Lectures On Inequality, Poverty and Welfare

Lecture Notes in Economics and Mathematical Systems 685
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
Prof. Dr. W. Trockel

Murat Sertel Institute for Advanced Economic Research
Istanbul Bilgi University
Istanbul, Turkey
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
ISSN 0075-8442 ISSN 2196-9957 (electronic)

Lecture Notes in Economics and Mathematical Systems
ISBN 978-3-319-45561-7 ISBN 978-3-319-45562-4 (eBook)
DOI 10.1007/978-3-319-45562-4
Library of Congress Control Number: 2016959767
© Springer International Publishing AG 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To Flavia
Preface
This monograph originates in a short course I taught at the Economics Department

of the European University Institute (Florence, Italy) in the first semester of 2015.
The course consisted of a series of five 2-hour lessons aimed at graduate students
pursuing their PhD in Economics. Its contents, though, can also be of interest to
graduates of other disciplines with a minimal mathematical training who are
interested in those topics.
The length of the course imposed severe restrictions on the material to be
covered but also constituted a challenge on the selection of topics. I chose to
focus those Lectures on understanding the logics and the construction of the main
indicators that permit one having sensible measures of inequality and poverty, with
special attention to their social welfare underpinnings. No attempt at covering the
whole spectrum of indices or measurement problems is made, and hence, many
complementary topics have been put aside, particularly those regarding the empir-
ical implementation. The reader will find suitable references along the Lectures.
The emphasis is on theory and applicability rather than on specific applications,
which were left as part of the students’ course activities. Nevertheless, I include
here some empirical illustrations that help understanding the possible uses of
inequality and poverty measures.
The first three lectures are largely based on a former monograph, written in
Spanish with F. Goerlich.1 The remaining ones benefit from some former contri-
butions of the author. Each Lecture provides its own list of references, to facilitate
the reader further insights, and tries to be self-contained.
I have finished the writing of these Lectures while I was visiting the OECD in
Paris as a Thomas J. Alexander Fellow. I would like to thank the Fellowship for the
support provided. Thanks are also due to the funding by the Spanish Ministry of
Economics and Innovation, under project ECO2015-65408-R (MINECO/FEDER).
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.
Paris Antonio Villar Ph.D.

June 2016
Contents
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
3.3.1 The First Index of Theil: From Information Theory to

Inequality Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.2 The Second Index of Theil (Mean Logarithmic
Deviation) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3.3 The Generalised Indices of Theil (Tθ) . . . . . . . . . . . . . . . . 44
3.4 Quantile Measures: The Palma Ratio . . . . . . . . . . . . . . . . . . . . . . 48
Appendix: Decomposability of the Generalised Indices of Theil . . . . . . 53
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4 Normative Inequality Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.1.1 The Arrowian Approach to Normative Inequality Indices . . . 57
4.1.2 The Social Evaluation Function Approach . . . . . . . . . . . . 60
4.2 Dalton’s Approach to Inequality Measurement . . . . . . . . . . . . . . . 60
4.3 Atkinson Inequality Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3.2 The Atkinson Family of Inequality Indices . . . . . . . . . . . . 64
4.3.3 The Atkinson and the Entropy Families of Inequality
Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5 Inequality of Opportunity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.2 Measuring Inequality of Opportunity by Theil’s Inequality
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2.3 The Second Index of Theil . . . . . . . . . . . . . . . . . . . . . . . . 80
5.3 Equality of Opportunity With Categorical Data . . . . . . . . . . . . . . 81
5.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.3.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.3.3 Empirical Illustration: Inequality of Opportunity in
Compulsory Education in the OECD, According to PISA
(2012) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6 Inequality and Welfare . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.2 Social Evaluation Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.3 Multidimensional Inequality and Welfare . . . . . . . . . . . . . . . . . . 101
6.3.1 Multidimensional Welfare Indicators . . . . . . . . . . . . . . . . 103
6.3.2 Multidimensional Inequality . . . . . . . . . . . . . . . . . . . . . . . 105
6.3.3 Weights and Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.4 The Human Development Index . . . . . . . . . . . . . . . . . . . . . . . . . 109
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Contents xi
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
1.1 Equality and Welfare
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.
1.1.1 We Have a Problem. . .
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
© Springer International Publishing AG 2017 1

A. Villar, Lectures on Inequality, Poverty and Welfare, Lecture Notes in Economics
and Mathematical Systems 685, DOI 10.1007/978-3-319-45562-4_1
2 1 Introduction
situation is alarming, due to the extremely unequal opportunities to cover basic

needs (nutrition, shelter, education) or access to essential facilities (water, sani-
tation, electricity or information).
A look at the distribution of wealth among the countries in the world gives a first
idea about how bad is the situation. The data, elaborated by the Research Institute of
the Credit Suisse (see Shorrocks, Davies and Lluberas, 2014), tell us that Europe
holds about one-third of the world’s total wealth, North America owns another third
and the rest of the world the remaining one. Africa, China, India and South America
only account for about half of that last third, which implies that more than half of
the world’s population only gets some 15 % of the total wealth. The whole continent
of Africa barely holds 1 % of the aggregate wealth.
Such unequal distribution of resources, which results in high levels of inequality
and poverty in a large fraction of the world’s population, involves an extremely
unfair distribution of opportunities. Think, in particular, of the case of children.
Their opportunities are very much affected by the country and the family in which
they are born, something outside their choice and therefore unfair. Inequality and
poverty affect consumption opportunities, nutrition, health, education, personal
development, family relations, work options, social cohesion, political influence
and the overall economic growth potential of societies. So inequality and poverty
affect both individuals and societies and involve ethical and efficiency issues.
Moreover, most of these factors (e.g., education) are in turn determinants of the
income distribution, closing a circle of social discrimination.
The concern for inequality and poverty seems to have risen during the last
decades and is now a key element in the political agenda of many countries and
international organisations, such as the United Nations, the World Bank, the
International Monetary Fund or the OECD, among others. The success of Piketty
and Ganser (2014) book on inequality and capitalism, further beyond the realm of
academics, underlines this new sensitivity. In 2010, the United Nations Develop-
ment Programme included inequality in the measurement of human development
and introduced a new approach to assess poverty (see UNDP, 2010). In 2015, the
United Nations member states agreed on 17 Global Goals for Sustainable Devel-
opment, aimed at guiding public policies from 2016 to 2030.1 Those goals include
the reduction of inequality within and among countries and the eradication of
poverty. To justify the relevance of those goals, the United Nations highlights
some facts that help understanding the extent of the problems addressed. Among
them we can mention the following:
– More than 75 % of the world’s population is living today in societies where
income is more unequally distributed than it was in the 1990s.
– Evidence from developing countries shows that children in the poorest 20 % of
the populations are still up to three times more likely to die before their fifth
birthday than children in the richest quintiles.
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
– 836 million people still live in extreme poverty.

– About one in five persons in developing regions lives on less than $1.25 per day.
The overwhelming majority of those people belong to two regions: Southern
Asia and sub-Saharan Africa.
– One in seven children under age five in the world has inadequate height for his or
her age.
The evolution of inequality and poverty is complex, and each presents particular
features. It is interesting to stress that there is evidence of a clear improvement in
extreme poverty, as the number of people living in extreme poverty has dropped by
more than half between 1990 and 2015. This is very good news, even though the
economic crisis and a number of political and military conflicts have produced an
increase of poverty in many developing and developed countries. Needless to say,
the higher poverty rates are often found in small, fragile and conflict-affected
countries.
Inequality presents a long-standing increasing path in most countries, both
developing and developed. The distribution of personal income was already wors-
ening long before the beginning of the crisis and has been made even worse in the
last years. There are structural and cyclical causes behind this pattern. The cyclical
component of the higher inequality is mostly due to the increment of unemploy-
ment, which affects more workers with less human capital and less stability in their
contracts.2 The most important of the structural causes is technological change.
This change has led to some polarisation between high- and low-skilled workers in
many countries and to a progressive thinning of what we might call “middle-
classes”. The demand for labour seems to concentrate progressively on highly
qualified workers with high wages and low-skilled manual and low-wage workers
who perform socially useful tasks (see Dabla-Norris et al., 2015). Technological
change affects not only the distribution of personal income but is also producing a
reduction of the share of labour in total income. It is partly related to the increase in
the turnover of capital, capital increasingly complex with greater obsolescence and
lower relative cost per unit of output. This entails higher depreciation and thus
reducing the remuneration of factors in an asymmetric way, changing the capital/
labour ratios (Karabarbounis & Neiman, 2013; Piketty & Ganser, 2014). The other
structural cause of this evolution of inequality within countries is associated with
the globalisation of the economy (even though it may reduce inequality between
countries).
Inequality and poverty have an impact not only on the present welfare of those
individuals or families involved but also affect the opportunities of their descen-
dants and the capabilities of the whole society. As United Nations points out in the
Sustainable Development Goals related to these aspects, evidence shows that,
beyond a certain threshold, inequality harms growth and poverty reduction, the
quality of relations in the public and political spheres and individuals’ sense of
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.
1.1.2 Measuring Inequality and Poverty
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
The purpose of this monograph is to provide some basic ingredients to measuring

inequality and poverty. Such a measurement refers ultimately to the degree of
unfairness of the distribution of opportunities in a population. The evaluation of
inequality and poverty involves, therefore, value judgements about what a better
distribution of resources and opportunities would be. Those value judgements can
be linked, directly or indirectly, to welfare assessments on the distribution of
personal and social opportunities. So, from a wider perspective, inequality and
poverty can be regarded as part of the welfare evaluation of the opportunities in a
given society. The backbone of this monograph is, precisely, the analysis of the
interconnections between inequality, poverty and social welfare.
Inequality and poverty are related but different aspects of the problem of a fair
distribution of opportunities in society. Inequality focuses on the asymmetry of
those opportunities among the whole population. Poverty refers to the situation of
that part of the population that fails to achieve some baseline opportunities. As a
consequence, the approach to measurement follows different paths. In both cases,
though, there is the implicit assumption that equality is an essential social value.
We shall focus on the measurement of inequality and poverty on the construction
of sensible indices that provide quantitative estimates of those phenomena, mostly
in terms of some objective variable (as opposed to qualitative assessments or in
terms of subjective perceptions). Inequality and poverty indices are critical ele-
ments for the analysis because they provide the measuring rods to such an analysis.
Those indices have a number of relevant applications, among which there are the
following:
1. Allow comparing the evolution of inequality and poverty in a given society, as
well as the distributional impact of policy measures.
2. Allow comparing equality and poverty between different societies.
3. Permit one introducing distributional judgements in welfare evaluation, thus
going beyond the GDP (the United Nations Human Development Index is a case
in point).
4. Facilitate the analysis of the origin and nature of inequality and poverty as some
of these indices permit decomposing the measurement between social groups or
regions.
6 1 Introduction
The relevance of those applications calls for a sound understanding of those

indices, formulae and properties, as not all of them will evaluate (or even rank)
income distributions the same way.
1.2 Equality of What?
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.
1.2.1 Income Distributions
Income distribution is a key element in the analysis of economic inequality and

poverty, as it is an observable variable directly related to consumption and invest-
ment opportunities. The notion of income distribution can be given a number of
different interpretations. It may refer to the differences in the retribution of eco-
nomic factors (i.e. the prices of the different types of labour and capital), to the
division of national income between different sources (wages, profits, rents, inter-
ests), to the division between different territories or social groups, etc. Along this
monograph, we shall interpret income distributions as referring to the distribution
of income between people (individuals or families), i.e. the distribution of personal
income.
There are several reasons why personal income is an outstanding candidate when
talking about economic inequality and poverty. First and foremost, personal income
is the main determinant of consumption opportunities and, therefore, a good proxy
of material well-being. Second, personal income can be affected by economic
policies in a predictable way, using standard instruments (e.g. taxes and transfers,
subsidies), whose effects can be monitored. And third, because income is a variable
on which there are usually regular, rich and trustable data. Yet personal income is
not everything that matters, especially regarding poverty. In many countries,
monetary incomes are not the best variable to approach economic opportunities
because they lack sufficiently developed markets and the basic needs, such as
sanitation, shelter, education, access to water, energy or information, represent
much more basic concerns and therefore call for different ways of measurement.
Relying on prices and markets implicitly assumes the existence of sufficient
1.2 Equality of What? 7
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.
1.2.2 Wealth, Income and Expenditures
From a theoretical viewpoint, we can think of an agent’s income in a given period as

the market value of his/her consumption opportunities relative to that period. That
would correspond to the actual consumption within the period plus the change in
wealth experienced during the period. In other words, income in this broad sense
involves two flow variables, earnings and expenditures, plus the change in a stock
variable (wealth). These simple considerations already point out that inequality and
poverty measurement is a complex exercise. The complications do not stop here.
Even if we have clearly defined those notions, it is not that immediate to transform
those ideas into operational variables, as there are some other concerns to be taken
into account.
To start with, note the income is a flow variable, that is, a magnitude relative to a
time interval. Which implies that the choice of the time interval may be relevant.
Moreover, there is evidence on the existence of a common pattern of earnings along
the life of the agents (increasing up to a point and then decreasing). So the observed
distribution pattern will reflect income differences that are partly due to the
population structure. This suggests recurring to some notion of permanent income,
in order to avoid this problem, even though this option is not always available (see
Lam, 1986, 1988; Paglin, 1975).
Secondly, there is the problem of reducing current values of different periods
and different countries or regions to “real values”. There are standard procedures to
deal with those issues, e.g. expressing the values in terms of the purchasing power
8 1 Introduction
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).
1.2.3 The Units of Reference: Equality Among Whom?
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
measure of the household standard of living, as it corresponds to the income that, if

enjoyed by a single-member household, would yield the same welfare as a repre-
sentative member of the original household.
Adjusting household expenditures using the equivalised size of the households
seems better than using per capita values, because in this way we are able to
compute the presence of economies of scale and the differential needs within
households.5
1.2.4 Needs and Deserts
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.
1.3 Plan of the Book
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
Chapter 8 develops an approach to poverty measurement based on the inter-

pretation of poverty as a welfare loss. Following the standard approach in the
normative theory of income inequality, poverty indices are derived here from a
social evaluation function and some poverty thresholds. A welfare poverty index is
defined as the relative welfare loss due to the insufficient welfare of those agents
whose achievements do not reach the minimum established. The construction of
those indices is formulated in a multidimensional context. We show that, under
conventional assumptions, those indices can be expressed as the product of the
incidence and the inequality-adjusted intensity of poverty. We include an appli-
cation to the measurement or educational poverty using the data from PISA.
Finally, Chap. 9 deals with the analysis of the achievements of a series of targets.
The reference problem is the following: a society, a firm or an institution defines a
series of targets to be accomplished in a given time period. How should we evaluate
the overall performance when some targets have been achieved or even surpassed
while others have not been reached? We show here that the methodology of
multidimensional poverty analysis can be applied to deal with this type of problem.
We characterise an elementary multidimensional evaluation formula and illustrate
how it has been applied to the measurement of progress in the United Nations Green
Economy Initiative.
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
Cowell, F. A. (1977). Measuring inequality. Oxford: Phillip Allan.

Cowell, F. A. (1995). Measuring inequality, LSE handbooks in economics series (2nd ed.).
Hemel Hempstead: Harvester Wheatsheaf.
Cowell, F. A. (2003, May). Theil, inequality and the structure of income distribution (Working Paper
DARP No. 67). STICERD, LSE.
Cowell, F. A., & Jenkins, S. P. (2000, June). Estimating welfare indices: Household weights and
sample design (Working Paper DARP No. 48). STICERD, LSE.
Cowell, F. A., & Mercader-Prats, M. (1999, Marzo). Equivalence scales and inequality
(Working Paper DARP No. 27). STICERD, LSE.
Dabla-Norris, M. E., Kochhar, M. K., Suphaphiphat, M. N., Ricka, M. F., & Tsounta, E. (2015).
Causes and consequences of income inequality: A global perspective. Washington, DC:
International Monetary Fund.
Deaton, A., & Zaidi, S. (2002). Guidelines for constructing consumption aggregates for welfare
analysis (LSMS Working Paper No. 135). Washington, DC: The World Bank. http://www.
worldbank.org/lsms
Duclos, J. Y., & Araar, A. (2006). Poverty and equity: Measurement, policy and estimation with DAD.
Boston/Dordrecht/London: Kluwer Academic Publishers.
Ebert, U. (1995). Income inequality and differences in household size. Mathematical Social Sci-
ences, 30(1), 37–53.
Ebert, U. (1997). Social welfare when needs differ: An axiomatic approach. Economica, 64(254),
233–244.
Ebert, U. (1999). Using equivalent income of equivalent adults to rank income distributions when
household types are different. Social Choice and Welfare, 16(2), 233–258.
Ebert, U., & Moyes, P. (2003). Equivalence scales reconsidered. Econometrica, 71(1), 319–343.
Foster, J., Seth, S., Lokshin, M., & Sajaia, Z. (2013). A unified approach to measuring poverty and
inequality. Washington, DC: The World Bank.
Goerlich, F., & Villar, A. (2009). Desigualdad y Bienestar. De la Teoría a la Pr actica. Fundación
BBVA.
Houghton, J., & Khander, S. R. (2009). Handbook of poverty and inequality. Jenkins: The World
Bank.
Jenkins, S. P., & Cowell, F. A. (1994). Parametric equivalence scales and scale relativities.
The Economic Journal, 104(425), 891–900.
Kakwani, N. (1980). Income inequality and poverty. Methods of estimation and policy appli-
cations. The World Bank, Oxford: Oxford University Press.
Karabarbounis, L., & Neiman, B. (2013). The global decline of the labor share (No. w19136).
National Bureau of Economic Research.
Krueger, D., & Perri, F. (2006). Does income inequality lead to consumption inequality?
Evidence and Theory, Review of Economic Studies, 73, 163–193.
Lam, D. (1986). The dynamics of population growth, differential fertility, and inequality.
The American Economic Review, 76, 1103–1116.
Lam, D. (1988). Lorenz curves and inequality comparisons under changing population composition.
Journal of Policy Modelling, 10, 141–162.
Lambert, P. J. (1993). The distribution and redistribution of income (2nd ed.). Manchester:
Manchester University Press.
Love, R., & Wolfson, M. C. (1976). Income inequality: Statistical methodology and Canadian illus-
trations. Ottawa: Statistics Canada.
McClements, L. D. (1977). Equivalence scales for children. Journal of Public Economics, 8(2),
191–210.
Osmani, S. (1982). Economic inequality and group welfare. Oxford: Clarendon.
Paglin, M. (1975). The measurement and trend of inequality: A basic revision. American Eco-
nomic Review, 65, 598–690.
Piketty, T., & Ganser, L. J. (2014). Capital in the twenty-first century. Cambridge, MA:
Harvard University Press.
References 15
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.

20 2 Inequality Indices
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).
2.2 Properties of Inequality Indices
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
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
individual incomes. In particular, the continuity property establishes that two

income distributions that are very close will have very similar inequality values.
Property 6: Scale Independence
Scale independence is a usual property that establishes, from a formal viewpoint,
that inequality indices are homogeneous functions of degree zero. That is, for all
λ > 0, all y 2 ℝþþ
n
, I ðλyÞ ¼ I ðyÞ.
This is a two-sided property. On the one hand, it establishes that a change in the
units in which incomes are measured should not affect inequality (e.g. λ can be
interpreted as a exchange rate). This is clearly a highly desirable property. On the
other hand, it also tells us that multiplying all incomes by a constant, that is,
increasing or decreasing all incomes by the same proportion, does not change
inequality. In other words, inequality is a relative measure. This is a much more
arguable principle, because increasing all incomes proportionally widens the abso-
lute income differences between rich and poor people.
This property provides a specific answer to the following general question: what
changes in income values leave inequality unaltered? One can think of two extreme
responses. One would correspond to scale independence, which establishes that
proportional changes in incomes do not affect inequality. This amounts to consid-
ering inequality as a relative concept and is sometimes regarded as a conservative or
rightist notion of equity. In the other extreme, one may argue that inequality does
not change if we add or subtract the same amount to all incomes. This entails an
absolute concept of inequality; those indices that exhibit this property are some-
times called leftists. See the classic works of Kolm (1976a, 1976b) and the
discussion in Blackorby and Donaldson (1980).3
Remark 2.3 All former properties were ordinal in nature, in the sense that if an
index satisfies them, then any monotone transformation of this index will also satisfy
those five properties. The scale independence property introduces a cardinal
feature in our indices.
Property 7: Additive Decomposability
This is another cardinal property. It refers to the analysis of inequality in a society
made of several population subgroups, defined by some criteria unrelated to the
income distribution process (regional location, gender, religion, etc.). It establishes
a sort of consistency requirement on the evaluation of inequality in the whole
society and in the constituent subgroups. It consists on being able to express total
inequality as the sum of two different components, inequality within and between
the groups, by applying the same index to all inequality measurements (total, within
and between).4 More precisely, we look for the following relationship:
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),
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
distributionof the whole society N can be regarded as a vector of vectors,

income
y ¼ y1 ; y2 ; . . . ; yG . Let now ~
μ ¼ ðμ1 , μ2 , . . . , μG Þ be the vector of mean incomes
of the different population subgroups, and let n ¼ ðn1 ; n2 ; . . . ; nG Þ be the vector of
population sizes. Finally, let 1ng stand for a unit vector of dimension ng, that is,

1ng ¼ 1, 1, . . . , 1 . We say that an inequality index satisfies additive decompos-
|fflfflfflfflfflffl{zfflfflfflfflfflffl}
ng
ability if it can be expressed as follows:
X
I ð yÞ ¼ ω ð~
g g
μ; nÞI ðyg Þ þ I ðμ1 1n1 , . . . , μG 1nG Þ
for a set of positive coefficients ωg, which may depend on ~

μ and n.
Remark 2.4 There are more general forms of decomposability that will not be
considered here. See for instance Goerlich and Villar (2009, Chs. 7, 8) for a
detailed discussion.
2.3 Inequality and Dispersion
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.
2.3.1 The Variance (σ2)
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:
Table 2.1 % of income and population in the two income distributions
Cumulative population Cumulative income

(ordered from less to more income) (%) Distribution y1 (%) Distribution y2 (%)
20 5 12
40 15 24
60 30 36
80 50 48
100 100 100
1X n
σ2 ¼ ðy μÞ2
n i¼1 i
ð2:1Þ
1X n
¼ y2i μ2
n i¼1
An alternative expression of the variance is the following (Kendall & Stuart,

1977):
n 2
11X n X
σ2 ¼ 2
yi yj ð2:2Þ
2 n i¼1 j¼1
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.
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).
2.3.2 The Coefficient of Variation (CV)

pffiffiffiffiffi
It is frequent using the square root of the variance, σ ¼ σ 2 (i.e. the standard
deviation), instead of the variance to measure dispersion. The reason is that the
variance is a measure in terms of the squares of the units of the variables whereas
the standard deviation is expressed in the same units. As this is a monotone
transformation of the variance, it preserves all its ordinal properties (normalisation,
symmetry, population replication, principle of transfers and continuity), but it
losses the additive decomposability property, as this is a cardinal property (see
the discussion in Chap. 3, however).
The standard deviation does not satisfy scale independence. Yet it is easy to get a
simple and intuitive transformation of this measure by simply dividing the standard
deviation by the mean. This is, precisely, the coefficient of variation (CV):
σ
CV ¼ ð2:4Þ
μ
As dividing by the mean is just another monotone transformation (linear in this

case), the coefficient of variation satisfies all properties presented in Sect. 2.1, other
pffiffiffiffiffiffiffiffiffiffiffi
than additive decomposability. It is easy to check that CV 2 0; n 1 and that
the impact of an Dalton’s transfer δ from agent i to agent j is given by:
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
Atkinson, A. B., & Bourguignon, F. (1982). The comparison of multidimensional distributions of

economic status. Review of Economic Studies, 49(2), 183–201.
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).
Chapter 3
Positive Inequality Indices
3.1 The Lorenz Curve
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).

32 3 Positive Inequality Indices
Fig. 3.1 The Lorenz curve L

(for a discrete distribution) 1
Line of perfect equality
% cummulative income
Lorenz
curve
0 1/n 2/n 3/n ... 1 F

% cummulative population
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
Fig. 3.2 Dalton’s transfers

and the Lorenz curve 1 A
L(p)
% cumulative income
between quantiles q and r
0 r q 1 F
% cumulative population
Fig. 3.3 Lorenz dominance

1
% cumulative income
yA
yB
yC
0 1 F
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
when the Lorenz curves cross each other.1 See Dasgupta, Sen, and Starrett (1973)
and Rothschild and Stiglitz (1973).
3.2 The Gini Index
3.2.1 From Lorenz to Gini
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
As A + B ¼ ½, we can also write:
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
Fig. 3.4 The Lorenz curve

and the Gini index 1
L(p)
% cumulartive income
A
0 1 F
Fig. 3.5 Area below the L

Lorenz curve 1
Line of perfect
equality
pi
% cumulative income
Lorenz curve
pi - 1
0 pi - 1 pi 1 F
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.
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
As p0 ¼ L0 ¼ 0, and bearing in mind that G ¼ 2A [see equation (3.1b)], we obtain:
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
Note that p0 ¼ L0 ¼ 0 and pn ¼ Ln ¼ 1, so that we can write:

p0 L0 ¼ pn Ln ¼ 0, and simplify the above equation. Since
Xn X
n
ðpi1 Li1 Þ ¼ ðpi Li Þ, we can write:
i¼1 i¼1
" #
1 Xn Xn
G¼ ðp Li Þþ ðpi1 Li1 Þ
n i¼1 i i¼1
or, put in a simpler way:
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
y ¼ ðy1 , y2 , . . . yn Þ. It makes explicit that all those differences between population

and income shares are equally important. We shall see that this property can be
weakened generating the family of generalised Gini indices.
Sen (1973) proposes an alternative way of expressing the Gini index (see
Goerlich and Villar, 2009, Ch. 4) that underlies the way in which this indicator
treats income differences. It is the following3:
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.
3.2.2 The Generalised Gini Index
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
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
This is a uni-parametric family of inequality measures whose members are

singularised by the parameter v > 1. For v ¼ 2, we have κi(v) ¼ 2 for all i, which
corresponds to the original Gini index (i.e. G2 ¼ G). For 1 < v < 2, this index gives
more weight to the rich than to the poor, whereas the contrary occurs for v > 2.
The parameter v can thus be interpreted as a degree of inequality aversion: the
higher its value the higher the relevance of what happens in the lower part of the
distribution. Indeed, one can show that, for any given distribution y,
y1
Gv ! 1 , when v ! 1
μ
Gv ! 0, when v ! 1
When using Sen’s formulation of the Gini index, we have:

n
v X v1 yi
Gv ¼ 1 ðn þ 1 iÞ ð3:6Þ
nv1 i¼1
nμ
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.
3.3 Theil’s Inequality Indices
3.3.1 The First Index of Theil: From Information Theory

to Inequality Measurement
Theil (1967) presents a different way of approaching inequality measurement,

adopting some ideas related to the notion of entropy in information theory (see
Khinchin, 1957; Kullback, 1959). The notion of entropy in this context refers to the
expected information of a given situation. This situation involves a collection of
random events with some probabilities w.
The informational content of the occurrence of an event is measured by an
information function, ϕ, that tells us how valuable is to know that something has
happened as a function of its probability, ϕ(w). Such a function is required to satisfy
two basic properties:
(i) The informational content of knowing the occurrence of an event is a decreas-
ing function of its probability, with zero informational value for the sure event.
That is, w > w0 ) ϕðwÞ < ϕðw0 Þ, with ϕð1Þ ¼ 0. This translates the idea that
the less likely is an event, the more important is to know that it has occurred.
(ii) The joint information of two independent events is the sum of the information
of each separate event: ϕðw w0 Þ ¼ ϕðwÞ þ ϕðw0 Þ.
It can be shown that, assuming that ϕ is differentiable, those properties deter-
mine completely the functional form of the information function, which is given by:

1
ϕðwÞ ¼ logðwÞ ¼ log
w
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
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.
e¼ 1 T
T ð3:9Þ
logn
with 0 Te 1. Note, however, that this normalisation implies renouncing to the

property of population replication, as the normalised index depends on the
population size.
Using elementary algebra and bearing in mind that si ¼ yi =nμ, we can rewrite
equation (3.8) as follows:
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
assuming that yj < yi . It therefore satisfies the principle of transfers.

With respect to additive decomposability, we can derive the following decom-
position that shows that the index of Theil satisfies this property (we use the same
notation as in the case of the variance in Sect. 2.3):
!
1X G X XG X
yi μ g X G X
ng ng ng
yi
g
y
g
yi
g g g
yi y
g
μg
T ¼ log i ¼ log ¼ log i þ log
n g¼1 i¼1 μ μ g¼1 i¼1
nμ μg μ g¼1 i¼1
nμ μ g μ
"n #
XG X g
yig yig X
ng
yig μg
¼ log þ log
g¼1 i¼1
nμ μg i¼1 nμ μ
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
Tg
7
See Kanbur (1984) for a different interpretation.
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.
3.3.2 The Second Index of Theil (Mean Logarithmic

Deviation)
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
Theil interprets T* as the expected informative value of an indirect message that

transforms income shares as a priori probabilities into posterior population shares.
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
assuming that yj < yi .

The decomposition of the second index of Theil is as follows:
8
Indeed, this index is not defined when si ¼ 0 for some i.
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
We have, therefore, the following:
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.
3.3.3 The Generalised Indices of Theil (Tθ)
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
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 decomposability of the square of the coefficient of variation, CV2, can be

easily realised by noting that this measure is nothing else than the variance of the
mean scaled income vector (i.e. an income vector in which all incomes are divided
by the mean). That is,
n 2
1 1X yi 1 Xn
σ2
y ¼ 1 ¼ 2 ðy μÞ2 ¼ CV 2
μ n i¼1 μ nμ i¼1 i
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.
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
The between component is the square of the weighted coefficient of variation

relative to the means of the different groups. The within component corresponds to
the weighted sum of the squared coefficients of variation of the groups. Note,
however, that now those coefficients do not add up to 1, as they correspond to the
values:
ng μ2g n2g μ2g n

¼
n μ2 n2 μ2 ng
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
The parameter θ is, therefore, similar in nature to the parameter v in the

generalised Gini index (and also to that of the Atkinson (1970) family of indices,
to be discussed in the next Lecture).
***
How does this family of indices fit with the information theory that lies behind
Theil’s inequality indices? We know that the information function ϕ(w) ¼ log(w)
is the only function that satisfies the criteria presented before: (a) w > w0 ) ϕ
(w ) < ϕ(w0 ); (b) ϕ(1) ¼ 0; and (c) ϕ(w∙w0 ) ¼ ϕ(w) þ ϕ(w0 ). So one of those prop-
erties will have to be weakened in order to get a generalised version of the function.
Note that property (c) is perfectly fine for random variables, but it is much less
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
The nature of this generalisation can be related to the introduction of different

degrees of progressivity of Dalton’s transfers, as it was the case with the generalised
Gini index. We have seen that the impact of a Dalton’s transfer in the first index of
Theil, dT, is given by12:
!
δ y
dT ¼ log i
nμ yj
ð3:19Þ
¼ ds logsi logsj
¼ ds ϕðsi Þ ϕ sj
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 *.
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.
3.4 Quantile Measures: The Palma Ratio
A much simpler approach to positive inequality measurement derives from the

comparison of different quantiles of the income distributions. One may consider the
share of total income in the hands of the richest 1 %, 5 % or 10 % of the population.
Yet quantile measures are usually formulated in terms of ratios. The most frequent
ones are the 80-20 ratio, the 10-90 ratio, the 20-20 (or quintile) ratio or the ratio
between the median and the mean, which provides a measure of the asymmetry in
the distribution. Note that these simple measures satisfy most of the properties one
may require for an inequality index, as stated in Chap. 2 (normalisation, symmetry,
population replication, a weak version of the principle of transfers,13 continuity and
homogeneity). As it was the case with the Gini index and the indices of Theil, those
measures can be derived directly from the Lorenz curve.
The advantage of quantile measures is that they are very intuitive and require
little information and elaboration. They are, therefore, accessible to the layman.
The inconveniencies are also clear. On the one hand, they only provide information
on one or two cuts of the distribution. On the other hand, changing the selected
quantiles changes substantially the figures that measure inequality, without having
good reasons to prefer one over the others.
In 2011, the Chilean economist Gabriel Palma published a paper in which he
showed a surprising empirical regularity regarding income distributions across
countries: the middle income groups, defined as those in the five deciles 5 to
9, get about 50 % of total income. The observed variability, therefore, would
correspond to the way in which the other half of the total income is distributed in
the complementary groups: the 10 % richest and the 40 % poorest (see Palma,
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.
Table 3.1 Several inequality indices in the world 2003–2012

Rank Country Quintile ratio Palma ratio Gini coefficient (%)
Very high human development
49 Argentina 11.3 2.4 44.5
21 Austria .. .. 29.2
21 Belgium .. .. 33.0
8 Canada .. .. 32.6
41 Chile 13.5 3.5 52.1
47 Croatia 5.2 1.4 33.7
33 Estonia 6.4 .. 36.0
24 Finland .. .. 26.9
6 Germany .. .. 28.3
29 Greece .. .. 34.3
43 Hungary 4.8 1.2 31.2
11 Ireland .. .. 34.3
19 Israel .. .. 39.2
26 Italy .. .. 36.0
48 Latvia 6.0 1.4 34.8
18 Liechtenstein .. .. ..
35 Lithuania 6.7 1.6 37.6
21 Luxembourg .. .. 30.8
4 The Netherlands 5.1 .. 30.9
1 Norway .. .. 25.8
35 Poland 5.2 1.3 32.7
31 Qatar 13.3 .. 41.1
37 Slovakia 3.6 0.9 26.0
25 Slovenia 4.8 .. 31.2
27 Spain .. .. 34.7
12 Sweden .. .. 25.0
3 Switzerland .. .. 33.7
14 United Kingdom 7.2 .. 36.0
5 United States .. .. 40.8
High human development
95 Albania 5.3 1.4 34.5
87 Armenia 4.6 1.2 31.3
76 Azerbaijan 5.3 1.4 33.7
53 Belarus 3.8 0.9 26.5
84 Belize 17.6 .. 53.1
86 Bosnia and Herz. 6.5 1.5 36.2
79 Brazil 20.6 4.3 54.7
58 Bulgaria 4.3 1.0 28.2
91 China 10.1 2.1 42.1
98 Colombia 20.1 4.5 55.9
68 Costa Rica 14.5 3.3 50.7
(continued)
3.4 Quantile Measures: The Palma Ratio 51
Table 3.1 (continued)

102 Dominican Republic 11.3 2.7 47.2
98 Ecuador 12.5 3.1 49.3
88 Fiji 8.0 2.2 42.8
79 Georgia 9.5 2.1 42.1
75 Iran 7.0 1.7 38.3
96 Jamaica 9.6 .. 45.5
77 Jordan 5.7 1.5 35.4
70 Kazakhstan 4.2 1.1 29.0
62 Malaysia 11.3 2.6 46.2
71 Mexico 10.7 2.7 47.2
51 Montenegro 4.3 1.0 28.6
65 Panama 17.1 3.6 51.9
82 Peru 13.5 2.9 48.1
54 Romania 4.1 1.0 27.4
57 Russian Federation 7.3 1.9 40.1
77 Serbia 4.6 1.1 29.6
71 Seychelles 18.8 6.4 65.8
73 Sri Lanka 5.8 1.6 36.4
100 Suriname 17.9 .. 52.9
89 Thailand 6.9 1.8 39.4
84 Republic of Macedonia 10.0 2.3 43.6
90 Tunisia 6.4 1.5 36.1
69 Turkey 8.3 1.9 40.0
83 Ukraine 3.6 0.9 25.6
50 Uruguay 10.3 2.5 45.3
67 Venezuela 11.5 2.4 44.8
Medium human development
142 Bangladesh 4.7 1.3 32.1
136 Bhutan 6.8 1.7 38.1
113 Bolivia 27.8 4.8 56.3
136 Cambodia 5.6 1.5 36.0
123 Cape Verde .. .. 50.5
140 Congo 10.7 2.8 47.3
110 Egypt 4.4 1.2 30.8
115 El Salvador 14.3 3.0 48.3
112 Gabon 7.8 2.0 41.5
138 Ghana 9.3 2.2 42.8
125 Guatemala 19.6 4.5 55.9
129 Honduras 29.7 5.2 57.0
135 India 5.0 1.4 33.9
108 Indonesia 6.3 1.7 38.1
120 Iraq 4.6 1.2 30.9
(continued)

125 Kyrgyzstan 5.4 1.3 33.4
139 Lao 5.9 1.6 36.7
103 Maldives 6.8 .. 37.4
124 Micronesia .. .. 61.1
114 Moldova 5.3 1.3 33.0
103 Mongolia 6.2 1.6 36.5
129 Morocco 7.3 2.0 40.9
127 Namibia 21.8 .. 63.9
132 Nicaragua 7.6 1.9 40.5
107 Palestine. State of 5.8 1.5 35.5
111 Paraguay 17.3 3.7 52.4
117 Philippines 8.3 2.2 43.0
142 Sao Tome and Principe .. .. 50.8
118 South Africa 25.3 7.1 63.1
118 Syrian Arab Republic 5.7 .. 35.8
133 Tajikistan 4.7 1.2 30.8
116 Uzbekistan 6.2 1.6 36.7
121 Viet Nam 5.9 1.5 35.6
141 Zambia 17.4 4.8 57.5
Low human development
169 Afghanistan 4.0 1.0 27.8
149 Angola 9.0 2.2 42.7
165 Benin 6.6 1.8 38.6
181 Burkina Faso 7.0 1.9 39.8
180 Burundi 4.8 1.3 33.3
152 Cameroon 6.9 1.8 38.9
185 Central African Rep. 18.0 4.5 56.3
184 Chad 7.4 1.8 39.8
159 Comoros 26.7 .. 64.3
186 Congo 9.3 2.4 44.4
171 Côte d’Ivoire 8.5 2.0 41.5
170 Djibouti .. .. 40.0
173 Ethiopia 5.3 1.4 33.6
172 Gambia 11.0 2.8 47.3
179 Guinea 7.3 1.8 39.4
147 Kenya 11.0 2.8 47.7
162 Lesotho 19.0 3.9 52.5
175 Liberia 7.0 1.7 38.2
155 Madagascar 9.3 2.3 44.1
174 Malawi 8.9 2.3 43.9
176 Mali 5.2 1.3 33.0
161 Mauritania 7.8 1.9 40.5
(continued)
Appendix: Decomposability of the Generalised Indices of Theil 53

178 Mozambique 9.8 2.5 45.7
145 Nepal 5.0 1.3 32.8
187 Niger 5.3 1.4 34.6
152 Nigeria 12.2 3.0 48.8
146 Pakistan 4.2 1.2 30.0
151 Rwanda 11.0 3.2 50.8
163 Senegal 7.7 1.9 40.3
183 Sierra Leone 5.6 1.5 35.4
166 Sudan 6.2 1.4 35.3
148 Swaziland 14.0 3.5 51.5
159 Tanzania 6.6 1.7 37.6
166 Togo 7.6 1.8 39.3
164 Uganda 8.7 2.3 44.3
154 Yemen 6.3 1.7 37.7
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).
Appendix: Decomposability of the Generalised Indices

of Theil
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:
"
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
which depend on the parameter θ of inequality aversion.

Note that those weights can be expressed in terms of population and income
shares, that is,
θ
ng μ g θ ng 1θ ng μ g
ωgG ðμ; nÞ ¼ ¼
n μ n nμ
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.
Gini, C. (1921). Measurement of inequality of incomes. The Economic Journal, 31, 124–126.
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.
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).
4.1.1 The Arrowian Approach to Normative Inequality

Indices
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

58 4 Normative Inequality Indices
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,
u : Φ ! ℝn , uðxÞ ¼ ½u1 ðxÞ, u2 ðxÞ, . . . , un ðxÞ
A social welfare function, relative to a society N whose individuals have utilities

u, is a mapping W u : Φ ! ℝ that evaluates social alternatives from the agents’
utilities. It is customary to assume that this function satisfies the following proper-
ties: universal domain (all individual preferences are admissible), unanimity (if all
agents agree that an alternative is better than another, then the social evaluation
should also concur with this judgement), anonymity (the only relevant information
of the agents is that corresponding to their preferences), and informational effi-
ciency (any two social alternatives can be evaluated out of the individual utilities
for those two alternatives, without needing to know the agents’ evaluation of the
remaining).
It is well established that those properties are incompatible unless we are ready
to perform interpersonally utility comparisons (see D’Aspremont, 1984, 1994;
D’Aspremont & Gevers, 1977). That is, assuming that we can meaningfully say
things like “agent i is better off under alternative x than agent j under alternative z”.
That comparison can be established in terms of levels (ordinal comparability) or in
terms of common units (cardinal unit comparability), each yielding a particular
social welfare function.1 Under ordinal level comparability, one basically has to
measure social welfare by the following function, known as leximin2:
W u ðxÞ ¼ mini f½u1 ðxÞ, u2 ðxÞ, . . . , un ðxÞg
Under cardinal unit comparability, the social welfare function corresponds to

classical utilitarianism. 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
The stronger the interpersonal comparability we assume, the more flexibility to

choose the functional form of social welfare functions. In the extreme, when all
utilities are cardinal and identical for all agents, we have the largest number of
options.
Let us now focus on the case in which utility only depends on income, to open
the path that connects inequality and welfare. The set Φ of social alternatives
corresponds in our context to the space of income distributions (the interior of
ℝþn , to be precise). The social welfare associated to an income distribution y 2
ℝþþ
n
is given by:
W u ðyÞ ¼ W ½u1 ðyÞ, u2 ðyÞ, . . . , un ðyÞ
(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:
W u ðyÞ ¼ F½μðyÞ, I ðyÞ
Function F is assumed to be decreasing in I(y) (more inequality implies lower

welfare) and, usually, increasing in μ (even though this is less obvious as one may
consider that increasing the mean income for high values of inequality may reduce
social welfare).
A standard way of ensuring that welfare decreases with inequality is assuming
that Wu(y) is strictly quasi-concave, that is, 8y 6¼ y0 2 ℝþþ
n
and for all λ 2 ð0; 1Þ, we
have:
W u ðλy þ ð1 λÞy0 Þ > minfW u ðyÞ, W u ðy0 Þg
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).
4.1.2 The Social Evaluation Function Approach
There is an alternative approach to normative inequality measurement that avoids

using individual utilities as arguments. It relies on the definition of a social
evaluation function, V : ℝþþn
! ℝ, that associates a real number V(y) to each
income distribution y ¼ ðy1 ; y2 ; . . . ; yn Þ. Value judgements can be introduced
explicitly as restrictions on the behaviour of this mapping that determine its
functional form (e.g. symmetry, quasi-concavity, homogeneity). The social welfare
function discussed above is just a particular case of this more general formulation.
Chapter 6 will be devoted to the analysis of inequality and welfare from this
more general perspective.
4.2 Dalton’s Approach to Inequality Measurement
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
Under these assumptions, social welfare can be measured by a utilitarian social

welfare function, that is, for a given income distribution, y ¼ ðy1 ; y2 ; . . . ; yn Þ, we have:
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
Clearly, D ¼ 0 when there is perfect equality, but it is hard to give further

properties of this index as it depends on the specific form of individual utilities.
In particular, changes in the representation of utilities (i.e. cardinal transformations
that represent the same utilities) will modify the value of the inequality index,
which is a flaw in the design of this indicator.
This can be illustrated when considering the effect of a Dalton’s transfer of size δ
from a rich individual i to poorer one j. The impact of the transfer on inequality is
given by:
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.
2 3
1 4∂uðyi Þ ∂u yj
δ 5 ð4:3Þ
nuðμÞ ∂yi ∂yj
Concavity ensures the negativity of this expression, as yi > yj implies that

∂uðyi Þ ∂uðy Þ
∂yi
< ∂y j . Yet the magnitude of this change depends on the specification of
j
the utility function.

A similar problem happens when considering changes in the units in which
income is measured: nothing ensures that the property of scale independence is
satisfied.
In summary, Dalton introduces a powerful approach to link inequality measure-
ment with social welfare measurement. Yet his inequality index suffers from some
conceptual inconsistencies, as his measure is not invariant with respect to linear
transformations of the utility functions involved (even though the ranking of
income distributions does not suffer from this problem).
4.3 Atkinson Inequality Measures
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
The simplest restriction on marginal utilities in this context is assuming constant

elasticity of the marginal utility. In this way, we can parameterise the inequality
aversion by a single number: the value of the elasticity of marginal utility, ε. This,
obviously, corresponds to the notion of constant relative risk aversion in expected
utility theory (e.g. Pratt, 1964).
The restriction of constant elasticity of the marginal utility translates into utility
functions with the following functional form5:
(
y1ε
uε ð y Þ ¼ ε 6¼ 1 ð4:5Þ
1ε
logy ε¼1
where concavity implies ε > 0.

To see that this is the case, let us take the corresponding first and second
derivatives of this utility function:
8 8 ε1
>
> 1
ε 6¼ 1 > εy
>
∂uε ðyÞ < yε ∂ uε ðyÞ < y2ε ε 6¼ 1
2
¼ 1 , ¼
∂y >
> ∂y2 >
> 1
: ε¼1 : 2 ε¼1
y y
The elasticity of this utility function, for ε 6¼ 1, is thus given by:
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.
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, . . .
4.3.2 The Atkinson Family of Inequality Indices
Atkinson (1970) takes up Dalton’s approach to inequality measurement and applies

some of the ideas developed in the literature on expected utility theory, which uses
cardinal utilities as the backbone (e.g. Pratt, 1964). He exploits the analogy between
risk aversion and that of inequality aversion by using the type of utility functions
with a constant relative Arrow–Pratt coefficient of risk aversion (which coincides
with the elasticity of marginal utility).
There is a key element in that theory that allows Atkinson to circumvent the
inconsistencies of Dalton’s formulation: the notion of certainty equivalent, which in
this context corresponds to the notion of egalitarian equivalent income (also
known as equally distributed egalitarian, EDE, income), which we denote by ξ.
The egalitarian equivalent income is defined as the per capita income such that, if
enjoyed by all individuals in society, it would yield the same social welfare as the
actual income distribution.6 In this way the egalitarian equivalent income is a
money measure of social welfare.
Assuming a utilitarian social welfare function, as in Dalton, the egalitarian
equivalent income, ξ, is given implicitly by the following equation:
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).
Fig. 4.1 Egalitarian yj

equivalent income
Line of perfect
equality
Income agent j
yH
yE
yA
yj
0 yi yi
Income agent i
The concavity of utility functions implies that ξ μ with ξ ¼ μ if and only if

yi ¼ μ for all i. This property allowed Atkinson (1970) to define the following type
of inequality index:
ξ
A¼1 ð4:7Þ
μ
with A ¼ 0, when ξ ¼ μ and A ! 1 as ξ ! 0, which implies that 0 A 1.

The difference μ ξ corresponds to the reduction of the per capita income that
could be realised without affecting social welfare provided total income is equally
distributed. The inequality index A can thus be interpreted as the social cost of
inequality. A value A ¼ 0.3, for instance, means that ξ is 70 % of μ, that is, we could
obtain the present welfare level with only 70 % of total income, by means of an
egalitarian distribution.
This index depends on the reference value ξ, which is determined by the social
welfare function, W. As a consequence, the value of inequality depends, besides the
income distribution vector, on the aggregation formula adopted by the social
welfare function, the specific functional forms of individual utilities (and, in
particular, on the degree of concavity), and the assumption of equal utilities for
all agents.
As already mentioned, Atkinson (1970) assumes the utilitarian social welfare
function, W ðyÞ ¼ Σ i¼1
n
uðyi Þ, and selects a family of utility functions with constant
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
The parameter ε corresponds to the degree of inequality aversion as it

controls the curvature of the social welfare indifference curves. For ε ¼ 0,
there is no concern for inequality, so that total welfare only depends on total
income no matter how it is distributed (a linear social welfare function with
linear indifference curves), so that A ¼ 0 for all distributions. As we increase the
value of ε, social welfare pays progressively more attention to distributive
aspects. The extreme case, when ε ! +1, corresponds to the Leximin social
welfare function that links social welfare to the income of the poorest agent,
W ðyÞ ¼ minyi ðy1 ; . . . ; yn Þ (see Hammond, 1975; Rawls, 1971). Figure 4.2 illus-
trates this case.
From (4.8), we can deduce the explicit form of the egalitarian equivalent
income. For ε 6¼ 1, we have:
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).
Fig. 4.2 Egalitarian yj

Line of perfect
equivalent income with equality
infinite inequality aversion
Income of agent j
yH
yE yA
yj
0 yi yi
Income of agent i
For the case ε ¼ 1, developing the same type of reasoning, we obtain:
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
The Atkinson inequality

index, forε ¼ 1, applied to the example discussed in
Sect. 2.3 yields A1 y1 ¼ 0:471, A1 y2 ¼ 0:458, so the second distribution is more
egalitarian than the first one, with some 10 % difference (in this case the magnitudes
of the index are similar to those of Gini, but the relative difference is similar to that
of the first index of Theil).
Different degrees of social aversion to inequality can be introduced by simply
modifying the parameter ε. Indeed, for a given income distribution, y, Aε is
increasing in ε > 0 (Cowell, 1995). As ε ! +1, Aε ! 1 ðy1 =μÞ (where y1
corresponds to the smallest income). For any (finite) value of this parameter,
Aε ¼ 0 under perfect equality. The upper bound of Aε, corresponding to the case
in which all income is held by a single agent, can only be determined for values
ε
0 < ε 1. For 0 < ε < 1, this upper bound is given by 1 n1ε , which tends to 1 as
n ! 1. For ε ¼ 1, the upper bound is exactly 1. For ε > 1, Aε, this upper bound is
not defined, even though one can show that Aε ! 1 as y ! 0.
The strict concavity of the utility function, ensured by positive values of the
parameter, implies that Atkinson inequality indices satisfy the principle of transfers.
It can be checked that a Dalton’s transfer δ, from an individual i to a poorer
individual j, changes Aε by the amount:
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.
4.3.3 The Atkinson and the Entropy Families of Inequality

Indices
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
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)

Gini Atkinson coefficient Gini/Mean Atkinson/Mean
LIS dataset coefficient (ε ¼ 1) (%) (%)
Taiwan 2010 0.318 0.163 91.3 77.5
United Kingdom 0.338 0.194 97.1 92.2
2010
United States 2013 0.392 0.250 112.6 118.8
Uruguay 2004 0.450 0.300 129.2 142.6
so that, letting θ ¼ 1 ε, we can write I1 ε, using Eq. (3.16) in Chap. 3, as:
1 h i
I 1ε ¼ 1 ð1 Aε Þ1ε ð4:10Þ
εð1 εÞ
This shows that I1 ε is a monotone increasing transformation of Aε for ε > 0.

For ε ¼ 1, L´Hôpital rule applied to (4.10) implies:
I 0 ¼ logð1 A1 Þ ¼ logμ log~

μ ð4:11Þ
So that I0 is a monotone increasing transformation of A1.

We find that Iθ and Aε, for ε > 0 and θ ¼ 1 ε, are ordinally equivalent and,
therefore, they rank all income distributions the same way (even though they will
attach different values which are not linearly related).
The Atkinson family of inequality indices satisfies normalisation, symmetry,
population replication, principle of transfers, differentiability and scale indepen-
dence, but it fails to satisfy additive decomposability.
References
Arrow, K. J. (1951). Social choice and individual values. Primera edici

on. New York: Wiley.
Atkinson, A. B. (1970). On the measurement of inequality. Journal of Economic Theory, 2,
244–263.
Bergson, A. (1938). A reformulation of certain aspects of welfare economics. The Quarterly
Journal of Economics, 52, 310–334.
Blackorby, C., Donaldson, D., & Auersperg, M. (1981). A new procedure for the measurement of
inequality within and among population subgroups. Canadian Journal of Economics, 14,
665–685.
Champernowne, D. G. (1953). A model of income distribution. The Economic Journal, 63(250),
318–351.
Champernowne, D., & Cowell, F. A. (1997). Economic inequality and income distribution.
Cambridge: Cambridge University Press.
Cowell, F. A. (1995). Measuring inequality. LSE handbooks in economics series (2nd ed.). Hemel
Hempstead: Harvester Wheatsheaf.
References 71
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
Inequality indices provide an assessment of the degree of unfairness of outcome

distributions. Those indices measure observed outcomes independently on its
origin. This can be sometimes too crude an approach as part of the observed
outcome differences may be just the product of different people’s choices and
have nothing of unfair. Think for instance of the case of two people and suppose
that one gets twice the income of the other because she works twice the number of
hours. Is that to be considered unfair, provided the two individuals have freely
chosen the number of working hours? The obvious answer is no.
Yet there are observed differences that cannot be regarded as a mere product of
“free choices”. Take again the case of the distribution of labour income in a given
society. The larger differences are mostly related to the level of education of the
workers. One can be tempted to conclude that those differences do not entail
inequality but rather are the product of the differential retribution of skills (e.g. a
surgeon gets better paid than a taxi driver because of the different investment in
human capital involved). This would be fine as long as those people have had the
chance of choosing freely the education obtained. Yet the data show that the family
environment affects the likelihood of achieving high levels of education, which
implies that part of the observed wage dispersion reflects differences in the family
origin and thus involve true inequality.
The methodological approach to take into account the role of free choices and
the role of circumstances is referred to as equality of opportunity, nowadays one of
the most relevant approaches to distributive justice. There is a wide spectrum of
views with respect to what is required for equality of opportunity, from the
non-discrimination viewpoint to the consideration that social provision should
compensate for all forms of disadvantages. Common to those views is that indi-
viduals are accountable, to some extent, for the achievement of the advantage in
question, whether this refers to health, education, income, utility or welfare. Indeed,

74 5 Inequality of Opportunity
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.
5.2 Measuring Inequality of Opportunity by Theil’s

Inequality Index
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
Table 5.1 Labour income Number of agents Characteristics Labour income

and family background
20 Educated Parents 100
(example)
20 Uneducated Parents 60
Table 5.2 Types and effort Type 1 Type 2

groups
Effort groups Size Labour income Size Labour income
1 20 100 20 60
2 20 90 20 50
3 20 75 20 40
4 20 50 20 25
5 20 30 20 10
5.2 Measuring Inequality of Opportunity by Theil’s Inequality Index 77
types. By construction, those groups corrspond to the quintiles of the distribution of

the different types.
In this case, we find two different types, depending on the education of their
parents, and five effort groups, depending on the distribution of labour earnings
within each type. To make things easier, we have designed the example in such a
way that all agents within each cell have exactly the same labour income, both types
have the same number of people and there are exactly five different outcome values
for each type. The key point is that differences withing columns, which correspond
to differences within types, are irrelevant whereas differences within rows, which
correspond to differences between types, express inequality of opportunity.
Let us now formulate the model that permit one to evaluate this type of problem.
5.2.2 The Model
Consider a society M ¼ {1,2, . . ., n} whose agents can be classified into τ types,

t ¼ 1, 2, . . ., τ, depending on their external circumstances. Also assume that we can
distinguish G degrees of effort (e.g. quantiles of the income distribution), which are
indexed by g. We define a cell as a group of agents with the same circumstances that
exert a comparable degree of effort. That is, cell (t, g) is the set of agents of type
t who exert a degree of effort g. We define effort group g as the set of agents in
M who exert a degree of effort g. In Table 5.2, for instance, the cell (1, 2) consists of
seven agents of type 2 who exert the maximum level of effort.
τ
Let yg ¼ [t¼1 yt, g denote the income distribution of effort group g, for
g ¼ 1,2,. . ., G, where yt,g is the income distribution of agents of type t who exert
effort g, i.e. the income distribution of cell (t, g). We denote by ntg the number of
τ
agents in cell (t, g), with ng ¼ Σ t¼1 nt g and nt ¼ Σ g¼1
G
nt g . Similarly, μ(yt, g) is the
mean income of agents in cell (t, g), μ(yg) is the mean income of effort group g and
μ(y) the mean income of the whole distribution.
Remark 5.1 From a theoretical viewpoint, one may well consider that all agents
in a cell should have the same income, as both circumstances and efforts are the
same. This requires assuming the absence of random factors (or that luck is part of
the agents’ external circumstances) and a fine enough partition of types and effort
degrees. Yet, one should expect in practice to observe income differences within
cells due to the imprecise nature of the partition of the population into effort groups
and/or the presence of random factors in the income generation process. We shall
keep here this more realistic formulation.
Note that, by construction, all income differences within an effort group corre-
spond to differences in opportunity, whereas income differences between effort
groups are ethically irrelevant. This implies that when we calculate the inequality of
an income distribution y ¼ ðy1 ; y2 ; . . . ; yn Þ , we are typically overestimating
inequality because we compute all observed inequality, no matter the origin. We
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
Now substituting into (5.2) we get:

5.2 Measuring Inequality of Opportunity by Theil’s Inequality Index 79
" 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
The second component is really our measure of inequality of opportunity, which

is given by the weighted sum of the mean deviations of each cell with respect to its
effort group, in logs. That is, we measure the inequality between the types within
each effort group and then we add up those terms for all groups. Note that this term
is zero when all types within each effort group have the same mean, i.e., when there
is no inequality of opportunity. This is, therefore, the measure of inequality of
opportunity we were looking for. Formally:
" t, g #
XG
ng μ ð yg Þ X τ
ntg μðyt, g Þ μ ðy Þ
IopðyÞ ¼ log ð5:5Þ
g¼1
nμ ð y Þ t¼1
n g μ ð y gÞ μ ð yg Þ
where Iop(y) stands for inequality of opportunity.

We can thus rewrite Eq. (5.3) as follows:
IopðyÞ ¼ T Iop ðyÞ ρðyÞ ð5:6Þ
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.
5.2.3 The Second Index of Theil
Additive decomposability by population subgroups is the key feature of the whole

entropy family of inequality indices. So, in principle, one could apply this type of
reasoning to any of those indices. Yet, let us recall that only the original first and
second indices of Theil admit a decomposition by population subgroups in which
the weights of the different subgroups add up to 1, which makes much easier
interpreting the decomposability. The main difference in this respect between
those two indices is that the first index of Theil uses income shares as weights
whereas the second index of Theil uses population shares.
If we replicate the discussion above, using the second index of Theil, T* [see
Eq. (3.13)], we end up with the following:

1X n
μ ð yÞ
T * ð yÞ ¼ log
n i¼1 yi
The corresponding decomposition is given by [see Eq. (3.15)]:
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
From which we deduce:
X
G
ng
T *Iop ðyÞ ¼ T * ð yg Þ
g¼1
n
Decomposing now each of those terms in the right-hand side, we get:
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
This type of decomposition in which the weights correspond to population

shares is very attractive. Moreover, it satisfies another decomposability property
known as path independence (see Foster & Shneyerov, 1999). Yet the second index
of Theil is somehow less intuitive than the first one and, as pointed out at the end of
the last chapter, it fails to satisfy the normative requirement of minimal equity.
5.3 Equality of Opportunity With Categorical Data3
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).
Table 5.3 Distribution of grades by family types (example)

A B C D E
Q1 0.05 0.2 0.25 0.25 0.25
Q2 0.1 0.25 0.25 0.2 0.2
Q3 0.2 0.25 0.25 0.2 0.1
Q4 0.25 0.3 0.3 0.1 0.05
3
We follow here Herrero and Villar (2014) where an application to the educational equality
of opportunity in Spain is presented, using PISA data.
Inequality of opportunity here is a measure of the differences between those

distributions. In particular, on how different are the values within columns because
it is there that the differences between social classes manifest. Differences within
rows, on the contrary, would reflect the distribution of effort. In other words, if all
rows were identical (all entries within each column are equal), inequality of
opportunity would be zero, even though outcome inequality would be positive
due to differences between columns.
Let us provide now a simple model to measure that type of inequality.
5.3.2 The Model
Consider, as before, a society M ¼ {1,2, . . ., n} whose agents can be classified into τ

types, t ¼ 1, 2, . . ., τ, depending on their external circumstances. In the example of
Table 5.3, those circumstances refer to family income but in real problems may
involve many more elements. The achievements of the individuals of this society
can be classified into a set C ¼ f1, 2, . . . , γ g of different categories (grades in the
example). A cell corresponds to a group of agents with the same circumstances that
have the same achievements. That is, cell (t, c) is the set of agents of type t whose
outcomes are of category c.
We want to assess the equality of opportunity, regarding a given aspect, in a
society made of m individuals of τ different types that may produce outcomes
belonging to one of γ different categories. The main idea is that the observed
differences in the distribution of individual outcomes across types reflect the
different opportunities that people enjoy.
Let ntc denote the number of agents in cell (t, c), i.e., the number of agents of type
τ γ
t with categorical outcome c, with nc ¼ Σ t¼1 nt c and nt ¼ Σ c¼1 nt c . We define now:
ntc
αtc ¼
nt
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
Row t of matrix H describes the outcome distribution of type t and is to be

interpreted as an expression of the differential effort of the agents of that type.
Column c of matrix H describes the distribution of the categorical outcome c across
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
This inequality of opportunity index can also be interpreted as a summary

measure of the differences between the rows of matrix H, which describe the
distribution of outcomes across types.4
Let us measure the dispersion within columns of matrix H by Atkinson’s
inequality index, for the value ε ¼ 1. We shall have:
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).
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
5.3.3 Empirical Illustration: Inequality of Opportunity

in Compulsory Education in the OECD, According
to PISA (2012)
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
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).
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)

Lopp Rel. Iopp (OECD ¼ 100)
Norway 0.0364 59
Poland 0.0814 133
Portugal 0.0943 153
Slovak Republic 0.1103 179
Slovenia 0.0757 123
Spain 0.0775 126
Sweden 0.0510 83
Switzerland 0.0578 94
Turkey 0.0590 96
United Kingdom 0.0652 106
United States 0.0761 124
OECD 0.0615 100
Source: OECD (2014a, b)
Appendix
Table 5.7 Distribution of students by levels of proficiency and quartiles of the ESCS for the
OECD (PISA 2012)
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)

Czech Republic Q1 0.0371 0.1130 0.2135 0.2614 0.2315 0.1435
Q2 0.0784 0.1734 0.2775 0.2560 0.1502 0.0644
Q3 0.1223 0.2212 0.2779 0.2186 0.1233 0.0368
Q4 0.2825 0.2822 0.2220 0.1309 0.0599 0.0224
Denmark Q1 0.0241 0.1196 0.2363 0.3194 0.2144 0.0862
Q2 0.0508 0.1782 0.3168 0.2757 0.1392 0.0392
Q3 0.1136 0.2160 0.3336 0.2300 0.0861 0.0208
Q4 0.2183 0.2906 0.2862 0.1480 0.0456 0.0113
Estonia Q1 0.0762 0.1773 0.3136 0.2744 0.1265 0.0321
Q2 0.1010 0.2198 0.3003 0.2492 0.1071 0.0226
Q3 0.1424 0.2506 0.3026 0.2150 0.0743 0.0150
Q4 0.2722 0.2978 0.2600 0.1374 0.0286 0.0039
Finland Q1 0.0807 0.1667 0.2878 0.2636 0.1358 0.0653
Q2 0.1086 0.2150 0.3065 0.2388 0.1060 0.0251
Q3 0.1583 0.2757 0.3004 0.1815 0.0644 0.0197
Q4 0.2709 0.2816 0.2687 0.1327 0.0388 0.0073
France Q1 0.0334 0.0845 0.2027 0.2764 0.2255 0.1774
Q2 0.0720 0.1528 0.2460 0.2658 0.1645 0.0989
Q3 0.1339 0.2239 0.2774 0.2255 0.0960 0.0433
Q4 0.2904 0.3101 0.2398 0.1131 0.0336 0.0129
Germany Q1 0.0661 0.1408 0.2191 0.2629 0.2033 0.1078
Q2 0.1284 0.1946 0.2762 0.2159 0.1303 0.0546
Q3 0.2201 0.2884 0.2414 0.1558 0.0732 0.0210
Q4 0.3597 0.2821 0.1981 0.1007 0.0395 0.0198
Greece Q1 0.0035 0.0436 0.1456 0.2742 0.2836 0.2494
Q2 0.0180 0.0792 0.2080 0.2848 0.2467 0.1633
Q3 0.0308 0.1212 0.2378 0.3030 0.1980 0.1092
Q4 0.1054 0.2063 0.2949 0.2266 0.1191 0.0477
Hungary Q1 0.0153 0.0577 0.1533 0.2682 0.2939 0.2116
Q2 0.0476 0.1256 0.2283 0.2996 0.2028 0.0961
Q3 0.0863 0.1515 0.2655 0.2784 0.1610 0.0573
Q4 0.2260 0.2455 0.2790 0.1687 0.0637 0.0171
Iceland Q1 0.0555 0.1311 0.2284 0.2720 0.1984 0.1146
Q2 0.0723 0.1519 0.2693 0.2732 0.1588 0.0745
Q3 0.1470 0.2007 0.2548 0.2279 0.1196 0.0501
Q4 0.1818 0.2515 0.2836 0.1722 0.0757 0.0352
Ireland Q1 0.0343 0.1129 0.2507 0.3050 0.1982 0.0990
Q2 0.0725 0.1664 0.2967 0.2683 0.1440 0.0521
Q3 0.1155 0.2316 0.3123 0.2198 0.0917 0.0292
Q4 0.2091 0.3061 0.2760 0.1612 0.0417 0.0059
Israel Q1 0.0160 0.0553 0.1423 0.2287 0.2670 0.2908
Q2 0.0575 0.1085 0.2033 0.2679 0.2043 0.1585
Q3 0.1158 0.1855 0.2454 0.2176 0.1399 0.0959
Q4 0.1991 0.2456 0.2558 0.1552 0.0926 0.0516
(continued)

Italy Q1 0.0346 0.0983 0.2027 0.2808 0.2358 0.1478
Q2 0.0757 0.1482 0.2440 0.2577 0.1811 0.0933
Q3 0.1173 0.1916 0.2675 0.2295 0.1347 0.0594
Q4 0.1718 0.2364 0.2735 0.1938 0.0893 0.0352
Japan Q1 0.1181 0.2047 0.2585 0.2282 0.1304 0.0601
Q2 0.1882 0.2450 0.2717 0.1892 0.0824 0.0235
Q3 0.2780 0.2494 0.2586 0.1389 0.0544 0.0208
Q4 0.3812 0.2617 0.2020 0.1098 0.0367 0.0086
Korea Q1 0.1634 0.2188 0.2676 0.2102 0.0934 0.0465
Q2 0.2431 0.2409 0.2379 0.1680 0.0807 0.0295
Q3 0.3508 0.2549 0.1960 0.1301 0.0484 0.0198
Q4 0.4815 0.2421 0.1556 0.0759 0.0321 0.0128
Luxembourg Q1 0.0229 0.0877 0.1863 0.2780 0.2515 0.1736
Q2 0.0610 0.1320 0.2494 0.2709 0.1955 0.0911
Q3 0.1298 0.2187 0.2785 0.2115 0.1124 0.0491
Q4 0.2381 0.3098 0.2358 0.1362 0.0541 0.0260
Mexico Q1 0.0012 0.0107 0.0657 0.2152 0.3568 0.3504
Q2 0.0020 0.0248 0.1156 0.2760 0.3405 0.2411
Q3 0.0043 0.0343 0.1357 0.3003 0.3240 0.2014
Q4 0.0178 0.0798 0.2111 0.3251 0.2538 0.1124
The Netherlands Q1 0.0842 0.1655 0.2560 0.2456 0.1822 0.0665
Q2 0.1366 0.2352 0.2717 0.2027 0.1212 0.0326
Q3 0.2229 0.2863 0.2362 0.1483 0.0766 0.0298
Q4 0.3425 0.2794 0.2077 0.1107 0.0442 0.0156
New Zealand Q1 0.0430 0.0875 0.1800 0.2796 0.2537 0.1561
Q2 0.1053 0.1811 0.2548 0.2419 0.1542 0.0626
Q3 0.1475 0.2237 0.2742 0.2031 0.1164 0.0350
Q4 0.3234 0.2465 0.2123 0.1382 0.0626 0.0170
Norway Q1 0.0468 0.1241 0.2139 0.2798 0.2236 0.1119
Q2 0.0645 0.1630 0.2626 0.2671 0.1659 0.0769
Q3 0.1170 0.2044 0.2764 0.2293 0.1271 0.0458
Q4 0.1572 0.2526 0.2832 0.1900 0.0795 0.0376
Poland Q1 0.0577 0.1327 0.2506 0.2940 0.1935 0.0715
Q2 0.1014 0.2095 0.2632 0.2620 0.1322 0.0316
Q3 0.1750 0.2308 0.2751 0.2124 0.0846 0.0221
Q4 0.3399 0.2824 0.2251 0.1146 0.0316 0.0064
Portugal Q1 0.0291 0.0873 0.1959 0.2657 0.2542 0.1678
Q2 0.0620 0.1493 0.2452 0.2733 0.1907 0.0794
Q3 0.0967 0.1902 0.2766 0.2416 0.1347 0.0602
Q4 0.2455 0.2912 0.2572 0.1354 0.0495 0.0212
(continued)

Slovak Republic Q1 0.0237 0.0690 0.1488 0.2414 0.2512 0.2661
Q2 0.0598 0.1364 0.2532 0.2857 0.1890 0.0758
Q3 0.1113 0.1923 0.2546 0.2401 0.1363 0.0652
Q4 0.2515 0.2663 0.2344 0.1534 0.0738 0.0206
Slovenia Q1 0.0436 0.0985 0.2207 0.3030 0.2459 0.0883
Q2 0.0945 0.1647 0.2398 0.2701 0.1732 0.0577
Q3 0.1398 0.2213 0.2559 0.2227 0.1232 0.0372
Q4 0.2764 0.2686 0.2443 0.1427 0.0544 0.0136
Spain Q1 0.0250 0.0842 0.1988 0.2954 0.2458 0.1509
Q2 0.0459 0.1402 0.2643 0.2809 0.1901 0.0785
Q3 0.0802 0.1994 0.2892 0.2552 0.1244 0.0516
Q4 0.1728 0.2855 0.2938 0.1652 0.0655 0.0172
Sweden Q1 0.0248 0.0906 0.1967 0.2872 0.2486 0.1522
Q2 0.0521 0.1551 0.2366 0.2687 0.1863 0.1012
Q3 0.0940 0.1914 0.2711 0.2434 0.1469 0.0533
Q4 0.1581 0.2321 0.2717 0.2003 0.1073 0.0305
Switzerland Q1 0.0918 0.1821 0.2561 0.2417 0.1518 0.0765
Q2 0.1560 0.2485 0.2626 0.1989 0.0972 0.0368
Q3 0.2350 0.2714 0.2528 0.1588 0.0640 0.0180
Q4 0.3795 0.2589 0.2102 0.1049 0.0380 0.0085
Turkey Q1 0.0135 0.0434 0.1214 0.2529 0.3253 0.2435
Q2 0.0281 0.0769 0.1658 0.2716 0.2938 0.1637
Q3 0.0537 0.0892 0.1650 0.2756 0.2793 0.1372
Q4 0.1416 0.1955 0.2111 0.2246 0.1576 0.0697
United Kingdom Q1 0.0398 0.1162 0.2360 0.2876 0.2004 0.1199
Q2 0.0697 0.1474 0.2618 0.2657 0.1682 0.0872
Q3 0.1257 0.2230 0.2652 0.2271 0.1112 0.0478
Q4 0.2520 0.2665 0.2443 0.1530 0.0640 0.0202
United States Q1 0.0250 0.0816 0.1868 0.2964 0.2648 0.1454
Q2 0.0384 0.1228 0.2261 0.3057 0.2197 0.0873
Q3 0.0966 0.1854 0.2601 0.2562 0.1531 0.0486
Q4 0.1966 0.2502 0.2657 0.1928 0.0676 0.0272
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
Source: OECD (2014a, b)
*
Levels 5 and 6 have been aggregated into a single one denoted by 5*
References 91
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.
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
The normative approach to inequality is very attractive because it permits giving an

explicit ethical content to its measurement, by interpreting inequality as a welfare
loss that is captured by specific social welfare functions. This approach also pro-
duces inequality indices that fit well with our intuition of what is inequality, both in
terms of dispersion and the concern for the less well off. Think for instance of the
case of Atkinson’s measure for ε ¼ 1 that corresponds to the normalised difference
between the arithmetic and the geometric mean (an idea that reminds of that
measure of the asymmetry in a distribution that consists of the ratio between the
mean and the median). We have also shown how the Atkinson family of indices
avoids some of the conceptual problems present in Dalton’s original formulation
and provides a way of controlling the degree of inequality aversion by means of a
single parameter.
This way of implementing the normative approach to inequality, though, has
some aspects that are subject to criticisms. We shall consider next four of those
arguable aspects.
The first critique is simple and clear: if individual welfare is the basis for the
normative evaluation of inequality, why focus our attention on income distributions
and not on welfare distributions? Note that the distribution of individual utilities,
n n
fuðyi Þgi¼1 , and the distribution of individual incomes, fyi gi¼1 , might be rather
different depending on the degree of concavity of the utility function (and even
more if we allow for different utilities for different individuals and the dependence
of the individual utility on the whole distribution vector). In other words, there is no
reason to think that income distribution is a good proxy of welfare distribution.
The second criticism refers to the fact that the measurement of inequality
strongly depends on the form of individual utilities, which are chosen in a rather
arbitrary way. Changing those utilities may change drastically the inequality
measure and we may not have very good reasons to choose between different utility

94 6 Inequality and Welfare
specifications. Even if we confine our choices to the family of constant inequality

aversion utilities, as proposed by Atkinson, we still have the problem of deciding on
the value of the parameter of inequality aversion, which is not a simple choice.
The third criticism refers to the assumption that all agents have identical utilities.
This assumption is key to get a non-dictatorial social welfare function, because it
entails a trivial form of interpersonal comparability. It is also essential to make the
egalitarian distribution the social optimum (i.e. in this case utilitarianism is fully
egalitarian!). Yet this assumption is far from being natural or justified. When
utilities differ between agents, we have to face the problem of defining interper-
sonally comparable utilities or fall into Arrow’s impossibility theorem.
Finally, let us point out a more technical aspect. In general, there is no one-to-
one correspondence between inequality indices and social welfare functions. This
problem can be avoided by assuming the homogeneity of the social welfare
function, and this depends in turn on the choice of individual utilities and the
aggregation rule [see the discussion in Blackorby and Donaldson (1978) and
Dutta and Esteban (1992)].
To avoid some of these problems, Sen (1973) proposes a slight reformulation of
the Dalton–Atkinson approach, based on the consideration of a more general
welfare evaluation, that we shall denominate Social Evaluation Function (SEF,
for short). This is a function V defined directly on the space of income distributions,
without going through individual utilities. The idea is that we can find reasonable
restrictions on this evaluation function that help finding ways of measuring inequal-
ity and welfare in a less arbitrary way. We shall show in this chapter that this
approach greatly facilitates the analysis of inequality in terms of its welfare bases.
6.2 Social Evaluation Functions
A Social Evaluation Function is a mapping V : ℝþþ n

! ℝ that associates real
numbers to income distributions, in the understanding that higher values describe
better distributions. That is, V ðyÞ ¼ V ðy1 ; y2 ; . . . ; yn Þ. One can then introduce the
pertinent value judgements as restrictions on the functional form of this mapping.
Sen assumes two basic properties1:
• Symmetry: Permuting the agents’ incomes does not change the welfare evalua-
tion. That is, if y0 ¼ π ðyÞ, where π(y) represents a permutation of the elements of
y, then V ðy0 Þ ¼ V ðyÞ.
• Strict quasi-concavity: 8y 6¼ y0 2 ℝþþ n
, 8λ 2 ð0; 1Þ,
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
V ðλy þ ð1 λÞy0 Þ > minfV ðyÞ, V ðy0 Þg
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
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
Then, it follows from equation (6.1) that:
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
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
where αi ðyÞ ¼ ∂V=∂yi . This is an immediate consequence of Euler’s Theorem for

homogeneous functions. The interest of this equation is that it expresses the welfare
measure of an income distribution as a weighted sum of individual incomes. Then it
is easy to introduce value judgements in terms of the weighting system. This
formulation corresponds to Sen’s (1976) personalised goods approach. Indeed
those weights control the shape of welfare indifference curves.
To summarise this discussion, let us formulate the following result:
Theorem 6.1 A social evaluation function V : ℝþþ n
! ℝ that satisfies the prop-
erties of symmetry, strict quasi-concavity, differentiability, homogeneity of degree
one and scale can be expressed as:
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
V G ðyÞ kμðyÞ½1 GðyÞ ð6:6Þ
where k is a positive constant and G is the Gini index .

Proof2
Making use of the definition in equation (3.7), we can write:
1 2
G ð yÞ ¼ 1 þ ½ny þ ðn 1Þy2 þ ðn 2Þy3 þ þ 2yn1 þ yn
n μðyÞn2 1
nþ1 2 X n
¼ ðn þ 1 iÞyi
n μðyÞn2 i¼1
nþ1 2 X n
¼ α G ðyÞyi
n μðyÞn i¼1 i
Therefore,
X
n
μðyÞn n þ 1
αiG ðyÞyi ¼ GðyÞ
i¼1
2 n
that can be rewritten as:

X
n
αiG ðyÞyi kμðyÞð1 GðyÞÞ
i¼1
where k ¼ n2 and we let nþ1

n approximately equal to 1 (assuming a large n).
Q.E.D.
These coefficients, which correspond to those in equation (3.7) divided by n,
give weights to the agents that are inversely proportional to their ranking. That is,
2
We assume implicitly that all incomes are different to avoid problems with differentiability
in the SEF.
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:
υðyÞ ¼ μðyÞ½1 T ðyÞ ð6:7Þ
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.
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:
∂αiT 1 μ ∂ðyi =μÞ

¼ <0
∂yi n yi ∂yi
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
Substituting the last expression into (6.7) we get:
υðyÞ ¼ μð1 T ðyÞÞ

X
G
ng μg
¼μ μ T ðy Þ þ log
g
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
So that the social evaluation of income distribution y can be expressed as the

sum of two different components,
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.
6.3 Multidimensional Inequality and Welfare
Personal income is a variable that captures a substantial part of economic well-

being because it is a good proxy of consumption opportunities and exhibits a high
correlation with several aspects of individual welfare (e.g. health, occupation,
social status). Yet not all relevant welfare dimensions can be properly represented
by a single variable as that correlation is far from perfect. A multidimensional
approach is thus a way of enriching our assessments of welfare and inequality. We
shall briefly refer here to this more complex approach.
The need for multidimensional welfare indicators has been recognised for many
years now. Indeed, there have been proposals in that sense since the very beginning
of the adoption of the GDP as a measure of economic achievement.3 Adopting a
multidimensional approach to well-being is usually addressed by looking for a
suitable real-valued function that combines in a sensible way all those dimensions
and produces a scalar measure. This implies defining a mapping from the space of
social realisations (now a matrix rather than a vector) into the real numbers [see
Foster, Lopez-Calva, and Szekely (2005), Alkire and Foster (2010), Herrero,
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).
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).
(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”.
6.3.1 Multidimensional Welfare Indicators
Consider a society N consisting of n individuals, N ¼ f1; 2; . . . ; ng, and suppose we

want to assess its global performance as a function of the achievements of its
members with respect to a set K ¼ f1; 2; . . . ; kg of different dimensions. A social
state is a matrix Y with n rows (one for each individual) and k columns (one for
each dimension) that describes the achievements of the members of this society in
all dimensions. That is, the element yij of matrix Y describes the value of the
variable j for individual i. We assume implicitly that all dimensions are continuous
quantitative variables, whereas the case of qualitative or subjective variables will
not be addressed here.
Let Ω denote the space of all finite-dimensional n k matrices and consider the
following:
Definition A Multidimensional Social Evaluation Function is a continuous
single-valued mapping W : Ω ! ℝ that provides a numerical evaluation of social
states.
A multidimensional social evaluation function maps social states into real
numbers, in the understanding that higher values of this index correspond to better
social states. Note that we keep here the approach of defining welfare directly on the
matrix of social outcomes rather than on the vector of agents’ utilities.
All the properties presented in Sect. 6.2 are immediately extended to this
context. That is:
• Symmetry: For all Y 2 Ω, let π R(Y) denote a permutation of the rows of matrix
Y. Then, W ðπ R ðYÞÞ ¼ W ðYÞ.
• Strict quasi-concavity: For all Y 6¼ Y0 2 Ω, all λ 2 ð0; 1Þ,
W ðλY þ ð1 λÞY0 Þ > minfW ðYÞ, W ðY0 Þg
• Differentiability: Function W has continuous derivatives.

0 of degree one:1For all Y 2 Ω, all λ 2 ð0; 1Þ, W ðλYÞ ¼ λW ðYÞ.
• Homogeneity
• Scale: W @1n q, 1n q, . . . , 1n q A ¼ q (where 1n is the unit vector of dimension n).

|fflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflffl}
k times
5
Some key contributions include Maasoumi (1986, 1989), Bourguignon and Chakravarty (1999),
Tsui (1995, 1999), Weymark (2004), Gajdos and Weymark (2005).
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

W ðYÞ ¼ W Yj , 1n ye Yj , yðjÞ
where 1n is the unit vector of dimension n.

When the social evaluation function W satisfies also separability, the egalitarian
equivalent value ye Yj , yðjÞ turns out to be independent of Yj , so that we can
simply write yej as a short-handed version of ye(y( j)). As it was the case before,
homogeneity and continuity ensure that these values are unique and well defined.
The specific cases of additive and multiplicative separability allow us to express
the welfare evaluation function as:
8 Xk
< a j W 1 n y e
j ðadditive separabilityÞ
W ðYÞ ¼ Y kj¼1 aj ð6:11Þ
: W 1n y e ðmultiplicative separabilityÞ
j¼1 j
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.
6.3.2 Multidimensional Inequality
Adopting the approach described in Sect. 6.3.1 makes it immediate to construct

multidimensional inequality indices, by simply deciding on the proper value of the
egalitarian equivalent variables (see also Gajdos and Weymark (2005)).
Extending the idea in (6.1) to a multidimensional context, we can formulate a
multidimensional inequality index as a function I : Ω ! ℝ given by:
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).
Under the assumption of additive or multiplicative separability, and bearing in

mind that we can write yje ¼ μj ð1 I ðyðjÞÞÞ for each single dimension, we have:
(a) Additive separability:
Xk
a μ I ðyðjÞÞ
j¼1 j j
I ðY Þ ¼ Xk ð6:13aÞ
aμ
j¼1 j j
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
Assuming multiplicative separability, instead, from (6.13b) we get:

Y k aj
j¼1
μ
ej
I Að1Þ ðYÞ ¼ 1 X k ð6:14bÞ
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.
6.3.3 Weights and Units
All multidimensional inequality and welfare formulae involve some unexplained

weighting system that reflects the relative importance of each dimension, as those
values determine the corresponding rates of substitution. Note that the larger the
number of dimensions, the more important the weighting system becomes for the
overall evaluation, whereas a small number of dimensions will yield, in general,
evaluations that are more robust.
How should we select those weights? There is no clear answer to that question,
among other things because it depends very much on the type of problem at hand.
We can think of four basic ways of dealing with this difficulty, which we comment
briefly.6
The first and simplest one is opting by an equal weighting, provided all dimen-
sions can be regarded a priori as of similar importance. In the absence of a criterion
to discriminate them, choosing aj ¼ 1=k for all j is a reasonable compromise. This is
usually the case when there are few dimensions involved and they all refer to very
basic aspects of well-being (e.g. the Human Development Index, discussed below).
The second way of choosing those weights is by giving them some explicit
normative content. This is the case, for instance, in the recent Index of Green
Economy Progress (UNEP 2017) where the weights incorporate initial values and
some relevant thresholds regarding the ecological impact of some variables.
The third way is obtaining some consensus either among the specialists or
between the users or the subjects of those evaluation protocols (e.g. one of the
strategies followed by the OECD in the aggregation of indicators that constitute the
How’s Life dashboard).
Finally, there are some statistical methods that permit one deciding on the
weights by “letting the data speak”. Yet the welfare basis of this approach is not
very clear (see Aaberge and Brandolini (2014, pp. 23–25) for a discussion).
In summary, agreeing upon a particular weighting scheme is not an easy task. It
is therefore prudent to check the robustness of our evaluations by comparing the
results under alternative weighting systems. Different tools for sensitivity and
robustness analyses have been developed recently in order to test the robustness
of rankings generated by the composite indices [see Nardo et al. (2008), Cherchye,
Ooghe, and Puyenbroeck (2008), Permanyer (2011), Foster, McGillivray, and Seth
(2009, 2013)].
A related question refers to the choice of units. In the single-dimensional case,
scale independence ensures that the units in which we measure our variable do not
affect the value of the inequality index. Things are different in the multidimensional
case because the mean values of the different variables are involved and thus affect
the impact of the different dimensions on the overall evaluation. Think of the case
of income, measured in dollars, and health, measured in years of life expectancy.
6
For a discussion on different techniques for setting weights for multidimensional indices, see
Decancq and Lugo (2013).
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
6.4 The Human Development Index
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.
using distributive sensitive indices, in 2010 a companion index was introduced in

which inequality plays its part. It corresponds to the welfare index that derives from
assuming multiplicative separability and takes yje ¼ μj ½1 A1 ðyðjÞÞ, where A1 is
Atkinson’s inequality index, for ε ¼ 1. The Inequality-adjusted Human Develop-
ment Index can be written, after some elementary algebra, as:
Y Ync 1=3nc
IHDI c ¼ j¼H, E, Y i¼1
yij ð6:18Þ
This conceptual scheme is modified in the actual calculation by two variations

that make it less convincing.
First, the income variable is measured in logs rather than in natural units. The
justification is that the welfare impact of an additional dollar is smaller the higher
the per capita GDP. This sounds convincing except that: (i) The same argument
could be applied to an additional year of expected life or an additional year of
schooling.8 (ii) If the log of the income is the relevant variable, the inequality
measure should be applied to the distribution of the logs of the incomes, which is
not the case.
Second, each variable is normalised in terms of relative gains, as in (6.16), so
that all the problems associated with this normalisation are present. We refer to
Herrero, Martı́nez, and Villar (2012) and Seth and Villar (2016) for a critical
discussion of those and other related aspects.
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 economıá 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
Osberg, L. (1985). The measurement of economic weil-being. In D. Laidler (Ed.), Approaches to

economic wellbeing (Vol. 26, p. 49). Toronto: University of Toronto Press.
Osberg, L., & Sharpe, A. (2002). An index of economic wellbeing for selected OECD countries.
Review of Income and Wealth, 48, 291–316.
Permanyer, I. (2011). Assessing the robustness of composite indices rankings. Review of Income
and Wealth, 57, 306–326.
Phillipson, T., & Soares, R. (2001). Human capital, longevity, and economic growth, mimeo.
University of Chicago.
Ram, R. (1982). Composite indices of physical quality of life, basic needs fulfillment, and income:
A ‘principal component’ representation. Journal of Development Economics, 11, 227–247.
Sagar, A. D., & Najam, A. (1999). The human development index: A critical review. Ecological
Economics, 25, 249–264.
Savaglio, E. (2006). Multidimensional inequality with variable population size. Economic Theory,
28(1), 85–94.
Sen, A. (1973). On economic inequality. Oxford: Oxford University Press.
Sen, A. K. (1976). Real national income. Review of Economic Studies, 43(1), 19–39.
Sen, A. K. (1985). Commodities and capabilities. Oxford: Oxford University Press.
Seth, S. (2013). A class of distribution and association sensitive multidimensional welfare indices.
Journal of Economic Inequality, 11(2), 133–162.
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.
Spence, R. (2009). Economic growth. In S. Deneulin & L. Shahani (Eds.), An introduction to
human development and capability approach. London: Earthscan.
Stiglitz, J. E., Sen, A. K., & Fitoussi, J-P (2009). Report by the commission on the measurement of
economic performance and social progress.
Tsui, K. Y. (1995). Multidimensional generalizations of the relative and absolute inequality
indices: The Atkinson-Kolm-Sen approach. Journal of Economic Theory, 67(1), 251–265.
Tsui, K. Y. (1999). Multidimensional inequality and multidimensional generalized entropy mea-
sures: An axiomatic derivation. Social Choice and Welfare, 16(1), 145–157.
Tsui, K. (2002). Multidimensional poverty indices. Social Choice and Welfare, 19, 69–93.
UNEP. (2017). The Green Economy Progress Measurement Framework. Geneva: UNEP.
United Nations Development Programme. (2010). Human development report. New York: Oxford
University Press.
Weymark, J. A. (2004). The normative approach to the measurement of multidimensional inequal-
ity (Working Paper No. 03-W14R). Vanderbilt University.
Zambrano, E. (2014). An axiomatization of the human development index. Social Choice and
Welfare, 42, 853–872.
Part II
Poverty
Chapter 7
Poverty Measurement
7.1 Introduction
Poverty is a complex phenomenon that refers to the difficulty of having access to

those goods and services that ensure living with dignity and developing a satisfac-
tory social and personal life. Insufficient income is, no question, one of the key
determinants of poverty. Yet it is not the only one nor it is able to capture all aspects
that poverty entails. Be as it may, note that all the elements mentioned above are as
intuitive as difficult to precise and measure. Even in the simplest case, where
poverty is identified with insufficient income, it is not clear how to define “insuf-
ficient”. So in a way poverty presents a more difficult field of study than inequality,
even though both involve concerns about the fairness of income distributions.
Indeed, poverty measurement can be addressed in a number of ways, which derive
from some basic methodological choices that are often connected.
To start with we have to decide focussing on one or several dimensions of
poverty. Even though the single-dimensional approach has been predominant, a
good deal of research has been done lately on multidimensional poverty. The basic
idea behind this approach is that money measures of poverty fail to capture some
key aspects of poverty, especially people’s deprivation of some essential goods and
services with low or no substitutability beyond some small enough values (current
water, medical services, access to education, etc.).
Second, we can consider poverty from a subjective or from an objective per-
spective. The subjective poverty approach takes as the main ingredient the agents’
perception regarding their situation. The objective poverty approach, on the con-
trary, deals with poverty in terms of some observable variables (e.g. income, access
to consumption goods). In both cases, one has to choose those variables that better
represent the notion of poverty we want to measure (e.g. disposable income) and
units of reference (e.g. consumption units).
Third, we have to define a criterion to determine who are the poor, as this will be
the population target of the measurement exercise. In the case of objective poverty,

116 7 Poverty Measurement
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.
7.2 Poverty Indices
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
where P[yq, z] is the poverty index of subgroup q when considered as an indepen-

dent society and nq/n is the corresponding population share.
Some of those properties have already been discussed when analysing inequality
indices. The ones that are specific of poverty measurement are those of Focus,
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).
Monotonicity and Decomposability by population subgroups. Focus establishes

that our measure does not depend on what happens to those agents with incomes
above the poverty threshold. So the index is only sensitive to part of the distribution.
Monotonicity plays here the role of the principle of transfers and tells us that
poverty increases when the income of the poor decreases, other things equal.
Decomposability by population subgroups establishes that the poverty of a group
is independent on the poverty of other groups. As in the case of inequality indices,
decomposability refers to the behaviour of the poverty measure when society can be
regarded as composed of a series of population subgroups, defined by some
sociological or demographic characteristic. Here again this property permits one
to describe the society’s overall poverty as a weighted sum of the poverty of the
constituent subgroups. From a mathematical point of view, this property imposes
relevant restrictions on the admissible poverty measures.
Decomposability by population subgroups is an arguable property that never-
theless is widely accepted because of its operational advantages. In particular, by
subdividing each subgroup until we have single individuals we have:
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.
7.2.1 Head Count Ratio
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
where p is the number of the poor and n the total population.3
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.
This index only takes into account the incidence of poverty and disregards the
other two aspects, intensity and inequality.
7.2.2 Poverty Gap Measures
Consider the following income distributions:
y ¼ ð 10, 10, 30, 50 Þ, y0 ¼ ð 2, 2, 30, 50 Þ
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.
7.2.3 The Sen’s Family of Poverty Measures
Consider now the following income distributions:
y ¼ ð 0, 14, 30, 50 Þ, y0 ¼ ð 7, 7, 30, 50 Þ
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.
decomposability can be defined when the inequality index satisfies additive

decomposability.
7.2.4 Decomposable Poverty Measures
The property of decomposability by population subgroups is relevant when we are

interested in knowing the impact of poverty depending on some demographic or
sociological characteristics, such as age, sex, level of studies, region, etc.
Probably, the best-known family of decomposable poverty indices is that known
as the FGT family, following the work of Foster, Greer, and Thorbecke (1984).
This family of indices is defined by a function that depends on a parameter α that
can be interpreted as the degree of poverty aversion. It is obtained from the
following individual poverty measure, which gives us the relative difference
between the poverty threshold and the agent’s income, to the power of α (provided
yi < z, and it is zero otherwise):
n z y α o
gðyi ; zÞ ¼ max 0; i
z
We then define the FGT poverty index as follows:
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
The case α ¼ 2 corresponds to what is usually known as the severity of poverty

and is closely linked to the square of the coefficient of variation, substituting the
mean value by the poverty threshold. That is,5
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
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
Another interesting decomposable poverty index is that proposed by Watts

(1968), which is related to Theil’s inequality index:

1X n
z
PW ðy; zÞ ¼ max 0, ln ð7:9Þ
n i¼1 yi
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
where e 2 ð0, 1Þ is a parameter of poverty aversion.

Remark 2 Note that the notion of incidence of poverty appears in those formulae
in an indirect way, by summing over p (the number of the poor) individual poverty
gaps, while dividing by n (the population size). The idea o inequality is much
difficult to precise but it can be partially captured by the weights attached to
those individual gaps or ratios (or the recourse of logs in the case of Watts’).
7.3 Multidimensional Poverty Indices
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,
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
The family of FGT in this context adopts the following format:

1X K X
yij α
p
α
PFGT ðY; z; bÞ ¼ bj 1 ð7:13Þ
n j¼1 i¼1 zj
The extension of Watts’ poverty index is given by:

!
1X K X p
zj
PW ðY; z; bÞ ¼ bj log ð7:14Þ
n j¼1 i¼1 yij
Another well-established multidimensional poverty index is that introduced by

Tsui (2002), given by:
2 !bj 3
1X K Y
4 p zj
PTsui ðY; z; bÞ ¼ 15 ð7:15Þ
n j¼1 i¼1 yij
Remark 3 Many other multidimensional indices do exist. The reader is referred to

the work of Chakarvarty (2009, Ch. 2) for a discussion.
7.4 Deprivation and Non-monetary Poverty Measures 125
7.4 Deprivation and Non-monetary Poverty Measures
Results on multidimensional poverty using variables different from income, some-

times referred to as multidimensional deprivation, appear in the late 70s, when a
series of studies were carried out trying to measure poverty and social exclusion in a
much broader sense, by introducing non-monetary indicators (see Townsend 1979;
Mack and Lansley 1985; Callan et al. 1993). In recent years, different international
institutions have developed statistical databases in order to provide measures of
multidimensional deprivation. It is easy to understand that those measures will very
much depend on the degree of development of the countries. Indeed, non-monetary
poverty measures may be relevant for two different reasons. First, because there are
aspects of poverty that are hardly captured by monetary indices. Second, because in
the less developed countries there is a relevant part of consumption goods and
services that are non-marketed.
We present here two different approaches to multidimensional non-monetary
measurement. The first one corresponds to the European Union definition of people
at risk of poverty or social exclusion (the AROPE indicator), which is conceived to
measure deprivation in the context of highly developed countries. The second one is
the United Nations’ Multidimensional Poverty Index, designed to measure depri-
vation in less developed countries, very much anchored in the collective research
work developed in the Oxford Poverty and Human Development Initiative (OPHI).
7.4.1 At Risk of Poverty or Social Exclusion (AROPE)
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
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.
Fig. 7.1 AROPE index in Europe (2011–2012)
7.4.2 The United Nations Multidimensional Poverty Index8
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)
Table 7.1 AROPE in Europe

Geo/Time 2005 2006 2007 2008 2009 2010 2011 2012 2013
European Union : : : : : 23.7 24.3 24.7 24.5
(28 countries)
European Union 25.7 25.3 24.4 23.8 23.3 23.6 24.2 24.7 24.5
(27 countries)
European Union 21.6 21.9 21.6 21.7 21.4 21.8 22.6 23.1 23.1
(15 countries)
New Member States 41.0 38.0 35.0 31.7 30.6 30.8 30.6 30.7 30.0
(12 countries)
Euro area (18 countries) 21.7 22.0 21.8 21.7 21.5 21.8 22.9 23.2 23.1
Euro area (17 countries) 21.5 21.8 21.7 21.6 21.4 21.7 22.7 23.2 23.0
Belgium 22.6 21.5 21.6 20.8 20.2 20.8 21.0 21.6 20.8
Bulgaria : 61.3 60.7 44.8 46.2 49.2 49.1 49.3 48.0
Czech Republic 19.6 18.0 15.8 15.3 14.0 14.4 15.3 15.4 14.6
Denmark 17.2 16.7 16.8 16.3 17.6 18.3 18.9 19.0 18.9
Germany 18.4 20.2 20.6 20.1 20.0 19.7 19.9 19.6 20.3
Estonia 25.9 22.0 22.0 21.8 23.4 21.7 23.1 23.4 23.5
Ireland 25.0 23.3 23.1 23.7 25.7 27.3 29.4 30.0 29.5
Greece 29.4 29.3 28.3 28.1 27.6 27.7 31.0 34.6 35.7
Spain 24.3 24.0 23.3 24.5 24.7 26.1 26.7 27.2 27.3
France 18.9 18.8 19.0 18.5 18.5 19.2 19.3 19.1 18.1
Croatia : : : : : 31.1 32.6 32.6 29.9
Italy 25.0 25.9 26.0 25.3 24.7 24.5 28.2 29.9 28.4
Cyprus 25.3 25.4 25.2 23.3 23.5 24.6 24.6 27.1 27.8
Latvia 46.3 42.2 35.1 34.2 37.9 38.2 40.1 36.2 35.1
Lithuania 41.0 35.9 28.7 27.6 29.6 34.0 33.1 32.5 30.8
Luxembourg 17.3 16.5 15.9 15.5 17.8 17.1 16.8 18.4 19.0
Hungary 32.1 31.4 29.4 28.2 29.6 29.9 31.0 32.4 33.5
Malta 20.5 19.5 19.7 20.1 20.3 21.2 22.1 23.1 24.0
The Netherlands 16.7 16.0 15.7 14.9 15.1 15.1 15.7 15.0 15.9
Austria 17.4 17.8 16.7 20.6 19.1 18.9 19.2 18.5 18.8
Poland 45.3 39.5 34.4 30.5 27.8 27.8 27.2 26.7 25.8
Portugal 26.1 25.0 25.0 26.0 24.9 25.3 24.4 25.3 27.5
Romania : : 45.9 44.2 43.1 41.4 40.3 41.7 40.4
Slovenia 18.5 17.1 17.1 18.5 17.1 18.3 19.3 19.6 20.4
Slovakia 32.0 26.7 21.3 20.6 19.6 20.6 20.6 20.5 19.8
Finland 17.2 17.1 17.4 17.4 16.9 16.9 17.9 17.2 16.0
Sweden 14.4 16.3 13.9 14.9 15.9 15.0 16.1 15.6 16.4
United Kingdom 24.8 23.7 22.6 23.2 22.0 23.2 22.7 24.1 24.8
Iceland 13.3 12.5 13.0 11.8 11.6 13.7 13.7 12.7 13.0
Norway 16.2 16.9 16.5 15.0 15.2 14.9 14.5 13.7 14.1
Switzerland : : 17.9 18.1 17.9 17.2 17.2 17.5 16.3
Macedonia : : : : : 47.2 50.4 50.3 :
Serbia : : : : : : : : 42.0
Turkey : 72.4 : : : : : : :
Source: Eurostat
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)
Next, a deprivation score ci is obtained for each person as:
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:
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
Table 7.3 UN Multidimensional poverty index

Index Headcount
Value (%)
Afghanistan 0.293 58.8
Albania 0.005 1.2
Argentina 0.015 3.7
Armenia 0.002 0.6
Azerbaijan 0.009 2.4
Bangladesh 0.237 49.5
Belarus 0.001 0.4
Belize 0.030 7.4
Benin 0.401 69.8
Bhutan 0.128 29.4
Bolivia 0.097 20.6
Bosnia and Herzegovina 0.006 1.7
Brazil 0.012 3.1
Burkina Faso 0.508 82.8
Burundi 0.442 81.8
Cambodia 0.211 46.8
Cameroon 0.260 48.2
Central African Republic 0.424 76.3
China 0.026 6.0
Colombia 0.032 7.6
Congo 0.192 43.0
Congo 0.399 74.4
Cote d’Ivoire 0.307 59.3
Djibouti 0.127 26.9
Dominican Republic 0.026 6.2
Egypt 0.036 8.9
Ethiopia 0.537 88.2
Gabon 0.073 16.7
Gambia 0.329 60.8
Georgia 0.008 2.2
Ghana 0.144 30.5
Guinea 0.548 86.5
Guinea-Bissau 0.495 80.4
Guyana 0.031 7.8
Haiti 0.242 50.2
Honduras 0.098 20.7
India 0.282 55.3
Indonesia 0.024 5.9
Iraq 0.052 13.3
Jordan 0.004 1.0
Kazakhstan 0.004 1.1
(continued)

Index Headcount
Value (%)
Kenya 0.226 48.2
Kyrgyzstan 0.013 3.4
Lao 0.186 36.8
Lesotho 0.227 49.5
Liberia 0.459 81.9
Madagascar 0.420 77.0
Malawi 0.332 66.7
Maldives 0.008 2.0
Mali 0.533 85.6
Mauritania 0.362 66.0
Mexico 0.024 6.0
Moldova 0.005 1.3
Mongolia 0.077 18.3
Montenegro 0.012 3.0
Mozambique 0.390 70.2
Namibia 0.200 42.1
Nepal 0.197 41.4
Nicaragua 0.088 19.4
Niger 0.584 89.8
Nigeria 0.239 43.3
Pakistan 0.237 45.6
Palestine 0.007 2.0
Peru 0.043 10.4
Philippines 0.038 7.3
Rwanda 0.352 70.8
Sao Tome and Principe 0.217 47.5
Senegal 0.390 69.4
Serbia 0.001 0.3
Sierra Leone 0.405 72.7
Somalia 0.500 81.8
South Africa 0.041 10.3
Suriname 0.033 7.6
Swaziland 0.113 25.9
Syrian Arab Republic 0.024 6.4
Tajikistan 0.031 7.9
Tanzania 0.335 66.4
Thailand 0.004 1.0
Macedonia 0.007 1.7
Timor-Leste 0.322 64.3
Togo 0.260 50.9
Trinidad and Tobago 0.007 1.7
(continued)
References 133

Index Headcount
Value (%)
Tunisia 0.006 1.5
Uganda 0.359 70.3
Ukraine 0.002 0.6
Uzbekistan 0.013 3.5
Vanuatu 0.135 31.2
Viet Nam 0.026 6.4
Yemen 0.191 37.5
Zambia 0.318 62.8
Zimbabwe 0.181 41.0
Source: Human Development Report (2014)
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
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.
poverty measurement and analysis. Chap. 1. Introduction (OPHI Working Paper No. 82).
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.
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.
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.
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
We have already mentioned that the measurement of poverty is usually assumed to

involve three different aspects: incidence, intensity and inequality. That is, how
many poor people are in society, how poor they are and how unequal is the
distribution of the achievements among the poor. As we have seen in Chap. 7, the
most common approach to poverty measurement starts by defining a poverty line,
as the minimum income deemed acceptable, and then proceeds to construct an
evaluation function that applies to those agents whose incomes are below that
threshold. A poverty index is a mapping that associates real numbers to income
distribution vectors, which incorporates some basic value judgements regarding our
ethical appraisal of poverty (see Chakravarty (2009, Ch. 2) for a more detailed
discussion).
Let us recall that poverty measurement may refer to the distribution of one or
several variables, adopt a subjective or an objective approach, refer to some
absolute or relative threshold to define who are the poor and consider quantitative
and/or categorical variables.
Our analysis here refers to objective poverty measurement for quantitative vari-
ables in a multidimensional context. Focussing on quantitative variables leaves
aside somehow the analysis of extreme poverty, which tends to be addressed
nowadays in terms of qualitative variables (see for instance flow of contributions
emanating from the Oxford Poverty and Human development Initiative).2 Yet
poverty analysis regarding quantitative variables is also very relevant, especially
after the substantial increase of inequality and poverty in most developed countries
due to the impact of the crisis. The multidimensional approach is important because
welfare and poverty may involve several aspects that cannot always be represented
1
This chapter is based on Villar (2015, 2016).
2
http://www.ophi.org.uk

136 8 Multidimensional Poverty and Welfare
by income or wealth variables. This is so because those aspects may evolve

differently within each society. See Chakravarty (2009), Duclos and Araar
(2006), Haughton and Khandker (2009), Wagle (2008) or the recent Alkire et al.
(2015) for a comprehensive discussions of the questions involved.
This chapter focuses on the construction of poverty indices based on the inter-
pretation of poverty as a welfare loss, following the ideas presented in Chap. 6, that
follow the Atkinson–Kolm–Sen approach to the normative theory of income
inequality (see also Chap. 4). Related ideas appear in the works of Blackorby and
Donaldson (1980), Clark, Hemming, and Ulph (1981), Lewis and Ulph (1988),
Pyatt (1987) and Vaughan (1987) among others [see also the discussion in Kakwani
(1997) and Chakravarty (2009, 2.3.3)].
In this chapter, we apply the welfare approach to a multidimensional context by
identifying a poverty index with a relative welfare loss, by means of a social
evaluation function defined on the space of multidimensional distributions. We
use the term welfare poverty indices to refer to this construct. All properties of those
poverty measures will be derived from the assumptions established on the under-
lying social evaluation function. Our approach can be regarded as combining what
Duclos and Araar (2006, Ch. 5) call “the EDE (equally distributed egalitarian)
approach” and “the poverty gap approach”. Indeed, in the single-dimensional case,
a welfare poverty index corresponds to a poverty gap measure, adjusted by inequal-
ity. Rather than looking for new poverty indices we shall provide here a relatively
easy venue to derive some familiar formulae from a normative approach parallel to
that used to discuss inequality and welfare.
The reference model is presented in Sect. 8.2. Section 8.3 introduces a series of
assumptions on the social evaluation function that provide structure to the poverty
measure (quasi-concavity, scale and factor decomposability). In particular, we
show that the welfare poverty measure can be expressed as the product of the
incidence and the intensity of poverty. We also introduce here a variant of the
decomposability principle that eventually yields the social evaluation function
corresponding to the inequality-adjusted human development index (the geometric
mean across dimensions of the geometric means across individuals). Section 8.4
describes two different venues to arrive at closed formulas of the welfare poverty
measure. Finally, Sect. 8.5 provides an empirical illustration analysing the educa-
tional poverty in the OECD, using the data corresponding to PISA 2012.
8.2 The Setting
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.
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
such that PW ðY; zÞ ¼ 0 if and only if there are no welfare-poor individuals.

The welfare poverty index is defined, therefore, as the relative welfare loss due
to the existence of individuals who do not reach the minimum admissible value of
the reference variables.
Note that PW ðY; zÞ 0, by definition, which entails that
W 1p z1 , . . . , 1p zK > W ðYp Þ when there is some poor individual and it is zero
otherwise. As a consequence, this index moves into the interval (0, 1).
Note that an individual is welfare-poor if and only if his corresponding welfare
poverty index is positive. That is,

W yi
PW y i , z ¼ 1 >0
W ðzÞ
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
8.3 Adding Structure
We shall introduce in this section some restrictions on the social evaluation

function that will permit to obtain more precise formulae for our poverty measure.
8.3.1 Quasi-concavity, Scale and Factor Decomposability
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.
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.
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
Moreover, quasi-concavity, scale and factor decomposability permit identifying

the social evaluation function with the weighted sum of the equally distributed
equivalent values of each dimension. By factor decomposability, we can write:
XK
W ðYÞ ¼ b W ðyðjÞÞ. Let now yej denote the equally distributed equivalent
j¼1 j
(EDE)
value
of the jth variable.5 That is, the value that solves the equation:
W 1n yje ¼ W ðyðjÞÞ. Since the homogeneity property implies the continuity of
W on ℝnKþþ and monotonicity along the rays, the intermediate value theorem ensures
that this value yej always exists. When the social evaluation function satisfies the
property of scale, we thus have: WðyðjÞÞ ¼ Wð1n yje Þ ¼ nyje .
We can, therefore, write the social evaluation function as follows:
XK
W ðYÞ ¼ n j¼1
bj yje ð8:2Þ
As a consequence, applying the definition of welfare poverty index in (8.1) when

the properties of quasi-concavity, scale and decomposability hold, we obtain the
following family of indices:
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.
8.3.2 Multiplicative Factor Decomposability
There are many problems in which assuming perfect substitutability between

welfare dimensions is not a sensible approach. One can introduce a multiplicative
variant of the factor decomposability property, which still provides an easy
approach to the construction of multidimensional indicators. It is the following:
Property 30
The social evaluation function W satisfies multiplicative factor decomposability.
That is,
YK βj
W ðYÞ ¼ j¼1
½W ðyðjÞÞ
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.
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).
8.4 Closing the Formula
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.
8.4.1 Anonymity, Inequality and Welfare
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.
8.4.2 Subgroup Decomposability
Suppose we can partition the population into G different subgroups according to

some socio-demographic characteristics (e.g. region of residence, race, sex, reli-
gion). Let Yg denote the achievements of the members of subgroup g ¼ 1, 2, . . ., G.
The property of subgroup decomposability states that the poverty of the whole
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
from which we obtain:

0 XK Xp 1
p by
i¼1 j ij
PðY; zÞ ¼ @1 A
j¼1
XK ð8:7Þ
n bz
j¼1 j j
We can consider now a multiplicative version of the subgroup decomposability

property.
Property 50
The social evaluation function W satisfies multiplicative subgroup decompos-
YG ng
ability. That is, W ðYÞ ¼ g¼1
½ W ðY g Þ n .
This property penalises the dispersion of welfare among population subgroups.
Replicating the reasoning above we conclude that multiplicative factor and sub-
group decomposability together imply:
Y K Y n 1=n βj
W ðYÞ ¼ j¼1 j¼1
yij ð8:8Þ
When all dimensions are equally important (8.8) becomes: W ðYÞ ¼

Y K Y n 1=nK
j¼1 j¼1
yij (the geometric mean across dimensions of the geometric
mean across individuals, which is the welfare evaluation formula that United
Nations (2010) adopts for the inequality-adjusted human development index).
8.5 Measuring Educational Poverty from PISA 145
The corresponding welfare poverty measures are:

0 0Y 1=p 1βj 1
p
pB YK B yij
C C
PðY; zÞ ¼ B A C
i¼1
@ 1 j¼1 @ A ð8:9aÞ
n zj
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
8.5 Measuring Educational Poverty from PISA
We present in this section an empirical application of the evaluation model

described in this chapter concerning educational poverty in the OECD countries,
using the data from the PISA 2012 regarding the three competencies covered:
mathematics, reading and science.8 We have already used PISA to illustrate the
extent of inequality of opportunity and now will use it again regarding the evalu-
ation of those students that do not reach level 2 of proficiency, which is considered
as the baseline level (se Sect. 5.3.3 for a description of the PISA data).
8.5.1 The Model
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).
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
This is the social evaluation

Yn function that leads to the welfare poverty index in
equation (8.9b). Since y
i¼1 ij
¼μeðyðjÞÞ is the geometric mean of the agents’
achievements regarding the jth dimension, for j ¼ m, r, s, our poverty measure can
be written as follows:
p
PðY; zÞ ¼ ρðY; zÞ ð8:11Þ
n
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
which corresponds to the coefficient that measures the inequality-adjusted intensity

of poverty.
Note that the two components of the index provide separate information on the
extent and intensity of educational poverty (where the intensity measure is to be
understood as adjusted by inequality). This is interesting because they permit
uncovering different patterns behind societies with similar values of educational
poverty.
We have already mentioned that PISA establishes six levels of proficiency,
parameterized in terms of the scores of the tests that students perform for each
subject. It is understood that Level 2 is the baseline level of proficiency for an
individual to be able to develop a reasonable integration in the labour market and,
more generally, in society. Students who do not reach that level are considered to
have an insufficient knowledge (low performers). It is only natural to interpret
insufficient knowledge as educational poverty and thus use the thresholds that
define those minimum levels to set the corresponding poverty lines in mathematics,
reading competence and science, respectively.
According to the PISA 2012 report, the thresholds that define low performance
in those aspects are: 420.1 test score points for mathematics (m), 407.5 points for
reading competence (r), and 409.5 points for science (s). Therefore, our vector z of
poverty lines is given by: z ¼ ð420:1, 407:5, 409:5Þ. We consider that those three
dimensions are equally important.
The micro-data of the PISA report provide information about the test scores of
individuals that conform the representative sample for each country. We take those
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
A student is considered educationally poor whenever this number is strictly

positive and non-poor otherwise. So our first step is computing this expression for
each individual student in every country. We then select those students for which
this expression is positive within each country, which gives us the set of poor
students. Once this set has been determined, we calculate all the elements required
to compute the poverty index.
8.5.2 The Results
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).
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)
Educational poverty is highly correlated (negatively) with the average scores of

the tests, with some 94 % of common variance. There is a positive but relatively
weak correlation between the Educational Poverty Index and the index of economic
and socio-cultural status (ESCS). This is not surprising since it is already well
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.
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
deviation of the relative incidence). Regarding equality-adjusted intensity, Israel,

Sweden, France, Iceland and Belgium present values which are more than one
standard deviation (15 points) above the OECD mean, whereas Estonia, Poland,
Slovenia, Canada, Ireland and the United States have values that are at least one
standard deviation below the mean. So even though there is a positive correlation
between incidence and equality-adjusted intensity, the common variance is quite
small (0.17), and there is a variety of situations regarding those variables.
8.5.3 Final Comments
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
working paper n 976).
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.
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.
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.
Tansel, A. (2015). Inequality of opportunities in educational achievements in Turkey over time
(IZA working paper DP 9005).
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).

156 9 The Evaluation of Relative Achievements
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
discussions when evaluating performances. The first interpretation (the intersection

approach) amounts to decide that no compensation is admitted. The second inter-
pretation (the union approach) entails full compensation, so that reaching one target
is enough.
We present here an axiomatic approach to the identification of intermediate
compensation patterns in terms of specific evaluation functions. We shall show that
four intuitive properties define a simple and reasonable evaluation criterion: the
average of the ratios between outcomes and targets.
This type of evaluation can be regarded as particular case of multicriterion
decision-making (e.g. Kenney & Raiffa, 1976; Yu, 1985), and the proposed solu-
tion corresponds to what the literature calls “compromise solutions” (see André,
Cardenete, & Romero, 2010, Ch. 9; Romero, 2001). Yet this approach follows the
literature on welfare evaluation and multidimensional poverty analysis presented in
former chapters (see also Bourguignon & Chakravarty, 2003; Chakravarty, 2003;
Herrero, Martinez, & Villar, 2010; Seth, 2010; Tsui, 2002; Villar, 2011).
9.2 The Basic Model
Consider an organisation consisting of several units whose performance is to be

evaluated with respect to a vector of targets or reference values previously set.
Depending on the problem under consideration, those targets may represent abso-
lute values, relative performance thresholds or a mixture of them. We can think of
the purpose of the evaluation as the allocation of some resources among those who
qualify. The evaluation procedure itself may be conceived as a simple dichotomous
criterion concerning the achievement of the targets (objectives reached or not); or it
may attempt at providing quantitative estimates of the overall degree of fulfilment
or something in between (e.g. classification in different categories).
We shall refer to the organisation as a society and to the incumbent units as
agents. The key feature of the problem is the existence of a society with many
agents whose performance is to be evaluated with respect to a given set of targets, to
be called standards. Note that in some cases, meeting the standards may imply
getting values below the thresholds.
Deciding who meets the standards in a multidimensional scenario is not imme-
diate. As we have already seen when presenting the example of the Hypothetical
University, two extreme positions can be considered. On the one hand, there is the
most demanding interpretation by which meeting the standards means achieving all
9.2 The Basic Model 157
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).
9.2.1 Measuring the Achievements
Let N ¼ f1; 2; . . . ; ng denote a society with n agents and let K ¼ f1; 2; . . . ; kg be a

set of reference dimensions, with k > 1. A realisation for this society is a matrix
Y with n rows, one for each agent, and k columns, one for each dimension. The
entry yij is a real number that describes the value of variable j for agent i. Therefore,
ℝnk is the space of realisation matrices and we assume implicitly that all dimensions
can be approximated quantitatively by real numbers. There is a parameter vector of
reference values z 2 ℝþþ K
that describes the standards fixed for the different
dimensions. We shall not discuss here how those thresholds are set, even though
the importance of that choice is more than evident.
In the case of example in Sect. 9.1, regarding the performance of the Hypothet-
ical University, we would have:
0 1
15 23 8
Y ¼ @ 26 40 4 A, z ¼ ð10, 15, 5Þ
10 0 12
We take ℝnK ℝþþ K

as the joint space of realisations and targets. An evaluation
problem, or simply a problem, is a point ðY; zÞ 2 ℝnK ℝþþ K
. To evaluate the
society’s outcomes given by matrix Y, relative to a vector of targets z, we look for a
continuous function φ : ℝnK ℝþþ K
! ℝ that associates to each problem (Y, z) a
real number φ(Y, z) that tells us about the fulfilment of the objectives. We shall
obtain this function from an intuitive set of axioms.
Our first axiom, symmetry, establishes that all agents are equally important. That
is, the relevant aspect refers to the outcomes obtained and not to who got them.
Formally,
• Symmetry: Let ðY; zÞ 2 ℝnK ℝþþ K
and let π(Y) stand for a permutation of the
rows of matrix Y. Then, φðY; zÞ ¼ φðπ ðYÞ, zÞ.
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,
φ½ðY þ ΔYÞ; z ¼ φðY; zÞ þ φðΔY, zÞ
The following result is obtained:

Theorem 9.1 A continuous function φ : ℝnK ℝþþ K
! ℝ satisfies the properties
of symmetry, neutrality, normalisation and additivity, if and only if:
1 Xn X K
yik
φðY; zÞ ¼ ð9:1aÞ
nK i¼1 k¼1 zk
Moreover, all those properties are independent.

Proof (i) It is easy to see that function (9.1a) satisfies all properties. Let us verify
the reciprocal.
For a given problem ðY; zÞ 2 ℝnK ℝþþ K
, let Δðyik Þ 2 ℝ
nK
denote a matrix
whose elements are all zero except entry (i, j) that is equal to yik. Applying
repeatedly the property of additivity, we have:
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:
φ½Δij ðaÞ, 1s ¼ φ½Δht ðaÞ, 1s, 8i, h 2 N, 8j, t 2 K
Let now Y(yik) denote a n K matrix all of whose entries are equal to yik. We
would have:
φðYðyik Þ, 1sÞ ¼ nKφðΔðyik Þ, 1sÞ
so that:
φðYðyik Þ, 1sÞ
φðΔðyik Þ, 1sÞ ¼
nK
We can thus write:
1 Xn X K
φðY; zÞ ¼ φðYðyik Þ, 1sÞ ð9:2Þ
nK i¼1 k¼1
Now observe that our assumptions imply that function φ is homogeneous of

degree 1, that is, φðλY, zÞ ¼ λφðY; zÞ, for all λ > 0. Define a new function f : ℝ
ℝþþ ! ℝ as follows:
f ðyik ; zk Þ :¼ φðYðyik Þ, 1zk Þ
As this function inherits the properties of homogeneity and normalisation, by

letting λ ¼ zk =yik , we have:
zk y
f ðyik , zk Þ ¼ f ðzk , zk Þ ¼ 1 ) f ðyik , zk Þ ¼ ik
yik zk
Substituting into Eq. (9.2) for all i, k, we get:
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
It satisfies all properties except normalisation.

yik
φB ðY; zÞ ¼ mini
zk
It satisfies all properties except additivity.

1X n X K
y
φ ðY; zÞ ¼
C
wi ik
K i¼1 k¼1 zk
X
with i
wi ¼ 1 and wi 6¼ 1=n for some i. It satisfies all properties except symmetry.

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
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.
9.2.2 Application to the Example of the Hypothetical

University
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:
Table 9.2 Relative values and evaluations of departments and targets

Ratio Ratio research Ratio Evaluation
publications funds grants departments
Department A 1.50 1.53 1.60 1.54
Department B 2.60 2.67 0.80 2.02
Department C 1.00 0.00 2.40 0.47
Evaluation 1.70 1.40 0.00 1.03
targets
Bold value is the overall significance that coincides with the sum of the last row and last column
(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.
9.3.1 Weighted Symmetry and Weighted Neutrality
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
The evaluation formula under generalised symmetry and generalised neutrality

would be given by a weighted average of relative achievements. That is, the sum of
all ratios yij/zj each of which pondered by the weight corresponding to the agent and
the objective, wi, bk. That is, the general version of (9.1) is given by:
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.
Table 9.3 Evaluation of Departments with targets of different importance

Departments’ evaluations
Department A 1.50 0.2 + 1.53 0.4 + 1.6 0.4 ¼ 1.55
Department B 2.60 0.2 + 2.67 0.4 + 0.80 0.4 ¼ 1.91
Department C 1.00 0.2 + 0.00 0.4 2.40 0.4 ¼ 0.76
Table 9.4 Evaluation of targets of asymmetric Departments

Publications Research funds Grants
1.50 0.3 + 2.60 0.2 1.53 0.3 + 2.67 0.2 1.60 0.3 + 0.80 0.2 –
+ 1.00 0.5 ¼ 2.47 + 0.00 0.5 ¼ 1.49 2.40 0.5 ¼ 1.26
Regarding the achievement of the different targets in asymmetric Departments

as set in the strategic plan, the application of Eq. (9.3) yields the following result:
φðY; zÞ ¼ 2:47 0:2 þ 0:4 1:49 0:4 1:26 ¼ 0:59
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).
9.3.2 Generalised Means
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.
9.3.3 A Graphical Illustration
Let us conclude this section by presenting a graphical illustration of the different

evaluation formulae presented, focussing on the case in which there are only two
targets.
In Fig. 9.1, we describe the relative values of the
agents’
achievements on the
yi1 yi2
Cartesian axis, where each point represents a value z1 ; z2 of an agent. Let us take
as reference the point (1, 1) that corresponds to the case in which the achievements
match the targets.
When we adopt the most demanding interpretation of what meeting the targets
means (i.e. the intersection approach), then we are saying that only those points in
the shaded area D are admissible for a positive evaluation. So those agents whose
relative realisations lie outside that area will be considered as failing to satisfy the
objectives. In the polar case in which reaching a target is enough (the union
approach), all points outside the area A correspond to satisfactory values (i.e. all
agents in the union of the areas B, C, D, E and F will be evaluated positively.
Our evaluation formula (9.1a) establishes that achieving the targets amounts to
get an average of relative performance larger than or equal to 1. That is,
Fig. 9.1 Comparison of

criteria of fulfilment

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).
9.4 An Application: The Green Economy Progress Index
9.4.1 The Setting
The Inclusive Green Economy is a United Nations Environmental Programme

(UNEP) initiative that aims at providing tools for delivering sustainable develop-
ment and a response to three sets of challenges facing humanity: persistent poverty,
inequitable sharing of the growing prosperity and overstepped planetary
9.4 An Application: The Green Economy Progress Index 167
Fig. 9.2 Evaluation with

the arithmetic and the
geometric mean
boundaries. So it is concerned, in particular for “sustained, inclusive and sustain-

able growth, full and productive employment and decent work for all” (Sustainable
Development Goal 8).
UNEP places a particular emphasis on those actions “that can mobilize finance
for building the new generation of assets or re-shaping the patterns of consumption,
investment, public spending, and trade, or both. Examples include fiscal policy
(e.g. reform of fossil fuels can encourage energy efficiency in both consumption
and production), industrial policy (e.g. government spending on R&D for renew-
able energy technologies), sustainable public procurement, labour training, social
safety nets and trade liberalization for environmental goods and services, among
others. Other policies such as rules, regulations, and standards remain important as
ever should complement the IGE policy instruments that focus more directly on
mobilizing finance for the transition towards an IGE” (Cf. UNEP, 2017).
In line with the United Nations well-known Human Development indicators,
UNEP has developed a Green Economy Progress measurement framework, as an
instrument that assesses country efforts towards achieving targets set within plan-
etary boundaries and allows cross-country comparison of progress towards achiev-
ing an inclusive green economy. It consists of an index of progress and a dashboard.
The Green Economy Progress index measures the changes in green economy
indicators, relative to desired changes, which directly or indirectly impact current
human well-being. The dashboard of sustainability indicators has the objective of
monitoring the sustainability of well-being (i.e. well-being of future generations).
The GEP index includes 13 multidimensional indicators that are associated with
the three challenges of Inclusive Green Economy: inequitable sharing of growing
prosperity, overstepped planetary boundaries and persistent poverty (Table 9.5).
Three remarks are in order before presenting the GEP index:
(i) Some of the variables included in the table are “goods” (boldfaced) whereas
others are “bads”. This means that progress for a given country derives from
increasing the values of the goods and decreasing the values of the bads.
(ii) The index focuses on the progress, i.e. the changes, rather than on the levels.
Table 9.5 Components of the GEP index

Indicator Description
Green trade Export of environmental goods according to OECD and APEC
(% of total export)
Green technology Patent publication in environmental technology by filing office
innovation (% of total patents)
Renewable energy sources Share of renewable energy supply (of total energy supply)
Energy use Energy use (kg of oil equivalent) per $1000 GDP (constant 2011
PPP).
Palma ratio Ratio of the richest 10 % of the population’s share of income
divided by the poorest 40 %’s share
Access to basic services It is a composite measure created by the average access to three
basic services with key social and environmental implications:
Access to improved water sources (% of total population),
Access to electricity (% of total population),
Access to sanitation facilities (% of total population)
Air pollution PM2.5 pollution, mean annual exposure (micrograms per cubic
meters)
Material footprint per capita Raw material consumption of used biotic and abiotic materials
(tonnes/person)
Marine and terrestrial Sum of terrestrial protected areas (% of total land area) and
protected areas marine protected area (% of territorial waters)
Gender inequality index A composite measure reflecting inequality in achievements
between women and men in three dimensions: reproductive
health, empowerment and the labour market
Pension coverage Share of population above statutory pensionable age receiving an
old age pension by contribution and sex
Education (Mean years of Average number of years of education received by people ages
schooling) 25 and older, converted from education attainment levels using
official durations of each level
Life expectancy Life expectancy at birth indicates the number of years a new-born
infant would live if prevailing patterns of mortality at the time of
its birth were to stay the same throughout its life
Source: UNEP (2017)
Boldfaced variables refer to “goods” and non-boldfaced to “bads”
(iii) Progress is measured relative to some standards: targets and thresholds.

Targets refer to desired changes, whereas thresholds define some critical
levels.
9.4.2 The Index
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
depending on whether it is a good (G) or a bad (B).

We denote by yj the desired progress of the jth variable of the country regarding
the evolution of this variable. Then, we define the country’s target for the jth
variable, depending on whether it corresponds to a good or a bad, as follows:

dy*j d y*j
zGj ¼ , zBj ¼
y0j 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
for the case of goods, and:

n o
y*j ¼ min tj , βj y0j , βj 1
for the case of bads.

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:
X y1j y0j X y0j y1j

GEP ¼ π
j2G j
þ π
j2B j
ð9:6aÞ
y*j y0j y0j y*j
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
In this way, we give higher importance to progress on those indicators in which

the initial condition is less favourable relative to the threshold. This can be
interpreted as an incentive to improve in those indicators in which a country is
relatively worse (further away from the threshold). Consequently, for each indica-
tor, the corresponding weight is set as a scaled version of the relation between
threshold and initial value. Then, by normalising those weights so that they add up
to 1, we get the following formula:
0 1
1 @X tj y1j y0j X y0j y0j y1j
GEP ¼ n o þ n oA ð9:6bÞ
S j2G y0 j2B t 0
j max tj , λj y0 y0
j
j y min tj , β y0
j j j j
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.
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.
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.

Lectures On Inequality, Poverty and Welfare

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lectures On Inequality, Poverty and Welfare

Uploaded by

Copyright:

Available Formats

Lecture Notes in Economics and Mathematical Systems 685

Prof. Dr. W. Trockel

ISSN 0075-8442 ISSN 2196-9957 (electronic)

Library of Congress Control Number: 2016959767

© Springer International Publishing AG 2017

Printed on acid-free paper

This Springer imprint is published by Springer Nature

This monograph originates in a short course I taught at the Economics Department

Paris Antonio Villar Ph.D.

3.3.1 The First Index of Theil: From Information Theory to

1.1 Equality and Welfare

1.1.1 We Have a Problem. . .

© Springer International Publishing AG 2017 1

situation is alarming, due to the extremely unequal opportunities to cover basic

– 836 million people still live in extreme poverty.

1.1.2 Measuring Inequality and Poverty

The purpose of this monograph is to provide some basic ingredients to measuring

The relevance of those applications calls for a sound understanding of those

1.2 Equality of What?

1.2.1 Income Distributions

Income distribution is a key element in the analysis of economic inequality and

1.2.2 Wealth, Income and Expenditures

From a theoretical viewpoint, we can think of an agent’s income in a given period as

1.2.3 The Units of Reference: Equality Among Whom?

measure of the household standard of living, as it corresponds to the income that, if

1.2.4 Needs and Deserts

1.3 Plan of the Book

Chapter 8 develops an approach to poverty measurement based on the inter-

Cowell, F. A. (1977). Measuring inequality. Oxford: Phillip Allan.

© Springer International Publishing AG 2017 19

2.2 Properties of Inequality Indices

individual incomes. In particular, the continuity property establishes that two

distributionof the whole society N can be regarded as a vector of vectors,

for a set of positive coefficients ωg, which may depend on ~

2.3 Inequality and Dispersion

2.3.1 The Variance (σ2)

Table 2.1 % of income and population in the two income distributions

Cumulative population Cumulative income

An alternative expression of the variance is the following (Kendall & Stuart,

2.3.2 The Coefficient of Variation (CV)

As dividing by the mean is just another monotone transformation (linear in this

Atkinson, A. B., & Bourguignon, F. (1982). The comparison of multidimensional distributions of

3.1 The Lorenz Curve

© Springer International Publishing AG 2017 31

Fig. 3.1 The Lorenz curve L

Line of perfect equality

0 1/n 2/n 3/n ... 1 F

Fig. 3.2 Dalton’s transfers

Fig. 3.3 Lorenz dominance

3.2 The Gini Index

3.2.1 From Lorenz to Gini

As A + B ¼ ½, we can also write:

Fig. 3.4 The Lorenz curve

Fig. 3.5 Area below the L

As p0 ¼ L0 ¼ 0, and bearing in mind that G ¼ 2A [see equation (3.1b)], we obtain:

Note that p0 ¼ L0 ¼ 0 and pn ¼ Ln ¼ 1, so that we can write:

or, put in a simpler way:

y ¼ ðy1 , y2 , . . . yn Þ. It makes explicit that all those differences between population

3.2.2 The Generalised Gini Index

This is a uni-parametric family of inequality measures whose members are

When using Sen’s formulation of the Gini index, we have:

3.3 Theil’s Inequality Indices

3.3.1 The First Index of Theil: From Information Theory

The concavity of utility functions implies that ξ μ with ξ ¼ μ if and only if

with A ¼ 0, when ξ ¼ μ and A ! 1 as ξ ! 0, which implies that 0 A 1.