Professional Documents
Culture Documents
Antonio Villar
Lectures on
Inequality,
Poverty and
Welfare
Lecture Notes in Economics
and Mathematical Systems 685
Founding Editors
M. Beckmann
H.P. K€
unzi
Managing Editors
Prof. Dr. G. Fandel
Fachbereich Wirtschaftswissenschaften
Fernuniversität Hagen
Hagen, Germany
and
Institut f€
ur Mathematische Wirtschaftsforschung (IMW)
Universität Bielefeld
Bielefeld, Germany
Editorial Board
H. Dawid, D. Dimitrov, A. Gerber, C-J. Haake, C. Hofmann, T. Pfeiffer,
R. Slowiński, W.H.M. Zijm
More information about this series at http://www.springer.com/series/300
Antonio Villar
Lectures on Inequality,
Poverty and Welfare
Antonio Villar
Department of Economics
Universidad Pablo de Olavide
Seville, Spain
1
See Goerlich, F., & Villar, A. (2009). Desigualdad y Bienestar: De la Teoría a la Pr
actica.
Fundación BBVA.
vii
viii Preface
Finally, I would like to thank Ricardo Martı́nez for his critical reading and his
numerous suggestions. Needless to say, all remaining errors and imprecisions are
my sole responsibility.
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Equality and Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 We Have a Problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Measuring Inequality and Poverty . . . . . . . . . . . . . . . . . . 4
1.1.3 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Equality of What? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 Income Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.2 Wealth, Income and Expenditures . . . . . . . . . . . . . . . . . . 7
1.2.3 The Units of Reference: Equality Among Whom? . . . . . . . 9
1.2.4 Needs and Deserts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.5 Multidimensionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Plan of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Part I Inequality
2 Inequality Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Properties of Inequality Indices . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Inequality and Dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.1 The Variance (σ 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.2 The Coefficient of Variation (CV) . . . . . . . . . . . . . . . . . . 29
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3 Positive Inequality Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1 The Lorenz Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 The Gini Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.1 From Lorenz to Gini . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.2 The Generalised Gini Index . . . . . . . . . . . . . . . . . . . . . . . 37
3.3 Theil’s Inequality Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
ix
x Contents
Part II Poverty
7 Poverty Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.2 Poverty Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.2.1 Head Count Ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.2.2 Poverty Gap Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.2.3 The Sen’s Family of Poverty Measures . . . . . . . . . . . . . . . 120
7.2.4 Decomposable Poverty Measures . . . . . . . . . . . . . . . . . . . 121
7.3 Multidimensional Poverty Indices . . . . . . . . . . . . . . . . . . . . . . . . 122
7.4 Deprivation and Non-monetary Poverty Measures . . . . . . . . . . . . 125
7.4.1 At Risk of Poverty or Social Exclusion (AROPE) . . . . . . . 125
7.4.2 The United Nations Multidimensional Poverty Index . . . . . 127
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8 Multidimensional Poverty and Welfare . . . . . . . . . . . . . . . . . . . . . . 135
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.2 The Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
8.3 Adding Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.3.1 Quasi-concavity, Scale and Factor Decomposability . . . . . 139
8.3.2 Multiplicative Factor Decomposability . . . . . . . . . . . . . . . 141
8.4 Closing the Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
8.4.1 Anonymity, Inequality and Welfare . . . . . . . . . . . . . . . . . 142
8.4.2 Subgroup Decomposability . . . . . . . . . . . . . . . . . . . . . . . 143
8.5 Measuring Educational Poverty from PISA . . . . . . . . . . . . . . . . . 145
8.5.1 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
8.5.2 The Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
8.5.3 Final Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
9 The Evaluation of Relative Achievements . . . . . . . . . . . . . . . . . . . . . 155
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
9.2 The Basic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
9.2.1 Measuring the Achievements . . . . . . . . . . . . . . . . . . . . . . 157
9.2.2 Application to the Example of the Hypothetical
University . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
9.3 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
9.3.1 Weighted Symmetry and Weighted Neutrality . . . . . . . . . . 162
9.3.2 Generalised Means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
9.3.3 A Graphical Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . 165
9.4 An Application: The Green Economy Progress Index . . . . . . . . . . 166
9.4.1 The Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.4.2 The Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
9.4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Chapter 1
Introduction
Equality is a recurrent issue in social thinking that involves many different aspects:
equality of rights, equality of opportunity, equality of income, etc. Economic
inequality and poverty constitute significant aspects of the asymmetric distributions
of opportunities and can be regarded as the main form of inequality in those
countries in which rights are guaranteed by constitutional laws, and the public
sector ensures the coverage of some minimal services (health, education, unem-
ployment benefits, pensions, etc.). Indeed, most democratic governments make an
explicit goal of reducing economic inequality and poverty in their political agendas.
We devote this introductory chapter to underline some basic ideas that put in
perspective the content of this monograph. First, that inequality and poverty are
relevant problems, in terms of magnitude, extent and implications. Second, that
those problems deserve attention at social, political and academic levels, as they
affect society’s welfare. Measuring inequality, poverty and welfare is therefore
important and complex as those measures involve ethical judgements (we move in
the realm of normative economics). And third, that there are a number of method-
ological alternatives when it comes to choose the variable whose distribution is to
be measured. Different choices imply different meanings of any evaluation
exercise.
Let us start by recalling a key fact: The differences in living standards across the
world are huge. They appear both between and within countries and are multi-
farious. Those differences range from rights and liberties to public facilities or the
distribution of wealth. Even if one focuses on purely economic inequalities the
1
This agenda extends the former Millennium Development Goals, which focused on developing
countries, and incorporates targets that are relevant for all countries.
1.1 Equality and Welfare 3
2
There is also the impact of the reduction of the prices of some basic assets (housing, in particular),
but the effect of that on inequality is less clear.
4 1 Introduction
fulfilment and self-worth. Policy makers from around the world acknowledged that
inequality in their countries is generally high and potentially a threat to long-term
social and economic development. It is important to stress, though, that inequality
and poverty are not unavoidable: they can be effectively reduced when applying
suitable policies. Several countries have already managed to contain or reduce
income inequality while achieving strong growth performance. It is also important
to note that achieving such a reduction implies reducing the underlying inequality
of opportunities and not only the inequality of outcomes.
Given the extent and relevance of inequality and poverty and their impact on
present and future personal and social welfare, it might seem unnecessary allude
to the need of measuring properly those phenomena. Yet, it is worth commenting on
some elements that convey the interest of such measurement.
The most obvious of those elements is recognising that measurement is an
essential part of the diagnosis: how important is the problem, how is it evolving,
how is the situation of a society relative to the others, who are the social groups
affected, what are the ultimate causes that produce inequality and poverty. From
this, it follows that measurement also helps the design and implementation of
policies to deal with those problems, and it is key to monitoring the effects of policy
measures: are being those measures effective? Do we meet the policy targets? Are
we progressing sufficiently in the reduction of inequality and poverty?
Besides those standard arguments, there is a subtler and eventually more impor-
tant aspect of developing regularly such a measurement. It refers to keep thinking
about people and keep thinking about those problems.
Inequality and poverty are widespread phenomena that have very negative
implications for individuals and societies and affect their possibilities for the future.
Keeping track of the magnitude of those problems helps maintaining them in the
political agenda and reminds us of the need of dealing with them. This might seem
a trivial aspect but it is not so, as less well-off people tend to be less present in the
social and political life, so that their problems may remain hidden as they lack
channels to get recognition and promote action. Poor people are not very visible in
the social and political life and are outside of those interest groups that conform
social needs and influence public policies.
Measuring inequality and poverty also help keeping those problems in the
research agenda. Addressing the diagnosis of the extent and nature of inequality
and poverty, as well as designing and implementing policies to reduce their impact
and monitoring the effectiveness of those policies, requires having appropriate tools
to measure those phenomena. Designing those tools calls for conceptual elabo-
ration, formal modelling and data. All that implies a social effort on the researchers
and the institutions to build up the ideas, models and resources that permit such a
measurement. One can think of the role played by the GDP in the economic
1.1 Equality and Welfare 5
analysis. In spite of the simplicity and limitations of this measure, it has become a
key reference for the evaluation of economic growth with a regular flow of data and
a constant reference on the economic success of the countries. Many economic
policy targets are set around this variable, which is already part of the usual
society’s concerns. Identifying the right measures of inequality and poverty may
lead to a similar outcome.
1.1.3 Purpose
Most of the studies on inequality and poverty take as the reference variable income
or wealth. We mostly follow here this tradition even though we have to make it
clear that: (i) The measurement of economic inequality and poverty may involve
several dimensions; (ii) The degree of development of the countries conditions the
significance of the analysis in terms of market measures; (iii) Inequality and poverty
indices can be applied to many different problems besides those regarding income
and wealth and (iv) Even if we confine our analysis to monetary income, there are a
number of methodological options that have to be made. We shall focus here on this
last aspect and will address the other ones along the different chapters.
substitutability between goods. Yet, when dealing with extreme forms of poverty,
the goods that cover essential needs can hardly be regarded as substitutes. So while
inequality measures are interpreted the same way in most cases, poverty measures
may refer to different aspects of this phenomenon depending on the degree of
development of the country under consideration.
Measuring inequality and poverty requires making some methodological
choices, before having to decide on which index to use. As those choices determine
the meaning of our evaluation exercise, understanding their nature and implications
is crucial to get a sensible interpretation of the results we obtain. Those methodo-
logical options involve two main elements. First, we have to decide on the variable
that approaches the standard of living, in particular on whether it refers to income or
expenditure. Second, we have to identify the reference units, that is, whether our
analysis refers to individuals, households or consumption units (households
adjusted by size and composition). Only after specifying those two elements can
we meaningfully speak of income distributions. Note that methodological choices
are not mere academic exercises but the way of conveying substance to the analysis
and the key to interpret the outcomes.
We shall briefly refer to those topics here just to raise a flag of warning on the
relevance of those aspects.
parity (PPP) with respect to a given base year. More difficult is computing real
values when relative prices change and affect differently different social groups.
This may have a relevant impact for the measurement of poverty in developing
countries in which markets are not fully developed or are not accessible to all social
groups. In those cases, prices express the relative evaluation of commodities
determined by those agents who act in the markets and do not reflect the demands
or needs of the less well-off groups. Moreover, the substitutability that the price
system implies only makes sense when the basic needs are covered (there is little or
no trade-off between primary goods).
Thirdly, the income variable can be regarded as a measure of earnings (which is
the standard interpretation) or as a measure of consumption (expenditures). These
two variables do not behave in a similar way so that our measurement of inequality
turns out to depend very much of this choice (Atkinson & Brandolini, 2001). There
is no general agreement on which variable is better, as there are pros and cons in
choosing one or another (e.g. Atkinson & Bourguignon, 2000; Blundell & Preston,
1996; Brewer & O’Dea, 2012; Deaton & Zaidi, 2002; Krueger & Perri, 2006;
Slesnick, 1991, 1993). The main reason to choose consumption as the reference
variable is that it incorporates the economic units’ saving and borrowing decisions
and is thus less sensitive to the economic cycle.3 This is also the reason why some
authors prefer identifying income with earnings: since the expenditures in con-
sumption involve changes in the wealth, using this variable in isolation may hide
relevant changes in the agents’ assets. This is problematic because the available
information on wealth distribution is very limited. Moreover, there is the problem
of the difference between consumption and consumption expenditures mostly due
to the difficulty of properly computing the services rendered by durables and public
facilities (Deaton & Zaidi, 2002). In some cases, we find data on income or
expenditures that estimate non-monetary earnings and the services of some dura-
bles (e.g. imputing the implicit rent of owned houses). As not all countries apply the
same protocols to make those imputations, it is important to be careful on the
precise definition of the reference variable.
Fourthly, one has to take into account the effect on people’s consumption
opportunities of taxes, transfers and the public goods provided. The pre-tax,
pre-transfer income is usually called market income, whereas the after tax and
transfer income is called disposable income. Taxes and transfers affect substantially
the distribution of income, reducing the observed inequality and poverty depending
on the degree of progressivity of the fiscal system. Much more difficult is to
attribute the implicit income, both for earnings and consumption expenditures,
derived from the free access to some public goods (e.g. education, health). Given
the size of the public expenditures relative to the GDP and the asymmetric distri-
bution of benefits and contributions, this is also a source of improvement of
3
There is some evidence that surveys on expenditures tend to be more reliable than surveys on
earnings, due to the tendency to hide part of the earnings for fiscal reasons.
1.2 Equality of What? 9
inequality and poverty. Yet this is an element hardly taken into account in a
systematic way in empirical applications.
And, of course, there is the difficulty of data availability, especially when trying
to cover long periods and many different countries. Yet there are nowadays many
databases with a wide coverage and a good level of detail, which permit many
evaluation exercises. Besides national statistical offices, we find data internation-
ally comparable in the main international economic institutions, such as Eurostat,
the World Bank, the United Nations or the OECD, to name a few, and specific
databases such as the Luxembourg Income Studies (LIS) or the Standarised World
Income Inequality Database (SWIID).
We have stressed that our focus is that of the distribution of personal income. Yet
one has to specify whether such a distribution refers to individuals or to households.
Households would seem the most natural option, as it is within households (includ-
ing single-member households) where consumption opportunities develop. Let Yh
denote the income of household h and let Y¼(Y1, Y2,. . .,Ym) stand for the distribu-
tion of income in a society made of m households. Note that focusing on this type of
income distributions amounts to disregard the differences in the size and composi-
tion of the households. Adjusting by the size of households is immediate and simply
requires computing the average income of the household, Yh/nh, where nh is the
number of members of household h (this is the so-called per capita approach). The
limitation of this approach is twofold. On the one hand, it ignores that households
with the same number of members may have different needs due to the differences
in their composition. On the other hand, it disregards the economies of scale
associated with the size of the households (mostly associated to the consumption
of durables).
The economic unit approach tries to avoid some of those problems by adjusting
the households’ income by size and composition according to some equivalence
scale. There are different ways of scaling households to define those consumption
units.4 A case in point is the convention used by Eurostat (also known as the
modified OECD), according to which the first adult in an economic unit counts as
1, all other adults as 0.5 and children as 0.3. Let nha, nhc stand for the number of
adults and children of household h. The income per consumption unit of this
household is thus given by Yh/neh , where nhe ¼ 1 þ 0:5ðnha 1Þ þ 0:3nhc is the
equivalised size of the household. The ratio Yh/neh can be regarded as a monetary
4
See Atkinson, Rainwater, and Smeeding (1995), McClements (1977), Pollak and Walles (1979),
Buhmann, Lee, Schmaus, and Smeeding (1988), Cowell and Jenkins (2000), Banks and Johnson
(1994), Jenkins and Cowell (1994), Ruiz-Castillo (1995), Cowell and Mercader-Prats (1999) and
Ebert and Moyes (2003). For a general discussion see Goerlich and Villar (2009, Ch. 13).
10 1 Introduction
The choice of the units of reference is just a particular case of a much more
important problem: how to evaluate inequality and poverty when agents’ are
heterogeneous? In particular, how to take into account the differences in agents’
characteristics, particularly regarding circumstances, needs and deserts?
Let us think of the case of income inequality to illustrate this point. Suppose that
we have already solved all the difficulties related to the measurement of that
variable and the definition of the economic units, which we shall refer to as agents.
Suppose, furthermore, that we have a function that measures the inequality that is
observed in the income distribution of a given society. Is that observed inequality
what we really want to measure? Or, put differently, is all observed inequality
unfair?
This problem can be easily illustrated as follows. Imagine a society in which all
workers have the same salary but they choose to work a different number of hours
per year, as the result of their preferences between income and leisure. Should we
interpret the resulting differences in income distribution as inequality? More
generally, one can think of the observed inequality as the result of the interplay
of different variables, such as effort, luck and circumstances. There is consensus on
the unfairness of income differences that are due to differential circumstances
(aspects that affect people’s capabilities but are beyond their decision possibilities).
There is some debate on the role of the other variables. The equality of opportunity
approach (Roemer, 1998), for instance, assumes that inequality due to differential
effort is ethically admissible and should be disregarded.
5
Note that both in the per capita approach and in the consumption unit approach, one has still to
decide if the relevant income vector is that in which we substitute the income of each household by
the corresponding mean (per capita or per equivalised number of members) or by nh or neh times
that value. See Ebert (1995, 1997, 1999) and Ebert and Moyes (2003) for a discussion.
1.3 Plan of the Book 11
1.2.5 Multidimensionality
There are many other aspects that may affect inequality in a society, aspects that
may interact but that do not follow a common pattern (e.g. income, education,
health). Dealing with this situation requires the design of multidimensional inequal-
ity indicators.
These Lectures are organised into two parts. The first part is devoted to the analysis
of income inequality by means of inequality indices. The second one deals with the
analysis of poverty, again centred on the study of poverty indices. We shall touch
only a limited number of issues and will try to get a sound understanding of them.
Those who are interested in a more extensive coverage may refer to the standard
monographs, among which we single out the following (references on specific
topics will be mentioned along the different Lectures): Sen (1973, 1992), Love
and Wolfson (1976), Cowell (1977, 1995, 2003), Kakwani (1980), Osmani (1982),
Anand (1983), Atkinson (1983), Chakravarty (1990, 2009), Lambert (1993),
Temkin (1993), Sen and Foster (1997), Silber (1999), Atkinson and Bourguignon
(2000), Duclos and Araar (2006), Goerlich and Villar (2009), Houghton and
Khander (2009), Foster, Seth, Lokshin, and Sajaia (2013) and Alkire et al. (2015).
Let us summarise now the content of the different chapters.
Chapter 2 presents the notion of inequality indices as functions that map the
space of income distributions into the real numbers. We address here two specific
questions. First, we discuss a set of properties, or requirements, that make of a
function a suitable candidate for an inequality index. Those properties include
Normalisation, Symmetry, Population Replication, Principle of Transfers, Conti-
nuity, Scale Independence and Additive Decomposability. Second, we illustrate the
difference between the notions of dispersion and inequality, by analysing the
properties and key features of the variance and its related measures (standard
deviation, coefficient of variation). We show that even though the variance satisfies
most of the properties we may require for an inequality index, it is not a suitable
inequality measure.
Chapter 3 discusses positive inequality measures, as functions that try to provide
descriptive estimates of the variability of income distributions that satisfy most of
the properties presented in Chap. 2 and are compatible with our intuitions about
what inequality means. There are no explicit references to social welfare even
though different measures incorporate implicitly different value judgements.
Indeed, one can think of those measures as particular ways of aggregating the
differences between individual incomes and a reference value, most of the times the
mean income. We review here four different ways of comparing income distribu-
tions from a descriptive point of view: the Lorenz curve, the Gini index, the Theil’s
12 1 Introduction
family of indices and the Palma ratio. We shall check the properties that those
measures satisfy in order to illustrate the different ways of valuing inequality. There
are many other ways of measuring inequality, but these are the most common ones.
Chapter 4 describes the standard approach to normative inequality measurement.
The key idea is to interpret inequality as a welfare loss, when social welfare is
measured by a conventional social welfare function. We focus on the analysis of
Atkinson’s family of inequality indices, which extends the initial ideas of Dalton by
applying some of the notions that are common in expected utility theory. Atkinson
uses the notion of equally distributed equivalent income to evaluate the actual
distribution and to assess the size of the welfare loss. Those inequality indices are
derived from a utilitarian social welfare function applied to individuals with
identical cardinal utility functions. Atkinson generates a family of indicators that
depend on a single parameter, to be understood as a measure of inequality aversion.
Chapter 5 deals with the welfare evaluation of income distributions when agents
are heterogeneous, regarding needs and deserts, focussing on inequality of oppor-
tunity rather than on inequality of outcomes. The basic idea behind this approach is
that individual incomes can be regarded as the result of two different elements: free
choices and external circumstances. It is the inequality associated with the agents’
external circumstances that is unfair, whereas differences derived from free indi-
vidual decisions on a common ground are ethically admissible. We present here two
different ways of approaching the measurement of inequality of opportunity. The
first one is based on the additive decomposability properties of Theil’s index. The
second one applies to the case in which the reference variable is categorical rather
than quantitative and provides an application to the measurement of inequality in
educational outcomes, using PISA data as a reference.
Chapter 6 provides an integrated way of approaching inequality measurement
from a normative viewpoint, by using the notion of social evaluation function,
instead of that of social welfare function. A social evaluation function is a mapping
that is defined directly on the space of income distributions, without going through
the intermediate step of individual utilities. This notion permits extending the
standard normative approach to inequality and provides a general framework in
which all the inequality indices can be confronted in terms of the properties that
imply on this social evaluation function. Besides, we introduce here the notion on
multidimensional inequality and welfare, that applies when more than one relevant
dimension is involved. We illustrate this venue by means of the Human Develop-
ment Index.
Chapter 7 refers to the measurement of poverty, focusing on objective poverty
measures, referred to a single period, taking “income” as the key reference variable
and using a relative poverty line to define who are the poor. It also discusses
multidimensional (objective) poverty measures. Following the standard approach,
a poverty index is conceptualised as a function that combines three different aspects
of poverty: Incidence (how many poor people are in society), Intensity (how poor
they are) and Inequality (how unequal are the poor). We also refer here to some
measures of deprivation and non-monetary poverty indicators.
References 13
References
Alkire, S., Foster, J., Seth, S., Santos, M. E., Roche, J. M., & Ballon, P. (2015). Multidimensional
poverty measurement and analysis OPHI. Oxford: Oxford University Press.
Anand, S. (1983). Inequality and poverty in Malaysia. Oxford: Oxford University Press.
Atkinson, A. B. (1983). The economics of inequality (Vol. 2). Oxford: Clarendon.
Atkinson, A. B., & Bourguignon, F. (2000). Introduction: Income distribution and economics. In
A. B. Atkinson & F. Bourguignon (Eds.), Handbook of income distribution (Vol. 1, Chap. 1).
Amsterdam: Elsevier Science Publishers B. V. North Holland.
Atkinson, A. B., & Brandolini, A. (2001). Promise and pitfalls in the use of ‘secondary’ data-sets:
Income inequality in OECD countries as a case study. Journal of Economic Literature, 39,
771–799.
Atkinson, A. B., Rainwater, L., & Smeeding, T. (1995). Income distribution in OECD countries:
Evidence from Luxembourg income study. Parı́s: OCDE.
Banks, J., & Johnson, P. (1994). Equivalence scale relativities revisited. The Economic Journal,
104, 883–890.
Blundell, R., & Preston, I. (1996). Income, expenditure and the living standards of UK households.
Fiscal Studies, 16, 40–54.
Brewer, M., & O’Dea, C. (2012). Measuring living standards with income and consumption:
Evidence from the UK (IFS Working Papers W12/12). Institute for Fiscal Studies.
Buhmann, B., Lee, R., Schmaus, G., & Smeeding, T. (1988). Equivalence scales, well-being,
inequality, and poverty: Sensitivity estimates across ten countries using the Luxembourg
Income Study (LIS) database. Review of Income and Wealth, 34, 115–142.
Chakravarty, S. R. (1990). Ethical social index numbers. Berlin: Springer.
Chakravarty, S. R. (2009). Inequality, polarization and poverty. Advances in distributional ana-
lysis. Economic Studies in Inequality, Social exclusion and well-being. New York: Springer.
14 1 Introduction
Pollak, R. A., & Walles, T. (1979). Welfare comparisons and equivalent scales. American Eco-
nomic Review, Papers and Proceedings, 69, 216–221.
Roemer, J. E. (1998). Equality of opportunity. Cambridge: Massachusetts Harvard University Press.
Ruiz-Castillo, J. (1995). Income distribution and social welfare: A review essay. Investigaciones
Econ omicas, 19, 3–34.
Sen, A. K. (1973). On economic inequality. Oxford: Clarendon.
Sen, A. K. (1992). Inequality reexamined. Oxford: Oxford University Press.
Sen, A. K., & Foster, J. E. (1997). On economic inequality (2nd ed.). Oxford: Clarendon.
Shorrocks, A., Davies, J., & Lluberas, R. (2014). Global Wealth Databook 2014.
Silber, J. (1999). Income inequality measurement: From theory to practice. Dordrecht: Kluwer Aca-
demic Publishers.
Slesnick, D. T. (1991). The standard of living in the United States. Review of Income and Wealth,
37, 363–386.
Slesnick, D. T. (1993). Gaining ground: Poverty in the postwar United States. Journal of Poli-
tical Economy, 10, 1–38.
Temkin, L. (1993). Inequality. Oxford: Oxford University Press.
United Nations Development Programme. (2010). Human development report 2010: The real wealth
of nations: pathways to human development. Palgrave Macmillan.
Part I
Inequality
Chapter 2
Inequality Indices
2.1 Introduction
Inequality indices are the yardsticks with which we try to assess the extent of
inequality between and within societies. The basic problem is easy to formulate:
given the information on the income distribution in two or more societies, evaluate
how unequal those distributions are.
Let us put this more formally. Consider two societies, N ¼ f1; 2; . . . ; ng, M
¼ f1; 2; . . . ; mg with income distribution vectors y ¼ ðy1 ; y2 ; . . . ; yn Þ and
z ¼ ðz1 ; z2 ; . . . ; zn Þ, respectively. Our basic problem is: (i) Can we say if y is
more or less equitable than z? And (ii) If so, can we say how much?
Answering those questions requires having an evaluation criterion that permits
one comparing income distributions with different dimensions and regarding dif-
ferent societies. Question (i) is an ordinal concern and requires defining just a
ranking (whether one distribution is more unequal than the other). Question
(ii) calls for a cardinal evaluation (measuring how much unequal they are). Obvi-
ously, (ii) implies (i). We shall address here the evaluation problem from a cardinal
viewpoint by recurring to the construction of inequality indices.
From this perspective, an inequality index can be understood as a function
I defined over the space of income distributions, Ω, which attaches a real number
to each conceivable income distribution. That is, a mapping I : Ω ! ℝ that
associates, to each possible income distribution y ¼ ðy1 ; y2 ; . . . ; yn Þ, a real number
that represents the inequality in this distribution.1 Such a function will have to
exhibit those value judgements and operational properties we deem adequate for
such an evaluation exercise. That is, being more unequal should imply a larger
value of the function I : Ω ! ℝ.
1
This is not the only way of thinking of the evaluation of income distributions, but it is the
dominant approach. Other approaches may look for a relative evaluation, that is, evaluating the
inequality of a distribution with respect to others, but not in isolation.
We shall consider next a set of properties that we may require for a function to be
considered as a possible inequality index. Those properties give us insights on the
nature of the evaluation we are performing. Indeed, all conventional indicators
satisfy most or all of those properties.
We focus on the case of discrete income distributions (after all each society
consists of a finite number of agents). To make the discussion easier we shall
assume, for the time being, that the societies we compare only differ in the
distribution of the income variable and the number of agents involved. We therefore
disregard all those differences in society that might affect people’s well-being
beyond income. We shall return on this point in later Lecture. In what follows,
therefore, we shall identify an inequality index with a mapping I : ℝþþ n
! ℝ, where
I(n, y) is the value of the index corresponding to the income distribution vector
y ¼ ðy1 ; y2 ; . . . ; yn Þ. When there is no confusion, we shall take the population size
as an implicit argument and simply write inequality indices as I(y).
Let us consider now some basic properties that one may require, in full or in part,
for a function to be considered an inequality index. We shall provide the rationale
for those properties and also the implications they entail.
Property 1: Normalisation
This property simply says that inequality is zero if and only if all incomes are equal.
Otherwise inequality is positive. That is, for a given society N with n agents, let
1X n
y 2 ℝþþ
n
denote an income distribution vector and let μðyÞ ¼ y stand for the
n i¼1 i
mean income (per capita income). Then I ðyÞ 0, with I ðyÞ ¼ 0 if and only if yi
¼ μðyÞ for all i. A distribution with zero inequality is called perfectly egalitarian.
Remark 2.1 This property sets a lower bound for the inequality index, which may
be unbounded above. Yet most indices will have an upper bound, and some of them
move into the interval [0, 1].
Property 2: Symmetry
This property (also known as anonymity)2 says that the only relevant information
that the index takes into account is the income of the different agents, and not other
personal characteristics. As a consequence, permuting incomes among agents will
not change the value of inequality. Formally: For a given society N, let y, y0 stand
for two income distributions such that y0 obtains as a permutation of y (usually
written as y0 ¼ π ðyÞ). Then, I(y) ¼ I(y0 ).
2
In a richer context, anonymity and symmetry are not exactly the same. Yet here both properties
are equivalent.
2.2 Properties of Inequality Indices 21
When this property holds, we can assume, without loss of generality, that any
income vector is ordered in a non-decreasing way, that is,
y1 y2 y3 yn1 yn . There are many instances in which this represen-
tation is useful.
Remark 2.2 There are cases in which we may find this property too strong or
inadequate. One is that involving agents with different personal characteristics that
one wants taking into account (e.g. differences in the size and composition of the
families). Another refers to a dynamic context in which we might be willing to keep
track of the past distributions. See Atkinson and Bourguignon (1982) for a more
general notion of symmetry.
Property 3: Population Replication
This property is key to comparing income distribution vectors corresponding to
societies with different number of agents. It can be motivated as follows. Suppose
we have two societies, A and B, with the same number n of agents and identical
income distributions. We would have: I(yA) ¼ I(yB). Now imagine that those two
societies merge together to form a new society C, with 2n agents and twice the
number of agents for each income value. It is reasonable to assume that the
inequality in this new society is the same as that of the original ones.
This idea was introduced by Dalton (1920) under the name of “principle of
population replication” (also known as replication invariance). It can be formalised
as follows.
Let N be a society with an income distribution vector y, and let Nk denote a new
society obtained by replicating k times society N with the kcorresponding
income
distribution, y . That is, for y ¼ ðy, y, . . . , yÞ we have: I y ¼ I ðyÞ.
k k
|fflfflfflfflfflffl{zfflfflfflfflfflffl}
k
What this property says is that the inequality index only depends on the relative
frequencies of the income distributions. This permits one comparing income dis-
tributions with different numbers of agents, as illustrated next.
Example Let A, B be two societies with two and three agents, respectively. The
corresponding income distributions are given by: yA ¼ ðyA1 ; yA2 Þ,
yB ¼ ðyB1 ; yB2 ; yB3 Þ. Let us replicate three times society A and twice society B.
We would have:
y3A ¼ ðyA1 ; yA2 ; yA1 ; yA2 ; yA1 ; yA2 Þ, y2B ¼ ðyB1 ; yB2 ; yB3 ; yB1 ; yB2 ; yB3 Þ
Since I yA ¼ I y3A , I yB ¼ I y2B , by the population replication principle, we
conclude that IðyA Þ IðyB Þ ⟺ Iðy3A Þ Iðy2B Þ. That is, we are able to compare
societies A and B by means of an implicit suitable replication.
Property 4: The Principle of Transfers
This is a property introduced by Dalton (1920), following an idea of Pigou (1912).
Consider a society N with n agents and an income distribution y ¼ ðy1 ; y2 ; . . . ; yn Þ.
We call a Dalton transfer from a “rich” agent to a “poor” one to an income transfer
22 2 Inequality Indices
between them such that their positions in the income ranking do not change. That is,
for two agents i, j with incomes yj < yi , making a Dalton’s transfer of size δ > 0
means reducing the income of agent i and increasing the income of agent j by
precisely that amount, under one proviso: yj þ δ yi δ (which implies 0 < δ
y y
i 2 j ). Note that a Dalton’s transfer does not affect the mean income.
The principle of transfers says that a Dalton transfer should reduce inequality.
Formally, for a given society N, let yA, yB be two income
distributions
such that yB
is obtained from yA by a Dalton’s transfer. Then, I yB < I yA .
This principle is generally admitted as one of the key normative properties of
inequality indices. Yet there are three aspects of this principle that deserve some
consideration.
(a) First, it may be regarded as a very weak principle because it establishes no
relationship between the size of the transfer and the change in the index. One
may be willing to get some kind of relationship between the amount of money
that is transferred and the reduction of inequality.
(b) Second, it may be considered as not very specific, as it does not establish
whether the change in inequality should depend only on the two agents
involved or may depend on the whole distribution (in particular regarding the
relative ranking of the individuals involved). See Cowell and Kuga (1981) and
Cowell (1995).
(c) Third, and partly related to the second consideration, this principle ignores the
local effect that a transfer may have. Think of the case of a large transfer from
an extremely rich person to an extremely poor one. The relative position of the
rich person may have only changed marginally, whereas that of the poor one
may have been dramatically altered, now becoming very rich relative to their
peers in the pre-transfer income range. Should this local increase of inequality
be necessarily smaller than the global reduction in inequality for the whole
society? The Dalton principle establishes that this is always the case. Yet, one
may argue that this is not so evident.
Property 5: Continuity/Differentiability
Continuity is a standard analytical requirement that conveys a simple and intuitive
idea: small changes in the variable induce small changes in the function. It is then
generally assumed that function I(y) is continuous. Yet in some cases, the stronger
differentiability assumption is introduced. Differentiability requires that small
changes in the variable induce small changes not only to the value of the function
but also to its rate of change. Formally, Function I(y) is differentiable in y 2 ℝþþ
n
.
Clearly differentiability, which amounts to requiring continuous derivatives,
implies continuity but the converse is not true.
The continuity and differentiability of the inequality index are, therefore,
requirements on the behaviour of the index with respect to small changes in
2.2 Properties of Inequality Indices 23
3
One can also think of intermediate situations. See Pfingsten (1986), Bossert and Pfingsten (1990),
Del Rı́o and Ruiz-Castillo (2000), Cowell (2003).
4
Note that one cannot explain total inequality as a weighted sum to within groups’ inequality. To
see this, simply think of the case of a society made of two different groups (e.g. men and women),
24 2 Inequality Indices
X
I ð yÞ ¼ ωg I ðyg Þ þ I ðÞ
g
|fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl} |{z}
between groups
within groups
Here, inequality within groups appears as the weighted sum of the inequality that
exists in each of those groups, considered as an independent society, where the
weights reflect their relative importance.5 The between groups’ inequality can be
measured applying the same inequality index to a fictitious society made of the
union of the populations of all subgroups, after cancelling all income differences
within the groups (i.e. substituting each original income for a member of the group
by the corresponding mean income).
Let y ¼ ðy1 ; y2 ; . . . ; yn Þ stand for the income distribution of a society N that is
made of G different population subgroups, g ¼ 1, 2, 3, . . . , G, defined by some
criterion different
from income. We denote by ng, the number of agents in group g,
g g
and by yg ¼ y1 ; y2 ; . . . ; yngg , its corresponding income distribution. Now the
This section aims at illustrating that inequality indices are something more than
dispersion measures. We shall see that even though the standard dispersion mea-
sures, such as the variance or the coefficient of variation, satisfy practically all of
such that all agents within a group have the same income but the income between agents of
different groups differ. Total inequality will be positive whereas within groups’ inequality is zero.
5
Those weights typically depend either on the population shares or on the income shares, or
on both.
2.3 Inequality and Dispersion 25
the properties described in the former section, they are not suitable indices of
inequality. In other words, measuring inequality is not the same as measuring
dispersion, even though both aspects are closely related. Let us consider a very
simple example that will serve the purpose of illustrating that difference.
Suppose we want to compare the inequality of two different income distributions
in a society made of five agents (we can think that those values correspond to the
mean values of five different regions of a given country, in two different years). The
income is expressed in thousands of constant euros, and the income distributions are
the following:
y1 ¼ ð10, 20, 30, 40, 100Þ y2 ¼ ð24, 24, 24, 24, 104Þ
Note that both distributions involve dividing a total of 200 income units between
five agents, i.e. the mean income is the same in both distributions.
The question is, which income distribution is more unequal?
Intuition suggests that the first income distribution is more unequal than the
second one. It can be argued that the rich person in the second distribution has
practically the same income than that of the rich agent in the first one, whereas all
other agents have equal incomes in distribution 2 and rather different in distribution
1. One can also defend this idea by observing that the second distribution is more
favourable for the less well off. Indeed, computing the % of the total income that is
enjoyed by the corresponding % of the population, arranged from less to more
income, one obtains the following:
Table 2.1 shows that the second distribution is more beneficial for the 60 % of
the population with less income.
So there are good reasons to think that y2 is more egalitarian than y1. Yet, this has
to be confirmed with some reasonable indicator that should also tell us how much
inequality is in each of those distributions. Let us consider now how the most
standard measures of dispersion evaluate those distributions and how do they fare
regarding the properties described in Sect. 2.2.
The variance is probably the best well-known and most widely used measure of
dispersion in all sciences. It is defined as the average of the squared differences
between the individual values and the mean and corresponds to the second moment
of the distribution of a random variable. The formula is given by:
26 2 Inequality Indices
1X n
σ2 ¼ ðy μÞ2
n i¼1 i
ð2:1Þ
1X n
¼ y2i μ2
n i¼1
In this way, the variance appears as one half of the quadratic mean of the
differences between all incomes, without explicit reference to a central value.
Those two equivalent expressions permit one checking easily some of the properties
that this measure satisfies. It is a differentiable function that satisfies normalisation
(the variance ranges between 0, when all incomes are equal, and (n 1)μ2, when all
income is held by a single agent), symmetry and population replication.
From Eq. (2.2), we can deduce that it also satisfies the Dalton principle of trans-
fers. Taking a small Dalton’s transfer δ between two agents, i and j, with yi > yj , the
change over the variance can be obtained by differentiation and is given by:
2
δ yj yi < 0:
n
Note that the impact of a Dalton’s transferon inequality only depends on the
income difference between the agents, yi yj , but not on the income levels yi, yj
at which this difference occurs. This can be criticised as a lack of sensitivity to what
happens in the lower part of the distribution.
The variance clearly fails to satisfy scale independence. Its value, therefore,
changes when the units in which we measure the income vary (e.g. from euros to
dollars or from thousands of euros to tens of thousands). This is unpleasant, but it is
not the main inconvenience (indeed the relative variance of two distributions is not
altered by this change of units). A more worrying aspect is, as pointed out by Sen
(1973), that since it depends on the mean income, we can find a distribution that
2.3 Inequality and Dispersion 27
exhibits a larger relative dispersion than another one and still shows a smaller
variance, provided the mean differences are large enough.6
Remark 2.5 Scale independence can be easily obtained by simply dividing the
variance by the square or the mean. This yields the square of the coefficient of
variation (we return to this measure in Chap. 3):
σ2
CV 2 ¼
μ2
The variance satisfies additive decomposability, and this is one of its most
appealing properties. To see this let N stand for a society with n agents and let y
¼ ðy1 ; y2 ; . . . ; yn Þ describe the income distribution. Suppose that this society is
made of G different population subgroups, g ¼ 1, 2,
. . ., G. Let ng denote the
g g
number of agents in group g, let yg ¼ y1 ; y2 ; . . . ; yngg stand for the corresponding
income distribution of this group, where ygi is the income of agent i in group g and
let ~
μ ¼ ðμ1 ; μ2 ; . . . ; μG Þ stand for the vector of mean incomes of the different
groups. We can write the overall mean income, μ, as the weighted average of the
groups’ mean incomes, with weights given by their population shares. That is,
1X 1X G X
1X X
n gn G G
g ng
μ¼ yi ¼ yi ¼ ng μg ¼ μ
n i¼1 n g¼1 i¼1 n g¼1 g¼1
n g
We can now establish the relationship of the variance of the population and that
of the different groups as follows:
ng 2
1X G X g 2 1X G X
ng
g
σ2 ¼ yi μ ¼ yi μg þ μg μ
n g¼1 i¼1 n g¼1 i¼1
XG X ng 2
1 g
2
g
¼ yi μg þ 2 yi μg μg μ þ μg μ
n g¼1 i¼1
X ng X 1 g 2
G ng
¼ yi μ g
g¼1
n i¼1 ng
|fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl}
σ 2g
G Xng X
2X G
ng 2
þ μg μ yig μg þ μg μ
n g¼1 i¼1 g¼1
n
|fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl}
¼0
6
The variance is an absolute inequality index because adding the same amount of income to each
agent does not change its value.
28 2 Inequality Indices
Therefore,
X
G
ng X
G
ng 2
σ2 ¼ σ 2g þ μg μ ð2:3Þ
g¼1
n g¼1
n
The first term of Eq. (2.3) gives us a measure of dispersion within the subgroups,
as a weighted average of their variances, with weights equal to the corresponding
population shares. The second term is a measure of variability between the groups,
as a weighted average of the squared mean deviations, again with weights given by
the population shares. We can write:
(i)
X
G
ng
σ 2W ¼ σ 2g ;
g¼1
n
(ii)
XG
ng 2
σ 2B ¼ μg μ
g¼1
n
so that we have the overall variance decomposed as the sum of two terms,
σ 2 ¼ σ 2W þ σ 2B , corresponding to the within and the between components. The ratios
σ 2W σ 2B
σ2 , σ2 provide the “percentage” of total variability explained by each of those
components.
The variance satisfies, therefore, all the properties presented in Sect. 2.1 other
than scale independence. Let us see how this measure evaluates the income
distributions in the example presented above, y1 ¼ (10, 20, 30, 40, 100),
y2 ¼ (24, 24, 24, 24, 104). Making the corresponding computations we obtain:
σ 2 y1 ¼ 1000, σ 2 y2 ¼ 1024
That’s a surprise! According to the variance there is more inequality in the second
distribution than in the first one, against our intuition. Why is that? In a nutshell: the
variance is a good measure of dispersion but a deficient measure of inequality. The
reason is that the variance is an average of the square differences with respect to the
mean and, therefore, income differences above or below the mean are treated alike.
Yet the very notion of inequality has some implicit value judgements regarding what
happens below the mean or below the median. It implies in a way a greater concern
for the less well off, which the variance cannot take properly into account (in spite of
the Dalton principle of transfers being satisfied!).
This example is a good illustration of two relevant aspects of the evaluation of
income distributions:
(a) Inequality analysis goes beyond the measurement of income dispersion, as it
involves value judgements regarding equity.
References 29
(b) When discussing different inequality measures, one has to observe how those
measures treat transfers at different parts of the distribution (or, put differently,
how sensitive are those measures to changes in the incomes of the poor).
δ
yj yi < 0
nμσ
This expression shows that CV inherits the property that the impact of a transfer
depends on the income differences between the agents involved but not on their
income levels. Indeed calculating the coefficient of variation for the example
presented
at the beginning
of this section yields the following values:
CV y1 ¼ 0:05, CV y2 ¼ 0:16. So here again we find the counterintuitive result
already observed before, as should be expected, but also a much larger relative
difference.
References
Blackorby, C., & Donaldson, D. (1980). A theoretical treatment of indices of absolute inequality.
International Economic Review, 21, 107–136.
Bossert, W., & Pfingsten, A. (1990). Intermediate inequality: Concepts, indices and welfare
implications. Mathematical Social Science, 19, 117–134.
Cowell, F. A. (1995). Measuring inequality (Second edition), LSE handbooks in economics series.
Hemel Hempstead: Harvester Wheatsheaf.
Cowell, F. A. (2003). Theil, inequality and the structure of income distribution (Working Paper
DARP No. 67). Sticerd, LSE.
Cowell, F. A., & Kuga, K. (1981). Inequality measurement. An axiomatic approach. European
Economic Review, 15(3), 287–305.
Dalton, H. (1920). The measurement of inequality of income. The Economic Journal, 30, 348–361.
Del Rı́o, C., & Ruiz-Castillo, J. (2000). Intermediate inequality and welfare. Social Choice and
Welfare, 17(2), 223–239.
Goerlich, F., & Villar, A. (2009). Desigualdad y Bienestar: De la Teorı́a a la Práctica Fundación
BBVA.
Kendall, M., & Stuart, A. (1977). The advanced theory of statistics, Volume 1: Distribution theory
(4ª edición). London: Charles Griffin & Company Limited.
Kolm, S.-C. (1976a). Unequal inequalities I. Journal of Economic Theory, 12, 416–442.
Kolm, S.-C. (1976b). Unequal inequalities II. Journal of Economic Theory, 13, 82–111.
Pfingsten, A. (1986). Distributionally-neutral tax changes for different inequality concepts. Jour-
nal of Public Economics, 30, 385–393.
Pigou, A. C. (1912). The economics of welfare. London: Transaction Publishers (Editado por
MacMillan, New York, en 1952, 4ª edición).
Sen, A. K. (1973). On economic inequality. Oxford: Clarendon.
Chapter 3
Positive Inequality Indices
One of the most basic approaches to evaluation, the relative fairness of income
distributions derives from the comparison of the percentages of total income and the
percentages of the people who enjoyed that income, as we did in Table 2.1. The idea
is due to Lorenz (1905) and has become a key reference for the analysis of
inequality because it incorporates an intuitive graphic instrument, the Lorenz
curve, which can also be given an analytical content.
Let y ¼ ðy1 , y2 , . . . yn Þ be an ordered income distribution, with y1 yn .
The Lorenz curve is built as follows. We order the cumulative shares of population,
from the poorest agent to the richest one, on the horizontal axis and the
corresponding cumulative shares of income on the vertical axis. This produces a
box of dimension 1 containing a curve that describes the income distribution. The
diagonal of this box represents the line of perfect equality (each given share of
people gets exactly the same share of total income), so that the difference between
this line and the actual distribution provides a measure of inequality. Figure 3.1
illustrates this construct.
More formally, for a given ordered income vector, the Lorenz curve is defined as
the set of points [pi, L( pi)], where pi ¼ i=n is the share of people with income equal
to or smaller than yi, and L( pi) is defined as follows :
Xi
y
j¼1 j
Lð p i Þ ¼
nμ
with L(0) ¼ 0, L(1) ¼ 1 (assuming, in this discrete version, that intermediate points
are obtained by linear interpolation). We shall also write Li as a shorthand version
of L( pi).
% cummulative income
Lorenz
curve
Note that the slope of the Lorenz curve between two consecutive points,
i =nμ
½pi1 , Lðpi1 Þ, [pi, L( pi)], corresponds to the relative income, y1=n ¼ yμi . This implies
that the slope increases as we move to the right (the richer part of the population).
The way in which this curve is built implies that a Dalton’s transfer moves the
curve inwards, approaching it to the line of perfect equality, thus unambiguously
reducing inequality (see Fig. 3.2). This allows comparing some distributions from
an ordinal point of view (assuming equal means): if the Lorenz curve of a distri-
bution is above the Lorenz curve of another, then the first one is more egalitarian
than the second one. This is so because in that case the first one can be obtained by
successive series of Dalton’s transfers from the second one. This is usually referred
to as Lorenz dominance.
This type of comparison presents two difficulties.
(i) It cannot be applied when the Lorenz curves cross each other. This is so
because the Lorenz dominance is an incomplete criterion. This is illustrated
in Fig. 3.3, where we present three different curves corresponding to three
distributions, yA, yB and yC. We can say that the first income distribution is
more egalitarian than the other two, yet we cannot say whether yA is more
egalitarian with respect to yB than with respect to yC or whether yB is more or
less egalitarian than yC.
(ii) It does not provide a measure on how much unequal is one distribution relative
to the other (Lorenz dominance is an ordinal notion).
Those problems point out the need of defining some cardinal measure from this
intuitive construct. There have been many proposals on this respect, the most
successful of which is the Gini index, which we analyse in the next section. Be as
3.1 The Lorenz Curve 33
L(p)
% cumulative income
between quantiles q and r
0 r q 1 F
% cumulative population
yA
yB
yC
0 1 F
% cumulative population
it may, note that the information contained in the Lorenz curve is very rich. For
instance, if we observe L(0.5) ¼ 0.2 this means that the 50 % of the poorer popu-
lation obtains 20 % of the total income. In general, the smaller the values of L( p) the
more unequal the distribution.
Remark 3.1 It can be shown that when a distribution Lorenz dominates another,
all standard inequality indices will rank them in the same way. This is not the case
34 3 Positive Inequality Indices
when the Lorenz curves cross each other.1 See Dasgupta, Sen, and Starrett (1973)
and Rothschild and Stiglitz (1973).
The Gini index, also known as the Gini coefficient, was proposed by Corrado Gini at
the beginning of the XXth Century (see, in particular, Gini, 1921). It is one of the
most popular and widely used inequality indices, probably because of its intuitive
appeal when interpreted geometrically using the Lorenz curve.
The index can be obtained in a number of ways and admits several alternative
and useful expressions, each of which illuminating some particular aspect. The
most intuitive formulation is the geometric one obtained directly from the Lorenz
curve. The Gini index corresponds to the ratio of the area between the Lorenz curve
and the 45 line and the whole area below this line. That is, looking at Fig. 3.4, we
can obtain the Gini index as:
A
G¼ ð3:1aÞ
AþB
G ¼ 2A ð3:1bÞ
From this simple definition, bearing in mind how the Lorenz curve is built, one
can immediately deduce that the Gini coefficient satisfies normalisation, symmetry,
population replication, the principle of transfers and continuity.
Let us now describe how to calculate the Gini index in the case of a discrete
income distribution y ¼ ðy1 , y2 , . . . yn Þ, as seen in Fig. 3.5. In view of equation
(3.1b), we can do that by simply computing the area A between the Lorenz curve
and the 45 line.
We start by measuring first the area between the Lorenz curve and the 45 line
for two adjacent points, [(i - 1), i ]. We do that in two steps. First, calculating the
whole shaded area below the 45 line for those adjacent points. Second, deducing
the area below the Lorenz curve at those points.
1
This is what happens with the Lorenz curves of the example in the Introduction, as can be easily
deduced from Table 2.1. We shall see that the Gini index and other inequality measures rank those
two distributions differently as how the variance or the coefficient of variation does.
3.2 The Gini Index 35
L(p)
% cumulartive income
A
0 1 F
% cumulative population
pi
% cumulative income
Lorenz curve
pi - 1
0 pi - 1 pi 1 F
% cumulative population
From Fig. 3.5 and elementary geometry, we can compute the whole shaded area
as follows2:
1 1
ðpi pi1 Þpi1 þ ðpi pi1 Þ2 ¼ ðpi pi1 Þðpi þ pi1 Þ
2 2
2
Note that is corresponds to the area of a rectangle minus the upper left triangle, which has two
equal sides.
36 3 Positive Inequality Indices
Moreover, the area below the Lorenz curve in two adjacent points can be
calculated as follows:
1 1
ðpi pi1 ÞLi1 þ ðpi pi1 ÞðLi Li1 Þ ¼ ðpi pi1 ÞðLi þ Li1 Þ
2 2
Deducing the part of the area below the Lorenz curve from the former expression
we get:
1 1
ðpi pi1 Þðpi þ pi1 Þ ðpi pi1 ÞðLi þ Li1 Þ
2 2
1
¼ ðpi pi1 Þ½ðpi þ pi1 Þ ðLi þ Li1 Þ
2
Now, adding up all those adjacent points, we can calculate the area A between
the Lorenz curve and the 45 line:
1X n
A¼ ðp pi1 Þ½ðpi þ pi1 Þ ðLi þ Li1 Þ
2 i¼1 i
X
n
G¼ ðpi pi1 Þ½ðpi þ pi1 Þ ðLi þ Li1 Þ
i¼1
1 Xn
¼ ½ðp þ pi1 Þ ðLi þ Li1 Þ
n i¼1 i
1X n
¼ ½ðp Li Þ þ ðpi1 Li1 Þ
n i¼1 i
2X n
G¼ ð p Li Þ ð3:2Þ
n i¼1 i
This expression tells us that the Gini index is twice the average of the difference
between population and income shares in the income distribution
3.2 The Gini Index 37
n þ 1 2X n
y
G¼ ðn þ 1 iÞ i ð3:3Þ
n n i¼1 nμ
This formula establishes that the Gini index is a linear function of the weighted
sum of income shares, with weights corresponding to the income ranking. The
poorest agent enters the formula with a weight equal to n, the second poorest agent
with a weight equal to n 1 and so on and so forth, until the richest one who has
weight equal to 1.
This formula makes it easier computing the impact of a Dalton’s transfer of size
δ, which is given by: δμn2 2 ½ j i < 0, assuming that j < i. We see, therefore, that the
impact of the transfer depends on the ranking of the agents involved but not on their
income differences.
We can also deduce that the Gini index satisfies the following properties:
normalisation, symmetry, population replication, principle of transfers, continuity
and scale independence. It does not satisfy additive decomposability.
If we calculate the Gini index for the two distributions of the example of Sect.
2.3, we find: G y1 ¼ 0:4, G y2 ¼ 0:32. The Gini index says that the first
distribution is more unequal than the second one, according to our intuition and
contrary to what the variance tells. The reason is that we are now giving more
weight in the evaluation to those agents with lower incomes. Also note that whereas
the variance is a function of the income differences with respect to the mean, the
Gini index takes the shares in total income, which ensures scale independence.
We have already mentioned, when presenting the formula of the Gini index in
equation (3.2), that all the differences between income and population shares enter
the index with the same weight. One may build a more flexible index by simply
allowing for different weights to those differences depending on the ranking in the
n
X
2i ðn þ 1Þ y
3
This expression can also be rewritten as: G ¼ 1n i
i¼1
n μ
38 3 Positive Inequality Indices
distribution. There are several ways of doing that.4 Yet the most widely used
weighting system is that proposed by Yitzhaki (1983),5 who suggested the follow-
ing uni-parametric family of weights:
v2
nþ1i
κ i ð v Þ ¼ v ð v 1Þ ð3:4Þ
n
with v > 1. It is clear that as v increases those agents with lower incomes get
progressively more weight.
Using this weighting system, Yitzhaki (1983) proposes the following General-
ised Gini index (also called “S-Gini”):
1X n
Gv ¼ κi ðvÞðpi Li Þ
n i¼1
" # ð3:5Þ
1X n
n þ 1 i v2
¼ v ð v 1Þ ð p i Li Þ
n i¼1 n
From this expression, we deduce that the impact of a Dalton’s transfer of size δ is
given by:
4
See Mehran (1976), Donaldson and Weymark (1980, 1983), Weymark (1981), Yaari (1988),
Chakravarty (1988), Ben-Porath and Gilboa (1994), Aaberge (2000) or Imedio Olmedo and
Bárcena Martı́n (2007).
5
The idea appears first in Kakwani’s (1980) extension of Sen’s (1976) poverty indices.
3.3 Theil’s Inequality Indices 39
v h v1 v1
i
δ ð n þ 1 i Þ ð n þ 1 jÞ <0
μnv
assuming that j < i and v > 1. Note that, as it happened with the original Gini index,
this number depends on the ranking of the individuals involved (now weighted
differently) but not on the income differences that exist between those individuals.
In a discrete scenario, with n possible events with probabilities w1, w2,. . ., wn, we
define the entropy of this situation, Q(w), as its expected informational content:
X
n X
n
1
QðwÞ ¼ wi ϕðwi Þ ¼ wi log ð3:7Þ
i¼1 i¼1
wi
40 3 Positive Inequality Indices
Note that, in this case, the maximum entropy is given by the uniform distribu-
tion, wi ¼ 1=n, 8i, so that the closer the probabilities to 1/n, the larger the entropy.6
As a consequence, the maximum value of the entropy is given by:
X n
1 1 1 1 1
Q ; ;...; ¼ logn ¼ logn ¼ ϕ
n n n i¼1
n n
Theil (1967) proposes applying this conceptual scheme to the analysis of income
distribution by introducing two variations. First, substituting the notion of proba-
bility by that of income shares, si ¼ yi =nμ, i ¼ 1, 2,. . ., n (which are all positive and
also add up to 1). Second, taking the difference betwee log n, the maximum entropy
that coincides with the perfectly egalitarian distribution, and the value Q(s)
corresponding to the actual income distribution [see Foster (1983) for an alternative
characterisation].
Formally, Theil’s measure is defined as follows:
1 1 1
T ¼ Q ; ;...; QðsÞ ¼ logn QðsÞ
n n n
where s is the vector of income shares. That is, the difference between the value of
the function Q for perfect equality and that of the observed distribution.
Xn X
n Xn
As QðsÞ ¼ si logs1i ¼ si logsi and logn ¼ s logn, we can write:
i¼1 i
i¼1 i¼1
X
n X
n
T¼ si logn þ si logsi
X
n
i¼1 i¼1
ð3:8Þ
¼ si lognsi
i¼1
Theil interprets this index as the expected information of a message that trans-
forms population shares into income shares. So T can be regarded as a distance
function between population shares and income shares that uses income shares as a
1
weighting system. Clearly T ¼ 0 when si ¼ , 8i, whereas the maximum inequality,
n
which corresponds to sn ! 1 and si ! 08i 6¼ n, yields T ! logn, since slogs ! 0 as
s ! 0. This suggests normalising this indicator so that its range becomes the
interval [0, 1], by letting:
Xn
6
It is immediate to check that solving the programme max w logð1=wi Þ subject to
i¼1 i
Xn
w
i¼1 i
¼ 1 yields logð 1=w i Þ ¼ log 1=w j , for all i, j, so that all w i must be equal and thus
correspond to 1/n.
3.3 Theil’s Inequality Indices 41
e¼ 1 T
T ð3:9Þ
logn
1X n
yi yi
T¼ log ð3:10Þ
n i¼1 μ μ
This is probably the most used form of the Theil’s index. Note that yi/μ is the slope
of the Lorenz curve at the quantile corresponding to yi. This index can, therefore, be
also obtained directly from the Lorenz curve.7
It is easy to check that T satisfies the properties of normalisation, symmetry,
population replication, differentiability and scale independence. Regarding the
principle of transfers, one can easily deduce that at Dalton’s transfer of size δ, the
impact on the index is given by:
1 yj
δ log <0
nμ yi
XG
ng μg X
ng
yi
g
y X
g G
ng μ g μ g
¼ log i þ log
g¼1
nμ i¼1 ng μg μg g¼1 nμ μ
|fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl}
Tg
7
See Kanbur (1984) for a different interpretation.
42 3 Positive Inequality Indices
The first term of the last equality is the weighted sum of the Theil indices within
the groups. The between groups component Xnweights inequality by the income
g g
ng μg y
¼ Xi¼1
i
shares o the groups in total income, nμ n :
y
i¼1 i
We can write:
XG
ng μ g g X G
ng μg μg
T¼ T þ log ð3:11Þ
g¼1
nμ g¼1
nμ μ
Or, put differently, we write the Theil index as the sum of within and between
inequality as T ¼ T W þ T B , with:
(i)
XG
ng μ g g
TW ¼ T ;
g¼1
nμ
(ii)
XG
ng μ g μ g
TB ¼ log
g¼1
nμ μ
The values of the Theil index in the example presented in Sect. 2.3, given by
y1 ¼ (10, 20, 30, 40, 100), y2 ¼ (24, 24, 24, 24, 104), are:
T y1 ¼ 0:276 T y2 ¼ 0:252
so that here again the second distribution appears as more egalitarian than the
first one.
Theil (1967) proposed a second inequality index, that we denote by T*, exchanging
the roles of income and population shares. That is,
X
n
1 1
T* ¼ log ð3:12Þ
i¼1
n nsi
This index also computes the differences between population and income shares,
but using now population shares as weights.
Note that nsi ¼ yi =μ so that we can rewrite equation (3.12) as follows:
1X n
μ
T* ¼ log ð3:13Þ
n i¼1 yi
As μ/yi is the inverse of the slope of the Lorenz curve at the quantile corresponding
to yi, T* can also be obtained directly from the Lorenz curve. The expression (3.13)
justifies the alternative name under which it is known, the mean logarithmic
deviation.
When the income distribution is perfectly egalitarian, income and population
shares are identical and then T* ¼ 0. As the income gets concentrated into a single
agent, sn ! 1 and si ! 0 8i 6¼ n, T* is not bounded above.8
The second index of Theil, T*, corresponds to the log of the ratio between the
arithmetic mean, μ, and the geometric mean, μ e, that is,
1X n
T* ¼ ðlogμ logyi Þ
n i¼1
1X n
¼ logμ logyi ð3:14Þ
n i¼1
¼ logμ loge μ
μ
¼ log
μ
e
This index also satisfies all the properties presented in Sect. 2.1. Regarding the
principle of transfers, the effect of a Dalton’s transfer δ on T* is given by:
!
1 1 1
δ <0
n yi yj
8
Indeed, this index is not defined when si ¼ 0 for some i.
44 3 Positive Inequality Indices
1X G Xn g
μ
T* ¼ log
n g¼1 i¼1 yig
G X
X ng
1 μg μ
¼ log g :
n yi μ g
g¼1 i¼1
!
G X
X ng
1 μg μ
¼ log g þ log
n y μg
" n i
g¼1 i¼1
#
Xng 1 X μg
G g
μ
¼ log þ log
g¼1
n ng i¼1 yig μg
Xng 1 X μg Xng
G n g G
μ
¼ : log g þ log
g¼1
n ng i¼1 yi g¼1 n μg
|fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl}
T*g
X
G
ng X
G
ng μ
T* ¼ T *g þ log ð3:15Þ
g¼1
n g¼1
n μg
This equation expresses total inequality as the sum of between and within
inequality, T * ¼ T *W þ T *B , with:
X
G
ng
T *W ¼ T *g ;
g¼1
n
X
G
ng μ
T *B ¼ log
g¼1
n μg
Here the weights of the partial components in each of those two terms corre-
spond to the population shares, ng/n.
The two indices of Theil presented above can be regarded as particular cases of a set
of uni-parametric functions known as the generalised entropy family. We shall
present first this family as the only one that satisfies all the properties in Sect. 2.1
3.3 Theil’s Inequality Indices 45
and then review it from the information theory viewpoint, by relaxing the functional
form of the information function ϕ presented above.9
Shorrocks (1980, Th. 5) shows that there is a unique family of functions that
satisfies all the requirements for a candidate to inequality index in Sect. 2.1.10 It can
be stated as follows:
Theorem 3.1 A function I : Ω ! ℝ satisfies the properties of normalisation,
symmetry, population replication, principle of transfers, differentiability, scale
independence and additive decomposability, if and only if it is of the form:
" #
1 1 X n
yi θ
I θ ðn; yÞ ¼ 1 ð3:16Þ
n θðθ 1Þ i¼1 μ
(or it is a proportional transformation of this function, αIθ(n, y) with α > 0), where θ
is a non-negative parameter that ranges over the whole real line.
For θ ¼ 1, L´Hôpital rule yields the first index of Theil, T1 ¼ T. For θ ¼ 0,
L’Hôpital rule yields n times the second index of Theil so that, by taking α ¼ 1=
n we get: T0 ¼ T*.
For θ ¼ 2, the resulting index corresponds to one half of the square of the
coefficient of variation. We can see that as follows:
" #
n 2
1 X n
yi 2 1 X yi μ 2 1 X n 1 σ2
T 2 ð yÞ ¼ 1 ¼ ¼ y2
μ2
¼ 2
2n i¼1 μ 2n i¼1 μ2 2nμ2 i¼1 i 2μ
The decomposition of CV2 can thus be obtained by simply applying the formula
of the decomposition of the variance:
9
Bourguignon (1979), Shorrocks (1980, 1984), Cowell (1980) and Cowell and Kuga (1981a) study
this family as the result of additive decomposability. An alternative approach is that of strength-
ening the principle of transfers Cowell and Kuga (1981b). The study of the family extending the
information function that we follow here appears in Cowell (1977, 1995) and Cowell and Kuga
(1981b).
10
See Cowell (2003) for an additional extension of these indices.
46 3 Positive Inequality Indices
X
G
ng μ2g σ 2g 1X G
ng 2
CV 2 ¼ þ μ μ
g¼1
n μ2 μ2g μ g¼1 n g
2
|{z}
CV 2g
Appendix presents the decomposition of this family of indices. We show there that
only the original indices of Theil have the property that the weights of the within
components add up to 1.
The parameter θ captures the sensitivity of the index with respect to transfers
depending on the part of the distribution in which those transfers take place. As θ
decreases, Iθ becomes more sensitive to the transfers in the lower part of the
distribution in the sense that the same amount of the transfer reduces more the
inequality the lower the income of the recipient. In the limit, as θ ! 1, the index
is only sensitive to transfers to the less well off. As θ increases, Iθ becomes more
regressive. Indeed, for θ > 2 the index is reduced only if we equalise the incomes of
the richer, in spite of the fact that the principle of transfers is always satisfied (for a
critique of this aspect, see Kolm, 1976a, b; Love & Wolfson, 1976).
To check this behaviour simply notice that the effect of a Dalton’s transfer of
size δ is given by:
8
>
>
1
yθ1 yθ1
<δ θ θ 6¼ 1
nμ ðθ 1Þ j i
ð3:17Þ
>
> 1 yj
: δ log θ¼1
nμ yi
compelling in the case of income distributions. Keeping the properties (a) and
(b) yields the following family of functions11:
8
< 1 wβ
ϕðβ; wÞ ¼ β 6¼ 0 ð3:18Þ
: β
logðwÞ β ¼ 0
δ
where ds ¼ .
nμ
This expression shows that the change in T depends on the distance between the
two agents involved in terms of their relative incomes, evaluated by the information
function ϕ(s). As a result, the larger the income differences between the agents
involved, the larger the inequality reduction for a given size of a Dalton’s transfer.
It can be shown that using the information function (3.18) and letting θ ¼ β þ 1
yields the following generalisation of the indices of Theil:
" #
n1θ 1 X n
yi θ
Tθ ¼ 1 8θ 6¼ 0, 1 ð3:20Þ
θðθ 1Þ n i¼1 μ
(with the corresponding limit expressions for θ ¼ 0 and θ ¼ 1), which is a linear
transformation of (3.16).
In summary, either using the axiomatic approach in Theorem 3.1 or applying the
generalisation of the information function in (3.18), we generate a uni-parametric
family of inequality functions that extends Theil’s indices, which is given by:
11
This family of functions corresponds to the (negative) class of non-linear transformations
introduced by Box and Cox (1964), widely used in statistics and econometrics.
12
The same argument here applies to the second index of Theil, T *.
48 3 Positive Inequality Indices
8 " #
>
> 1 1 X
n
y θ
>
> i
1 θ 6¼ 0, 1
>
> nθðθ 1Þ i¼1 μ
>
>
>
< X
1 n μ
I θ ðn; yÞ ¼ log θ¼0 ð3:21Þ
>
>
> n i¼1 yi
>
>
> 1X n
>
> yi yi
>
:n log θ¼1
i¼1
μ μ
Under perfect equality Iθ ¼ 0, 8θ. Yet the maximum value of the index changes
θ1
1
with θ. For θ > 0, the upper bound is given by nθðθ1 Þ , if θ 6¼ 1 and log n for θ ¼ 1. For
θ 0, there is no upper bound.
13
Clearly, any transfer that does not affect the shares of the relationship between the selected
quantiles does not affect the index. Transfers that affect the selected quantiles move according to
the principle of transfers. This is what we mean by “weak version”.
3.4 Quantile Measures: The Palma Ratio 49
2011). From these data it follows that, roughly speaking, much of the distributional
conflict concentrates on the battle of the rich and the poor for the half of the cake
that is not enjoyed by the middle classes. Indeed, empirical results also show that
the distribution of that half between the rich and the poor is very different among
countries.
The Palma Ratio is defined as the ratio between the aggregate income of the
richest 10 % of the population and that of the poorest 40 %.
This proposal has received a lot of attention because it provides a simple quantile
measure that is not arbitrary as the cutting points derive from relevant empirical
regularities. Those regularities were later confirmed in the work of Cobham and
Sumner (2013), who showed the robustness of Palma’s main results over time: the
remarkable stability of the middle class capture across countries coupled with much
greater variation in the 10/40 ratio. Moreover, these authors showed that there is a
very high correlation between the Palma ratio and the Gini index. They suggest that
the Palma might be a better inequality measure than the Gini index because it is
intuitively easier to understand for policy makers and citizens alike.14
From a theoretical viewpoint, the Palma ratio overcomes the excessive sensitiv-
ity of the Gini index in the middle of the distribution and its relative insensitivity to
changes at the top and bottom. By focussing on those parts of the distribution where
the differences concentrate, it becomes an index that clearly speaks about the
struggle between rich and poor and how changes in the middle class affect that
division.15 The Palma index should be seen, therefore, as a measure of group
inequality, in the spirit of measures of gender or spatial inequality, say.
The judgement of Cobham and Sumner (2013) about this index vis-a-vis the Gini
index is clear: “While the Gini and the Palma (indices) are closely correlated. . . the
Palma (index) should be strongly preferred as being “over”-sensitive to changes in
the distribution at the extremes, rather than in the relatively inert middle, since this
is what matters to policymakers. . . The differences in sensitivity, combined with
the relative stability of the intermediate deciles’ income share, militate in favour of
the Palma over the Gini. In addition, the clarity of the Palma favours its use for
policy targets where popular engagement may be important for accountability.”
In March 2013, a group of 90 well-established economists (now including Nobel
Prize-winning economist Joseph E. Stiglitz), academics and development experts
urged a key United Nations economic development panel to considering inequality
as one of the key concerns of economic and social development. They strongly
suggested using the Palma ratio to measure it. Subsequently, Doylan and Stigliz
(2014) proposed to add “Eliminating Extreme Inequality” as a ninth Sustainable
Development Goal for the post-2015 process.
14
These authors also provide empirical results that point out a link between countries’ Palma ratios
and their rates of progress on the major Millennium Development Goal (MDG) poverty targets.
15
As there are three groups clearly defined, any change in the middle one will affect the relative
position of the other two.
50 3 Positive Inequality Indices
The United Nations Development Program, the unit in charge of elaborating the
Human Development Reports, has included lately data on the quintile and the
Palma ratio, besides the Gini index. Table 3.1 below reproduces the data included
in the 2014 report. Countries are classified in different groups according to their
human development index (whose ranking is given in the first column, as a
reference). We have eliminated from the list all those countries for which none of
the three inequality measures was available.
The criticism to this ratio is the conventional one for quantile measures: this type
of index disregards the effect of changes within the first decile, the 2–5 deciles and
the 6–10 deciles. So any redistribution within those parts of the income distribution
does not affect the index. Moreover, the observed stability of the share of the middle
classes may not hold on the long run (see Milanovic, 2015).
Let us describe how the generalised index of Theil can be decomposed. The formula
of this uni-parametric family of indices is give by:
" #
1 1 X n
yi θ
I θ ðn; yÞ ¼ 1
n θðθ 1Þ i¼1 μ
with θ 6¼ 0,1 (the 0 and 1 cases have already been discussed). We can proceed as
follows:
54 3 Positive Inequality Indices
"
ng g θ
#
1 1 X G X
yi
I θ ðn; yÞ ¼ 1
n θðθ 1Þ g¼1 i¼1 μ
" n θ #
1 X 1 X
G g g
yi
¼ ng
θðθ 1Þ g¼1 n i¼1 μ
2 2 ! 3 !θ 3
g θ
1 X ng μ g θ 4 1 X
ng
G
4 i y μ
¼ 15 þ 1 5
θðθ 1Þ g¼1 n μ ng i¼1 μg μg
2 ! 3 " #
X g θ
1 X 1 X ng μ g θ
G ng G
n g μg θ 1 4 y 5
¼ i
1 þ 1
g¼1
n μ ng θðθ 1Þ i¼1 μg θðθ 1Þ g¼1 n μ
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
g
Iθ
The first term of this expression is a weighted sum of the indices Iθwithin groups.
The weights are given by:
ng μg θ
ωgG ðμ; nÞ ¼
n μ
This expression is interesting because it shows that only when θ ¼ 0 and θ ¼ 1 (the
first and second indices of Theil), those weights add up to 1.
The second term of that expression corresponds to Iθ applied to the mean values
of each group, that is, it describes the between groups component.
We can thus write:
" #
X
G
ng μ g θ 1 XG
ng μg θ
I θ ðn; yÞ ¼ I θg þ 1 ð3:22aÞ
g¼1
n μ θðθ 1Þ g¼1 n μ
or,
I θ ¼ I θ, W þ I θ , B ð3:22bÞ
where,
References 55
(i)
X
G
ng μ g θ
I θ, W ¼ I θg
g¼1
n μ
(ii) " #
1 XG
ng μg θ
I θ, B ¼ 1
θðθ 1Þ g¼1 n μ
References
Aaberge, R. (2000). Characterizations of Lorenz curves and income distributions. Social Choice
and Welfare, 17(4), 639–653.
Atkinson, A. B. (1970). On the measurement of inequality. Journal of Economic Theory, 2(3),
244–263.
Ben-Porath, E., & Gilboa, I. (1994). Linear measures, the Gini index, and the income-equality
trade-off. Journal of Economic Theory, 64, 443–467.
Bourguignon, F. (1979). Decomposable income inequality measures. Econometrica, 47(4),
901–920.
Box, G. E. P., & Cox, D. R. (1964). An analysis of transformations. Journal of the Royal Statistical
Society, Series B, 26(2), 211–243.
Chakravarty, S. R. (1988). Extended Gini indices of inequality. International Economic Review,
29(2), 147–156.
Cobham, A., & Sumner, A. (2013). Puntting the Gini back in the bottle? “The Palma” as a policy
relevant inequality measure. Working paper.
Cowell, F. A. (1977). Measuring inequality. Oxford: Phillip Allan.
Cowell, F. A. (1980). On the structure of additive inequality measures. Review of Economic
Studies, 47, 521–531.
Cowell, F. A. (1995). Measuring inequality, LSE handbooks in economics series (2nd ed.). Hemel
Hempstead: Harvester Wheatsheaf.
Cowell, F. A. (2003, Mayo). Theil, inequality and the structure of income distribution (Working
Paper DARP No. 67). STICERD, LSE.
Cowell, F. A., & Kuga, K. (1981a). Additivity and the entropy concept: An axiomatic approach to
inequality measurement”. Journal of Economic Theory, 25(1), 131–143.
Cowell, F. A., & Kuga, K. (1981b). Inequality measurement. An axiomatic approach. European
Economic Review, 15(3), 287–305.
Dasgupta, P. S., Sen, A. K., & Starrett, D. A. (1973). Notes on the measurement of inequality.
Journal of Economic Theory, 6(2), 180–187.
Donaldson, D., & Weymark, J. A. (1980). A single parameter generalization of the Gini indices of
inequality. Journal of Economic Theory, 22, 67–68.
Donaldson, D., & Weymark, J. A. (1983). Ethically flexible Gini indices for income distribution in
the continuum. Journal of Economic Theory, 29(4), 353–358.
Doylan, M. W., & Stigliz, J. E. (2014). Eliminating extreme inequality: A sustainable development
goal 2015-2030. http://www.ethicsandinternationalaffairs.org/2014/eliminating-extreme-
inequality-a-sustainable-development-goal-2015-2030/
Foster, J. E. (1983). An axiomatic characterization of the Theil measure of income inequality.
Journal of Economic Theory, 31(1), 105–121.
Gini, C. (1921). Measurement of inequality of incomes. The Economic Journal, 31, 124–126.
56 3 Positive Inequality Indices
Goerlich, F., & Villar, A. (2009). Desigualdad y bienestar. Fundación BBVA: De la teorı́a a la
práctica.
Imedio Olmedo, L. J., & Bárcena Martı́n, E. (2007). Dos familias numerables de medidas de
desigualdad. Investigaciones Econ omicas, 31(1), 191–217.
Kakwani, N. (1980). On a class of poverty measures. Econometrica, 48(2), 437–446.
Kanbur, R. (1984). The measurement and decomposition of inequality and poverty. In F. van der
Ploeg (Ed.), Mathematical methods in economics. New York: Wiley.
Khinchin, A. (1957). Mathematical formulations of information theory. New York: Dover
Publications.
Kolm, S.-C. (1976a). Unequal inequalities I. Journal of Economic Theory, 12, 416–442.
Kolm, S.-C. (1976b). Unequal inequalities II. Journal of Economic Theory, 13, 82–111.
Kullback, S. (1959). Information theory and statistics. New York: Wiley.
Lorenz, M. O. (1905). Methods of measuring the concentration of wealth. Publications of the
American Statistical Association, 9, 209–219.
Love, R., & Wolfson, M. C. (1976). Income inequality: Statistical methodology and Canadian
illustrations. Ottawa: Statistics Canada.
Mehran, F. (1976). Linear measures of income inequality. Econometrica, 44(4), 805–809.
Milanovic, B. (2015). Why I’m still not excited by the Palma index. Blog Globalinequality. http://
glineq.blogspot.com/2015/06/why-i-am-still-not-excited-by-palma.html
Palma, J. G. (2011). Homogeneous middles vs. heterogeneous tails, and the end of the ‘Inverted-
U’: The share of the rich is what it’s all about. Cambridge Working Papers in Economics
(CWPE) 1111. Cambridge University.
Rothschild, M., & Stiglitz, J. E. (1973). Some further results on the measurement of inequality.
Journal of Economic Theory, 6(2), 188–203.
Sen, A. K. (1973). On economic inequality. Oxford: Clarendon.
Sen, A. K. (1976). Poverty: An ordinal approach to measurement. Econometrica, 44(2), 219–231.
Shorrocks, A. F. (1980). The class of additively decomposable inequality measures.
Econometrica, 48(3), 613–626.
Shorrocks, A. F. (1984). Inequality decomposition by population subgroups. Econometrica, 52(6),
1369–1386.
Theil, H. (1967). Economics and information theory. Amsterdam: North-Holland.
Weymark, J. A. (1981). Generalized Gini inequality indices. Mathemathical Social Sciences, 1,
409–430.
Yaari, M. E. (1988). A controversial proposal concerning inequality measurement. Journal of
Economic Theory, 44(4), 381–391.
Yitzhaki, S. (1983). On an extension of the Gini inequality index. International Economic Review,
24(3), 617–628.
Chapter 4
Normative Inequality Indices
4.1 Introduction
Inequality analysis always involves value judgements, even though they are often
implicit (e.g. the impact of transfers between rich and poor agents on inequality). It
is then natural to consider this field of analysis a part of normative economics. The
normative approach to economic problems is characterised by explicitly dealing
with value judgements, which in this case translates into modelling inequality as a
social welfare loss. By linking inequality indices with social welfare, we can
discuss their normative content in terms of the principles that lead to the particular
welfare functions behind those indices.
We can distinguish two different approaches in this regard. First, the evaluation
of income distributions by means of social welfare functions, a la Arrow, which
aggregate individual utilities into a welfare indicator (this is the arrowian approach
to social evaluation). Second, a more general venue in which social evaluation is
defined directly on the space of income distributions, without having to go through
individual utilities (what will be called the social evaluation function approach).
We shall devote this chapter to discuss the arrowian approach to inequality
measurement, focused on the family of Atkinson’s inequality indices. The social
evaluation function approach will be discussed later on (see Chap. 6).
The social welfare function is a way of evaluating social alternatives using infor-
mational inputs as the utilities or preferences of the agents involved. Its origins go
back to authors such as Bentham, Mill, Condorcet or Borda. The modern approach
starts with Bergson (1938), but it is with the contribution of Kenneth Arrow (1951)
when this theory acquires maturity. Arrow showed the surprising result that there is
no social welfare function capable of satisfying a few mild and reasonable require-
ments (Arrow’s impossibility theorem).
Let Φ denote a set of social alternatives for a society N ¼ f1; 2; . . . ; ng, and let
ui : Φ ! ℝ denote agent i’s utility function, for i 2 N. Let u denote the vector of all
agents’ utility functions in this society, that is,
X
n
W u ðxÞ ¼ ui ðxÞ
i¼1
1
Those types of comparability reflect the class of invariant transformations in utilities that are
admissible. Under ordinal level comparability, any monotone increasing function that is common
to all agents is admissible. Under cardinal unit comparability, any linear transformation with
common slope for all agents is admissible.
2
We are implicitly assuming here that all utilities are different. When there are ties in the minimum
values, one chooses the option that is preferred for the first agent in the ranking from less to more
utility who is not indifferent.
4.1 Introduction 59
(where ui(y) is the utility of agent i that may depend on the whole distribution
vector). When measuring social welfare in this context, it is customary to look for
functional forms that permit one to express total welfare as a function of the “size”
and the “division” of total income. The size is most frequently associated with the
mean income (to cancel the effect of differences in population size) and the division
can be measured by some inequality index, I. That is, a function of the form3:
Note that for income vectors with the same mean, a convex combination yields
an income vector with a smaller dispersion, so that quasi-concavity reflects prefer-
ence for equality.
This property follows when all individual utilities are equal for all agents, only
depend on each one’s incomes, and are strictly concave. In that case, a transfer from
a rich to a poor enhances social welfare as the reduction in the utility of the rich is
more than compensated by the increase in the utility of the poor.
3
See Blackorby, Donaldson, and Auersperg (1981), Dutta and Esteban (1992), Lambert (1993,
Ch. 5) and Champernowne and Cowell (1997).
60 4 Normative Inequality Indices
Dalton (1920) introduced the idea of interpreting inequality as a social welfare loss
and argued for the need of an explicit formulation in terms of social welfare
functions. The basic idea is that of comparing the social welfare associated with
the actual income distribution with that welfare that could be reached under perfect
equality. Let us formulate these ideas more precisely.
Starting from the general framework in which agent i’s utility function is given by
ui ðyÞ ¼ ui ðy1 ; y2 ; . . . ; yn Þ, which initially may depend on the whole income distribu-
tion, Dalton introduced the following two basic assumptions on this function:
1. The utility of an agent only depends on the income of that agent
(no externalities), and it is the same for all agents: ui ðyÞ ¼ ui ðyi Þ ¼ uðyi Þ, 8i.
2. Utility is differentiable, positive, increasing and concave:
2
∂uðyÞ ∂ uð y Þ
uðyÞ 0, > 0, <0
∂y ∂y2
X
n
W ð yÞ ¼ uð y i Þ ð4:1Þ
i¼1
Note that
Xn
∂uðyi Þ
dW ¼ dyi
i¼1
∂yi
4.2 Dalton’s Approach to Inequality Measurement 61
so that those partial derivatives, ∂uðyi Þ=∂yi , can be regarded as the weights with
which individuals enter social welfare. The concavity of utility functions ensures
that those weights are decreasing in income so that this formulation introduces a
clear “preference for equality” in the evaluation of social welfare.
It is easy to check that the egalitarian distribution is the income distribution that
maximises social welfare for any given level of aggregate income, Y0. This obtains
from solving the following program:
X
n 9
>
>
Maxfyi gi¼1
n uðyi Þ >
=
i¼1
X
n
>
s:t: : yi ¼ Y 0 >
>
;
i¼1
The first order conditions of this program imply that λ ¼ ∂uðyi Þ=∂yi , 8i, where λ
is the corresponding Lagrange multiplier. Since u( y) is increasing and concave, in
the optimum, all incomes will be equal y1 ¼ y2 ¼ ¼ yn ¼ μ, and this condition
turns out to be also sufficient. That is, the maximum welfare associated with an
aggregate income Y 0 ¼ nμ is given by: Σ i¼1 n
uðyi Þ ¼ nuðμÞ.
Remark 4.1 Note that assuming equal utility functions for all agents is key to
obtain this result, as it implies that equalising all marginal utilities is a necessary
condition for the maximum.
The inequality measure proposed by Dalton (1920), denoted by D, can be
associated with the relative welfare loss due to inequality. That is 4:
X
n
nuðμÞ uðyi Þ
i¼1
D¼ ð4:2Þ
nuðμÞ
X
n
uðyi Þ
4
Dalton’s original formulation is given by i¼1n:uðμÞ , which takes on value 1 for zero inequality. We
use this formulation to satisfy the principle of normalisation.
62 4 Normative Inequality Indices
2 3
1 4∂uðyi Þ ∂u yj
δ 5 ð4:3Þ
nuðμÞ ∂yi ∂yj
4.3.1 Preliminaries
We shall address here some properties of utility functions that will be most useful to
modify Dalton’s formulation in a more consistent way. The concavity of utility
functions plays a key role in determining that the egalitarian allocation is the
income distribution that maximises social welfare, for all distributions with the
same mean. Moreover, the impact of a Dalton’s transfer described in Eq. (4.3)
depends on the curvature of the utility function (i.e. on the behaviour of marginal
utility). Larger values of Eq. (4.3) imply a greater concern for equality. The degree
of inequality aversion can thus be directly linked to the behaviour of marginal
utility, which is described by the second derivative of the individual utility function.
A conventional way of imposing restrictions on the behaviour of marginal utility
is in terms of its elasticity, ε, which is given by:
2
∂ð∂u=∂yÞ y ∂ y=∂y2
εð y Þ ¼ ¼ y ð4:4Þ
∂y ∂u=∂y ∂u=∂y
The elasticity of marginal utility measures the relative change of marginal utility
with respect to the relative change of individual income. This type of restriction is
robust to linear transformations and, therefore, consistent with cardinal utilities
(something that does not happen with marginal utilities). Consequently, this elas-
ticity provides a consistent index of inequality aversion.
4.3 Atkinson Inequality Measures 63
εyε1
y2ε y εyε1 yε y
εðyÞ ¼ ¼ ¼ε
1=yε y2ε
and, for ε ¼ 1,
1
y2 y
εð y Þ ¼ ¼1
1=y
Notice that ε ¼ 0 implies that u0 ðyÞ ¼ y, so that the utility is not strictly concave
which amounts to say that there is no concern whatsoever for equality. The case
ε ¼ 0 can thus be regarded as the limit for inequality aversion. Values ε > 0 entail
strictly concave utilities and, therefore, a positive degree of inequality aversion that
increases with ε. For ε ¼ 1, we have a logarithmic utility function, uðyÞ ¼ logy.
Note that marginal utilities determine the contribution of the changes in indi-
vidual incomes to social welfare, so that they can be associated with the implicit
weights given to the agents. Except for the case ε ¼ 0, those weights are decreasing
functions of individual incomes, as a direct consequence of the concavity assump-
tion. Yet, it is interesting to observe how fast those weights fall down as ε increases
(e.g. for ε > 2, the inequality aversion is very pronounced). To illustrate this note
5
Indeed, any linear transformation will also do because utilities are cardinal.
64 4 Normative Inequality Indices
that if we interpret income levels on the horizontal axis in relative terms (i.e. as y/μ),
we find that an agent with income ½ μ has a weight 4 times that of one with the
mean income when ε ¼ 2; it is 8 times higher for ε ¼ 3, 16 times higher for ε ¼ 4,
32 times higher for ε ¼ 5, . . .
X
n
nuðξÞ ¼ W ðyÞ ¼ uð y i Þ ð4:6Þ
i¼1
From this equation, once the utility function is known, we can easily deduce ξ.
Figure 4.1 provides a graphical illustration of this notion, where some indiffer-
ence curves of the social welfare function are depicted. The actual income distri-
bution is yA. The egalitarian equivalent income is obtained as the projection of
distribution yE (whereas yH describes the egalitarian distribution of the total
available income and permits one to deduce the mean income). We observe that
the convexity of indifference curves implies that ξ < μ, and that this difference is a
measure of inequality that depends on the degree of concavity of the social welfare
function (the curvature of the social indifference curves). When utility functions are
concave the social welfare function has convex indifference curves, whose curva-
ture is determined by the degree of inequality aversion.
6
This approach also appears in Kolm (1969). The notion of egalitarian equivalent income was
formerly introduced by Champernowne (1953).
4.3 Atkinson Inequality Measures 65
Income agent j
yH
yE
yA
yj
0 yi yi
Income agent i
ξ
A¼1 ð4:7Þ
μ
inequality aversion, uε ðÞ, which implies that the inequality index will satisfy scale
independence.7 Those two modelling options imply that the explicit form of the
social welfare function is given by:
8X n
>
> y1ε
>
<
i
ε 6¼ 1
1 ε
W ðyÞ ¼ Xi¼1
ð4:8Þ
>
>
n
>
: logyi ε¼1
i¼1
ξ1ε
nuðξÞ ¼ n
1ε
Xn Xn
y1ε
uðyi Þ ¼ i
i¼1 i¼1
1 ε
Hence,
" #1ε
1
ξ1ε Xn
yi1ε 1X n
n ¼ )ξ¼ y 1ε
1ε i¼1
1ε n i¼1 i
Substituting this value into (4.7) yields the following inequality index (for ε 0,
ε 6¼ 1):
" #1ε " #1
n 1ε 1ε
1
1 1X n
1X yi
Aε ¼ 1 y1ε
¼1
μ n i¼1 i n i¼1 μ
7
See Pratt (1964) and Lambert (1993, Ch. 4, Th. 4.2).
4.3 Atkinson Inequality Measures 67
Income of agent j
yH
yE yA
yj
0 yi yi
Income of agent i
nuðξÞ ¼ nlogξ
X
n X
n
uð y i Þ ¼ logyi
i¼1 i¼1
Hence,
X
n Yn
nlogξ ¼ logyi ) logξ ¼ log i¼1
ðyi Þ1=n
i¼1
)ξ¼μ
~ ð yÞ
where μ
~ ðyÞ is the geometric mean of the income distribution y.
Atkinson’s inequality index for ε ¼ 1 is thus given by:
μ
~ ðyÞ
A1 ¼ 1
μ ð yÞ
So the complete family of Atkinson inequality indices, for ε > 0, is given by:
8 " #1
>
> Xn 1ε 1ε
>
>1 1 y
< i
ε 6¼ 1
Aε ¼
n
i¼1
μ ð4:9Þ
>
> μ
~ ð Þ
>
> y
:1 ε¼1
μðyÞ
68 4 Normative Inequality Indices
ð1 Aε Þε h ε ε
i
δ y i y j <0
nμ1ε
Table 4.1 below provides information about the value of the Atkison index for
ε ¼ 1, which is probably the value of the parameter easiest to interpret, vis a vis the
Gini index. Note that in both cases the range of the indices is the interval [0, 1]. We
observe that the values of the Atkinson index are systematically lower than those of
the Gini index. Yet one has to be careful in the interpretation of those results. The
last two columns of Table 4.1 provide data of the relative values of both indices,
that is, the value of the index of each country over the mean of the index (expressed
as %). We observe that the Atkinson index discriminates more, as it tends to give
smaller values to those countries with inequality below the mean and higher values
to those above.
Let us conclude this chapter by analysing the relationship between the Atkinson and
the entropy family of inequality indices, discussed in Lecture 3. For ε 6¼ 1, it follows
from expression (4.9) that:
n 1ε
1X yi
¼ ð1 Aε Þ1ε
n i¼1 μ
4.3 Atkinson Inequality Measures 69
Table 4.1 Comparison of the Gini index and Atkinson’s index (for ε ¼ 1) for a selected sample of
countries in different years, from the Luxembourg Income Study (LIS) database
Gini Atkinson coefficient Gini/Mean Atkinson/Mean
LIS dataset coefficient (ε ¼ 1) (%) (%)
Australia 2010 0.333 0.186 95.6 88.4
Austria 2004 0.269 0.120 77.3 57.0
Belgium 2000 0.316 0.172 90.8 81.7
Brazil 2011 0.465 0.320 133.5 152.1
Canada 2010 0.318 0.170 91.3 80.8
China 2002 0.509 0.397 146.2 188.7
Colombia 2010 0.519 0.398 149.1 189.2
Czech Republic 0.256 0.106 73.5 50.4
2010
Denmark 2010 0.254 0.124 72.9 58.9
Egypt 2012 0.513 0.382 147.3 181.6
Estonia 2010 0.325 0.175 93.3 83.2
Finland 2010 0.263 0.114 75.5 54.2
France 2010 0.291 0.146 83.6 69.4
Germany 2010 0.288 0.135 82.7 64.2
Greece 2010 0.334 0.195 95.9 92.7
Guatemala 2006 0.509 0.381 146.2 181.1
Hungary 2012 0.291 0.143 83.6 68.0
Iceland 2010 0.246 0.105 70.7 49.9
India 2004 0.514 0.401 147.6 190.6
Ireland 2010 0.294 0.147 84.4 69.9
Israel 2010 0.387 0.238 111.1 113.1
Italy 2010 0.327 0.189 93.9 89.8
Japan 2008 0.302 0.154 86.7 73.2
Luxembourg 2010 0.269 0.120 77.3 57.0
Mexico 2010 0.442 0.316 126.9 150.2
The Netherlands 0.257 0.117 73.8 55.6
2010
Norway 2010 0.249 0.123 71.5 58.5
Peru 2004 0.524 0.414 150.5 196.8
Poland 2010 0.313 0.169 89.9 80.3
Romania 1997 0.284 0.132 81.6 62.7
Russia 2010 0.357 0.214 102.5 101.7
Serbia 2013 0.336 0.212 96.5 100.8
Slovak Republic 0.265 0.122 76.1 58.0
2010
Slovenia 2010 0.252 0.119 72.4 56.6
South Africa 2010 0.614 0.527 176.3 250.5
Spain 2010 0.334 0.211 95.9 100.3
Sweden 2005 0.239 0.099 68.6 47.1
Switzerland 2004 0.268 0.137 77.0 65.1
(continued)
70 4 Normative Inequality Indices
1 h i
I 1ε ¼ 1 ð1 Aε Þ1ε ð4:10Þ
εð1 εÞ
References
D’Aspremont, C. (1984). Axioms for social welfare orderings. In L. Hurwicz, D. Schmeidler, &
H. Sonnenschein (Eds.), Social goals and social organization. Essays in memory of Elisha
Pazner. Cambridge: Cambridge University Press.
D’Aspremont, C. (1994). Welfarism and interpersonal comparisons. Investigaciones Econ omicas,
18, 3–17.
D’Aspremont, C., & Gevers, L. (1977). Equity and the informational basis of collective choice.
Review of Economic Studies, 44, 199–209.
Dalton, H. (1920). The measurement of inequality of income. The Economic Journal, 30, 348–361.
Dutta, B., & Esteban, J. M. (1992). Social welfare and equality. Social Choice and Welfare, 9,
267–276.
Hammond, P. J. (1975). A note on extreme inequality aversion”. Journal of Economic Theory, 11,
465–467.
Kolm, S. C. (1969). The optimal production of social justice. In J. Margolis & H. Guitton (Eds.),
Public economics (pp. 145–200). London: Macmillan.
Lambert, P. J. (1993). The distribution and redistribution of income (2nd ed.). Manchester:
Manchester University Press.
Pratt, J. W. (1992). Risk aversion in the small and in the large. In G. Dionne & S. E. Harrington
(Eds.), Foundations of insurance economics (pp. 83–98). Netherlands: Springer.
Rawls, J. (1971). A theory of justice. Oxford: Oxford University Press.
Chapter 5
Inequality of Opportunity
5.1 Introduction
the issue of responsibility has become prominent in some of the recent develop-
ments within the areas of political philosophy and welfare economics.
The bottom line behind the equality of opportunity principle is that people who
are relatively disadvantaged due to external circumstances deserve some kind of
recognition or compensation. And, complementarily, that we should not be
concerned for those outcome differences among people with the same circum-
stances, as those differences derive from people’s autonomous choices (e.g. differ-
ential effort). Needless to say, the concepts of external circumstances and effort
have fuzzy boundaries, vary from one to another problem and may open some
difficult conceptual issues (e.g. are individuals fully responsible for their prefer-
ences? is luck-type dependent?). Yet, those notions refer to relevant categories that
matter for the ethical evaluation of income distributions.
For a sound discussion, the reader is referred to the works of Dworkin (1981a,
1981b), Arnesson (1989), Cohen (1989), Van de gaer, D (1993), Fleurbaey (1995),
Bossert (1995), Bossert and Fleurbaey (1996), Roemer (1996, 1998), Fleurbaey and
François (2001), Peragine (2002, 2004), Ruiz-Castillo (2003), Villar (2005), Bour-
guignon, Ferreira, and Menéndez (2007), Fleurbaey (2008), Lefranc, Pistolesi, and
Trannoy (2008, 2009), Checchi and Peragine (2010), Almas, Cappelen, Lind,
Sorensen, and Tungodden (2011), Calo-Blanco and Garcı́a-Pérez (2012), Fleurbaey
and Peragine (2013) and the literature cited therein. Roemer and Trannoy (2013,
2015) provide comprehensive discussions on this topic.
Following Roemer (1998), we shall assume that an income distribution may be
regarded as the result of (at least) two different effects: effort and opportunity.
Effort has to do with responsibility and involves people’s autonomous choices on a
common “playing field” (e.g. choice of occupation, investment in human capital,
length and intensity of work, healthy lifestyle, etc.). Opportunity refers to the
agents’ external circumstances, which may include genes, race, gender, family
socioeconomic and cultural background, and other aspects for which agents cannot
be held responsible. A fair society should compensate agents for differences in
opportunity but not for those differences derived from autonomous personal
decisions.1
We consider here a society with a finite number of individuals who are classified
into different types. Each type corresponds to a population subgroup with the same
opportunity (similar circumstances). By construction, therefore, the income distri-
bution of the agents of the same type is determined by their effort decisions. In other
words, agents of the same type have the same opportunity, and all income differ-
ences within a type correspond to differences in people’s effort decisions so that
they are ethically irrelevant.
Note that effort is typically an unobservable variable whose distribution may be
type-dependent, which means that a direct comparison of effort across types may be
problematic. To overcome this problem, we shall assume that there is some index
that allows comparing the degree of effort exerted by people with different
1
See Lefranc, Pistolesi, and Trannoy (2009) for a discussion on the role of luck.
5.1 Introduction 75
circumstances. A case in point is that proposed by Roemer (1998) who assumes that
effort is a single-valued variable that is positively correlated to income. So, even
though the effort distribution is a characteristic of the type, we can take the
quantiles of the outcome distribution within types to compare the degrees of effort.
That is to say, two individuals of different types exert a comparable degree of effort
if their incomes belong to the same quantile of the income distribution of their
corresponding types. Sometimes the effort variable is measured by some indirect
sociological trait as level of studies, labour experience, type of activity, etc.
We shall define an effort group as a set of agents who exert a comparable degree
of effort. Therefore, the income differences within effort groups provide a suitable
measure of the inequality of opportunity whereas income differences between effort
groups only represent diverse rewards of people’s autonomous choices and will not
be considered unfair.2 The basic principle underlying our analysis is that the income
distribution satisfies the principle of equality of opportunity whenever all people
who exert the same degree of effort receive the same income. And, by the same
token, those differences due to diverse degrees of effort are irrelevant. The impli-
cation is that income differences within effort groups reflect the inequality of
opportunity whereas income differences between effort groups are of no avail.
In the first part of this chapter, we address the measurement of inequality of
opportunity by taking advantage of the decomposability properties of Theil’s
inequality indices. The starting point is a double partition of the population by
circumstances (types) and effort (effort groups), which defines a collection of cells.
We then focus on effort groups and apply a decomposition of total inequality in
inequality within and between effort groups. Inequality of opportunity refers to that
within effort groups. Decomposing further the inequality within effort groups into
their corresponding within and between cells, the within component provides a
measure of the imprecision (and the impact of luck) in our choice of the double
partition.
The second part of the chapter provides an elementary way of applying the
equality of opportunity approach to the evaluation of societies when individual
outcomes are categorical rather than quantitative (e.g. when the data that describe
their achievements correspond to ordinal perceptions, positions in a ranking or
quality levels). To do so we start by grouping people according to their circum-
stances so that society is partitioned into a finite number of types, each of which
gathers individuals with similar circumstances. In that way all people of the same
type share the same opportunity and therefore, outcome differences within types
will be deemed irrelevant from an ethical viewpoint. The differences in the distri-
bution of individual achievements across types, on the contrary, can be regarded as
differences in people’s opportunities. This is the kind of inequality we are
interested in.
2
This is somewhat an extreme position that simplifies the analysis. In reality, the rewards of
different degrees of effort are dependent on the economic system and will typically change from
one to another society.
76 5 Inequality of Opportunity
5.2.1 Motivation
We shall apply in this section the decomposability properties of the (first) Theil’s
inequality index to the analysis of equality of opportunity along the lines laid down
above. To motivate the discussion, let us start by presenting a very simple example,
before going into the details of the formal model. Suppose we have a population
consisting of 200 individuals and we want to assess equality of opportunity regard-
ing labour earnings, taking into account a single circumstance: whether those
workers come from educated or uneducated families. The data of the corresponding
income distribution, keeping track of family characteristics, appears in Table 5.1.
To make things simple, we consider the case in which all groups of agents are of the
same size.
In this simplified world, we can define two different types of workers: those with
educated parents (Type 1) and those with uneducated parents (Type 2). We now
rearrange the data in Table 5.1 according to the distribution of labour income by
types and effort groups in Table 5.2. Each effort group is defined here by the
different levels of the income ranking, which we take as the relevant proxies of
effort for the different types. That is, two people belong to the same effort group if
their incomes belong to the same position in the ranking of incomes for the different
aim here, however, at measuring that part of inequality that derives from unequal
opportunities, not that which is the result of different effort. To do so we rely on the
property of decomposability by population subgroups of the first index of Theil,
which permits one to separate that part of the income dispersion that is due to
different opportunities from that due to different efforts.
Let us consider the decomposition of the Theil’s inequality index [see
Eq. (3.11)]:
X X g
G
ng μðyg Þ g G
ng μ ð yg Þ μ ðy Þ
T ð yÞ ¼ T ðy Þ þ log ð5:1Þ
g¼1
nμ ð y Þ g¼1
nμ ð y Þ μ ð yÞ
That is, we write the Theil index as the sum of two terms. The first one
corresponds to the within inequality, and it is the weighted average of the effort
groups’ inequality indices, with weights given by the corresponding income shares.
The second term measures the inequality between the groups and corresponds to the
weighted sum of the mean deviations, in logs.
Since we have assumed that inequality due to differential effort is not ethically
relevant, we can define Theil’s inequality of opportunity index, TIop(y), as the result
of cancelling the second term of the decomposition in (5.1) because that term
captures, precisely, the difference due to effort decisions. We can thus write,
XG
ng μðyg Þ g
T Iop ðyÞ ¼ T ðy Þ ð5:2Þ
g¼1
nμðyÞ
that is, the weighted sum of the partial indices of inequality of the effort groups,
with weights corresponding to the income shares.
It is important bearing in mind that the significance of this measure depends on
the definition of types and the partition of the population into effort groups. That is,
on the conceptual distinction between effort and circumstances, a difficult topic that
is to be addressed very carefully in each particular application.
We can get further insights on equality of opportunity by decomposing again
Theil’s inequality of opportunity index in terms of the constituent cells of each
effort group, as each effort group is made of different types. Applying the same
decomposability procedure to T(yg) in (5.2), we obtain:
Xτ Xτ t, g
ntg μðyt, g Þ tg ntg μðyt, g Þ μðy Þ
T ð yg Þ ¼ T ð y Þ þ log
t¼1
n g μ ð y gÞ
t¼1
n g μ ð y gÞ μðyg Þ
" t, g #
XG
ng μ ð yg Þ X τ
ntg μðyt, g Þ tg X τ
ntg μðyt, g Þ μ ðy Þ
T Iop ðyÞ ¼ T ðy Þ þ log : ð5:3Þ
g¼1
nμ ð y Þ t¼1
n g μ ðy gÞ
t¼1
n g μ ð y gÞ μðyg Þ
Equation (5.3) describes Theil’s inequality of opportunity index as the sum of two
components. The first component measures the inequality within cells. Ideally, each
term T(yt, g) of that weighted sum should be zero or close to zero, as all agents within
a cell will have similar incomes, provided the partition is sufficiently fine. We can
therefore take this factor as a measure of imprecision in the description of types and
effort groups (or, more precisely, as a measure of the variability that is not captured
by the variables in the partition). Let us call ρ(y) to this “error term”. That is,
" #
XG
ng μðyg Þ Xτ
ntg μðyt, g Þ
ρð yÞ ¼ T ðy Þ
t, g
ð5:4Þ
g¼1
nμðyÞ t¼1 ng μðyg Þ
When all relevant aspects are captured by the double partition between types and
effort levels, and there is no influence of luck, we have: ρðyÞ ¼ 0. Consequently,
when the partition is fine enough, the ratio ρ(y)/Tlop(y) is a measure of the influence
of luck (i.e., the % of the outcome that is due to the diverse realisations of the same
random variable). And, complementarily, in a deterministic world such a ratio can
be regarded as a relative measure of the roughness of our measurement, due to the
choice of the variables that define types and effort groups. Trivially, when there is
neither inequality of opportunity nor imprecision, TIop (.) ¼ 0.
80 5 Inequality of Opportunity
X
G
ng X
G
ng μ ð yÞ
T * ð yÞ ¼ T * ð yg Þ þ log
g¼1
n g¼1
n μðyg Þ
X
G
ng
T *Iop ðyÞ ¼ T * ð yg Þ
g¼1
n
X
τ
ntg X
τ
ntg μðyg Þ
T * ð yg Þ ¼ T * ðyt, g Þ þ log
t¼1
ng t¼1
ng μðyt, g Þ
Hence:
!
X
G
ng X
τ
ntg X
τ
ntg μðyg Þ
T *Iop ðyÞ ¼ T ðy Þ þ
* t, g
log ð5:7Þ
g¼1
n t¼1
ng t¼1
ng μðyt, g Þ
with
X
G
ng X
τ
ntg
ρð yÞ ¼ T * ðyt, g Þ
g¼1
n t¼1
ng
5.3 Equality of Opportunity With Categorical Data 81
X
G
ng X
τ
ntg μðyg Þ
IopðyÞ ¼ log
g¼1
n t¼1
ng μðyt, g Þ
5.3.1 Motivation
Let us start here with another simple example to illustrate the nature of the problem.
We have to evaluate the equality of opportunity regarding scholastic performance
in a given society. To fix ideas, suppose that our population of reference is that of
students at the end of compulsory education. Students’ final grades are given in
terms of a categorical scale consisting of five levels: A, B, C, D, E. We want to
analyse the relationship between students’ achievements and their family origin. In
a fair society, students’ results should not depend on their family origin. So equality
of opportunity here amounts to ensuring that the distribution of outcomes does not
depend on the students’ origins.
Table 5.3 presents the data of the evaluation problem. To keep things simple, we
assume that students are divided into four types (social groups) that correspond to
the quartiles of the family income distribution in this society. Each row describes
the distribution of the members of a type (a social class in this context) across
grades. Each column tells us how each grade is distributed among the types. Each
cell of the table says the fraction of students of a given type that obtain a specific
grade (e.g. the first cell of the table tells us that only 5 % of the students coming
from families in the first quartile get top marks).
3
We follow here Herrero and Villar (2014) where an application to the educational equality
of opportunity in Spain is presented, using PISA data.
82 5 Inequality of Opportunity
that is, αtc describes the share of individuals of type t ¼ 1, 2, . . ., with outcome
c ¼ 1, 2, . . ., γ.
Consider now the following matrix:
2 3
α11 α12 ... α1γ
6 α21 α22 ... α2γ 7
H¼6
4 ...
7 ð5:8Þ
... ... ... 5
ατ1 ατ2 ... ατγ
types. The differences in the values within a column correspond in our framework
to differences among types due to their differential circumstances.
Our target here is to define a measure that captures the inequality of opportunity
associated with a matrix H of relative frequencies of individual outcomes among a
population consisting of τ different types. Ideally, the distribution of outcomes
across types should be uniform. That is, in a fair society, the differences in people’s
external circumstances should not affect their realisations. As a consequence, the
observed differences in the distribution of realisations across types derive from
characteristics that involve diverse opportunities for the members of this society.
The inequality that is relevant for our purposes is, therefore, that within the
columns of matrix H, which describes the agents’ chances of having a given
categorical outcome depending upon their type. Clearly, there is no point in
fostering an egalitarian distribution within a type, as those internal differences are
ethically irrelevant because we interpret them as the result of autonomous
decisions.
Let α0 ðcÞ ¼ ðα1c ; α2c ; . . . ; ατc Þ stand for the (transpose of the) cth column vector
of matrix H. The dispersion of these shares tells us about the inequality of
opportunity with respect to categorical value c. We can apply a conventional
inequality measure I(α0 (c)) to assess that dispersion. The overall inequality of
opportunity associated with matrix H can thus be obtained as the weighted sum
of the inequality across categories, with weights equal to the corresponding popu-
lation shares.
That is,
X
γ
nc
IopðHÞ ¼ I ðα0 ðcÞÞ ð5:9Þ
c¼1
n
eðα0 ðcÞÞ
μ
I Að1Þ ðα0 ðcÞÞ ¼ 1 ð5:10Þ
μðα0 ðcÞÞ
(where μ
e stands for the geometric mean).
Consequently, the measure of inequality of opportunity will be given by:
4
For a more precise evaluation of the outcome distribution involving categorical variables, see
Herrero and Villar (2013), Villar (2016).
84 5 Inequality of Opportunity
Table 5.4 Computation of the elements of Inequality of opportunity measure (from Table 5.3)
A B C D E
μ 0.150 0.250 0.263 0.188 0.150
μ
e 0.126 0.247 0.262 0.178 0.126
1 ðμ
e = μÞ 0.1617 0.0102 0.0032 0.0516 0.1617
X
γ
nc eðα0 ðcÞÞ
μ
IopAð1Þ ðHÞ ¼ 1 ð5:11Þ
c¼1
n μðα0 ðcÞÞ
Applying this evaluation protocol to the example in Table 5.3 yields the follow-
ing (Table 5.4):
The associated inequality of opportunity measure is IopAð1Þ ðHÞ ¼ 0:0616
The Programme for International Student Assessment (PISA) provides the broadest
dataset for the evaluation of schoolchildren performance and the characteristics of
their schooling and family environment. It is a triennial worldwide test of 15-year-
old schoolchildren’s scholastic performance, the implementation of which is coor-
dinated by the OECD. The aim of the PISA study is to test and compare
schoolchildren’s performance across the world, with a view to improving educa-
tional methods and outcomes. The age of the students corresponds to the end of
compulsory education in most participating countries, so that those data can be
regarded as a good proxy of the basic knowledge that different countries ensure to
their citizens.
PISA surveys started in 2000 with the aim of evaluating the students’ ability,
about the end of compulsory education, in three different domains: reading, math-
ematics and science. Every period of assessment specialises in one particular
category, but it also tests the other two main areas studied. The 2012 report, the
one taken here as reference, focused on mathematics.
Achievement in PISA is primarily measured as the result of a test, graded in a
1000-point scale with a mean originally set at 500 and a standard deviation of 100.
Besides, the Programme establishes six levels of proficiency that try to approximate
the ability of the young to deal with different tasks. Each level corresponds to a
different capacity or set of skills (see OECD 2014a, 2014b, vol 1, p. 61 for details).
Those proficiency levels are conceived as intrinsically qualitative, even though they
are parameterised in terms of intervals of the average scores of the tests that
students realise in each subject. The data regarding the distribution of the students
5.3 Equality of Opportunity With Categorical Data 85
among those levels of proficiency shows that there is a large diversity among
countries, which is not captured by average scores. Those levels thus provide
relevant information on the structural features of the different educational systems.
Another feature of the PISA that makes it an extraordinary database refers to the
information collected on the students’ family and school environment. In particular,
one can use data regarding socio-economic conditions to analyse the degree of
equity of educational systems, in the understanding that more equity means less
dependence of the results on the family environment. The OECD provides a
summary variable that gathers the key information about the family environment:
the index of Economic and Socio-Cultural Status (ESCS). The ESCS is a composite
measure made of the following variables: the International Socio-Economic Index
of Occupational Status (ISEI); the highest level of education of the student’s
parents, converted into years of schooling; the PISA index of family wealth; the
PISA index of home educational resources and the PISA index of possessions
related to “classical” culture in the family home.
The OECD provides several measures of the degree of equity of educational
systems. Yet the distribution of the students among the proficiency levels is not
taken into account when analysing the equitability of educational systems. Here
again we find that those distributions provide information on the equitability of
educational systems, as we can compare the characteristics of those distributions by
social groups.
We shall illustrate here the evaluation model for equality of opportunity
presented above, by interpreting levels of proficiency as categories. We use the
distribution of results in the OECD countries regarding mathematics. As for the
definition of social groups (the types), we shall emphasise here the role of family
background, as it is regarded as the main explaining factor of differential outcomes.
More specifically, we divide the students of each country into four different types,
according to the quartile distribution of the index of Economic and Socio-Cultural
Status (ESCS). So we shall consider the population of each OECD country divided
into four different types, which correspond to the quartiles of the ESCS country-
specific distribution. Those types will be labelled Q1 (bottom 25 %), Q2, Q3 and Q4
(top 25 %).5
Table 5.5 describes the distribution of the students by proficiency levels
depending on their type, for the OECD as a whole. Table 5.7 in the Appendix to
this chapter provides the information for all OECD countries.
5
Most of the literature regarding educational equity adopts the equality of opportunity perspective,
with different ways of implementing that idea. Among the recent studies on equality of opportu-
nity in education, let us mention the works of Lefranc et al. (2009), Checchi and Peragine (2010),
Ferreira et al. (2011), Gamboa and Waltenberg (2012), De Carvalho et al. (2012), Carvalho,
Gamboa, and Waltenberg (2015) or Tansel (2015). Those works use average scores and different
ways of defining social groups to analyse the dependence of the outcomes on those conditioning
variables. The measurement of educational achievements or cognitive skills by comparing the
distributions of outcomes by levels or proficiency, using a different procedure, can be found in
Herrero and Villar (2013), Herrero, Méndez, and Villar (2014) and Villar (2014, 2016).
86 5 Inequality of Opportunity
The data send a clear message: the family background is an important factor in
the achievement of competences, especially for those at the top and the bottom ends
of the distribution of the socioeconomic conditions. The degree of equality of
opportunity, according to the evaluation formula (5.11), appears in Table 5.6,
both in absolute and relative terms. Belgium, Chile, France, Hungary, Israel,
Luxembourg, New Zealand, Portugal and the Slovak Republic exhibit values of
inequality of opportunity more that 40 % points higher than the OECD mean.
Canada, Estonia, Finland, Iceland, Japan, Korea, Mexico and Norway present
values more than 25 % points below that mean.
Table 5.5 Distribution of students by levels of proficiency and quartiles of the ESCS for the
OECD (PISA 2012)
Level 5* Level 4 Level 3 Level 2 Level 1 Level < 1
OECD Q1 0.0463 0.1096 0.2062 0.2663 0.2246 0.1470
Q2 0.0843 0.1605 0.2476 0.2552 0.1710 0.0813
Q3 0.1385 0.2072 0.2577 0.2195 0.1249 0.0522
Q4 0.2468 0.2586 0.2441 0.1554 0.0695 0.0256
*
Levels 5 and 6 have been aggregated into a single one denoted by 5*
Table 5.6 Inequality of opportunity in educational outcomes in OECD countries (PISA, 2012)
Lopp Rel. Iopp (OECD ¼ 100)
Australia 0.0609 99
Austria 0.0736 120
Belgium 0.0975 159
Canada 0.0462 75
Chile 0.1061 173
Czech Republic 0.0786 128
Denmark 0.0759 123
Estonia 0.0433 70
Finland 0.0438 71
France 0.1200 195
Germany 0.0838 136
Greece 0.0748 122
Hungary 0.1129 184
Iceland 0.0374 61
Ireland 0.0730 119
Israel 0.0980 159
Italy 0.0472 77
Japan 0.0437 71
Korea 0.0456 74
Luxembourg 0.1022 166
Mexico 0.0444 72
The Netherlands 0.0562 92
New Zealand 0.0917 149
(continued)
5.3 Equality of Opportunity With Categorical Data 87
Appendix
Table 5.7 Distribution of students by levels of proficiency and quartiles of the ESCS for the
OECD (PISA 2012)
Level 5* Level 4 Level 3 Level 2 Level 1 Level < 1
Australia Q1 0.0590 0.1210 0.2203 0.2704 0.2189 0.1104
Q2 0.1015 0.1675 0.2685 0.2478 0.1550 0.0598
Q3 0.1803 0.2223 0.2616 0.2057 0.0958 0.0343
Q4 0.2659 0.2649 0.2446 0.1476 0.0585 0.0187
Austria Q1 0.0460 0.1212 0.2083 0.2852 0.2229 0.1164
Q2 0.1160 0.1799 0.2510 0.2454 0.1470 0.0607
Q3 0.1503 0.2516 0.2621 0.2098 0.0955 0.0306
Q4 0.2634 0.2922 0.2460 0.1337 0.0488 0.0159
Belgium Q1 0.0602 0.1220 0.2234 0.2545 0.2006 0.1393
Q2 0.1220 0.1911 0.2463 0.2275 0.1435 0.0697
Q3 0.2321 0.2462 0.2483 0.1569 0.0830 0.0335
Q4 0.3834 0.2773 0.1893 0.0952 0.0386 0.0162
Canada Q1 0.0775 0.1691 0.2683 0.2680 0.1523 0.0648
Q2 0.1297 0.2062 0.2726 0.2482 0.1131 0.0302
Q3 0.1781 0.2488 0.2829 0.1886 0.0815 0.0201
Q4 0.2872 0.2899 0.2420 0.1289 0.0413 0.0106
Chile Q1 0.0008 0.0086 0.0540 0.1871 0.3542 0.3953
Q2 0.0068 0.0264 0.1252 0.2572 0.3388 0.2455
Q3 0.0068 0.0518 0.1687 0.3050 0.3109 0.1568
Q4 0.0498 0.1626 0.2737 0.2654 0.1738 0.0747
(continued)
88 5 Inequality of Opportunity
References
Almas, I., Cappelen, A., Lind, J. T., Sorensen, E., & Tungodden, B. (2011). Measuring unfair (in)
equality. Journal of Public Economics, 95(7–8), 488–499.
Arnesson, R. (1989). Equality and equal opportunity for welfare. Philosophical Studies, 56,
159–194.
Bossert, W. (1995). Redistribution mechanisms based on individual characteristics. Mathematical
Social Sciences, 29, 1–17.
Bossert, W., & Fleurbaey, M. (1996). Redistribution and compensation. Social Choice and Welfare,
13(3), 343–355.
Bourguignon, F., Ferreira, F. H., & Menéndez, M. (2007). Inequality of opportunity in Brazil.
Review of Income and Wealth, 53(4), 585–618.
Calo-Blanco, A., & Garcı́a-Pérez, J. I. (2012). On the welfare loss caused by inequality of
opportunity. The Journal of Economic Inequality. doi: 10.1007/s10888-013-9247-6.
Carvalho, M., Gamboa, L. F., & Waltenberg, F. D. (2015). Equality of educational opportunity:
Taking both achievement and access into account (Ecineq Working Paper 2012-277).
Checchi, D., & Peragine, V. (2010). Inequality of opportunity in Italy. Journal of Economic Inequal-
ity, 8, 429–450.
Cohen, G. A. (1989). On the currency of egalitarian justice. Ethics, 99(4), 906–944.
De Carvalho, M., Gamboa, L. F., & Waltenberg, F. (2012). Equality of educational opportunity
employing PISA data: Taking both achievement and access into account. Universidad del
Rosario, Serie Documentos de Trabajo, 130.
Dworkin, R. (1981a). What is equality? Part 1: Equality of welfare. Philosophy and Public Affairs,
10, 185–246.
Dworkin, R. (1981b). What is equality? Part 2: Equality of resources. Philosophy and Public Affairs,
10, 283–345.
Ferreira, F. H. G., Gignoux, J., & Aran, M. (2011). Measuring inequality of opportunity with
imperfect data: The case of Turkey. Journal of Economic Inequality, 9, 651–680.
Fleurbaey, M. (1995). Equality and responsibility. European Economic Review, 39(3–4), 683–689.
Fleurbaey, M. (2008). Fairness, responsibility, and welfare. Oxford: Oxford University Press.
Fleurbaey, M., & François, M. (2001). Compensation and responsibility. In K. J. Arrow, A. Sen, &
K. Suzumura (Eds.), Handbook of social choice and welfare (Vol. 2). Amsterdam: North-
Holland.
Fleurbaey, M., & Peragine, V. (2013). Ex ante and ex post equality of opportunity. Economica, 80,
118–130.
Foster, J. E., & Shneyerov, A. A. (1999). A general class of additively decomposable inequality
measures. Economic Theory, 14, 89–111.
Gamboa, L. F., & Waltenberg, F. D. (2012). Inequality of opportunity in educational achievement
in Latin America: Evidence from PISA 2006-2009. Economics of Education Review, 31,
694–708.
Herrero, C., Méndez, I., & Villar, A. (2014). Analysis of group performance with categorical data
when agents are heterogeneous: The evaluation of scholastic performance in the OECD through
PISA. Economics of Education Review, 40, 140–151.
Herrero, C., & Villar, A. (2013). On the comparison of group performance with categorical data.
PLoS One, 8(12), e84784.
Herrero, C., & Villar, A. (2014). Equality of opportunity with categorical data. Lecturas de Eco-
nomía, 81, 9–30.
Lefranc, A., Pistolesi, N., & Trannoy, A. (2008). Inequality of opportunities vs inequality of
outcomes: Are western societies all alike? Review of Income and Wealth, 54, 513–546.
Lefranc, A., Pistolesi, N., & Trannoy, A. (2009). Equality of opportunity and luck: Definitions and
testable conditions, with an application to income in France. Journal of Public Economics,
93(11–12), 1189–1207.
92 5 Inequality of Opportunity
OECD. (2014a). PISA 2012 results: What students know and can do: Student performance in
mathematics, reading and science (Vol. I). Paris: OECD Publishing.
OECD. (2014b). PISA 2012 results: Excellence through equity (Vol. II). Parı́s: OECD Publishing.
Peragine, V. (2002). Opportunity egalitarianism and income inequality: The rank-dependent
approach. Mathematical Social Sciences, 44, 45–64.
Peragine, V. (2004). Measuring and implementing equality of opportunity for income.
Social Choice and Welfare, 22, 1–24.
Roemer, J. (1996). Theories of distributive justice. Cambridge: Massachusetts Harvard University
Press.
Roemer, J. (1998). Equality of opportunity. Cambridge: Massachusetts Harvard University Press.
Roemer, J., & Trannoy, A. (2013). Equality of opportunity, Cowles foundation discussion Paper
no 1921.
Roemer, J., & Trannoy, A. (2015). Equality of opportunity: Theory and measurement. Journal of
Economic Literature, forthcoming.
Ruiz-Castillo, J. (2003). The measurement of inequality of opportunities. Research in Eco-
nomic Inequality, 9, 1–34.
Tansel, A. (2015). Inequality of opportunities in educational achievements in Turkey over time
(IZA Working Paper DP 9005).
Van de gaer, D. (1993). Equality of opportunity and investment in human capital. PhD Dissertation,
Catholic University of Leuven.
Villar, A. (2005). On the welfare evaluation of income and opportunity. Contributions to Theo-
retical Economics, 5(1): article 3. http://www.bepress.com/bejte/contributions/vol5/iss1/art3
Villar, A. (2014). Education and cognitive skills in the Spanish adult population. Intergenerational
comparison of mathematical knowledge from PIAAC Data. Advances in Social Sciences
Research Journal, 1(1), 72–88.
Villar, A. (2016). Levels of proficiency, performance and equity in educational outcomes:
The OECD according to PISA 2012, mimeo.
Chapter 6
Inequality and Welfare
6.1 Introduction
1
He also assumes monotonicity (i.e., higher aggregate income implies higher social welfare), but
this is a property we may be willing to drop to give more weight to equity considerations.
6.2 Social Evaluation Functions 95
We have already discussed those properties, so we shall not insist here on their
interpretation.
Sen (1973) defines the generalised egalitarian equivalent income, ye, as the per
capita income that satisfies the following:
Vð1n ye Þ ¼ VðyÞ
with 1n ¼ ð1; 1; . . . ; 1Þ . That is, this notion corresponds to the former ξ obtained
|fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}
n
from a social welfare function defined on a different space.
As V(y) is symmetric and strictly quasi-concave, it follows that ye μðyÞ. Then
we can define the following family of inequality indices:
ye
SV ¼ 1 ð6:1Þ
μ ð yÞ
This formula differs from Atkinson’s (4.7) in that now the egalitarian equivalent
income does not depend on the individual utilities and represents, therefore, a more
general expression. The sub-index V is introduced to remind that ye depends on the
specific SEF adopted.
This change of formulation gives us much more flexibility to choose the func-
tional form of the welfare function V. Moreover, assuming that this function is
homogeneous of degree one (scale independence) ensures a one-to-one relationship
between welfare functions and inequality measures. Needless to say, V ðyÞ ¼ Σ i¼1 n
uðyi Þ implies SV ¼ A.
As in the case of Atkinson, the value of Sen’s inequality measure SV can be
interpreted in terms of a money metric that provides an estimate of the welfare loss
due to inequality.
This approach, though, opens the question of how can we find reasonable
restrictions on those social evaluation functions in order to obtain sensible inequal-
ity indices. Of course one can always plug a standard inequality index in the left
hand side of equation (6.1) and find out the kind of egalitarian equivalent income
that it implies and then try to relate that to a specific social evaluation function.
This discussion can be very much simplified if we introduce the following three
additional assumptions on the welfare function:
• Differentiability: Function V has continuous derivatives. That is, 8y 2 ℝþþ
n
,
8i 2 N, ∂V=∂yi is a continuous function.
• Homogeneity of degree one: 8y 2 ℝþþ n
, 8λ > 0, V ðλyÞ ¼ λV ðyÞ.
• Scale: V ðμ; μ; . . . ; μÞ ¼ μ.
The first assumption is a technical requirement that facilitates the handling of the
formulae and has already been discussed. The homogeneity property establishes
96 6 Inequality and Welfare
that if we multiply all incomes by a given constant, then the associated welfare is
multiplied by this constant (it is a parallel requirement to that of scale indepen-
dence). Note that it implies monotonicity along rays. Finally scale is an assumption
that defines the units in which we are going to measure welfare. It establishes that
when all incomes are equal, we can take the mean income as the welfare value of
the income distribution (which amounts to adopt a money metric for welfare
analysis).
Remark 6.1 Note that under scale, the concept of inequality described in (6.1)
corresponds to one minus the ratio between actual welfare and the welfare that
would be obtained under perfect equality.
Those three properties (differentiability, homogeneity and scale) ensure that, for
all y 2 ℝþþn
, we can find a unique scalar γ ðyÞ 2 ℝþþ such that V ðyÞ ¼ γ ðyÞ. To see
this, notice that, by homogeneity, for any given y 2 ℝþþ n
we can find λ0 , λ1 > 0 such
that V ðλ0 1n μÞ < V ðyÞ < V ðλ1 1n μÞ, where 1n is the unit n-vector. Then, as V is
continuous (by differentiability), the intermediate value theorem ensures that there
will be some γ ðyÞ 2 ðλ0 ; λ1 Þ such that V ðyÞ ¼ V ð1n γ ðyÞÞ (uniqueness derives from
the monotonicity feature induced by homogeneity). Then, scale implies that
V ð1n γ ðyÞÞ ¼ γ ðyÞ.
Assuming symmetry and strict quasi-concavity, we can identify this value γ(y)
with the egalitarian equivalent income, ye, which describes the inequality adjusted
mean income. Consequently, we have:
V ðyÞ ¼ ye ð6:2Þ
V ð yÞ ¼ μ ð yÞ ½ 1 S V ð yÞ ð6:3aÞ
This is a simple and very intuitive expression that links nicely inequality and
welfare. For each inequality index, this equation defines a welfare function that
permits one to evaluate income distributions as the per capita income deflated by
inequality. Indeed, the term μ(y)SV(y) describes the per capita welfare loss due to
inequality, expressed as a money metric. This can be interpreted as the additional
per capita income that could be enjoyed by society in the absence of inequality.
Needless to say, V ðyÞ ¼ μðyÞ if and only if there is perfect equality.
Adopting this approach to inequality measurement permits one to construct
welfare measures for income distributions, using either positive or normative
inequality indices. As an illustration, take Atkinson’s inequality index for ε ¼ 1.
We would have: SV ¼ A1 ðyÞ ¼ 1 ðμ e=μÞ, where μe stands for the geometric mean.
By substituting in (6.3), we obtain:
V A1 ðyÞ ¼ μ
e ð yÞ ð6:3bÞ
6.2 Social Evaluation Functions 97
which is the welfare measure used by the United Nations to introduce equity
considerations in the Human Development Index (United Nations Development
Programme, 2010), briefly discussed in a later section.
Note that when the inequality index takes values larger than 1 the welfare index
(6.3) becomes negative, and the welfare function is not monotone in the mean
income. This would express the idea that our concern for equality is so strong that
when inequality goes beyond this threshold, any increase in the mean income
produces a reduction of social welfare. This may happen (more theoretically than
in practice) for the Theil’s family of inequality indices.
Since we have adopted a money metric of social welfare, the value of function
V will depend on the income units. This problem disappears when we measure the
welfare loss in relative terms. That is:
μðyÞSV ðyÞ SV ð yÞ
Z V ð yÞ ¼ ¼ ð6:4Þ
V ð yÞ 1 S V ð yÞ
This expression is simply the relative welfare loss and tells us the share of
additional income that could be enjoyed by the society in the absence of inequality.
Moreover, the trivial expression,
μðyÞSV ðyÞ
SV ð yÞ ¼
μ ð yÞ
permits one interpreting the magnitude of the inequality index, SV(y), as the welfare
loss per unit of income.
***
There is another interesting venue to linking inequality and welfare along these
lines. The properties of differentiability and homogeneity of degree one allow
writing the social evaluation function as follows:
X
n
V ð yÞ ¼ αi ðyÞyi ð6:5Þ
i¼1
Xn
V ð yÞ ¼ i¼1
αi y yi ¼ μðyÞ½1 I ðyÞ
where αi(y) is a function homogenous of degree zero and I(y) is a relative inequality
index.
An immediate way of realising the interest of this formulation is by means of the
following result, which gives us the social evaluation function associated with the
Gini index:
Proposition 6.1: (Sen, 1976)
The weighting system αiG ðyÞ ¼ ðn i þ 1Þ=n yields a social evaluation function
Therefore,
X
n
μðyÞn n þ 1
αiG ðyÞyi ¼ GðyÞ
i¼1
2 n
2
We assume implicitly that all incomes are different to avoid problems with differentiability
in the SEF.
6.2 Social Evaluation Functions 99
n1 1
α1G ðyÞ ¼ 1, α2G ðyÞ ¼ , . . . , αnG ðyÞ ¼
n n
Note that, as it happened with the effect of a Dalton’s transfer in the Gini
coefficient, those weights are independent on the magnitude of the income differ-
ences (a property sometimes referred to as homothetic distributivity).
Another relevant case is that of Theil’s first inequality index for its special
properties regarding decomposability. Let denote by υ(y) the social evaluation
function associated with that inequality index, that is, υðyÞ ¼ μðyÞ½1 T ðyÞ.
The following result is easy to obtain and clarifies the normative content of
Theil’s index in terms of the weights given to the individuals depending on their
incomes.
Proposition 6.2: (Herrero & Villar, 1989)
The weighting system αiT ðyÞ ¼ ½1 logðyi =μÞ=n yields the following social eval-
uation function:
Proof
(i) Making use of the definition of the first index of Theil in equation (3.10), we
have:
1X n
yi yi
T¼ log
n μ μ
" i¼1 #
1X n
yi yi
¼1 1 log
n i¼1 μ μ
" #
1X n
y i 1X n
yi yi
¼1 log
n i¼1 μ n i¼1 μ μ
n
1X y
¼1 1 log i yi
nμ i¼1 μ
1 X n
¼1 α T ðyÞyi
μ i¼1 i
Therefore,
X
n
αiT ðyÞyi ¼ μð1 T Þ
i¼1
Q.E.D.
100 6 Inequality and Welfare
The index of Theil is therefore a way of valuing inequality in which agents with
income equal to the mean are given a weight of 1/n whereas this value decreases
progressively as the income grows above the mean and increases progressively as
income gets smaller than the mean. That is, as ∂ðyi =μÞ=∂yi is positive, we have:
Observe that for incomes above eμ(y) (where e is the base of natural logarithms),
those coefficients become negative. So we can say informally that Theil’s social
evaluation function gives negative weights to all those incomes above three times
the mean income.
The decomposability of Theil’s inequality index can be translated to the
corresponding evaluation function υ(y). To see this assume that society is made
of G different groups, g ¼1, 2, 3, . . . , G,
each one with ng members and an income
distribution vector yg ¼ y1g ; y2g ; . . . ; yngg (where ygi is the income of agent i in
μ ¼ ðμ1 ; μ2 ; . . . ; μG Þ denote the vector of mean incomes, where μg is
group g). Let ~
the mean income of group g. We can write the society’s mean income as:
X
G
ng
μðyÞ ¼ μ g ð yg Þ
g¼1
n
Let us recall that Theil’s index can be decomposed as follows [see Eq. (3.11)]:
X
G
ng μ g X
G
ng μ g μg
T ð yÞ ¼ T ð yg Þ þ log
g¼1
nμ g¼1
nμ μ
ð6:8Þ
1 X
G
μg
¼ ng μg T ðyg Þ þ log
nμ g¼1 μ
g¼1
n g μ
X
G
ng μg
¼ μg 1 T ðyg Þ log
g¼1
n μ
X
G
ng X
G
ng μg
¼ μg ð1 T ðyg ÞÞ μg log
g¼1
n g¼1
n μ
6.3 Multidimensional Inequality and Welfare 101
X
G
ng
υðyÞ ¼ υðyg Þ υB ð ~
μÞ ð6:9Þ
g¼1
n
The first component corresponds to the weighted sum of the social evaluations of
the different social groups, taken in isolation, with weights equal to their population
shares. Note that each individual evaluation, υ(yg), already computes the social
welfare loss due to inequality within that group. The second component is a
measure of the welfare loss due to inequality between the different groups and is
given by the weighted sum of the logs of the ratios between the mean incomes of the
different groups and the overall mean income, with weights given by the per capita
income of each group.
It is an interesting exercise deducing other indices and comparing how those
inequality measures implicitly give weights to individual incomes in the welfare
function. Observe that with this formulation, we are able to give an explicit welfare
content to those inequality measures initially belonging to the descriptive class.
3
Let us mention the United Nations 1954 report on the standards of living, the “basic needs
approach” fostered by the International Labour Organization in 1974, the Physical Quality of Life
Index (PQLI), due to Morris (1979) [reformulated by Ram (1982)] or that proposed by the Daj
Hammarskj€old Foundation. For more recent critiques, see Boarini, Johansson, and Mira d’Ercole,
(2006), Stiglitz, Sen, and Fitoussi (2009) or Fleurbaey (2009).
102 6 Inequality and Welfare
Martı́nez, and Villar (2010a, 2010b) and Seth (2013)]. From a practical point of
view, this is done either by aggregating first the dimensions for each individual into
some utility value and then aggregating utilities into a social welfare value, or
assuming some type of separability across dimensions that permits one to evaluate
first each dimension independently and then aggregate them into a real-valued
indicator.
There are pros and cons in both approaches. Aggregating first all dimensions by
individuals permits one capturing the cross effect of those dimensions on each agent
and keeping track of whether insufficient achievements in different dimensions
refer to the same or different individuals. This requires that all the data
corresponding to the different dimensions are obtained from the same sample or
at least of samples that can be regarded as perfect substitutes. This is a very
demanding requirement that is not very often matched in practice. Another diffi-
culty refers to the choice of individual utility functions to proceed to such aggre-
gation (we have briefly referred to this difficulty in Sect. 6.1). The separability
approach is formally easier and accommodates better data coming from different
sources. The key limitation is that it simplifies extremely the interplay between the
different variables and ignores the cumulative effect of shortages over the same
individuals. We shall follow here this easier approach, which we may regard as an
extension of the welfare evaluation function approach presented above.4
Building a social evaluation function in a multidimensional context involves
some difficult issues that call for agreement and compromise. In particular:
(i) Which are the most relevant dimensions to be considered?
(ii) How can we approximate those dimensions by means of specific variables
whose data are available?
(iii) How should those variables be aggregated into a single index in order to get a
systematic evaluation criterion? More specifically,
The complexity of reaching a general agreement on those issues may partly
explain the persistence of uni-dimensional welfare and equity measures. Yet, a
number of multidimensional indicators have been developed recently and are
becoming new references for the evaluation (e.g. United Nations Human Develop-
ment Index or the OECD How’s Life proposal). The relevance of this approach has
brought the OECD to issue a manual on the construction of composite indices
(Nardo et al., 2008).
Multidimensional inequality is a complex topic that will not be discussed here
except in an extremely elementary way. We refer to Lugo (2005), Chakravarty
4
There is still the possibility of not aggregating the different dimensions, either by using a
dashboard of variables or applying dominance criteria to get partial orderings (e.g., Kolm (1977),
(a) How should we treat the cross effects among dimensions on individuals and society?
(b) How should we weight the different dimensions?
(c) How should we choose the right units of measurement?
Atkinson and Bourguignon (1982); see Savaglio (2006) for a discussion).
6.3 Multidimensional Inequality and Welfare 103
(2009, Ch. 5) and the recent survey by Aaberge and Brandolini (2014) for a
thorough discussion of this topic.5 Moreover, as Aaberge and Brandolini (2014,
p. 8) point out, “the theoretical literature on the multidimensional measurement of
inequality and poverty has been growing very rapidly in the last quarter of a
century, and is still far from consolidation”.
We have already discussed those properties, so we shall not insist here on their
interpretation. Note, though, that strict quasi-concavity in this context is a very
weak form of the Dalton principle of transfers.
One of the main difficulties that appears when dealing with multidimensional
welfare evaluation refers to the presence of cross effects between dimensions. A
way of dealing with this difficulty (actually a way of circumventing it) is by
introducing some notion of separability across dimensions, which permits one to
evaluate multidimensional welfare as a relatively simple function of the different
dimensions considered in isolation. Separability is a property closely related to the
preferential independence axiom in utility theory (e.g. Keeney & Raiffa, 1976,
Ch. 3). It establishes that if social state Y is considered at least as good as social
state Y’, when there is a common distribution of one dimension (both have an
identical column), then this relation holds for all common values of this column.
To state formally this property, we shall write Y ¼ ðYj , yðjÞÞ, where y( j)
denotes the jth column of matrix Y and Yj the n ðk 1Þ matrix with the
remaining columns. Then,
• Separability. For all Y, Z 2 Ω and all j 2 K,
W Yj , yðjÞ W Zj , yðjÞ ) W Yj , y0 ðjÞ W Zj , y0 ðjÞ
Two particular cases of separability are worth considering: additive and multi-
plicative separability. Additive separability is defined by the following property:
X
k
W ðYÞ ¼ aj W ðyðjÞÞ ð6:10aÞ
j¼1
that is, total welfare appears as a weighted sum of the welfare of each dimension
considered in isolation. The coefficients aj 2 ℝþ measure the relative importance of
Xk
each of those dimensions, with a ¼ 1.
j¼1 j
Multiplicative separability adopts the form:
Yk aj
WðYÞ ¼ j¼1
½WðyðjÞÞ ð6:10bÞ
The key difference between these two particular instances of separability refers
to the substitutability between dimensions. Additive separability implies constant
marginal rates of substitutions whereas multiplicative separability yields the more
conventional decreasing marginal rates of substitution.
Given a welfare evaluation function W that satisfies symmetry
and strict
quasi-
concavity, and a social state matrix Y, we denote by ye Yj , yðjÞ 2 ℝ the
egalitarian equivalent
associated with the distribution of the jth dimension
value
in Y. That is, y Yj , yðjÞ is implicitly defined by the following equation:
e
6.3 Multidimensional Inequality and Welfare 105
W ðYÞ ¼ W Yj , 1n ye Yj , yðjÞ
When the social evaluation function satisfies the property of scale, we get the
following more precise forms:
8 Xk
< aj yje ðadditive separabilityÞ
W ðYÞ ¼ Y kj¼1 aj ð6:12Þ
: yje ðmultiplicative separabilityÞ
j¼1
which correspond to the arithmetic mean and the geometric mean of the egalitarian
equivalent values, when we assume that all dimensions are equally important.
W ðY Þ
I ðYÞ ¼ 1 X k
aμ
j¼1 j j
where μj stands for a short-handed version of μ(y( j)). That is, one minus the ratio
between total welfare, as measured by our social evaluation function, and the mean
value of the variables, which corresponds to the welfare evaluation of the achieve-
Xk
ments when there is no inequality; that is, a μ ¼ W ð1n μ1 , . . . , 1n μk Þ (see
j¼1 j j
Remark 6.1).
106 6 Inequality and Welfare
That is, the multidimensional version of an inequality index I(∙), under additive
separability, corresponds to the ratio of the welfare loss due to inequality and
the average value of the achievements. Such a welfare loss is simply the
weighted sum of the single-dimensional inequality indices applied to each
dimension separately, with weights given by the scaled means.
(b) Multiplicative separability:
Yk aj
j¼1
μj ð1 I ðyðjÞÞÞ
I ðYÞ ¼ 1 Xk ð6:13bÞ
aμ
j¼1 j j
That is, the multidimensional version of an inequality index I(∙), under multipli-
cative separability, can be written as 1 minus the ratio between the inequality
adjusted geometric mean of the welfare of each dimension and the aggregate
mean, all pondered by the corresponding coefficients aj.
Let us take the case of Atkinson’s inequality index, for the value ε ¼ 1, in order
to illustrate this formulation. Under additive separability in (6.13a), we obtain:
Xk
j¼1
aj μ
ej
I Að1Þ ðYÞ ¼ X k ð6:14aÞ
aμ
j¼1 j j
Note that in each case we find that total inequality is a function of inequality within
dimensions and inequality between dimensions.
6.3 Multidimensional Inequality and Welfare 107
6
For a discussion on different techniques for setting weights for multidimensional indices, see
Decancq and Lugo (2013).
108 6 Inequality and Welfare
Using those units implies giving much more weight to the income dimension just
because of its higher mean.
This problem can be avoided by using some kind of normalisation. For instance,
dividing each original variable yij by some reference value, vj, of the corresponding
dimension (e.g. the mean, the median, the maximum of a given range of values),
thus yielding values ^y ij ¼ yij =vj . This change of variables leaves unaltered the
inequality index of each dimension and neutralises the effect of the different units.
Yet, the decision on the reference value matters because the mean of the scaled
variables depends on that reference.
The case of normalising the original values by using the mean value of each
dimension, ^y ij ¼ yij =μj , is particularly interesting. This is so because in that case the
Xn
mean of the scaled variable is always equal to 1, μ ^ j ¼ ð1=nÞ i¼1 ^y ij ¼ 1, so that
(6.13a) and (6.13b) become:
Xk
^ ¼
I Y aj I ðyðjÞÞ ð6:15aÞ
j¼1
Yk
^ ¼
I Y ½I ðyðjÞÞaj ð6:15bÞ
j¼1
There is another normalisation venue that we find often in the literature, based on
the idea of confining the range of all variables to the interval [0, 1]. This is done by
choosing a minimum and a maximum value for each dimension and transforming
the original variables into relative gains, by means of the following transformation:
yij ymin
j
^y ij ¼ ð6:16Þ
ymax
j ymin
j
where ymax
j , ymin
j represent the maximum and minimum values for the jth dimension,
respectively.
This type of normalisation presents some disadvantages in the context of
multidimensional welfare and inequality. In particular, inequality measures become
dependent on the particular choice of minimum values, even in the single dimen-
sional case. This implies, when comparing different societies, that the ranking of
those societies in terms of inequality or welfare depends on the choices of those
minimum values. Moreover, the use of minimum goalpost in the normalisation may
have a very large effect on the marginal rates of substitution, under multiplicative
separability, due to the behaviour of the slope of Cobb-Douglas functions when a
given component approaches zero. Therefore, subtracting whatever amount to an
already close to zero magnitude will increase substantially and artificially the
associated marginal rates of substitution.
6.4 The Human Development Index 109
The Human Development Index (HDI, for short) is probably the most successful
multidimensional welfare indicator nowadays. It was proposed by the United
Nations in 1990 in order to assess the well-being of a society, based on Amartya
Sen’s idea of functionings and capabilities [see Sen (1985)]. It identifies three basic
dimensions related to human welfare, which are regarded as equally important:
health, education and material well-being. Achievements in health are associated
with the variable life expectancy at birth, measured in years. Achievements in
education are approximated by a mixture of two variables: mean years of schooling
and expected years of schooling. Finally, the achievements in material well-being
are measured through the log of the standard per capita GDP (actually the per capita
gross national product). All those average values are normalised so that they range
over the interval [0, 1].
Between 1990 and 2009, the United Nations adopted the additive separability
approach and defined the Human Development Index as the arithmetic mean of the
average values of the variables that approximate the achievements in the three basic
dimensions, after normalisation. Due to the critiques on the design of the HDI,7 and
in particular to the assumption of constant rates of substitution between dimensions,
in 2010 the HDI was substantially amended and the multiplicative separability
approach was adopted so that now is the geometric mean of the mean values of
those variables. Some other innovations were introduced in that year, such as the
change of the variable that measured education, the construction of consistent time
series and, most especially, the addition of an inequality-adjusted Human Devel-
opment Index.
Let μcH , μcE , μcY stand for the mean values of the variables that measure health,
education and income, respectively, for country c (we ignore the normalisation
process for the time being). Then the Human Development index for this country is
defined as follows:
1=3
HDI c ¼ μHc μEc μYc ð6:17Þ
The HDI is a simple and intuitive construct that refers to very relevant aspects of
the socio-economic performance and uses data that are available in most countries.
Those features allow for widespread international comparisons that are accessible
to non-specialists. That probably explains its popularity and the relevance given by
the media to the yearly publication of each new wave of data.
Note that equation (6.17) disregards completely the unequal distribution of
achievements within the country. That is, inequality is set to zero by definition, as
it is implicitly assumed that yje ¼ μj . To deal with a long-standing requirement of
7
See Osberg (1985), Anand and Sen (1994a, 1994b), Hicks (1997), Phillipson and Soares (2001),
Osberg and Sharpe (2002), Chakravarty (2003), Becker, Philipson, and Soares (2005), Foster et al.
(2005), Herrero et al. (2010a, 2010b) for a critical appraisal and some alternative formulations.
110 6 Inequality and Welfare
References
Aaberge, R., & Brandolini, A. (2014). Multidimensional poverty and inequality (Banca d’Italia
Working Paper No. 976).
Alkire, S., & Foster, J. E. (2010). Designing the inequality—Adjusted human development index
(IHDI) (OPHI Working Paper 37). University of Oxford: Oxford Poverty and Human Devel-
opment Initiative.
Anand, S., & Sen, A. K. (1994a). Human development index: Methodology and measurement
(Human Development Report Office Occasional Paper 12). New York: United Nations Devel-
opment Program.
Anand, S., & Sen, A. K. (1994a). The income component of the human development index.
Journal of Human Development, 1, 83–107.
Atkinson, A. B., & Bourguignon, F. (1982). The comparison of multidimensional distributions of
economic status. Review of Economic Studies, 49(2), 183–201.
Becker, G. S., Philipson, T. J., & Soares, R. R. (2005). The quantity and quality of life and the
evolution of world inequality. The American Economic Review, 91(1), 277–291.
Blackorby, C., & Donaldson, D. (1978). Measures of relative equality and their meaning in terms
of social welfare. Journal of Economic Theory, 18(1), 59–80.
8
The only reasonable argument that I know to treat income differently is due to Zambrano (2014),
who interprets that health and education are direct capabilities whereas income is only an indirect
one that refers to the command over resources.
References 111
Boarini R., Johansson, A., & Mira d’Ercole, M. (2006). Alternative measures of well-being (OECD
Social, Employment and Migration Working Papers No. 33). Paris: OECD.
Bourguignon, F., & Chakravarty, S. R. (1999). A family of multidimensional poverty measures. In
Advances in econometrics, income distribution and scientific methodology (pp. 331–344).
Physica-Verlag HD.
Chakravarty, S. R. (2003). A generalized human development index. Review of Development
Economics, 7, 99–114.
Chakravarty, S. R. (2009). Inequaltiy, polarization and poverty. New York: Springer.
Cherchye, L., Ooghe, E., & Puyenbroeck, T. V. (2008). Robust human development rankings.
Journal of Economic Inequality, 6, 287–321.
Decancq, K., & Lugo, M. A. (2013). Weights in multidimensional indices of well-being: An
overview. Econometric Reviews, 32, 7–34.
Dutta, B., & Esteban, J. M. (1992). Social welfare and equality. Social Choice and Welfare, 9,
267–276.
Fleurbaey, M. (2009). Beyond GDP: The quest for a measure of social welfare. Journal of
Economic Literature, 47(4), 1029–1047.
Foster, J. E., Lopez-Calva, L. F., & Szekely, M. (2005). Measuring the distribution of human
development: Methodology and an application to Mexico. Journal of Human Development, 6,
5–29.
Foster, J. E., McGillivray, M., & Seth, S. (2009). Rank robustness of composite indices (OPHI
Working Paper 26). University of Oxford: Oxford Poverty and Human Development Initiative.
Foster, J. E., McGillivray, M., & Seth, S. (2013). Composite indices: Rank robustness, statistical
association and redundancy. Econometric Reviews, 32, 35–56.
Gajdos, T., & Weymark, J. A. (2005). Multidimensional generalised Gini indices. Economic
Theory, 26, 471–496.
Herrero, C., & Villar, A. (1989). Comparaciones de renta real y evaluación del bienestar. Herri
ekonomiaz. Revista de economı´a pública, 2, 79–102.
Herrero, C., Martı́nez, R., & Villar, A. (2010a). Multidimensional social evaluation An application
to the measurement of human development. Review of Income and Wealth, 56, 483–497.
Herrero, C., Martı́nez, R., & Villar, A. (2010b). Improving the measurement of human develop-
ment (Human Development Reports Research Papers 2010-12).
Herrero, C., Martı́nez, R., & Villar, A. (2012). A newer human development index. Journal of
Human Development and Capabilities, 13, 247–268.
Hicks, D. A. (1997). The inequality adjusted human development index: A constructive proposal.
World Development, 25(8), 1283–1298.
Keeney, R. L., & Raiffa, H. (1976). Decisions with multiple objectives: Preferences and value
tradeoffs. New York: Wiley.
Kolm, S. C. (1977). Multidimensional egalitarianisms. The Quarterly Journal of Economics, 91,
1–13.
Lugo, M. A. (2005). Comparing multidimensional indices of inequality: Methods and application.
Society for study of economic inequality. Working Paper, 14.
Maasoumi, E. (1986). The measurement and decomposition of multi-dimensional inequality.
Econometrica, 54(4), 991–997.
Maasoumi, E. (1989). Continuously distributed attributes and measures of multivariate inequality.
Journal of Econometrics, 42, 131–144.
Morris, M. D. (1979). Measuring the condition of the world’s poor: The physical quality of life
index. Washington, DC: Overseas Development Council.
Nardo, M., Saisana, M., Saltelli, A., Tarantola, S., Hoffman, A., & Giovannini, E. (2008).
Handbook on constructing composite indicators: Methodology and user’s guide. Paris: Joint
Research Centre (JRC) of the European Commission and OECD.
Neumayer, E. (2011). Sustainability and inequality in human development (Human Development
Research Paper 2011/04).
OECD. How’s life. http://www.oecd.org/statistics/how-s-life-23089679.htm
112 6 Inequality and Welfare
7.1 Introduction
which the one considered in this Lecture, this criterion is usually linked to the
notion of a poverty line. A poverty line defines a threshold that determines who is
going to be considered poor and non-poor. This threshold typically represents an
estimate of the consumption opportunities considered as essential (or the minima
admissible values of a series of variables). We find two broad categories or those
poverty lines, absolute poverty lines and relative poverty lines. Absolute poverty
lines provide an estimate of some consumption threshold that permits satisfying the
most basic needs (food, shelter, clothing), assuming that they are somehow univer-
sal. An example that is used commonly nowadays is that of “two dollars per person
and day” in terms of parity purchasing power. Relative poverty lines, which are the
most frequently used, are defined for each society as a function of their own
parameters. The idea behind is that needs, even basic needs, are relative to the
society of reference so that who is poor depends on the social environment. Relative
poverty lines are usually defined in terms of a given percentage of the mean income
or, more often, the median. This implies that proportional income changes do not
affect the number of the poor in a society (a homogeneity property similar to that of
inequality indices). In some cases, one considers a relative poverty line in a given
point in time (e.g. a given year) and holds it fixed for a period of analysis (e.g. the
years of the crisis).
Fourth, one has to specify whether the measurement exercise refers to one or
several periods. In the second case, it is important to know not only how poverty
evolves but also whether the poor are always the same or they keep changing. The
reason is that poverty in the long run entails social exclusion as a major side effect, a
feature that depends on the degree of social mobility. When social mobility is low,
falling beyond the poverty threshold may be very dramatic because it involves a
difficult return path.
In this lecture, we focus on the analysis of objective poverty measures, referred
to a single period, taking “income” as the key reference variable, and using a
relative poverty line to define who are the poor. We shall also refer to
multidimensional (objective) poverty measures. As it was the case with the study
of inequality, we shall assume that the right choices of the money variable and the
units of reference have already been made.
In this scenario, it is customary to assume that poverty measurement involves a
direct or indirect appraisal of three different aspects, known as Sen’s three I’s of
poverty (Sen 1976): Incidence, Intensity and Inequality. Incidence refers to how
many poor people are in society (usually expressed as a fraction of the total
population), which is determined by those who are below the poverty line. The
most common standard nowadays is that of considering poor in a given society to
those whose income is below 60 % of the median income of that society. Intensity
tries to capture how poor are the poor. This is important because being below the
poverty line conveys no information about how far away are people from that
threshold. Inequality aims at measuring the income dispersion within the poor. All
those aspects are captured, in different ways, by a series of conventional poverty
indices that are explained in the next sections.
Before presenting those poverty indices, it might be worth taking into account
that we shall be referring to very different realities when talking about poverty in
7.2 Poverty Indices 117
developing and developed countries. Not only because their average standards of
living are very diverse but because of the penetration of market institutions, the
degree of substitutability between primary goods, and hence the relevance of
monetary values, may also be very different.
Our reference is a society with n agents (e.g. consumption units) with an income
vector y ¼ ðy1 , y2 , . . . , yn Þ. We shall assume that a poverty line z has been
established in order to define what insufficient income means.
A poverty index is a function P : Ω ℝ ! ℝ, where Ω stands for the set of all
possible income distributions. For each pair (y, z), where y is an income distribution
vector and z a given poverty threshold, P(y, z) is a quantitative measure of poverty
that fits our intuition of this phenomenon.
It is common to require that this function satisfies some basic properties, such
as1:
• Normalisation: Pðy; zÞ ¼ 0 , yi z8i (the index is zero if and only if there are
no poor agents in society).
• Symmetry: Pðy; zÞ ¼ Pðπ ðyÞ, zÞ (permuting agents’ incomes does not change
the value of the index).
• Continuity: P is a continuous function of the incomes of the poor.
• Focus: Let y ¼ ðyp ; ynp Þ, where yp, ynp are the income
vectors
ofthe poor and the
_np
non-poor, respectively. Then, Pððyp ; ynp Þ; zÞ ¼ P yp ; y ; z for all possible
_np
y (i.e. the index is independent on the incomes of the non-poor).
• Monotonicity: Reducing the income of a poor agent, other things equal,
increases the value of the index.
• Scale independence: Pðλy, λzÞ ¼ Pðy; zÞ, 8λ > 0.
• Decomposability by population subgroups: Suppose that society is made of Q
different population subgroups. Then,
X
Q
nq
Pðy; zÞ ¼ P½yq ; z ð7:1Þ
q¼1
n
1
There is a long collection of properties that have been discussed in the literature. We mention here
the most common ones. For a review of all those properties, see Chakravarty (2009, Ch. 2).
118 7 Poverty Measurement
1X n
Pðy; zÞ ¼ gð y i ; z Þ ð7:2Þ
n i¼1
where g(yi, z) ¼ P(yi, z) corresponds to individual i’s poverty measure, with g(yi,
z) ¼ 0 if and only if yi z. So we can attach a poverty index to each individual,
that depends only on her income and the poverty threshold, by using the same
evaluation function for all of them.2
Let us consider now some families of poverty indices that satisfy all or most of
those properties.
The head count ratio, PH, is probably the simplest and one of the most frequent
measures of poverty, also known as poverty risk. It is defined as the share of the
poor in society, that is,
p
PH ðy; zÞ ¼ ð7:3Þ
n
2
Decomposability is a particular case of consistency (see Foster and Shorrocks, 1991).
3
Note that p depends on the whole income distribution y and the poverty line. Precision would
require writing p(y, z). Yet we shall renounce to that precision to get a lighter notation.
7.2 Poverty Indices 119
This index only takes into account the incidence of poverty and disregards the
other two aspects, intensity and inequality.
and assume that the poverty threshold is given by z ¼ 12.5. If we apply the head
count ratio we conclude that poverty is the same in both societies, as the two of
them have half the population below the poverty threshold (and also inequality
within the poor is zero in both cases). Yet the two situations are extremely different
because the poor in the first distribution are very close to the poverty threshold
whereas those in the second distribution are very far away.
Poverty gap indicators introduce a measure of the distance between the income
of the poor and the poverty line. The simplest of those indices is the poverty gap
index, PPG, a measure of absolute distances given by:
X
p
PPG ðy; zÞ ¼ ðz y i Þ ð7:4aÞ
i¼1
The value of this index can be interpreted as the amount of income required to
allow the poor reaching the poverty line. This can be better visualised by rewriting
(7.4) as follows:
X
p
PPG ðy; zÞ ¼ pz y i ¼ p z μ p ð yÞ ð7:4bÞ
i¼1
(where μp(y) is the mean income of the poor). In this format, the poverty gap
describes the difference between the mean income of the poor and the poverty gap,
times the number of the poor. Note that this measure does not compute the
incidence (as it does not keep track of the size of the whole population) and does
not satisfy scale independence.
We can obtain a relative poverty measure that avoids those inconveniencies by
simply dividing the above expression by nz, which corresponds to the minimum
income that society would require to avoid poverty. The measure so obtained is
know as the poverty gap ratio, PPGR, and is defined as follows:
PPG ðy; zÞ p μp ðyÞ
PPGR ðy; zÞ ¼ ¼ 1 ð7:5Þ
nz n z
This is an index that provides a nice expression combining the incidence and
intensity of poverty. Note that it is an index that satisfies decomposability by
population subgroups.
120 7 Poverty Measurement
and let z ¼ 14.1 (which corresponds to the 60 % of the mean income in both
distributions). Calculating the poverty gap ratio, we get:
p μ p ð yÞ 2 7
PPGR ðy; zÞ ¼ 1 ¼ 1 ¼ 1:0072
n z 4 14:1
p μ p ð y0 Þ 2 7
PPGR ðy0 ; zÞ ¼ 1 ¼ 1 ¼ 1:0072
n z 4 14:1
That is, both distributions exhibit identical poverty levels according to that indica-
tor, thus missing the extreme differences in income distributions within the poor. So
this poverty index does not pay attention to the third element in poverty measure-
ment: inequality.
Sen (1976) suggested a way of introducing inequality into poverty measurement
by substituting the mean income of the poor in equation (7.5) by the mean income
deflated by Gini inequality index applied to the income distribution of the poor,
μp ðyÞ½1 GðyÞ. This idea was further extended by other authors (e.g. Blackorby &
Donalson, 1980, Chakravarty 1983, 1990), by suggesting different ways of
substituting the mean income by some other reference income.
For our purposes, a general way of introducing the inequality dimension into the
measurement of poverty is by substituting the mean income of the poor by the
egalitarian equivalent income of the poor,
h i
yIe yðpÞ ¼ μp 1 I yðpÞ
where I(y( p)) is a relative inequality index referred to the income distribution of the
poor, y( p).
We have, then, the following family of indices4:
p μ p ð yÞ ½ 1 I ð yÞ
PI ðy; zÞ ¼ 1 ð7:6Þ
n z
Remark 1 Note that this family of inequality measures does not satisfy decom-
posability by population subgroups, precisely because of the relative nature of the
inequality index that enters the formula. Yet a different version of subgroup
4
We discuss further this family of indices in the next Lecture.
7.2 Poverty Indices 121
1X y α
p
α
PFGT ðy; zÞ ¼ 1 i ð7:7Þ
n i¼1 z
That is, the formula corresponds to the aggregate individual relative poverty gaps,
to the power of α, over the total population.
Note that some particular values of α yield familiar formulae, as it is the case for
the following:
p
0
PFGT ðy, zÞ ¼ ¼ PH ðy, zÞ
n
p μp ðyÞ
1
PFGT ðy, zÞ ¼ 1 ¼ PPGR ðy, zÞ
n z
5
Note that we can write the square of the coefficient of variation as:
p
σ 2 1 X ðyi μÞ2 1 X
p
y2i 2μyi
CV 2 ¼ ¼ ¼ 1 þ
μ2 n i¼1 μ2 n i¼1 μ2
122 7 Poverty Measurement
p 2 p
1X y 1X y2 2zy
2
PFGT ðy, zÞ ¼ 1 i ¼ 1þ i 2 i ð7:8Þ
n i¼1 z n i¼1 z
In the same vein, we find Chakravarty’s (1983) decomposable index, given by:
1Xh y e i
p
PCh ðy; zÞ ¼ 1 i ð7:10Þ
n i¼1 z
There is evidence that not all elements that affect people’s well-being can be
reduced to income. Think for instance of the case of education which exhibits a
rather complex connection with income or GDP. Some of the countries with better
outcomes in the OECD’s PISA study are far from being among the richest countries
(see OECD, 2014). The multidimensional approach to well-being is nowadays a
standard of evaluation, with some solid initiatives into play, as it is the case with the
United Nations human development index [see also the Stiglitz, Sen, and Fitoussi
(2009) report].
It is then natural to consider the multidimensional approach to poverty measure-
ment, since poverty is a phenomenon with many faces that can hardly be reduced to
a metric of insufficient income. To start with, note that the consumption possibil-
ities depend not only on monetary income but also on the availability of
non-marketed goods (something very important in the less developed countries),
the access to basic services (water, electricity, communications), the availability of
durables, the housing conditions, the public services locally provided, etc.
7.3 Multidimensional Poverty Indices 123
Moreover, for low levels of income one finds that some goods hardly admit
substitution so that some additional income cannot compensate their accessibility.
Poverty is also related to social exclusion and insufficient participation in the labour
market, which are again aspects beyond low income. Finally, one should take into
account that market prices may not be suitable aggregators as they reflect average
consumption patterns that may be very different from those accessible to the poor.
See Dardadoni (1995), Tsui (2002), Bourguignon and Chakravarty (2003), Aaberge
and Brandolini (2014) or Alkire et al. (2014, 2015) for a discussion.
All those considerations have led to developing a multidimensional approach to
poverty measurement as well as a series of more qualitative indicators that try to
capture deprivation.
Note that moving from one to several dimensions, when approaching the mea-
surement of poverty, opens a number of difficult issues as already mentioned:
(i) Which are the most relevant dimensions to be considered besides (or instead
of) monetary income? (ii) How can we approximate those dimensions by means of
specific variables whose data are available? (iii) How should those variables be
aggregated into a single index in order to get a systematic evaluation criterion?
Those are the key questions to be addressed in any multidimensional analysis.
We shall focus here on the last one, i.e. the formulae of multidimensional indices,
assuming that the other two issues have already been solved and also that all
dimensions can be approached by numerical variables.
In a multidimensional context, the achievements of an agent i are described by a
K-vector yi of realisations. Consequently, the information regarding the whole
society will be given by an n K matrix Y in which rows represent the agents’
realisations, yi, and columns, y( j), the distribution of each variable in society.
Similarly, we shall find now a K-vector z of thresholds that describe the different
poverty levels, one for each dimension. And, unless we assume that all dimensions
are equally important, we would need to determine a vector of weights, b, which
tells us about the relative importance of each dimension.
A multidimensional poverty index is, therefore, a mapping P(Y, z, b) that
applies the triplet “realisations, thresholds, weights” into the real numbers.
Note that now “counting the poor” is a more difficult exercise since we can find
people that are above the poverty thresholds in some dimensions and below in some
others. There are two extreme positions in the literature dealing with this question.
On the one hand, considering poor those agents who fall below some poverty
threshold (what is known as the union approach). On the other hand, declaring
poor only those who fall below all poverty thresholds (the so called intersection
approach). There are also some intermediate, and somehow arbitrary, approaches
that consider poor to those who are below a certain number of poverty thresholds.
(We take up this point again in the next chapter).
All the properties presented in the former section have an immediate extension
to the case of multidimensional poverty. We shall introduce a new requirement,
124 7 Poverty Measurement
factor decomposability, that makes much easier to translate the standard poverty
indices to the multidimensional case, by assuming that there are no cross effects
between dimensions worth considering.
Formally:
• Factor decomposability: A multidimensional poverty index P(Y, z, b) satisfies
factor decomposability when it can be written as:
X
K
PðY; z; bÞ ¼ bj P yðjÞ, zj ð7:11Þ
j¼1
where P(y( j), zj) is the poverty index relative to the jth dimension.
When a multidimensional poverty index satisfies both population and factor
decomposability, it can be written as:
1X K X n
PðY; z; bÞ ¼ bj g yij ; zj ð7:12Þ
n j¼1 i¼1
Following the 2020 European Strategy, people at risk of poverty or social exclusion
are those who meet at least one of three different criteria: insufficient income
(equivalised income below 60 % of the median), low work intensity (below 20 %
of their labour capacity) or severe material deprivation (inability to afford 4 or more
out of 9 items). People who suffer more than one of those sources of poverty are
computed just once.
Let us explain how those notions are defined.
The at-risk-of-poverty rate is the share of people with an equivalised disposable
income (after social transfers) below the at-risk-of-poverty threshold, which is set at
60 % of the national median equivalised disposable income after social transfers.
This indicator does not measure wealth or poverty, but low income in comparison to
other residents in that country, which does not necessarily imply a low standard of
living.
The indicator “persons living in households with low work intensity” is defined
as the number of persons living in a household having work intensity below a
threshold set at 0.20. The work intensity of a household is the ratio of the total
number of months that all working-age household members have worked during the
126 7 Poverty Measurement
reference year and the total number of months the same household members
theoretically could have worked in the same period.6
Material deprivation refers to a state of economic strain and durables, defined as
the enforced inability (rather than the choice not to do so) to pay unexpected
expenses, afford a one-week annual holiday away from home, a meal involving
meat, chicken or fish every second day, the adequate heating of a dwelling, durable
goods like a washing machine, colour television, telephone or car, being confronted
with payment arrears (mortgage or rent, utility bills, hire purchase instalments or
other loan payments).
The material deprivation rate is an indicator in EU-SILC7 that expresses the
inability to afford some items considered by most people to be desirable or even
necessary to lead an adequate life. The indicator distinguishes between individuals
who cannot afford a certain good or service, and those who do not have this good or
service for another reason, e.g. because they do not want or do not need it.
The indicator adopted by the Social protection committee measures the percent-
age of the population that cannot afford at least three of the following nine items:
1. To pay their rent, mortgage or utility bills
2. To keep their home adequately warm
3. To face unexpected expenses
4. To eat meat or proteins regularly
5. To go on holiday
6. A television set
7. A washing machine
8. A car
9. A telephone
Severe material deprivation rate is defined as the enforced inability to pay for
at least four of the above-mentioned items.
Figure 7.1 below provides an illustration of the situation in Europe according to
this way of measuring poverty.
Table 7.1 gives us further details on the evolution of the AROPE index
6
A working-age person is a person aged 18–59 years, with the exclusion of students in the age
group between 18 and 24 years. Households composed only of children, of students aged less than
25 and/or people aged 60 or more are completely excluded from the indicator calculation.
7
The European Union Statistics on Income and Living Conditions (EU-SILC) is an instrument
aiming at collecting timely and comparable cross-sectional and longitudinal multidimensional
microdata on income, poverty, social exclusion and living conditions. This instrument is anchored
in the European Statistical System (ESS). The EU-SILC provides two types of data: (a) Cross-
sectional data pertaining to a given time or a certain time period with variables on income, poverty,
social exclusion and other living conditions. (b)Longitudinal data pertaining to individual-level
changes over time, observed periodically over a four-year period. Social exclusion and housing
condition information is collected mainly at household level while labour, education and health
information is obtained for persons aged 16 and over. The core of the instrument, income at very
detailed component level, is mainly collected at personal level.
7.4 Deprivation and Non-monetary Poverty Measures 127
In the 2010 Human Development Report, the United Nations Development Pro-
gram (UNDP) introduced “Multidimensional Poverty Index (MPI)” for developing
countries, following the ideas in Alkire and Santos (2010). This shows the UNDP’s
willingness to mark a clear departure from the use of composite indices to
multidimensional indices that are able to capture joint distributions across the
population. The MPI considers three different dimensions, education, health and
standard of living, which are captured by ten categorical indicators. The indicators
and their deprivation cut-offs are reported in Table 7.2. The health dimension and
the education dimension consist of two indicators each, and the standard of living
dimension consists of six indicators.
The construction of the MPI, which is an adaptation of the Adjusted Headcount
Ratio proposed by Alkire and Foster (2007, 2011), can be illustrated as follows.
Consider a society with n individuals and d indicators (ten in case of the MPI). Let
xij denote the achievement of individual i with respect to dimension j. The society’s
achievements will be described by a n X 10 matrix X. The weight attached to
indicator j is denoted by bj > 0 such that b ¼ 1. Each indicator has its own
j j
deprivation cut-off. A person failing to meet the cut-off is identified as deprived in
that dimension. The deprivation cut-off of indicator j is denoted by zj. Subject to the
deprivation cut-off, person j is assigned a deprivation status score in indicator j,
which is denoted by qij such that qij ¼ 1 if person i is deprived in indicator j and
qij ¼ 0 otherwise.
8
Extracted from Seth and Villar (2016)
128 7 Poverty Measurement
Table 7.2 Dimensions, indicators, deprivation cut-offs and weights of the MPI
Dimension Indicator A person in a household is deprived if . . .
Health Nutrition Any woman or child in the household with nutritional information
is undernourished
Mortality Any child has died in the household
Education Schooling No household member has completed five years of schooling
Attendance Any school-aged child in the household is not attending school up
to class 8
Standard of Electricity The household has no electricity
Living Sanitation The household’s sanitation facility is not improved or it is shared
with other households
Water The household does not have access to safe drinking water or safe
water is more than 30 min walk round trip
Flooring The household has a dirt, sand or dung floor
material
Cooking The household cooks with dung, wood or charcoal
fuel
Assets The household does not own more than one of: radio, telephone,
TV, bike, motorbike or refrigerator and does not own a car or
truck
Source: Alkire, Roche, Santos, and Seth (2011)
X
K
ci ¼ bj qij
j¼1
The deprivation score of each person is the weighted average of deprivation status
scores, with ci ¼ 0 if person i is not deprived in any indicator and ci ¼ 1 if person i is
deprived in all indicators, and so ci 2 ½0; 1.
Not all those who are deprived in some indicator are identified as poor though.
The identification step involves a poverty cut-off p. A person is identified as poor
whenever ci p and non-poor whenever ci < p. If the value of p is positive, but
lower than the minimum weight assigned to any indicator such that
0 < p < minfb1 ; . . . ; bK g, then the identification approach is referred as the
union approach. By union approach, a person is identified as poor, even when the
person is deprived in a single indicator. On the other extreme, an intersection
approach identifies a person as poor only if the person is deprived in all indicators
or when p ¼ 1. Both these approaches may be too stringent and in that case an
alternative middle ground may be found by using an intermediate approach, such
that minfb1 ; . . . ; bK g < p < 1. Once, individuals are identified as poor and
non-poor, then a censored distribution of deprivation scores are obtained, such
that ci ðpÞ ¼ ci if ci p and ci ðpÞ ¼ 0 for all ci < p. The adjusted headcount ratio,
denoted by M0, is computed from the censored distribution scores as:
130 7 Poverty Measurement
Xn
c ð pÞ
i¼1 i
M0 ¼
n
The MPI uses a particular set of indicators and deprivation cut-offs, a particular
set of weights and a certain value of poverty cut-off. The three dimensions, ten
indicators and the corresponding deprivation cut-offs are already outlined in
Table 7.2. The MPI weights each dimension equally and furthermore weight within
each dimension is equally distributed across indicators. For example, the mortality
indicator in health dimension is assigned a weight equal to 1/6; whereas the assets
indicator in the standard of living dimension is assigned a weight equal to 1/18. The
poverty cut-off for the MPI is equal to one-third of weighted indicators or p ¼ 1/3.
Thus, a person within a household is identified as poor if the household’s depriva-
tion score is equal to or larger than 1/3.
Note that the identification takes place at the household level, but not at the
individual level because it is difficult to obtain data at the individual level. Because
the identification takes place at the household level, it is not possible to capture the
difference in achievements that may exist within a household. Despite this short-
coming, the construction of the MPI is a big leap forward in the measurement of
poverty.
The MPI also has certain useful properties. First, it can be expressed as a product
of two terms. One is the multidimensional headcount ratio (H ), which is the
proportion of the population living in households that are deprived in one-third of
weighted indicators or with deprivation scores equal or larger than one-third. If we
denote the number of poor by q, then H ¼ q=n. The other is the average deprivation
scores among the poor (A). By definition, H lies between zero and one: it is equal to
one when everybody is identified as poor and is equal to zero when there is no poor
at all. The range of A is, however, not straightforward. Whenever there is at least
one poor, A lies between p and one, but when there is no poor in the society, then
A cannot be defined. The second useful feature is that the MPI can be expressed as a
weighted average of censored headcount ratios of the ten indicators. The censored
headcount ratio of an indicator is the proportion of the population who is identified
as multidimensionally poor and is simultaneously deprived in that indicator. The
third useful property is that it is decomposable across any population subgroup,
which means that the overall MPI can be expressed as a weighted average of
subgroup MPIs where the weight attached to each subgroup is equal to its relative
population share.
Table 7.3 provides the latest data on this multidimensional poverty index.
Let us conclude this section by pointing out a couple of critiques to this approach
to poverty measurement. First, note that the MPI is not sensitive to inequality across
the poor. There are various alternative poverty measures that use binary indicators
as the MPI does, but are sensitive to inequality across the poor (see Bossert,
Chakravarty, & D’Ambrosio, 2009; Jayaraj & Subramanian, 2009; Rippin, 2011).
However, there is a trade-off: these inequality-sensitive poverty indices do not
allow the overall indices to comprehend the contribution of each indicator or
7.4 Deprivation and Non-monetary Poverty Measures 131
dimension to the overall poverty, which is crucial for policy analysis. Which of
these two properties is more important and whether it is possible to find a way out
without sacrificing any of these two properties is a subject for further research.
Second, there is the need of developing a test of the robustness of ranking and
country comparisons with respect to the choice of parameters in MPI’s construc-
tion. Although Alkire and Santos (2010) test the robustness of country rankings
with respect to a few alternative weighting schemes, a range of poverty cut-offs and
a few different alternative set of deprivation cut-offs, a more sound and concrete
approach is required.
References
Aaberge, R., & Brandolini, A. (2014). Multidimensional poverty and inequality (Banca d’Italia
Working Paper No. 976).
Alkire, S., & Foster, J. E. (2007). Counting and multidimensional poverty measures (OPHI
Working Paper 7). Oxford: Oxford Poverty and Human Development Initiative, University
of Oxford.
Alkire, S., & Foster, J. E. (2011). Counting and multidimensional poverty measurement. Journal
of Public Economics, 95, 476–487.
Alkire, S., Foster, J., Seth, S., Santos, M. E., Roche, J. M., & Ballon, P. (2014). Multidimensional
poverty measurement and analysis. Chap. 1. Introduction (OPHI Working Paper No. 82).
Alkire, S., Foster, J., Seth, S., Santos, M. E., Roche, J. M., & Ballon, P. (2015). Multidimensional
poverty measurement and analysis (Chap. 2). The Framework. (OPHI Working Paper No. 83).
Alkire, S., Roche, J. M., Santos, M. E., & Seth, S. (2011). Multidimensional Poverty Index 2011:
Brief Methodological Note. Oxford: Oxford Poverty and Human Development Initiative,
Oxford University.
Alkire, S., & Santos, M. E. (2010). Acute multidimensional poverty: A new index for developing
countries (Working Paper 38). Oxford: Oxford University, Oxford Poverty and Human Devel-
opment Initiative.
Atkinson, A. B. (1987). On the measurement of poverty. Econométrica, 55, 749–764.
Blackorby, C., & Donalson, D. (1980). Ethical indices for the measurement of poverty.
Econometrica, 48, 1053–1060.
134 7 Poverty Measurement
Bossert, W., Chakravarty, S. R., & D’Ambrosio, C. (2009). Multidimensional poverty and material
deprivation with discrete data. Review of Income and Wealth, 59, 29–43.
Bourguignon, F., & Chakravarty, S. R. (2003). The measurement of multidimensional poverty.
Journal of Economic Inequality, 1, 25–49.
Chakravarty, S. R. (1983). A new index of poverty. Mathematical Social Sciences, 6(3), 307–313.
Chakravarty, S. R. (1990). Social ethical index numbers. New York: Springer.
Chakravarty, S. (2009). Inequality, polarization and poverty. Berlin: Springer.
Callan, T., Nolan, B., & Whelan, C. T. (1993). Resources, deprivation and the measurement of
poverty. Journal of Social Policy, 22(2), 141–172.
Clark, S., Hemming, R., & Ulph, D. (1981). On indices for the measurement of poverty. The
Economic Journal, 91, 515–526.
Dardadoni, V. (1995). On multidimensional poverty measurement. Research on Economic
Inequality, 6, 201–207.
Foster, J., Greer, J., & Thorbecke, E. (1984). A class of decomposable poverty measures.
Econometrica, 52, 761–766.
Foster, J., & Shorrocks, A. (1991). Subgroup consistent poverty indices. Econometrica, 59,
687–709.
Jayaraj, D., & Subramanian, S. (2009). A Chakravarty-D’Ambrosio view of multidimensional
deprivation: Some estimates for India. Economic and Political Weekly, 45, 53–65.
Jenkins, S. P., & Lambert, P. J. (1997). Three “I’s of Poverty” curves, with an analysis of
U.K. poverty trends. Oxford Economics Papers, 49, 317–327.
Mack, J., & Lansley, S. (1985). Poor britain. London: Allen & Unwin.
OECD. (2014). PISA 2012 results: What students know and can do: Student Performance in
mathematics, reading and science (Vol. I). Paris: OECD Publishing.
Rippin, N. (2011). A response to the weaknesses of the multidimensional poverty index (MPI): The
correlation sensitive poverty index (CSPI). German Development Institute.
Sen, A. (1976). Poverty: An ordinal approach to measurement. Econometrica, 44, 219–231.
Seth, S., & Villar, A. (2016, forthcoming). The measurement of human development and poverty.
In: C. D’Ambrosio (Ed.), Handbook of research on economic and social wellbeing. Edward
Elgar.
Stiglitz, J., Sen, A., & Fitoussi, J. P. (2009). The measurement of economic performance and social
progress revisited. Reflections and overview. Commission on the measurement of economic
performance and social progress, Paris.
Townsend, P. (1979). Poverty in the United Kingdom: A survey of household resources and
standards of living. Berkeley: University of California Press.
Tsui, K. Y. (2002). Multidimensional poverty indices. Social Choice and Welfare, 19(1), 69–93.
United Nations Development Programme, & Malik, K. (2014). Human development report 2014:
Sustaining human progress-reducing vulnerabilities and building resilience (PDF). UN.
Watts, H. (1968). An economic definition of poverty. In D. P. Moynihan (Ed.), On understanding
poverty: Perspective from social science. Nueva York: Basic Books.
Chapter 8
Multidimensional Poverty and Welfare
8.1 Introduction1
1
This chapter is based on Villar (2015, 2016).
2
http://www.ophi.org.uk
Let us start by assuming that poverty may involve several dimensions (see Aaberge
& Brandolini, 2014; Alkire et al., 2015; Bourguignon & Chakravarty, 2003;
Dardadoni, 1995; Tsui, 2002). To be precise, we shall consider here the evaluation
of poverty in a society N consisting of n agents with respect to K different
dimensions, all of which can be measured in terms of quantitative variables. We
8.2 The Setting 137
assume from the outset, when comparing different societies in terms of our poverty
measure, that the set of dimensions involved is exactly the same for all societies.
We denote by Y the n K matrix that describes the achievements of the n agents
with respect to the K welfare dimensions and by a boldfaced z the K-vector of
poverty thresholds, one threshold for each dimension. All entries in Y and z are
assumed to be strictly positive. A poverty index is a function that applies the space
of pairs (Y, z), for any (finite) dimension, into the real numbers. Following the
normative approach, we shall deduce our poverty indices from a social evaluation
function, W, defined on the space of achievements. The poverty measure so
obtained will be called a welfare poverty index.
Let Ω stand for the family of all finitely dimensional positive matrices, Y. We
adopt the convention of describing each agent’s achievements as a row and the
distribution of the variable that captures each dimension as a column. In particular,
agent i’s achievements will be denoted by the (row) K-vector yi whereas the
distribution of dimension j by the (column) n-vector y( j).
Consider now the following:
Definition 8.1 A social evaluation function is a mapping W : Ω ! ℝþ , homo-
geneous of degree one (i.e. 8λ > 0, W ðλYÞ ¼ λW ðYÞ).
That is, in our formulation, a social evaluation function is a homogeneous
function W(.) that associates a non-negative real number to any n K positive
matrix Y, for finite values of n and K. Homogeneity is a cardinal property that
ensures a consistent connection between welfare, inequality and poverty measure-
ment (Blackorby & Donaldson, 1978); it also implies a weak form of monotonicity
as higher achievements are better.3
Armed with this tool, we can define those individuals who are welfare-poor as
follows:
Definition 8.2 An individual i 2 N with achievements yi 2 ℝþþ K
is welfare-poor,
relative to a social evaluation function W and a vector of thresholds z 2 ℝþþ
K
, if and
only if W yi < W ðzÞ.
A welfare-poor individual is thus an individual whose personal welfare is below
the minimum associated to the thresholds. Note that the agent’s welfare is measured
by our social evaluation function applied to a society consisting of this individual
alone.
We shall now define our poverty index as the relative welfare loss of the poor,
measured by the social evaluation function restricted to the set of the poor. Let
p denote the number of the poor (i.e. the number of agents i 2 N for which W yi
< W ðzÞ), and let 1m be the unit vector of dimension m. The very definition of the
vector of poverty thresholds implies that W(1pz1, 1pz2, . . ., 1pzK) is the minimum
3
We require homogeneity rather than homoceticity to be able to define the welfare poverty index
appropriately.
138 8 Multidimensional Poverty and Welfare
welfare that society would like to ensure for the poor individuals. Yet the actual
welfare of the poor is given by W(Yp), where Yp is a p K matrix that describes the
achievements
of the poor. The difference between those two values,
W 1p z1 , 1p z2 , . . . , 1p zK W ðYp Þ, tells us how far away is this society from
ensuring the minimum admissible welfare to all its members (i.e. the absolute
welfare loss due to poverty). We shall identify our multidimensional welfare
poverty index, PW(Y, z), with the ratio between that difference and the minimum
welfare admissible for all society, W(1nz1, 1nz2, . . ., 1nzK) (i.e. the relative welfare
loss due to poverty). Formally:
Definition 8.3 A welfare poverty index, relative to a society N with n members,
regarding K welfare dimensions, is a mapping PW : ℝnKþþ ℝþþ ! ℝþ given by:
K
W 1p z1 , 1p z2 , . . . , 1p zK W ðYp Þ
PW ðY, zÞ ¼ ð8:1Þ
W ð1n z1 , 1n z2 , . . . , 1n zK Þ
Remark 8.1 The definition of the individual poverty measure does not involve any
subgroup decomposability axiom (see Chakravarty, 2009, 2.2 and the discussion in
Chap. 7). It is just the application of the definition to a society consisting of a single
individual.
It is interesting to point out that, within this framework, the determination of who
are the poor is resolved in a simple and natural way. Let us recall here that counting
the poor in a multidimensional context is not immediate because the poverty
threshold is a vector with K > 1 components, and we may find that the achievements
of some agents exceed the threshold levels in some dimensions and fall short in
some others. There are two extreme positions in the poverty literature regarding this
problem: the union approach, that declares poor anyone who is below the reference
value in some dimension, and the intersection approach, according to which one
person is poor only if all her achievements are simultaneously below the reference
values.
8.3 Adding Structure 139
The first property we consider is quasi-concavity. This can be regarded as one of the
most basic value judgments in welfare evaluations, as it is an expression of concern
for equality. It implies that redistributions of achievements enhance social welfare.
Formally:
Property 1
The social evaluation function W satisfies quasi-concavity. That is,
8λ 2 ½0; 1, Y, Y0 2 ℝnK 0
þþ , W ðλY þ ð1 λÞY Þ minfW ðYÞ, W ðY Þg.
0
The second property, scale, serves the purpose of defining the units in which
welfare is measured. It is a convention that establishes that in a society consisting of
n individuals with identical achievements in all dimensions, the associate welfare
equals n times that common value. Formally,4
Property 2
The 1 function W satisfies scale. That is, for any positive scalar a,
0 social evaluation
W @1n a, . . . , 1n a A ¼ na.
|fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl}
K times
One of the standard assumptions introduced in the analysis of multidimensional
welfare, inequality and poverty, in order to simplify the discussion and get sensible
formulae, is that of factor decomposability (Chakravarty, Mukherjee, & Ranade,
1998 and the discussion in Chaps. 6 and 7). It can be defined as follows:
Property 3
The social evaluation function W satisfies factor decomposability. That is,
XK
W ðYÞ ¼ j¼1
bj W ðyðjÞÞ
where W(y( j)) is the social evaluation function relative to the jth dimension
XK
considered in isolation, and b ¼ 1.
j¼1 j
As it is clear from the definition, factor decomposability is a convenient assump-
tion that extraordinarily simplifies the cross effects of the different dimensions on
the overall welfare measure. In particular, it implies a constant rate of substitution
4
Note that this notion of scale is slightly different from that in Chap. 6. This is so because we want
to keep track of the size of the total population and of the population of the poor.
140 8 Multidimensional Poverty and Welfare
between those dimensions, no matter the values of the variables, given by the ratios
of the corresponding weights bj. Those weights represent the relative importance
attached to the different dimensions in the social evaluation function.
Note that the properties of decomposability and scale imply that:
XK XK
W ðzÞ ¼ j¼1
bj z j , W yi ¼ by ;
j¼1 j ij
so that it follows that agent i’s individual welfare poverty index, PW(yi, z), is given
by:
8 0 XK 19
< b j y ij
=
yi ; z ¼ max 0; @1 X K A
j¼1
PWi
: bz ; j¼1 j j
Consequently, the number of the poor, p, corresponds to those agents for which
the ratio of the weighted averages of achievements and thresholds is smaller than
1. There is, therefore, compensation between dimensions in terms of constant rates
of substitution.
Also observe that the combination of scale and decomposability implies:
XK XK
W 1p z 1 , 1p z 2 , . . . , 1p z K ¼ b W 1p zj ¼ p j¼1 bj zj
j¼1 j
5
One should really write ye(y( j)) to be precise. Yet we shall use a less cumbersome notation.
8.3 Adding Structure 141
0 XK eðpÞ
1
p@ by
j¼1 j j
PðY, zÞ ¼ 1 XK A ð8:3Þ
n bj zj j¼1
eðpÞ
where yj is the EDE of the poor relative to the jth dimension; that is,
eðpÞ
W 1p yj ¼ W ðyp ðjÞÞ.
where W(y( j)) is the social evaluation function relative to the jth dimension
XK
considered in isolation, and β ¼ 1.6
j¼1 j
By replicating the reasoning above, using Property 30 instead of Property 3, we
immediately deduce that the set of the poor is given by those agents i 2 N for which:
Y K βj
j¼1
yij
Y K β < 1
j¼1
zj j
Within this context, there is scope for compensation between dimensions, which is
fully determined by the expression above and involves the familiar decreasing
marginal rates of substitution.
It is also easy to check that the corresponding family of welfare poverty
measures is given by:
6
The geometric mean of the welfare dimensions has been characterized in terms of intuitive and
simple properties (alternative characterizations appear in Foster et al. (2005), Herrero, Martı́nez,
and Villar (2010) or Seth (2013), among others). The geometric mean exhibits better properties as
a welfare indicator, as it does not imply constant rates of substitution between welfare dimensions.
142 8 Multidimensional Poverty and Welfare
0 !βj 1
eðpÞ
p YK yj
PðY; zÞ ¼¼ @1 A ð8:4Þ
n j¼1 zj
Remark 8.2 Factor decomposability and its multiplicative version are both strong
assumptions, which help getting practical evaluation formulae. Those properties
can be regarded as particular instances of the standard separability principle that
appears in utility theory (e.g. Barten and B€
ohm, 1982, #6.2; see also the discussion
in Chap. 7).
We can think of several venues to arrive at a closed formula, while keeping the
properties discussed above. We shall consider here two of them. The first venue
consists of exploiting the properties of the social evaluation function in terms of the
equally distributed equivalent values of the variables, yej , j ¼ 1, 2, . . ., K. This
provides ways of selecting W in terms of inequality measures. The second venue
goes through the properties of decomposability by population subgroups.
Consider now the following property, anonymity, which establishes that the welfare
evaluation depends only on the agents’ achievements but not on other idiosyncratic
features. Formally:
Property 4
The social evaluation function W satisfies anonymity. That is, for all Y, Y0 2 ℝnK þþ
such that Y0 is a reshuffling of the rows of Y, we have: W ðYÞ ¼ W ðY0 Þ.
The quasi-concavity of the social evaluation function implies that yje μðyðjÞÞ,
where μ(y( j)) is the mean value of the jth variable. So the difference between those
two magnitudes can be regarded as a measure of the impact of inequality on the
welfare evaluation. Indeed, under anonymity one can write:
yje ¼ μðyðjÞÞ½1 I W ðyðjÞÞ, where IW(y( j)) is some relative inequality index, pro-
vided I W ðyðjÞÞ 2 ½0; 1.
Assuming quasi-concavity, scale, decomposability and anonymity, therefore,
the welfare poverty index appears as the product of the incidence and the intensity
of poverty, once intensity has been adjusted by inequality. That is, the share of poor
people times how poor and unequal they are. This is an intuitive formula, derived
from a standard normative approach, easy to interpret, which integrates nicely and
rather explicitly the three key aspects of poverty measurement. Then, choosing a
8.4 Closing the Formula 143
particular inequality index closes the formula depending on the type of decompos-
ability we adopt (Property 3 or Property 30 ).
As an example, take the Gini index and assume factor decomposability. Then we
obtain a multidimensional version of the standard Sen’s (1976) poverty measure
given by:
0 XK 1
p p
p b μ
j¼1 j j
1 Gj
PG ðY; zÞ ¼ @1 XK A ð8:5aÞ
n b z
j¼1 j j
where μpj , Gpj are shorthand versions of the mean of the poor in the jth dimension
and the corresponding Gini index, μ(yp( j)) , G(yp( j)), respectively.
If we take Atkinson’s inequality index for ε ¼ 1 and use Property 30 instead, we
obtain:
0 !β j 1
p@ YK ejp
μ
PAð1Þ ðY; zÞ ¼¼ 1 j¼1 A ð8:5bÞ
n zj
where μ ejp is the geometric mean of the achievements of the poor regarding
dimension j.
Trivially, when all dimensions are equally important those expressions become:
0 X K p p
1 0 !1=K 1
p@ μ j 1 G j p YK ejp
μ
PG ðY; zÞ ¼ 1
j¼1
XK A, PAð1Þ ðY; zÞ ¼¼ @1 A
n zj n j¼1 zj
j¼1
respectively.
Remark 8.3 The social evaluation function associated with Atkinson’s inequality
index for ε¼1 with equal weights, corresponds to the Equality Adjusted Human
Developed Index, as introduced in the 2010 edition of the Human Development
Report (United Nations Development Program, 2010), so that the welfare poverty
measure PA(1)(.) can be regarded as the associated counterpart, when measuring
poverty in a multidimensional context with quantitative variables.
society can be expressed as the weighted average of the poverty of those population
subgroups, with weights corresponding to the population shares. Formally:
Property 5
The social evaluation function W satisfies subgroup decomposability. That is,
XG
W ðYÞ ¼ g¼1
W ðYg Þ.
This property implies that all the social evaluation functions do not care about
the distribution of the welfare between population subgroups but just about the
aggregate value. By applying Property 5 to those subgroups consisting of a single
individual, we get a utilitarian social evaluation function:
Xn
W ðYÞ ¼ i¼1
W yi
which shows that this property implies anonymity and disregards inequality among
individuals.
Combining factor and subgroup decomposability we get:
Xn XK
W ðYÞ ¼ i¼1 j¼1
bj yij ð8:6Þ
Y K Y n 1=n βj
W ðYÞ ¼ j¼1 j¼1
yij ð8:8Þ
and
0 0Y 1=p 11=K 1
p
pB YK B yij C C
PðY; zÞ ¼ B A C
i¼1
@ 1 j¼1 @ A ð8:9bÞ
n zj
These are intuitive and easy to handle formula, based on clear-cut assumptions,
which correspond to the poverty counterpart of the standard Atkison’s social
evaluation function for the unit value of the inequality aversion parameter.7
The social evaluation function in this context for each OECD country is a mapping
þ ! ℝ that associates to each positive n 3 matrix, Y, a real number, W(Y).
W : ℝn3
Each row of matrix Y describes the achievements of an individual of that country in
the three educational dimensions evaluated, denoted by m, r, s (mathematics,
7
This index can also be regarded as a derivation of Watts (1968) poverty measure, under the
assumption of equally important dimensions.
8
This section is based on Villar (2016).
146 8 Multidimensional Poverty and Welfare
reading comprehension and sciences). And each column gives us the distribution of
each dimension across the whole society.
We shall assume that our social evaluation function corresponds to that linked to
the Atkinson’s inequality index for ε ¼ 1. That is,
W ðYÞ ¼ n½μ
eðyðmÞÞ μ eðyðsÞÞ1=3
e ð yð r Þ Þ μ ð8:10Þ
where the first term, p/n, measures the incidence of poverty (the share of the poor in
society) and the second term is given by:
μeðyðmÞÞ μe ð yð r Þ Þ μeðyðsÞÞ 1=3
ρðY; zÞ ¼ 1 ð8:12Þ
zm zr zs
individual micro-data as our starting point, focusing on the test scores on mathe-
matics, reading and science. Out of these data we are able to compute, for each
student i in the sample of every OECD country, the corresponding individual
poverty index. That is, the number:
y
yir yis 1=3
max 0; 1 im
ð8:13Þ
420:1 407:5 409:5
We present now the results that are obtained when analysing educational poverty in
the OECD using the methodology described above.
The main results are reported in Table 8.1. The table contains both the value of
the Educational Poverty Index (EPI) in absolute terms (i.e. the computation of
equation (8.11)) and the value of the index in relative terms (as percentages of the
OECD mean). Using relative values is helpful because those numbers are easier to
interpret since the EPI values are very small, as they correspond to the product of
two numbers smaller than one. Figure 8.1 illustrates those values ordering the
countries from best to worse.
Table 8.1 also provides information about the rank of the different countries
regarding achievements (understood as the geometric mean of the test scores,
ordered from more to less) and the poverty index (arranged in an increasing
order). The comparison of those rankings shows that poverty analysis provides
some information about the performance of the educational systems that is badly
captured by the average scores. Even though the rank correlation is high, there are
substantial differences in particular cases such as Belgium, Denmark, France,
Luxembourg, New Zealand, Slovenia, Spain, Turkey, United Kingdom and the
United States.
A prominent feature of educational poverty, as clearly illustrated in Fig. 8.1, is
the large variability that exhibits among OECD countries: while the coefficient of
variation of the test scores is very low, around 0.054, the coefficient of variation of
educational poverty is ten times larger (0.537). The data exhibit a similar structure
for the whole set of the countries participating in PISA, even though substantially
amplified (the coefficient of variation of the test scores is 0.1 while that of
educational poverty index jumps up to 0.9).
148 8 Multidimensional Poverty and Welfare
Table 8.1 OECD Educational Poverty Index (EPI), relative EPI (OECD mean ¼ 100) and
ranking of countries by EPI and mean test scores
EPI Relative EPI Ranking test scores Ranking EPI
Australia 0.0177 75 11 13
Austria 0.0195 83 15 15
Belgium 0.0245 104 12 21
Canada 0.0113 48 5 6
Chile 0.0470 199 33 32
Czech Republic 0.0180 76 16 14
Denmark 0.0172 73 19 11
Estonia 0.0051 22 4 1
Finland 0.0104 44 3 4
France 0.0284 120 17 25
Germany 0.0162 69 10 10
Greece 0.0363 154 31 30
Hungary 0.0253 107 26 23
Iceland 0.0301 127 27 27
Ireland 0.0120 51 9 7
Israel 0.0487 206 29 33
Italy 0.0255 108 24 24
Japan 0.0107 45 2 5
Korea 0.0081 34 1 2
Luxembourg 0.0302 128 22 28
Mexico 0.0621 263 34 34
The Netherlands 0.0151 64 7 9
New Zealand 0.0219 93 13 18
Norway 0.0245 104 20 20
Poland 0.0091 38 6 3
Portugal 0.0250 106 25 22
Slovak Republic 0.0439 186 30 31
Slovenia 0.0176 75 18 12
Spain 0.0217 92 23 17
Sweden 0.0339 144 28 29
Switzerland 0.0131 56 8 8
Turkey 0.0291 123 32 26
United Kingdom 0.0226 96 14 19
United States 0.0210 89 21 16
OECD (2014)
Fig. 8.1 Educational poverty in the OECD according to PISA 2012 (OECD mean ¼ 100)
established that the correlation between the ESCS index and average test scores is
not very high (about one-third of common variance).9
Remark 8.4 It is worth noting that those data underestimate educational poverty
in some countries, as all the information refers to the students who actually keep
attending formal education. There are some countries in which the rate of 15-year
olds who have abandoned school is very high (in particular Mexico and Turkey),
which implies that educational poverty would be substantially higher. For a
discussion on how to combine data regarding access and achievement,
seeCarvalho, Gamboa, and Waltenberg (2015), Ferreira and Gignoux (2011),
Ferreira, Gignoux, and Aran (2011), Gamboa and Waltenberg (2012) and Tansel
(2015).
Equation (8.11) describes the Educational Poverty Index as the product of two
terms. The first one, p/n, captures the incidence of educational poverty. The second
one, ρ(Y, z), is a measure of the intensity of educational poverty adjusted by
inequality. The correlation between both components is positive but moderate
(a coefficient of 0.4), which indicates that OECD countries exhibit rather different
mixes of both ingredients. The variability of those two components is also quite
diverse. The coefficient of variation of the incidence is three times that of the
inequality-adjusted intensity (0.476 versus 0.153).
9
Note, however, that this correlation refers to the link between low performance and socio-
economic conditions between countries. Things are different when we analyse low performance
within countries by social groups.
150 8 Multidimensional Poverty and Welfare
Table 8.2 Incidence and inequality-adjusted Intensity of educational poverty in the OECD (PISA
2012)
Country p/n ρ(Y,z) Relative p/n Relative ρ(Y,z)
Australia 0.14 0.12 78 98
Austria 0.17 0.11 93 91
Belgium 0.17 0.15 92 115
Canada 0.10 0.11 55 88
Chile 0.40 0.12 218 94
Czech Republic 0.15 0.12 84 93
Denmark 0.15 0.12 80 93
Estonia 0.06 0.08 35 63
Finland 0.09 0.12 49 92
France 0.19 0.15 103 119
Germany 0.14 0.12 77 92
Greece 0.26 0.14 140 113
Hungary 0.21 0.12 113 97
Iceland 0.20 0.15 111 117
Ireland 0.11 0.11 59 88
Israel 0.28 0.17 153 138
Italy 0.20 0.13 107 104
Japan 0.09 0.12 48 98
Korea 0.07 0.12 37 94
Luxembourg 0.22 0.14 120 109
Mexico 0.48 0.13 261 103
The Netherlands 0.13 0.11 73 90
New Zealand 0.17 0.13 92 103
Norway 0.18 0.14 96 111
Poland 0.09 0.10 52 76
Portugal 0.20 0.13 107 101
Slovak Republic 0.27 0.16 148 129
Slovenia 0.16 0.11 89 86
Spain 0.17 0.13 94 100
Sweden 0.22 0.15 122 120
Switzerland 0.12 0.11 63 91
Turkey 0.28 0.10 155 81
United Kingdom 0.16 0.14 88 111
United States 0.19 0.11 104 88
Relative values take the OECD mean equal to 100
Table 8.2 provides the data regarding the values of the two components of the
Educational Poverty Index, both in absolute and relative terms. Mexico, Chile,
Turkey, Israel and Greece exhibit relative values of the incidence variable more
than 40 % higher than the OECD mean, whereas Korea, Japan, Finland, Poland and
Ireland are at least 40 % below the incidence average (40 is about one standard
References 151
Educational poverty has been identified here as the product of two different factors:
the incidence of educational poverty, given by the share of students who do not
reach level 2 of proficiency, and the inequality-adjusted intensity, which measures
how unequal and far away they are from the threshold defining level 2.
The Educational Poverty Index is to be interpreted as quantitative assessment of
the welfare loss due to the failure of educational systems to provide a minimal
knowledge to all citizens. This is so because the basic data (the PISA scores) are
collected at the end of the period of compulsory education.
The data show that the OECD countries are much more diverse regarding
educational poverty than with respect to the average scores. In all cases, we find
that most of the observed differences in the Educational Poverty Index are due to
the differences in the incidence of educational poverty, whereas the impact of the
equality-adjusted intensity is much smaller (with a positive but not too high
correlation between those variables).
The high negative correlation between educational poverty and average scores
tells us that reducing low performance is the most effective way of improving the
overall educational outcomes. In other words, making the educational system more
inclusive is the best strategy to get simultaneously higher equity and higher
efficiency.
Finally, let us mention that the analysis of the possible causes and remedies of
low performance, a topic outside the scope of this work, has been elaborated by the
PISA team in OECD (2016).
References
Aaberge, R., & Brandolini, A. (2014). Multidimensional poverty and inequality (Banca d’Italia
working paper n 976).
Alkire, S., Foster, J., Seth, S., Santos, M. E., Roche, J. M., & Ballon, P. (2015). Multidimensional
poverty: Measurement and analysis. Oxford: Oxford University Press.
Barten, A. P., & B€ohm, W. (1982). Consumer theory. In K. J. Arrow & M. D. Intriligator (Eds.),
Handbook of mathematical economics (Vol. II). New York: North Holland.
152 8 Multidimensional Poverty and Welfare
Blackorby, C., & Donaldson, D. (1978). Measures of relative inequality and their meaning in terms of
social welfare. Journal of Economic Theory, 18, 59–80.
Blackorby, C., & Donaldson, D. (1980). Ethical indices for the measurement of poverty.
Econometrica, 48, 1053–1060.
Bourguignon, F., & Chakravarty, S. R. (2003). The measurement of multidimensional poverty.
Journal of Economic Inequality, 1, 25–49.
Carvalho, M., Gamboa, L. F., & Waltenberg, F. D. (2015). Equality of educational opportunity:
Taking both achievement and access into account (Ecineq working paper 2012-277)
Chakravarty, S. R. (2009). Inequality, polarization and poverty. New York: Springer.
Chakravarty, S. R., Mukherjee, D., & Ranade, R. (1998). The family of subgroup and factor decompo-
sable measures of multidimensional poverty. Research on Economic Inequality, 8.
Clark, S., Hemming, R., & Ulph, D. (1981). On indices for the measurement of poverty.
The Economic Journal, 91, 515–526.
Dardadoni, V. (1995). On multidimensional poverty measurement. Research on Economic Inequal-
ity, 6, 201–207.
Duclos, J.-Y., & Araar, A. (2006). Poverty and equity: Measurement, policy and estimation with
DAD. New York: Springer.
Ferreira, F., & Gignoux, J. (2011). The measurement of educational inequality: Achievement and
opportunity (Ecineq working paper n 240).
Ferreira, F. H. G., Gignoux, J., & Aran, M. (2011). Measuring inequality of opportunity with
imperfect data: The case of Turkey. Journal of Economic Inequality, 9, 651–680.
Foster, J. E., Lopez-Calva, L. F., & Szekely, M. (2005). Measuring the distribution of
human development: Methodology and an application to Mexico. Journal of Human Devel-
opment, 6(1), 5–25.
Gamboa, L. F., & Waltenberg, F. D. (2012). Inequality of opportunity in educational achievement
in Latin America: Evidence from PISA 2006–2009. Economics of Education Review, 31,
694–708.
Haughton, J., & Khandker, S. R. (2009). Handbook of poverty and inequality. Washington, DC:
The World Bank.
Herrero, C., Martı́nez, R., & Villar, A. (2010). Multidimensional social evaluation. An application to
the measurement of human development. Review of Income and Wealth, 56, 483–497.
Kakwani, N. (1997). Inequality, welfare and poverty: Three interrelated phenomena (Working paper
97/18). Kensington, NSW: The University of New South Wales.
Lewis, G. W., & Ulph, D. T. (1988). Poverty, inequality and welfare. The Economic Journal, 98,
117–131.
OECD. (2014). What students know and can do. Student performance in mathematics, reading and
science (Revised edition, Vol. 1). Paris: PISA, OECD Publishing.
OECD. (2016). Low performing students: Why they fall behind and how to help them succeed.
Paris: PISA, OECD Publishing. 10.1787/9789264250246-en.
Pyatt, G. (1987). Measuring welfare, poverty and inequality. The Economic Journal, 97, 459–467.
Sen, A. K. (1976). Poverty: An ordinal approach to measurement. Econometrica, 44, 219–231.
Seth, S. (2013). A class of distribution and association sensitive multidimensional welfare indices.
Journal of Economic Inequality, 11, 133–162.
Tansel, A. (2015). Inequality of opportunities in educational achievements in Turkey over time
(IZA working paper DP 9005).
Tsui, K. (2002). Multidimensional poverty indices. Social Choice and Welfare, 19, 69–93.
United Nations Development Program. (2010). The real wealth of Nations: Pathways to
human development. New York: UNDP.
Vaughan, R. N. (1987). Welfare approaches to the measurement of poverty. The Economic Journal,
97, 160–170.
Villar, A. (2015). Multidimensional welfare-poverty indices (mimeo).
References 153
Villar, A. (2016). Educational poverty as a welfare loss: Low performance in the OECD according
to PISA 2012. Modern Economy, 7, 441–449. doi:10.4236/me.2016.74049.
Wagle, U. (2008). Multidimensional poverty measurement. New York: Springer.
Watts, H. (1968). An economic definition of poverty. In D. P. Moynihan (Ed.), On understanding
poverty (pp. 316–329). New York: Basic Books.
Chapter 9
The Evaluation of Relative Achievements
9.1 Introduction
We shall present here a way of using some of the ideas developed in the field of
poverty measurement to the case of evaluating the achievements of a society with
respect to some given targets. The bottom line is to change the role of poverty
thresholds into multidimensional targets, so that performance can be assessed in
terms of some distance or ratio between achievements and targets.1
Let us motivate this problem by means of a simple example that will be used as a
reference to illustrate different aspects of this model.
Example 9.1 The Hypothetical University
The Hypothetical University consists of three Departments, A, B and C, with equal
size in terms of staff. The strategic plan of the University established a series of
research objectives that are summarised in terms of the following specific yearly
targets: (i) Increase refereed publications by 10 %, (ii) increase the research funds
obtained competitively from external sources by 15 % and (iii) increase by 5 % the
number of grants for graduate students. The data reported by the three Departments
at the end of the year, in terms of annual variations, are shown in Table 9.1.
If we interpret that “meeting the targets” is getting results above the thresholds in
all dimensions, what is known as the intersection approach, then only Department
A exhibits a satisfactory performance. If, on the contrary, we interpret that meeting
the targets is reaching any of them, known as the union approach, then all
Departments’ achievements are satisfactory. Most people would agree that none
of those interpretations of what “meeting the targets” means is really satisfactory,
because they imply evaluating equally very different performances.
The question is, obviously, how to determine the type of compensation between
targets we find admissible. This is actually one of the points that lead to hotter
1
This chapter follows closely the contribution in Villar (2011).
Table 9.1 Yearly increments in the target variables in the Hypothetical University
Publications (%) Research funds (%) Grants (%)
Department A 15 23 8
Department B 26 40 4
Department C 10 0 12
target values simultaneously. On the other hand, there is the other extreme inter-
pretation according to which achieving some target is a sufficient criterion. Each of
those polar views makes the decision on who meets the standards rather trivial. The
drawback is that in both cases we may find very unfair outcomes, because we might
be treating equally highly different performances. The difficult problem is, as
already mentioned, how to handle the intermediate cases. That is, when agents in
society exceed some of the prescribed targets but fail to reach some others
(a relevant case in practice and a usual source of conflicts). The bottom line is
whether we admit or not compensations among achievements, both across dimen-
sions and across agents, and what kind of compensations should be considered
(we shall refer to this feature as the substitutability problem).
The second axiom, neutrality, says that all objectives are equally important (the
principle of symmetry applied to the dimensions).
• Neutrality: Let ðY; zÞ 2 ℝnK ℝþþ K
and let π c(Y), π c(z) denote a permutation of
the columns of matrix Y and of the corresponding dimensions of the vector of
targets z. Then, φðY; zÞ ¼ φðπ c ðYÞ, π c ðzÞÞ.
The third property, normalisation, determines a scale for our evaluation function.
It says that the value of the function is zero when all outcomes are zero and the
value is equal to one when all outcomes match exactly the targets.
• Normalisation: Let 0 denote the matrix all of whose components are equal to
zero, and let Z be the matrix whose columns are all equal to the target vector z.
Then, φð0; zÞ ¼ 0, φðZ; zÞ ¼ 1.
Our last property, additivity, establishes conditions on the behaviour of the
evaluation function when the matrix of the agents’ achievements changes from
Y to Y0 ¼ Y þ ΔY, for some ΔY 2 ℝnK . The property requires the change of the
index to correspond to the evaluation of that change by the very same evaluation
function. This is a very natural property that is most useful when the data on the
agents’ performance is collected from several sources, or across different time
periods, or when there are mistakes to be corrected. The new data can be integrated
by simply computing the value of that change and adding up the result to the
original value of the index. Formally:
• Additivity: Let ðY; zÞ 2 ℝnK ℝþþ
K
and let ΔY 2 ℝnK . Then,
1 Xn X K
yik
φðY; zÞ ¼ ð9:1aÞ
nK i¼1 k¼1 zk
00 1 1
y11 0 ... 0
BB 0 0 ... 0 C C
φB B
@@ . . .
C; zC ¼ φðΔðy11 Þ, zÞ þ φð0; zÞ ¼ φðΔðy11 Þ, zÞ
... ... ...A A
00 0 0 ... 0 1 1
y11 y12 ... 0
BB 0 0 ... 0 C C
φB B
@@ . . .
C; zC ¼ φðΔðy12 Þ, zÞ þ φðΔðy11 Þ, zÞ
... ... ...A A
0 0 ... 0
...
...
n X
X K
φðY; zÞ ¼ φðΔðyik Þ, zÞ
i¼1 k¼1
Let now [1, 1, . . ., 1]a be a matrix with all entries equal to a. Take the special
case in which z ¼ 1s, for a given s, where 1 is the unit vector of the corresponding
dimension and suppose that yik ¼ yth . Then, by symmetry and neutrality, we have:
Let now Y(yik) denote a n K matrix all of whose entries are equal to yik. We
would have:
so that:
φðYðyik Þ, 1sÞ
φðΔðyik Þ, 1sÞ ¼
nK
1 Xn X K
φðY; zÞ ¼ φðYðyik Þ, 1sÞ ð9:2Þ
nK i¼1 k¼1
zk y
f ðyik , zk Þ ¼ f ðzk , zk Þ ¼ 1 ) f ðyik , zk Þ ¼ ik
yik zk
1 Xn X K
yik
φðY; zÞ ¼
nK i¼1 k¼1 zk
(ii) In order to separate those properties, let us consider the following indices:
n X
X K
y
φA ðY; zÞ ¼ ik
i¼1 k¼1
zk
1X n X K
y
φD ðY; zÞ ¼ bk ik
n i¼1 k¼1 zk
X
with b ¼ 1 and bk 6¼ 1=K for some k. It satisfies all properties except
k k
neutrality.
Q.e.d.
This theorem shows that assuming the principles of symmetry, neutrality,
normalisation and additivity amount to choose a precise and very intuitive evalu-
ation function: We evaluate a matrix of outcomes with respect to a vector of targets
as the arithmetic mean of the average of relative achievements across dimensions.
The separability of this evaluation formula makes it very useful to analyse
particular aspects. For instance, let now
9.2 The Basic Model 161
1X n
yik 1X K
yik
μk ðY; zÞ ¼ , μi ðY; zÞ ¼
n i¼1 zj K k¼1 zj
denote the mean value of the relative achievements of society in the kth dimension
and the mean value of agent i’s relative achievements, respectively. We can write
Eq. (9.1a) as follows:
1X K
φðY; zÞ ¼ μk ðYÞ ð9:1bÞ
K k¼1
1X K
φðY; zÞ ¼ μ ðYÞ ð9:1cÞ
n i¼1 i
From this, it follows that agent meets the targets whenever μi ðY; zÞ 1 (i.e. the
agent’s performance is satisfactory when her average relative achievements is
larger than or equal to 1). Similarly, we can say that a given target has been reached
by society, when μj ðY; zÞ 1.
To help fixing ideas, let us apply this model to the example introduced in Sect. 9.1,
relative to the performance of the Hypothetical University with respect to the
targets set in its strategic plan.
Table 9.2 presents the relative outcomes, that is, cell (i, j) corresponds to the
value yij/zj (the ratio between outcomes and targets). In the first cell, for instance, we
find the number 1.5 that corresponds to the ratio between the increase in publica-
tions of Department A (y11 ¼ 15) and the first target (z1 ¼ 10). The last column
provides the evaluation of the Departments whereas the last row provides the
evaluation of each of the three targets. The last cell corresponds to the overall
evaluation of the University.
From those data, we can deduce:
(i) Department B is the one with a better performance, even though it falls short in
one of the targets (increasing grants by 5 %).
(ii) Departments A and B would obtain positive evaluations whereas Department
C would not.
(iii) The University has reached the first two targets, but has failed to reach the
third one.
(iv) Overall, the Hypothetical University has achieved the targets of the strategic
plan, as the global evaluation is 1.03.
9.3 Extensions
The discussion of the example we have just developed is useful to understand how
this evaluation criterion operates and also to consider how to deal with more
complex situations. Upon reflection, one can think of three questions that need to
be answered in order to adopt this evaluation strategy in more general scenarios:
(a) How to cope with units of different size? (e.g. what happens when the Depart-
ments differ in the number of researchers?). Common sense suggests dropping,
or at least adjusting, the property of symmetry because in this context may not
be desirable.
(b) How to deal with targets of different importance? (e.g. what happens if the
University is willing to ponder more publications relative to other objectives?).
In this case, one has to change the neutrality property, which establishes that all
objectives are equally important.
(c) How to change the type of compensation between objectives that the formula
establishes? (i.e. how to allow for variable rates of substitution between
research dimensions?). To get this flexibility, one has to allow for a more
general notion of mean, by substituting the additivity property by some other.
The two first questions have relatively simple answers. We can generalise the
notions of symmetry and neutrality by introducing the idea of weighted symmetry
X
and weighted neutrality. Let us call wi the weight attached to agent i, with w
i i
¼ 1 ; the notion of weighted symmetry says that all that matters is the agents’
outcomes and their corresponding weights, but not other characteristics. When wi
¼ 1=n8i we are back to the symmetry axiom.
The most common case in which agents will enter the evaluation with different
weights is when those agents are groups of different numbers of individuals, as it is
usually the case with university departments. In this context, the weights will
simply correspond to the population shares. More generally, one can think of agents
9.3 Extensions 163
with different needs or merits. A more difficult situation is that in which the
outcomes of the agents may be interdependent, as it is the case in a society with a
network structure, or a hierarchy (think of a research team, for instance). Here, the
different weights could be linked to some measure of “centrality” (e.g. Ballester,
Calvó-Armengol, & Zenou, 2006; Ruhnau, 2000).
A similar approach can be adopted with respect to differences in the importance
of the objectives. Assuming that we have some criterion that permits give different
weights to these Xobjectives, we can introduce them in terms of coefficients bk, k ¼ 1,
2, . . ., K, with b ¼ 1. The neutrality axiom corresponds to the especial case in
k k
which bk ¼ 1=K for all k. These weights can be obtained in some cases by specific
surveys (e.g. Buela-Casal et al. 2010) or from some value judgements on certain
parameters. Different weights also appear naturally in the context of inter-temporal
evaluations, usually to give higher weights to more recent outcomes.
In this more general scenario, we shall say that agent i has met the standards
when the weighted sum of her relative outcomes is larger than or equal to 1. That is,
X
K
yik
bk 1
k¼1
zk
Similarly, we shall say that target k has been reached when the weighted sum of
the agents’ relative outcomes is larger than or equal to 1. That is,
X
n
yik
wi 1
i¼1
zk
X
n X
K
y
φðY; zÞ ¼ wi bk ik ð9:3Þ
i¼1 k¼1
zk
Let us return to the example of the Hypothetical University and see how the
evaluation would change when applying those criteria. Suppose now that Depart-
ment A is made of 30 individuals, Department B is made of 20 individuals and
Department C is made of 50 individuals. That naturally induces a weighting system
for the Departments of 0.3, 0.2 and 0.5, respectively.
Suppose, furthermore, that the University values most the research funds and the
number of grants obtained and relatively less the increment in publications. Let
b1 ¼ 0:2, b2 ¼ b3 ¼ 0:4 be the corresponding coefficients.
The evaluation of the results presented in Table 9.1 are now described in
Tables 9.3 and 9.4.
164 9 The Evaluation of Relative Achievements
So, in this case, the University would have not met its targets (mostly due to the
relevance of the grants, which is the variable that exhibits worst results and now has
become much more important, and the different size of the Departments, being the
larger one that with worst outcomes).
The additive structure of the formula implies that objectives can be substituted at a
constant rate, irrespective of the level of the variable (linear indifference curves of
the social evaluation function). This property may not be natural or convenient in
some contexts, perhaps because one would prefer to have similar results in all
dimensions, rather than very disperse values (for instance, the Hypothetical Uni-
versity might prefer to have three Departments of medium quality rather than one
very good and two very bad). Or it may be willing to have agents (Departments in
the example) with a balanced performance, rather than high outcomes in some
dimension and low in others. A way to obtain this is by substituting the arithmetic
mean by the geometric mean, either with respect to the agents, the dimensions or
both. When the evaluation is willing to promote balancedness both across agents
and dimensions, assuming symmetry and neutrality, we would have:
Y K 1=nK
n Y
y
φðY; zÞ ¼ ik
i¼1 k¼1
zk
9.3 Extensions 165
This is the approach followed by the United Nations when evaluating human
development after the 2010 edition of the Human Development Report (see UNDP
2010). We have already seen this evaluation function in Chap. 8.
Note that both the arithmetic mean and the geometric mean are particular cases
of the generalised means of order α, a parameter that controls the importance of the
dispersion of the variables (see Appendix 2 to Lecture 1 for a discussion). The
general model is analysed in Villar (2011) where the following general formula is
obtained:
8" #1=α
>
>
n X
X K
>
> wi bj ðyik =zk Þ α
α 6¼ 0
< ,
φα ðY; zÞ ¼ i¼1 k¼1 ð9:4Þ
>Y
>n YK
>
> ðyik =zk Þ wi b k
α¼0
: ,
i¼1 k¼1
Remark 9.1 Generalised means introduce flexibility in the evaluation formula. Yet
the simplicity of the arithmetic mean has some relevant advantages. First, it applies
an elementary aggregation principle, easy to understand and operate (this is
something to take into account in many situations where the evaluation criterion
is to be accepted by the evaluated parties). Second, the arithmetic mean allows
dealing with positive and negative components, something that is not possible with
other means.
yi1
1
2 z1 þ yzi22 1. This condition corresponds in Fig. 9.1 to the area above the line
that goes through the points (2, 0) and (0, 2), that is, the area corresponding to the
union of the sets C, D and E. This is clearly an intermediate compensation criterion
between the other two extreme cases. Different weights for agents and dimensions
would simply translate in a change in the units.
The change of the arithmetic mean by the geometric can be captured graphically
in terms of the indifference curves of the corresponding evaluation functions.
Figure 9.2 shows that comparison and makes it clear how the geometric mean
entails a more demanding criterion. The agents who fulfil the objectives are those in
the area G above the curve for the geometric mean, whereas in the case of the
arithmetic mean are those in the area above the straight line.
Let us recall that the arithmetic mean corresponds to a generalised mean of order
1 whereas the geometric mean is a generalised mean of order 0. As the order of the
mean increases, the indifference curves become more and more “open” until they
reach the frontier of the area A in Fig. 9.1 for α ¼ þ1. And they keep closing for
negative values until they become right angles for α ¼ 1 (the frontier of area D in
Fig. 9.1). In other words, by changing α we move between all compensation
possibilities, from full compensation for α ¼ þ1 (the union approach) until no
compensation for α ¼ 1 (the intersection approach).
We shall focus here on the case of a single country, to facilitate the discussion. Let
y1j , y0j stand for the present and past average values of the jth reference variable. We
denote by dyj ¼ y1j y0j the absolute change for the case of goods and by d yj
¼ y0j y1j the absolute change for the case of bads. This formulation translates the
9.4 An Application: The Green Economy Progress Index 169
idea that progress means increments in the goods and reductions in the bads. The
progress experienced in dimension j is given by:
dyj d yj
pGj y1j ; y0j ¼ 0 , pBj y1j ; y0j ¼
yj y0j
Applying now our evaluation procedure (9.1b) to this case, we obtain the
following:
!
1 X yj yj X yj yj
1 0 0 1
GEP ¼ j2G y* y0
þ j2B y0 y*
ð9:5Þ
K j j j j
where G and B represent the set of variables that are goods and bads, respectively.
Besides the targets, that indicated desired values, in the context of the green
economy there is another relevant element to be taken into account: the existence of
thresholds. Those thresholds describe critical values that represent minimal or
maximal levels of the reference variable, depending on whether we consider
goods or bads. Let us see now how the GEP index integrates those values. In a
first instance, the thresholds play a role in the definition of the targets. They will
also appear in the definition of the weights of the different dimensions. Let us see
how it is done.
We have seen that, within this framework, progress is considered as relative to
the targets set for the different variables. The choice of those values yj are,
therefore, relevant modelling choices for the evaluation. A first way of introducing
the thresholds into this formulation is by defining the targets as follows:
n o
y*j ¼ max tj , λj y0j , λj 1
This formulation indicates that, for the case of goods, the desired values corre-
spond to an increase of the initial value that has to reach at least the threshold. For
the case of bads, the target should aim at a reduction of the initial value that should
respect that upper boundary.
The role of thresholds can be given a higher relevance by weighting differently
the variables involved. The formula in (9.5) can be extended as follows:
Now we can choose the weights that ponder how relevant is progress in one
indicator vis-a-vis the others, as the ratios between initial levels of the variables
and the corresponding thresholds. That is, by letting π^ j denote the weight for the jth
dimension, before normalisation, we have:
8t
>
>
j
if j 2 G
< y0 ,
j
π^ j ¼
>
> y0j
: , if j 2 B
tj
where:
X tj X y0j
S¼ j2G y0
þ j2B t
j j
References 171
9.4.3 Results
UNEP (2017) provides a first set of empirical results on the evaluation of progress
in the inclusive green economy, for the time period 2000–2010. To do so, the
targets for each dimension and each country are set by calculating the values λj , βj,
corresponding to goods and bads, as the average values of the 10 % best performers
in each relevant country group (in particular, countries with similar human devel-
opment according to the HDI). In this way, each country has been assigned a target
that is based on an increase in yj that is at least as good as the one achieved by the
10 % best performing countries in its relevant comparison group. Similarly, for
bads, the target of the country is set to achieve a reduction as significant as the
reduction of the 10 % best performing countries in the relevant comparison group.
Thresholds are determined both on the basis of the data and on internationally
recognised scientific sources. For goods (resp. bads), the value of the threshold is set
at the value of the 25th (resp. 75th) percentile of the distribution in 2000. Countries
should never go below (above) the value achieved by the bottom 25 % (top 75 %) of
countries in 2000 for this indicator. For environmental indicators, internationally
recognised scientific sources are used (see UNEP (2017) for the right references).
References
André, F., Cardenete, M. A., & Romero, C. (2010). Designing public policies. Berlin: Springer.
Ballester, C., Calvó-Armengol, A., & Zenou, Y. (2006). Who’s who in networks. Wanted: The key
player. Econometrica, 74, 1403–1417.
Bourguignon, F., & Chakravarty, S. R. (2003). The measurement of multidimensional poverty.
Journal of Economic Inequality, 1, 25–49.
Buela-Casal, G., Bermúdez, M. P., Sierra, J. C., & Quevedo-Blasco, R. (2010). Ranking de 2009
en investigación de las universidades públicas españolas. Psicothema, 22(2), 171–179.
Chakravarty, S. R. (2003). A generalized human development index. Review of Development
Economics, 7, 99–114.
Herrero, C., Martinez, R., & Villar, A. (2010). Multidimensional social evaluation. An application
to the measurement of human development. The Review of Income and Wealth, 56, 483–497.
Kenney, R. L., & Raiffa, H. (1976). Decisions with multiple objectives: Preferences and value
trade-offs. New York: Wiley.
Romero, C. (2001). A note on distributive equity and social efficiency. Journal of Agricultural
Economics, 52, 110–112.
Ruhnau, B. (2000). Eigenvector-centrality: A node centrality? Social Networks, 22, 357–365.
Seth, S. (2010). A class of sensitive multidimensional welfare indices (mimeo). Nashville, TN:
Vanderbilt University.
Tsui, K. (2002). Multidimensional poverty indices. Social Choice and Welfare, 19, 69–93.
United Nations Development Programme. (2010). Human development report 2010: The real
wealth of nations: Pathways to human development. New york: Palgrave Macmillan.
UNEP. (2017). The Green Economy Progress Measurement Framework. Geneva: UNEP.
Villar, A. (2011). Who meets the standards? A multidimensional approach. Modern Economy, 2,
614–624.
Yu, P. L. (1985). Multiple criteria decision making: Concepts, techniques and extensions.
New York: Plenum.