56 districts to work with.
We augment household survey data with some external information on district
based education, health and road density variables that represent government
investment on human and physical infrastructure. This information is obtained from
Provincial Development Reports and Punjab Highway Department. If non-income
indicators of welfare called social services are improved, they are expected to lower
poverty. We employ principal component analysis to construct six education and
health indices and a road density variable based on the data of Punjab province.
3.3 Calculating Poverty, Inequality and Polarization Indices
To calculate headcount poverty ratio, we use inflation-adjusted poverty lines
suggested by the cut off of 2350 calories per adult equivalent per day. The inflation-
adjusted poverty lines used for various survey years are given in Table 2. In our
poverty calculations, we also account for regional heterogeneity in prices and
normalize the external effects on prices of food items by using the Paasche index. We
follow a consistent methodology for all the survey years. However, Gini index of
inequality and Wolfson polarization index measures are calculated by using per capita
household consumption expenditures.
Table 3 presents changes in overall poverty, inequality and polarization indices for
Pakistan and shows that the estimates of poverty, inequality and polarization changed
considerably during the period of the study. The headcount ratio of poverty is the
percentage of population living below the poverty line. The percentage of people
living below the poverty line in rural areas increased from 35.1% in 1992-93 to 37%
in 1993-94 and declined to 29.2% in 1996-97 before rising to 46% in 1998-99. Rural
poverty has consistently declined since 1998-99.
From Table 3 it is abundantly clear that the trend of Gini index and Wolfson
polarization index is not uniform as they do not always move in the same direction.
Gini index increases from 1992-93 to 1998-99 showing increase in income inequality
in the country. During the same period, the polarization index first remains constant
then gently decreases from 1993-94 to 1996-97 (an indication of increase in the
middle class). Polarization increases from 1996-97 to 1998-99 indicating a sharp drop
in the middle class. Likewise, from 2000-01 to 2005-06 when income inequality
roughly remains constant, polarization index again increases to suggest another
episode of decline in the middle class. Table 4 presents changes in poverty, inequality
and polarization in Punjab province which are qualitatively similar to the trends
observed on Pakistan’s data. These trends suggest that there are indeed issues of
distributional change that are missed out by Gini index, but are better described by the
Wolson index that need to be investigated.
3.4 Econometric Results for The Basic Model
Definition of variables are found in Table 5 while Table 6 reveals the descriptive
statistics of all the individuals in our sample, their household and individual
characteristics and other control variables used in probit regressions. The sample
consists of 349295 individuals from 56 rural districts. About 35% of them were below
poverty line, which is similar to other developing countries and the countries in the
region.
We first present basic regression results using the empirical methodology
specified in Eq.(1) for the incidence of poverty. The results of this probit regression
appear in Table 7 from column (1) to column (4). In all regressions, we include
household and individual controls as the set of regressors, which consist of four
dummy variables to capture the education level of the head of household (the omitted
category is illiterate household head); three dummy variables to capture the effects of
household size (the omitted category is small-sized households consisting of 1 to 3
members); four dummies to capture household landholding (the omitted category is
no landholding); one dummy variable each to indicate if household owns livestock
and farm equipment, the value of financial assets owned by household; age of the
individual; a dummy for gender; days worked during the last month; and twelve
sector of employment dummies (the omitted category is unemployed).
The estimation results presented in column (1) are marginal effects from probit
regression obtained without including any covariates other than household and
individual characteristics. Column (2) presents results when we include in regression
fifty-six district indicators to capture time-invariant changes across districts (the
omitted category is Vehari district). The estimation results in column (3) are obtained
by deleting from the regression district effects but including twenty-four
province × year indicators to control for province-specific time-varying factors (the
excluded category is Sindh × 2005-06). Column (4) presents results when we control
for both province × year fixed effects and district fixed effects. Finally, the numbers in
parenthesis in Table 10 indicate asymptotic t-values obtained from robust standard
errors, corrected for clustering at the district level.
The estimation results presented in columns (1)–(4) of Table 7 indicate that the
parameter estimates are highly robust to alternative model specifications. The
estimated coefficients indicate consistent negative impact of education of household
head and ownership of household assets (including land, livestock, farm equipment
and financial assets) on poverty. For instance, there is unequivocal evidence to show
that poverty coexists with illiteracy of heads of household and lack of asset
ownership. Our results show that the incidence of poverty dramatically falls with
increase in education of household head. The results in column (4) suggest that
compared with illiterate heads, poverty falls by 8.9% when head has one to five years
of schooling, 15.5% when heads has six to ten years of schooling and 25% when head
has more than secondary education. This evidence suggests that making education
more attractive and accessible can be used as an important policy tool to reduce rural
poverty. Thus, policies that seek universal access to education by increasing the
quantity and quality of schools and colleges have a strong power to reduce rural
poverty.
Greater land wealth is usually positively correlated with household’s income.
Likewise, the results show that the probability of being poor progressively falls as
household land wealth increases from subsistence landholding, to economical
landholding and then to large landholding. Ownership of farm equipment is also
negatively associated with the incidence of poverty in rural Pakistan. Due to small
landholdings, ownership of farm equipment is not viable for a majority of farm
producers in Pakistan. Therefore, a vibrant equipment rental market has fast emerged
during the last three decades, which renders negative effect on incidence of poverty.
While households’ ownership of livestock has long been recognized as a sign of
economic stability and insurance for millions of resource poor farmers in South Asia,
deeper links of local dairy producers with organized milk supply chain of milk
processing industry in recent times has further increased the potential of this sector in
poverty alleviation. Therefore, livestock households are significantly less likely to be
poor than non-livestock households. Two studies based on the data of Africa and
Peruvian Andes also reach similar conclusions [Peacock (2005), Kristjanson et al.
(2007)]. Not surprisingly, the significantly negative coefficient on household financial
assets indicates that a one standard deviation (i.e., 0.128) increase in these assets
lowers the probability of poverty by 23%.
As far as household level factors are concerned, the family size of a household
makes the most significant difference on poverty. Increase in family size consistently
increases poverty in our sample. Relative to small sized households (1 to 3 members),
the likelihood of poverty increases by 36% for individuals coming from medium sized
households (4 to 8 members) and to almost 59% for individuals from large sized
households. In other words, scale economies often attributed to larger households do
not exist in rural Pakistan.
In the individual level control variables, the coefficient on age is significantly
negative suggesting that younger population is more likely to be poor. Given that the
majority of Pakistani population is young, this result is alarming. The incidence of
poverty on male and female population is roughly similar, except that males are
0.06% less likely to be poor than females. Increase in the number of days worked
significantly decreases the probability of poverty in our sample.
4. PUBLIC INVESTMENTS ON E&H AND POVERTY
4.1 Introduction
Recent literature on poverty in developing countries suggests that expenditures on
education and health care can be instrumental in alleviating poverty [Klasen (2008),
Gustafsson and Shi (2004)], but empirical evidence on the nature of disparities across
regions and their impact on poverty is far from clear. Moreover, empirical studies on
the impact of regional level human capital infrastructure on poverty, especially under
varying degrees of income inequality and polarization, do not exist.
This Chapter presents information on public investment on education and health
indicators in Pakistan and uses the principal component analysis to construct six
principal components and identifies the most and least developed districts on the basis
of post-primary school system and hospital index. In the next step, we use an
augmented probit model where in addition to the survey data we employ the selected
principal component variables to examine their association with poverty. Finally, we
study the impact of post-primary school system on poverty under different inequality
and polarization regimes.
4.2 Modeling Education and Health Indices with Principal Components
Therefore, we collect regional level data on a number of education and health
indicators that capture government spending on these sectors, construct indices to
identify most developed and least developed regions and use these indices in
regression analysis to determine their impact on rural poverty.
We use a multivariate statistical weighting approach known as the method of
principal components (PC), which transforms original variables into linear
combinations that explain as much variation as possible with no loss of relevant
information [Manly (1994), Sharma (1996), Greene (1997)]. The weights in the
principal component method are dictated by the data in an objective manner. The first
principal component is the linear combination accounting for the largest variance, the
second principal component accounts for the greatest amount of the remaining
variance, and so on. Only small number of principal components from large number
of variables is chosen that contribute highly in explaining the variance. Components
associated with smallest eigenvalues are discarded because they are least informative.
We adopt Kaiser-Gutman Rule whereby only principal components associated with
eigenvalues greater than one are retained. The retained principal components are
uncorrelated with each other, which is useful in principal component regression
analysis.
4.3 Education and Health Data Used for Principal Component Analysis
We collect district level data of education and health indicators from various
reports of Provincial Development Statistics published by the Bureaus of Statistics of
the Provincial Governments of Punjab, Sindh, Khyber Pakhtunkhwa and Balochistan.
The data on 26 education and health indicators was collected for 56 districts of
Pakistan from 1991–2006. Table 8 shows the 20 indicators chosen for inclusion in the
analysis along with their definition. The first five variables capture the size of school
and college system in each district. Since these institutions are not homogenous in
size and vary by their enrolment levels, we multiply the number of schools/colleges
by their respective enrolments and weigh them by district area and population. Here
we assume that public spending is positively associated with size of these institutions.
The next five variables represent quality of the school system captured by teacher-
student ratios in respective districts. We use teacher-student ratios (instead of student-
teacher ratios) because public spending on teachers is also positively associated with
these ratios just like the size of school and college system while student-teacher ratio
moves in the opposite direction. Public spending on hospitals also varies by their size,
measured by the number of hospital beds. Therefore, we construct hospital size
variable by multiplying the number of hospitals with beds in hospitals weighted by
area and population. The rest of the health indicators are used as numbers and
weighted only by population because their size and coverage is based on a uniform
policy for rural localities.
We were guided by three factors in selecting these variables. Firstly, our policy
interest was to generate education and health infrastructure variables that reflect
government spending at the regional level. Secondly, we considered the availability
and quality of published data on the relevant variables. Finally, we included equal
number of variables for education and health sectors because inclusion of unequal
number of variables may have weighed the PC order in favor of the sector with more
variables. For each of the chosen indicators we had 896 observations from 56 districts
for 16 years.
4.4 Six Principal Components Retained
The principal components were obtained by using the Factor command in Stata
software version 10. After varimax rotation, six principal components were retained
by using Kaiser eigenvalue criterion6 which accounts for 74.47% of total variance.
The six retained principal components along with rotated factor loadings derived from
the correlation matrix of the twenty education and health variables are presented in
Table 9. Numbers in bold indicate high positive coefficient. It needs to be clarified
that the order in which the six principal components are listed in the table only
reflects the order in which they were extracted and it does not necessarily indicate
their relative importance in explaining poverty.
The first factor (PC1) accounts for 22.47% weight in the total variance explained
in the data. PC1 represents principal component basic health, dispensary & rural
health index as it is characterized by high positive loadings on basic health units
(0.934), dispensaries (0.910), rural health centers (0.909), mother child care center
(0.841), rural health center beds (0.789) and dispensary beds (0.567). Except for
hospitals, TB clinics and beds in basic health units, PC1 has high correlations with all
other health variables. The second factor (PC2) was characterized by high positive
factor loadings on degree college size (0.850), hospital size (0.823), high school size
(0.817), intermediate college size (0.667) and middle school size (0.512). It explains
6
Kaiser criterion suggests to retain those factors with eigenvalues equal to or higher than 1.
15.27% of the total variation and represents principal component post-primary school
system and hospital index. The third factor (PC3) represents principal component
teachers in high schools and colleges index as it is characterized by high factor
loadings in teacher student ratio in high schools (0.827) and degree colleges (0.774)
and accounts for a further 12.55% of the total variance. Similarly, the fourth factor
(PC4) represents principal component quality of middle and primary schools with
high factor loadings in teacher student ratio in middle schools (0.824) and teacher
student ratio in primary schools (0.677) and accounts for 12.32% of the total variance.
The fifth factor represents principal component primary school index as it accounts
for 6.71% of the total variance and is characterized by high factor loading in primary
school size (0.986). Finally, the sixth factor represents principal component beds in
TB clinic index with high factor loading (0.961) and accounts for 5.15% of the total
variance.
4.5 Spatial Deprivation: Most and Least Developed Districts
Of all the factors, factor 2 or principal component post-primary school system and
hospital index stands out as the most important measure in terms of coverage and
share in government spending on education and health sectors. This is because the
total number of schools and colleges, and hospitals covered in this category far exceed
the total number of primary schools and other health facilities accounted for by other
factors.
Table 10 lists the 20 most and the 20 least developed districts for 1997-98 and
2005-06 on the basis of scores obtained from principal component post-primary
school system and hospital index, while figures 1 and 2 display complete mapping of
districts on the basis of index for 2005-06 along with development rank (in
parenthesis) for these districts. We can clearly see that the most developed districts
consist of eight metropolitan cities, e.g., Karachi, Lahore, Faisalabad, Multan,
Gujranwala, Rawalpindi-Islamabad, Peshawar and Quetta, and districts in Northern
and Central Punjab. Most of these districts are concentrated by large-scale
manufacturing, SME clusters, upcoming agri-business supply chains (e.g., milk,
citrus, vegetables, etc.), and growing services sector (e.g., commerce & trade,
telecom, transportation, financial services). The thriving economic activity of these
districts places them well ahead of other districts in terms of all economic indicators.
By contrast, the 20 least developed districts mostly consist of districts in Southern
Punjab, interior of Sindh and remotely located districts in KP and Balochistan. In
other words, these districts locate far away from the metropolitan areas, where
employment turns out to be concentrated in agriculture, agro-based industries,
handicrafts and a few resource-based randomly scattered plants (e.g., petroleum
plants, sugar mills, ginning factories, marble & minerals, and the like). It is interesting
to note that our human capital infrastructure based rankings of districts closely
overlap with two other recent studies (e.g., the development rank ordering of districts
by Wasti and Siddiqui (2008) and district-wise index of multiple deprivations by
Jamal et al. (2003). If any thing, these evidences help in understanding the
relationship between infrastructure development and poverty and in determining the
areas of need.
Unfortunately, the non-availability of district-based poverty mapping forces

policy makers in Pakistan to allocate resources across districts on ad hoc basis [Jamal
et al. (2003)]. Perhaps for the same reason, the human capital infrastructure gap
overtime between the most developed districts closes down. In the same period,
however, the gap between the most-developed and the least-developed districts
widens. For example, in 1997-98 in Punjab the gap between the first and the fifth
district was 28-times that fall down to around 12-times in 2005-06 (Table 11).
However, the gap between the most developed and the least developed districts was
increasing in the same period: the gap in PC ranks reported in Tables 11 and 12 was
−1.35 and 6.9 in 1997-98 that increased to −1.72 and 14.45 in 2005-06. These results
suggest that if balanced development is the objective of government policy then the
decision-makers may want to adopt a policy of “geographical targeting” whereby the
development funds are disproportionately allocated to least developed areas on the
basis of non-income poverty mapping of districts. This view has been corroborated by
Jamal et al. (2003) who has previously noted that due to widening spatial deprivation
of Pakistan, geographical targeting of scarce funds in least developed areas may be a
viable option for poverty alleviation.
4.6 How Spatial Deprivation in E&H Affects Poverty
Improvements in public education and health services in a region can help in
poverty alleviation profile of that region. This view has been subscribed to by a
number of recent studies [e.g., Ferreira and Lanjouw (2001), Ravallion and Dutt
(2002), Gustafson and Shi (2004), Klasen (2008), but the magnitude of the effect of
these non-income factors on poverty remains ambiguous. In this sub-section, we
begin our econometric analysis by examining the relationship between six principal
component education and health indicators and poverty. Table 13 presents the
descriptive statistics of the relevant policy variables of interest. We consider the
relationship by estimating an augmented probit model where in addition to the
variables considered in the basic model we also introduce six factor component
education and health indices. We rely on district level factor component education
and health indices that are more apt to be exogenous to a household. Because factor
component education and health indices vary across districts and are correlated with
district dummy variables, we do not include district effects in these regressions.
Moreover, all standard errors account for district level clustering and also allow
general form of heteroskedasticity.
As before, the unit of analysis is the individual level poverty measured by a binary
variable, and we control for household characteristics, individual characteristics and
time-varying province-year dimensions. As Table 14, column 1 reveals,
improvements in local education and health facilities do not produce uniform effects
on poverty alleviation. Only principal component post-primary school system &
hospital index is negatively and significantly correlated with rural poverty: an
increase in this index by 2.48 (its standard deviation) decreases the probability of
poverty by 1.98%. However, the probability of rural poverty is not affected by
increase/decrease in local supply of any other education or health indicator including
primary school index, health facilities through basic health units, dispensaries and
rural health facilities. The impact of increased supply of primary schools and TB
clinics is also negative, but no statistical significance could be attached to these results.
The obvious concern with these findings is whether we can interpret them as the
causal effect of improvements in public education and health facilities on rural
poverty. It may be argued that investments in education and health infrastructure may
be linked through lagged effects on poverty. To address this concern, we also run a
regression where we include only the initial condition education and health
infrastructure variables. Our results reported in column (2) indicate that qualitatively
there is no change in the results, except that principal component beds in TB clinics
index variable also becomes significantly negative. These results indicate that
improvements in the supply of TB clinics have a lagged effect on rural poverty, which
is understandable given the long-term treatment needed for cure in TB disease. For
subsequent analysis we select only principal component post-primary school system
index to examine its effects on poverty under various local level inequality and
polarization regimes.
4.7 Post-Primary School Systems and Poverty: Does Polarization Matter?
A number of recent studies confirm the importance of inequality in efforts to
alleviate poverty in developing countries (Ravallion (1997), Jayaraman and Lanjouw
(1999), Jha (2000), Borooah et al. (2006), Jamal (2006), Ravallion and Chen (1997),
Chambers et al. (2008) and Wang et al. (2009)]. These studies suggest that for given
rates of economic growth, prevalence of higher initial inequality leads to lower
poverty reducing effects. However, the existing literature provides no evidence on
how local level improvements in provision of public schools affect rural poverty
under different inequality and polarization regimes. We take this step and empirically
examine the relationship between district level investments on public schools under
different inequality and polarization regimes and poverty. We introduce three
interaction terms by interacting principle component post-primary school system with
three dummy variables to represent (1) low inequality/polarized districts; (2) medium
inequality/polarized districts, and (3) high inequality/polarized districts. The districts
are placed into low- and high-inequality/polarization regions if they are ½ standard
deviations below or above the mean, respectively. All other districts are classified as
medium inequality/polarization regions.
Table 15 provides the results where the three interaction terms between principal
component post-primary school system × inequality/polarization are the variables of
interest. Column (1) presents the results when districts are segmented by the Gini
index into low-, medium- and high-inequality regions while column (2) evaluates this
effect when districts are classified into low, medium and high-inequality regions on
the basis of initial inequality of 1992-93. In general, the results in column (1) and (2)
provide unequivocal support to the view that investments in post-primary school
system do not produce pro poor effects in high inequality districts. The initial level of
Gini inequality has stronger pro poor effects in low and medium inequality regions.
The magnitude of the effect suggests that a one SD increase in post-primary school
system index leads to a 0.26% reduction in rural poverty in low-inequality districts as
compared with 0.08% poverty reduction in medium-inequality districts. The impact of
present levels of Gini inequality on poverty is relatively weak in our sample where
low inequality districts do not immediately benefit from these investments while the
impact on medium-inequality districts roughly remains unchanged.
In highly polarized districts the society is sharply divided into extremely rich and
extremely poor households where public investments in post-primary school system
tend to benefit the upper income groups and those at the lower end of the distribution
fail to make significant gains. The results of the Wolfson polarization index in column
(3) and (4) confirm that on average school system investments in highly polarized
districts do not affect the probability of poverty reduction, which suggests that public
investments across regions do not bring about uniform returns. This is a powerful
result that has direct policy implications on future school system investments. As
before, starting level of polarization index is relatively more powerful and significant
in explaining the decline in rural poverty in less polarized districts where dominance
of lower and middle income groups allows them to take control of key decisions
impacting their socioeconomic status and in turn their poverty profiles. The results
further suggest that if starting levels of polarization and inequality are not high then
increased local investments in post-primary school system produce roughly similar
long run prospects of poverty alleviation in these districts.
4.8 Conclusions
Overall, columns (2) and (4) suggest that public sector investments on post-
primary school systems are associated with decline in rural poverty throughout all but
highly unequal and highly polarized districts where public sector investments on post-
primary school systems are not associated with declines in rural poverty. These results
suggest that public sector investments do not necessarily lead to welfare gains and
that the returns on investments significantly differ across regions.
5. IMPACT OF ROAD DENSITY ON POVERTY
5.1 Introduction
In the development literature, roads have long been regarded as catalyst of
economic growth [see Rostow (1962)]. The belief on the power of roads to remedy
rural disadvantage is resurfacing [Ali and Pernia (2003)]. Growing concerns about
uneven spatial development in developing countries points to the danger that physical
isolation may be sustaining poverty and vulnerability [Minot et al. (2003)]. It is
logical to assume that investment in road infrastructure could result in more informed
production decisions based on comparative advantage, increased agricultural
productivity, more non-farm employment, higher wages and employment and, in turn,
more welfare [Ali and Pernia (2003)]. Moreover, from another angel, better market
access to rural markets also results in better supply of goods and services and lower
prices that lead to increased well being of the poor. Therefore, some recent studies
suggest that improvements in roads at the regional level can significantly contribute to
the pursuit of socially inclusive growth and be instrumental in alleviating poverty
[see, among others, Khandker et al. (2009), Jacoby and Minten (2009)].
However, the available literature on the impact of road density on rural poverty in
developing countries does not provide a clear picture. Moreover, how the impact of
regional road disparities on poverty varies by regional income inequality and income
polarization is also unclear from the existing literature. In this chapter, we empirically
examine the relationship between local road density and rural poverty under various
regional inequality and polarization regimes. We relate this spatial and inter-temporal
variation in the relative road density to variation in rural poverty using the individual
and household level data of Punjab province from 1992-93 to 2005-06 supplemented
by external information on district level road density for the corresponding data
points. The use of individual and household level data is an attractive alternative to
study the nexus between infrastructure investment, inequality and poverty to studies
that examine this effect on various outcomes with cross-country data.
5.2 Road Density and Spatial Deprivation in Punjab
A consistent time-series data on road density at the district level is not available
from any published source. Part of the problem is that more than a dozen government
institutions at the federal, provincial, district and municipal government level, besides
armed forces and state-owned corporations, are involved in construction and
maintenance of road infrastructure in the country, which makes the data collection
exercise extremely cumbersome. While Provincial Development Statistics
occasionally report district level data on road density, but concerns on its reliability
and consistency make it suspect for all practical purposes. Fortunately, the Punjab
Highway Department maintains a consistent district time-series data on national
highway roads, farm to market roads and district government roads. However, they do
not cover the road network maintained by cantonment boards and defense housing
authorities mostly located in Karachi, Lahore and Islamabad. We obtain district level
road density data of the districts of Punjab from Punjab Highway Department from
1992 to 2006. Since we are concerned about the impact of road density on rural
poverty, we assume that the omission of data on cantonment and DHA roads would
not directly affect poverty incidence of rural people.
Figure 3 displays mapping of the districts of Punjab on the basis of road density
for 2005-06 as well as their relative rank from most dense to least dense. It appears
from there that a policy of unequal government investment has generated regional
concentration of road infrastructure. Many of the districts located in Southern and
Central Punjab, where poverty levels are generally high, are most deprived in road
density. Figure 4 plots road density of 35 districts relative to the road density of
Lahore in 1992-93 and in 2005-06. Here we consider road density of each district
divided by the road density of Lahore. We have chosen Lahore because it had the
highest road density in both 1992-93 and 2005-06 due to which the units of relative
road density can be readily interpreted. Changes in road density across districts are
illustrated by departures from the 45-degree line. The share of Lahore in road density
relative to Lahore is 100%. Districts above the 45-degree line experienced
improvement over time while those below the line experienced decline in their
relative shares. Some Southern districts, e.g., Rajanpur, D.G. Khan, Bhakkar, Layyah
and Rahimyar Khan, with relative road density of less than 40% of Lahore district in
1992-93 experienced no change in their relative road density by 2005-06. Sargodha,
Faisalabad and Rawalpindi districts improved their relative shares from around 70%
of Lahore to more than 80% mainly due to construction of 400 km motorway and
other ancillary roads. Jhang, Chakwal, Muzaffargarh and Mianwali districts are other
noticeable exceptions where relative road density has substantially increased during
this 13-year period. Sialkot, Gujranwala, Sahiwal, Multan and Gujrat districts are
some of the districts that have experienced a large decline in their relative road
density. Sialkot had the second highest road density after Lahore in 1992-93 with road
density equal to 90% of Lahore. By 2005-06, Sialkot’s road density had fallen to less
than 50% of Lahore.
5.3 Empirical Relationship between Road Density and Poverty

Figure 5 sets the scene with a plot relating district level road density of Punjab
with poverty headcount ratio. The measure on the horizontal axis is road density of
each district; the vertical axis shows the poverty headcount ratio of each district
multiplied by 100. In the scatter plot each data point represents a district. With a
correlation of −0.093 the general pattern is clear: the gently downward sloping line
shows a link between road density and poverty headcount at the district level, without
controlling for other factors that may also be affecting poverty ratios.
Next, we estimate probit regressions using the household survey data for
individuals who live in Punjab. Descriptive statistics of the relevant variables
presented in Tables 16 and 17. The sample consists of 139944 individuals. We run
several regressions by varying the empirical specifications to evaluate the effect of
road density on poverty under alternative assumptions. Results presented in Table 18
correspond to the model where road density and starting road density of 1992-93 are
used as variables of interest where the set of regressors include household and
individual controls, sector and employment status dummies and province-year fixed
effects. Column (1) presents our estimation results for road density while column (2)
presents results for starting road density of 1992–93.
Table 18 shows a strong negative relationship between regional road density and
individual poverty: the higher the road density the lower is the probability of poverty.
This negative relationship is statistically significant at the 99% confidence level or
better. The magnitude of the association between road density and poverty is large. A
one standard deviation (0.125) increase in road density decreases the probability of
poverty by 4.2% which is substantially higher than the effect of regional investment
on post-primary school system. Column (2) estimates the impact of initial level or
starting level road density of districts in 1992-93 on long run poverty. Our results
indicate that the starting level road density is associated with a decline in poverty of a
larger magnitude: a one SD increase in starting road density leads to 6.4% decline in
poverty or almost 2.2% point more poverty reduction in the long run than the short
run decline. Thus, it is clear from these results that investments in road infrastructure
are most pro poor.
5.4 The Nexus between Road Density, Inequality and Poverty
Polarization refers to distributional changes that are “more spread out” or where
the distributions have “more hallowed-out middle” or when population becomes
grouped into clusters so that members within each cluster are very similar, but
different across clusters. While Gini coefficient is most commonly used measure of
inequality, it does not allow such clustering within distribution due to which the
polarization and inequality measures diverge and at times move in the opposite
directions (see Tables 3 and 4).
How investment in regional road infrastructure affects rural poverty under
different inequality and polarization regimes? We investigate the effects of changes in
regional road infrastructure on poverty by classifying districts into three categories of
inequality and polarization. Low inequality/polarized districts are those where Gini
index and Wolfson index is 0.5 standard deviations below the mean; high
inequality/polarized refer to districts where Gini index and Wolfson index is 0.5
standard deviations above the mean and medium inequality/polarized districts refer to
all other districts. We incorporate road density in respective districts into the model by
interaction of these three dummy variables with the road density to capture the slope
differentials of road density. Thus the effect of a change in the district level index of
road density on the probability of poverty under alternative inequality/polarization
regimes is captured by these interaction terms.
Regression results shown in column (1) of Table 19 reveal that holding all else as
constant increase in road density significantly decreases the probability of poverty in
all the low, medium and high inequality districts. For example, a one standard
deviation increase in road density (i.e., 0.125) demonstrates highest levels of poverty
reduction in the low-inequality districts (5.8%), followed by the medium-inequality
districts (3.9%) and then the high-inequality districts (3.5%). Column (2) presents the
results of increase in road density with starting levels of inequality and shows that a
one standard deviation increase in road density leads to 9.3%, 4% and 4.3% reduction
in poverty in low-, medium and high-inequality districts, respectively. By implication,
these results suggest that the long run poverty reduction potential of investment in
road infrastructure almost doubles when we move from high-inequality and medium-
inequality districts to low-inequality districts. In other words, policies that support
more equal income distribution offer high gains in poverty alleviation.
In Table 19, columns (3), we present the regression results for the Wolfson index
and see that due to a one standard deviation increase in road density, the probability of
poverty reduction in less polarized districts is highest (6.3%), followed by medium
polarized districts (3.7%) and then highly-polarized districts (3.75%). If anything, the
results in column (3) when compared with results in column (1) show that less
polarized districts have more potential to reduce poverty than less unequal districts.
Similarly, column (4) captures the long run effects of road density with starting levels
of polarization. Here we find that one standard deviation increase in road density in
less polarized districts leads to highest probability of poverty reduction of 10%. In
other words, the long run effects of road density far exceed the short run effects.
Moreover, poverty reduction profile of less polarized districts is far greater than less
unequal districts.
5.5 Conclusions
In sum this chapter suggests that persistent unequal allocation of funds on road
network has led to spatial concentration of road infrastructure in the Northern districts
of Punjab at the cost of most deprived Southern districts where income poverty levels
are generally high. Trends for the last 13-years suggest that most deprived districts
with their relative road density of less than 40% of Lahore in 1992-93 have
experienced no change in their relative share. Our empirical results suggest a strong
negative relationship between road density and probability of poverty, however, long
run decline in poverty far exceeds the short run decline. Our results further suggest
that the long run poverty alleviation potential of investment on roads almost doubles
when we move from high-inequality/polarized and medium-inequality/polarized
districts to low-inequality/polarized districts. In addition, we find that the poverty
reduction potential of less polarized districts far exceeds the poverty reduction
potential of less unequal districts. Therefore, public policies that seek more regional
equality and less polarization are desirable.
6. CONCLUDING COMMENTS
The main purpose of this study was to explore how spatial variation in public
sector infrastructural investments affects rural poverty under various inequality and
polarization regimes. We use data from six rounds of Pakistan’s household survey
from 1992-93 to 2005-06 to examine the incidence of rural poverty, inequality and
polarization and the determinants of rural poverty. As expected, the percentage of
rural people living below the poverty line change over survey years. Gini inequality
index and Wolfson polarization index do not always move in the same direction,
which confirms that there are indeed issues of distribution change that are missed out
by the inequality index but are better described by the polarization index.
We also investigate that given regional infrastructural investments how variation
in inequality and polarization affect poverty reduction potential. Using six rounds of
the household survey consisting of 349295 individual observations and a probit
maximum likelihood model our basic results strongly support the general view that,
among other things, poverty in rural Pakistan coexists with illiteracy of household
heads and lack of asset ownership of the households. Policies that seek universal
access to education by increasing the quantity and quality of schools and colleges
seem to have a strong power to reduce rural poverty.
We also map the spatial deprivation on the basis of principal component post-
primary school system and hospital index, which shows that most developed districts
are located near the metropolitan cities while most deprived districts are remotely
located. Moreover, the education and health infrastructure gap between the developed
districts is closing down while the same gap between the least developed and most
developed districts is increasing over time.
Our econometric results suggest that increased public sector investments on
education and health infrastructure are associated with decline in rural poverty
throughout all but highly unequal and polarized districts where public sector
investments on post-primary school systems are not associated with declines in rural
poverty. Moreover, these results suggest that public sector investments do not
necessarily lead to welfare gains and that the returns on investments significantly
differ across regions. Likewise, persistent unequal allocation of funds on road
network produces spatial concentration of road infrastructure in few developed
districts at the cost of most deprived districts where otherwise income poverty levels
are generally high. Our results suggest a strong negative relationship between road
density and probability of poverty, however, the long run decline in poverty far
exceeds the short run decline. The long run poverty alleviation potential of investment
on roads almost doubles when we move from high-inequality/polarized and medium-
inequality/polarized districts to low-inequality/polarized districts. It is also significant
to note that with investment on road network, the poverty reduction potential of less
polarized districts far exceeds the poverty reduction potential of less unequal districts:
a confirmation that issues of distributional change are indeed missed out by inequality
index. Therefore, public policies that seek more regional equality and less polarization
are desirable for pro poor growth policies.
7. REFERENCES
Ali, I., E.M. Pernia (2003). Infrastructure and Poverty Reduction: What is the Connection? ERD
Policy Brief Series, Economics and Research Department, No.13, Manila: Asian Development
Bank.
Anwar, T. (2005). Long Term Changes in Income Distribution in Pakistan: Evidence Based on
Consistent Series Estimates. Discussion Paper Series No.3. Islamabad: Centre for Research on
Poverty and Income Distribution.
Anwar, T., S.K. Qureshi (2002). Trends in Absolute Poverty in Pakistan: 1990–2001. Pakistan
Development Review, 36(1), 39–68.
Arif, G.M. (2006). Targeting Efficiency of Poverty Reduction Programs in Pakistan. Working Paper
No. 4. Islamabad: Asian Development Bank. 1–70.
Atkinson, A.B. (1970). On the Measurement of Inequality, Journal of Economic Theory, 2.
Basu, K. (2006). Globalization, Poverty, and Inequality: What is the Relationship? What Can be Done?
World Development, 34 (8), 1361–1373.
Besley, T., R. Burgess, B.Esteve-Volart (2005). Operationalising Pro-Poor Growth: India Case Study,
Mimeo.
Binswanger, Hans P., J. von Braun. (1991). Technological Change and Commercialization in
Agriculture: The Effect on the Poor. The World Bank Research Observer, 6 (1), 57–80.
Borooah, Vani K., B. Gustafsson, L. Shi (2006). China and India: Income Inequality and Poverty North
and South of the Himalayas. Journal of Asian Economics, 17, 797–817.
Burki, Abid A., Mushtaq A. Khan (2008). Impact of Higher Wheat Prices on Poverty in Pakistan:
Futuristic of Food Security, Paper in National Conference on Socioeconomic Challenges
Faced by Pakistan. Islamabad: International Islamic University, June 2–3.
Chambers, D., Y. Wu, H. Yao (2008). The Impact of Past Growth on Poverty in Chinese Provinces.
Journal of Asian Economics. 19, 348–357.
Duclos, J.-Y., D. Echevin (2005). Bi-polarization Comparisons, Economics Letters, 87, 249–258.
Duclos, J.-Y., J. Esteban, D. Ray (2004). Polarization: Concepts, Measurement, Estimation.
Econometrica, 72(6), 1737–1772.
Duro, Juan A. (2005). Another Look to Income Polarization Across Countries, Journal of Policy
Modeling, 27, 1001–1007.
Ercelawn, Aly A. (1986). Poverty in Pakistan: A Study of Villages. In I. Nabi (ed.) The Quality of Life
in Pakistan. Lahore: Vanguard.
Ercelawn, Aly A. (1989). Poverty in Pakistan: Choice of Poverty Criteria. University of Karachi,
Applied Economics Research Centre.
Ercelawn, Aly A. (1990). Absolute Poverty in Pakistan: Poverty Lines, Incidence, Intensity. University
of Karchi, Applied Economics Research Centre.
Esteban, J., D. Ray (1994). On the Measurement of Polarization. Econometrica, 62(4), 819–851.
Fan, S., P. Hazell, S. Thorat (2000). Government Spending, Growth and Poverty in Rural India.
American Journal of Agricultural Economics, 82 (4), 1038–1051.
Fedorov, L. (2002). Regional Inequality and Regional Polarization in Russia, 1990–99, World
Development, 30(3), 443–456.
Ferreira, Francisco H.G., P. Lanjouw (2001). Rural Non-Farm Activities and Poverty in the Brazalian
NorthEast. World Development, 29 (3), 509–528.
Foster, J., J. Greer, E. Thorbecke (1984). A Class of Decomposable Poverty Measures. Econometrica,
52(3), 76–166.
Government of Pakistan (2007). Pakistan Economic Survey, 2006-07, Islamabad: Finance Division,
Economic Adviser’s Wing.
Gradin, C. (2000). Polarization by Sub-populations in Spain, 1973–91, Review of Income and Wealth,
46(4), 457–474.
Graff, G., D. Roland-Holst, D. Zilberman (2006). Agricultural Biotechnology and Poverty Reduction in
Low-income Countries. World Development, 34 (8), 1430–1445.
Greene, W.H. (1997). Econometric Analysis, New Jersy: Prentice Hall International.
Gustaffson, B., L. Shi (2004). Expenditures on Education and Health Care and Poverty in Rural China.
China Economic Review, 15, 292–301.
Haq, Zahoor ul, H. Nazli, K. Meilke (2008). Implications of High Food Prices for Poverty in Pakistan,
Agricultural Economics, 39(1) Supplement 1, Special Issue on the World Food Crisis, 477–
484.
Harrison, Glenn W., Thomas F. Rutherford and David G. Tarr (2003). Trade Liberalization, Poverty
and Efficient Equity. Journal of Development Economics, 71(1), 97–128.
Irfan, M., R. Amjad (1984). Poverty in Rural Pakistan. In Azizur Rehman Khan and Eddy Lee (eds.),
Poverty in Rural Asia. Bangkok: International Labour Office/Asian Employment Programme.
Jacoby, H.G., B. Minten (2009). On Measuring the Benefits of Lower Transport Costs, Journal of
Development Economics, 89, 28–38.
Jafri, S.M. Younis (1999). Assessing Poverty in Pakistan. In A Profile of Poverty in Pakistan.
Islamabad: Mahbub-ul-Haq Centre for Human Development and UNDP.
Jamal, H. (2003). Poverty and Inequality during the Adjustment Decade: Empirical Findings from
Household Surveys, Pakistan Development Review, 42(2), 125–136.
Jamal, H. (2006). Does Inequality Matter for Poverty Reduction? Evidence from Pakistan’s Poverty
Trends, Pakistan Development Review, 45(3), 439–459.
Jamal, H., A.J. Khan, I.A. Toor, N. Amir (2003). Mapping the Spatial Deprivation of Pakistan,
Pakistan Development Review, 42(2), 91–111.
Jayaraman, R., P. Lanjouw (1999). The Evolution of Poverty and Inequality in Indian Villages. Oxford
University Press.
Jha, R. (2000). Growth, Inequality and Poverty in India: Spatial and Temporal Characteristics.
Economic and Political Weekly, 35(11), 921–928.
Khandker, S.R., Z. Bakht, G. B. Koolwal (2009). The Poverty Impact of Rural Roads: Evidence from
Bangladesh, Economic Development and Cultural Change, 57, 685–722.
Klasen S. (2008). Economic Growth and Poverty Reduction: Measurement Issues using Income and
Non-income Indicators. World Development, 36 (3), 420–445.
Kraay, A. (2006). When is Growth Pro-poor? Evidence from a Panel of Countries. Journal of
Development Economics, 80 (1), 198–227.
Kristjanson, P., A.Krishna, M. Rageny, J. Kuan, G. Quilca, A. Sanchez-Urrelo, C. Velarde (2007).
Poverty Dynamics and the Role of Livestock in the Peruvian Andes, Agricultural Systems, 94,
294–308.
Kruger, Diana I. (2007). Coffee Production Effects on Child Labor and Schooling in Rural Brazil,
Journal of Development Economics, 82, 448–446.
Malik, Sohail J. (1993). Poverty in Pakistan, 1984-85 to 1987-88. In M. Lipton and J. Van der Gaag
(eds.) Including the Poor. Washington, D.C.: The World Bank.
Malik, Sohail J. (2005). Agricultural Growth and Rural Poverty: A Review of the Evidence. Pakistan
Resident Mission Working Paper No. 2, Islamabad.
Manly, B. (1994). Multivariate Statistical Methods: A Primer, London: Chapman and Hall.
Minot, N., Baulch, B. and Epprecht, M. (2003) Poverty and Inequality in Vietnam: Spatial Patterns
and Geographic Determinants. Washington, DC: IFPRI and Brighton: IDS.
Minot, N., F. Goletti. (1998). Export Liberalization and Householf Welfare: The Case of Rice in
Vietnam. American Journal of Agricultural Economics, 80 (4), 738–749.
Muller, C. (2002). Prices and Living Standards: Evidence from Rwanda. Journal of Development
Economics, 68, 187–203.
Naseem, S.M. (1973). Mass Poverty in Pakistan: Some Preliminary Findings, Pakistan Development
Review, 13(4), 317–360.
Naseem, S.M. (1977). Rural Poverty and Landlessness in Pakistan, In ILO Report on Poverty and
Landlessness in Asia. Geneva: International Labor Organization.
Peacock, C. (2005). Goats: A Pathway out of Poverty. Small Ruminant Research, 60, 179–186.
Ravallion, M. (1997). Can High Inequality Developing Countries Escape Absolute Poverty?
Economics Letters, 56, 51–57.
Ravallion, M. (2002). On the Urbanization of Poverty. Journal of Development Economics, 68, 435–
442.
Ravallion, M., G. Datt (2002). Why Has Economic Growth Been More Pro-poor in Some States of
India than Others? Journal of Development Economics, 68, 381–400.
Ravallion, M., S. Chen (1997). What Can New Survey Data Tell Us About Recent Changes in
Distribution and Poverty? The World Bank Economic Review, 11 (2), 357–382.
Ravallion, M., S. Chen (2003). Measuring pro-poor growth. Economics Letters, 78, 93–99.
Rostow, W. W. (1962) The Process of Economic Growth. New York: W. W. Norton and Co.
Seshanna, S., S. Decornez (2003). Income Polarization and Inequality Across Countries: An Empirical
Study, Journal of Policy Modeling, 25, 335–358.
Sharma, S. (1996). Applied Multivariate Techniques, New York: Wiley.
Suryahadi, A., D. Suryadarma, S. Sumarto (2009). The Effects of Location and Sectoral Components
of Economic Growth on Poverty: Evidence from Indonesia. Journal of Development Economics,
89, 109–117.
Wagle, Udaya R. (2010). Does Low Inequality Cause Low Poverty? Evidence from High-Income and
Developing Countries, Poverty and Public Policy, 2(3), 29–52.
Wang, Z., R. Smyth, Y-K. NG. (2009). A New Ordered Family of Lorenz Curves with an Application
to Measuring Income Inequality and Poverty in Rural China. China Economic Review,
doi:10.1016/j.chieco.2008.12.003.
Wasti, S.A., M.U. Siddiqui (2008). Development Rank Ordering of Districts of Pakistan: Revisited,
Pakistan Journal of Applied Economics, 18(1&2), 1–29.
Winters, L. Alan, N. McCulloch, A. McKay (2004). Trade Liberalization and Poverty: The Evidence so
Far, Journal of Economic Literature, 42 (1), 72 – 115.
Wolfson, Michael C. (1994). When Inequalities Diverge? American Economic Review, 84(2), 353–358.
Wolfson, Michael C. (1997). Divergent Inequalities: Theory and Empirical Results, Review of Income
and Wealth, 43(4), 401–421.
Zhang, X., R. Kanbur (2001). What Differences Do Polarization Measures Make? An Application to
China, Journal of Development Studies, 37, 85–98.
TABLE 1
Distribution of Rural Sample by Household Survey
Survey year Individual observations
HIES 1992-93 57725
HIES 1993-94 47638
HIES 1996-97 45310
HIES-PIHS 1998-99 65029
HIES-PIHS 2000-01 66445
PSLM 2005-06 67148
Total sample 349295
TABLE 2
Inflation-Adjusted Poverty Lines Based on Cut off of 2350
Calories per Adult Equivalent per Day
Survey year/type Poverty line (Rs)

HIES 1992-93 400
HIES 1993-94 501
HIES 1996-97 618
HIES-PIHS 1998-99 673.40
HIES-PIHS 2001-02 723.40
PSLM 2005-06 948.47
TABLE 3
Changes in Poverty, Inequality and Polarization in Pakistan
1992-93 1993-94 1996-97 1998-99 2000-01 2005-06
Poverty Headcount ratio 0.351 0.370 0.292 0.462 0.391 0.273
Gini index 0.321 0.325 0.339 0.343 0.304 0.305
Wolfson Index 0.114 0.115 0.108 0.126 0.116 0.120
TABLE 4
Changes in Poverty, Inequality and Polarization in Punjab
1992-93 1993-94 1996-97 1998-99 2001-02 2005-06

Poverty Headcount ratio 0.356 0.356 0.289 0.481 0.361 0.214
Gini index 0.326 0.334 0.348 0.348 0.300 0.304
Wolfson Index 0.118 0.118 0.107 0.129 0.121 0.119
TABLE 5
Definition of Variables
Variables Description
Dependent variable
Poverty =1 if individual falls below the basic need poverty line
Independent variables
Household level controls:
Head is illiterate =1 if individual is from a household where head is
illiterate
Lowly educated head, 1 – 5 years =1 if individual is from a household where head has 1–
5 years of schooling
Medium educated head, 6 – 10 years =1 if individual is from a household where head has 6–
10 years of schooling
Highly educated head, 11 years or more =1 if individual is from a household where head has 11
or more years of schooling
Small sized household (1–3 members) =1 if individual belongs to a household comprising of
1–3 members
Medium sized household (4–8 members) =1 if individual belongs to a household comprising of
4–8 members
Large sized household ( 9 or more =1 if individual belongs to a household comprising of
members) more than 8 members
No land =1 if individual belongs to a household that does not
have any inherited or rented-in land
Household has subsistence landholding, 1– =1if individual belongs to a household that has 1–12.5
12.5 acres acres of inherited or rented-in land
Household has economical landholding, =1if individual belongs to a household that has 12.6–25
12.6–25 acres acres of inherited or rented-in land
Household has large landholding, more =1if individual belongs to a household that has more
than 25 acres than 25 acres of inherited or rented-in land
Livestock owned (yes=1, no=0) =1 if individual belongs to a household that owns
livestock
Farm equipment owned (yes=1, no=0) =1 if the individual belongs to a household that owns
farm equipment
Financial assets owned by household (Rs. Value of financial assets owned by an individual’s
million) household in millions of rupees
Individual level controls
Age Age of the individual in years
Male =1 if an individual is a male
Female =1 if an individual is a female
Days worked during last month Number of days worked by the individual during the
last month
Employment in agriculture or forestry =1 if individual is employed in agriculture or forestry
sector
Employment in mining =1 if individual is employed in mining sector
Employment in electricity or gas =1 if individual is employed in electricity or gas supply
sector
Employment in manufacturing =1 if individual is employed in manufacturing sector
Employment in construction =1 if individual is employed in construction sector
Cont. Table 5
Employment in wholesale or retail trader =1 if individual is employed in wholesale and retail
trading sector
Employment in transport =1 if individual is employed in transport sector
Employment in real estate =1 if individual is employed in real estate sector
Employment in social or personal services =1 if individual is employed in provision of social or
personal services
Unemployed =1 if individual is unemployed
Self employed =1 if individual is self-employed
Paid employee =1 if individual is a paid employee
Unpaid family worker =1 if individual is an unpaid family worker
No employment status =1 if individual has no employment status
Province-year Effects
Survey year 1992-93 × Punjab =1 if individual belongs to the province of Punjab and
the survey year is 1992-93
Survey year 1992-93 × Sindh =1 if individual belongs to the province of Sindh and
Survey year 1992-93 × Balochistan =1 if individual belongs to the province of Balochistan
and the survey year is 1992-93
Survey year 1992-93 × NWFP =1 if individual belongs to the province of NWFP and
Cont. Table 5
Survey year 2001-02 × Balochistan =1 if individual belongs to the province of
Balochistan and the survey year is 2001-02
Survey year 2005-06 × Balochistan =1 if individual belongs to the province of
Balochistan and the survey year is 2005-06
District effects
Abbottabad =1 if individual belongs to Abbottabad district
Attock =1 if individual belongs to Attock district
Badin =1 if individual belongs to Badin district
Bahawalnagar =1 if individual belongs to Bahawalnagar district
Bahawalpur =1 if individual belongs to Bahawalpur district
Bannu =1 if individual belongs to Bannu district
Bhakkar =1 if individual belongs to Bhakkar district
Chakwal =1 if individual belongs to Chakwal district
Dadu =1 if individual belongs to Dadu district
Dera Ghazi Khan =1 if individual belongs to Dera Ghazi Khan district
Dera Ismail Khan =1 if individual belongs to Dera Ismail Khan district
Dir =1 if individual belongs to Dir district
Faisalabad =1 if individual belongs to Faisalabad district
Gujranwala =1 if individual belongs to Gujranwala district
Gujrat =1 if individual belongs to Gujrat district
Hyderabad =1 if individual belongs to Hyderabad district
Jacobabad =1 if individual belongs to Jacobabad district
Jhang =1 if individual belongs to Jhang district
Jhelum =1 if individual belongs to Jhelum district
Kalat =1 if individual belongs to Kalat district
Karachi East =1 if individual belongs to Karachi East district
Karak =1 if individual belongs to Karak district
Kasur =1 if individual belongs to Kasur district
Khairpur =1 if individual belongs to Khairpur district
Khanewal =1 if individual belongs to Khanewal district
Khushab =1 if individual belongs to Khushab district
Kohat =1 if individual belongs to Kohat district
Lahore =1 if individual belongs to Lahore district
Larkana =1 if individual belongs to Larkana district
Layyah =1 if individual belongs to Layyah district
Mansehra =1 if individual belongs to Mansehra district
Cont. Table 5
Mardan =1 if individual belongs to Mardan district
Mekran =1 if individual belongs to Mekran district
Mianwali =1 if individual belongs to Mianwali district
Multan =1 if individual belongs to Multan district
Muzzafargarh =1 if individual belongs to Muzzafargarh district
Nawabshah =1 if individual belongs to Nawabshah district
Okara =1 if individual belongs to Okara district
Peshawar =1 if individual belongs to Peshawar district
Quetta =1 if individual belongs to Quetta district
Rahimyar Khan =1 if individual belongs to Rahimyar Khan district
Rajanpur =1 if individual belongs to Rajanpur district
Rawalpindi =1 if individual belongs to Rawalpindi district
Sahiwal =1 if individual belongs to Sahiwal district
Sanghar =1 if individual belongs to Sanghar district
Sargodha =1 if individual belongs to Sargodha district
Sheikhupura =1 if individual belongs to Sheikhupura district
Shikarpur =1 if individual belongs to Shikarpur district
Sialkot =1 if individual belongs to Sialkot district
Sibi =1 if individual belongs to Sibi district
Sukkur =1 if individual belongs to Sukkur district
Swat =1 if individual belongs to Swat district
Tharparker =1 if individual belongs to Tharparker district
Thatta =1 if individual belongs to Thatta district
Toba Tek Singh =1 if individual belongs to Toba Tek Singh district
Vehari =1 if individual belongs to Vehari district
TABLE 6
Descriptive Statistics – Full Sample
Explanatory variables Mean Std. Min Max
Dev.
Poverty 0.351 0.47 0 1
Head is illiterate 0.630 0.48 0 1
Lowly educated head, 1 – 5 years (yes=1, no=0) 0.16 0.37 0 1
Medium educated head, 6 – 10 years (yes=1, no=0) 0.150 0.357 0 1
Highly educated head, 11 years or more (yes=1, no=0) 0.049 0.217 0 1
Medium sized household, 4 – 8 members (yes=1, no=0) 0.553 0.497 0 1
Large sized household, 9 or more members (yes=1, 0.379 0.485 0 1
no=0)
Household has subsistence landholding, 1 – 12.5 acres 0.305 0.460 0 1
(yes=1, no=0)
Household has economical landholding, 12.6 – 25 acres 0.048 0.213 0 1
(yes=1, no=0)
Household has large landholding, more than 26 acres 0.169 0.128 0 1
(yes=1, no=0)
Cont. Table 6

Dev.
Livestock owned (yes=1, no=0) 0.297 0.45 0 1
Farm equipment owned (yes=1, no=0) 0.007 0.083 0 1
Financial assets owned by household (Rs. million) 0.021 0.128 0 9.734
Individual level controls
Age 22.83 18.68 0 99
Male (yes=1, no=0) 0.510 0.499 0 1
Days worked during last month 6.517 11.528 0 30
Employment in agriculture or forestry (yes=1, no=0) 0.171 0.377 0 1
Employment in mining (yes=1, no=0) 0.0009 0.312 0 1
Employment in manufacturing (yes=1, no=0) 0.019 0.136 0 1
Employment in electricity or gas (yes=1, no=0) 0.0017 0.041 0 1
Employment in construction (yes=1, no=0) 0.022 0.146 0 1
Employment in wholesale or retail trader (yes=1, no=0) 0.0233 0.151 0 1
Employment in transport (yes=1, no=0) 0.0141 0.118 0 1
Employment in real estate (yes=1, no=0) 0.0008 0.029 0 1
Employment in social or personal services (yes=1, no=0) 0.035 0.184 0 1
Self employed (yes=1, no=0) 0.105 0.307 0 1
Paid employee (yes=1, no=0) 0.098 0.297 0 1
Unpaid family worker (yes=1, no=0) 0.089 0.285 0 1
Survey year 1992-93 × Balochistan 0.018 0.135 0 1
Survey year 1992-93 × NWFP 0.036 0.186 0 1
Survey year 1992-93 × Punjab 0.073 0.260 0 1
Survey year 1992-93 × Sindh 0.037 0.189 0 1
Survey year 1993-94 × NWFP 0.030 0.172 0 1
Survey year 1996-97 × NWFP 0.031 0.174 0 1
Survey year 1998-99 × NWFP 0.042 0.201 0 1
Survey year 2001-02 × NWFP 0.041 0.198 0 1
Survey year 2005-06 × NWFP 0.045 0.208 0 1
Cont. Table 6

Dev.
Abbottabad 0.023 0.150 0 1
Attock 0.011 0.104 0 1
Badin 0.012 0.112 0 1
Bahawalnagar 0.012 0.109 0 1
Bahawalpur 0.0148 0.120 0 1
Bannu 0.016 0.128 0 1
Bhakkar 0.010 0.101 0 1
Chakwal 0.006 0.081 0 1
Dadu 0.018 0.135 0 1
Dera Ghazi Khan 0.010 0.102 0 1
Dera Ismail Khan 0.017 0.130 0 1
Dir 0.028 0.166 0 1
Faisalabad 0.02 0.145 0 1
Gujranwala 0.015 0.133 0 1
Gujrat 0.0202 0.140 0 1
Hyderabad 0.017 0.132 0 1
Jacobabad 0.019 0.137 0 1
Jhang 0.017 0.132 0 1
Jhelum 0.006 0.077 0 1
Kalat 0.031 0.173 0 1
Karachi East 0.009 0.096 0 1
Karak 0.009 0.097 0 1
Kasur 0.014 0.118 0 1
Khairpur 0.017 0.130 0 1
Khanewal 0.010 0.101 0 1
Khushab 0.007 0.085 0 1
Kohat 0.012 0.111 0 1
Lahore 0.008 0.090 0 1
Larkana 0.0216 0.145 0 1
Layyah 0.006 0.078 0 1
Mansehra 0.019 0.139 0 1
Mardan 0.025 0.156 0 1
Mekran 0.019 0.136 0 1
Mianwali 0.007 0.086 0 1
Multan 0.019 0.138 0 1
Muzzafargarh 0.015 00.122 0 1
Nawabshah 0.027 0.162 0 1
Okara 0.015 0.122 0 1
Peshawar 0.028 0.166 0 1
Quetta 0.052 0.223 0 1
Rahimyar Khan 0.020 0.140 0 1
Cont. Table 6

Dev.
Rajanpur 0.005 0.073 0 1
Rawalpindi 0.020 0.142 0 1
Sahiwal 0.020 0.140 0 1
Sanghar 0.015 0.122 0 1
Sargodha 0.014 0.118 0 1
Sheikhupura 0.019 0.136 0 1
Shikarpur 0.011 0.106 0 1
Sialkot 0.022 0.149 0 1
Sibi 0.037 0.188 0 1
Sukkur 0.022 0.147 0 1
Swat 0.045 0.209 0 1
Tharparker 0.025 0.158 0 1
Thatta 0.012 0.112 0 1
Toba Tek Singh 0.012 0.109 0 1
Vehari 0.012 0.109 0 1
Number of observations 349295 -- -- --
TABLE 7
Effects of Household and Individual Characteristics on Rural
Poverty in Pakistan, Basic Results
Probit regressions (average marginal effects)
Explanatory variables
(1) (2) (3) (4)
Lowly educated head, 1–5 years (yes=1, no=0) -0.101*** -0.089*** -0.093*** -0.089***
(-11.71) (-12.38) (-11.64) (-11.49)
Medium educated head, 6–10 years (yes=1, no=0) -1.166*** -0.159*** -0.161*** -0.155***
(-17.65) (-18.48) (-17.80) (-17.72)
Highly educated head, 11 years or more -0.256*** -0.252*** -0.251*** -0.249***
(yes=1, no=0) (-19.12) (-19.00) (-20.21) (-20.48)
Medium sized household, 4–8 members 0.347*** 0.355*** 0.347*** 0.356***
(yes=1, no=0) (33.71) (34.98) (31.64) (32.45)
Large sized household, 9 or more members 0.574*** 0.584*** 0.574*** 0.585***
(yes=1, no=0) (36.74) (40.83) (39.74) (40.57)
Household has subsistence landholding, 1–12.5 acres -0.109*** -0.103*** -0.095*** -0.092***
(yes=1, no=0) (-8.49) (-7.85) (-7.69) (-7.51)
Household has economical landholding, 12.6–25 acres -0.148*** -0.159*** -0.146*** -0.157***
(yes=1, no=0) (-5.60) (-7.26) (-7.03) (-7.88)
Household has large landholding, more than 25 acres -0.193*** -0.205*** -0.184*** -0.198***
(yes=1, no=0) (-5.93) (-6.99) (-5.47) (-6.51)
Livestock owned (yes=1, no=0) -0.034** -0.048*** -0.041*** -0.051***
(-2.45) (-4.04) (-3.63) (-4.33)
Farm equipment owned (yes=1, no=0) -0.101** -0.098** -0.098*** -0.097***
(-2.38) (-2.16) (-2.96) (-2.79)
Financial assets owned by HH (Rs. million) -1.74*** -1.784*** -1.777*** -1.738***
(-6.19) (-5.88) (-5.81) (-5.49)
Age (years) -0.0009*** -0.0008*** -0.0008*** -0.0008***
(-9.39) (-9.26) (-11.02) (-10.07)
Male (yes=1, no=0) -0.007*** -0.007*** -0.007*** -0.006***
(-3.17) (-3.21) (-3.42) (-3.11)
Days worked during last month -0.0004 -0.0002 -0.0002 -0.0002
(-0.58) (-0.40) (-0.43) (-0.70)
12 sectors and employment status indicators Yes Yes Yes Yes
District level controls included No Yes No Yes
Province × year fixed effects included No No Yes Yes
Pseudo R2 0.109 0.126 0.129 0.153
Number of observations 349295 349295 349295 349295
Notes: All regressions are estimated by probit maximum likelihood. Numbers in parenthesis are asymptotic t-values obtained from robust standard
errors adjusted for clustering at the district level. The regressions include intercept terms, but they are not reported. **, and *** denote
statistical significance at the 5 and 1% levels, respectively. Sector of employment indicators are agriculture & forestry; mining;
manufacturing; electricity or gas; construction; wholesale and retail trade; transport; real estate; and social and personal services while
employment status indicators are self-employed, paid-employees and unpaid family workers. Province*year fixed effects are 24 province-
year interaction terms to introduce province level trends, which capture variation in economic conditions at the province level during each
year of the survey that may affect rural poverty. District fixed effects are specified by introducing district dummy variables, one each for
each district, which are meant to capture the effects of time-invariant district specific characteristics on poverty in our sample.
TABLE 8
District Level Variables Used in Principal Component Analysis
Explanatory variable Definition
1 Primary school size Number of primary schools × enrollment in

primary schools/(district area × district
population)
2 Middle school size Number of middle schools × enrollment in middle
schools/(district area × district population)
3 High school size Number of high schools × enrollment in high
schools/(district area × district population)
4 Intermediate college size Number of intermediate colleges × enrollment in
intermediate colleges/(district area × district
population)
5 Degree college size Number of degree colleges × enrollment in degree
colleges/(district area × district population)
6 Teacher student ratio in primary schools Teaching staff in primary schools/enrollment in
primary schools
7 Teacher student ratio in middle schools Teaching staff in middle schools/enrollment in
middle schools
8 Teacher student ratio in high schools Teaching staff in high schools/enrollment in high
schools
9 Teacher student ratio in intermediate Teaching staff in intermediate
colleges colleges/enrollment in intermediate colleges
10 Teacher student ratio in degree colleges Teaching staff in degree colleges/enrollment in
degree colleges
11 Hospital size Number of hospitals × beds in hospitals/(district
area × district population)
12 Dispensaries per 10,000 population Number of dispensaries/district population
13 Dispensary beds per 10,000 population Beds in dispensaries/district population
14 Rural health centers per 10,000 population Number of rural health centers/district population
15 Rural health center beds per 10,000 Beds in rural health centers/district population
population
16 T.B. clinics per 10,000 population Number of TB clinics/district population
17 T.B. clinic beds per 10,000 population Beds in TB clinics/district population
18 Basic health units per 10,000 population Number of basic health units/district population
19 Basic health units beds per 10,000 Beds in BHUs/district population
population
20 Mother-child care centers per 10,000 Number of mother-child care centers/district
population population
TABLE 9
Principal Component Rotated Factor Loadings
Explanatory Variable PC1 PC2 PC3 PC4 PC5 PC6
Primary school size -0.180 0.060 0.047 -0.024 0.896 0.032

Middle school size 0.030 0.512 -0.230 -0.599 0.368 -0.042
High school size 0.013 0.817 -0.184 -0.222 0.171 -0.052
Intermediate college size -0.047 0.667 0.088 0.056 0.239 0.048
Degree college size -0.159 0.850 0.046 0.074 -0.115 0.012
Teacher student ratio in primary schools 0.046 0.103 -0.272 0.677 -0.191 0.145
Teacher student ratio in middle schools 0.215 -0.041 -0.052 0.824 -0.060 -0.061
Teacher student ratio in high schools 0.153 -0.119 0.827 0.0186 0.039 -0.013
Teacher student ratio in intermediate 0.002 -0.008 -0.274 -0.199 -0.378 -0.137
colleges
Teacher student ratio in degree colleges 0.279 -0.086 0.774 -0.076 -0.024 -0.033
Hospital size -0.055 0.823 0.088 -0.069 -0.059 -0.043
Dispensaries per 10,000 population 0.910 0.023 0.229 0.135 -0.057 -0.048
Dispensary beds per 10,000 population 0.567 0.015 0.369 -0.401 -0.163 0.148
Rural health centers per 10,000 0.909 -0.169 0.185 0.140 -0.015 -0.007
population
Rural health center beds per 10,000 0.789 -0.182 -0.268 0.093 -0.091 0.049
population
T.b. clinics per 10,000 population 0.265 -0.105 -0.099 0.784 0.265 -0.108
T.b. clinic beds per 10,000 population -0.011 -0.036 0.021 -0.039 0.029 0.961
Basic health units per 10,000 population 0.934 -0.156 0.120 0.063 -0.035 0.012
Basic health units beds per 10,000 -0.181 -0.210 -0.784 0.171 -0.049 -0.104
population
Mother-child care centers per 10,000 0.841 0.314 0.185 0.042 -0.053 -0.039
population
Proportion (relative weight of each 22.47 15.27 12.55 12.32 6.71 5.15
factor in the total variance)
Note: Numbers in bold face indicate high positive loadings
TABLE 10
Most and Least Developed Districts in Pakistan Based on Principal
Component Post-Primary School System and Hospital Index
Rank/city Principal component value, Rank/city Principal component value,
1997-98 2005-06
20 most developed districts
Lahore 14.85 Lahore 20.232
Karachi 10.59 Karachi 8.552
Rawalpindi 3.57 Rawalpindi 7.420
Hyderabad 2.47 Faisalabad 2.105
Peshawar 1.494 T.T singh 1.511
Sibi 1.452 Sibi 1.201
Faisalabad 0.976 Peshawar 0.700
Jhelum 0.423 Abbottabad 0.583
Multan 0.377 Gujrat 0.513
Thatta 0.230 Jhelum 0.503
Quetta 0.216 Multan 0.405
Sargodha -0.006 Quetta 0.354
Attock -0.158 Sargodha 0.276
Sialkot -0.184 Sialkot 0.245
Abbottabad -0.201 Gujranwala 0.056
T.T singh -0.223 Khanewal -0.207
Chakwal -0.271 Jhang -0.427
Khanewal -0.281 Attock -0.537
Gujrat -0.541 Hyderabad -0.554
Tharparker -0.557 Bannu -0.633
20 least developed districts
Badin -0.990 Mardan -1.373

Kasur -1.011 Larkana -1.438
Layyah -1.013 Sanghar -1.439
Jhang -1.038 Dir -1.558
D.G khan -1.052 D.G khan -1.585
Okara -1.056 Bhakkar -1.595
Mardan -1.061 Shikarpur -1.650
Karak -1.077 R.Y. Khan -1.719
Sukkur -1.117 Swat -1.726
R.Y. Khan -1.133 Tharparker -1.730
Rajanpur -1.239 Muzzafargarh -1.857
Khairpur -1.378 Dadu -1.863
Jacobabad -1.407 Mansehra -1.869
Bhakkar -1.414 Rajanpur -1.953
Swat -1.418 Sukkur -1.993
Vehari -1.420 Jacobabad -2.086
Mansehra -1.449 Thatta -2.131
Muzzafargarh -1.507 Khairpur -2.198
Nawabshah -1.640 Badin -2.320
Dir -1.721 Nawabshah -2.669
TABLE 11
Most Developed Districts in Punjab Based on Principal Component
Post-Primary School System and Hospital Index
Rank/city Principal component Rank/city Principal component

value, 1997-98 value, 2005-06
1. Lahore 6.9 1. Lahore 14.45
2. Rawalpindi 3.5 2. Rawalpindi 8.87
3. Faisalabad 0.62 3. T.T. Singh 2.43
4. Sargodha 0.46 4. Faisalabad 2.19
5. T.T. Singh 0.25 5. Gujrat 1.24
6. Sialkot 0.17 6. Sargodha 0.86
7. Gujrat -0.07 7. Jhelum 0.70
8. Jhelum -0.09 8. Sialkot 0.60
9. Khanewal -0.15 9. Gujranwala 0.24
10. Chakwal -0.33 10. Chakwal 0.06
TABLE 12
Least Developed Districts in Punjab Based on Principal Component Post-Primary
School System and Hospital Index
Rank/city Principal component Rank/city Principal component

value, 1997-98 value, 2005-06
1. Muzaffargarh -1.35 1. Rajanpur -1.72
2. Vehari -1.33 2. Muzaffargarh -1.68
3. D.G. Khan -1.20 3. D.G. Khan -1.48
4. Bhakkar -1.15 4. R.Y. Khan -1.48
5. Khushab -1.11 5. Bhakkar -1.34
6. Rajanpur -1.09 6. Khushab -0.94
7. R.Y. Khan -0.96 7. Vehari -0.90
8. Bahawalpur -0.94 8. Mianwali -0.89
9. Jhang -0.94 9. Layyah -0.82
10. Layyah -0.88 10. Bahawalnagar -0.82

TABLE 13
Descriptive Statistics of Policy Variables in Full Sample
Dev.
Principal component basic health, dispensary & rural health index 0.667 6.690 -7.213 33.737
Principal component post-primary school system and hospital index -0.205 2.477 -2.669 25.541
Principal component teachers in high schools & colleges index 0.540 2.648 -4.726 12.051
Principal component teachers in middle and primary schools index 0.108 2.077 -3.244 8.402
Principal component primary school index 0.128 1.182 -4.287 4.837
Principal component beds in TB clinics index -0.038 0.928 -1.587 8.978
Gini index 0.246 0.065 0.141 0.618
Low Gini, ½ SD below yearly mean (yes=1, no=0) 0.301 0.459 0 1
Medium Gini, around mean (yes=1, no=0) 0.520 0.500 0 1
High Gini, ½ SD above yearly mean (yes=1, no=0) 0.179 0.383 0 1
Low Wolfson, ½ SD below yearly mean (yes=1, no=0) 0.310 0.463 0 1
Medium Wolfson, around mean (yes=1, no=0) 0.402 0.490 0 1
High Wolfson, ½ SD above yearly mean (yes=1, no=0) 0.287 0.453 0 1
Low Gini 1992-93, ½ SD below mean (yes=1, no=0) 0.381 0.485 0 1
Medium Gini 1992-93, around mean (yes=1, no=0) 0.385 0.487 0 1
High Gini 1992-93, ½ SD above mean (yes=1, no=0) 0.234 0.423 0 1
Low Wolfson 1992-93, ½ SD below mean (yes=1, no=0) 0.298 0.457 0 1
Medium Wolfson 1992-93, around mean (yes=1, no=0) 0.417 0.493 0 1
High Wolfson 1992-93, ½ SD above mean (yes=1, no=0) 0.284 0.451 0 1
Principal component post-primary school index × low Gini -0.111 0.397 -2.669 20.232
Principal component post- primary school index × medium Gini -0.052 1.780 -2.320 24.827
Principal component post-primary school Index × high Gini -0.042 1.019 -1.943 25.542
Principal component post-primary school index × low Wolfson -0.065 1.545 -2.669 20.232
Principal component post- primary school index × medium Wolfson -0.109 1.210 -1.969 24.827
Principal component post-primary school Index × high Wolfson -0.030 1.519 -2.320 25.542
Principal component post-primary school index × low Gini 1992-93 0.154 2.298 -2.348 25.542
Principal component post- primary school index × medium Gini -0.169 0.638 -2.131 2.105
1992-93
Principal component post-primary school Index × high Gini 1992-93 -0.190 0.640 -2.669 3.350
Principal component post-primary school index × low Wolfson 0.067 1.403 -2.348 15.241
1992-93
Principal component post- primary school index × medium Wolfson -0.041 1.951 -2.198 25.542
1992-93
Principal component post-primary school Index × high Wolfson -0.231 0.586 -2.669 2.105
1992-93
TABLE 14
Impact of Education and Health Infrastructure on Rural Poverty in Pakistan
Explanatory variables Probit Regressions
(average marginal effects)
E&H Indices Starting E&H
(1) indices
of 1992–93
(2)
Principal component basic health, dispensary & rural health index 0.002 (0.76) 0.002 (0.84)
Principal component post-primary school system & hospital index -0.008**(-2.20) -0.008* (-1.81)
Principal component teachers in high schools & colleges index -0.005 (-0.72) 0.004 (0.37)
Principal component teachers in middle & primary schools index 0.005 (0.51) -0.002 (-0.21)
Principal component primary school index -0.014 (-1.32) -0.007 (-0.72)
Principal component beds in TB clinics index -0.003 (-0.44) -0.012** (-2.23)
Household characteristics included Yes Yes
12 sector and employment status dummies Yes Yes
Individual level controls included Yes Yes
District level controls included No No
Province × year fixed effects included Yes Yes
Pseudo R2 0.131 0.131
Number of observations 349295 349295
Notes: All regressions are estimated by probit maximum likelihood. Numbers in parenthesis are asymptotic t-
values obtained from robust standard errors adjusted for clustering at the district level. The regressions
include intercept terms, but they are not reported. **, and *** denote statistical significance at the 5 and
1% levels, respectively. For definition of control variables, see Table 7.
TABLE 15
Post-Primary School System and Poverty under Different
Inequality and Polarization Regimes
Explanatory variables Gini index Starting Gini of Wolfson index Starting Wolfson
(1) 1992–93 (3) of 1992–93
(2) (4)
Principal component post-primary school -0.008 -0.017*** -0.009* -0.014**
index × low inequality/polarization (-1.27) (-2.65) (-1.81) (-2.2)
Principal component post- primary school -0.009** -0.005** -0.009 -0.005**
index × medium inequality/polarization (-2.33) (-2.39) (-1.3) (-2.31)
Principal component post-primary school -0.008 -0.012 -0.008 -0.021
Index × high inequality/polarization (-1.25) (-0.93) (-1.36) (-1.23)
Lowly educated head, 1–5 years (yes=1, no=0) -0.093*** -0.093*** -0.093*** -0.092***
(-11.66) (-11.58) (-11.6) (-11.75)
Medium educated head, 6–10 years (yes=1, -0.159*** -0.158*** -0.159*** -0.158***
no=0) (-18.53) (-18.54) (-18.49) (-18.81)
Highly educated head, 11 years or more (yes=1, -0.251*** -0.251*** -0.251*** -0.251***
no=0) (-20.54) (-20.46) (-20.5) (-20.72)
Medium sized household, 4–8 members (yes=1, 0.348*** 0.349*** 0.348*** 0.349***
no=0) (31.89) (31.94) (31.9) (32.02)
Large sized household, 9 or more members 0.575*** 0.575*** 0.575*** 0.575***
(yes=1, no=0) (39.61) (39.15) (39.55) (39.49)
Household has subsistence landholding, 1–12.5 -0.098*** -0.099*** -0.098*** -0.097***
acres (yes=1, no=0) (-7.93) (-8.04) (-7.93) (-8.31)
Household has economical landholding, 12.6– -0.149*** -0.150*** -0.149*** -0.149***
25 acres (yes=1, no=0) (-7.2) (-7.31) (-7.19) (-7.28)
Household has large landholding, more than 26 -0.187*** -0.188*** -0.187*** -0.187***
acres (yes=1, no=0) (-5.65) (-5.71) (-5.65) (-5.66)
Livestock owned (yes=1, no=0) -0.042*** -0.042*** -0.042*** -0.043***
(-3.67) (-3.69) (-3.67) (-3.82)
Farm equipment owned (yes=1, no=0) -0.100*** -0.101*** -0.100*** -0.102***
(-3.03) (-2.91) (-3.05) (-3.01)
Financial assets owned by household (Rs. -1.782*** -1.784*** -1.782*** -1.776***
million) (-5.79) (-5.77) (-5.78) (-5.77)
Age -0.001*** -0.001*** -0.001*** -0.001***
(-11.29) (-11.28) (-11.26) (-11.31)
Male (yes=1, no=0) -0.007*** -0.007*** -0.007*** -0.007***
(-3.5) (-3.48) (-3.5) (-3.46)
(-0.35) (-0.36) (-0.35) (-0.27)
12 sector and employment status dummies Yes Yes Yes Yes
Individual level controls included Yes Yes Yes Yes
District level controls included No No No No
Province × year fixed effects included Yes Yes Yes Yes
Pseudo R2 0.130 0.131 0.1301 0.1306
Notes: All regressions are estimated by probit maximum likelihood. Numbers in parenthesis are asymptotic t-values obtained
from robust standard errors adjusted for clustering at the district level. The regressions include intercept terms, but they
are not reported. **, and *** denote statistical significance at the 5 and 1% levels, respectively. For definition of
control variables, see Table 7.
TABLE 16
Summary Statistics of Punjab’s Data
Explanatory variables Mean Std. Dev. Min Max

Poverty 0.326 0.468 0 1
No educated head 0.618 0.485 0 1
Lowly educated head, 1 – 5 years 0.168 0.374 0 1
(yes=1, no=0)
Medium educated head, 6 – 10 years 0.181 0.385 0 1
(yes=1, no=0)
Highly educated head, 11 years or more 0.031 0.175 0 1
(yes=1, no=0)
Small sized household 0.760 0.265 0 1
Medium sized household, 4 – 8 members 0.621 0.485 0 1
(yes=1, no=0)
Large sized household, 9 or more members 0.302 0.459 0 1
(yes=1, no=0)
Household has no landholding 0.615 0.486
Household has subsistence landholding, 1 – 12.5 acres 0.316 0.465 0 1
(yes=1, no=0)
Household has economical landholding, 12.6 – 25 acres (yes=1, no=0) 0.481 0.214 0 1
Household has large landholding, more than 26 acres 0.020 0.140 0 1
(yes=1, no=0)
Livestock owned 0.366 0.481 0 1
(yes=1, no=0)
Farm equipment owned 0.009 0.094 0 1
(yes=1, no=0)
Financial assets owned by household (Rs. million) 0.018 0.059 0 2.28
Age 24.300 19.633 0 99
Male (yes=1, no=0) 0.503 0.499 0 1
Days worked during last month 6.951 11.869 0 30
Employment in agriculture or forestry 0.185 0.388 0 1
(yes=1, no=0)
Employment in mining 0.0003 0.017 0 1
(yes=1, no=0)
Employment in manufacturing 0.028 0.167 0 1
(yes=1, no=0)
Employment in electricity or gas 0.001 0.034 0 1
(yes=1, no=0)
Employment in construction 0.021 0.144 0 1
(yes=1, no=0)
Employment in wholesale or retail trader 0.026 0.160 0 1
(yes=1, no=0)
Employment in transport 0.013 0.114 0 1
(yes=1, no=0)
Employment in real estate 0.0009 0.030 0 1
(yes=1, no=0)
Employment in social or personal services 0.034 0.183 0 1
(yes=1, no=0)
Self employed 0.121 0.326 0 1
(yes=1, no=0)
Paid employee 0.098 0.298 0 1
(yes=1, no=0)
Unpaid family worker 0.096 0.295 0 1
(yes=1, no=0)
Cont. Table 16

Time dummy variables
Survey year 1992-93 0.182 0.386 0 1
Survey year 1993-94 0.150 0.357 0 1
Survey year 1996-97 0.139 0.346 0 1
Survey year 1998-99 0.173 0.378 0 1
Survey year 2001-02 0.174 0.379 0 1
Survey year 2005-06 0.180 0.384 0 1
District dummy variables
Attock 0.027 0.164 0 1
Bahawalnagar 0.030 0.171 0 1
Bahawalpur 0.036 0.188 0 1
Bhakkar 0.026 0.159 0 1
Chakwal 0.016 0.127 0 1
Dera Ghazi Khan 0.026 0.160 0 1
Faisalabad 0.053 0.225 0 1
Gujranwala 0.045 0.208 0 1
Gujrat 0.050 0.218 0 1
Jhang 0.044 0.206 0 1
Jhelum 0.014 0.121 0 1
Kasur 0.035 0.184 0 1
Khanewal 0.025 0.159 0 1
Khushab 0.018 0.133 0 1
Lahore 0.020 0.141 0 1
Layyah 0.015 0.123 0 1
Mianwali 0.018 0.135 0 1
Multan 0.049 0.216 0 1
Muzzafargarh 0.038 0.191 0 1
Okara 0.038 0.191 0 1
Rahimyar Khan 0.050 0.218 0 1
Rajanpur 0.013 0.116 0 1
Rawalpindi 0.052 0.222 0 1
Sahiwal 0.049 0.217 0 1
Sargodha 0.035 0.185 0 1
Sheikhupura 0.047 0.212 0 1
Sialkot 0.057 0.232 0 1
Toba Tek Singh 0.030 0.170 0 1
TABLE 17
Descriptive Statistics of Policy Variables in Punjab
Road Density 0.265 0.125 0.044 0.697
Low Gini, ½ SD below yearly mean (yes=1, no=0) 0.319 0.413 0 1
Medium Gini, around mean (yes=1, no=0) 0.405 0.491 0 1
High Gini, ½ SD above yearly mean (yes=1, no=0) 0.377 0.485 0 1
Low Wolfson, ½ SD below yearly mean (yes=1, no=0) 0.170 0.376 0 1
Medium Wolfson, around mean (yes=1, no=0) 0.403 0.490 0 1
High Wolfson, ½ SD above yearly mean (yes=1, no=0) 0.427 0.494 0 1
Low Gini 1992-93, ½ SD below mean (yes=1, no=0) 0.187 0.390 0 1
Medium Gini 1992-93, around mean (yes=1, no=0) 0.558 0.497 0 1
High Gini 1992-93, ½ SD above mean (yes=1, no=0) 0.256 0.436 0 1
Low Wolfson 1992-93, ½ SD below mean (yes=1, no=0) 0.130 0.336 0 1
Medium Wolfson 1992-93, around mean (yes=1, no=0) 0.450 0.497 0 1
High Wolfson 1992-93, ½ SD above mean (yes=1, no=0) 0.420 0.494 0 1
Road Density × low Gini 0.060 0.130 0 0.698
Road Density × medium Gini 0.105 0.145 0 0.611
Road Density × high Gini 0.100 0.152 0 0.666
Road Density × low Wolfson 0.045 0.116 0 0.390
Road Density × medium Wolfson 0.103 0.145 0 0.698
Road Density × high Wolfson 0.116 0.158 0 0.440
Road Density × low Gini 1992-93 0.046 0.110 0 0.591
Road Density × medium Gini 1992-93 0.167 0.173 0 0.698
Road Density × high Gini 1992-93 0.052 0.104 0 0.440
Road Density × low Wolfson 1992-93 0.033 0.095 0 0.591
Road Density × medium Wolfson 1992-93 0.132 0.169 0 0.698
Road Density × high Wolfson 1992-93 0.100 0.140 0 0.611
Observations 139944 -- -- --
TABLE 18
Impact of Road Density on Poverty in Punjab
Explanatory variables Probit Regression Marginal
Effects
Road Density Starting Road
(1) Density of
1992–93
(2)
Road Density -0.336*** -0.513***
(-3.63) (-4.55)
Lowly educated head, 1 – 5 years -0.081*** -0.080***
(yes=1, no=0) (-7.59) (-7.68)
Medium educated head, 6 – 10 years -0.155*** -0.155***
(yes=1, no=0) (-11.46) (-11.30)
Highly educated head, 11 years or more (yes=1, no=0) -0.246*** -0.246***
(-9.52) (-9.63)
Medium sized household, 4 – 8 members (yes=1, no=0) 0.301*** 0.302***
(22.78) (22.56)
Large sized household, 9 or more members (yes=1, no=0) 0.560*** 0.561***
(35.61) (34.87)
Household has subsistence landholding, 1 – 12.5 acres (yes=1, no=0) -0.146*** -0.145***
(-11.12) (-10.86)
Household has economical landholding, 12.6 – 25 acres (yes=1, no=0) -0.189*** -0.190***
(-6.66) (-6.81)
Household has large landholding, more than 26 acres (yes=1, no=0) -0.213*** -0.215***
(-4.06) (-4.17)
Livestock owned (yes=1, no=0) -0.032*** -0.033***
(-2.76) (-2.84)
Farm equipment owned (yes=1, no=0) -0.087* -0.084*
(-1.82) (-1.76)
Financial assets owned by household (Rs. million) -2.340*** -2.316***
(-6.10) (-6.07)
Age -0.001*** -0.001***
(-8.32) (-8.27)
Male (yes=1, no=0) --0.005* -0.006*
(-1.91) (-1.95)
Days worked during last month --0.0004 -0.001
(-1.07) (-1.05)
12 sector and employment status dummies Yes Yes
Individual level controls included Yes Yes
District level controls included No No
Year fixed effects included Yes Yes
Pseudo R2 0.154 0.155
Number of observations 139944 139944
Notes: All regressions are estimated by probit maximum likelihood. Numbers in parenthesis are asymptotic t-
values obtained from robust standard errors adjusted for clustering at the district level. The regressions
include intercept terms, but they are not reported. **, and *** denote statistical significance at the 5 and
1% levels, respectively. For definition of control variables, see Table 7
TABLE 19
Road Density and Poverty under Different Inequality and Polarization Regimes
Explanatory variables Gini index Starting Gini of Wolfson index Starting Wolfson
(1) 1992–93 (2) (3) of 1992–93
(4)
Road density × low Gini/Wolfson -0.467*** -0.741*** -0.505*** -0.799***
(-4.00) (-7.77) (-4.09) (-7.87)
Road density × medium Gini/Wolfson -0.313*** -0.320*** -0.294*** -0.326***
(-3.01) (-5.19) (-2.83) (-4.37)
Road density × high Gini/Wolfson -0.278*** -0.344*** -0.300*** -0.347***
(-3.42) (-3.14) (-3.41) (-4.45)
Lowly educated head, 1–5 years (yes=1, -0.079*** -0.075*** -0.079*** -0.074***
no=0) (-7.33) (-6.52) (-7.2) (-6.3)
Medium educated head, 6–10 years -0.153*** -0.148*** -0.153*** -0.145***
(yes=1, no=0) (-11.95) (-11.61) (-12.08) (-11.14)
Highly educated head, 11 years or more -0.245*** -0.242*** -0.245*** -0.243***
(yes=1, no=0) (-9.52) (-9.67) (-9.6) (-9.68)
Medium sized household, 4–8 members 0.301*** 0.299*** 0.301*** 0.299***
(yes=1, no=0) (22.61) (22.45) (22.71) (22.27)
Large sized household, 9 or more 0.560*** 0.558*** 0.560*** 0.558***
members (yes=1, no=0) (35.07) (34.02) (35.01) (34.04)
Household has subsistence landholding, -0.145*** -0.143*** -0.144*** -0.144***
1–12.5 acres (yes=1, no=0) (-11.21) (-10.85) (-10.9) (-10.75)
Household has economical landholding, -0.188*** -0.188*** -0.187*** -0.188***
12.6–25 acres (yes=1, no=0) (-6.52) (-6.77) (-6.54) (-6.86)
Household has large landholding, more -0.212*** -0.214*** -0.210*** -0.215***
than 26 acres (yes=1, no=0) (-3.95) (-4.2) (-3.91) (-4.28)
Livestock owned (yes=1, no=0) -0.034** -0.036*** -0.034** -0.037***
(-2.94) (-3.24) (-2.9) (-3.28)
Farm equipment owned (yes=1, no=0) -0.092** -0.095** -0.092** -0.093**
(-1.97) (-2.03) (-1.94) (-1.98)
Financial assets owned by household -2.34*** -2.329*** -2.340*** -2.310***
(Rs. million) (-6) (-6.10) (-6.05) (-6.13)
Age -0.001*** -0.001*** -0.001*** -0.001***
(-8.29) (-7.78) (-8.4) (-7.6)
Male (yes=1, no=0) -0.005** -0.005* -0.005** -0.005*
(-1.95) (-1.86) (-1.89) (-1.82)
(-0.99) (-1.13) (-1.08) (-1.15)
12 sector and employment status Yes Yes Yes Yes
dummies
Individual level controls included Yes Yes Yes Yes
District level controls included No No No No
Year fixed effects included Yes Yes Yes Yes
Pseudo R2 0.1551 0.1592 0.1553 0.1586
Notes: All regressions are estimated by probit maximum likelihood. Numbers in parenthesis are asymptotic t-values obtained
from robust standard errors adjusted for clustering at the district level. The regressions include intercept terms, but they
are not reported. **, and *** denote statistical significance at the 5 and 1% levels, respectively. For definition of control
variables, see Table 7.
Figure 1: Geographic Location of 20 Most Developed Districts Based on

Figure 2: Geographic Location of 20 Least Developed Districts Based on

Figure 3: Geographic Concentration of Road Density in Punjab, 2005-06
Figure 4: Relative Road Density of Punjab’s Districts, 1992-93 vs. 2005-06
Relative road density in Punjab, 1992-93 vs. 2005-06

LHR
1
SRG
FSD
RWP
Relative road density, 2005-06
VHR
OKR
.8
TTS
KHW
JHG KSR
SHP
.6
CHK JHL
GJW
SIA
MNW MZF BHW SHW
MLT
RYK
.4
GJT
KHB
BKR ATK
LYH
.2
DGK
BHP
RJN
.2 .4 .6 .8 1
Relative road density, 1992-93
Figure 5: Road Density and Poverty Headcount in Punjab
80
Poverty Headcount Ratio
20 400 60
0 .2 .4 .6 .8
Road density
Poverty headcount ratio Fitted values

The study has been carried out under the 11th Round of the SANEI Regional Research
Competition (RRC) made possible with a financial grant from the
Global Development Network (GDN).
Journal of Economic Perspectives—Volume 30, Number 2—Spring 2016—Pages 3–28
Consumption Inequality
Orazio P. Attanasio and Luigi Pistaferri
M
uch of the debate over the rising levels of inequality in the United States
and other developed countries is phrased in terms of income, or in terms
of components of income like wages and earnings. But for economists,
a basic utility function of individuals typically refers to consumption and leisure,
not income.
The distinction between income and consumption could make a mean-
ingful difference in thinking about inequality if the distribution of consumption
at a given point in time is less wide than that of income, or if its evolution over
time is smoother than that of income. Consumption can differ from income if
consumers borrow or save, or if they receive transfers from other family members
or the g
overnment in response to income shocks. The joint analysis of consumption
and income inequality can be informative in several ways. It can show the p resence
(or lack) of such consumption-smoothing mechanisms. It can shed light on
the nature of income shocks, and in particular the extent to which they should
be understood as temporary (which may be easier to smooth out for consumption
■ Orazio P. Attanasio is Jeremy Bentham Professor of Economics, University College London,

London, United Kingdom, and Research Fellow, Institute for Fiscal Studies, both in London,
United Kingdom. Luigi Pistaferri is Professor of Economics and Ralph Landau Senior
Fellow, Stanford Institute for Economic Policy Research (SIEPR), both at Stanford University,
Stanford, California. Both authors are Research Associates, National Bureau of Economic
Research, Cambridge, Massachusetts, and also Research Fellows, Centre for Economic Policy
Research (CEPR), London, United Kingdom. Their email addresses are o.attanasio@ucl.
ac.uk and pista@stanford.edu.
†
For supplementary materials such as appendices, datasets, and author disclosure statements, see the
article page at
http://dx.doi.org/10.1257/jep.30.2.3 doi=10.1257/jep.30.2.3
purposes) or permanent. If one is interested in the effects of inequality on those in

the poorest segments of society, consumption might reveal different insights than
income—for example, because of different dynamics in the relative prices of goods
consumed by rich and poor households. Finally, higher consumption of leisure
could partly offset lower consumption of goods when it comes to overall welfare
measurement.
In this essay, we begin with a discussion of the sources of consumption data,
and some of the issues that arise when looking at data on consumer spending while
trying to infer the economically relevant concept of consumption. We then offer
our interpretation of the research that has compared trends in income inequality
with trends in consumption inequality. The narrative has evolved very sharply from
arguing that trends in consumption inequality are quite different from those of
income inequality to concluding that they track each other closely. This change in
findings has been shaped in substantial part by the adoption of strategies aimed at
dealing with measurement problems in consumption data, as well as by some rein-
terpretation of the underlying economic forces.
We then discuss some additional aspects of consumption. We look at specific data
on inequality in consumption of food, ownership of major household appliances,
leisure, and persistence in consumption across generations. These comparisons
suggest ways in which aggregate consumption inequality fails to tell the entire story.
In the concluding section, we take stock of the evidence and summarize challenges
for future work. Our main conclusion is that researchers interested in measuring
inequality in well-being need to go beyond the fact that consumption is unequally
distributed and realize that a full picture of the evolution in welfare requires taking
a stand on quality concerns and on the value that people attach to leisure, among
other things.
Consumption Data: Sources and Concepts
What Consumption Data Do We Have?

US researchers who want to study income inequality have a considerable
array of data sources from which to choose. If they want household survey data on
incomes, the Current Population Survey, Panel Study of Income Dynamics, Survey
of Income and Program Participation, National Longitudinal Survey of Youth, and
even the decennial Census (for studying long-run trends) all offer large, consis-
tent samples and detailed information on income and its components. Researchers
may also have access to administrative-level data (where measurement error issues
may be less important with regard to income data), such as data from the Internal
Revenue Service and data from the W-2 forms that employers use to report income
paid. Overall, datasets with measures of household income resources (such as wages,
earnings, and income) are more frequently available, typically have larger samples,
and have more consistent variable definitions than datasets containing information
on consumption (Pistaferri 2015).
Orazio P. Attanasio and Luigi Pistaferri 5
In contrast, household surveys on household expenditure are rare, small, and

lack a consistent longitudinal component. The Consumer Expenditure Survey
(CE), the only dataset with comprehensive and detailed information on household
expenditure and its components, is available on a continuous basis since 1980. It
is used by the Bureau of Labor Statistics primarily to form weights placed on price
changes of goods in the computation of the overall Consumer Price Index. The CE
is composed of two distinct surveys. In the Interview survey, respondents are sampled
every three months for a total of four quarters. In the Diary survey, respondents fill
a two-week diary of their expenditures and are sampled only once. In producing
aggregate means, the Bureau of Labor Statistics routinely uses and tabulates certain
items from the Interview survey and others from the Diary survey.
The other dataset that is widely used by academic researchers to study consump-
tion behavior is the Panel Study of Income Dynamics (PSID), which is available on an
annual basis from 1968 to 1997, and on a biannual basis since then. The initial goal
of the PSID was to study income dynamics (and poverty) between and across gener-
ations. For this reason, information on consumption was considered ancillary, and
until 1997, the PSID collected information only on a few consumption items: food
(at home and away from home), home rent, and (occasionally) utility payments.
Starting with the 1999 wave, however, the PSID began to collect information on a
larger range of items, covering about 70–90 percent of the spending collected in the
Consumer Expenditure Survey. Respondents typically report spending for broad
categories, with the reference period being (with some exceptions) the previous
calendar year. Blundell, Pistaferri, and Saporta-Eksten (2016) show that these data
track national accounts aggregates well.
Comprehensive administrative data on consumption spending are not avail-
able, but some partial sources do exist. For example, some researchers have used
spending data from credit card expenditure records (as in Gross and Souleles 2002;
Aydin 2015). Others have used data on spending, income, and assets for consumers
using online financial aggregators such as Mint.com or Check.com (Baker 2014;
Gelman et al. 2014). Finally, there is spending data from checkout scanners from
the Nielsen Homescan datasets (Handbury 2014; Broda and Weinstein 2008), which
refer primarily to grocery store items. While these new sources of administrative
data on consumption constitute remarkable steps ahead, they are either not repre-
sentative of all households or not representative of all the goods that people buy.1
Looking at Spending, Thinking about Consumption

Consumption can be harder to measure accurately than income, and measure-
ment errors may be differently severe in different parts of the income distribution.
1
For some Nordic countries, researchers have proposed using longitudinal administrative tax record
information on income and wealth to create consumption using the intertemporal budget constraint, so
that expenditure can be derived as income minus the change in wealth: see Browning and Leth-Petersen
(2003) and De Giorgi, Frederiksen, and Pistaferri (2015) for Denmark; Koijen, Van Nieuwerburgh, and
Vestmanz (2015) for Sweden; and Autor, Kostøl, and Mogstad (2015) for Norway.
Survey data like the Consumer Expenditure Survey typically report consumer
spending, which may not coincide with consumption for at least four reasons. First,
consumption is overstated relative to spending for those who have made durable
purchases in the current period, and understated for those who made durable pur-
chases in the past. Most surveys of household consumption have no information on
the stock of existing durables owned by the household. In the Consumer Expen-
diture Survey, the only exception is cars, as consumers report the year, type, and
make of the cars they own. However, there is no information on the resale values,
which must instead be estimated. For other durables, there is some information on
ownership and number of items owned, but no information on current values. As
for housing, the survey contains information on imputed services, as homeowners
report how much their house would rent for. Second, some consumption is received
in kind, through transfers from friends, relatives, private institutions (charities or
churches), or the government (in the form of in-kind or voucher-provided benefits
like food stamps, school lunches, health care services through Medicaid or Medi-
care, rent subsidies, and so forth). Third, some consumption is produced at home
using time and good inputs, like child care provided by parents or siblings. Finally,
the conversion from spending to consumption requires knowledge of prices paid
for the goods people consume, a requirement that is usually solved by assuming
that households face the same prices. This assumption is violated in practice for
many nontradeable goods (such as housing), and even for tradeable, homogenous
goods due to shop-specific effects, bulk purchases, or loyalty cards. Ignoring the
distinction between consumption and spending can either understate or overstate
differences in well-being across individuals with different levels of spending.
Survey nonresponse and measurement errors create a different set of issues.
Sabelhaus et al. (2015) study nonresponse rates in the Consumer Expenditure
Survey and conclude that they have risen over time, especially among the high-
income population. Sabelhaus and Groen (2000), using a variety of techniques,
argue that the ratio of consumption to income for richer households is downward
biased. This may affect the measurement of trends in aggregate consumption and
consumption inequality, respectively. In this journal, Meyer, Mok, and Sullivan
(2015) conclude that measurement error in survey data has also increased over
time. In principle, there is no obvious reason to expect errors in reporting spending
to be worse than errors in reporting income. On the one hand, the changing nature
of spending modes and patterns may heighten reporting errors; on the other hand,
the move from cash to e-commerce may facilitate the collection of administrative
data on spending. Several papers discuss strategies to elicit consumption informa-
tion in general-purpose surveys (for instance, Crossley and Winter 2015).
The combination of these issues makes any measure of consumption naturally
problematic, no matter how much effort researchers put into making it accurate.
In contrast, income measured in surveys is arguably closer to the relevant economic
concept (except in cases involving businesses). However, we should also note
that income is not easier to measure than consumption for every household. For
low-income individuals (in both developed and developing countries) income can
Consumption Inequality 7
be complex to measure, because it includes a myriad of different sources including

wages, interpersonal and government transfers, and so on. In comparison,
consumption for the poorest households may be fairly straightforward. The situa-
tion is probably reversed for well-off households for which administrative income
data might be accessible and reliable, while their consumption can be complex and
varied and difficult to measure (Deaton and Grosh 2000).
Survey Data versus National Accounts Data on Consumption

Consumer spending data available in the Consumer Expenditure Survey
appears increasingly detached from the Personal Consumption Expenditure (PCE)
data collected by the Bureau of Economic Analysis (which forms the basis for the
national income and product accounts data). For example, Passero, Garner, and
McCully (2015) report that the ratio of total expenditures in the CE data compared
to the PCE data has declined from 0.70 in 1992 to 0.58 in 2010. Of course, when-
ever two different methods of measuring a similar economic concept give different
answers, it’s a matter for concern.
Some of the discrepancy is due to the fact that the two series measure different
concepts and cover different entities. The Personal Consumption Expenditure
data includes the value of goods and services purchased by US resident households
(including imputed rents for owner-occupied housing) and by nonprofit organi-
zations on behalf of households (typically, employer-paid health insurance and
medical care, and expenses associated with life insurance and pension plans). It
also includes purchases by US government civilian and military personnel stationed
abroad and US residents traveling or working abroad for one year or less. For most
consumption categories, the PCE is estimated using a “commodity-flow” method.
This approach computes the value of domestic output based on data from the
Census of Manufactures, which looks at the value of manufacturers’ shipments and
inventories. Next, domestic consumption (denominated in producers’ prices) is
estimated by adding imports and subtracting exports and changes in inventory.
Finally, the value of consumer purchases is converted from producers’ prices to
purchasers’ prices by adding wholesale margins and taxes, transportation costs, and
retail margins, and taxes. Clearly, many steps in the calculation of the PCE are also
likely to contain sizeable measurement error.
Both the population coverage and the methodology for obtaining total
spending is different in the Consumer Expenditure Survey (as discussed in Slesnick
1991). The Personal Consumption Expenditure data includes institutionalized
individuals; the Consumer Expenditure Survey does not. The Consumer Expen-
diture Survey excludes spending made by US residents abroad and by nonprofit
institutions on behalf of households (with the most obvious difference being the
value of Medicare and Medicaid spending, which tripled in real terms between 1990
and 2014). The PCE concept includes imputed rents on owner-occupied housing,
while the Consumer Expenditure Survey aggregates typically exclude them. Indeed,
the discrepancy between the two measures is less dramatic when comparing items
that are conceptually comparable and definitionally similar. Passero, Garner, and
McCully (2015) compare different components of consumption and conclude that

“non-durables are most alike for the CE and PCE with about 93 percent of total
non-durable expenditures identified as comparable within the CE and within the
PCE.” Their conclusion is that “focusing on comparable goods and services only,
CE to PCE ratios have steadily decreased,” but slightly less than when comparing
unadjusted statistics. For example, the CE–PCE ratio for total comparable goods
and services decreased from 84 percent for 1992 to 74 percent for 2010 (as opposed
to 70 and 58 percent, respectively, in unadjusted figures). They write: “The greatest
decline in CE to PCE ratios is for durables, with a decrease of 24 percentage points,”
from 0.82 to 0.62. The decline is smaller for services (0.95 to 0.86) and for nondu-
rables (0.70 to 0.63).
Bee, Meyer, and Sullivan (2015) assess the performance of the Consumer
Expenditure Survey on a good-by-good basis and report three findings. First, in
general the Interview survey performs better than the Diary survey in matching
numbers from the Personal Consumption Expenditure data for some categories.
Second, the coverage ratios are excellent for some goods (food at home, rent,
and utilities) and, in those areas, have not changed appreciably over time; on the
other hand, the coverage ratios for other items (such as clothing or alcoholic bever-
ages) are low and declining. Finally, some durable stocks and durable purchases
appear to be reported sufficiently well (new vehicles), while the quality of others
has worsened considerably (furniture). Overall, they conclude that the consump-
tion categories that tend to be reported poorly are those that involve small and
infrequent purchases, while large and regular purchases are reported sufficiently
well. Moreover, they write that “based on observable characteristics, the [Consumer
Expenditure Survey] appears to be fairly representative, although there is strong
evidence of under-representation at the top of the income distribution and under-
reporting of income and expenditures at the top.” This is another reason for the
growing discrepancy between the CE aggregates and the PCE from the National
Accounts. If consumption growth is higher at the top of the distribution, declining
survey response among the rich may easily explain the deterioration of the match
between CE aggregates and the PCE.
Quantity versus Prices

Survey data like the Consumer Expenditure Survey measure expenditure,
which is the product of prices and quantities. To make comparison of consump-
tion across periods meaningful, researchers deflate expenditure using an overall
measure of the cost of living, typically the Consumer Price Index (CPI). Indeed,
as mentioned above, the CE is collected primarily to compute the weights for the
CPI. However, average weights may not be relevant for all households. It is possible
that the composition of the consumption basket varies substantially and systemati-
cally across different households, as a direct consequence of differences in access to
resources as well as needs and tastes. Luxuries will be more prominent in the expen-
diture basket of the rich, while necessities will account for a larger share of poor
households’ expenditure. Health costs may be more relevant for older individuals
and certain types of entertainment more relevant for younger ones. Therefore
changes in relative prices can have distributional consequences.
Moreover, there may be price differences across space (or stores within a given
geographical location), or across time within space (because of high-frequency
sales) even for relatively homogenous goods. Because of differences in prices over
space, consumers might have an incentive to search for the best deals and these
incentives may vary for individuals with different values of time.
The presence of differences in prices for homogeneous goods and changes
in the relative prices of goods that are more or less relevant for different groups of
individuals might lead researchers to overstate or understate the level and trends
of inequality. If the poor live in relatively cheaper areas (or if they shop in relatively
less-expensive stores), or if the prices for the goods that they typically purchase
grow less than the prices of the goods typically purchased by rich households, then
inequality in consumption (that is, spending deflated by an index that accounts
for household-specific price differences) will be less (and grow more slowly) than
inequality in spending deflated with a common price index. There are two reasons
why this may be the case. First, increasing trade with low-wage countries lowers
the price of imported goods. This may reduce consumption inequality even in
the absence of any change in income inequality if goods imported from low-wage
countries are relatively more important in the consumption basket of low-income
individuals than in the basket of high-income individuals. Moreover, the diffusion
of mega-stores (such as Walmart) has likely benefited low-income individuals more
than high-income individuals. These differences are partly attenuated by the consid-
eration that better quality and the experience of shopping in certain stores have an
amenity value.
Datasets where researchers can disentangle the two components of expenditure
are hard to come by. The Nielsen Homescan data is one exception, but it is limited
to groceries and a few other items. Other data sources (such as the ACCRA Cost of
Living Index produced by the Council for Community and Economic Research)
provide city-specific indexes on a few categories of interest. In the Consumer Expen-
diture Survey, the geographical detail is very limited due to confidentiality concerns
(and in any case, it would miss information on the type of store where goods are
purchased). Later in the paper, we discuss some recent work on the implications of
price inequality.
Does Consumption Inequality Track Income Inequality?
Consumption Smoothing: Why and How

Is consumption inequality a better measure of changes in welfare than income
inequality? For economists, using consumption inequality has theoretical appeal.
The life-cycle hypothesis of Modigliani and Brumberg (1955) and the permanent
income hypothesis of Friedman (1957), which constitute the workhorse theory of
how people make their consumption decisions, suggest that risk-averse households
prefer a smooth to a variable consumption flow. Hence, households would choose

consumption to be a constant fraction of their permanent or lifetime income, not
current income. Because current income can be highly volatile from one year to the
next, it may give a partial snapshot of people’s living standards. The extent to which
households can achieve a smooth consumption flow depends on the tools they have
to move resources over time and states of nature. Savings can be used to absorb
certain income shocks and can be accumulated for such a purpose. Other tools for
consumption smoothing may include access to credit and insurance markets, and
interpersonal and government transfers.
The ability to move resources across time and states explains why consump-
tion may not track income. Consumption may exceed current income because a
consumer is borrowing (permanent income is above current income, as in the case
of a medical student taking out loans in the expectation of higher future earnings)
or it may be below current income because the consumer is saving (and the doctor
is now repaying medical school loans). Large wealth effects can also have a consider-
able influence on consumption independently of income. It is then possible for the
income distribution to reveal no changes in well-being even though the underlying
consumption distribution is shifting in response to wealth effects. Consumption
may vary from income for other reasons as well. For example, consumption falls
below current earnings and wages because of taxes paid, and above them because
of government transfers—a different form of consumption smoothing especially
relevant for households at the bottom of the distribution. Even if full smoothing is
not feasible, perhaps because of borrowing restrictions and imperfections in insur-
ance and credit markets, some consumption smoothing would still occur. Recent
surveys on consumption (or marginal utility) smoothing are Browning and Crossley
(2001) and Attanasio and Weber (2010).
These considerations imply that how consumption (and welfare) react to
changes depends on the tools available for consumption smoothing and on the
nature of income changes. In general, we can think of current income as including
two components: one reflects long-term or permanent factors such as the level of
skills and human capital, while the other reflects temporary or transitory factors
(like being out of the labor force due to human capital investments, fertility, job
loss, and the like). For any given household, permanent income changes slowly
over time. When it does, it is because of unpredictable events, like the way in which
new technologies affect the price and quantity demanded of one’s skills (for better
or worse), along with unexpected promotions or demotions, a change in the local
economy that affects wage levels, and other factors. MaCurdy (1982) and Abowd
and Card (1989) are two representative papers in a vast labor literature that tries
to decompose income (or earnings, or wages) into transitory and permanent
components.
The distinction between temporary and persistent shifts in the wage or
income distribution has important policy and welfare implications. Policies
aimed at reducing inequality under the two scenarios are very different. In the
first case, it is probably necessary to reduce inequality in the endowments of
human capital, whilst in the latter it may be sufficient to improve the access to
smoothing mechanisms. In theory, permanent shocks are harder to absorb and
insure and are thus more likely to be reflected in substantial changes in consump-
tion and welfare. In contrast, temporary shocks are easier to smooth through
borrowing or running down accumulated assets. Hence, if all changes in income
inequality were of a transitory nature, we could expect no large changes in
consumption inequality.
The Evidence on Consumption and Income Inequality

The mainstream narrative on consumption inequality has evolved considerably
over time, from earlier uncertainty over whether consumption was rising less than
income inequality to the current belief that it has been rising just about as much
as income inequality.
The first few papers in the earlier literature that looked at different dimen-
sions of inequality in consumption were Cutler and Katz (1991, 1992) and Slesnick
(1994). These papers used the Consumer Expenditure Survey data (in an era in
which the measurement error issues discussed above were less well-known than they
have since become). Cutler and Katz (1991) found that “changes in the distribution
of consumption parallel changes in the distribution of income.” In contrast, Slesnick
(1994) found that consumption inequality had grown more modestly than income
inequality.2 A number of later studies (for example, Krueger and Perri 2006) found
evidence similar to Slesnick (1994). In the terminology of consumption smoothing
and the permanent income hypothesis, this finding can be interpreted as implying
that a sizeable proportion of shocks to income were both temporary and insurable.
Evidence on this comes from two sources: direct evidence on the income process,
and indirect evidence coming from the response of consumption to income changes
of different natures.
On the first front, researchers working on income inequality were finding
strong evidence that the rise in income inequality was partly of transitory nature—
which in turn implies that some portion of the rise in inequality could be more
smoothed in consumption. Gottschalk and Moffitt (1994) argued that the rise in
the variance of the transitory component of income (what they called “income
instability”) represented about one-third to one-half of the overall rise in income
inequality observed in the 1980s and 1990s. (Income instability is often measured
using the variance of income changes, or growth. This is because the change in
income across two periods is approximately equal to the change in the transitory
component if permanent income evolves along the expected path.)
The fact that a good chunk of the rise in income inequality was of a transitory
nature would not matter much (in terms of separating income from consumption
inequality) if consumers were unable to smooth transitory shocks. However, several
2
The differences between the two studies arise partly from their different consumption definitions.
Cutler and Katz (1991) include spending on durables (other than housing and vehicles), while Slesnick
(1994) imputes service flows. Moreover, Slesnick adjusts for topcoding in some spending categories.
papers show that consumers are able to smooth short-run shocks (Dynarski and
Gruber 1997), although less so if they have low assets or low education (Blundell,
Pistaferri, and Preston 2008). Attanasio and Davis (1996) focused on the relation-
ship between relative wages and relative consumption across different groups in
the US population, where groups were defined on the basis of the year of birth
of the household head and on their educational achievement. They found that
long-run relative movements in wages across these groups were mirrored in relative
movements in consumption. This correlation was driven by the relative move-
ments across education groups: the increases in the return to education in terms
of wages and earnings seemed to be reflected in increases in the return to educa-
tion in terms of consumption. The online appendix available with this paper at
http://e-jep.org contains an update of Attanasio and Davis (1996), extending the
data to 2012. As in the original paper, we find that when we consider the impact
on consumption of one-year wage changes, the variability of which is probably
dominated by temporary fluctuations in wages that can be smoothed in some way,
we do not find a significant relationship between relative changes in wages and
consumption. However, when considering longer (five- and eight-year) horizons,
where instead persistent wage factors are more likely to be at play, the relation-
ship between changes in consumption inequality and income inequality becomes
strongly significant.
In keeping with these two pieces of evidence, Krueger and Perri (2006) show
that while in the 1980–2003 period the variance of log income increases from 0.35
to 0.57, the variance of log consumption increases only from 0.18 to 0.24. In other
words, inequality rises in both income and consumption, but the rise in income
inequality is much larger.
More recently, however, researchers have started to question the evidence
about consumption inequality, rethinking the measurement issues that arise from
considering measures of expenditure in the Consumer Expenditure Survey. The
survey seems to be affected by serious nonclassical measurement error whose impor-
tance is increasing over time. One possible strategy is to focus on the components
of the Consumer Expenditure Survey that appear to be measured most accurately
and to use alternative datasets for other components of consumption. The chal-
lenge of course is that one would like to make statements about inequality in
overall consumption, not necessarily about inequality in some components. Once
one corrects for the measurement problems afflicting the Consumer Expenditure
Survey, uses alternative data sources when these seem preferable, or measures
consumption in alternative ways, consumption inequality seems to rise by more
than previously believed, and to track income inequality closely.
A number of papers develop this view using a variety of data sources and empir-
ical approaches (Attanasio, Battistin, and Ichimura 2007; Aguiar and Bils 2011;
Attanasio and Pistaferri 2014). Figure 1 gives an overall view of the evolution of
consumption inequality over time and across papers (and empirical strategies). In
this figure, consumption inequality is measured by the variance of log consump-
tion (deflated by the Consumer Price Index and expressed in per capita terms).
Figure 1
The Evolution of Consumption Inequality over Time as Measured by Different
Papers
.45
Aguiar and Bils (2015)
Attanasio, Battistin, and Ichimura (2007)
.40 Attanasio and Pistaferri (2014)
Variance of log consumption
Heathcote, Perri, and Violante (2010)
.35
.30
.25
.20
1980 1985 1990 1995 2000 2005
Note: Heathcote, Perri, and Violante (2010) used the Interview survey of the Consumer Expenditure
Survey. Attanasio, Battistin, and Ichimura (2007) combined consumption items from the Interview
survey and the Diary survey (in the attempt of picking the survey component that best measures each
item). Aguiar and Bils (2015) used the Consumer Expenditure Survey but computed consumption as
the difference between disposable income and active savings. In Attanasio and Pistaferri (2014), we used
Panel Study of Income Dynamics consumption data available from 1999 onward, estimated an inverse
demand function for food for the 1999–2009 period, and then used the estimated coefficients to predict
consumption for the period before 1999 (when only food data were available). Consumption inequality
is measured by the variance of log consumption (deflated by the Consumer Price Index and expressed
in per capita terms).
Heathcote, Perri, and Violante (2010) used the Interview survey of the Consumer
Expenditure Survey, and their findings reproduce the flat profile of consumption
inequality shown by Krueger and Perri (2006). Attanasio, Battistin, and Ichimura
(2007) combined consumption items from the Interview survey and the Diary
survey (attempting to pick the survey component that best measures each item),
with results showing a more marked increase in inequality. Aguiar and Bils (2015)
used the Consumer Expenditure Survey but computed consumption as the differ-
ence between disposable income and active savings, and they find an even larger
increase in inequality. Finally, in Attanasio and Pistaferri (2014), we used Panel
Study of Income Dynamics consumption data available from 1999 onward, esti-
mated an inverse demand function for food for the 1999–2009 period, and then
used the estimated coefficients to predict consumption for the period before 1999
(when only food data were available). We found that inequality also increases more
than the Heathcote, Perri, and Violante measure, especially in the last years of the
sample period.3
To obtain a sense of how much consumption inequality grows relative to income
inequality, and how the response depends on the methodology used to measure
consumption, consider the following calculation. Over the period considered in
the figure, the variance of the log of family income from the PSID (deflated by the
Consumer Price Index and expressed in per capita terms) increases by 27 points (or
about 20 points when using an after-tax measure available in the Consumer Expen-
diture Survey, as reported by Heathcote, Perri, and Violante 2010). If we take the
Aguiar and Bils measure of consumption inequality shown in Figure 1 as the most
credible, the variance of log consumption increases by about 18 points over the
same time period. In contrast, the Heathcote, Perri, and Violante measure would
suggest an increase of only about 10 points. Meyer and Sullivan (2013) show that
the tracking between income and consumption inequality is stronger at the top of
the distribution (as measured by the 90th–50th percentile difference) and in the
1980s and 1990s than in subsequent years. Aguiar and Bils’s (2015) core exercise is
actually to measure consumption inequality by looking at how high- and low-income
households allocate spending to luxuries and necessities. In particular, inequality in
the luxury/necessity spending ratio (scaled by the difference in demand elasticities,
which can be obtained from estimation of a demand system) is shown to provide a
measure of consumption inequality that is robust to measurement error in overall
spending, as well as to household-specific measurement errors (for example, more
severe underreporting by high-income households) and good-specific measure-
ment errors (more severe underreporting for some goods than others). Using
this alternative metric, Aguiar and Bils confirm that over the 1980–2007 period,
inequality in consumption grows as much as income inequality.
The common element of the papers above is that once one makes an attempt
to move away from the traditional measurement of consumption inequality using
the Interview component of the Consumer Expenditure Survey, and tries to correct
for the measurement problems, then the trends in consumption inequality appear
much steeper than initially believed. Of course, the conclusion reached by these
papers may also be premature, because the strategies adopted, while ingenious, are
based on data that may have different types of measurement problems.
One aspect that seems to militate in their favor, however, is that the change
in the consensus about the trends in consumption inequality has been accompa-
nied by changes in the consensus on the evolution of income inequality, based on
3
An important caveat is that the consumption series used in the four papers are not identical. For
example, the Attanasio and Pistaferri (2014) measure is limited by the fact that the PSID collects a
limited amount of information on expenditure, while the Aguiar and Bils (2015) measure, by definition,
does not use any consumption information. The Heathcote, Perri, and Violante (2010) and Attanasio
and Pistaferri (2014) measures include out-of-pocket spending on health and education, while the Atta-
nasio, Battistin, and Ichimura (2007) measure excludes them. The differences between the series should
thus be seen as illustrative. We normalize all series to equal the Heathcote, Perri, and Violante (2010)
value in 1982 (the first year in which we observe all four series).
improved income data. Recent work using administrative data about income—which
is less prone to measurement error issues than survey data—finds that most of the
increase in wage and earnings variance has been structural, or of a more permanent
nature (for example, DeBacker et al. 2013; Kopczuk, Saez, and Song 2010; Guvenen,
Ozkam, and Song 2014). Kopczuk, Saez, and Song (2010) use Social Security data
over a very long time horizon and present a formal decomposition between total,
transitory, and persistent earnings variances. Using this decomposition, the rise in
total variance during the period of interest is primarily driven by a rise in structural
factors. In contrast, there is very little evidence of a rise in the variance of the transi-
tory component.
These recent findings about the nature and dynamics of income inequality are
consistent with the revised thinking in the dynamics of consumption inequality. If
income volatility is stable, it means the variance of the transitory component has
not increased. Hence, the bulk of the change in income inequality has occurred
because of a rise in the variance of the permanent component. It is possible for
consumption inequality to rise less than income inequality even in a setting in which
income volatility is stable. This is because consumers may be able to insure even
some shocks to their permanent income, at least partially. For example, the Disability
Insurance program seeks to attenuate the economic cost of permanent shocks to
health that result in permanent inability to work. But it is clearly more difficult to
smooth changes in permanent income, and as a consequence, it is not surprising
that consumption inequality rises by roughly as much as income inequality. Indeed,
their rise is explained by the same forces (absent strong insurance mechanisms).
Inequality in Prices
As mentioned earlier, recent research has started to look at data on individual
purchases and has documented the existence of important heterogeneity in prices
of even very homogenous goods, both in different stores and within a store over
short periods of time, through the use of sales and discounts. One of the first papers
to document the existence of substantial heterogeneity in the prices of very homo-
geneous goods is Aguiar and Hurst (2007). They correlate observed prices and
shopping behavior with consumer characteristics. They show that older consumers
are more likely to shop longer and more frequently and, probably as a consequence,
pay lower prices for similar goods. Griffith, Leibtag, Leicester, and Nevo (2008)
use British scanner data to show that low-income households realize considerable
savings by buying in bulk and by buying economy brands, while savings from sales,
coupons, and the like are nonlinear in income—specifically, higher at the top and
bottom of the income distribution. More recently, Nevo and Wong (2015) show
that during the Great Recession, consumers switched to buying more on sale, using
more coupons, buying more generics and larger pack sizes and these changes were
larger in states that suffered larger increases in unemployment rates.
Kaplan and Menzio (2015) use scanner data on a large sample of US households
covering grocery stores purchases in 54 geographical markets over the 2004–2009
period. They find that the distribution of prices is symmetric and with fatter tails
than the normal distribution, and its average standard deviation is between 19 and
36 percent. They also show that, when decomposing the variability of prices of
homogeneous goods into a store component, a store-good component, and a trans-
action component (the dispersion of prices within a store), most of the variability
of prices in their sample is explained by the latter two factors. They suggest that
price dispersion is more likely to be driven by intertemporal price discrimination
and search frictions than differences in amenities or marginal costs across stores.
This hypothesis is explored more fully in Kaplan, Menzio, Rudanko, and Trachter
(2016).
The extent to which differences in prices actually paid affect the dynamics of
consumption inequality, either through differences in consumer baskets or through
price heterogeneity induced and sustained by frictions and retailer behavior, is an
open question and one of considerable interest. The availability and use of scanner
data on individual transactions can be very useful in this respect, as is the development
of models that incorporate price discrimination and frictions in price-setting behavior.
More broadly, the measurement of consumption and income inequality is a lively area
of research. The existing work is undoubtedly subject to improvement as better data
or more creative approaches to overcome measurement issues come along.
Inequality in Components of Consumption
Looking at inequality of consumption across specific components of consump-

tion may be interesting for several reasons. First, the measurement of some
components of consumption is of better quality than others, thus alleviating
concerns about whether results are affected by measurement error. Second, the
analysis of different groups of commodities with different income elasticities can
be informative about the nature of shocks and about mechanisms for smoothing
consumption. Third, changes in the patterns of expenditure on durables can be
informative about the perception of future shocks, because individuals know that
commodities such as furniture or cars provide services for long periods and can
be sold only subject to large transaction costs. Finally, disparities in consumption
necessities such as food may be more worrying from a welfare point of view than
disparities in the consumption of luxuries, such as exotic vacations.
Food
In Figure 2, we use data from the Panel Study of Income Dynamics and plot
the difference between the 90th and the 10th percentile of the logarithm of food
consumption distribution over the 1977–2012 period. Food consumption is defined
as the sum of spending on food at home, food away from home, and the monetary
value of food stamps. Data are in real terms and adjusted for family composition
by dividing by an OECD scale (defined as $1 + 0.7(n – 1) + 0.3k, where n is the
number of adults and k the number of kids). PSID food data exist before 1977, but it
is only in 1977 that the Food Stamps Act established national standards of eligibility.
Figure 2
The 90th–10th Percentile Log Food Difference
1.7
90th–10th percentile log difference
Adding food stamps
1.6
And correcting for price differences
1.5
1.4
1.3
1.2
1.1
1975 1980 1985 1990 1995 2000 2005 2010 2015
Source: Authors using data from the Panel Study of Income Dynamics.
Note: This figure plots the difference between the 90th and the 10th percentiles of the logarithm of food
consumption distribution from 1977 to 2012.
Our sample includes all households whose head is aged 25–85. The sample
includes the poverty subsample of the Panel Study of Income Dynamics and
hence sampling weights are used throughout. To emphasize the distinction
between consumption and spending (which in the case of food may be particu-
larly relevant due to government transfers), we plot the 90th–10th percentile
difference both including and excluding food stamps from our definition of
food consumption.
Clearly, inequality in food consumption is rising. Most of the rise is coming
from a decline in spending at the bottom (not shown separately here). The differ-
ence between the top and intermediate lines shows clearly the insurance value of
government transfers. In particular, during the Great Recession the availability of
food stamps allowed poor households to maintain their food consumption, while
spending declined substantially.
We should also note that some of the lower spending on food by the poorest
households may be due to a decline in the prices of the food items that they purchase
(Broda, Leibtag, and Weinstein 2009). To have some sense about the importance
of price differentials, we also plot the 90th–10th percentile difference allowing the
price deflator to be good-specific (that is, food at home plus food stamps, and food
away from home). Correcting for price differentials has a small effect, although it
is more pronounced in recent years. Because the price of food at home (a neces-
sity consumed in large fractions by households at the bottom of the distribution)
has been steadily declining relative to the price of food away from home (a luxury
consumed in large fractions by households at the top of the distribution), inequality

in consumption is lower when adjusting for these price differences.
The decline in spending on food consumption at the bottom of the distri-
bution may not indicate a decline in caloric intakes. Households may spend
less on food without modifying the caloric intake of the food purchases they
make (as argued by Aguiar and Hurst 2005). Indeed, Singh et al. (2009) report
that energy intake is not statistically different between US adults with income
below the poverty line and those with income above 500 percent of the poverty
line. After all, sugars and fats, which are high in calories, can be consider-
ably less expensive than diets based on vegetables, fruits, whole grains, and
lean meat.
The different qualities of food raise a question: Should an assessment of
inequality in food consumption be based on its monetary cost, its energy content,
its healthfulness, or some other measure of quality? The US Department of Health
and Human Services and the US Department of Agriculture have proposed to
measure diet quality with an index known as the Healthy Eating Index (HEI). The
index gives a 0–10 score to 12 food components (like Total Fruits, Whole Fruits,
and so on). For some “good” components (like Total Vegetables) a higher intake
means a higher score, while for some “bad” components (like Saturated Fat) the
opposite is true. Wang et al. (2014) use data from the 1999–2010 National Health
and Nutrition Examination Survey, and compare the HEI index for people of
different socioeconomic background and education. They find that in the popula-
tion at large the quality of food consumed increases monotonically over the sample
period. However, individuals with low socioeconomic status (defined by those with
less than high school and income below 130 percent of the poverty line, the eligi-
bility threshold for food stamps) make no progress in terms of HEI from 1999 to
2010, while most of the improvements are concentrated among medium and high
socioeconomic status groups. We want to stress that while these differences reflect
changes in the “quality” of food consumed between rich and poor individuals,
they are silent regarding the reasons. One possibility is that tastes for healthy food
changed differently for rich and poor individuals (or that the rich were more recep-
tive or attentive to “eating healthy” campaigns). Another possibility is that salience
was similar but the higher price of healthy food or its lower availability in poor
neighborhoods represent significant “barriers to entry” in healthier eating habits
for poor individuals.
Durable Goods
Consumer durables reflect an element of standard of living that may not be
captured by current spending (as people buy them infrequently). Hence, another
way to look at consumption inequality is to see how many and what type of house-
holds own certain home appliances and durable goods. Also in this case—as we did
in the discussion of food consumption—we stress that the quality of what is being
consumed or purchased can matter substantially in thinking about the welfare
consequences of inequality.
Figure 3
Share Owning Durables in the Top and Bottom Income Deciles
A: Cooking durables B: Refrigerators C: Dishwashers

1.00 1.00 1.00
.98 .80
.95 .96 .60
.94 .40
.90
.92 .20
19
19
19
20
20
20
19
19
19
20
20
20
19
19
19
20
20
20
85
90
95
00
05
10
85
90
95
00
05
10
85
90
95
00
05
10
D: Washer and dryer E: Vehicles F: Entertainment durables
1.00 1.00
.90
.80 .90
.90
.70 .80
.80
.60 .70
.50 .70 .60
.40 .60 .50
19
19
19
20
20
20
19
19
19
20
20
20
19
19
20
20
20
85
90
95
00
05
10
85
90
95
00
05
10
90
95
00
05
10
Top income deciles
Bottom income deciles
Source: Authors using data from the Consumer Expenditure Survey

Note: For different categories of durables, the figure compares ownership rates of the bottom and the
top after-tax income deciles.
Given that the Panel Study of Income Dynamics has no information on durables
except cars, we use the Consumer Expenditure Survey, which contains consistent
series on durable goods ownership over a long period time. For some appliances
like refrigerators, washing machines, and others, “availability” is probably a more
appropriate term than “ownership” if such items are attached to the housing unit.
Figure 3 offers a comparison of ownership rates for the bottom and the top after-tax
income deciles. For some categories, we have a long series from 1984 to 2012; for
others, the series starts in 1989. For most categories, there is evidence of catching
up. For example, at the beginning of the time period, ownership of cooking dura-
bles (stoves or microwaves) and refrigerators is almost universal among the top 10
percent households, while the proportion of households in the bottom income
decile owning such appliances is below 90 percent. For refrigerators at the start of
the period, the difference is less but still noticeable. By 2012, these differences have
largely disappeared. While there is convergence for these categories, the catch-up
rates for dishwashers and for washers and dryers are much slower. Ownership of
cars has also converged, albeit at a slower rate than food-related appliances. Finally,
there is only a small difference in the fractions owning entertainment durables

(TVs, sound systems, DVD players, PCs, and so on) throughout the entire period.
There are two caveats to these findings. First, a convergence in ownership does
not imply convergence in the number of appliances owned. Indeed, for the dura-
bles for which this information is available (vehicles and entertainment), there is no
evidence of convergence. Moreover, there may be a large quality difference between
high-end and low-end appliances, but the existing data are not rich enough to
measure the quality of the durables owned by socioeconomic status.
Inequality in the Consumption of Leisure
Economists traditionally write the utility function of individuals as comprising

consumption and leisure. Perhaps greater inequality in the consumption of goods
and services is being partially offset by greater equality in leisure time?
Measuring leisure is complex. Aguiar and Hurst (2009) have looked at trends
in time use as a way of measuring leisure time. We follow a similar strategy here.
In particular, we use surveys collecting information on time use over the last 50
years: the 1965–66 Americans’ Use of Time, the 1975–76 Time Use in Economics
and Social Accounts, the 1985 American Use of Time, and the 2003–11 integrated
American Time Use Survey. The datasets do not include detailed or consistent
information on income or on other measures of economic resources. Thus, we will
use education as a rough measure of socioeconomic status.
To display the sharpest differences, we consider only the top and bottom educa-
tion categories: individuals with less than a high school degree, and those with at
least some college. In these datasets, people report the number of minutes they
spend in various activities in the previous 24 hours. The main time use categories
are: “work” (including time spent searching for jobs), “chores” (all household activi-
ties such as cooking, cleaning, and others), “child care,” “social” (watching sports,
going to movies, partying, and so on), “organizational” (for example, volunteering,
religion), “personal care” (sleeping, eating, and so on), “shopping,” “education,”
“active leisure” (sport activities, playing games, and others), and “passive leisure”
(watching television, listening to radio, relaxing, and others). All figures are
weighted with sampling weights (except 1965–75 where no weights were released)
and expressed in hours per week.
In Figure 4, we plot trends in total leisure time—the sum of social activities, active
and passive leisure, and time devoted to personal care—controlling for the day of the
week the diary was filled in. This measure of leisure may not be without problems. For
example, it includes time spent assisting or helping adult household members (which
for some people may represent “chores”); it excludes gardening or cooking (which for
some individuals may represent a form of leisure). However, excluding personal care
does not affect the main trends. The less-than-high-school education group has more
leisure, and in the last few decades most of the growth of leisure has happened for this
group, too. Moreover, growth has been stronger for men than women.
Figure 4
Trends in Total Leisure Time
A: Total leisure, men B: Total leisure, women

130 130
120 120
110 110
100 100
1960 1970 1980 1990 2000 2010 1960 1970 1980 1990 2000 2010
Less than high school

Some college +
Source: Authors using data from the 1965–66 Americans’ Use of Time, the 1975–76 Time Use in
Economics and Social Accounts, the 1985 American Use of Time, and the 2003–11 integrated American
Time Use Survey.
Note: In the figure, we plot trends in total leisure time—the sum of social activities, active and passive
leisure, and time devoted to personal care.
How do we interpret these trends? It is tempting to argue that the rise in

consumption inequality has been to some extent counterbalanced with increasing
leisure time among the poor, implying that the increase in the inequality of well-
being is less severe than what consumption data alone may suggest. But any such
conclusion needs to be hedged around with cautions. Some of the increase in
leisure is involuntary, due to lack of job market opportunities. Excluding recession
years from the analysis gives the same broad picture of Figure 4. Moreover, if we
repeat the analysis only for the employed, we find that the differences are smaller
but still significant (especially in the 2000s). Of course, an analysis that conditions
on employment does not solve the problem of making welfare comparisons across
income groups that include leisure, as employment itself results from and is affected
by a combination of supply and demand factors. It is also possible that there is
substantial heterogeneity in preferences for leisure—which may also help to explain
different educational choices in the first place.
What component of leisure time is driving these trends? In Figure 5, we decom-
pose total leisure into three components: personal care, active leisure (plus social
activities), and passive leisure. There are some notable trends. First, for both low-
educated women and (especially) men, the increase in total leisure time visible from
Figure 5
Decomposing Total Leisure
(in hours per week)
A: Personal care, women B: Passive leisure, women C: Active leisure, women

82 30 12
80
25 10
78
76 20 8
74
15 6
19
19
19
19
19
19
20
20
19
19
20
20
19
19
19
19
20
20
60
70
60
70
80
90
00
10
80
90
00
10
60
70
80
90
00
10
D: Personal care, men E: Passive leisure, men F: Active leisure, men

80 35
12
78 30 11
76 10
25
9
74 20 8
72 15 7
19
19
19
19
19
19
19
19
19
19
19
19
20
20
20
20
20
20
60
70
60
70
60
70
80
90
80
90
80
90
00
10
00
10
00
10
Less than high school

Some college +
Source: Authors using data from the 1965–66 Americans’ Use of Time, the 1975–76 Time Use in
Economics and Social Accounts, the 1985 American Use of Time, and the 2003–11 integrated American
Time Use Survey
Note: Here we decompose total leisure into three components: personal care, active leisure (plus social
activities), and passive leisure.
Figure 4 is coming primarily from an increase in time devoted to passive leisure

activities. Second, time spent on active leisure and social activities also increases,
resulting in greater similarity between high- and low-educated individuals. Finally,
time spent on personal care is stable. It is possible that these changes represent an
evolution of preferences for leisure across education (income) groups. It is also
possible that increasing availability of durable goods in the lower-income groups
(documented earlier) frees up time previously devoted to housework.
Consumption Mobility
While there is a popular image of the United States as a land where high rates
of mobility across the income distribution are possible, in practice intergenerational
income mobility has not changed much over the last 40 years (Chetty, Hendren,
Kline, and Saez 2014). In fact, some European countries display more intergenera-
tional income mobility than does the United States (Black and Deveraux 2011).
There are also vast geographical differences in mobility across US regions.
Do the trends in intergenerational mobility in consumption mirror those found
for income? To study this topic, we need longitudinal information on consump-
tion that follows multiple generations. The Panel Study of Income Dynamics offers
such data. In particular, for each household, it follows the children, the “splitoff”
households, when they leave the parental home. A few authors have looked at the
intergenerational dimension of the PSID data in the context of risk-sharing within
the family (for example, Hayashi, Altonji, and Kotlikoff 1996; Attanasio, Hurst, and
Pistaferri 2015).
We first construct a measure of household consumption using the Panel Study
of Income Dynamics data. Specifically, we define consumption in this data as the
sum of spending on food, rent, health, home insurance, utilities, car insurance, car
repair, gasoline, parking and transportation, education, child care, clothing, vaca-
tion, and entertainment (with the last three categories only available since 2005).
We add the monetary value of food stamps and imputed rents for homeowners and
free-rent households. To obtain a measure of consumption and income, we deflate
by the Consumer Price Index and as before use the OECD adult equivalence scale.
This measure of consumption is available for the survey years 1999–2013. To reduce
the impact of measurement error, we take moving averages across three subsequent
surveys for both consumption and income.
For each year in which the household is observed, we compute the percentile
occupied by the household relative to the head’s reference birth cohort (born in the
1900s, 1910s, and so on). We do this for the father and his children. Next, we look
graphically at the relationship between the average percentile occupied by the chil-
dren and the percentile occupied by the father. If there is no relationship between
the ranks of parents and children, the (local) regression lines we plot separately
for consumption and income in Figure 6 should be flat; on the other hand, perfect
correlation between the ranks of parents and children would give a 45-degree line.4
We find that the slope of the local regression line for income gradient is
higher than that for consumption, implying greater intergenerational mobility in
consumption than income. This finding is especially true at the bottom of the distri-
bution. Hence, as consumption is more equally distributed than income, there is
also more intergenerational mobility when looking at consumption than income.5
4
If we follow the suggestion of Chetty, Hendren, Kline, and Saez (2014) of conditioning on the age of
parent and child, we get similar, though less-precise, results. Interestingly, the relationship we plot in
the right panel of Figure 6 is remarkably similar to that reported by Chetty et al., despite the enormous
differences in sample sizes.
5
Wodon and Yitzhaki (2002) extend the traditional Sen (1973) welfare function to the dynamic case.
Sen’s welfare function increases with aggregate income and declines with inequality. Wodon and Yitzhaki’s
social welfare function increases with intergenerational mobility. Hence, social welfare is higher when
considering consumption than when considering income not only because of less unequal distribution
of consumption (relative to income), but also because of higher intergenerational mobility.
Figure 6
Intergenerational Mobility in Consumption and Income
A: Mobility in consumption B: Mobility in income

80 80
Average consumption percentile of child
Average income percentile of child

70 70
60 60
50 50
40 40
30 30
20 20
Bandwidth = 0.8 Bandwidth = 0.8

10 10
0 20 40 60 80 100 0 20 40 60 80 100
Consumption percentile of father Income percentile of father
Source: Authors using Panel Study of Income Dynamics data.

Note: If there is no relationship between the ranks of parents and children, the regression lines we plot
separately for consumption and income should be flat; on the other hand, perfect correlation between
the ranks of parents and children would give a 45-degree line. See text for details.
The explanation of the former phenomenon is in all likelihood the tendency to

smooth-out income shocks whenever possible (through saving and borrowing,
public programs, or informal mechanisms). As for intergenerational mobility, one
can conjecture that parents transfer genetic endowments of ability (which will
be reflected in both consumption and income) as well as preferences (which will be
reflected primarily in consumption). However, the extent of similarity between
the consumption of parents and the consumption of children also depend on the
credit and insurance market frictions faced by the two generations. (Similarities
can also depend on the point of the life-cycle we are observing father and child,
but we neglect this complication here.) In the end, whether there is more or
less intergenerational mobility in income or consumption is an empirical matter,
and the data we present constitute one of the first pieces of relevant evidence
in this regard.
Conclusion
The goal of this paper has been to discuss what we do and do not know about
the evolution of consumption inequality in the United States, while contrasting
it with trends in income inequality. There is now some cumulating evidence
showing that increasing disparities in income are approximately replicated by

increasing disparities in consumption. These findings suggest that a substantial
portion of the nature of the shocks to income and wages that have generated
the observed and well-documented increase in income inequality over the
last 35 years should be viewed as permanent rather than temporary, and that
households have only a limited ability to absorb such shocks for more than a
short period.
While much attention in discussions of inequality has been given to the top
of the income and consumption distribution, the left tail is also of considerable
interest, both from a scientific and policy point of view. We have considered
different components of consumption, inequality in leisure, and also the inter-
generational transmission of consumption inequality. When looking at indi-
vidual components, some of them show greater equality, and some raise difficult
questions of how to adjust for quality changes. Ownership of major durables,
which in principle raise living standards, has also been converging rapidly
between low- and high-permanent-income households. While inequality in
food consumption has increased, there is little evidence of growing inequality
in caloric intakes—partly as a result of assistance provided by government
programs (like food stamps) supplementing private spending, partly from price
declines of some food items, and probably in part because low-income people
spend more time searching for lower prices. The latter trend involving time
use is actually more general: the consumption of leisure has increased among
low-socioeconomic status individuals at a faster pace than among the higher
educated.
What do we conclude about whether disparities in well-being have increased?
Our opinion is that, despite the fact that some studies have suggested the opposite,
inequality in the consumption of nondurables and services has increased substan-
tially over the last few decades and has paralleled the increase in inequality in income
and earnings. A consequence of this is that the increase in income inequality is
reflected in an increase in inequality in welfare and well-being.
Some important caveats, however, are in order. We have provided evidence
on specific goods, leisure, durable ownership, and even mobility across genera-
tions that points to relative utility gains realized by the lower-income groups. These
relative gains arise as a consequence of lower prices for the goods they typically
purchase, increasing availability of leisure time, increasing durable ownership, and
improved consumption opportunities for their children. Obviously, assigning a
value to these utility gains is hard, and we do not attempt to do this. It should also
be pointed out that most of these gains have been in quantity terms, not quality
terms.
This discussion suggests that if using consumption is in principle a better way
than income to measure the well-being of households, a complete welfare analysis
will need to go beyond looking at aggregate categories of household expenditure,
and consider in addition the value that people assign to time and the quality of
goods they consume, among other factors.
References
Abowd, John M., and David Card. 1989. “On Baker, Scott R. 2014. “Debt and the Consump-
the Covariance Structure of Earnings and Hours tion Response to Household Income Shocks.”
Changes.” Econometrica 57(2): 411–45. April. http://web.stanford.edu/~srbaker/Papers/
Aguiar, Mark A., and Mark Bils. 2011. “Has Baker_DebtConsumption.pdf.
Consumption Inequality Mirrored Income Bee, Adam, Bruce D. Meyer, and James X.
Inequality?” NBER Working Papers 16807. Sullivan. 2015. “The Validity of Consumption
Aguiar, Mark, and Mark Bils. 2015. “Has Data: Are the Consumer Expenditure Interview
Consumption Inequality Mirrored Income and Diary Surveys Informative?” In Improving the
Inequality?” American Economic Review 105(9): Measurement of Consumer Expenditures, edited by
2725–56. Christopher D. Carroll, Thomas F. Crossley, and
Aguiar, Mark, and Erik Hurst. 2005. “Consump- John Sabelhaus. University of Chicago Press.
tion vs. Expenditure.” Journal of Political Economy Black, Sandra, and Paul Deveraux. 2011.
113(5): 919–48 “Recent Developments in Intergenerational
Aguiar, Mark, and Erik Hurst. 2007. “Measuring Mobility.” Chap. 16 in Handbook of Labor Economics,
Trends in Leisure: The Allocation of Time over vol. 4, Part B, edited by Orley Ashenfelter and
Five Decades.” Quarterly Journal of Economics David Card. North Holland Press, Elsevier.
122(3): 969–1006. Blundell, Richard, Luigi Pistaferri, and Ian
Attanasio, Orazio, Erich Battistin, and Hide- Preston. 2008. “Consumption Inequality and
hiko Ichimura. 2007. Chap. 17 in “What Really Partial Insurance.” American Economic Review
Happened to Consumption Inequality in the 98(5): 1887–1921.
United States?” In: Hard-to-Measure Goods and Blundell, Richard, Luigi Pistaferri, and Itay
Services: Essays in Honor of Zvi Griliches, edited by Saporta-Eksten. 2016. “Consumption Inequality
Ernst E. Berndt and Charles R. Hulten. National and Family Labor Supply.” American Economic
Bureau of Economic Research. Review 106(2): 387–435.
Attanasio, Orazio P., and Steven J. Davis. 1996. Broda, Christian, Ephraim Leibtag, and
“Relative Wage Movements and the Distribution of David E. Weinstein. 2009. “The Role of Prices in
Consumption.” Journal of Political Economy 104(6): Measuring the Poor’s Living Standards.” Journal of
1227–62. Economic Perspectives 23(2): 77–97.
Attanasio, Orazio P., Erik Hurst, and Luigi Broda, Christian, and David E. Weinstein. 2008.
Pistaferri. 2015. “The Evolution of Income, Prices, Poverty and Inequality: Why Americans Are
Consumption, and Leisure Inequality in the US, Better Off Than You Think. Washington, DC: AEI
1980–2010.” Chap. 4 in Improving the Measurement Press.
of Consumer Expenditures, edited by Christopher D. Browning, Martin, and Thomas F. Crossley.
Carroll, Thomas F. Crossley, and John Sabelhaus. 2001. “The Life-Cycle Model of Consumption and
University of Chicago Press. Saving.” Journal of Economic Perspectives 15(3): 3–22.
Attanasio Orazio, and Luigi Pistaferri. 2014. Browning, Martin, and Søren Leth-Petersen.
“Consumption Inequality over the Last Half 2003. “Imputing Consumption from Income and
Century: Some Evidence Using the New PSID Wealth Information.” Economic Journal 113(488):
Consumption Measure.” American Economic Review F282–F301.
104(5): 122–26. Chetty, Raj, Nathaniel Hendren, Patrick Kline,
Attanasio, Orazio, and Guglielmo Weber. and Emmanuel Saez. 2014. “Where is the Land of
2010. “Consumption and Saving: Models of Inter- Opportunity? The Geography of Intergenerational
temporal Allocation and Their Implications for Mobility in the United States.” NBER Working
Public Policy.” Journal of Economic Literature 48(3): Paper 19843.
693–751. Crossley, Thomas F., and Joachim K. Winter.
Autor, David, Andreas Ravndal Kostøl, and 2015. “Asking Households about Expenditures:
Magne Mogstad. 2015. “Disability Benefits, What Have We Learned?” Chap. 1 in Improving the
Consumption Insurance, and Household Labor Measurement of Consumer Expenditures, edited by
Supply.” https://bfi.uchicago.edu/research/ Christopher D. Carroll, Thomas F. Crossley, and
working-paper/disability-benefits-consumption- John Sabelhaus. University of Chicago Press.
insurance-and-household-labor-supply-0. Cutler, David M., and Lawrence F. Katz. 1991.
Aydin, Deniz. 2015. “The Marginal Propensity “Macroeconomic Performance and the Disadvan-
to Consume Out of Liquidity.” http://stanford. taged.” Brookings Papers on Economic Activity no. 2,
edu/~daydin/DAydin_MPCL.pdf. pp. 1–74.
Cutler, David M., and Lawrence F. Katz. 1992. Dynamics 13(1): 15–51
“Rising Inequality? Changes in the Distribution of Kaplan, Greg, and Guido Menzio. 2015. “The
Income and Consumption in the 1980s.” NBER Morphology of Price Dispersion.” International
Working Paper 3964. Economic Review 56(4): 1165–1206.
Deaton, Angus, and Margaret Grosh. 2000. Kaplan, Greg, Guido Menzio, Leena Rudanko,
“Consumption.” Chap. 5 in Designing Household and Nicholas Trachter. 2016. “Relative Price
Survey Questionnaires for Developing Countries: Dispersion: Evidence and Theory.” NBER Working
Lessons from 15 Years of the Living Standards Measure- Paper 21931.
ment Study, Vol. 1, edited by Margaret Grosh and Koijen, Ralph, Stijn Van Nieuwerburgh, and
Paul Glewwe. Washington, DC: World Bank. Roine Vestmanz. 2015. “Judging the Quality
DeBacker, Jason, Bradley Heim, Vasia Panousi, of Survey Data by Comparison with ‘Truth’ as
Shanthi Ramnath, and Ivan Vidangos. 2013. “Rising Measured By Administrative Records: Evidence
Inequality: Transitory or Persistent? New Evidence from Sweden.” Chap. 11 in Improving the
from a Panel of U.S. Tax Returns.” Brookings Papers Measurement of Consumer Expenditures, edited by
on Economic Activity, Spring, pp. 67–142. Christopher D. Carroll, Thomas F. Crossley, and
De Giorgi, Giacomo, Anders Frederiksen, and John Sabelhaus. University of Chicago Press.
Luigi Pistaferri. 2015. “Consumption Network Kopczuk, Wojciech, Emmanuel Saez, and Jae
Effects.” November 23. https://www.economics. Song. 2010. “Earnings Inequality and Mobility in
utoronto.ca/index.php/index/research/ the United States: Evidence from Social Security
downloadSeminarPaper/60199. Data since 1937.” Quarterly Journal of Economics
Dynarski, Susan, and Jonathan Gruber. 1997. 125(1): 91–128.
“Can Families Smooth Variable Earnings?” Brooking Krueger, Dirk, and Fabrizio Perri. 2006.
Papers on Economic Activity 1: 229–305. “Does Income Inequality Lead to Consumption
Friedman, Milton. 1957. A Theory of the Consump- Inequality? Evidence and Theory.” Review of
tion Function. Princeton University Press. Economic Studies 73(1): 163–93.
Gelman, Michael, Shachar Kariv, Matthew MaCurdy, Thomas E. 1982. “The Use of Time
Shapiro, Dan Silverman, and Steven Tadelis. 2014. Series Processes to Model the Error Structure of
“Harnessing Naturally Occurring Data to Measure Earnings in a Longitudinal Data Analysis.” Journal
the Response of Spending to Income.” Science, July of Econometrics 18(1): 82–114.
11, 345(6193): 212–15. Meyer, Bruce D., Wallace K. C. Mok, and James
Gottschalk, Peter, and Robert Moffitt. 1994. X. Sullivan. 2015. “Household Surveys in Crisis.”
“The Growth of Earnings Instability in the U.S. Journal of Economic Perspectives 29(4): 199–226.
Labor Market.” Brookings Papers on Economic Activity Meyer, Bruce D., and James X. Sullivan. 2013.
25(2): 217–72. “Consumption and Income Inequality in the U.S.
Griffith, Rachel, Ephraim Leibtag, Andrew since the 1960s.” https://www3.nd.edu/~jsulliv4/
Leicester, and Aviv Nevo. 2008. “Timing and Quan- Inequality3.6.pdf.
tity of Consumer Purchases and the Consumer Modigliani, Franco, and Richard Brumberg.
Price Index.” NBER Working Paper 14433. 1955. “Utility Analysis and the Consumption Func-
Gross, David B., and Nicholas S. Souleles. tion: An Interpretation of Cross-Section Data.”
2002. “Do Liquidity Constraints and Interest Rates Chap. 15 in Post Keynesians Economics, edited by
Matter for Consumer Behavior? Evidence from Kenneth K. Kurihara. Rutgers University Press.
Credit Card Data.” Quarterly Journal of Economics Nevo, Aviv, and Arlene Wong. 2015. “The
117(1): 149–85. Elasticity of Substitution between Time and
Guvenen, Fatih, Serdar Ozkan, and Jae Song. Market Goods: Evidence from the Great Reces-
2014. “The Nature of Countercyclical Income sion.” http://sites.northwestern.edu/awo760/
Risk.” Journal of Political Economy 122(3): 621–60. files/2015/10/Paper_Aug19_2015-y5zgke.pdf.
Handbury, Jessie. 2014. “Are Poor Cities Cheap Passero, William, Thesia I. Garner, and Clinton
for Everyone? Non-Homotheticity and the Cost of McCully. 2015. “Understanding the Relationship:
Living Across U.S. Cities.” Unpublished paper. CE Survey and PCE.” Chap. 6 in Improving the
Hayashi, Fumio, Joseph Altonji, and Lawrence Measurement of Consumer Expenditures, edited by
Kotlikoff. 1996. “Risk-Sharing between and within Christopher D. Carroll, Thomas F. Crossley, and
Families.” Econometrica 64(2): 261–94. John Sabelhaus. University of Chicago Press.
Heathcote, Jonathan, Fabrizio Perri, and Pistaferri, Luigi. 2015. “Household Consump-
Gianluca L. Viuolante. 2010. “Unequal We Stand: tion: Research Questions, Measurement Issues,
An Empirical Analysis of Economic Inequality in and Data Collection Strategies.” Journal of Economic
the United States, 1967–2006.” Review of Economic and Social Measurement 40(1–4): 123–49.
Sabelhaus, John, and Jeffrey A. Groen. 2000. Kerr, and Connie M. Weaver. 2009. “Comparison
“Can Permanent-Income Theory Explain Cross- of Self-Reported, Measured, Metabolizable Energy
Sectional Consumption Patterns?” Review of Intake with Total Energy Expenditure in Over-
Economics and Statistics 82(3): 431–38. weight Teens.” American Journal of Clinical Nutrition
Sabelhaus, John, David Johnson, Stephen Ash, 89(6): 1744–50.
David Swanson, Thesia I. Garner, John Greenlees, Slesnick, Daniel T. 1991. “The Standard of
Steve Henderson. 2015. “Is the Consumer Expen- Living in the United States.” Review of Income and
diture Survey Representative by Income?” Chap. Wealth 37(4): 363–86.
8 in Improving the Measurement of Consumer Expen- Slesnick, Daniel T. 1994. “Consumption, Needs
ditures, edited by Christopher D. Carroll, Thomas and Inequality.” International Economic Review
F. Crossley, and John Sabelhaus. University of 35(3): 677–703.
Chicago Press. Slesnick, Daniel T. 2001. Consumption and Social
Sen, Amartya. 1973. On Economic Inequality. Welfare. Cambridge Books.
Clarendon Press. Wang, Dong D., Cindy W. Leung, Yanping Li,
Singh, Gopal K., Mohammad Siahpush, Robert Eric L. Ding, Stephanie E. Chiuve, Frank B. Hu,
A. Hiatt, and Lava R. Timsina. 2011. “Dramatic and Walter C. Willett. 2014. “Trends in Dietary
Increases in Obesity and Overweight Prevalence Quality among Adults in the United States, 1999
and Body Mass Index among Ethnic-Immigrant through 2010.” JAMA Internal Medicine 174(10):
and Social Class Groups in the United States, 1587–95.
1976–2008.” Journal of Community Health 36(1): Wodon, Quentin, and Shlomo Yitzhaki. 2002.
94–110. “Inequality and Social Welfare.” A Sourcebook for
Singh, Rajni, Berdine R. Martin, Yvonne Poverty Reduction Strategies, Vol. 1: Core Techniques
Hickey, Dorothy Teegarden, Wayne W. Campbell, and Cross-Cutting Issues, edited by Jeni Klugman,
Bruce A. Craig, Dale A. Schoeller, Deborah Anne 75–104.

Econ 4413 Poverty and Income Distribution

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Econ 4413 Poverty and Income Distribution

Uploaded by

Copyright:

Available Formats

Reading Package

Poverty and Income Distribution

Fall Semester 2019-20

Lahore University of Management Sciences

Professor Abid A. Burki Phone: 042-35608076, Ext. 8076

Grading Breakup and Policies:

Recommended Texts & Readings:

Sessions and Readings:

2. INCOME STANDARDS AND INEQUALITY

4. ESTIMATING WAGE INEQUALITY

7. UNDERSTANDING DETERMINANTS OF POVERTY

Income Inequality in Department of Economics

Pakistan May 10, 2018

Evolution of Kuznets has been

Taxes and top

Source: World Inequality Report 2018

All inter-temporal inequality Lack of data is a major hurdle

Probability density function

Median 2015 0.0014

Region 2001-02 2015-16 Growth 2001-02 2015-16 Growth Change

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Note: Author’s calculations based on HIES 2001-02 and PSLM-HIES 2015-16.

Bottom 50% captured 32.4% of total

• Effects of fathers’ earning on sons earnings is measured by inter-generational earning elasticity.

Lack of tax revenue puts

• Income inequality in top 1% of incomes in Pakistan is unknown.

The Top 1 Percent in International and

Facundo Alvaredo, Anthony B. Atkinson, Thomas

■ Facundo Alvaredo is Research Fellow at Nuffield College and Department of Economics,

We should start by emphasizing the factual importance of the top 1 percent.

The United States Top 1 Percent in International Perspective

A: Top 1 Percent Income Shares in English-speaking Countries (U-Shape)

Source: The World Top Incomes Database.

Taxes and Top Shares

Elasticity = .47 (.11)

–40 –30 –20 –10 0 10

Richer Models of Pay Determination

Top Tax Rates and Growth

Capital Income and Inheritance

The analysis just cited focused—like much of the literature— on what is

wealth-holders to sustain their preeminence. The key factor in determining the

disposable income.6 Two methods are employed: a constructive calculation from

countries than in others.7 Another source of difference between countries could

Joint Distribution of Earned and Capital Income

A: Percent of top 1% capital incomes in various top labor income groups

B: Percent of top 1% labor incomes in various top capital income groups

Source: Aaberge, Atkinson, Königs, and Lakner (forthcoming).

To understand the changing relationship between earned and capital incomes,

Income Standards, Inequality,

This chapter complements the introductory chapter by providing a detailed

1. Identification, in which individuals are identified as poor or nonpoor

However, to identify individuals as poor or nonpoor, we need to select a

1. Space selection, which is described below

• An individual’s consumption is the destruction of goods and services

Whether it is income or consumption expenditure, welfare indica-

There are various types of equivalence scales and economies of scale.

Suppose our reference society X consists of N people, where the income of

There are two different ways to represent an income distribution:

• The simplest income distribution is a vector of incomes. We denote

Another, more intuitive, presentation of the cdf is the quantile function,

An income distribution’s density function reports the percentage of the popu-

Figure 2.1: Probability Density Function

distribution x. Hence, 50 percent of the area underneath fx lies to the

In addition, a density function can be a useful tool for understanding the