Bok:978 81 322 2166 1 PDF

Economic Studies in Inequality, Social Exclusion
and Well-Being
Series Editor: Jacques Silber
SatyaR.Chakravarty
Inequality,
Polarization
and Conflict
An Analytical Study
Economic Studies in Inequality, Social Exclusion

and Well-Being
Volume 12
Series editor
Jacques Silber
More information about this series at http://www.springer.com/series/7140
Satya R. Chakravarty
Inequality, Polarization
and Conflict
An Analytical Study
Economic Research Unit
Indian Statistical Institute
Kolkata, India
ISSN 2364-107X
ISSN 2364-1088 (electronic)
Economic Studies in Inequality, Social Exclusion and Well-Being
ISBN 978-81-322-2165-4
ISBN 978-81-322-2166-1 (eBook)
DOI 10.1007/978-81-322-2166-1
Library of Congress Control Number: 2015933995
Springer New Delhi Heidelberg New York Dordrecht London
Springer India 2015
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
Springer (India) Pvt. Ltd. is part of Springer Science+Business Media (www.springer.com)
In memory of
Professor Nikhilesh Bhattacharya,
my favorite teacher and mentor
Preface
The major concern of my two earlier Springer books Ethical Social Index Numbers
and Inequality, Polarization and Poverty: Advances in Distributional Analysis was
measurement of inequality, poverty and well-being. Only one chapter of the second
monograph was devoted to an analysis of income polarization. However, research
on polarization has gained impetus in the last decade because of the pivotal role of
polarization in analyzing the evolution of the distribution of income, economic
growth and social conflicts. Policy advisers in many countries now insist on looking
at polarization as a source of social conflict. In view of this, the present monograph
makes a systematic treatment of theory and methodology of alternative notions of
polarization and related issues. A wide coverage of inequality, polarization and
conflict is provided in the book. It gives an overall view of the recent developments
in the subject.
There are two approaches to the measurement of income polarization: bipolarization and multi-polar polarization. According to the first approach, polarization is
the shrinkage of the middle class; on the other hand, the later approach regards
polarization as clustering around local means of the distribution, wherever these
local means are located on the income scale. In order to make a clear distinction
between inequality and polarization, in Chap. 1 there will be a discussion on income
inequality measurement. Then Chap. 2 of the monograph goes on to analyze
alternative approaches to the measurement of bipolarization rigorously.
An analysis of multi-polar polarization indices is presented in an axiomatic
framework in Chap. 3. Then in Chap. 4, there will be a formal discourse on
reduced-form indices which are increasingly related to between-group component
and decreasingly related to within-group component of a subgroup decomposable
inequality index. Social polarization refers to the widening of gaps between specific
subgroups of people in terms of their social circumstances and opportunities.
Chapter 5 of this monograph studies social polarizations using a rigorous and
analytical structure.
vii
viii
Preface
It is now well-known that human well-being is a multidimensional phenomenon.

While some of the dimensions correspond to ratio scale variables (e.g., income,
wealth), dimensions like health and literacy are represented by ordinal variables.
Study of polarization for an ordinal dimension of human welfare is the subject of
Chap. 6 of the book. Chapter 7 of the book analyzes the question of the effects of
inequality, fractionalization and polarization on social conflict in a broad structure.
I am indebted to Nachiketa Chattopadhyay, Conchita DAmbrosio, Bhargav
Maharaj, Amita Majumder, Sonali Roy and Claudio Zoli for the benefit I derived
from them as my coauthors. I gave seminars on several sections of the book at
Bar-Ilan University, Ramat-Gan, Israel; Bocconi University, Milan, Italy;
Statistics-Mathematics Unit of Indian Statistical Institute, Kolkata, India;
Jawaharlal Nehru University, New Delhi, India; University of International
Business and Economics, Beijing, China; and Yokohama National University,
Yokohama, Japan. I am grateful to the seminar participants for their comments
and suggestions. I have also interacted with Rolf Aaberge, Sabina Alkire, Yoram
Amiel, Tony Atkinson, Charles Blackorby, Walter Bossert, Francois Bourguignon,
Frank A. Cowell, Koen Decancq, Joseph Deutsch, David Donaldson, Jean-Yves
Duclos, Bhaskar Dutta, Indranil Dutta, Udo Ebert, Gary S. Fields, Marc Fleurbaey,
James E. Foster, Tomoki Fujii, Carlos Gradin, Nanak C. Kakwani, Ravi Kanbur,
Serge-Christophe Kolm, Peter J. Lambert, Casilda Lasso de la Vega, Maria Ana
Lugo, Francois Maniquet, Laurence Roope, Amartya K. Sen, Tony Shorrocks,
Jacques Silber, Kai-Yuen Tsui, Gaston Yalonetzky, Shlomo Yitzhaki and Buhong
Zheng. It is a pleasure for me to express my sincere gratitude to all of them.
I thank Md. Aslam, Debasmita Basu, Nandish Chattopadhyay, Ranajoy Guha
Neogi, Doyel Kayal and Pradip Maiti for generating the figure files in different
chapters. I also thank my wife Sumita for carefully reading the manuscript and my
son Ananyo for helpful cooperation.
Kolkata, India
Contents
Measuring Income Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Postulates for an Index of Inequality . . . . . . . . . . . . . . . . . . . .
1.4 Ethical Approaches to the Measurement of Inequality . . . . . . .
1.5 Subgroup Decomposable Indices of Inequality . . . . . . . . . . . . .
.
.
.
.
.
.
1
1
2
4
15
26
On the Measurement of Income Bipolarization . . . . . . . . . . . . . . . .

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Measuring the Middle Class . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Axioms for an Index of Income Bipolarization . . . . . . . . . . . . . .
2.4 A Bipolarization Ordering and Some Relative
Bipolarization Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Alternative Notions of Bipolarization Invariance
and the Associated Orderings . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6 Welfare Theoretic Approaches to the Measurement
of Bipolarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
33
35
37
Measurement of Income Multipolar Polarization . . . . . . . . . . . . .

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Income Multipolar Polarization for Discrete Distributions . . . . .
3.3 Income Multipolar Polarization for Continuous Distributions . .
3.4 Some Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
53
53
54
57
60
Reduced-Form Indices of Income Polarization . . . . . . . . . . . . . . .

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 The Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4 An Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Analysis of Reduced-Form Indices and Their
Eventual Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
63
63
65
66
68
71
40
48
50
ix
Contents
Social Polarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 A Discrete Metric-Based Index . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Some Characterizations and a Generalized Index . . . . . . . . . . .
5.4 Some Alternative Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 A Social Polarization Ordering . . . . . . . . . . . . . . . . . . . . . . . .
Measuring Polarization for a Dimension of Human

Well-being with Ordinal Significance . . . . . . . . . . . . . . . . . . . . . .
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Axioms for an Index of Bipolarization for a Dimension
Measurable on Ordinal Scale . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 Some Indices of Bipolarization for a Dimension
of Well-being with Ordinal Information . . . . . . . . . . . . . . . . . .
6.4 The IdentificationAlienation Approach to Polarization
Measurement for a Dimension with Ordinal Representation . . .
6.5 Bipolarization Orderings for an Ordinal Dimension . . . . . . . . .
.
.
.
.
.
.
77
77
80
83
87
91
.
.
97
97
98
. 104
. 105
. 107
Fractionalization, Polarization, and Conflict . . . . . . . . . . . . . . . . . .

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Fractionalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3 Inequality and Fractionalization . . . . . . . . . . . . . . . . . . . . . . . . .
7.4 Inequality, Polarization, and Fractionalization as Indicators
of Conflict: A Behavioral Model . . . . . . . . . . . . . . . . . . . . . . . .
109
109
111
115
116
Glossary of Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

Extended Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
About the Author
Satya R. Chakravarty is a Professor of Economics at the Indian Statistical

Institute, Kolkata, India. He has articles published in many internationally known
journals and edited volumes on welfare issues, cooperative game theory, industrial
organization and mathematical finance; and books published by Cambridge University Press and Springer. He is an Associate Editor of Social Choice and Welfare,
a member of the Editorial Board of Journal of Economic Inequality and a Co-editor
of Economics E Journal. He worked as a consultant to the Asian Development
Bank, an external reviewer of the World Bank, and also as an adviser of the
National Council of Social Policy Evaluation, Mexico. He was awarded the
Mahalanobis memorial prize by the Indian Econometric Society in 1994 and is a
fellow of the Human Development and Capability Association.
xi
Chapter 1
Measuring Income Inequality
1.1
Introduction
Inequality in an income distribution in a society delineates disparities of incomes

among the individuals in the society. Indicators of inequality are often employed to
judge the distributional effects of a particular economic policy or evaluate a
particular distribution. For instance, government policy advisors may be interested
in knowing whether implementation of a suggested economic policy has led the
economy to a lower level of inequality over a certain period of time. In order to
reduce social tensions or conflicts, a societys objective may be to reduce the level
of inequality that currently exists between different ethnic or social subgroups.
Which particular social subgroup or region is a major source of current level of
income discrepancy in the country? Has a particular ethnic subgroup in the society
become more cohesive because of reduction of inequality in the subgroup? For any
partitioning of the population with respect to some socioeconomic attribute, does
more overall inequality, measured by a subgroup decomposable index, make the
society more polarized in the sense that there is higher between-group inequality
but lower within-group inequality so that the between-group component is dominant over the within-group component? Will a highly progressive tax system be
able to make the income distribution more equitable and generate sufficient funds
for financing the provision of a public good?
In order to answer all such questions and related enquiries, a rigorous discussion
on the measurement of inequality is necessary. This is the objective of this chapter.
After presenting some preliminaries in the next section, we discuss the axioms for
an index of inequality in Sect. 1.3. This discussion will enable us to make a
systematic comparison between indices of inequality and polarization. There will
be a discussion on ethical approaches to the measurement of inequality, including
stochastic dominance, in Sect. 1.4 because it will be useful for developing a similar
approach to the measurement of bipolarization. Since subgroup decomposable
Springer India 2015

S.R. Chakravarty, Inequality, Polarization and Conflict, Economic Studies in
Inequality, Social Exclusion and Well-Being 12, DOI 10.1007/978-81-322-2166-1_1
inequality indices form the basis of reduced-form polarization indices and the
related ordering presented in Chap. 4, we analyze such inequality indices in
Sect. 1.5.
1.2
Preliminaries
For a population of size n, an income distribution is represented by a vector x

x1 , x2 , . . . xn 2 Dn ; where Dn is the nonnegative part of the n-dimensional
Euclidean space Rn with the origin deleted. Here, xi denotes the income of individual i of the population. We can write Dn explicitly as Dn Rn =f0:1n g, where Rn is
the nonnegative part of the n-dimensional Euclidean space, 1n is the n-coordinated
vector of ones, and n is any arbitrary positive integer. The set of all possible income
distributions is given by D [ Dn , where N is the set of positive integers. Let Dn
n2N
be the positive part of Dn. The sets of all possible income distributions
corresponding to Rn and Dn are denoted by R and D , respectively. Observe
that for all n 2 N, each of the three sets Dn , Rn and Dn is convex, that is, if x and
y are any two elements of any of these sets, then tx 1 ty is also an element of
that set, where 0 t 1 is arbitrary.
Unless specified, we will assume that D is the set of all possible income
distributions. We will adopt the following notation. For all n 2 N, for all x 2 Dn ,
n
P
x (or, simply ) stands for the mean of x, 1n xi : For all n 2 N, for all x 2 Dn , let x^
i1
be the illfare-ranked or nondecreasingly ordered permutation of x, that is, x^ 1 x^ 2

. . . x^ n : The distribution x is used to denote the welfare-ranked or nonincreasingly ordered permutation of x, that is, x1 x2 . . . xn , where x 2 Dn
is arbitrary. By an inequality index, we mean a nonconstant function I : D ! R1 .
This general definition of an inequality index allows inequality comparisons of
distributions of income whose totals as well as population sizes are different. If the
domain of I is simply Dn, then we can make only comparisons of inequality for a
fixed population size n. Inequality is not defined if n 1. Consequently, we assume
that n 2.
Definition 1.1 A function H : D ! R1 is called concave if for all n 2 N, x, y 2 Dn
and for all 0 t 1, Htx 1 ty tH x 1 tH y : The function H : Dn
! R1 is called strictly concave if Htx 1 ty > tH x 1 tH y for all
0 < t < 1 and for all x, y 2 Dn , where x 6 y: The function H : Dn ! R1 is defined as
convex (strictly convex) if H : Dn ! R1 is concave (strictly concave).
Definition 1.2 A function H : D ! R1 is called S-concave if for all n 2 N, x 2 Dn
and for all bistochastic matrices A of order n, H xA H x, where an n n matrix
A with nonnegative entries is called a bistochastic matrix order n if each of its rows
1.2 Preliminaries
and columns sums to unity.1 Strict S-concavity of H requires that the weak
inequality is to be replaced by a strictly inequality whenever xA is not a reordering
or permutation of x. A function H : D ! R1 is defined as S-convex (strictly
S-convex) if H : Dn ! R1 is S-concave (strictly S-concave).
Definition 1.3 A function H : D ! R1 is called symmetric if all n 2 N, x 2 Dn ,
H x Hy, where y is any permutation of x, that is, y x, where is any
permutation matrix of order n.
Symmetry requires invariance of the value of the function under reordering of
incomes. It is an anonymity principle. All S-concave functions are symmetric.
Definition 1.4 For all n 2 N, x, y 2 Dn , we say that x is obtained from y by a Pigou
(1912)Dalton (1920) progressive transfer (progressive transfer, for short), which
we denote by xTy, if for some i, j and c > 0,
xi yi c xj ,
xj yj c;
and xk yk for all k 6 i, j.
That is, x is obtained from y by a transfer of c units of income from a rich person
j to a poor person i that does not make the donor poorer than the recipient.
Equivalently, we can say that y is obtained from x by a regressive transfer.
Definition 1.5 For all n 2 N, x, y 2 Dn , x is said to be obtained from y by a simple
increment if yj c xj for some j and xi yi for all i 6 j, where c > 0.
That is, x and y are identical except that the jth income in x is obtained by
increasing the corresponding income in y by the amount c. We denote this by the
inequality x y.
Definition 1.6 A function H : D ! R1 is called increasing in individual arguments
(increasing, for short) if for all n 2 N, x, y 2 Dn , H x > H y whenever x is
obtained from y by a simple increment.
Definition 1.7 For any x 2 Dn ;y 2 Dnl , where each income in x appears l times in y,
is called an lfold replication of x, where l 2 is an integer.
For ordered distributions, ordering should be maintained in the replicated versions as well. For instance, if x 2 Dn ; then the income distribution
y^ x^ 1 , x^ 1 , . . . , x^ 1 , x^ 2 , . . . x^ 2 , . . . , x^ n , where each x^ i appears l times, is a l
fold replication of x^ .
Definition 1.8 A function H : D ! R1 is called population replication invariant if
for all n 2 N, x 2 Dn , H x H y ; where y 2 Dnl is an l fold replication of x,
l 2 being any integer.
1
A bistochastic matrix order n with exactly one positive entry in each row and column is called a
permutation matrix of order n.
The population replication invariance property enables us to view the functional

value of H as an average concept. For instance, the mean income is an average
function.
1.3
Postulates for an Index of Inequality
In this section, we present alternative postulates for an index of inequality and

examine some implications of these postulates. Necessity of these postulates arises
for distributional comparisons of inequality.
The general definition of an inequality index I : D ! R1 permits variability of
total incomes and population sizes. However, our concern should be inequality
comparison of the distributions of a given total over a given population size. The
issue of differing totals can be taken care of using some notion of inequality
invariance concept.
An inequality index I : D ! R1 is a relative or scale invariant index if for all
n 2 N, x 2 Dn ,
I cx I x;
1:1
where c > 0 is any scalar. That is, proportional changes in all the incomes do not
change inequality. In other words, I is homogeneous of degree zero. In contrast, an
absolute index A : D ! R1 is invariant under equal absolute changes in all the
incomes, that is, for all n 2 N, x 2 Dn ,
Ax c1n Ax;
1:2
where c is a scalar such that x c1n 2 Dn . Such an index is called translation

invariant or translatable of degree zero.
To understand these two notions of inequality invariance from normative perspectives, let us consider the two-person income distribution (1, 2000). By any
relative index, this distribution and the distribution (2, 4000) are inequality equivalent because the latter is obtained from the former by doubling the incomes.
However, out of the additional total income of 2001, the poorer person receives
only 1 unit, whereas the richer person receives 2000 units. That is why Kolm
(1976a, b) referred to relative inequality invariance as a rightist concept. On the
other hand, if each individuals income increases by one unit, the resulting distribution becomes (2, 2001). In this case, the poorer persons income increases by
100 % and the richer persons income increases by 0.05 %. But because of equal
absolute change in the two incomes, the absolute inequality remains unchanged. In
view of this, Kolm (1976a, b) regarded absolute invariance as a leftist concept. A
compromise condition between these two notions of inequality invariance is a
centrist concept, which requires that a multiplication of individual incomes by a
1.3 Postulates for an Index of Inequality
positive scalar should increase inequality and a translation of all incomes by the
same amount should decrease inequality. One particular type of centrist invariance
is the BossertPfingsten (1990) notion of intermediate invariance.
As noted by Bossert and Pfingsten (1990), a natural generalization of (1.1) and
(1.2) is an intermediate condition, which stipulates that a convex mixture of relative
and absolute variations in incomes should keep inequality unchanged. That is, an
inequality index I : D ! R1 represents an intermediate view of inequality if for all
n 2 N, x 2 Dn ,
I x cx 1 1n I x;
1:3
where , 0 1, is a parameter which reflects

an evaluators view on inequality

equivalence and c is a scalar such that x c x 1 1n 2 Dn . The vector 1n is

expressed in income units so that y x c x 1 1n becomes well defined.
Obviously, for all non-negative finite values of c, y 2 Dn . The relative and absolute
concepts of inequality equivalence emerge as polar cases of the intermediate notion
given by (1.3) when takes on the values 1 and 0, respectively.
In order to illustrate (1.1), (1.2), and (1.3) graphically, consider a society
consisting of two individuals with income levels x1 and x2 (Fig. 1.1).
Income x2
b
a
x2
x1
Fig. 1.1 Different inequality concepts
Income x1
0 0
0
The vector x x1 ; x2 is an initial distribution of income. The three straight
lines in the figure, denoted, respectively, by a, b, and q, specify the sets of all
0 0
0
two-person income distributions that are inequality equivalent to x x1 ; x2 with
respect to the absolute, intermediate, and relative notions of inequality. Observe
that for 0 < < 1, b lies between a and q. As the value of increases (decreases) to
one (zero), we become more concerned about relative (absolute) invariance.
As we have argued, the rightist concept of inequality benefits the richer section
of the society and the leftist notion benefits the poorer section in two different ways.
However, the free parameter enables the planner to decide on the trade-offs
between the increase and decrease of inequalities implied by the above two extreme
notions. In other words, a centrist position, more precisely the BossertPfingsten
position, provides an instrument in the planners hand.
However, the BossertPfingsten approach has some shortcomings. For instance,
in some situations for a very low value of , the associated line of invariance may be
closer to the relative invariance ray than the absolute invariance ray. This is simply
the opposite of what is expected for a low value of the parameter . Del Rio and
Alonso-Villar (2008) have investigated this issue in details. As the size (total
income) of a distribution increases, the BossertPfingsten notion approaches the
relative view (see also Seidl and Pfingsten (1997) and Del Rio and Ruiz Castillo
(2000)). Therefore, one can argue that the results obtained using the -inequality
invariance concept can be almost indistinguishable from that when one chooses the
relative view of inequality. They demonstrated that the inequality invariance line
for the distribution x 20; 80 corresponding to 0:5 almost coincides with that
corresponding to the relative concept of inequality. That is, inequality remains
almost unchanged if the free parameter that gives rise to intermediate invariance as
a convex combination of relative and absolute notions takes on the values that
generate the relative case and a pure intermediate situation. Thus, we do not have
any additional information on inequality equivalence by shifting our analysis from
the extreme situation 1 to the intermediate situation 0:5. This undesirable
feature of -inequality invariance concept raises doubts about the interpretation of
the parameter : In contrast, Del RioRuiz-Castillos (2000) v; notion of
inequality invariance, where v is an element of the n-dimensional simplex that
fixes the distribution of reference and 0 1 is a parameter that combines
relative and absolute rays associated with v as a convex combination, does not
suffer from this problem. However, a clear merit of the -inequality invariance
notion is that it has established itself as a standard invariance concept through its
long usage and its easy exposition. Furthermore, one can visualize it graphically
quite quickly. In order to use it, the value of should be chosen such that this kind of
problem can be avoided.
In an interesting contribution, Zoli (1999) characterized a general notion of
intermediate invariance, which demands that for any income distribution x in a
x1n
given domain, x should be equally unequal to the distribution x
, where
x1
0 1 and 0 1 are constants. For 0, the Zoli invariance condition
coincides with the absolute inequality criterion for all values of , whereas for
1, it represents the relative notion of inequality. The Krtscha (1994)

criterion drops out as a particular case of the Zoli condition for 1 and
0 < < 1. The Zoli criterion gets related to the BesleyPreston (1988) intermediate condition if 0 < < 1 and 1.
It may be important to note that the relative and absolute notions of inequality
invariance are independent, that is, we cannot find any non-negative valued nonconstant function defined on the set of income distributions that satisfies these two
notions simultaneously. Since the intermediate view contains the relative and
absolute views as special cases, components of each of the two pairs (relative,
intermediate and absolute, intermediate) are not independent.
Zheng (2007a) introduced the notion of unit consistency that enables us to rank
income distributions when incomes are expressed in different units. An inequality
index I : D ! R1 is called unit consistent if for all n 2 N, x, y 2 Dn , I x < I y
implies I cx < I cy; where c > 0 is any scalar. Suppose x and y are income
distributions of countries A and B, respectively, and x is more unequal than y when
they are expressed in the currency of country A. Now, if the two distributions are
converted using the currency of country B, then the unit consistency axiom
demands that x should still be more unequal than y. Thus, the unit consistency
axiom does not lead to contradictory conclusions in inequality rankings under
changes in the unit of measurement of incomes. Evidently, a relative inequality
index is unit consistent. However, the converse is not true. We will demonstrate in
Sect. 1.4 that there exits an absolute inequality index that satisfies unit consistency.
Thus, unit consistency is not independent of relative and absolute inequality
concepts.
Del Rio and Alonso-Villar (2008) have also analyzed the issue of simultaneous
satisfaction of the unit consistency axiom and -inequality invariance concept. The
slope of the inequality invariance line x 1 1n is increasing with the total
income in x, given that the extent of relative inequality remains unaltered. Observe
that the invariance lines are not parallel. Consequently, it is not possible to claim
that -inequality rankings remain the same when the currency unit changes.
However, when incomes are measured in a single currency, the unit consistency
axiom and the -inequality invariance concept can be employed simultaneously.
Such a situation can arise when we compare inequalities of two regions of a country
where incomes are measured in the same currency.
For a meaningful comparison of inequality across distributions, the underlying
population sizes should be the same. While using an inequality invariance property
we can make the sizes of two different distributions over a given population size
identical, the following property shows how we can compare inequalities of two
distributions with different population sizes.
Daltonian Principle of Population (DPP) The inequality index I : D ! R1 is
population replication invariant.
Suppose our objective is to compare inequalities of two distributions x1 and x2
with population sizes n1 and n2, respectively. Such a situation may arise for
inequality comparison of two different societies or of the same society over two
periods. If we consider a n2 fold replication of x1 and a n1 fold replication of x2,
then the replicated distributions, which we denote by y1 and y2, respectively, have a
common population size of n1n2. If the inequality index satisfies DPP, then comparison of inequalities of y1 and y2 is essentially the same as comparison of
inequalities of the original distributions x1 and x2.
If we assume that any characteristic other than income, say the names of the
individuals or their marital status, is irrelevant to the measurement of inequality,
then a property that can ensure this is symmetry.
Symmetry (SYM) The inequality index I : D ! R1 is symmetric.
Using SYM, we can define the inequality index directly on nondecreasingly or
nonincreasingly ordered income distributions.
An inequality index should be sensitive to any redistribution of income between
two persons. That is, it should respond properly if there is a progressive (regressive)
transfer between two persons.
PigouDalton Transfer Principle (PDT) For all n 2 N, y 2 Dn , if x is obtained
from y by a progressive transfer, then I x < I y.
Likewise, the value of an inequality index should increase under a regressive
transfer. If an inequality satisfies SYM, then only rank-preserving transfers are
allowed.
A progressive transfer of income can be stated alternatively using a Pigou
Dalton matrix. We may consider an example to motivate this seemingly different
formulation of a progressive transfer. Let z1 3; 5; 7 and z2 4; 5; 6 be two
income distributions so that z2 is obtained from z1 by a progressive transfer of 1 unit
of income from the richest person to the poorest person. Now, the transformation
that
takes
us2 0 from
z1
to
z2 1can
be
expressed
as
1
0
3
1 0 0
0 0 1
4; 5; 6 3; 5; 7434@ 0 1 0 A 14@ 0 1 0 A5:The first matrix within the
0 0 1
1 0 0
third bracketed term on the right-hand side is a 3 3 identity matrix whose diagonal
entries are 1 and non-diagonal entries are 0, and the second matrix is a permutation
matrix of order 3, obtained by interchanging the first and third coordinates of the
identity matrix. In other words, this permutation matrix is one in which entries in
the first-row third column and third-row first column are 1 and in the second row
only diagonal entry is one. The second row corresponds to the person unaffected by
the transfer. Post-multiplication of (3, 5, 7) with a weighted average of these two
3
1
3 3 matrices, where the weights are, respectively,
and , generates the
4
4
distribution
(4,
5,
6)
from
the
distribution
(3,
5,
7).
The
weighted
average
0 0
1
0
11
1 0 0
0 1 0
@3@ 0 1 0 A 1@ 1 0 0 AA is known as a PigouDalton matrix of
4
4
0 0 1
0 0 1
order 3. (In the literature, it is also known as a T-transformation; see Marshall

et al. 2011.)
In general, for any two income distributions x^ and y^ over a given population size
n, if x^ is obtained from y^ by a progressive transfer from person j to person i, then
x^ y^ M, where M tInn 1 tij , 0 < t < 1, I nn is the n n identity matrix,
and ij is the permutation matrix of order n that interchanges the i and j coordinates.
The n n matrix M is a PigouDalton matrix of order n. One can easily check that
x^ i t^
y i 1 t^
y j , x^ j 1 t^
y i t^
y j , and x^ l y^ l for all l 6 i, j (see Marshall
et al. 2011; Weymark 2006; Chakravarty 2009a).
Strict S-convexity implies SYM and reduction in inequality under rankpreserving progressive transfers. Conversely, under SYM, only rank-preserving
progressive transfers are permissible, and inequality reduction under such a transfer
implies strict S-convexity (Hardy et al. 1934; Dasgupta et al. 1973; Marshall
et al. 2011, pp. 1567). From this, it follows that strict S-convexity of an inequality
index is sufficient to incorporate egalitarian bias into ethical assessments of income
distributions.2
There is no inequality in a distribution if each individual possesses the same
level of income. In such a case, the value of an inequality index should be zero. The
converse is also assumed, that is, if the value of the inequality index is zero, then
income is equally distributed across persons. This cardinal property of an inequality
index is referred to as normalization.
Normalization (NOM) For all n 2 N, x 2 Dn ; I x 0 if and only if x c1n for
some c > 0:
Since we have assumed at the outset that an inequality index is nonnegative
valued, NOM implies that the level of inequality of an unequal distribution is
positive.
Minor observational errors in income should not change the value of an inequality index abruptly. The following postulate ensures this requirement.
Continuity (CON) For all n 2 N, I is a continuous function on Dn.
In order to motivate the study of some implications of the above postulates
involving relative inequality indices, we first present some examples of such
indices. (Analogous discussion on orderings involving absolute and intermediate
inequality indices is relegated to the next section.) A well-known relative inequality
index that verifies DPP, SYM, PDT, and NOM is the Gini index IG, which for any
n 2 N, x 2 Dn is defined as
Variants of PDT were suggested, among others, by Fleurbaey and Michal (2001), Chateauneuf
and Moyes (2006), and Chakravarty (2009a, Chapter 3). Fleurbaey and Maniquet (2011) provide
discussion on several related issues.
10
8
n X
n

>
1 X
>
xi xj
>
>
2
>
2n x i1 j1
>
>
>
>
n X
n
>

>
1 X
>
>
min xi ; xj
>1 2
>
>
n
>
i1 j1
>
>
< 1 X
n X
n

I G x
max xi ; xj min xi ; xj :
2
> 2n x i1 j1
>
>
>
>
n
>
1 X
>
>
>
1 2
2i 1xi
>
>
n x i1
>
>
>
n
>
>
1 X
>
>
xi
>
: 1 n2 x 2n i 1^
1:4
i1
When a person in a society compares his/her income with that of a richer person,
he/she may have a feeling of unhappiness, which may be specified in terms of the
difference between the two incomes. Given the population size and the mean
income, the average of all such extents of unhappiness in all pairwise comparisons
n P
n

P
xi xj (Foster and Sen
becomes directly related to the Gini index 2n21x
i1 j1
1997; Chakravarty 2009a). Equivalence between 1 n2 1x
n P
n
P

min xi ; xj and
i1 j1
n P
n

P
1
xi xj can be established using the fact that min xi ; xj
2
2n x
i1 j1

xi xj xi xj
xi xj xi xj
: Since max xi ; xj
, it follows that xi xj
2
2
n P
n

P
xi xj shows
max xi ; xj min xi , xj , which on substitution into 2n21x
i1 j1
that I G x
1
2n2 x
n P
n

P
max xi ; xj min xi ; xj . If for any two persons i and j,
i1 j1
we measure
of the income distribution (xi, xj) by

inequality

max xi , xj min xi , xj , the excess of the maximum income over the minimum
income, then given the population size and the mean income, the average of
all such excesses for all two-person income distributions of the type (xi, xj), where
i, j 1, 2, . . . , n; leads us to the Gini index. For welfare-ranked permutation of
incomes, IG is a linear function of incomes with the weights being the odd integers
in increasing order. Increasingness of the weight sequence f2i 1g is necessary and
sufficient for IG based on x to satisfy PDT. Similarly, when the income distribution
is illfare ranked, the necessary and sufficient condition for the satisfaction of PDT
by IG is decreasingness of the sequence f2n i 1g.
Since there is no strong support in favor of using the weight sequence f2i 1g in
x, we can as well construct a generalized Gini index by taking any arbitrary positive
increasing sequence {ani } of weights, where the superscript n in ani shows explicit
dependence of the weights on the population size. This generalization of the
11
Gini index, which has been suggested and characterized by Weymark (1981),
is defined as
I WG x 1
x
1
n
X
n
X
ain xi :
1:5
ain i1
i1
If income of each individual is broken down into two or more sources and the rank
order of incomes is the same for each type of income, then the overall generalized
Gini inequality index is a weighted average of component-wise generalized Gini
inequality indices, where the weights are the proportions of total incomes from
different sources.
Donaldson and Weymark (1980) noted that a natural restriction on the members
of IWG is the requirement that weights ani are independent of n. They assumed that
a1 1 and ain ai for all n 2 N, maintaining the increasingness assumption
of {ani }. This defines a single-series Gini index, a subclass of the generalized
Gini indices in (1.5). For the single-series Gini, they defined the function
f : N [ f0g ! R1 by
8
m0
>
<0
m
X
:
1:6
f m
ai , m 2 N
>
:
i1
It then follows that ai f i f i 1. Donaldson and Weymark (1980) demonstrated that the resulting single-series Gini index satisfies DPP if and only if
f i i , where > 1 is a parameter. This defines a single-parameter Gini index,
popularly known as the S-Gini index, given by
x 1
I DW
n h
i
1 X
i
i

1
xi :
xn i1
1:7
The inequality restriction > 1 is necessary and sufficient for I DW

to satisfy PDT. For
2, I DW coincides with the Gini index. As increases, more weight is assigned to

minfxi g
i
; the
approaches 1
lower incomes in the aggregation. As ! 1, I DW
x
relative maximin index of inequality. Donaldson and Weymark (1980) also characterized a family similar to (1.7) based on illfare-ranked incomes.3
Donaldson and Weymark (1983) studied several properties of the continuous version of I DW
. This
was also investigated further by Yitzhaki (1983) and Kakwani (1980a). Chakravarty (1988)
suggested a family of generalized Gini indices, called E-Gini indices, that satisfies the diminishing
transfer principle, a postulate under which more weight is attached to transfers at the lower end of
the distribution. Ebert (1988a, b, c) characterized two families of inequality measures which are
generalizations of the Gini index.
12
Bonferroni (1930) suggested an index, which is also linear with a given rank
order of incomes. It is based on the comparison of the partial means and the general
mean of an income distribution and is defined as
I B x 1
n
i
1 X
1X
x^ j :
nx i1 i j1
1:8
It satisfies PDT and a stronger redistributive principle known as the principle of

positional transfer sensitivity, which demands that a progressive income transfer
between two individuals with a given rank difference will reduce inequality by a
larger amount the lower the income of the donor is (Mehran (1976), Zoli (1999),
and Aaberge (2000, 2007)). However, the Bonferroni index IB(x) is a violator of
DPP so that it is not suitable for cross-population comparisons of inequality
(Chakravarty 2007, 2009a).
Two other well-known relative inequality indices that fulfill all the postulates
considered in the section are the coefficient of variation ICV and the Atkinson (1970)
index I , where for all n 2 N, x 2 Dn ,
I CV x
s
n
1X
xi 2
n i1
1:9
and, for all n 2 N, x 2 Dn ,
I A x
8
>
>
>
>
>
>
>
>
>
<1
n
1X
x
n i1 i
n
>
Y
1
>
>
>
xi n
>
>
>
>
i1
>
,
:1
x
!1
,
< 1,
6 0,
1:10
0:
While for IG the effect of progressive transfer between two persons depends on their
rank difference, for ICV it depends on the difference between the incomes of the
concerned persons. The parameter in I A represents transfer sensitivity in the sense
that under a progressive transfer, the index will decrease by a larger amount the
lower is the income of the recipient of the transfer. The reduction in the value of I A
resulting from a progressive transfer will be higher the lower is the value of . For a
given distribution x , an increase in the value of decreases inequality unambiguously. As ! 1, I A approaches the relative maximin index of inequality.
13
Observe that IG, IB, and ICV are compromise relative indices in the sense that when
multiplied by the mean income, they become absolute inequality indices. These
three absolute indices are, respectively, the absolute Gini index, the absolute
Bonferroni index, and the standard deviation. However, I A is not a compromise
index.
There is no a priori reason why different inequality indices should rank
two different income distributions in the same way. For instance, for the two
11
distributions xo 4; 18; 10; 8 and xp 3; 15; 24; 18, I G xo I G xp ,
40
r
13
I CV x0 I CV xp
, but for 0, I A xo 0:129 and I A xp 0:212.
50
Thus, while the Gini index and the coefficient of variation regard the two distributions as equally unequal, a member of the Atkinson family rank them in a
different way.
We can use an ordering that relies on the well-known Lorenz curve for judging
whether different inequality indices can rank alternative distributions of income in
the same direction. The Lorenz curve of a distribution is the plot of the proportions
of total income possessed by the bottom t 0 t 1 proportions of the population.
j
P
For any given income distribution x 2 Dn ,
x^ i
i1
nx ,
the share of the total income
j
n
population proportion, is the ordinate of the Lorenz

enjoyed by the bottom
j
curve of x at n, where n 2 N is arbitrary. Then, the Lorenz curve of x,
0 j
0j1 1
1
P
P

x^ i C
x^ i C
B
B
j
Bi1 C
Bi1 C
LCx; t, t 2 0; 1, is defined as L x;
1 B
C B
C,
@nxA
@nxA
n
where LCx; 0 0, 1 j n 1, and 0 1. LCx; 0 0 ensures that
the graph is closed. Note that LCx; 1 1. When all the incomes are equal so that
every tth percentile of the population holds t % of the total income, the curve
becomes a straight line with a slope of 45 , and in this case, it is known as the
line of equality. For an unequal distribution, the curve falls below the line of
equality, and as the distribution becomes more and more unequal, the curve
moves more and more to the horizontal axis until we have the extreme situation
where only one person has positive income and all other persons have zero income.
This extreme situation is the case of maximal inequality and here the curve will
coincide with the horizontal axis for all the persons with zero income, and then, for
the richest person, it will rise perpendicularly. The Lorenz curve is increasing and
strictly convex (see Kakwani 1980b; Chakravarty 1990). Twice the area enclosed
between the Lorenz curve and the line of equality is known as the Lorenz ratio.
Interestingly enough, the Gini index coincides with the Lorenz ratio.
Often, it becomes useful to look at Lorenz curves of income distributions defined
for a continuum of population. Consider an income distribution represented by the
14
cumulative income distribution function F : 0; 1 ! 0; 1. For any v 2 0; 1,

F(v) gives the proportion of persons with income less than or equal to v . By
definition, the distribution function F is nondecreasing, F0 0 and FvF 1
for some vF < 1. The inverse distribution function F1 : 0; 1 ! 0; 1 is defined
as F1 t inf fv : Fv tg for all t 2 0:1. The inverse function F1 is also
known as the quantile function associated with the cumulative income distribution
function F. For any t 2 0:1, F1 t is the income of the person at the tth percentile.
R1
The mean of F is F1 d. The Lorenz curve L(F, t) of F is the share of the
0
total income enjoyed by the poorest t proportion of the population:

Rt 1
F d
LF; t 0
(Gastwirth 1971).
For x, y 2 D, we say that x Lorenz dominates y (xL y, for short) if Lx; t L
y; t holds for all 0 t 1, with > for some 0 < t < 1. That is, the Lorenz curve of
x is nowhere below and at some places (at least) inside that of y. This definition of
xL y does not require equality of the total incomes and population sizes of the
distributions. The Lorenz dominance relation L is transitive and incomplete, where
transitivity demands that for any x1 , x2 , x3 2 D, if x1 L x2 and x2 L x3 hold, then
x1 L x3 must hold. On the other hand, incompleteness is demonstrated by the
observation that we can get two distributions x and y for which the Lorenz curves
cross so that neither xL y nor yL x holds.
In Fig. 1.2, the Lorenz curves of three income distributions x1, x2 and x3
are shown. In the figure, we note that both x1 L x3 and x2 L x3 hold, but none of
x1 L x2 and x2 L x1 holds. Thus, L is a quasi-ordering, that is, it is transitive but
incomplete. One can check that the Lorenz curves of the distributions xo and xp
cross. However, the distribution (8, 4, 16, 12) Lorenz dominates the distribution
(27, 6, 15, 12).
For any x, y 2 D, xL y holds if and only if I x < I y for all relative inequality
indices I : D ! R1 which satisfy DPP, SYM, and PDT (Foster 1985; Chakravarty
1990, 2009a and Foster and Sen 1997). The result shows that of two income
distributions, if one Lorenz dominates the other, then calculation of any inequality
index is not necessary to judge whether the former has lower inequality than the
latter. However, as we have seen, if the Lorenz curves of the two distributions cross,
then no such an unambiguous conclusion about inequality ranking can be deduced.
It is possible to have two different inequality indices for which the distributional
rankings will not coincide. If the distributions under consideration have the same
mean (population size), then the result holds without the scale invariance condition
(1.1) (DPP). In particular, for x, y 2 Dn , where x y, xL y holds if and only if
I x < I y for all I : Dn ! R1 that satisfy SYM and PDT, that is, strict
S-convexity.
1.4 Ethical Approaches to the Measurement of Inequality
15
Line of equality
x1
x2
Cumulative
proportions
of income
x3
Cumulative population shares
Fig. 1.2 Lorenz curve and Lorenz dominance
1.4
Ethical Approaches to the Measurement of Inequality
Simple comparison of the income inequality levels of two distributions (using some
index of inequality) does not take into account the differences in mean incomes of
the distributions and therefore neglects an important factor which affects the wellbeing of a population. The influence of mean income on well-being of a population
may be so high that it can reverse the inequality ranking between two populations.
To illustrate this, let us consider the income distributions x 9; 3; 9; 3; 9; 3 and
1
e
x 20; 4; 20; 4; 20; 4. The values of Gini index for these two distributions are
4
1
and , respectively. In fact, x Le

x for all relative
x holds and hence I x < I e
3
6
1
inequality indices I : D ! R that satisfy SYM and PDT. But as we will see later
in this section, the value of any welfare function satisfying certain desirable criteria
is lower for the former distribution than that for the latter one.
From our observations in Sect. 1.2, we can say that for any x, y 2 Dn , where
x y, xL y holds if and only if W x > W y, where I W is strictly
S-concave. Since the value of W increases under a rank-preserving progressive
transfer, we can regard W as a social welfare function and strict S-concavity of a
welfare function as a postulate that shows preference for equity. Likewise, a social
welfare function should demonstrate preference for efficiency as well, that is, of
two income distributions, if one is obtained from the other by some simple
increment(s), then social welfare for the former should be higher than that of the
latter. We say that a social welfare function W : D ! R1 is increasing or satisfies the
16
strong Pareto principle if W x > W y where x is obtained from y 2 Dn by a simple

increment, n 2 N being arbitrary.4
The following theorem, whose detailed proof can be found in Foster and
Shorrocks (1988), enables us to rank income distributions using welfare functions
satisfying the strong Pareto principle.
Theorem 1.1 Let x1 and x2 be two arbitrary income distributions over the population sizes n1 and n2, respectively. That is, let x1 2 Dn1 and x2 2 Dn2 be arbitrary.
Denote the respective n2- and n1-fold replications of x^ 1 and x^ 2 , the illfare-ranked
permutations of x1 and x2, respectively, by x^ 3 and x^ 4 . Assume that x^ 3 and x^ 4 are
unequal, that is, x^ 3i 6 x^ 4i for at least one i 1, 2, . . . , n1 n2 . Then, the following
conditions are equivalent:
(i) x^ 3 can be obtained from x^ 4 by a finite, nonempty sequence of rank-preserving
simple increments.
(ii) x^ 3 x^ 4 , that is, x^ 3i x^ 4i for all i 1, 2, . . . , n1 n2 , with > for some i.
(iii) W x1 > W x2 for all social welfare functions W : D ! R1 that are increasing, symmetric, and population replication invariant.
n1
n2

P
P
(iv) n11 U x1i > n12 U x2i ; where the individual utility function U is
i1
i1
increasing.
Since an increase in a persons income is welfare increasing, from (ii) it follows
that W x^ 3 > W x^ 4 . By the population replication invariance property of W, we
have W x^ 3 W x^ 1 and W x^ 4 W x^ 2 . Since W is also symmetric, W x^ 1
W x1 and W x^ 2 W x2 . Thus, W x1 > W x2 : The converse can also be
checked similarly. Condition (iv) is essentially a restatement of condition (iii) under
additivity of the welfare function. It says that x1 is preferred to x2 by the symmetric
average utilitarian rule where the individual utility functions are increasing.
Condition (ii) says that x^ 3 Pareto rank dominates x^ 4 (Saposnik 1981, 1983).
Assume for simplicity that the inequality x^ 3i x^ 4i > 0 holds for i j and
for all other values of i equality holds. Then, we can definitely arrive at x^ 3 from
x^ 4 by a finite, nonempty sequence of simple increments. Likewise, (i) implies (ii).
It may be noted that x^ 3; 3; 3; 9; 9; 9 is Pareto rank dominated by
^
e
x^ 4, 4, 4, 20, 20, 20, and hence x has lower welfare than x by all symmetric
increasing welfare functions. A sufficient condition for Pareto rank dominance is
Pareto absolute dominance. The distribution x^ 3 is said to Pareto absolute dominate
the distribution x^ 4 if x^ 3 x^ 4 , where is any permutation matrix of order
n (McMlelland and Rohrbaugh 1978). If is the n n identity matrix with ones on
In contrast, W : D ! R1 satisfies the weak Pareto principle if W x > W y, whenever each

income in x is obtained from the corresponding income in y 2 Dn by a simple increment, that is,
xi > yi for all i 1, 2, . . . , n, where n 2 N is arbitrary. Evidently, the strong principle implies the
weak principle.
17
the diagonal and zeros elsewhere, then x^ 4 x^ 4 , and hence Pareto absolute
dominance becomes Pareto rank dominance. If is the matrix obtained by
4
4
interchanging the ith and (n i 1)th rows of the identity
4 matrix, then x^ x ,
4
3
3
where 1 i n, so that mini x^ i x^ 1 x1 maxi xi .
Conditions stipulated in Theorem 1.1 are concerned with higher efficiency, more
is preferred to less. But higher efficiency may be accompanied by higher inequality.
To illustrate this, consider the distributions u1 0; 1; 8 and u2 0; 1; 5. Then,
u1 Pareto rank dominates u2, and if we regard the mean income of a distribution as
the welfare function, then their welfare values are 3 and 2, respectively. However,
5
the value of Gini index for u1 is 16
27, which is higher than 9, the Gini index value for
2
u . Therefore, the welfare function should also be concerned with equity. To discuss
this further, we first explain the role of the stochastic dominance criteria as an
ordering device.
The stochastic dominance rules are used extensively in finance for ranking
uncertain prospects on the basis of their expected returns. They can also be used
for ranking alternative distributions of income with respect to social welfare.5 In
order to define these criteria formally, we assume continuous-type income distributions. Let F : 0; 1 ! 0; 1 be the cumulative income distribution function.
Assume that F is continuously differentiable. We can define the integrals F2(s) and
F3(s) using the following recurrence relation:
Zs
Fr s
Fr1 vdv for all s 2 0; 1;
1:11
where r 2 is an integer and F1 s Fs. For all n 2 N, x 2 Dn , the discrete

nP
s;x
s x^ i r1 for all
counterpart to any Fr(s) in (1.11) is given by Frd s r11 !n
i1
s 2 0; 1, where ns; x #fi 2 f1; 2; ::; ngjx^ i sg is the number of persons with
incomes less than or equal to s in the distribution x^ . To illustrate this, let
x 2; 5; 3. Then, x^ 2; 3; 5, and for s 2 0; 2, F1 s F2 s F3 s 0:
2
s2
For s 2 2; 3, F1 s 13, F2 s s2
6 . When s 2 3; 5,
3 , and F3 s
2
s2
s32
2s2 10s13
F1 s 23, F2 s s23 s3 2s5

,
and
F
.
3
6
6
3
3s2 20s38
3s10
Finally, for s 2 5; 1, F1 s 1, F2 s 3 , and F3 s
.
6
Of two income distributions F and G defined on 0; 1, F is said to rth order
stochastic dominate G if and only if Fr s Gr s for all s 2 0; 1 with < for some
R1
R1
s 2 0; 1. This is equivalent to the condition that U zdFr z > U zdGr z
0
5
See, for example, Hadar and Russell (1969), Whitmore (1970), Fishburn and Lavalle (1995),
Foster and Sen (1997), Levy (2006), Chakravarty and Zoli (2012), and Chakravarty (2013).
18
for all utility functions U : 0; 1 ! R1 such that 1i1 Ui > 0, where

i 1, 2, . . . , r and Ui is the ith order derivative of the utility function U. (We assume
that U is differentiable up to any desired degree.) First-order stochastic
dominance of F over G means that F is preferred to G by the symmetric average
utilitarian rule showing preference only for higher incomes (Hadar and Russell
1969). Observe that the stochastic dominance relations based on discrete distribution functions are population replication invariant so that they can be used for
ranking distributions over differing population sizes. Each of the four conditions
stated in Theorem 1.1 is equivalent to the statement that x3, the n2-fold replication
of x1, first-order stochastic dominates x4, the n1-fold replication of x2. A lower-order
stochastic dominance implies all its higher-order stochastic dominances, but the
converse is not true.
In order to identify the conditions under which distributional judgments using
the symmetric average utilitarian rule become equivalent to the second-order
stochastic dominance criterion, we consider the generalized Lorenz curve, the
product of the mean income, and the Lorenz curve. Formally, the generalized
Lorenz curve GL(x, t) of the distribution x 2 Dn is defined as xLCx; t
(Kolm 1969; Shorrocks 1983). For x, y 2 D, x is said to generalized Lorenz
dominate y (xGL y, for short) if we have GLx; t GLy; t for all 0 t 1,
with > for some t. By construction, the generalized Lorenz dominance relation GL
is a quasi-ordering. In Fig. 1.3, the generalized Lorenz curves of three distributions x4, x5 and x6 are depicted, and we have both x4 GL x5 and x4 GL x6 , but neither
x5 GL x6 nor x6 GL x5 holds.
The following theorem shows that many seemingly unrelated conditions for
welfare ranking of income distributions with different mean turn out to be equivalent.
Fig. 1.3 Generalized Lorenz curve and generalized Lorenz dominance
19
Theorem 1.2 x1 and x2 be two arbitrary income distributions over two populations
of sizes n1 and n2, respectively. That is, let x1 2 Dn1 and x2 2 Dn2 be arbitrary.
Denote the respective n2- and n1-fold replications of x^ 1 and x^ 2 , the illfare-ranked
permutations of x1 and x2, respectively, by x^ 3 and x^ 4 . Assume that x^ 3 and x^ 4 are
unequal. Then, the following conditions are equivalent:
(i) x^ 3 can be obtained from x^ 4 by a finite, nonempty sequence of rank-preserving
simple increments and/or rank-preserving progressive transfers.
(ii) x1 GL x2 :
(iii) W x1 > W x2 for all social welfare functions W : D ! R1 that are increasing, strictly S-concave and population replication invariant.
n1
n2

P
P
(iv) n11 U x1i > n12 U x2i ; where the individual utility function U is
i1
i1
increasing and strictly concave.

(v) x3 second-order stochastic dominates x4, where x3 and x4 are, respectively,
the n2 and n1-fold replications of x1 and x2.
(vi) x^ 3 x^ 4 A for some bistochastic matrix of order n.
(vii) There exists a set {M1, M2., . . ., Mk} of finitely many PigouDalton matrices
of order n1n2 such that x^ 3 x^ 4 M1 M2 . . . Mk .
Theorem 1.2 has been stated and proved in the literature in different forms
(Kolm 1969; Shorrocks 1983; Marshall et al. 2011; Foster and Shorrocks 1988;
Chakravarty 2009a). Equivalence between conditions (iv) and (v) means that
second-order stochastic dominance of x3 over x4 is the same as the requirement
that x1 is regarded as better than x2 by the symmetric average utilitarian rule, which
clearly indicates preference for efficiency as well as equity6 (Hadar and Russell
1969). This is tantamount to the statement that there is a unanimous welfare verdict
for x1 over x2 according to all increasing, population replication invariant, symmetric, and equality preferring social welfare functions (condition (iii)). The additional
equivalence with the generalized Lorenz ordering offers an implementation process
for second-order stochastic dominance; one just needs to check if the two generalized Lorenz curves cross. The first condition, a component-wise vector comparison,
shows how we can move from one distribution to another by a sequence of transformations involving simple increments and/or equity. The post-multiplication of
x^ 4 by the bistochastic matrix A makes the resulting distribution x^ 4 A u^ , say,
more equal than x^ 4 in the sense that u^ can be obtained from x^ 4 by a sequence of
6
We do not discuss here higher-order dominances. For third-order stochastic dominance, the
equivalent condition for ranking by the symmetric average utilitarian rule requires that the
marginal utility is positive, decreasing, and strictly convex (see Whitmore 1970; Foster and Sen
1997; Levy 2006; Shaked and Shanthikumar 2006; and Chakravarty 2009a, 2013).
20
rank-preserving progressive transfers (Dasgupta et al. 1973). This is same as the

condition that u^ L x^ 4 . Since x^ 4 and u^ are two distributions of the same total income
over and a population of size n1n2, u^ L x^ 4 is the same as the stipulation that u^ GL
x^ 4 holds.7 Now, x^ 3 u^ means that x^ 3i u^ i for all i, with > for some i,
1 i n1 n2 . Therefore, it follows immediately that x^ 3 GL u^ holds. By transitivity
of GL , we then have x^ 3 GL x^ 4 , which, by the population replication invariance
property of the generalized Lorenz curve, is equivalent to x1 GL x2 . The converse is
true as well, that is, condition (ii) implies condition (vi). (A formal demonstration of
the equivalence between conditions (ii) and (vi) can be found in Marshall
et al. (2011, p. 156).) The distribution x^ 4 M1 becomes more equal than x^ 4 under
post-multiplication by the PigouDalton matrix M1 (see Marshall et al. (2011, p. 32)
and Sect. 1.3 of this chapter). Next, x^ 4 M1 , when post-multiplied by the
PigouDalton matrix M2, generates the distribution x^ 4 M1 M2 , which is more equal
than x^ 4 M1 . Hence, x^ 4 M1 M2 is more equal than x^ 4 . Continuing this way, we can
conclude that x^ 4 M1 M2 . . . Mk v^ , say, is more equal than x^ 4 , that is, v^ L x^ 4 .
The inequality x^ 3 v^ along with the facts that v^ L x^ 4 and the means of v^ and x^ 4
are the same leads us to the observation that x^ 3 GL x^ 4 , which is x1 GL x2 . The
converse is also true, that is, condition (ii) implies condition (vii) (see Marshall
et al. 2011, p. 43).
^
For the two distributions x 9; 3; 9; 3; 9; 3 and x 20; 4; 20; 4; 20; 4
^
considered
at the beginning of the section, it follows that x GL x . Consequently,
W x
> W x for all equity-oriented (strictly S-concave) social welfare func-
tions that also indicate preference for higher efficiency (mean). The higher mean of
^
x is sufficient to offset its higher inequality so that x has unambiguous welfare
dominance over x .
If the mean incomes of the two distributions x1 and x2 in Theorem 1.2 are the
same, then x1 GL x2 coincides with x1 L x2 . This is the same as the condition that
W x1 > W x2 for all social welfare functions W : D ! R1 that are strictly
S-concave and population replication invariant. Thus, when efficiency considerations are absent (mean is fixed), the Lorenz ordering agrees with welfare ranking
for all welfare functions that are equity oriented and remain invariant under income
by income replications of the population (Dasgupta et al. 1973). Equivalence of this
with the similar ranking of the distributions generated by the average symmetric
utilitarian rule, where the identical utility function is strictly concave, was demonstrated by Atkinson (1970). Thus, an advantage of GL over L is that the former
In general, the Lorenz and generalized Lorenz orderings of two distributions coincide if their
mean incomes are the same, even if the population sizes corresponding to the two distributions are
different.
21
can be used for welfare ordering of distributions even when the means of the
distributions are not identical.
Observe that Pareto absolute dominance and Pareto rank dominance imply GL
but not L . Also note that L is neither necessary nor sufficient GL . However,
xL y along with x > y is sufficient for xGL y.
The two distributions (0, 1, 5) and (1, 1, 4.9) cannot be ranked by the Pareto rank
dominance criterion, but the latter generalized Lorenz dominates the former. This
establishes that the generalized Lorenz dominance criterion does not imply the
Pareto rank dominance rule. This example also displays that the Pareto rank
dominance rule is an incomplete relation, although it is transitive.
Moyes (1987) developed a highly interesting application of GL . He defined the
generalized Lorenz curve of the distribution x x1n as the absolute Lorenz
curve of x. At any population proportion t, the negative of this curve indicates a
normalized value of the shortfall of total income of the bottom t % of the population
from that in the egalitarian distribution in which everybody enjoys the equal
income, where the normalization is done by dividing with the population size.
Moyes showed that an unambiguous ranking of two income distributions with
respect to absolute inequality indices can be obtained by comparing their absolute
Lorenz curves.8
In order to present a discussion on welfare theoretic approaches to the measurement of inequality, we assume that W : D ! R1 is continuous (i.e., for all n 2 N,
W is continuous on Dn), is strictly S-concave, and satisfies the strong Pareto
principle. We refer to such a welfare function as regular. Continuity will ensure
that minor changes in incomes will generate minor changes in welfare.
The Atkinson (1970)Kolm (1969)Sen (1973) equally distributed equivalent
(ede) income xe corresponding to x 2 Dn is defined as that level of income which, if
given to everybody, will make the existing distribution ethically indifferent.
Formally,
W xe :1n W x:
1:12
Given regularity of W, we can solve (1.12) uniquely for xe and write it as xe Ex,
where E, being a particular numerical representation of W, is also regular. Further,
xe < x : Since E represents an average, Ec1n c for all c > 0 :
In order to interpret relative indices ethically, we assume that W is regular
and

homothetic. According to homotheticity, for all n 2 N, x 2 Dn , W x W x ;
where W is linear homogeneous, that is, W cx cW x, where c > 0 is any
constant and is increasing in its argument. The Atkinson (1970)Kolm (1969)
Sen (1973) relative index is then defined as
8
For intermediate indices, similar orderings were considered by Chakravarty (1989) and Del Rio
and Ruiz Castillo (2000). Yoshida (2005) applied Krtschas intermediate concept to the Lorenz
ordering. Zheng (2007a) showed that the Krtscha-type Lorenz ordering is the only intermediate
Lorenz ordering that satisfies unit consistency.
22
I AKS x 1
xe
W x
1
;
x
W 1n x
1:13
so that here Ex WW1xn : Since x is linear homogeneous, IAKS is a relative index.

It is continuous, strictly S-convex, and bounded between zero and one, where the
lower bound is achieved if all the incomes are equal. It indicates the fraction of total
income that could be saved if the society distributed incomes equally without any
welfare loss (see Blackorby and Donaldson 1978).
Using (1.13) and (1.12), we can express E(x) as Ex x1 I AKS x. As
noted, the function E itself or any increasing transformation function of it can be
regarded as a regular social welfare function. Thus, E implies and is implied an
inequality index. This welfare function is represented as an increasing function of
the product of the mean income and the index of equality, the shortfall of the
inequality index IAKS from unity. It expresses welfare as a trade-off between equity
and efficiency (mean income). For instance, an increase in income of a person will
definitely increase the mean income, but it may increase inequality as well. The
welfare function will give an unambiguous verdict about the direction of welfare
change by taking into account the underlying trade-off explicitly. Such a welfare
function is referred to as an abbreviated- or reduced-form welfare function because
its arguments summarize the entire distribution in terms of the mean income and
inequality.
We now provide some examples of social welfare functions associated with
some relative AtkinsonKolmSen inequality indices.
For the Atkinson index, the

corresponding welfare function is EA x x 1 I A x ; the symmetric mean of
order < 1, defined explicitly as
EA x
8
>
>
>
>
<
n
X
1
xi
n
!1
,
< 1,
6 0,
i1
n
>
Y
1
>
>
>
: x i n
1:14
0;
i1
where x 2 Dn . For 1, EA x depends only on the efficiency component x, and

for 0, it becomes the geometric mean. The harmonic mean drops out as a
particular case of EA x if 1. On the other hand, as ! 1, EA x
approaches the Rawlsian (1971) maximin welfare function mini{xi}, which corresponds to the relative maximin index.
Another example
is the DonaldsonWeymark
S-Gini welfare function defined as

x x 1 I DW
x . More explicitly,
EDW
EDW
x
n h
i
1X
i
i

1
xi :
n i1
1:15
23
As ! 1, EDW
x approaches the symmetric linear social welfare function x. In
x approaches the Rawlsian maximin rule. With an

contrast, as ! 1, EDW
increase in the value of , more importance is assigned to weights attached to
lower incomes in the aggregation. For 2, EDW

x becomes the Gini welfare
n
P
1
function n2 2i 1xi : While this formula is based on x, the welfare-ranked
i1
permutation of x, the formula for the Gini welfare function based on x^ , the
n
P
illfare-ranked permutation of x, is given by n12 2n i 1xi :
i1
From (1.4), it appears that the Gini welfare function can also be expressed as
n P
n

P
1
min xi ; xj . For any two persons i and j, if we measure the welfare of the
n2
i1 j1
income distribution (xi, xj) by min(xi, xj), the Rawlsian maximin criterion, then the
Gini welfare function is simply the average of the maximin welfare functions of all
two-person income distributions of the type (xi, xj), where i, j 1, 2, . . . , n :
A third example is the Bonferroni welfare function EB x x1 I B x,
given formally by
EB x
n
i
1X
1X
x^ j :
n i1 i j1
1:16
Although EA and EDW

are invariant under replications of the population, EB is not so.
, and IB are AKS indices so that they have a natural

The three indices I A , I DW
upper bound of 1. However, some relative indices are not bounded above naturally
by 1. An example is the coefficient of variation. In such a case, for the welfare
function to be increasingly related to efficiency and equity, the abbreviated welfare
function may be taken as xeIx . This welfare function may not fulfill the strong
Pareto principle, but it fulfills the scale improvement condition which requires
increasingness of welfare under equi-proportionate increase in all incomes
(Shorrocks 1983). It also meets Blackorby and Donaldsons (1984) minimal
increasingness postulate, which says that if all the individuals possess the same
income, more is preferred to less, a weaker condition than the strong Pareto
principle and the scale improvement condition.9
The values of their Gini welfare function for the distributions u1 0; 1; 8 and
1
2
8
u 0; 1; 5, considered earlier in the section, are 11
9 and 9, respectively. u has a
For other possibilities and related discussion, see, among others, Newbery (1970), Sheshinski
(1972), Kats (1972), Chipman (1974), Sen (1974), Kondor (1975), Graaff (1977), Kakwani (1985),
Lambert (1985, 2001), Ebert (1987, 1988a), Chakravarty (1988, 1990, 2009a, b), Shorrocks
(1988), Dagum (1990), Dutta and Esteban (1992), Ben -Porath and Gilboa (1994), Champernowne
and Cowell (1998), Amiel and Cowell (2003), and Blackorby et al. (2005).
24
higher welfare value than u2 since u1 GL u2 holds. But while in the earlier case,
when welfare was concerned only with efficiency, the dominance of u1 over u2 in
terms of welfare value was 1, now the dominance has reduced to 13 because of
welfares concern with equity as well.
In order to relate absolute indices to welfare functions, we assume that welfare
functions are regular and translatable. A social welfare function W : R ! R is

^ x ,
called translatable, if for all n 2 N, x 2 Rn , it can be written as W x W
^ is unit translatable and is increasing in its argument. Unit translatability
where W
^
^ x c1n W
^ x c, where c is a scalar such that x c1n
of W means that W
n
^ x W
^ 01n ; which is unit translatable by
2 R : Then, we have xe Ex W
^ . BlackorbyDonaldson (1980)Kolms (1976a) absolute
unit translatability of W
index can now be defined as
ABDK x x Ex;
1:17
where ABDK gives the per capita income that could be saved if the society
redistributed incomes equally with no loss of welfare. Unit translatability of x
ensures that ABDK is an absolute index. This continuous index fulfills PDT and takes
on the value zero if incomes are equally distributed. We can retrieve E(x) using
(1.17) and (1.12) and write it as Ex x ABDK x so that welfare is related to
efficiency and inequality increasingly and decreasingly, respectively.
If the inequality is measured by the Kolm (1976a)Pollak (1971) index,
n
1 1X
AKP
x x log
e xi ;
n i1
1:18
then, the abbreviated welfare function is given by

n
1 1X
x log
e xi :
EKP
n i1
1:19
The parameter 2 0; 1 attaches higher weight to progressive income transfers
x ! x mini fxi g; the absolute maximin

lower down the scale. As ! 1, AKP
index of inequality and the corresponding welfare function is the Rawlsian
maximin rule. Examples of other absolute indices are the S-Gini and Bonferroni
n
P
absolute indices given, respectively, by ADW

x x EDW
x x n1
h
i1
i
n
i
P
P
i i 1 xi and AB x x EB x x 1n 1i x^ j . For 2, ADW

is
i1 j1
the absolute Gini index of inequality
1
2n2
25
n P
n

P
xi xj . Some absolute indices
i1 j1
cannot be accommodated directly within the BlackorbyDonaldsonKolm framework (1.17). An example is the variance. For such indices, the welfare function may
be defined as xeAx . Given that A is an absolute index, this welfare function
satisfies the incremental improvement condition (Shorrocks 1983), which demands
increasingness of welfare under equal absolute increase in all incomes, a postulate
weaker than the strong Pareto principle.
The welfare functions corresponding to the S-Gini and Bonferroni indices are
distributionally homotheticthey are both homothetic and translatable. Consequently, they can be used to derive both relative and absolute indices. Formally, a
social welfare function W :
R !R1 is called distributionally homothetic if for all
e x , where is increasing in its argument and W
e
n 2 N, x 2 R n , W x W
e cx z1n cW
e x z, with c > 0 and z being scalars
satisfies the condition W
n
n
such that cx z1 2 R . The absolute indices whose associated welfare functions
satisfy distributional homotheticity are compromise absolute indices; they can be
converted into relative indices by dividing with the (positive) mean income of the
distribution.
Distributional homotheticity combined with population replication invariance
and a separability condition characterizes the single-parameter Gini welfare function (Bossert 1990). An interesting application of such a welfare function was
suggested by Chakravarty and Dutta (1987) for measuring distance between two
arbitrary income distributions that reflects the degree of welfare of one distribution
relative to the other. If welfare evaluation is done with respect to the S-Gini welfare
function, then for x 2 Rm and y 2 Rn , the ChakravartyDutta distance function is
given by

1X
m

X
1

1:20
x; y
i i 1 x i
i i 1 yi :
dCD

m i1
n i1
For 2, this distance function may be referred to as the Gini distance function.
For 1, it is simply the absolute value of the gap between the sizes (means) of the
distributions without showing any concern for their equity levels. As ! 1,
dCD
x; y ! jxm yn j, the maximin distance function. The distance function dCD
is always nonnegative; it takes on the value zero only when the well-beings of the
two populations are the same.10
10
See also Shorrocks (1982). For a recent discussion, see Yalonetzky (2012).
26
1.5
Subgroup Decomposable Indices of Inequality
The objective of this section is to analyze subgroup decomposable indices of

inequality. A subgroup decomposable inequality index is one, which, for a
partitioning of the population into subgroups by a characteristic such as by race,
religion, sex, ethnic groups, age, etc., can be broken down into the within-group and
between-group components. The within-group term aggregates inequalities within
different subgroups, and the between-group term is a consequence of variations in
mean incomes across these subgroups. Such decomposition becomes useful when a
policy maker becomes interested in determining the significance of income variations corresponding to characteristics like race, occupation, region, age, etc.
The subgroup decomposability postulate can be formally stated as:
Subgroup Decomposability (SUD) For all k 2 and for all x1 , x2 , . . . , xk 2 D,
I x
k
X

i n, I xi I 1 1n1 , 2 1n2 , . . . , k 1nk ;
1:21
i1
k
P
ni ; i
where ni is the population size associated with the distribution xi; n
i1

xi mean of the distribution xi; 1 ; 2 ; . . . ; k ; n n1 , n2 , . . . , nk); i n,
is the positive weight
to inequality in xi, assumed to depend on the vectors
1 attached
n and ; and x x ; x2 ; . . . ; xk . The population here has been partitioned into k
2 subgroups with respect to some characteristic, where k is arbitrary.
SUD provides a breakdown of overall inequality into within-group and betweengroup terms. The between-group component I 1 1n1 , 2 1n2 , . . . , k 1nk is the level of
inequality that would arise if each individual in a subgroup enjoys the mean income
k

P
of the subgroup, and the within-group component
wi , n I xi is the weighted
i1
sum of inequalities within different subgroups. In the literature, subgroup decomposable indices are also referred to as additively decomposable or, simply, additive
indices (see Foster 1983, 1985 and Chakravarty 1990, 2009a). Since the concept of
inequality is vacuous for n 1, for SUD to be well defined, we need k 2 and
ni 2 for all 1 i k. Consequently, we must have n 4.
Suppose an income distribution is characterized by a low within-group inequality and a high between-group inequality. Since the between-group inequality
measures the inequality in the mean incomes of the subgroups, this indicates that
a policy makers recommendation should concentrate on enhancement of the
welfare of the subgroups with low mean incomes. This can be done by proportionate increases of the incomes of individuals in such particular subgroups. Alternatively, increasing the incomes of the poor persons in the subgroups may be a
targeted policy. In contrast, with a high within-group inequality and a low
between-group inequality, the policy planners priority should be a recommendation for improvement of well-being of subgroups characterized by high inequality.
1.5 Subgroup Decomposable Indices of Inequality
27
Shorrocks (1980, 1984) demonstrated that the only family of relative subgroup
decomposable indices is the generalized entropy family defined as
8
n h
c
i
X
1
xi
>
>
1 ,
>
>
>
ncc 1 i1
>
>

>
n
< 1X
c
log
,
I S x
n
x
>
i
>
i1
> n h

i
>
>
1X x i
xi
>
>
>
log
,
:n
i1
c 6 0, 1,
c 0,
1:22
c 1:
n 2 N, x 2 Dn are arbitrary. The parameter c is an indicator of different perceptions

of inequality. The particular cases of IcS (x) corresponding to c 0 and c 1 are,
respectively, the Theil (1972) mean logarithmic deviation index and the Theil
(1967) entropy index of inequality. For c 2; the resulting index is simply half
the squared coefficient of variation. IcS satisfies POP, NOM, SYM, and PDT for all
values of c. For any c 2 1, 1 [ 1; 2, IcS attaches more weight to transfers lower
down the distribution. (See also Cowell (1980) and Cowell and Kuga (1981) for a
related discussion.)
The weight associated with the inequality of subgroup i in the decomposition is

c
given by wi , n nni i . Thus, only for the two Theil indices, sum of the weights
attached to inequality levels of different subgroups becomes 1. Bourguignon (1979)
characterized the Theil mean logarithmic deviation using population shares
of
different subgroups as the weights in the within-group component, that is, wi , n
nni : On the other hand, a characterization of the entropy index was developed by

Foster (1983) using wi , n nni i . The Atkinson (1970) index becomes monotonically increasingly related to IcS through the transformation
(
I A x

1
1 cc 1I Sc x 1 c ,
c
1 eIS x ,
c < 1, c 6 0,
c 0:
1:23
If we allow partitions of the population to be nonoverlapping, for instance, if we

partition the ordered distribution x^ 2 Dn into ordered distributions x^ 1 ,
x^ 1 , x^ 2 , . . . , x^ k , where the maximum income in subgroup j is less than the minimum
income in subgroup j 1, 1 j k 1, then the Gini index can be neatly
subdivided into between- and within-group components (Bhattracharya and
Mahalanobis 1967; Mookherjee and Shorrocks 1982). The Gini index for such a
partitioned distribution can be written as
I G x^

k
2
k X
k
X
i 1X
ni
x^ i
ni nj x^ i x^ i
^
IG x
:
x^
2 i1 j1 n n
x^
n
i1
1:24
28
Thus, the coefficients in the between-group decomposition depend on the

population proportions of the subgroups. Ebert (1988b) demonstrated that if in
(1.21) we restrict attention to nonoverlapping partitions, then the generalized
entropy class and the Gini index turn out to be the only relative inequality indices
that fulfill SUD.
Since the Bonferroni index does not meet DDP, which is used by both Shoorocks
(1980) and Ebert (1988b) in their characterizations, it is not included as a member
of the generalized entropy family and the Ebert family. Barcena-Martin and Silber
(2013) decomposed the Bonferroni index for a population partitioned according to
income classes and showed that in addition to between- and within-group components, a residual term, reflecting the effect of ranking and the indexs failure to
fulfill DPP, appears in the decomposition.
Subgroup or additive decomposability is extremely helpful for analyzing
inequality by population subgroups. However, replacement of the subgroup arithmetic mean by the subgroup symmetric mean of any arbitrary order q in the
calculation of the within-group component is a natural generalization of this
postulate. According to Foster and Shneyerov (1999), an inequality index I : D
! R1 satisfies generalized additive
if there is a q 2 R1 and a
decomposability
q 1
q k
such that for any
sequence of positive weights wi EA x , . . . , EA x , n
1 2
k
x , x , . . . , x 2 D ,
I x
k
X

wi EAq x1 , . . . :EAq xk , n I xi I EAq x1 1n1 , . . . , EAq xk 1nk ;
i1
1:25

where x x1 ; x2 ; . . . ; xk . Here, EqA (xi) is the symmetric mean of order q of the
income distribution of subgroup i. Thus, aggregate inequality is the weighted sum
of subgroup inequality levels plus the inequality in a smoothed distribution where
each subgroups income distribution is replaced by the one in which everybody in
the subgroup enjoys the symmetric mean of order q of the subgroup income
distribution. The weights are dependent on the population sizes of the subgroups
and their symmetric means of order q.
As Foster and Shneyerov (1999) demonstrated, the only relative inequality index
that satisfies the generalized decomposability property (1.25) is a positive multiple
of the following two-parameter family of inequality indices:
8

n c
1 X
E A x c
>
>
>
1 ,
c 6 0, q 6 c,
>
>
cc q i1 EAq x
>
>
>

>
n
>
1X
E q x
>
>
<
log A0
,
c 0, q 6 c,
q i1
EA x
1:26
I cq
FS x

q
n
>
>
xi
xi
> 1X
>
log q
, c 6 0, q c,
>
>
>
nq i1 EAq x
EA x
>
>
>
>
>
: 1I VL x,
c 0, q c:
2
29
E0A (x) is the geometric mean of the distribution x and
2
n
P
log E0xix
is the variance of the logarithm of incomes. The
I VL x 1n
where
i1
population replication invariant family Icq

FS coincides with the generalized entropy
family when q 1. It satisfies PDT if and only if either (i) q 1 and c 1 or (ii)
1
q 1 and c 1. Therefore, the particular case I VL x, which arises when
2
q c 0, is a violator of PDT (see also Foster and Ok 1999).
The weight assigned to the inequality of subgroup i is given by

c
q 1
ni EAq xi
q k
wi EA x , . . . , EA x , n
;
n EAq x
1:27
which depends on the subgroup population proportion only when c 0. This

weighting scheme bears similarity with that of the generalized entropy family,
but the symmetric mean of order q is used in place of the arithmetic mean.
The class of absolute indices that satisfies SUD is given by
A x
n h
i
1X
exi 1 , 6 0;
n i1
AV x
n
1X
x2 2 :
n i1 i
1:28
The index AV in (1.28) is the variance, which assigns equal weight to a transfer from
a rich to a poor at all income positions. In contrast, for any < 0, A favors transfers
at the lower end of the distribution. The opposite happens for > 0. For 0, A
becomes zero for all distributions. The Kolm (1976a)Pollak (1971) index of
x is increasingly related to A via the transformation

absolute inequality AKP
log1 A x AKP x, where

> 0
(see Chakravarty and
Tyagarupananda 2009; Bosmans and Cowell 2010).
The weights assigned to the inequality of the ith subgroup in the decomposition

n
ie i
i
of A and AV are given by i n, nne
and i n, n , respectively. Clearly, in
this case, the sum of the weights across subgroups becomes 1 only for the variance.
As Zheng (2007a) demonstrated, the family of unit-consistent inequality indices
that fulfills SUD is a positive multiple of
8
n
X
1
>
>
xi c xc ,
>
>
>
nc
c

1
>
>
i1
>
n
<
1 X
x
c
log
,
I Z x
nx i1
xi
>
>
>

>
n
>
X
1
xi
xi
>
>
>
log
,
:
x
nx1 i1 x
c 6 0, 1,
c 0,
c 1;
1:29
30
where c and are constants. This family coincides with the generalized entropy
family if c. From Zhengs characterization, it emerges that the variance is an
absolute unit-consistent decomposable inequality index, since for c 2 and 0;
c
I c
Z becomes the half of the variance. For c 2 and 1; I Z is half the Krtscha
index, an intermediate index. The member of Zheng family I c
Z corresponding to
c 2 and 0 < < 2 is also an intermediate index satisfying unit consistency, and it
can be regarded as a generalized Krtscha index.
The absolute Gini index also satisfies a weak decomposability postulate,
suggested by Ebert (2010). This postulate demands that for a 2-subgroup
partitioning (x1, x2) of the income distribution x,the
can be
overall inequality

decomposed into a within-group component w1 n Ax1 w2 n Ax2 and a

n1 P
n2
P
between-group component u n
A x1i ; x2j , where the weight functions
i1 j1

w1 n , w2 n , and u n are positive. Formally,
Weak Decomposability (WDE) For every n n1 ; n2 , where n1 1 and n2 1

are integers, there exist positive weight functions w1 n , w2 n , and u n such that
n1 X
n2

X

A x1i ; x2j ;
A x1 ; x2 w1 n A x1 w2 n A x2 u n
1:30
i1 j1
where xj is the income distribution over the population with size nj, j 1, 2. While
the first two terms on the right-hand side correspond to the usual within-group
component used in the literature, the third term depends on inequality between all
pairs of individuals, one belonging to subgroup 1 and the other belonging to
subgroup 2. Ebert (2010) developed a characterization of the following family of
weakly decomposable population replication invariant absolute inequality indices
AE x
n X
n

2X
A xi ; xj :
n2 i1 j1
1:31
r

In the above form, if we choose A xi ; xj xi xj , where r 1, then
n P
n

P
xi xj r , which are positive multiples of the absolute Gini
AE x n22
i1 j1
index and the variance for r 1 and r 2, respectively. Thus, the absolute Gini
index, although not subgroup decomposable, is weakly decomposable.
Blackorby et al. (1981) suggested an ethical approach to the inequality decomposition by population subgroups. Following Blewett (1982) and Blackorby
et al. (1999), we define the AtkinsonKolmSen between-group inequality index
as the proportion of the total income that could be saved by moving from the
subgroup-wise ethically ideal distribution x1e 1n1 , . . . , . . . , xek 1nk in which everybody in a subgroup enjoys his/her subgroup ede income to the population-wise
31
ethically ideal distribution (xe1n) in which everybody enjoys the society ede
Pk
ni x i nxe
BI
income. Formally, AAKS i1nxe
, where xie is the ede income of the ith
subgroup whose income distribution is xi; ni and i are, respectively, its population
Pk
size and mean income, and
i1 ni n. The within-group AtkinsonKolmSen
index is the proportional
saving
of

the total income in the movement from the actual
distribution x x1 ; x2 ; . . . ; xk to the subgroup-wise ideal distribution. Formally,
Pk
P
P k nj j
nx
n xi
n j j
k
xej
j
i1 i e
I WI
1

AKS
j1 n
j1 n I AKS x : Thus, the withinnx
j
group inequality is simply the income share weighted average of subgroup indices.
The AtkinsonKolm Sen index I AKS x 1 xxe for the population as a whole is
the sum of these two subindices. (See also Ebert 1999.) Ethical decomposition of
inequality is meaningful only when the underlying
income
ede
function is addin
P
1 1
gxi , where g : D1 ! R1
tively separable, that is, it is of the form xe g
n
i1
is continuous, increasing, and strictly concave. This is required by the condition that
the ede income of any subgroup is independent of the incomes in other subgroups.
The only relative index for which the ede income verifies additive separability is the
Atkinson index.
For the absolute index, Blackorby et al. (1981) defined the within-group inequality index as the per capita income saved in moving from the actual distribution x

x1 ; . . . ; xk 2 Rn to the subgroup-wise ethically ideal distribution. Formally,
k
k

P
P
ni xei
ni
i
AWI
BDK x
n
n i xe ; the population share weighted average of
i1
i1
BlackorbyDonaldsonKolm subgroup inequality indices. The saving in per capita

income for movement from subgroup-wise ethically ideal distribution to the
population-wise ethically ideal distribution is the between-group inequality index,
k
P
ni xei
BI
that is, ABDK
n xe : The overall per capita index ABDK is the sum of the

i1
BI
subindices AWI
BDK and ABDK . The only absolute index for which this aggregation
holds is the KolmPollak index.
Chapter 2
On the Measurement of Income

Bipolarization
2.1
Introduction
Polarization has recently attracted a great deal of attention in economics and some
other branches of social sciences. In the context of income distribution, polarization
is concerned with the division of the society into subgroups as a possible explanatory factor for social conflicts. In order to monitor the degree of polarization, it is
necessary to inquire whether the poor are becoming poorer and the rich are
becoming richer along with the investigation how distant the subgroups are from
one another.
Taking as a starting point that polarization is a matter of subgroups, two views
are distinguished in the literature. While the first opinion is concerned with the
shrinking middle class, where the middle class is defined in terms of concentration of population with incomes in some range around the median income (Wolfson
(1994, 1997) and Foster and Wolfson (2010)), the alternative notion deals with the
clustering of population into different income subgroups (Esteban and Ray (1994)
and Duclos et al. (2004)). In this chapter, we will analyze the median-based
approach to the measurement of polarization. Since it subdivides the population
into two subgroups, with incomes below and above the median, respectively, we
refer to this as the case of bipolarization. The notion of bipolarization corresponds
to the situation where the society has a significant proportion of poor persons and
there is also a non-negligible proportion of persons with very high income. In
contrast, the size of the middle class is rather low. The other approach that considers
bunching of the population into any number of income subgroups will be discussed
in the next chapter. We refer to the underlying concept as income multipolar
polarization. Esteban and Ray (2012) use the term polarization proper for this
general notion.
An indicator of bipolarization looks at the spread and dispersion of the distribution from the middle position. There are several important reasons for looking at
polarization from this perspective. A large and rich middle class contributes
Springer India 2015
33
34
On the Measurement of Income Bipolarization
significantly to the well-being of a society in many ways, particularly, in terms of

high economic growth, better health status, higher level of education, a sizeable
contribution to the countrys tax revenue and a better infrastructure, and more social
cohesion resulting from fellow feeling. High incomes resulting from high economic
growth provide higher chances for better institutions and state capacity, which in
turn may reduce chances for conflict (McBride et al. 2011). For instance, using
rainfall data as a vehicle for higher economic growth, Miguel et al. (2004) found
evidence for higher incomes reducing conflict, at least for Africa.
A person with a low income may regard the median as a reference income and be
optimistic to achieve this income by working hard. Given that gradually such a
person will be in the middle spectrum of the income distribution, he/she will be in
the main engines for economic growth and is not likely to feel disgusted about the
society because of his/her economic and social position. This in turn demonstrates
that for a society with rich middle class, chances of political and social instability
are rare. Thus, social tensions between the rich and the poor can be moderated by a
visible middle class. This also ensures possibilities for more social modernization
and more democracy. The middle class provides a bridge, using which the poor can
become rich. In the words of Aristotle (350) . . .the best political economy is
formed by citizens of the middle class, and that those states are likely to be welladministered, in which the middle class is large. On the other hand, a society
characterized with a small middle class and more persons away from the median
may lead to a strained relationship between the subgroups on the two sides of the
median which ultimately can generate social unrest and civil war. A small and weak
middle class implies weak institutions and hence non-sustainable growth (Birdsall
2007).
The study of bipolarization also becomes relevant if we accept the view that the
overall income distribution is a mixture of several distributions. Attempts have been
made to study the world income distribution as a mixture distribution (see, e.g.,
Paap and van Dijk 1998). Quah (1996) observed an emerging twin peaks shape in
the income distribution across countries. That is, the income distribution was
polarizing into twin peaks, the very rich and the very poor, while the middle class
of countries was disappearing. In the literature, this is referred to as two-peak or
two-component hypothesis. It provides a description of the divergence between
the incomes of the rich and the poor. Pittau et al. (2010) used the Penn World Data
for the period 19602000 to observe evidence of three-component densities: low,
average, and high mean income groups, with increased discrepancy between group
means over time. According to Grier and Mynard (2010), as early as the 1950s and
1960s, the per capita income distribution is most likely composed of two
components. . .. By 1980, the distribution changes into one that is most likely
composed of three components, with a large gap opening between the means of
the poorest and the richest components. . ..
In the next section of this chapter, we make a brief discussion on the size of the
middle class. Section 2.3 analyzes the axioms for an index of bipolarization. The
relative bipolarization ordering and several relative indices are discussed in
Sect. 2.4. A relative index remains unaltered under scale transformations of
2.2 Measuring the Middle Class
35
incomes. Indices satisfying alternative notions of invariance are presented in

Sect. 2.5. Finally, an analysis of the measurement of bipolarization using social
welfare functions is introduced in Sect. 2.6.
2.2
Measuring the Middle Class
As we have pointed out, the size and composition of the middle class of a society is
highly important for several reasons (see, e.g., Lawrence 1984; Thurow 1984;
Davis and Huston 1992; Landes 1998; Easterly 2001; Pressman 2001). The first
studies to analyze the decline of the middle class were made in 1980s. As in the
literature on inequality and poverty measurement, much of the attempts were
devoted to the construction of indices. As Duclos and Taptue (2014) pointed out,
four distinct but unrelated steps are involved in an exercise of this type. The first
step is to specify the space for disuniting the distribution across the middle position.
Two alternative notions of space are income and population spaces. According to
the income space, the middle class is defined with respect to a particular income
range, and all individuals whose incomes belong to this income range are treated as
middle-class persons. On the other hand, the population space looks at the income
range, covered by a given percentage of the population. The second step is to define
the middle class analytically. The third step involves specification of the boundaries
for the middle class. Finally, the fourth step requires aggregation of the characteristics of the middle class.
The indices to be discussed in this section rely on the median. For any
n 2 N, x 2 Dn , the median of x is denoted by m(x) (or, simply m). If n is odd,
th
th
m(x) is the n1
observation in x^ . But if n is even, the arithmetic mean of the n2
2

th
and the n2 1 observations in x^ is taken as the median.
Income space-based indices under alternative specifications of the income range
were suggested by several authors, including Thurow (1984), Blackburn and Bloom
(1995), Horrigan and Haugen (1988), Leckie (1988), and Beach et al. (1997). In
Thurow (1984), the middle class has been defined as the proportion of the population with incomes between 75 and 125 % of the median, whereas Blackburn and
Bloom (1995) extended the middle income range from 60 to 225 % of the median.
Leckie (1988) preferred to use the range 85115 % of the median wage to identify
the middle class. All these indices can be expressed as the share of the population
proportion whose incomes are in an interval of the form pm s1 , qm s2 , where
0 < p < 1 and q > 1 are constants. Horrigan and Haugen (1988) considered several
values of the two bounds p and q to define the middle class.
In order to illustrate these indices graphically, following Duclos and Taptue
(2014), let us assume that income follows a continuous-type distribution and denote
the cumulative distribution function by F. Then, F(v) gives the cumulative proportion of the population with income less than or equal to v. For the income range (s1,
s2), the associated index is the vertical distance Fs2 Fs1 , denoted by V (see
36
Cumulative distribution function F (v )
s1
s2
Income v
Fig. 2.1 The middle class
Fig. 2.1). Decline of the middle class in this context refers to a situation with fewer
people having incomes in the given interval (s1, s2).
Population space-based indices were suggested, among others, by Levy (1987)
and Beach (1989). In Levy (1987), the magnitude of the middle class has been
defined as the middle three-fifth of the population, that is, the size of the population
lying between 20 and 80 % of the total population, and the corresponding index is
the proportion of the total income enjoyed by this part of the population. For any
income distribution function F, this index can be formally defined in terms of the
R0:8 1
F d:
Lorenz curve as 0:2

: For any n 2 N, x 2 Dn ; it is simply Lx; 0:8 Lx; 0:2.
(See Sect. 1.3 for definition of the Lorenz curve.)
The Levy index can also be represented using the quantile function
corresponding to the cumulative income distribution function F. Since
R1 1
F d, the area under the quantile curve, equals the mean income , this
0:
index is simply the ratio between the hatched area and the area under the curve in
Fig. 2.2.
To illustrate the Levy index, let x 1; 3; 5; 7; 9 and y 2; 4; 6; 8; 10 be two
distributions of income. They are symmetric around their median (also mean)
incomes, 5 and 6, respectively. (A symmetric distribution is one all of whose
equidistant points from the line of symmetry (say, mean or median) have identical
frequency.) Note that y Lorenz dominates x so that y is less unequal than x by all
relative inequality indices that are strictly S-convex. The Levy index for x is
1
Lx; 0:8 Lx; 0; 2 16
25 25 0:6. By a similar calculation, the Levy index for
2.3 Axioms for an Index of Income Bipolarization
37
Qunatile function
1.0
0.2
0.5
0.8
1.0
Fig. 2.2 The Levy index
y also becomes 0.6. In fact, for any symmetric distribution x, we have Lx, 1 t
Lx; t 1 2t for all t. If t 0:2, the Levy index turns out to be 0.6, which is its
value for all symmetric distributions. This is true irrespective of how unequal the
income distribution is. This is an undesirable feature of the Levy index.
Beach (1989) used the middle 60 % of the population as the width of the middle
class. (See also Beach et al. (1997) for related indices.) Shrinkage of the middle
class here will mean loss of economic power of the population in the concerned
interval.
2.3
Axioms for an Index of Income Bipolarization
Bipolarization is concerned with the division of a population into two subgroups

using the median. By a bipolarization index, we mean a function P : D ! R1 . For
any n 2 N, x 2 Dn , P(x) gives the extent of polarization associated with x. Let
n n1
2 . We write x and x for the subvectors of x that include xi for xi < mx and
xi > mx, respectively. Thus, for any n 2 N, x 2 Dn , x^ ^x ; ^x if n is even and
x^ x^ , mx, x^ if n is odd.
As Wolfson (1994, 1997) and Foster and Wolfson (2010) pointed out, two
postulates that are regarded as innate to the concept of bipolarization are increased
spread and increased bipolarity. Increased spread demands that as the income
moves away from the middle position to the tails of the income distribution, the
distribution becomes more polarized. In other words, polarization is increasingly
38
related to the spread of the distribution from the middle position. Increased bipolarity says that polarization goes up under a bunching of incomes below or above
the median. That is, polarization has an increasing relationship with reduction of
gaps between any two incomes, above or below the median. To understand
these more explicitly, let x 1; 2; 3; 4; 5 so that mx 3. Now, the distribution
u 0:5; 2; 3; 4; 5:1 is obtained from x by reducing the minimum income by 0.5
and increasing the maximum income by 0.1. Each of these two changes increases
the spread of the distribution from the median; hence, by the rule of increased
spread, u is more polarized than x. On the other hand, v 1:3; 1:7; 3; 4:2; 4:8 is
generated from x by a progressive transfer of 0.3 units of income from 2 to 1 and a
progressive transfer of size 0.2 from 5 to 4. Thus, x is transformed into v by
clustering of incomes on both sides of the median. The increased bipolarity
condition requires that v is more polarized than x. Note that both these postulates
are specified under the assumption that the median remains unchanged. The two
examples clearly indicate that bipolarization incorporates an inequality-like factor,
the increased spread method, under which both inequality and polarization increase,
and an equality-like factor, the increased bipolarity rule, which augments bipolarization, but reduces inequality. This demonstrates that although polarization and
inequality involve a common component, the increased spared criterion, they
represent two different aspects of distributional issues. A general observation that
we can make here is that under a regressive income transfer from someone with
income below the median to someone with income above the median, there is an
increased spread away from the median. The two subgroups, the first, with individuals having incomes below the median and, the second, consisting of persons
with incomes above the median, move apart, and heterogeneity between them
increases.
A bipolarization index should satisfy some other postulates along with increased
spread and increased bipolarity. Below we formally present the postulates for an
index of bipolarization rigorously.
Increased Spread (IS) For all n 2 N, if x, y 2 Dn , where mx my, are related
through anyone of the following cases, (i) x^ y^ , y^ x^ , (ii) x^ y^ ,
x^ y^ , and (iii) y^ x^ , x^ y^ , then I n x > I n y:
Increased Bipolarity (IB) For all n 2 N, if x, y 2 Dn , where mx my, are
related through anyone of the following cases, (i) x^ T y^ , x^ y^ , (ii) x^ T y^ ,
x^ y^ , and (iii) x^ T y^ , x^ T y^ , then I n x > I n y:
Symmetry (SM) The polarization index P : D ! R1 is symmetric.
Principle of Population (PP) The polarization index P : D ! R1 is population
replication invariant.
Continuity (CN) For all n 2 N, P is a continuous function on Dn.
For mx my, x^ y^ , y^ x^ means that the income distributions above
the median are identical in x and y, and the income distribution below the median in
2.3 Axioms for an Index of Income Bipolarization
xi t
x i
39
x j
x j + s
Income
Fig. 2.3 Increased spread
Income
Fig. 2.4 Increased bipolarity
x is obtained from the corresponding part of y by reducing at least one income

so that there is at least one median-preserving spread below the median. As a
result, some of the poor are becoming poorer. Thus, y^ x^ is same as the
condition that y Pareto rank dominates x . In other words, y^ first-order stochastic
dominates x^ . Likewise, x^ y^ , x^ y^ means equality of income distributions below the median and Pareto rank dominance of x over y so that some of
the rich are becoming richer, which means that x^ first-order stochastic dominates
y^ (see also Duclos and Echevin 2005). Equivalently, there is at least one medianpreserving spread above the median. In (iii), there are median-preserving spreads
both below and above the median. For each type of spread, we have I n x > I n y:
In Fig. 2.3, the income x^ i below the median reduces to x^ i t and the income x^ j
above the median increases to x^ j s, where the positive incomes t and s are such
that the rank orders of incomes remain preserved. Therefore, we have medianpreserving spreads both below and above the median. Consequently, polarization
goes up.
In IB, x^ T y^ , x^ y^ says that x^ is obtained from y^ by at least one
bipolarity increasing progressive transfer below the median, whereas incomes
above the median are the same in x and y. For x^ y^ , x^ y^ and x Ty
mean Pareto rank dominance of x over y combined with at least one bipolarity
increasing transfer above the median. This is equivalent to second-order stochastic
dominance of x^ over y^ .
Figure 2.4 illustrates one simple case of increased bipolarity. The consecutive
incomes u and v below the median, where v > u, cluster at their mean income by a
progressive transfer of the size vu
2 from v to u. This clustering makes the income
distribution in the subgroup of persons with incomes below the median more
homogeneous, and as a result, polarization increases.
The postulates SM, PP, and CN are polarization counterparts to the respective
inequality postulates and need no further explanation.
Often we assume that a bipolarization index is normalized. That is, it takes on the
value zero when incomes are distributed equally. Formally,
40
Normalization (NM) For all n 2 N, x 2 Dn ; Px 0 if x c1n , where c > 0 is

any scalar.
We conclude this section by demonstrating that the Levy index is a transgressor
of
IS
and
IB.
For
this,
consider
the
income
distribution
y 1; 3; 5; 7; 9; 11; 13; 15; 17; 19, whose median (also mean) is 10. The distribution y1 0:5; 3; 5; 7; 9; 11; 13; 15; 17; 19:5 is obtained from y by increasing the
maximum income by 0.5 and reducing the minimum income by 0.5. Thus, y is
transformed into y1 by increasing spread both below and above the median. But the
value of the Levy index for both y and y1 is 0.6. This is a clear violation of IS by this
index. Next, y is transformed into y2 2; 2; 5; 7; 9; 11; 13; 15; 18; 18 by a bipolarity increasing progressive transfer on each side of the median. But the Levy index
for both y and y2 is 0.6, which is an indication of its transgression of IB.
2.4
A Bipolarization Ordering and Some Relative

Bipolarization Indices
Since the Lorenz ordering is inappropriate for ranking income distributions using
bipolarization indices, following Wang and Tsui (2000), we consider a bipolarization ordering, an equivalent formulation of the FosterWolfson bipolarization
ordering. Given that the median is the point of demarcation in bipolarization
measurement, the ordering relies on deviations of incomes from the median.
For any x, y 2 Dn , mx my m, x is more polarized than y in the Foster
Wolfson sense, what we write xFW y if and only if
X
X
jx^ i mj
jy^ i mj, 1 j < n,
and
ji<n
ji<n
nij
jx^ i mj
2:1
jy^ i mj,
n < j n;
nij
with at least one of the inequalities above being strict, where n n1

2 .
For any x 2 Dn , the cumulative income shortfall (excess) of the bottom nj
population proportion, j < n (j n) from (over) the corresponding total in the
imaginary distribution in which
! all the individuals enjoy the median income is
P
P
jx^ i mj
jx^ i mj . For any two distributions x and y over a given
ji<n
nij
population size, if the median income is the same and these cumulative shortfalls for
the latter are not higher than those for the former, and strictly less in at least one case,
2.4 A Bipolarization Ordering and Some Relative Bipolarization Indices
41
then the former is regarded as more polarized than the latter in the FosterWolfson
sense. In order to illustrate this, let the income distribution be x 3; 9; 6; 4; 1.
Then,
x^ 1; 3; 4; 6; 9, the median income mx 4 and n 3. The sums
P
jx^ i mj, 1 j < n ; are given, respectively, 4 and 1. On the other hand,
ji<n
P
the sums
jx^ i mj, n < j n; are given by 2 and 7, respectively. Next, for the
nij
distribution,
y 5; 10; 3; 2; 4, y^ 2; 3; 4; 5; 10, my 4, and n 3. The sums
P
jy^ i mj, 1 j < n; become 3 and 1, respectively, while the sums
ji<n
P
jy^ i mj, n < j n; turn out to be 1 and 7, respectively. x^ is obtained
nij
from y^ by reducing the income of the poorest person by 1 unit, whereas x^ is

obtained from y^ by a progressive transfer of 1 unit of income from the richest
person to the second richest person. Evidently, xFW y holds.
Violation of one or more of the inequalities in (2.1) indicates that we cannot rank
the underlying distributions by the relation FW . That is, FW is an incomplete
relation, although it is transitive. Hence, it is a quasi-ordering.
Given mx my m and y^ 6 x^ , the second inequality in (2.1) says that
^
x GL y , that is, x^ second-order
stochastic
P
P dominates y . By Theorem 1.2, this
is alike the condition that
x^ i >
y^ i for any increasing and strictly
nij
nij
concave function . For the first inequality in (2.1), define ui x^ ni , vi y^ ni
if n is odd. When n is even, define ui x^ n0:5i P
and vi P
xn0:5i . Then, for
y^ 6 x^ , the first inequality in (2.1) means that
ui
vi , 1 j < n,
1ij
1ij
with at least one >. This is same as the condition that x^ is submajorized by y^ , that
is, x^ y^ A1 for some
matrix A1 of appropriate order. This is identical
PbistochasticP
to the statement that
x^ i <
y^ i for any increasing and strictly convex
ji<n
ji<n
function (Marshall et al. 2011, pp. 1213). In view of Theorem 1.2 and the above
discussion, we are now in a position to present the following variant of a theorem of
Wang and Tsui (2000).
Theorem 2.1 For any unequal x, y 2 Dn , mx my m, the following statements are equivalent:
(i) xFW y.
(ii) x^ can be obtained from y^ by a sequence of spread-increasing movements
away from the median and/or bipolarity increasing progressive transfers;
and/or x^ can be obtained from y^ by a sequence of spread-increasing
movements away from the median and/or bipolarity increasing progressive
transfers.
(iii) x^ y^ A1 and/or y^ A2 x^ , for some bistochastic matrices A1 and A2 of
appropriate orders.
42
(iv) There exist two sets {M11 , M12 ., . . ., M1k } and {M21 , M22 ., . . ., M2q } of finitely
many PigouDalton matrices of appropriate orders such that x^ y^ M11
M12 . . . M1k and/or x^ y^ M21 M22 . . . M2q .
(v) x generalized Lorenz dominates y and /or x generalized Lorenz
dominates y .
(vi) x second-order stochastic dominates y and /or x second-order
stochastic
dominates
y .
P
P
^
(vii)
x i <
y^ i for any increasing and strictly convex function
ji<n
ji<n
P
P
x^ i >
y^ i for any increasing and strictly concave
and/or
nij
nij
function .
(viii) x^ < y^ for any increasing, strictly S-convex function and/or
x^ > y^ for any increasing, strictly S-concave function .
Bossert and Schworm (2008) showed that all polarization orderings satisfying IS
and IB are supersets of an ordering that can be defined in terms of generalized
Lorenz dominance. More precisely, a polarization ordering yPO x satisfies IS
and IB if and only if O PO , where yO x means that y GL x and
y GL x . If the medians of the two distributions under comparison are the
same, then O , FW .
Since inequality and polarization are two different concepts, it is incorrect to use
inequality indices for measuring polarization. Foster and Wolfson (2010) suggested
an index that can be interpreted in terms of the relative bipolarization curve, a plot
of a normalized value of the distance between the income share enjoyed by any
population proportion and the corresponding share that it would enjoy if everybody
possesses the median income against the population proportion. Their index is
defined as
PFW x
22K I G x
;
m
2:2
x
where K x 2
and IG is the (relative) Gini index of the income distribution x
x
(See Wolfson (1994).) This index is a relative index, that is, PFW cx PFW x for
all c > 0. It is the area under the relative bipolarization curve and satisfies all the
postulates for a bipolarization index. It takes on the minimum value zero for a
perfectly equal distribution. On the other hand, the maximum value 1 is achieved
when the distribution is perfectly bimodal, that is, when half the population has zero
income and the remaining half has income equal to 2.
Now, consider a partitioning of the population into two subgroups, S1 and S2,
where all the persons in S1 have incomes below median income and all the persons
in S2 have incomes above the median income. Then, the Gini index IG can be
written as the sum of the between-group factor IBI
G , which has no within-group
inequality, and within-group factor IWI
,
a
population
share weighted sum of Gini
G
43
indices within the two subgroups (see Chap. 1). IBI

G is the Gini index of the income
distribution where each person in S1 enjoys its mean income and each person in S2
possesses its mean income. The difference IG I GBI is an indicator of within-group
inequality, as measured by the Gini index. Then, PFW can be rewritten as

2 I GBI x I WI
G x
PFW x
:
m
2:3
The decomposition enables us to understand how PFW is influenced by the axioms

IS and IB. An increase in IBI
G increases both inequality and polarization, whereas an
WI
increase in IG increases inequality but reduces polarization.
As we have noted, the partial ordering FW is based on distances of incomes
from the median. This motivated Wang and Tsui (2000) to suggest the following
relative index of bipolarization:
PWT x

n
1X
x^ i m
;

n i1
m
2:4
where , defined on the set of nonnegative real numbers, is continuous. Wang

and Tsui (2000) demonstrated that PWT satisfies IS and IB if and only if is
increasing and strictly concave. However, one problem with PWT is the choice of
a particular .
Chakravarty (2009a) considered a normalized version of a variant of PWT
defined as
"
n
X
1 1n
#
jm x^ i j
i1
PC x
2:5
Given 0 0 and assuming that incomes are not unbounded, PC is a continuous

and bounded function with a lower bound of zero. However, PC need not be relative
index in general. The only member of PC that becomes a relative index is given by
1
n
P x
n
X
jm x^ i j
i1
!1
,
0 < < 1:
2:6
The inequality restriction 0 < < 1 is necessary and sufficient for IS and IB to be
satisfied. The index also satisfies all the other axioms. P is the symmetric mean of
order of deviations of incomes from the median as a fraction of the median itself.
We may refer to this as the relative mean deviation of order about the median. As
the value of reduces, a bipolarity augmenting progressive transfer on the either
side of the median increases P by a larger amount. For 1, P x becomes the
44
relative mean deviation about the median,

which satisfies
IS but is a violator of IB.1
P
P
For a symmetric distribution x,
m xi
xi m : Therefore, under
xi m
xi m
symmetry of the distribution, from policy point of view, for 1, mnP2 x gives
the total amount of money required to be transferred from all the persons above the
median income to all the persons below the median income so that polarization
becomes 0. This amount of money may, therefore, be termed as the cost of
polarization for a symmetric distribution. Alternatively, this is the total amount of
money required to put all the poor persons (with incomes below the median
income) at the median income itself, when the income distribution is symmetric.
Wang and Tsui (2000) also suggested a generalized Gini-type relative polarization index defined as
PWTG x
n
1X
bi x^ i ;
m i1
2:7
where {bi} is a sequence of real numbers. With a given rank order of incomes, PWTG
is linear in incomes. As Wang and Tsui (2000) have demonstrated, PWT satisfies IS
if and only if bi < 0 (bi > 0) for all x^ i < m (^
x i m). Wang and Tsui (2000) also
showed that PWT satisfies IB if and only if the weights are decreasingly ordered.
Finally, if normalization is included in the list of axioms, then PWTG becomes the
n
P
normalized version PNWTG x m1 bi jm x^ i j of PWTG. However, PWTG is not
i1
suitable for cross-population comparisons of bipolarization; it is a violator of PP.

Bossert and Schworm (2008) employed an independence property along with IS
and IB to characterize a family P of bipolarization indices. According to the
independence property, x is strictly separable from x in P and conversely, x is
strictly separable from x in P (see Blackorby et al. 1978, Chapter 3). This property
enables us to obtain polarization as a function that can be expressed using two
functions, one applied to x and the other applied to x . Formally,
Independence (IND) For any x, y 2 Dn ,

P y ; y P x ; y , Py ; x Px ; x ;

P y ; y Py ; x , P x ; y Px ; x :
2:8
Bossert and Schworm (2008) showed that a polarization index satisfies IS, IB,
and IND if and only if there exists an increasing function : R2 ! R1 ,
If in P x, given by (2.6), we replace the median income m by the mean income , then the
resulting inequality index reduces under a progressive income transfer for all > 1. This
inequality index becomes the coefficient of variation and the relative mean deviation about the
mean for 2 and 1, respectively. As ! 1, it approaches the relative maximin index.
45
a decreasing strictly S-concave real-valued function 1 defined on the income

distributions of the persons with incomes below the median and an increasing
strictly S-concave function real-valued function 2 defined on the income distributions of the persons with incomes above the median such that

P y ; y 1 y , 2 y :
2:9
The function 2 can be regarded as an inequality-sensitive welfare indicator for the

high-income group and 1 may be treated as a reverse welfare indicator for the
low-income group, where the median is the line of demarcation that separates the
high-income group from the low-income group. The representation (2.9) enables us
to calculate polarization by aggregating functions of the income distributions of two
subgroups of population partitioned using the median. An example of the general
BossertSchworm index of the type (2.9) is the index given by (2.4), which in
addition to IS, IB, and IND fulfills symmetry and scale invariance.
The indices of polarization, we have analyzed above, are of relative variety.
However, ranking of different income distributions by alternative indices need not
coincide. A unanimous ordering of different income distributions by all relative
polarization indices can be obtained by comparing the relative bipolarization curves
of the distributions. The
P construction of the curve relies on the relation FW . For
any x 2 Dn , the sum
m x^ i when normalized by nm, the total income in the
ji<n
hypothetical distribution where each individual enjoys the median income

P
1
m becomes RB x; nj nm
m x^ i . This quantity is the ordinate of the relative
ji<n
bipolarization curve of x at the population proportion nj , where 1 j < n and

n n1
incomes are not below the median, the corresponding
2 . Likewise, when
P
1
ordinate is given by nm
x^ i m. When n is odd, the RBC of x, RB(x, t), where
nij
t 2 0; 1, is defined as

j
j
j1
1 RB x;
RB x;
;
RB x;
n
n
n
2:10
0 1 and 1 j n 1, where RBx; 0 1.

If n is even, the curve is defined as

j
j
j1
RB x;
1 RB x;
RB x;
;
n
n
n
for all 0 1 and 1 j n 1, j 6 n, where RBx; 0 1,
2:11
46
and

n 0:5
n 0:5
n 0:5
RB x;
1 RB x;
RB x;
;
n
n
n
2:12
for all 0 1.
Note that when n is even, xn is not an income in the distribution x. However, the
ordinate corresponding to the population proportion nn is well defined (see
Chakravarty et al. 2007). The assumption RBx; 0 1 makes the graph closed.
For a perfectly equal income distribution, the RBC coincides with the horizontal
axis which represents the cumulative population proportions. In general, for an

unequal income distribution, the curve decreases gradually over the interval 0; nn .
At the population proportion nn, it coincides with the horizontal axis. The curve

increases monotonically over nn; 1 .
In order to illustrate the construction of RBC, let x 2; 5; 3; 4; 1. Then, m 3,
n 3; x^ 1; 2, and x^ 4; 5. The ordinates of the RBC of x for the population
3 1
1
3
proportions 5j , where j 1, 2, . . . , 5, are given, respectively, by 15
, 15, 0, 15
and 15
.
The next figure gives an example of a relative bipolarization curve (Fig. 2.5).
For x1 , x2 2 D, x1 is said to dominate x2 with respect to relative bipolarization
(what we write x1 RB x2 ) if

RB x1 ; t RB x2 ; t
2:13
for all t 2 0; 1, with strict inequality for some t. According to the relation x1 RB x2 ,
the relative bipolarization curve of x1 cannot be below that of x2 and must be above
at some places (at least).
The following theorem, whose proof can be found in Chakravarty (2009a),
shows that x1 RB x2 is the same as the stipulation that x1 has higher bipolarization
x2 for all relative bipolarization indices that verify IS, IB, SM, and PP. Formally,
Theorem 2.2 Let x1 2 Dk , x 2 Dl be arbitrary. Then, the following conditions are
equivalent:
(i) x1 RB x2 .
(ii) Px1 > Px2 for all relative bipolarization indices P : D ! R1 that satisfy IS,
IB, SM, and PP.
Theorem 2.2 shows that if the relative bipolarization domination relation
between two distributions holds, it is not necessary to calculate any relative
bipolarization index satisfying IS, IB, SM, and PP to judge whether one distribution
is characterized with higher polarization than the other. Since no assumption has
been made about the equality of the medians and the population sizes, this is the
most general result we can have for bipolarization ranking of distributions. However, if the two curves cross, such an unambiguous verdict about polarization
ordering of distributions cannot be made. Thus, RB is a quasi-relation.
47
Fig. 2.5 Relative

bipolarization curve
RB(x, t)
n
n
Cumulative population proportion
If the population size of the distributions under comparison is the same, then it is
obvious that PP can be dropped from the list of axioms required in Theorem 2.2.
Consequently, we have:
Theorem 2.3 Let x1 , x2 2 Dk be arbitrary. Then, the following conditions are
equivalent:
(i) x1 RB x2 .
(ii) Px1 > Px2 for all relative bipolarization indices P : Dk ! R1 that satisfy
IS, IB, and SM.
If the median income is fixed across the distributions, scale invariance property
of the indices is no longer an essential property. Given that the population size is
variable, in this particular case, the domain of definition of bipolarization indices is
Dm fx 2 Djmx mg. The following equivalence theorem can now be stated:
Theorem 2.4 Let x1 , x2 2 Dm be arbitrary. Then, the following conditions are
equivalent:
(i) x1 RB x2 .
(ii) Px1 > Px2 for all bipolarization indices P : D ! R1 that satisfy IS, IB, SM,
and PP.
If variability of the population size is also not permitted, PP is no longer
necessary.
Duclos and Echevin (2005) introduced bipolarization dominance tests for

income distributions. For any income distribution x, let qd x^ i 1 x^mi and
48
V x c 1n
n
P
IDqd x^ i c, where ID() is an indicator function that takes on the
i1
value 1 or 0 according as qd x^ i c or not. Evidently, Vx(c) represents the

proportion of population for which the proportional income distance qd from the
median exceeds c. The authors argued that a reasonable way of testing bipolarization dominance using proportional distance-based bipolarization indices of the two
distributions x and y is to compare Vx(c) and Vy(c) for all c > 0. The population
sizes and medians of the distributions need not be the same for this result to hold.
2.5
Alternative Notions of Bipolarization Invariance

and the Associated Orderings
An alternative to the relative notion of bipolarization invariance considered in the

earlier section is absolute invariance, which requires uninterruption of polarization
indices under equal absolute changes in all incomes. An ordering of alternative
income distributions by absolute indices is obtainable if and only if their absolute
bipolarization curves do not intersect. An absolute bipolarization curve of a distribution x, which we denote by AB(x, t), where 0 t 1, is obtained by scaling up its
relative bipolarization curve by the median. Formally, ABx; t mRBx; t 0 t
1: The area under this curve is given by mPFW, the absolute sister of the Foster
Wolfson index. Thus, it is possible to convert the FosterWolfson index into its
absolute counterpart by multiplying with the median. The mean of the transformed
n
P
absolute deviations about the median, 1n jx^ i mj, where is increasing and
i1
strictly concave, is as well an absolute index. It is a member of the Bossert

Schworm family. Other examples of absolute bipolarization indices are mP x

1
n
n
P
i1
jm x^ i j
1
and mPNWTG x
n
P
i1
bi jm x^ i j, where and the sequence
{bi} are the same as in (2.6) and (2.7), respectively. For 1, mP becomes the
absolute mean deviation about the median, a violator of IB but not of IS.
Between any two distributions x1 and x2, we can define absolute bipolarization
dominance of x1 over x2 (x1 AB x2, for short) by demanding that ABx1 ; t AB
x2 ; t for all t 2 0; 1, with > for at least one t. We can then develop absolute
versions of results that parallel Theorems 2.2, 2.3, and 2.4 for absolute bipolarization indices (see Chakravarty et al. 2007). Ranking of the distributions by RB may
not coincide with that generated by AB . To understand this, suppose that of two
distributions x1 and x2, where mx1 > mx2 , the relative bipolarization curve of x1
intersects that of x2 exactly once from below at a point t0 less than the midpoint of
2.5 Alternative Notions of Bipolarization Invariance and the Associated Orderings
49
the horizontal axis. That is, up to t0, the curve of x1 lies below and then never lies
below that of x2. But on multiplication of the curves by the corresponding medians,
it may emerge that x1 AB x2 holds. The reason behind this is that on multiplication,
the higher median m(x1) may be able to push the lower curve upward so that x1 AB
x2 appears unambiguously. Thus, some inconclusive rankings in the relative case
may be conclusive in the absolute situation. (See Chakravarty et al. (2007) for
further discussion.)
In view of our discussion on alternative notions of inequality invariance in
Chap. 1, we can say that a more general notion of polarization invariance can be
mbrosio (2010)
the BossertPfingsten intermediate criterion. Chakravarty and DA
noted that the following simple generalization of the relative and absolute notions
of bipolarization curves satisfies the BossertPfingsten intermediate condition:

j
IBC x; ;
n
1 X m x^ i
nji<n m 1 ,
1 j n,
1 X x^ i m
,
nnij m 1
n j n;
2:14

j
where , 0 1, is a value judgment parameter and n n1
2 . IBC x; ; n defines
the ordinate of the intermediate bipolarization curve at the population proportion nj ,
for a given value of . The curve can be defined rigorously the way we have defined

the relative bipolarization curve above. IBC x; ; nj becomes the ordinates of the
relative and the absolute bipolarization curves at the corresponding population
fraction for 1 and 0, respectively. We can rank alternative distributions
of income by intermediate polarization indices that agree with the ordering generated by nonintersecting intermediate bipolarization curves.
mbrosio (2010) indicated, the extension of the Foster
As Chakravarty and DA
Wolfson index to the intermediate set up is given by
IPFW x
2x x AG x
;
m 1
2:15
where AG(x) is the absolute Gini index of the distribution x. PIFW(x) is the area
under the intermediate bipolarization curve of x. It coincides with the Foster
Wolfson index or its absolute version whenever 1 or 0.
An alternative intermediate index of polarization that verifies all the desirable
postulates for an index of bipolarization is the following intermediate extension of
the relative index given by (2.6):
n
X
1
n
IP x
jm x^ i j
i1
m 1
!1
;
2:16
50
which for 0 and 1 becomes, respectively, the absolute and relative mean
deviation of order about the median.
Lasso de la Vega et al. (2010) identified unit-consistent bipolarization indices
under a particular structural assumption. For any x 2 Dn , they defined Zheng-type
intermediate bipolarization curve IBZ x; ; t as a quasilinear mean of the relative
and absolute bipolarization curves. Formally,

IBZ x; ; t f f 1 RBx; t 1 f 1 ABx; t ;
2:17
where 0 1 and f is continuous and increasing in its argument. Evidently, RB

(x, t) and AB(x, t) drop out as polar cases of IBZ x; ; t when 1 and 0,
respectively. For f t t, the curve is a simple weighted average of relative and
absolute curves. Lasso de la Vega et al. (2010) demonstrated that the ordering
generated by two nonintersecting Zheng-type intermediate bipolarization curves
t
obeys unit consistency if and only if f t e for some constant 6 0 and > 0.
t
Plugging f t e into (2.17), we get a Krtscha-type curve. In fact, the

corresponding ordering coincides with the ordering of the underlying distributions
by all unit-consistent bipolarization indices that meets IS and IB.
2.6
Welfare Theoretic Approaches to the Measurement

of Bipolarization
The indices of bipolarization we have analyzed so far are of descriptive nature; they
were not related to any social welfare function. The objective of this section is to
present some bipolarization indices that are derived directly using welfare theoretic
considerations. However, neither of the two approaches undermines the other. The
normative approach that relies on social welfare functions is an alternative way of
looking at the issue of measurement of bipolarization. In fact, some of descriptive
indices can be easily related to some specific social welfare functions.
Chakravarty and Majumder (2001) and Chakravarty et al. (2007) suggested,
respectively, relative and absolute indices of bipolarization that are based directly
on social welfare functions. In their framework, bipolarization is measured in terms
of welfare related to the given distribution. Throughout the section, each income is
assumed to be drawn from [a, b], a nondegenerate interval in the positive part R1
of the real line R1. The ChakravartyMajumder relative index is defined as
PCM x
Ex 2x Ex B1 mx
B2 m;
2m
2m
2:18
where the continuous, increasing, and strictly S-concave ede income function
E satisfies linear homogeneity (see Chap. 1). The two continuous normalization
2.6 Welfare Theoretic Approaches to the Measurement of Bipolarization
51
coefficients B1(m) and B2(m) whose choices are dependent on the welfare functions
have to be selected such that PCM fulfills different postulates for a bipolarization
index. When efficiency considerations are absent (the mean incomes in the segments above and below the median are fixed), the two welfare indicators Ex and
Ex are directly related to PCM(x), and hence satisfaction of IB is ensured. The
positive and negative relationships of the index with the mean incomes above and
below the median, respectively, were assumed with the objective of capturing the
view that greater distancing from the median leads to higher polarization.
To every homothetic social welfare function, we have a relative bipolarization
index of the type PCM. Difference between any two indices will arise in the way we
aggregate incomes of individuals to arrive at an indicator of welfare. For instance,
let the welfare evaluation be done with respect to the Gini welfare function. Then,
under the assumptions that B1 m 4 and B2 m 0, PCM becomes the Foster
Wolfson index for even values of n. Given that all the incomes are positive if we
employ the symmetric mean of order < 1 as the ede income function and use
1
1
B1 m ma
and B2 m 12 ma
2, then the resulting form of PCM turns out
to be an Atkinson-type index of bipolarization. The parameter indicates sensitivity
to bipolarization increasing progressive transfer on each side of the median. More
precisely, for the same transfer, a lower value of makes the distribution more
bipolarized.
Chakravarty et al. (2007) analyzed the following absolute ditto of PCM:
PCMR x
Ex 2x Ex B3 mx
B4 m;
2
2
2:19
where the continuous, increasing, strictly S-concave ede income function E now
satisfies unit translatability. B3 and B4 are the normalizing factors. Since some
functions, including the Gini ede income, are both unit translatable and linear
homogeneous, PCMR becomes the absolute variant of the FosterWolfson index
(for even values of n) when we use the Gini welfare function in (2.19) along with
the additional assumptions that B3 m 4 and B4 m 0. If we use the Kolm
Pollak welfare function involving the single parameter > 0 along with the
am
restrictions that B3 m eam and B4 m me 2 2m, then PCMR becomes
a translation of the KolmPollak absolute index of inequality into an index of
bipolarization. This is an example of an absolute bipolarization index whose
associated welfare function is nonlinear. A higher value of the positive parameter
will increase bipolarization by a larger amount, for any clustering of incomes on
the either side of the median, the lower the incomes are.
Chapter 3
Measurement of Income Multipolar

Polarization
3.1
Introduction
As mentioned in Chap. 2, often it becomes necessary to study polarization of the

distribution of income characterized by more than two poles. This general notion of
polarization, which we refer to multipolar polarization, started with Esteban and
Ray (1994), Duclos et al. (2004), and Esteban et al. (2007). It has been devised to
analyze the formation of any arbitrary number of poles.
A distinction similar to bipolarization and multipolar polarization has been made
in the theory of international relations. For instance, Mansfield (1993) argued that
. . .bipolarity is characterized by the approximate equality of the two largest
states and a wide power disparity between the smallest pole and any remaining
state; and multi-polarity is characterized by the approximate equality of more than
two particularly powerful states and a wide power disparity between the smallest
pole and any other state in the system (op. cit., pp. 109110). Mansfield (1993)
also argued that an analysis of international relations involves examining both
polarity and concentration of the distribution of power.
The formulation of multipolar polarization relies on the concepts identification
and alienation. It is assumed that there is partitioning of the population into
significantly sized income subgroups such that each subgroup has members with
similar incomes and members of different subgroups have dissimilar incomes. Each
subgroup is assumed to represent a particular pole. Identification refers to the idea
that individuals of the same subgroup identify themselves with each other. In
contrast, alienation means that any member of a subgroup feels alienated with
respect to a member of a different subgroup. Polarization is increasing in both the
components; higher identification and higher alienation correspond to an increased
level of polarization.
In the next section of the chapter, we analyze the EstebanRay (1994) index. The
DuclosEstebanRay (2004) polarization index defined for income distributions on
Springer India 2015

53
54
3 Measurement of Income Multipolar Polarization
the continuum is presented in Sect. 3.3. Section 3.4 discusses some alternatives and
variants of the EstebanRay index.
3.2
Income Multipolar Polarization for Discrete

Distributions
This section begins with a discussion on several axioms that are intended to capture
the central idea underlying multipolar polarization. In presenting these axioms we
follow mostly Esteban and Ray (1994). The income distribution is assumed to have
been split into a finite number of income classes (subgroups). Each subgroup has an
income and all the individuals possessing the same income will identify themselves
with the same subgroup. Different values of income give rise to alienation between
members of different subgroups.
For a population
of size n, a typical income distribution with distinct incomes is
given by a pair n; ^x , where x x1 , x2 , . . . xk 2 Dk , x^ is the illfare-ranked
permutation of x, and n n1 ; n2 ; . . . ; nk . Here k stands for the number of subgroups, ni is the number of persons in subgroup i, x^ i is the income of persons in this
k
P
subgroup, and n
ni . Identification felt by individuals in subgroup i is an
i1
increasing function g(ni) of the number of individuals in the subgroup, where

g0 0. Thus, the feeling of identification of a person depends explicitly on the
number of persons who are identical with him
of income. The continuous
in terms
and nondecreasing alienation function A x^ i x^ j , with A0 0, represents
alienation of individual i with respect to individual j. That is, alienation between
two individuals
i and

j is simply represented by a nonnegative transformation of the

distance x^ i x^ j between their incomes. The alienation relation is a symmetric
function in the sense that alienation of individual i with respect to individual j is the
same as that of individual j with respect to individual i. The effective antagonism
felt by person i toward person j is given by a continuous function
H gni , Ajx^ i x^ j, where H gni , 0 0 and it is increasing in alienation on
the strictly positive part of the domain. The effective antagonism relation
is not
a

symmetric relation because although it has a symmetric component A x^ i x^ j , it
depends as well on person is identification representation g(ni).
Esteban and Ray (1994) assumed the following function form of the multipolar
polarization index:
k X
k

X

PER n; ^x
ni nj H gni , A x^ i x^ j :
3:1
i1 j1
That is, PER is the sum of effective antagonisms of different individuals in the
society weighted by respective population sizes of subgroups taken into account for
defining the effective antagonisms.
3.2 Income Multipolar Polarization for Discrete Distributions
55
Esteban and Ray (1994) characterized a particular form of the general index
defined in (3.1) using a population homotheticity condition and three axioms.

According to the population homotheticity condition, for any two pairs n; ^x and
0

0

0
n ;^
y if PER n; ^
x PER n ; ^y then PER cn, x^ PER cn , y^ , where c > 0 is a
constant. This condition demands invariance of polarization ranking of two different distributions with respect to the population size.
The three axioms are based on an income distribution consisting of three distinct
values x^ 1 0, x^ 2 and x^ 3 , and the corresponding population sizes n1, n2 and n3.
Axiom ER1 Let n1 > n2 n3 > 0. Fix n1 > 0 and x^ 2 > 0. There exists c1 > 0
and c2 > 0 (possibly depending on n1 and x^ 2 ) such that if jx^ 2 x^ 3 j < c1 and
x^ 2 x^ 3
,
n2 < c2 n1 , then joining of the masses n2 and n3 at their midpoint,
2
increases polarization.
Since incomes x^ 2 and x^ 3 are moving closer, there should be a reduction in
average alienation. However, higher homogeneity of the subgroups size should
increase identification. Axiom ER1 demands that the reduction in alienation is
dominated by increased identification; consequently, polarization increases. See
Fig. 3.1 for a graphical representation of the merger.
Axiom ER2 Let n1, n2, n3 > 0; n1 > n3 and jx^2 x^3 j < x^ 2 : There exists c3 > 0
such that if n2 is moved to the right, toward n3, by an amount not exceeding c3,
polarization should increase.
There is a reduction in alienation between individuals at x^ 2 and x^ 3 . On the other
hand, alienation increases between individuals at x^ 2 and x^ 1 . Axiom ER2 demands
that polarization should get augmented because of heterogeneity among the subgroups. The movement of the population mass n2 is depicted in Fig. 3.2.
Axiom ER3 Let n1 , n2 , n3 > 0; n1 n3; x^ 2 x^ 3 x^ 2 c4 : Any new distribution
formed by shifting population mass from the central mass n2 equally to the two
lateral masses n1 and n3, each c4 units of distance away, should increase
polarization.
Fig. 3.1 Effect of reduction

of local alienation on
polarization
56
Fig. 3.2 Movement of x^ 2 to

x^ 3 increases polarization
Fig. 3.3 Polarization

increases under the
movement of the middle
class toward the extremes
Axiom ER3 demands increasingness of polarization as the middle class moves to

the higher and lower categories. That is, this axiom indicates that polarization
should increase as the middle class disappears. This disappearance of the middle
class is shown graphically in Fig. 3.3.
Esteban and Ray (1994) demonstrated that the only polarization index of the
form (3.1) satisfying the population homotheticity condition and axioms ER1, ER2,
and ER3 is given by
k X
k
X

PER
n; x K
ni1 nj x^ i x^ j ;
3:2
i1 j1
where the multiplicative constant K is positive and 2 0; 1:6. This index uses an
aggregation similar to that employed in the Gini index. In their formulation,
Esteban and Ray (1994), in fact, used the logarithm of incomes instead of absolute
incomes. In such a case x x1 , x2 , . . . xk 2 Dk . This makes the index scale
invariant type. An individuals feeling of deprivation arises from percentage
income differences not from absolute differences. If we use absolute incomes in
(3.2) and if the parameter takes on the value zero, PER

corresponds to the absolute
Gini index. However, with log of incomes, such a comparison is not valid. The
constant K is used for population normalization. Any positive value of makes the
3.3 Income Multipolar Polarization for Continuous Distributions
57
identification function ni nonconstant. Consequently, it becomes helpful in bringing out the distinguishing features between inequality and polarization. With an
increase in the value of , the departure of inequality from polarization increases.
Therefore, may be regarded as a polarization sensitivity parameter. Given n and k,
PER
takes on its maximum value when the population is equally split between the
is achieved if the
lowest- and highest-income classes. The minimum value of PER
entire population is concentrated at one income. As Esteban and Ray (1994) have
shown, if in (3.2) we impose another axiom which says that polarization increases
under a migration from a very small population mass at a low-income to a higherincome class of moderate size, then cannot be less than 1, that is, 2 1; 1:6. For
absolute incomes we may take K 1 to make the index scale invariant, where > 0
is the mean income. Esteban and Ray (1999) indicated that polarization indices of
arise naturally in behavioral models in which pattern and extent of

the form PER
social conflict are linked to the society-wide distribution of individual
characteristics.
3.3
Income Multipolar Polarization for Continuous

Distributions
Often for statistical examination of polarization, for example, whether the change in
polarization between two periods has been statistically significant, it is necessary to
assume that relevant distributions are represented by income density functions.
Duclos et al. (2004) characterized an index of multipolar polarization for income
distributions defined on the continuum. In this section following Duclos
et al. (2004), Esteban and Ray (2012), and Duclos and Taptue (2014), we present
a discussion on this.
Let F be the cumulative income distribution function on the domain 0; 1. It is
assumed that F is absolutely continuous, that is, it has a continuous derivative. The
continuous derivative of F, which we denote by f, is the corresponding income
density function. The authors assumed that the multipolar polarization index
PDER(F) should be proportional to the sum of effective antagonisms:
Z1 Z1 h
i
0

0
0
PDER F
H f v, v v f vf v dvdv :
0
3:3
f(v) appearing in (3.3) represents identification. The distance function

The density

v v0 summarizes alienation between individuals with incomes v and v0 . An
increased
value of the distance increases the antagonism function and

0
H 0; v v H f v, 0 0.
58
Let f be any income density function on the given domain with mean .
We refer to f as the basic density function. For any 0 < < 1, a squeeze of
f is defined as follows:

1
v 1
f v f
:
3:4
For any 0 < < 1, f is a density function with mean . A squeezing reduces the
variance of the basic density by 1 2 . In the extreme case 1, f coincides
with f. For 0 < < ^ < 1, f second order stochastic dominates f ^ .
Duclos et al. (2004) proposed four axioms for developing a characterization of
the continuous analogue of (3.2). These axioms are stated in terms of an arbitrary
multipolar polarization index P, not necessarily using the form (3.3). The first
axiom is concerned with the impact of squeeze of basic densities on polarization.
Axiom DER1 A squeeze of a distribution consisting of a single basic density
should not increase polarization.
A squeeze means compression of the distribution toward its mean, and this
should indicate no increased polarization. See Fig. 3.4 for a squeeze of a basic
density. A squeeze definitely increases identification but at the same time alienation
is reduced for centrally located individuals. Axiom DER1 then demands that the
increased impact of identification is countervailed by the reduced impact of
alienation.
Axiom DER2 For a symmetric distribution with three nonoverlapping poles, a
squeeze of two noncentral poles does not decrease polarization.
Income
Fig. 3.4 Polarization does not increase under a squeeze of the basic density
3.3 Income Multipolar Polarization for Continuous Distributions
59
Income
Fig. 3.5 Polarization is nondecreasing under a local double squeeze
Because of symmetry the two extreme poles are equidistant from the middle
pole. Thus, axiom DER2 reflects the impact of local squeeze instead of global
squeeze, as considered in Axiom DER1. This axiom is similar to the increased
bipolarity axiom for bipolarization indices and clearly shows distinction between
inequality and polarization. Under local squeeze inequality should reduce but
polarization does not reduce. A squeeze of local poles is depicted graphically in
Fig. 3.5.
Axiom DER3 For a symmetric distribution composed of four nonoverlapping
poles, if each of the two central poles moves closer equally to the corresponding
extreme pole, then polarization increases.
This axiom bears similarity with the increased spread axiom of bipolarization
indices. It represents the impact of increase in inequality across subgroups. Because
of the shift of middle poles to the extreme ones, inequality increases. Increase in
polarization under the shifts clearly demonstrates involvement of inequality in the
measurement of polarization. Figure 3.6 graphically describes the shifts considered
in the axiom.
The final axiom considered by Duclos et al. (2004) is similar to the EstebanRay
(1994) population homotheticity axiom.
Axiom DER4 Let F and G be two income distribution functions. If PF PG,
then for any positive scalar PcF PcG, where cF and cG stand for identical
population scaling of F and G, respectively.
When these four axioms are invoked jointly, the following form of polarization
index emerges:
60
Income
Fig. 3.6 Symmetric outward shift increases polarization
PDER
F
Z1 Z1
0
0
0
f v1 f v v v dvdv ;
3:5
where 2 0:25; 1. This index is an absolute index. However, it has a compromise
property. If we multiply it by 1, then the resulting index turns out to be a relative
index. For 0, PDER

coincides with the absolute Gini index. However, axiom
overcomes the limitation of PER

which
DER2 is violated if 0. The index PDER
requires a population to be clustered into relevant subgroups.
3.4
Some Variants
In PER
the assumption of identification of individuals within a subgroup ignores
along
disparity within each subgroup. The same assumption is maintained in PDER
with the further restriction that the number of possible subgroups can be infinity. As
Duclos and Taptue (2014) pointed out, one can consider a given finitely many
number of subgroups whose positions are chosen with the objective of maximizing
internal subgroup homogeneity. This ensures maximization of local (within subgroup) identification and minimization of local (within subgroup) alienation.
Esteban et al. (2007) adopted such an approach and suggested a variant of PER
that takes care of the above issues. Consider an income distribution represented
by the distribution function F and let the corresponding density be f. Given
Rx^ i
Rx^ i
that there are k income classes, let i
f v dv and i 1i
vf v dv,
x^ i1
x^ i1
respectively, be the population frequency and mean income of the income class
3.4 Some Variants
61
^x i1 ; ^
x i , i 1, 2, . . . , k: We denote the vectors of population proportions and
mean incomes of different income classes by and , respectively.
Bunching of a continuous-type income distribution into a finite number of
income classes generates approximation error in the measurement of polarization
of the given distribution. The error is decreasingly related with internal homogeneity of subgroups. As the error increases, the local identification reduces
and local
alienation increases. The EstebanGradinRay (2007) index LEGR

, is PER
,
applied to the discrete grouping considered here, with a correction for within-group
inequality. More precisely it is defined as
k X
k

X

LEGR
,
i1 j i j we , ;
3:6
i1 j1

where e , stands for the measurement error and w is the nonnegative weight
assigned to the error. The negative association of the error term ensures that LEGR

, is decreasing in within-group and increasing in between-group disparities.
The error function has been chosen as the average of income distances within all
subgroups. That is, the error term is given by the sum of all within-group Gini
indices, equivalently, the sum of within-group alienation. They then considered the
problem of clustering the population such that the error function is minimized. If we
denote the solution to the minimization problem by F^ , then the associated Esteban
GradinRay (2007) index is obtained by substituting F^ for Fin (3.6).
If we assume
that there are only two subgroups and w 1, then LEGR

, coincides with m2
times the FosterWolfson index.
A second variant of PER

was suggested by Lasso de la Vega and Urrutia (2006).
Their index is defined as
k X
k

X
,
,
1
j 1 I Gi logi log j ;
PLU
i
3:7
i1 j1
where IiG is the Gini index of the ith subgroup and 0 is a parameter that
represents sensitivity of the Gini index of equality of the ith subgroup to polariza

tion. The term i 1 I Gi in (3.7) can be regarded as a measure of the sense of
identification of each member of the subgroup i. Since the subgroup Gini index IiG
,
has a negative association with PLU
, it is clear that the polarization is a decreasing
function of within-group dispersion. On the other hand, it is increasingly related to
since intergroup transformed mean differences
between-group inequality

logi log j have a positive association with P, . However, (3.7) does not
LU
take into account the approximation error.
An alternative variant
of PER was proposed by DAmbrosio (2001). She replaced

the distance function x^ i x^ j used in (3.2) by the Kolmogorov measure of distance
62
1
2

R1
f v f vdv, where fi and fj are, respectively, the densities of income
i
j
0
distributions corresponding to these subgroups. The measure of distance

R1
1
f v f vdv is maximized when fi and fj do not overlap. On the other
i
j
2
0
hand, it achieves the minimum value 0 if fi and fj coincide.

The objective of the index was to make the alienation function sensitive to the
income distribution-based disparity between subgroups, which is not considered in
(3.2). It is defined as
Z1
k X
k
1X

PD ; f
i1 j f i v f j vdv;
2 i1 j1
3:8
where f stands for the vector ( f1, f2, . . ., fk). The error correction is not considered in
this index as well. The index takes on its minimum value zero if there is complete
overlap between fi and fj for all distinct pairs (i, j).
Chapter 4
Reduced-Form Indices of Income

Polarization
4.1
Introduction
Often two or more economic indicators can abbreviate the entire income distribution
of a society. For instance, a reduced-form welfare function represents social
welfare as an increasing function of efficiency (mean income) and a decreasing
function of inequality (see Sect. 1.3 of Chap. 1). Likewise, essential to the construction of an abbreviated or reduced-form polarization index is a (population)
subgroup decomposable income inequality index. A reduced-form polarization
index is increasingly related to the between-group term and decreasingly related
to the within-group term of a subgroup decomposable inequality index. The central
idea underlying this is that lower inequality within a subgroup increases homogeneity within the subgroup because it represents lower spread of the subgroup
income distribution. On the other hand, higher distance across subgroups, that is,
higher between-group inequality, indicates more heterogeneity across subgroups.
This framework can be reinterpreted in terms of the identificationalienation
structure; within-group inequality indicates a loss of identification, and betweengroup inequality can be taken as a proxy for alienation. The reduced-form or
abbreviated polarization indices can be employed to characterize the trade-off
between the alienation and identification components of polarization.
Zhang and Kanbur (2001) proposed an index of polarization, which is given by
the ratio between the between-group and within-group components of a subgroup
decomposable index of inequality. In a similar approach, Rodriguez and Salas
(2003) considered the partitioning of the population using the median income and
defined a bipolarization index as the difference between the between-group and
within-group components of the DonaldsonWeymark (1980) S-Gini index of
inequality. Deutsch et al. (2007) considered the bipartitioning of the population
using the median and suggested an index which is increasingly related to the
ZhangKanbur index, applied to the Gini index.
Springer India 2015

63
64
4 Reduced-Form Indices of Income Polarization
Chakravarty and Maharaj (2011a) characterized several reduced-form polarization indices, including a generalization of the RodriguezSalas form. The structure
of a normalized ratio form index bears similarity with the ZhangKanbur index.
They then demonstrated that it is possible to start with a specific form of a
polarization index and identify the corresponding inequality index which would
generate the given polarization index. This may be treated as the dual of the
characterization results for polarization indices.
For such indices, one characteristic such as age, sex, race, etc. is used to define
the subgroups, whereas income clustering (within subgroups) and income distances
(between subgroups) are used to define identification and alienation respectively.
Esteban and Ray (2012) refer to such indices as multidimensional indices. The
reason behind this is that an additional characteristic, for defining the subgroups, is
used for determination of alienation and identification. That is, the distribution of
income and some other qualitative characteristic, say region or race, have a
meaningful appearance while defining the subgroups between which polarization
is desired to be measured. They argued that apart from income inequality, increased
alienation across subgroups can as well be a consequence of issues like intolerance
between religions and attitudes toward foreigners. Because of high within-group
inequalities, individuals in some subset may be tempted to employ individuals in
another subset by providing financial support to get involved into a conflict. That is,
the richer subset provides financial support, while the poorer subset provides labor
support to the conflict. In other words, the poor are subsidized by the rich to provide
conflict labor.1 Since such indices are based on socioeconomic partition of the
population, Duclos and Taptue (2014) and Permanyer (2014) refer to them as
socioeconomic polarization indices and hybrid indices, respectively. In Chap. 5,
we deal with indices of polarization that are solely based on such characteristics.
Chakravarty and Maharaj (2011a) also developed an ordering of income distributions using reduced-form indices. This ordering enables us to say whether one
income distribution has higher or lower polarization than another by all abbreviated
polarization indices that satisfy certain conditions. If the population is bipartitioned
using the median, then this ordering becomes close to the FosterWolfson (2010)
bipolarization ordering.
After discussing the background material in the next section, Sect. 4.3 presents
some preliminary concepts. Section 4.4 analyzes the polarization ordering. A
detailed discussion on alternative reduced-form indices is provided in Sect. 4.5.
1
This issue is elaborated further in Esteban and Ray (2008a, 2011), and Chap. 7 of this book
presents a brief discussion on this.
4.2 The Background
4.2
65
The Background
This section presents the backdrop for defining a reduced-form index of polarization rigorously. For a population of size n, let xi > 0 be the income of person i so
that the vector x x1 ; x2 ; . . . ; xn 2 Dn represents the distribution of income for
this population. Consider any subgroup decomposable inequality index that defines
between-group inequality in terms of variations of mean incomes across subgroups.
We assume throughout the chapter that the number of subgroups (k) is arbitrary and
exogenously given. Under a progressive transfer of income between two persons in
a given subgroup, inequality within the subgroup reduces without upsetting the
between-group inequality. But polarization increases because of higher-income
homogeneity of individuals within the subgroup. Given two subgroups, for a
proportionate (an absolute) reduction in all the incomes of the one with lower
mean, the subgroup relative (absolute) inequality remains unchanged but its mean
income decreases further. Similarly, a proportionate (an absolute) increase in the
incomes of the other subgroup will raise its mean but does not induce any change in
its relative (absolute) inequality. This implies that between-group inequality
increases. Equivalently, we can say that a greater distancing between subgroup
means, without affecting within-group inequality, enhances between-group
inequality making the subgroups more heterogeneous. A sufficient condition
that

ensures satisfaction of this requisite is that the decomposition coefficient i n, ,
the weight assigned to inequality in subgroup i , for the generalized entropy family
(1.22) and the absolute subgroup decomposable family (1.25), is given by nni , the
subgroup population share. The only subgroup decomposable indices for which this
condition are verified are the Theil mean logarithmic deviation index IcS for c 0
and the variance IV (see Eqs. (1.22) and (1.28)). Note that both these indices are unit
consistent (see Eq. (1.29)).
If we assume that the inequality index satisfies the FosterShneyerov (1999)
generalized decomposability condition (1.25), where in the calculation of betweengroup inequality the subgroup mean is replaced by the subgroup symmetric mean of
any arbitrary order q, then the only inequality
q index
for which the decomposition

EA x
cq
ni
coefficient i n, equals n is I FS x log E0 x , c 0, q 6 0; where EqA (x)
A
is the symmetric mean of order q (see Eq. (1.26)). This index will satisfy the Pigou
Dalon Transfers Principle if and only if q 1. As the next possibility if we consider
the BlackorbyDonaldsonAuersperg (1981) ethical approach to the inequality
decomposition by population subgroups, then the only inequality index for which
the within-group factor is the population share weighted average of subgroup
inequality levels is the KolmPollak absolute index (see Eq. (1.18)). In this case
the equally distributed equivalent incomes of the subgroups form the basis of
between-group inequality. In the remainder of this chapter, which is mostly based
on Chakravarty and Maharaj (2011a), we restrict our attention to these four
inequality
is denoted by SD, that is,
n indices and the set of these four indices
o
SD I Sc , c 0; I V ; I cq
FS , c 0, q 1; I KP , > 0 . All of them are onto functions,
66
that is, for each nonnegative real number, we can get an income distribution whose
inequality is equal to the considered real number. Also they vary continuously over
the entire nonnegative part of the real line.
4.3
Preliminaries
We assume at the outset that all income distributions over a given population size
are partitioned into arbitrary number of subgroups using some homogeneous
characteristic. A polarization index P is a real-valued function of income distributions of such a partitioned population. Formally,
Definition 4.1 A polarization
index is a continuous function P : ! R1 , where
!
Q
Dni and N=f1g.
[
k2 ni 2 , 1ik

For any x x1 , x2 , . . . xk 2 , k 2 , the real number P(x) is an indicator the
extent of polarization associated with x.
Definition 4.2 A continuous function P : ! R1 is called
or
1 a reduced-form

2
k
abbreviated form polarization index whenever for all x x ; x ; . . . ; x 2 , k
2 , Px can be written as Px f BI x, WI x, where I 2 SD is arbitrary, BI
and WI are respectively the between-group and within-group components of I and
the real-valued function f defined on R2 is continuous.
The function f considered in Definition 4.2 can be referred to as a characteristic
function. If the inequality index I 2 SD is the Theil mean logarithmic deviation or
the FosterShneyerov (1999) index, then the polarization index P will be of relative
type. On the other hand, P will be of absolute variety if I 2 SD is the variance or the
KolmPollak index. The only member of SD that makes P a unit consistent absolute
index is the variance.
Homogeneity within a subgroup and heterogeneity across subgroups are congenital characteristics of the notion of polarization. These two concepts can be
taken into account properly if we assume that the function f is increasing in BI and
decreasing in WI. We refer to such indices of polarization indices as feasible
indices. Formally,
Definition 4.3 A reduced-form polarization index P : ! R1 is called feasible if
the function
in the formulation Px f BI x, WI x, where I 2 SD,
1 2 f considered

k
x x ; x ; . . . ; x 2 and k 2 are arbitrary, is increasing in BI and decreasing
in WI.
We can motivate feasibility of a reduced-form index further with the graphical
representation of the between-group and within-group components of the Gini
index for a population bipartitioned using the median income as the line of
demarcation (see Fig. 4.1). For any income distribution x, the three values L(x, 0),
4.3 Preliminaries
67
Line of equality
a2
a3
Cumulative
proportions
of income
a1
a4
0
1/2
Cumulative population shares
Fig. 4.1 Between-group and within-group components
L(x, 0.5), and L(x, 1) on the Lorenz curve L of x are joined to form a piecewise linear
curve. This curve is the between-group Lorenz curve. The within-group Gini index
WIG(x) is twice the area enclosed between the original Lorenz curve L and the
between-group Lorenz curve, that is, WI G x 2a3 a4. On the other hand, the
area of the triangle formed by line of equality and the between-group Lorenz curve,
that is, a1 a2, is simply 0:5L2x;0:5. Hence the between-group Gini index BIG(x)
is given by BI G x 2a1 a2 0:5 Lx; 0:5. Given a1 a2, bipolarity
increasing progressive transfers reduces WIG(x), that is, the original Lorenz curve
moves closer to the line of equality. Consequently, PFW(x) in (2.3) increases. In
contrast, given other things, a greater distancing between the subgroups in terms of
higher gap between their means will increase a1 a2 so that the between-group
Lorenz curve moves away from the line of equality. This as well leads to an
increase in PFW(x). These properties of PFW support the feasibility assumption
made in Definition 4.3.
It will now be worthwhile to make a systematic comparison of this notion of
polarization with the EstebanRay (1994) multipolar concept. While in both cases
the alienation factor is based on income distances, the identification component in
the EstebanRay case is a function of population proportions, and in the reducedform situation, it is inversely related to within-group inequality involving both
68
subgroup sizes and subgroup inequality levels. Consequently, the impact of a small
subgroup on identification and hence on polarization becomes low if the subgroup
is delineated with high inequality and makes high contribution to within-group
inequality.
Since in the EstebanRay framework, the axioms are formulated in terms of
population shift, minimum polarization arises when there is perfect population
homogeneity, that is, the whole population is centered in a subgroup. This is the
situation of maximum identification. In contrast, in the reduced-form formulation,
the minimum polarization arises if BI 0 and WI is maximum. This follows
directly from Definition 4.3. This definition also implies that maximum polarization
takes place when identification is maximum WI 0 and alienation (BI) is
maximum as well. Redistribution of incomes is an important determinant of
polarization in the reduced-form structure. On the other hand, in the EstebanRay
formulation, maximum polarization is symbolized under equal division of the entire
population into two subgroups and each of the remaining subgroups possessing zero
population. These differences emerge in view of different basic formulations. The
reason behind this is that in the former formulation, the number of subgroups is
assumed to be given a priori, whereas the latter allows the possibility of variability
of subgroups along with shifts of population across subgroups.
When the number of subgroups k increases and coincides with the population
size n, each subgroup contains only one person which means that within-group
inequality becomes undefined. In this case inequality is designated by the betweengroup factor, which in turn implies that inequality gets related to polarization
directly. For simplicity, here an inequality indicator itself can be treated as a
reduced-form index of polarization. There may be different impacts of a merger
of two equally sized groups at the midpoint on these two notions of polarization. In
the reduced-form situation, this will lead to an unambiguous reduction of inequality
as well as polarization. On the other hand, in the EstebanRay approach, this is
determined by the shape of the distribution. This difference follows because while
in the former, polarization is treated in terms inequality with a fixed number of
subgroups, in the latter, as we have observed earlier, the formulation allows
variability of subgroups as well as shifts of population across subgroups. However,
polarization is a multifaceted aspect of an income distribution and the two notions
depict two alternative views. Given that each approach has its own way of incorporating the two components, identification and alienation, in different ways, each
has got its self appeal. Therefore, our objective is definitely not to unseat one
formulation with the other and vice versa.
4.4
An Ordering
The number of polarization indices satisfying Definition 4.3 is rather large since the
number of f -functions is not finite. For two different f -functions, polarization
ranking of two different distributions may or may not be the same. The objective of
4.4 An Ordering
69
this section is to develop a unanimous ranking of two income distributions by all

reduced-form polarization indices.
In order to formulate such an ordering of alternative income distributions, for
arbitrary
k 2,
ni 2,
1 i k, consider the distributions
k
1 2

Q
x x ; x ; . . . ; xk , y y1 ; y2 ; . . . ; yk 2
Dni . Then we say that x is more
i1
polarized than y (xP y, for short) if Px > Py for all feasible polarization indices
k
Q
P:
Dni ! R1 . This definition of P does not require equality of the total incomes
i1
of the distributions.
k

Q
Given the distribution y y1 ; y2 ; . . . ; yk 2
Dni , we can bring about a dis-
tribution x x1 ; x2 ; . . . ; x

k
k
Q
i1
Dni by one of the following transformations:
i1
(i) decreasing WI without affecting BI, (ii) increasing BI without changing WI, and
(iii) decreasing WI and increasing BI. Each of these three transformations is a
polarization augmenting transformation. These three transformations can be
expressed more compactly as BI x BI y and WI x WI y with strict
inequality in at least one case.
The following theorem of Chakravarty and Maharaj (2011a) shows that this
requirement is the same as the condition xP y.
k

Q
Theorem 4.1 Let x x1 ; x2 ; . . . ; xk , y y1 ; y2 ; . . . ; yk 2
Dni , where k 2,
i1
ni 2, 1 i k are arbitrary. Then the following conditions are equivalent:

(i) xP y:
(ii) BI x BI y and WI x WI y for any inequality index I in SD, with strict
inequality in at least one case.
Given any inequality index in the set SD, condition (ii) of Theorem 4.1 can be
verified quite easily. The theorem says that if condition (ii) holds, then we can
unambiguously claim that the distribution x has higher polarization than the distribution y by all reduced-form polarization indices that are increasing in BI and
decreasing in WI. The result holds irrespective of the magnitudes of the total
incomes of the two distributions, that is, total incomes across the distributions
need not be the same for the theorem to hold. This is because each inequality
index in the set SD is either scale or translation invariant and this property enables
us to make unequal total incomes equal without altering the level of inequality.
We demonstrate now that P is incomplete and transitive. Let us first consider
the bipartitioned distributions x 2; 6; 10; 4; 12 and y 2; 6; 10; 4; 8,
where (2, 6, 10) and (4, 12) are respectively the subgroup income distributions
in x and so on. For these two distributions WAV x 64
5 and WAV y 8, whereas
70
BAV x 24
25 and BAV y 0, where BAV and WAV denote respectively the
between-group and within-group components of the variance AV. Then x has higher
between-group and within-group inequality components, that is, BAV x > BAV y
and WAV x > WAV y. Consequently, x and y are not comparable with respect to
P . Next, consider three distributions x, y and z over a given population size,
partitioned using the same characteristic into equal number of subgroups. Suppose
we have xP y and yP z. Then it follows that xP z holds, which proves transitivity
of P . This establishes that P is a quasi-ordering.
Next, we demonstrate that P may not agree with inequality ordering of
distributions.
For
this,
consider
the
bipartitioned
distributions
y a1 ; a2 ; a3 ; a4 and x a1 , a3 ; a2 , a4 , where a1 < a2 < a3 <
2
a4 and 0 < < a3 a
2 . For these choices of x and y, we have BAV y < BAV x but
WAV y > WAV x. Consequently, for all feasible polarization indices P, we have
Py < Px, but in view of the PigouDalton transfer principle, it follows that
AV y > AV x:
k

Q
Next, suppose the income distribution y y1 ; y2 ; . . . ; yk 2
Dni is generated
from the distribution x x1 ; x2 ; . . . ; x

k
k
Q
i1
j
i1
Dni by the following transformation:
yi xi for all i 6 j and yj is obtained from x by a progressive transfer of income

between two persons in subgroup j. Evidently, for an additively decomposable
index I, BI x BI y and WI x > WI y. Hence I x > I y. In contrast, for any
feasible polarization index P, Py > Px: Thus, the two examples clearly demonstrate opposite directional rankings of distributions by inequality and polarization. The underlying reasoning is that while both the components BI and WI are
increasingly related to inequality, the former components relationship with polarization is increasing but with the latter it is decreasing.
Evidently, the ordering P is sensitive to the choice of the characteristic used to
define the partitioning of the population. Thus, while for ethnic partitioning one
distribution may be regarded as more polarized than another, for geographic
location partitioning, the opposite may happen. The reasoning behind this is that
identification of the subgroups depends on the characteristic used for defining the
partitioning.
For the median-based bipartitioning of the distribution, higher alienation means
increasing the distance between the subgroups below and above the median and this
is achievable by increasing (decreasing) incomes proportionately above (below) the
median. This in turn shows that alienation has a similar spirit as the increased
spread axiom used in the bipolarization ordering. Next, identification increases
under a progressive transfer of incomes between two individuals on the same side of
the median. This indicates that the identification criterion represents the same
concept as the increased bipolarity axiom of bipolarity measurement. Therefore,
the two polarization orderings FW and P are equivalent when the median of the
income distribution is used to form two population subgroups. (See Chakravarty
and Maharaj (2011a), for a detailed discussion.)
4.5 Analysis of Reduced-Form Indices and Their Eventual Differences
4.5
71
Analysis of Reduced-Form Indices and Their Eventual

Differences
As we have observed, a polarization ordering may not provide unambiguous

ranking of two income distributions. However, a particular index of polarization
can completely order the distributions. A rigorous discussion on such indices
becomes helpful for getting insight into the implicit value judgements involved in
the choice of an index. This is the objective of the present section.
A characterization of a particular polarization index enables us to understand the
index in terms of the axioms employed in the characterization exercise. The axioms
constitute a set of necessary and sufficient conditions that isolate the index
uniquely. In fact, aximoatizations become helpful to separate two different indices
on the basis of the axiom sets.
It will be assumed throughout the section that all polarization indices considered
for our discussion are feasible. A society may become turbulent if its polarization
level exceeds certain threshold level/tolerance limit. In such a situation, there is
a possibility that even a minor increase in polarization, resulting from an increase in
alienation/identification, will make the society highly tensed which in turn may lead
to conflict. As Esteban and Ray (1994) argued, when a population is already
largely bunched at the two extreme points, additional bunching may increase
polarization. Now, for a quite peaceful society identified with a low level of
polarization, the tolerance limit is likely to be low. On the other hand, the tolerance
limit is expected to be high if the society already possesses a high degree of
polarization. It is also likely that net change in polarization will not be low for a
society marked with a higher level of polarization. Consequently, we can assume
that the increase in polarization has a non-decreasing relationship with alienation
and identification over the entire domain.
Chakravarty and Maharaj (2011a) proposed the following axioms for a feasible
polarization index. In view of the basic structure, they are formulated in terms of
inequality change.

Axiom (A1) For any x x1 ; x2 ; . . . ; xk 2 , k 2 , and for any nonnegative real
number , f BI x , WI x f BI x, WI x BI x, WI xg for some
continuous functions : R2 ! R1 and g : R1 ! R1 , where is non-decreasing in
BI, g is increasing, g0 0, and I 2 SD.

Axiom (A2) For any x x1 ; x2 ; . . . ; xk 2 , k 2 , and for any nonnegative real
number , f BI x, WI x f BI x, WI x BI x, WI xh for some
continuous functions : R2 ! R1 and h : R1 ! R1 , where is non-decreasing in
WI, h is increasing, h0 0, and I 2 SD.

Axiom (A3) For any k 2 , if x x1 ; x2 ; . . . ; xk 2 is of the form xi c1ni ,
where ni 2 for all 1 i k and c > 0 is any arbitrary scalar, then for any
I 2 SD, f BI x, WI x 0.
72
Axiom (A1) says that an increase in polarization resulting from an increase in

between-group inequality can be decomposed into two components, one involving
a non-decreasing function of the original values of the between-group and the
within-group factors of inequality and the other is an increasing transformation of
the increment in the between-group factor. It is natural to assume that is
non-decreasing in BI. A similar view, in a stricter way, is represented by
increasingness of g. Given other things, the higher is the level of increment in
BI, the higher is the increase in polarization. If there is no change in BI, then given
WI, there should be no change in polarization. This is reflected by the assumption
that g0 0. If we assume that the function f is differentiable, then this axiom
means convexity of the polarization index in BI. We can provide an analogous
explanation for axiom (A2). The functions g and h indicate sensitivity to alienation
and identification respectively. Axiom (A3) is a normalization condition which says
that a polarization index takes on the value zero when the income distribution
possesses no inequality. Since for a perfectly equal distribution x we have
BI x WI x 0, an equivalent formulation of (A3) is f 0; 0 0.
We may note here that a bipolarization index can as well satisfy decompositions
of the type stipulated in axioms (A1) and (A2). To see this formally, consider the
5-person income distribution x^ x^ 1 , x^ 2 , x^ 3 m, x^ 4 , x^ 5 , and let the absolute
5
P
1 jx^ i mj
symmetric normalized bipolarization index be Qx^ 1 5 e i1
.
This index satisfies the increased spread and increased bipolarity axioms. Now,
define the illfare ranked income distribution y^ y^ 1 , y^ 2 , y^ 3 m, y^ 4 , y^ 5 , where
y^ i x^ i , for 1 i 4 and y^ 5 x^ 5 c, c > 0 being a constant. That is, y^ is
obtained from x^ by simple (rank preserving) increment. Then by the increased
spread axiom bipolarization increases. The resulting increase in bipolarization is
0
1
5
P
jx^ i mj
1
A1 ec . That is, the increase in bipolarQy^ Qx^ @1
e i1
5
ization is the product of the original polarization level and an increasing function of
the increment in the highest income. The essential idea underlying this and axiom
(A1) is the same.
Chakravarty and Maharaj (2011a) demonstrated that a feasible polarization
index P : ! R1 satisfies axioms (A1), (A2) and (A3) if and only if it is of one
of the following forms, for some arbitrary positive constants c1 and c2:
(i) P1 x c1 BI x c2 WI x;
BIx

c1
a
1 c2 WI x, a > 1;
(ii) P2 x loga

c
1
WI x c2 WI x, 0 < a < 1, c2 0;
(iii) P3 x aBIx 1 loga
WIx

c2
b
1 , b > 1;
(iv) P4 x c1 BI x logb

c
2
BI x , 0 < b < 1, c1 0;
(v) P5 x c1 BI x bWIx 1 logb
73
BIx

WIx

c1
c2
(vi) P6 x loga
a
1 logb
b
1 , a > 1, b > 1;
BIx

WIx

c1
c2
(vii) P7 x loga
a
1 logb
b
1 aBIx 1 bWIx 1 , a > 1,
0 < b < 1, 0 loga c1 ;
BIx

WIx

c1
c2
(viii) P8 x loga
a
1 logb
b
1 aBIx 1 bWIx 1 ,
0 < a < 1, b > 1, c2 logb 0
BIx

WIx

c1
c2
(ix) P9 x loga
a
1 logb
b
1 aBIx 1 bWIx 1 , 0 < a,
c1
c2
b < 1, loga
logb
;
1 2

where x x ; x ; . . . ; xk 2 , k 2 and I 2 SD are arbitrary.
The constants c1 and c2 can be interpreted as scale parameters in the sense that
under ceteris paribus conditions, an increase in c1 (a reduction c2) increases
polarization. These parameters cast significance of alienation and identification in
the aggregation. Similar interpretations can be provided for other parameters.
Consider the bipartitioning of the distribution into two nonoverlapping subgroups using the median as the line of separation. For this bipartitioning to be
well defined, the income distribution is assumed to be welfare ranked. In such a
case, we can subdivide the DonaldsonWeymark S-Gini inequality index I DW

into
between-group and within-group components. If in P1 we use the between-group
and within-group components of I DW

under the additional assumption that
c1 c2 1, then the resulting form of P1 becomes the RodriguezSalas (2003)
index of polarization. A necessary condition for this index to satisfy the increased
bipolarity axiom is that 2 3. Recall that for 2, I DW

coincides with the Gini
index. Thus, the RodriguezSalas (2003) index, which we can refer to as the S-Gini
index of polarization, shows an alternative application of the Gini index to polarization measurement.
The RodriguezSalas index regards all income distributions with equal betweengroup and within-group factors of inequality as equally polarized. For instance, an
income distribution x with BI x WI x :05 becomes identically polarized as
the perfectly egalitarian income distribution y with BI y WI y 0: Therefore,
in situations of the type where BI WI, this problem for P1 can be avoided if
different choices of c1 and c2 are made. In fact,this observation
applies as well to

choices of a1 and a2 in the index Pa1 , a2 x
BI x
a1
WI x
a2
1 , which drops out as a
special case of P7, where a1 > 1 and a2 > 1: are parameters. To see that Pa1 a2 is a
c1
c2
particular case of P7, choose loga
logb
1 in P7. Then it follows that
BIx

BIx WI x
b
1, which can also be expressed as P7 x aa21WIx 1 ,
P7 x a
where a1 a > 1 and 1b a2 > 1:
It is now necessary to discuss the eventual differences among the indices
P1 P9 . For this purpose, Chakravarty and Maharaj (2011a) considered the
following properties:
74
Property 1 P is strictly convex in BI.

Property 2 P is strictly concave in WI.
These properties become helpful in making policy choices about the rate of
increase in identification and alienation components. P1 is a violator of these
properties and hence may be treated as a rather weak indicator. In contrast, P2
and P3 are violators of only the second property, whereas P4 and P5 are violators of
Property 1 but not of Property 2. On the other hand, each of the indices P6 P9
fulfils both the properties. Consequently, they can be regarded as strong indicators. Indices P7 P9 are identical in the sense that they vary only in terms of
parametric restrictions. It is evident that if we are concerned with polarization
ranking of two distributions using the ordering P and the ranking is unambiguous,
then no difference arises among these indices.
The ratio cc21 in P1 is the marginal rate of substitution of alienation (BI) for
identification (WI) along an iso-polarization contour, the locus of (BI, WI) combinations that generate the same value of polarization. This ratio indicates how WI
can be exchanged for BI along the contour. We can incorporate this alienation
identification collision in a more general way. Suppose all the incomes in the
subgroup with the minimum subgroup mean are reduced proportionately or by
the same absolute amount. This change increases the differences between subgroup
means and hence BI for an additively decomposable index I, that is, alienation
increases by some positive amount , say. If we wish to keep the level of polarization constant, then this increase in polarization has to be compensated by a
reduction in identification. This can be achieved through a sequence of regressive
transfers within one or more subgroups. This reduction is a function of and we
denote it by g1(). That is, in order to remain on the same iso-polarization contour,
an increase in BI by , has to be followed by an increase WI by some amount g1().
A parallel argument will establish that if WI increases by , then a corresponding
increase in BI by g2 , say, has to take place to keep the level of polarization
unchanged (see also Esteban and Ray 1994, p. 828, pp. 8456 and Chakravarty
et al. 2010, for a related discussion).
In view of the above argument, the following axiom can be stated:

Axiom (A4) For any x x1 ; x2 ; . . . ; xk 2 , k 2 and for any nonnegative ,
f BI x, WI x f BI x , WI x g1 f BI x g2 , WI x for
some continuous functions g1 , g2 : R1 ! R1 , where g1 0 g2 0 0.
The assumption g1 0 g2 0 0 ensures that if there is no change in identification or alienation, polarization remains unaltered.
Chakravarty and Maharaj (2011a) established that a feasible polarization index
P : ! R1 satisfies axioms (A1) (or (A2)), (A3), and (A4) if and only if it is of one
of the following forms:
(x) Pc1 , c2 x c1
BI x c2
WI x for some arbitrary constants c1 , c2 > 0,
(xi) Pa1 , a2 x c
BI x
a1
WI x
a2
1
for some arbitrary constants c > 0, a1 , a2 > 1,
75

where x x1 ; x2 ; . . . ; xk 2 , k 2 and I 2 SD are arbitrary.
This is a joint characterization of the normalized ratio form index Pa1 , a2 and the
difference form index P1.
For all x x1 ; x2 ; . . . ; xk 2 , k 2 and I 2 SD, the ZhangKanbur index is
BI x
defined by the ratio PZK x WI
x. The structure of this index is similar to the
transformed ratio form index 1 Pa1 , a2 , which is derived from (xi) for c 1.
However, PZK is undefined if WI x 0. This situation arises if inequality within
each subgroup is 0, even when subgroup mean incomes may be different. But the
transformed index 1 Pa1 , a2 avoids this problem. However, the alienation and
identification factors of polarization are embodied correctly in the formulation of PZK.
Chakravarty and Maharaj (2011a) demonstrated that the axioms in each of the three
sets S1 fA1; A2; A3g, S2 fA2; A3; A4g, and S3 fA1; A3; A4g
are independent. According to the independence property of a given set of postulates,
none of the postulates in the set implies or is implied by any postulate in the subset of
the remaining postulates. Consequently, if one postulate is dropped from the set, the
remaining postulates in the set will not characterize the corresponding indicator.
Hence demonstration of independence of the postulates in S1 fA1; A2; A3g
requires that if one postulate, say (A1), is dropped from S1, there exist at least one
reduced-form indicator of income polarization, which is different from the indices
i ix stated above, that will fulfil the postulates in {(A2), (A3)} but not (A1). This
has to be demonstrated for each postulate in the set S1. Similar statements hold for the
sets of axioms S2 and S3.
Deutsch et al. (2007) also used the Gini index to suggest a polarization index for
a bipartitioned income distribution, where the median has been taken as the line of
demarcation. As before, we denote the between-group and within-group components of the Gini index for the bipartitioned distribution by BIG and WIG respectively. Then the functional form of the DeutschHanokaSilber polarization index
is given by
PSDH x
BI G WI G
:
BI G WI G
4:1
As desired, this index is increasing in BIG and decreasing in WIG. Clearly, it has a
structure similar to the ZhangKanbur index PZK. In fact, PSDH is increasingly
1
related to PZK via the continuous transformation PSDH x PPZK
, given that the
ZK 1
income distribution is divided into two nonoverlapping subgroups using the median
and the underlying inequality index is the Gini index of inequality. However,
it avoids PZKs discontinuity problem when WI x 0. The authors noted that
1 PSDH x can be regarded as an indicator of kurtosis of the income distribution
(Berrebi and Silber 1989). An indicator of kurtosis is a measure of the extent of
the degree of steepness or peakedness of the distribution.
As we have noted in Chap. 1, it is a common practice in the literature on incomeinequality measurement to retrieve the underlying social welfare function from an
ethical index of inequality. Chakravarty et al. (1985) made a similar attempt to
76
determine the veiled social welfare function from the knowledge of the ethical
income mobility index suggested by them.
An analogous problem is to bring about the inequality index from a given
polarization index. That is, given a polarization index satisfying some structural
conditions, we procreate the hidden subgroup decomposable inequality index. This
is done by developing an appropriate algorithm. This issue may be regarded as the
dual of whipping up polarization indices from inequality indices.
For this purpose Chakravarty and Maharaj (2011a) assumed that for any fixed
k
Q
k 2 and n1 ; n2 ; . . . ; nk 2 k , the polarization index P :
Dni ! R1 satisfies the
i1
following axiom:
k

Q
Axiom (A5) For any x x1 ; x2 ; . . . ; xk 2
Dni , Py Px vi n, gxi ,
i1

where y y1 ; y2 ; . . . ; yk with yi xi 1ni and yj xj for j 6 i; vi is a positive real

number, assumed to depend on the vector n, ; and g is a nonnegative valued
k
function defined on [ Dni .

i1
Since subgroup i possesses less inequality in the distribution y than in the

distribution x and inequalities in all other subgroups in the two distributions are the
same, y should not have less homogeneity and hence less polarization than x. Axiom
(A5) is a very simple representation of this view. Given that the transformation that
takes us from x to y does not alter the distributions in all subgroups other than
subgroup i, it is natural to assume that the function g that summarizes the transformation does not depend on unaffected subgroups distributions. The positive coefficient vi makes the change in polarization dependent on the vectors of subgroup means
and population sizes, which we may regard as a general assumption.
Chakravarty and Maharaj (2011a) proved that if a continuous polarization index
k
Q
P:
Dni ! R1 satisfies axiom (A5), then there exists a corresponding subgroup
i1
k

k
Q ni
decomposable continuous inequality index I :
D [ [ Dni ! R1 of the
type I 1 1n1 , 2 1n2 , . . . , k 1nk
k
P
i1
i1

i n, I xi that assumes the value zero for the
i1
k
ni
perfectly equal distribution on [ D .

i1
This result demonstrates the existence of a subgroup decomposable inequality

index I that can be parented by an algorithm from the polarization index satisfying
Axiom (A5). If P and g are symmetric, then I is so as well. If i n, nni , then
given the domain, this inequality index is some member of SD.
Chapter 5
Social Polarization
5.1
Introduction
In all the earlier chapters, we have assumed that income is the only attribute that
distinguishes one subgroup of a population from another. However, one or more of
cultural, biological, educational, occupational, lingual, color, historical, and political or some other social characteristic of an individual in a society may make him
feel alienated from others in the society. The notion of polarization that applies
when the individuals distinguish among themselves in terms of one or more
characteristics of this type that partitions the population under consideration into
several subgroups is labeled as social polarization. Some classifications may simply
be based on language; some classifications may mix language with color or
language with race difference. More precisely, social polarization means widening
of the gaps between subgroups of a population in terms of some well-defined social
characteristic. For instance, one example of political polarization may be the
situation if peoples political views move from a moderate position to opposing
non-compromising views (see, e.g., Dixit and Weibull 2007).
Thus, one of the major reasons for a separate study of social polarization is that
in many situations income may not be the only characteristic that maintains
individuals identities and deviations. It may be necessary to look at the role of
some other social characteristic, which may be a causal factor of social tensions and
conflict.
Possibly religion is one of the most important characteristics of this type since
sometimes religion may foster stronger loyalty and commitment than any other
social attribute (Juergensmeyer 1993). Religious beliefs have always been those
that people were most willing to sacrifice, fight, and die-and live for (Grew 1997,
p. 20). Religion may act as a catalyst to increase the desire for shouldering the cost
of conflict (Lacina 2006). Throughout recorded human history, religion and
conflict have coincided (McBride and Richardson 2012, p. 113). In ethnic conflict
differences in opinions can be identified in terms of differences in religious beliefs.
Springer India 2015
77
78
5 Social Polarization
According to Horowitz (1985) religious differences have more significance than

language differences as a factor of social separation that can develop a conflict. Seul
(1999) also argued that often religion turns out as the dominant cultural feature
responsible for subgroups getting involved into conflicts.
A highly ethnically expanded society may provoke tensions in the society which
sooner or later may spark conflicts (see, among others, Braubaker and Laitin 1998;
Fearon and Laitin 2003a; Esteban and Ray 2008a, 2011). Using data on the
calamities across space and over time in districts of Nepal, Do and Iyer (2010)
found evidence that greater caste diversity is associated with higher level of
conflict. This shows that social and political implications of ethnic dissimilarity
have received widespread attention in the recent past.
Social instability arising from ethnic heterogeneity may be a factor for a low
economic growth in a society (Easterly and Levine 1997; Montalvo and ReynalQuerol 2005b). Other detrimental consequences of high ethnic diverseness can be
high corruption (Mauro 1995), low social adherence in the sense of less participation in associations (Alesina and La Ferrara 2000), low contribution to local public
goods (Alesina et al. 1999; Alesina and La Ferrara 2005; Bossert et al. 2011), and
reduced literacy and school (Alesina et al. 2003). There are evidences of negative
correlation between ethnic diversity and quality of government (La Porta
et al. 1999) or the size of the government social expenditure relative to gross
domestic product (Alesina et al. 2001) and social capital (Collier and Gunning
2000). Alesina et al. (2011) considered indices of artificiality of states based on
how straight boundaries break ethnic subgroups into different alongside countries.
Economic and political success has been found to be correlated with this index.
Matuszeski and Schneider (2006) constructed an indicator of ethnic diversity and
clustering with the objective of measuring diversity among ethnic subgroups within
a country using digital maps over 7000 linguistic subgroups and 190 countries. The
reason behind this formulation of the authors is that geographical overlaps among
different ethnic subgroups can be regarded as an anticipated provenance for conflict. A natural objective of a society should, therefore, be to reduce ethnic distinctiveness to the maximum possible extent.
For the measurement of ethnic polarization, which is based on ethnic diversity,
the only available information is whether a person belongs to a particular ethnic
category or not. As Duclos et al. (2004) argued, a dichotomous identification of this
type may be quite relevant in many situations. The use of a 0-1 indicator function
becomes appropriate to indicate whether the person belongs to a specific ethnic
subgroup. In other words, in the EstebanRay (1994) framework, the distance
across subgroups is now measured by a (1-0) discrete metric (Montalvo and
Reynal-Querol 2008). Likewise, in the context of cultural fractionalization, Fearon
(2003) defined a resemblance factor that takes on the value 1 if the two groups
speak exactly the same language and its value becomes zero if the languages of the
two groups come from different families. (This interesting proposal also appears in
Caselli and Coleman 2006.).
Reynal-Querol (2002) modified the EstebanRay (1994) income polarization
index to the case of ethnicity and proposed an index of ethnic polarization using a
5.1 Introduction
79
discrete metric. In the literature it is known as the RQ index. From now on, in order
to maintain notational consistency, we will denote it by PRQ. It is a measure of the
extent to which the actual ethnic distribution is far from the bipolar situation, which
says that polarization is maximized if the entire population is split equally between
two ethnic subgroups only and each of the remaining subgroups has zero population
size. Reynal-Querol (2002) observed that religious divisions are important than
language divisions in explaining social ethnic conflicts. Montalvo and ReynalQuerol (2005a) demonstrated that the power of PRQ as a causal factor for the
occurrence of civil wars is increasingly related to the intensification of conflicts.
Montalvo and Reynal-Querol (2008) found that a strong relation exists between
ethnic polarization and risk of genocide. However, Harff (2003) concluded that the
ethnic diversity as a causal factor of genocide is not statistically significant.
Chakravarty and Maharaj (2012) developed a generalization of PRQ. Montalvo
and Reynal-Querol (2008) and Chakravarty and Maharaj (2011b, 2012) investigated several theoretical properties of this index. It is clear that this index can as
well be applied to analyze polarization when the society is subdivided using a social
characteristic of the type mentioned at the beginning of the chapter. For instance, in
Desmet et al. (2008) and Desmet et al. (2012), computation of intergroup distance is
based on linguistic distance, whereas Spolaore and Wacziarg (2013) used genetic
distance. Therefore, unless specified, from now on, we will use the term social
polarization instead of ethnic polarization.
One assumption underlying PRQ is that the distance between any pair of ethnic
subgroups is the same irrespective of the subgroups under comparison. That is, any
two subgroups are equally alienated with respect to each other. However, given the
subgroup sizes, higher distance may give rise to higher social conflict. Often an
individuals involvement with a subgroup may be an important factor in the
measurement of social polarization. Parmanyer (2012) refers to this feeling of
involvement as radicalism and suggests a quantification of this. He proposed
some formulations that take into account the alienationdistance notion in different
ways. In one case, it is assumed that feeling of alienation of individuals belonging to
different subgroups is measured in terms of a microlevel distance function that
depends on the intensity of identification of individuals with respective subgroups.
An alternative formulation assumes that the distance concept is important not only
for measuring polarization between subgroups but also for polarization measurement within subgroups. In a different approach, Collier and Hoeffler (1998) introduced the index of dominance as a dummy variable that assumes the value 1 when
the size of the largest subgroup lies between 45 and 60 %. In a later article, Collier
and Hoeffler (2004), dominance has been defined as a setting in which the size of
the largest subgroup varies between 45 and 90 %.
A particular index of social polarization will completely order two different
populations in a complete manner. However, two alternative indices of social
polarization may not rank two populations in the identical direction. Consequently,
it becomes worthwhile to develop a social polarization ordering that will produce
the same ordering of populations partitioned into social subgroups. For instance, we
may be interested in knowing whether a society is more religiously polarized than
80
ethnically polarized. Another line of investigation can be whether one region of a

country is less polarized than another for a particular social attribute. Intertemporal
comparisons of social polarization may as well be an issue of concern. An ordering
of this type specifies the necessary and sufficient conditions for one population to be
regarded as more or less polarized than another by all social polarization indices
satisfying some common postulates. Chakravarty and Maharaj (2012) made an
attempt along this direction.
The next section defines the Reynal-Querol index PRQ, whose formulation relies
on a discrete metric, and analyzes it from different perspectives. Section 5.3
discusses some characterizations and generalizations of this index. Some alternatives and variants of this index are presented in Sect. 5.4. Section 5.5 deals with the
social polarization ordering.
5.2
A Discrete Metric-Based Index
To discuss the discrete metric-based index PRQ analytically, we consider a population consisting of k social subgroups, where k 2 N=f1g, N being the set of
positive integers. We write i for the proportion of individuals in subgroup i. Hence
k
P
0 i 1, 1 i k, and
i 1, k 2 being arbitrary. This gives rise to a
i1
probability distribution 1 ; 2 ; . . . ; k , which we can refer to as a social

distribution. For any k 2 N, we write k for the set of all discrete probability
distributions of dimension k on the real line R1.
Since essential to the construction of PRQ is the EstebanRay index (1994) of
income polarization, we begin with a specification of the latter in terms of popu0
lation fractions and illfare ranked incomes x^ i s:
k X
k
X

PER
, x K0
1
j x^ i x^ j ;
i
5:1
i1 j1
where the multiplicative constant K 0 > 0 and 2 0; 1:6.

In order to identify the status of an individual with respect to the social
characteristic, it is necessary to check whether he/she belongs to a particular
subgroup or not. As observed by Montalvo and Reynal-Querol
2008), in

(2005a,
a situation of this type, the Euclidean metric d ^x i ; ^x j x^ i x^ j should be
replaced by the discrete metric defined as

d ^x i ; ^x j
0
1
0
if x^ i 6 x^ j
if x^ i x^ j :
5:2
5.2 A Discrete Metric-Based Index
81
The value 1 assigned to the metric d0 indicates that we measure the distance
between individuals belonging to two different subgroups i and j by 1. On the
other hand, there is no distance between individuals belonging to the same subgroup
and

in this case the value of the metric is 0. Substituting the Euclidean metric
d ^
xi; ^
x j by the discrete metric in (5.1), we get a social polarization index of the
form
k X
k
X

i1 j :
P , K0
5:3
i1 j1

To demonstrate how P , leads to PRQ, Montalvo and Reynal-Querol (2008)
considered three properties, which are taken directly from Esteban and Ray (1994)
and are redefined in terms of sizes of the subgroups. The first property is:
Property 1 If there are three subgroups of sizes p, q, and r, and p > q and q r,
then if we merge the two smallest subgroups into a new subgroup with size e
q , the
new distribution is not less polarized than the original one. That is,
POLp; q; r POLp; e
q , with e
q q r , where POL stands for an index of
polarization.
This property demands that for three subgroups with relative frequencies p, q,
and r, where p > q and q r, polarization should not reduce if there is a merger of
the two smallest subgroups. It is social polarization counterpart to the EstebanRay
(1994)

Axioms 1 and 2. Montalvo and Reynal-Querol (2008) showed that the index
P , in (5.3) satisfies Property 1 if and only if 1.
As Montalvo and Reynal-Querol (2008) demonstrated the lower bound restriction, 1 is obtainable if Property 1 is replaced by Property 1b, whose formulation
does not rely on the assumption that the number of subgroups is three.
Property 1b Suppose there are two subgroups with sizes 1 and 2. Take any one
subgroup, say, the one with size 2, and split it into m 2 subgroups in such a way
that 1 e1 ei for all i 2, . . . , m 1, where e is the new vector of
m1
P
population shares, and clearly
ei 2 . Then the polarization for e is not higher
than that for 1 ; 2 .
i2
Chakravarty and Maharaj (2011b) considered the following minor variant of

Property 1b.
Property 1a Suppose there are two subgroups with sizes 1 and 2. Take any one
subgroup, say, the one with size 2, and split it into m 2 groups in such a way that
1 e1 ei for all i 2, . . . , m 1, with > for at least one i, where e is the new
m1
P
vector of population shares, and clearly
ei 2 . Then polarization for e is not
higher than that for 1 ; 2 .
i2
82
The following property, which bears similarity with Property 1b, has been
analyzed by Esteban and Ray (1999).
Property 1c Consider a distribution of the population across k 3 subgroups.
Suppose i 6 j for some i and j. Then polarization is increased by merging any of
the k 1 smallest subgroups into one. However, if the initial distribution of
population is uniform, then polarization is unchanged.
The next property considered by Montalvo and Reynal-Querol (2008) is based
on Axiom 3 of Esteban and Ray (1994). According to this property, if there are
three subgroups, of which two have the same population share, then polarization
should not decrease under shift of population mass from the subgroup with unequal
size equally to the other two subgroups. Formally,
Property 2 Assume that there are three subgroups of sizes p, q, and p. Then if we
shift mass from the subgroup with size q equally to the other two subgroups,
polarization does not decrease. That is, POLp; q; p POLp r, q 2r, p r ,
where 0 < r < q2 and POL stands for an indicator of polarization.
Montalvo and Reynal-Querol (2008) also demonstrated
that 1 is necessary

and sufficient for satisfaction of Property 2 by P , . Given 1, if we desire the
index tobe bounded
between 0 and 1, then we need to choose K 0 4. The resulting

index P , becomes the Reynal-Querol index PRQ, defined as

k X
k
k
X
X
X

0:5 i 2
2
2
PRQ 4
i j 4
i 1 i 1
i :
0:5
i1 j6i
i1
i1
5:4
PRQ achieves its minimum value, zero, if there is complete homogeneity in the
sense that all the individuals are concentrated in a particular social subgroup. That
is, minimum polarization corresponds to the perfectly homogeneous situation. On
the other hand, PRQ takes on the maximum value, 1, if there is an equal splitting of
the entire population into two subgroups. This is the situation of perfect bipolarity.
A polarization index satisfying this condition can be regarded as treating polariza
tion in terms of deviations from the extreme social distribution 12; 12; 0; . . . ; 0 (see
also Esteban and Ray 1994).
For i 1k for all i, PRQ becomes a decreasing function of k. This is intuitively
reasonable. Given the size of the population, as the number of subgroups increases,
the subgroups become weaker in terms of population shares and hence polarization
is likely to go down.
Another property of PRQ is zero-frequency independence, which says that for all
k 2 , 2 k , we have P 1 ; 2 ; . . . ; k P 1 ; 2 ; . . . ; k ; 0. According to this
property, if the population share of a subgroup added to the set of existing subgroups is zero, the significance of the existing subgroups does not get affected.
Consequently, polarization remains unaltered. Likewise, the deletion of a subgroup
with zero population share does not change the level of polarization.
5.3 Some Characterizations and a Generalized Index
The formulation 1
k

P
0:5 i 2
i1
0:5
83

i for PRQ is simply the difference between
unity and a weighted sum of population fractions, where the weights are squared
deviations of population shares of different subgroups from the population share 0.5
as a proportion of 0.5 itself. An increase in the number of subgroup or spreads of
sizes of the subgroups from 0.5 reduces PRQ. It is worth noting here that PRQ in (5.4)
is a positive multiple of the probability for any three randomly selected individuals
in the partitioned population; any two will belong to the same subgroup. To see this
explicitly, note that out of three randomly selected persons, the probability that two
will belong to the same subgroup i is 2i and the probability that the third person will
not belong to subgroup i is 1 i . Consequently, 2i 1 i is the probability
that one person will not belong to subgroup i, given that two other persons will
be in this subgroup. Four times the sum of these probabilities across all subgroups
k
P
4 2i 1 i gives us PRQ.
i1
5.3
Some Characterizations and a Generalized Index
A characterization of an index becomes helpful in understanding it in greater details

through the axioms used in the characterization. In this section we present some
characterizations of PRQ using several sets of intuitively reasonable axioms.
Given that k is the set of all discrete probability distributions of dimension k on
1
the real line R1, we write [ k for the set of all probability distributions. In
k2
order to characterize PRQ formally, we begin by specifying a formal definition of a

social polarization index.
Definition 5.1 By social polarization index, we mean a real-valued function
P defined on , that is, P : ! R1 , which
satisfies symmetry in its arguments
(i.e., for all k 2 , 2 k , P P , where is any permutation matrix of
order k) and which also satisfies zero-frequency independence (i.e., for all k 2 ,
2 k , P 1 ; 2 ; . . . ; k P 1 ; 2 ; . . . ; k ; 0).

For any k 2 , 1 ; 2 ; . . . ; k 2 k , P expresses the level of social
polarization associated with . Symmetry means that given the partitioning of the
population with respect to the social characteristic, in measuring polarization, we
need to be concerned only with the population shares of different subgroups.
Evidently, this postulate is satisfied by PRQ.
A social polarization index P : ! R1 will be called additive if P is of the
following form:
84
k
X
P
i ;
5:5
i1
where : 0; 1 ! R1 is continuous, k 2 , and 2 k are arbitrary. The functional

value pi may be interpreted as embodying the impact of subgroup i on overall
polarization. We refer to the function as the influence function corresponding to
the additive social polarization index P. An additive social polarization gets
uniquely determined by its influence function and hence it makes a social polarization index well defined. If we assume that i 2i 1 i , then the additive
index in (5.5) becomes a positive multiple of PRQ (see Chakravarty and Maharaj
2011b).
In order to present the characterization theorems formally, Chakravarty and
Maharaj (2011b) considered the following axioms for a general social polarization
index P, which were considered by Montalvo and Reynal-Querol (2005a, 2008):

Axiom 1 For all k 2 N, 2 k , 0 P 1.

Axiom 2 For all k 2 N, P 0 if 2 k is some permutation of (1,, 0, . . ., 0).

Axiom 3 For all k 2 , P 1 if 2 k is some permutation of 12; 12; 0; . . . ; 0 .
These axioms are the formal statements of the boundary condition and the
situations of minimum and maximum polarizations respectively for a general social
polarization index (see Montalvo and Reynal-Querol, 2005a, 2008) and, as
observed, they are satisfied by PRQ.
In order to specify the next axiom, we consider two subgroups, say 1 and 2, with
unequal population sizes and redistribute the total population proportion of these
two subgroups between themselves equally. One motivation for this type of redistribution can be complete elimination of employment segregation between the two
subgroups if the individuals belonging to these two groups are suitable for a specific
type of employment.
That is, analytically we consider the social distributions 1 ; 2 ; . . . ; k

0
2 1 2
2 k and 1
2 ; 2 ; 3 ; . . . ; k 2 k , where the latter arises from the
former after equal redistribution of the populations of subgroups 1 and 2 between
themselves. If we look at these two subgroups in isolation, then their initial
1
2
population shares are 1
and 1
respectively. After the redistribution, the
2
2
population share of each of them becomes 12. Under the additive formulation (5.6),
0

the resulting change in polarization P P depends only the population
shares of subgroups 1 and 2. Since we can treat this notion of marginal polarization
as well as an indicator of polarization, following Montalvo and Reynal-Querol
(2008), we
difference is dependent on a measure of distance
can say that the
1 1
1
2
between 1 2 ; 1 2 and 2; 2 .
5.3 Some Characterizations and a Generalized Index
85
0

While under additivity, the dependence of P P is well understood
because of the formulation; for a general social polarization index, let us consider a
situation where 1 2 is quite large (close to unity, say). The redistribution then

0
brings close to 12; 12; 0; . . . ; 0 , which in turn shows that polarization increases. For
small 1 2 , each of 1 and 2 will be small. In view of an observation made by
Esteban and Ray (2007), we can say that polarization is reduced under equalization
of populations. This indicates an explicit dependence of polarization difference on
1 2 .
For simplicity of exposition, we assume that the distance is measured by

2
2
1
2
1
1
square of the Euclidean distance, that is, by
2
2
1 2
1 2
2
1 j 1 2 j
2 1 2 2 .
Alternative possibilities definitely exist, but this is a very simple and
popular form. We can also use the positive square root of this form. But since the
former is obtained from the latter by taking a squared transformation, they are
ordinally equivalent and hence convey the same information essentially. Another
advantage of the former over the latter is that the former is continuously differentiable. Consequently, under constancy
0 of 1 2 , we can assume that the

polarization difference P P is proportional to d 2 j 1 2 j2 , the
0

square of the Euclidean distance. Formally, P P d2 , where , the
constant of proportionality, is dependent on 1 2 .
In view of the above discussion, we can state the following axiom, under
additivity:
0
Axiom 4 Consider the social distributions and defined above. Then, for any
additive polarization
index P : ! R1 of the form (5.5), the polarization difference
0

P P can be expressed as
0

P P 1 2 2 f 1 2 ;
5:6
where f : 0; 1 ! R1 is continuous.
PRQ satisfies this axiom under the particularization
f 1 2
2 3 1 2 .
We have noted that the under equal splitting of the population across
k subgroups, PRQ becomes a decreasing function of k. We rigorously stipulate
this for a general social polarization index in the following axiom.

Axiom 5 P 1k; 1k; 1k; . . . ; 1k becomes arbitrarily small for sufficiently large k.
It is possible to characterize PRQ using alternative combinations of Property 1,
Property 2, and Axioms 1, 2, 3, 4, and 5 (Chakravarty and Maharaj 2011b). This is
formally stated in the following theorem.
86
Theorem 5.1 Let P : ! R1 be an additive social polarization index of the form

(5.5), for which the influence function is twice continuously differentiable. Then
the following statements are equivalent:
(i)
(ii)
(iii)
(iv)
(v)
P satisfies Axioms 1, 2, 3, and 4 and Property 1.

P satisfies Axioms 1, 2, 3, and 4 and Property 1a.
P satisfies Axioms 2, 3, and 4 and Property 2.
P satisfies Axioms 2, 3, 4 and 5.
P is precisely the index PRQ.
Equivalence between statements (i) and (iv) of Theorem 5.1 means that PRQ is
the only social polarization index that satisfies Axioms 1, 2, 3, and 4 and Property 1.
Similar statement can be made for the sets of postulates specified in statements
(ii) and (iii).
Since the Theorem 5.1 is a demonstration of characterization of PRQ using
different sets of axioms, a natural investigation is to verify whether the axioms
employed in each of the four characterizations (iiv) are independent. Chakravarty
and Maharaj (2011a) demonstrated the following theorem in this context.
Theorem 5.2
(i)
(ii)
(iii)
(iv)
Axioms 1, 2, 3, and 4 and Property 1 are independent.

Axioms 1, 2, 3, and 4 and Property 1a are independent.
Axioms 2, 3, and 4 and Property 2 are independent.
Axioms 2, 3, 4 and 5 are independent.
The next theorem of Chakravarty and Maharaj (2011b) requires an axiom which
is specified under the assumption that there are only two subgroups. It relies on
Montalvo and Reynal-Querols (2005a, p. 799) argument that the proportional
contribution of the largest group in the index of polarization is larger. . .and the
reverse happens to the smallest group. In (5.5) if we assume that there are only two
subgroups, then i represents the contribution of subgroup i to the total
polarization and its proportional contribution is 2 i , where i 1, 2, 0 < 1 ,
P
j
j1
2 < 1, 1 2 1 : In view of Montalvo and Reynal-Querols argument, this

proportional contribution may be assumed to be directly proportional to the population proportion of the subgroup, that is, 2 i c i ; where c > 0 is a constant.
P
j
j1
This then gives

following:
1
2
1
2
; 0 < 1 , 2 < 1, 1 2 1. We can now state the
Axiom 6 If there are two subgroups with population proportions 1 and 2 ,

respectively, where 0 < 1 , 2 < 1, 1 2 1 : Then for any additive social
polarization index P : 2 ! R1 of the form (5.5), 12 12 :
Chakravarty and Maharaj (2011b) developed an alternative characterization that
does not make use of Properties 1, 1a, and 2 but uses Axiom 6. Formally,
5.4 Some Alternative Indices
87
Theorem 5.3 Let P : ! R1 be an additive social polarization index of the form

(5.5). Then P satisfies Axioms 2, 3, and 4 and Axiom 6 if and if it coincides with
PRQ.
Chakravarty and Maharaj (2011a) showed the axioms used in Theorem 5.3 are
independent. Another attractive feature of this characterization is that the influence
function is not assumed to be differentiable.
Given the simplicity of PRQ, there exists scope for its generalizations from
different perspectives. For instance, Chakravarty and Maharaj (2012) have recently
characterized the following generalization of PRQ:
k
X

4
i 2 1 i
PRQ
i1
i1 i2 i3 ;
5:7
1i1 <i2 <i3 k

where 2 0, 3 is an arbitrary constant, k 2 and 1 ; 2 ; . . . ; k 2 k are
arbitrary. While the first term of (5.7) is a positive multiple of the probability that
out of three randomly selected persons,
Xtwo will belong to one subgroup and the
i1 i2 i3 is a nonnegative multiple of
third to another, the second term
1i1 <i2 <i3 k
the probability that all the three individuals belong to three different subgroups. In a
perfectly homogeneous society, all the three individuals should belong to the same
social subgroup. Hence PRQ

linearly combines two nonnegative expressions, each
of which indicates an extent of social heterogeneity in the population. Conse
quently, PRQ
can be regarded as an index of social polarization. For 0, PRQ
coincides with PRQ. Hence, we can refer to PRQ

as a Generalized Reynal-Querol
index of order .
5.4
Some Alternative Indices
The discrete distance metric defined in (5.2) is an oversimplification of the distance

between two different social subgroups. Permanyer (2012) argued that social
polarization gets highly affected by the distance that exists between different subgroups and also by the sense of identity of an individual with the particular
subgroup of which he/she is a member. He suggested a modified formulation that
takes this into account in a more general way. Each individuals feeling of interest
that arises from his identification with his own subgroup is indicated by a degree of
radicalism denoted by a nonnegative real number v. The notion of radicalism
serves two purposes. First, it indicates the strength with which a person belonging to
a subgroup compares himself with persons in the other subgroups. It is an indicator
of the force with which an individual belonging to a subgroup will defend his
88
subgroup interest. Second, it also influences the difference he feels with other
persons in his own subgroup.
This gives rise to a general formulation where an individuals identification can
depend on the size of the individuals subgroup and/or on the extent of individual
radicalism. On the other hand, his/her feeling of alienation can be levels of
radicalism of individuals in the other subgroups and extents of radicalism of the
individuals in the same subgroup. Thus, this general structure shows some clear
deviations from the EstebanRay framework in the sense that in the latter there is
no scope for variations in a persons identification within the subgroup of which
he/she is a member. This divergence exists from the Reynal-Querol structure as
well. Another important difference of the Permanyer formulation from both these
structures is the role of members of the same subgroup in the context of alienation.
For each social subgroup i, there is a non-normalized density function fi that
measures how extents of radicalism are distributed within the subgroup, where the
support of fi is R1 , the nonnegative part of the real line. It is stipulated that an
individual, pertaining to subgroup i with radicalism v, possesses a feeling of
identification proportional to fi(v). We write f for the vector ( f1, f2, . . ., fk).
Permanyer (2012) then employed the identificationalienation framework to
define a general social polarization index initially under the assumption that there
is no alienation within a subgroup. Consequently, for individuals in subgroup i,
identity can be assumed to depend only on ni, the population size of the subgroup.
Alienation between any two subgroups with radicalism degrees v and v0 is taken as

0
0
an increasing function of the sum v v . The sum v v can be regarded as an
indicator of constriction between the two subgroups. Increasingness of the alien
0
ation function in v v reflects the view that as the individuals belonging to the
two subgroups increase their force for defending their subgroup interests, bitterness
between the subgroups will increase. This alternative formulation of the
identificationalienation structure is a parting from the DuclosEstebanRay
(2004) framework.
In view of the above discussion, the form of general social polarization index
advocated by Permanyer (2012) becomes
1 1
k X
k Z Z
X

0
0
0
H ni , v v f i vf j v dvdv ;
PP f
i1 j6i
5:8
where, as before, H is an antagonism function. Three axioms along with the

population homogeneity axiom, considered in Chap. 3, were imposed on this
general form to derive a more specific form of a polarization index.
According to the first axiom, suppose a particular subgroup is further subdivided
into two clusters, where the bigger cluster has a smaller level of radicalism, whereas
the lower clusters radicalism is higher. If the degrees of radicalism of the two
clusters are respectively increased and decreased equally, then within-group
5.4 Some Alternative Indices
89
f i (v )
Degree of radicalism
Fig. 5.1 Effect of within-group increase in radicalism on polarization
radicalism and polarization should increase. This within-group axiom shows that

0
the antagonism function H is concave in its second argument v v (Fig. 5.1).
The second axiom is a between-group axiom. It says that of two unequal subgroups, if the smaller one is becoming less radical and the larger one is becoming
more radical, then polarization should go up. Thus, the impact on polarization of an
increase in the radicalism of the bigger subgroup should be more than that of a
reduction in radicalism of the smaller subgroup. This axiom indicates convexity of

0
H in v v (Fig. 5.2).
The first two axioms of Permanyer (2012) are concerned respectively with
within-group and between-group alienation changes. They bear some similarity
with Axiom DER3 of Chap. 3. Permanyers (2012) third axiom requires that
polarization should not decrease under an equal movement of population from a
large subgroup to two equally smaller subgroups, where all the three subgroups
have the same normalized density. The equal population shift from the large
subgroup makes the relative forces of the subgroups more equal, which in turn
may generate higher tension. Consequently, polarization is unlikely to go down.
This axiom parallels Axiom ER3 of Chap. 3.
Permanyer (2012) showed that the only social polarization index of the form
(5.8) that satisfies these three axioms and the population homogeneity condition is a
positive multiple of
k X
k
X

PP f
i1 j i j ;
i1 j6i
5:9
90
Fig. 5.2 Effect of between-group increase in radicalism on polarization
where 2 0; 1 and i is the mean value of the distribution represented by fi. The
Permanyer index in (5.9) is a generalization of the polarization index P ,

defined in (5.3) and hence of PRQ as well. While in P , alienation between

subgroups is assumed to be constant, in its generalized form PP f , it is made
sensitive to the extent of radicalism
and resentment across subgroups. If is are the
same across subgroups, PP f coincides with a proportion of P , . The

parameter has the same interpretation as in the EstebanRay (1994) framework;
it is a polarization sensitivity parameter. If we impose the axiom that polarization
should decrease as the number of identical subgroups k (with relative frequency 1k),
then the lower bound of becomes 0.71 (Permanyer 2012). The reasoning behind
this axiom is that as the subgroups are becoming smaller, their members will be less
powerful which may lead to a reduction in social bitterness. Hence polarization is
expected to go down.
The index analyzed in (5.9) ignores within-group alienation. Alienation will
exist in a subgroup if members of the subgroup differ with respect to their
radicalisms. A more radical member may like to maintain distance from a member
with lower level of radicalism. Permanyer (2012) incorporated this view by assuming that alienation within
is measured by an increasing transformation of
a subgroup
0

radicalism distance v v , whereas, as before, alienation across subgroups is

0
increasingly related to the sum v v . Under this formulation, the social polarization index takes the form
5.5 A Social Polarization Ordering
91
1 1
k Z Z
X

0

0
0
PPE f
H f i v, v v f i vf i v dvdv
i1
1 1
k X
k Z Z

0
X
0
0
H f i v, v v f i vf j v dvdv :
i1 j6i
5:10
Permanyer (2012) invoked population homogeneity condition, axioms similar to

DER2 and DER3, and a new axiom concerning polarization change under a shift of
population mass across two initially identical subgroups, to axiomatize the following form of social polarization index:
PPE
1 1
k Z Z
X
0
0
0
f
f i1 vf i v v v dvdv
i1
1 1
k X
k Z Z
X
i1 j6i
0

0
0
f 1
vf j v v v dvdv ;
i
5:11
where is bounded between 12 and 1. The restriction on lower bound is guided again
by the requirement that polarization
should diminish as the number of identical

f
subgroups increases. PPE
is quite similar to the DuclosEstebanRay income
polarization index given by (3.5). However, the major difference between them is
that the latter does not take into consideration the notion of radicalism in the
formulation. In an empirical illustration, Permanyer (2012) provided a discussion
on estimation of religious radicalism.
5.5
A Social Polarization Ordering
In Sect. 5.3 several characterizations of the Reynal-Querol index and its generalizations using alternative sets of axioms have been presented. However, the ordering of two different social distributions by all these indices may not be the same.
Therefore, a natural investigation is to check whether two social distributions can
be ranked unambiguously by a class of social polarization indices satisfying certain
desirable postulates.
Chakravarty and Maharaj (2012) showed how some classes of Reynal-Querol
type polarization indices can be employed to rank alternative social distributions. In
this section we briefly analyze a simple ordering, suggested by these authors. For
this purpose, it is assumed that the given k social subgroups (k 3) are ordered
non-increasingly with respect to their population sizes. The set of all
non-increasingly ordered social profiles for k social subgroups is denoted by k .
92
Following Esteban and Ray (1994), it is also assumed throughout the section that
there is a small number of significantly sized groups.
The following property of a social polarization index is assumed for the purpose
at hand.
Property CI A rank-preserving population shift from subgroup i to subgroup
j (with j > i 2) cannot increase polarization.
Since an equalization of population mass across subgroups reduces population
concentration, given that the most dominant subgroup, subgroup 1, remains unaffected by the shift, polarization is likely to go down. This is also supported by
Esteban and Rays (2008b; p. 175) argument that if the two groups are small
enough, the equalization of population will reduce the intensity of conflict. The
Reynal-Querol index PRQ satisfies this property. This follows from the observation
that i i1 23 since each of i and i1 can be at most 13. If 1 > 13, we have
1 i i1 > 1, a contradiction.
The next property is specified in terms of the population shift between the first
two groups.
Property CII A rank-preserving population shift from subgroup 1 to subgroup
2 cannot decrease polarization.
If the sizes of other subgroups are not significantly large, social polarization will
be maximized when sizes of the two largest groups come close to one another
(Montalvo and Reynal-Querol 2005a; p. 798). Given that we have a small number
of significantly sized groups, an equalization of the sizes of the two largest subgroups makes them adjacent in terms of population size so that polarization is
expected to go up. If both subgroups are larger than 13, this again can be supported by
the Esteban and Ray (2008b, p. 175) observation that a transfer of population from
a group to a smaller one increases intensity of conflict. The index PRQ verifies this
property if 1 2 23. In particular, PRQ satisfies it for k 3.
The following definition formalizes the notion of social polarization ordering in
a specific framework.
Definition 5.2 For two social distributions
1 ; 2 ; . . . ; k ,
0 0
0
0
0
0

1 ; 2 ; . . . ; k 2 k , we say that is at most as polarized as , (P ,
for brevity), if for all social polarization index P : k ! R1 satisfying Axioms 1, 2,
0

and 3 and Properties CI and CII, we have P P .
An example of a social polarization
index that verifies Axioms 1, 2, and 3 and

Properties CI and CII is P0 4 1 2 .
The following theorem specifies some simple necessary conditions for P
to hold. (See Chakravarty and Maharaj (2011b), for a detailed analysis.)
0
Theorem 5.4 Let , 2 k be arbitrary. Then a set of necessary conditions for

0
P to hold is given by:
2 2 and
2 2
l
X
93
!
2 2
i3
l
X
!
0
i for all 3 l k:
5:12
i3
The necessary conditions given by (5.12) involving k 1 inequalities, stated in

terms of population proportions of the two distributions, can be checked very easily.
0
0
Of these the last inequality implies that 1 2 1 2 . However, these
0
necessary conditions are by no means sufficient for P to hold. To demonstrate
this explicitly, let 0:4; 0:3; 0:3 and e 0:34; 0:34; 0:32. Then all the
inequalities specified in (5.12) are satisfied. But observe that 1 2 > e1 e2 so that

for the social polarization index P0 4 1 2 , we have P0 > P0 e . This in
turn shows that P e does not hold.
Often it becomes useful to search for an easily verifiable sufficient condition for
P to be fulfilled. Such a condition can be stipulated using the population concentration curve of a social distribution. The population concentration curve of a social
distribution is the graph of the cumulative population shares against the cumulative
number of subgroups, given that the subgroups are ranked from
the
largest to the
smallest. Formally, for any 2 k , its concentration curve CC ; l is obtained by
l

P
i for all l 0, 1, 2, . . . , k, where it is assumed that CC ; 0 0. The
plotting
i0

assumption CC ; 0 0 ensures that the concentration curve is a closed graph. As

l increases from 0 to k, CC ; l increases. The curve is concave as well. In the case
of complete concentration, that is, when only one subgroup has the entire population fraction 1, the curve will be of the shape OAB in Fig. 5.3. If there is minimal
concentration, that is, when population fractions are equally distributed across
subgroups, the curve coincides with the diagonal line OB of the figure.
0
0
Given two social distributions , 2 k , we say that concentration curve
dominates if the concentration curve of the former lies nowhere below that of the
l
l
P
P
0
i
i for all 0 l k 1. However, the concentration
latter, that is,
i0
i0
curve dominance relation is not complete, although it is transitive. In Fig. 5.3, the
solid curve OC is the concentration curve for a two-subgroup population and the
dashed curve is a concentration curve for a three-subgroup population. Clearly, the
former dominates the latter.
In the following theorem of Chakravarty and Maharaj (2012) a set of sufficient
conditions are provided for satisfaction of P .
0
Theorem 5.5 For arbitrary , 2 k , the following inequalities provide a set of

0
sufficient conditions for P to hold:
94
A
Cumulative population proportions
1
2
3
Cumulative number of subgroups
Fig. 5.3 Population concentration curve
1 1 and
l
X
i0
i
l
X
i for all 2 l k 1:
5:13
i0
According to Theorem 5.4, if the population share of the largest ethnic group in the
0
profile is at least as large as that in the profile and the population concentration
0

0
0
0
0
curve of 1 2 , 3 , . . . , k dominates that of 1 2 , 3 , . . . , k , then is
regarded at least as polarized as by all social polarization indices that satisfy Axioms
1, 2, and 3 and Properties CI and CII. The conditions clearly indicate that even the
simple concentration curve dominance relation between the partially merged profiles
0
0
0
0
0
1 2 , 3 , . . . , k and 1 2 , 3 , . . . , k cannot be sufficient for P .
This should as well become evident from the observations that while maximum
concentration corresponds to minimum polarization, no such relation exists between
maximum polarization and minimum concentration. Since the inequality system laid
down in the theorem is very easy to check, the sufficient condition (5.13) can be
quickly implemented. Consequently, the sufficient condition may be treated as a
useful tool for ranking alternative distributions by ordering P .
Toverify that the set of conditions
laid down in (5.13) is not necessary, let

1 12; 12; 0 and 2 13; 13; 13 2 3 . For these two profiles, condition (5.13) is
violated, but by definition we have 2 P 1 .
The ordering P defined on is incomplete. Since a necessary condition for
0
satisfaction of the ordering P is specified by the system (5.13), incompleteness
of the ordering follows as a consequence of violation of one or more of the k 1
conditions. Given two social distributions, there is no guarantee for fulfillment of
these conditions and the ordering becomes incomplete.
95
Given , 0 2 k , in order to verify the ranking relation P 0 , we need to check

at the outset the conditions laid down in (5.12). If one or more of them are violated,
the satisfaction of P 0 is ruled out. Once they are satisfied, we move for
verification of the conditions in (5.13). If they hold as well, the profiles under
comparison can be ranked by P . In of one or more of them are not fulfilled, no
specific conclusions can be drawn about the ranking of the profiles.
Given two different social characteristics, say religion and language, often we
may be interested in knowing whether a society is more polarized with respect to
one than the other. Since a population concentration curve satisfies zero-frequency
independence, we can extend the corresponding dominance relation to the case
where the number of social subgroups is a variable. That is, suppose 2 k and
0 2 q , where q < k, say. Then we can add k q social subgroups, each with
zero frequency, to the profile 0 . We denote this extended profile by e0 so that both
and e0 now have the same number of subgroups. The concentration curve of e0
coincides with that of 0 so that P 0 is same as P e0 . Since the social
polarization indices considered under P are insensitive
to the addition or deletion

of subgroups with zero frequencies, Pe0 P 0 . Thus, we can use Theorems
5.4 and 5.5 to compare the profiles and e0 , or, more precisely, the profiles and 0 .
Chapter 6
Measuring Polarization for a Dimension

of Human Well-being with Ordinal
Significance
6.1
Introduction
In Chaps. 1, 2, 3, and 4, we have considered income as the only dimension of human

well-being. But the well-being of an individual depends as well on many dimensions other than income. Examples of such dimensions are health, literacy, environment, social cohesion, housing, and so on. Now, income is a ratio scale variable
in the sense that it has a natural zero and is unique up to a positive scale transformation. But some of the dimensions like health and literacy are measurable on
ordinal scale.
An ordinal variable has similarity with a categorical variable like color, race,
gender, and religion, possessing one or more categories or types. However, for a
variable, measurable on ordinal scale, a well-defined ordering rule has to be
followed. The ordering should remain invariant under increasing transformations.
For instance, in a self-assessed health survey, an individual can be asked to choose
one of the five categories that best describes his/her health status, namely, poor,
fair, good, very good, and excellent. Numerical values are assigned to
these categories in a way consistent with the ordering of preferences, more precisely, in an increasing order so that a higher number is assigned to a better
category. This assignment of numerical values to different categories is arbitrary.
Any increasing transformation of these numbers preserves the ordering of the
categories (see Allison and Foster 2004). Another example can arise in the context
of literacy ordering. Here we consider the problem of ordering of achievement
levels of education of individuals in a community commencing from zero education
(illiteracy) to university education by designating numbers in an increasing manner
to education levels (see Chakravarty and Zoli 2012).
The objective of this chapter is to analyze alternative approaches to the measurement of polarization for a variable measurable on ordinal scale. The basic
Springer India 2015

97
98
6 Measuring Polarization for a Dimension of Human Well-being with Ordinal. . .
ingredients for a polarization index for an ordinally measurable dimension are the
proportions of persons that belong to different categories corresponding to the
dimension. A clear distinction has to be made between the bipolarization axioms
for a ratio scale variable and an ordinal variable. To understand this explicitly,
consider the increased bipolarity axiom, analyzed in Chap. 2. According to this
axiom, a progressive transfer of income between two individuals on either side of
the median income increases bipolarization. However, in the current context, it is
difficult to imagine a meaningful counterpart to a progressive transfer. Transfer of
literacy from a more literate person to a less literate person or transfer of health
from someone with a good health status to a less healthy person is not a meaningful
concept.
Allison and Foster (2004) suggested a median-based ordinal approach for
calculating health inequality using self-assessed health data. In this article inequality is treated as a spread away from the median category. It demonstrates the
usefulness of first-order stochastic dominance for ranking two or more population
health statuses. Chakravarty and Zoli (2012) clearly indicated the role of integer
variable-based second-order stochastic dominance in this context (see also
mbrosio 2006). Apouey (2007) made the first attempt to
Chakravarty and DA
measure health bipolarization for self-assessed health data, where the median
category has been employed as the reference point. She suggested several intuitively reasonable axioms for a health bipolarization index and proposed a class of
mbrosio and Permanyer (2013) suggested
indices. In a recent contribution, DA
social polarization indices using the identificationalienation approach, where the
dimension of interest is measured on ordinal scale and the population is partitioned
into two or more subgroups using a characteristic like religion, ethnicity, etc.
In the next section of the chapter, we present the axioms for an index of
bipolarization for a dimension of human well-being with ordinal characteristic.
Section 6.3 discusses some relevant bipolarization indices. Section 6.4 analyzes the
indices based on the identificationalienation approach. Finally, Sect. 6.5 deals with
some quasi-orderings in the bipolarization framework.
6.2
Axioms for an Index of Bipolarization for a Dimension

Measurable on Ordinal Scale
For the population under consideration, let there be q categories c1, c2, . . ., cq,
arranged in ascending order, of some ordinal variable, where q 2 N, the set of
positive integers, is arbitrary. For instance, in the case of self-assessed health status,
c1 represents poor, c2 corresponds to fair, and so on. Now, given that ci1 stands
for a better category than ci, i 1, 2, 3, 4, we can assign the numbers (1, 2, 3, 4, 5) to
these categories. If we consider the increasing transformation f t t2 , where
6.2 Axioms for an Index of Bipolarization for a Dimension Measurable. . .
99
t 2 f1; 2; 3; 4; 5g, then the assigned numbers change to (1, 4, 9, 16, 25). Evidently,
these newly allotted numbers keep the ranks of the categories unchanged. As we
will see later, for stating some of our axioms, we need that q should be at least
3. We, therefore, assume at the outset that q 3.
Let i stand for the proportion of individuals in ci. Therefore, 0 i 1,
q
P
i 1, where
q2N
is arbitrary. The vector
1 i q, and

i1
1 ; 2 ; . . . ; q represents a discrete probability distribution, which we can
refer to as ordinal distribution. Let q stand for the set of all (ordered) discrete
probability distributions of dimension q and for the set of all probability distri1
butions, that is, [ q . The cumulative proportion of the population which

q3
belongs to the category c and lower ones in the distribution is defined as

c
P
Fc
i . For any q 2 N, 2 q , we denote the corresponding cumulative
i1

population share vector F1 ; F2 ; . . . ; Fq1 ; 1 by F . We write m for the median
category so that for any 2 , we have Fm1 < 12 and Fm 12. To illustrate this, let
the proportions of persons in five health categories considered above be
0:2, 0:2, 0:1, 0:3, 0:2. Then Fpoor 0:2, Ffair 0:4, Fgood 0:5, and so
on. The median category here is good. Since the ordering of the categories
remains invariant under any increasing transformation applied to the categories,
the position of the median category also remains invariant under such a
transformation.

As we have noted, corresponding to each 1 ; 2 ; . . . ; q 2 q , there is a

vector of cumulative population proportions F F1 ; F2 ; . . . ; Fq1 ; 1 . This F is

an element of
x1 ; x2 ; . . . ; xq 2 q 0 x1 x2 . . . xq 1 . More preq

cisely, given any 1 ; 2 ; . . . ; q 2 q , we have a vector of cumulative

population proportions F F1 ; F2 ; . . . ; Fq1 ; 1 2

q (and vice versa). Let
1
[
q.
q3
By a bipolarization index for a dimension measurable on ordinal scale, we mean

a continuous real-valued function PO defined on , that is, PO : ! R1 .
Continuity assumption is made to ensure that we do not observe any drastic change
in the value of PO under small changes in population proportions in categories. For
any F 2 , PO F is an indicator of the extent of bipolarization corresponding to
F .
Essential to the concept of bipolarization for a variable possessing ordinal
significance are increased spread (INS) and increased bipolarity (INB). As in
Chap. 2, the former is a monotonicity condition, which demands that bipolarization
increases under an expansion of the distribution away from the median category.
To
understand
this,
let
F 0:2, 0:4, 0:5, 0:8 , 1:0.
Define
vector
of
population
proportions
F 0:3, 0:4, 0:5, 0:8 , 1:0. The
100
0:3, 0:1, 0:1, 0:3, 0:2 associated with F is obtained from the vector
0:2, 0:2, 0:1, 0:3, 0:2 that matches F by reducing population proportion in
category 2 by 0.1 and increasing the corresponding proportion in category 1 by
0.1. Likewise, if we derive ^ from by reducing the population proportion in the
fourth category by 0.1 and increasing the proportion in the fifth category by 0.1,
then ^ 0:2, 0:2 , 0:1, 0:2, 0:3. The corresponding vector of cumulative popu^
lation proportions is F 0:2, 0:4, 0:5, 0:7 , 1:0. Thus, while in the former
situation, we increase the population proportion in the poor category at the
cost of a reduction in the proportion in a better category fair; in the latter
situation we increase the proportion in the category excellent by reducing the
^
proportion in a worse category very good. We obtain both F and F from F by
stretching the distribution away from the median category. In each case of
stretching, there is now widening of the distribution and such changes should
increase bipolarization. The axiom INS demands this. The statement of the axiom
relies on the following definition.
Definition 6.1 For all q 2 N, 2 q , 2 q is obtained from by a median
preserving spread if and have the same median category m; Fc Fc for all c < m
and Fc Fc for all c m, with at least one category c < m such that Fc > Fc or one
category c m such that
Fc
<
Fc .
Apouey (2007) took the following form of INS from Allison and Foster (2004).
Increased Spread (INS) For all q 2 N, 2 q , if 2 q is obtained from by a
median preserving spread, then PO F < PO F .

We illustrate the postulate INS graphically in Fig. 6.1. The health status of a
proportion of persons in a population reduces from the fair category to the poor
category and that of the entire population share in the very good category
improves to the excellent category. The original and the changed distributions
of the health status of the population are shown respectively in panels (a) and (b) of
the figure. The initials in the figure represent poor (P), fair (F), good (G), very good
(VG) and excellent (E). The two movements on both sides of the median category
(good) are median preserving shifts away from the median. INS demands that
polarization is higher in situation (b) than in situation (a).
Increased bipolarity is a clumping principle. Formal statement of this criterion
relies on the following two definitions.
Definition 6.2 For all q 2 N, given 2 q , we say that the distribution
0 0

0
0
1 ; 2 ; . . . ; q is obtained from through a transfer if for two given
categories c and c such that c^ 1 c 1, there are shifts of population mass of
0
size from c to c 1 and from c to c^ 1, where c , c^ , and i i for all
0
0
0
i 6 c^ , c^ 1, c, c 1 ; that is, c^ 1 c^ 1 , c^ c^ , c c ,
Population proportions
101
G
VG
Health categories
F
G
VG
Health categories
Fig. 6.1 Spread away from the median category for an ordinal variable
c1 c1 . In terms of distribution functions, Fi Fi for all i 6 c, c^ 1 ;
Fc Fc and Fc^ 1 Fc^ 1 .

To
understand
Definition
6.2,
we
consider
the
distribution
0:1, 0:2, 0:1, 0:05, 0:05, 0:1, 0:1, 0:1 , 0:2of population shares in 9 categories
associated with an ordinal dimension. The median category is the fifth category
from the left. Let c be the fourth category, that is, c^ 4. Assume also that c 1
^ 0:09, 0:21, 0:11, 0:04, 0:05, 0:1, 0:1, 0:1 , 0:2 is
and 0:01. Then
obtained from through a transfer. We can as well say that we get
F^ 0:09 , 0:3 , 0:41, 0:45, 0:5, 0:6, 0:7, 0:8, 1 from F 0:1, 0:3, 0:4, 0:45,
0:5, 0:6, 0:7, 0:8, 1 through a transfer. Since c^ < m and c^ 1 c 1,
this transfer is below the median category. Note that here we have Fi Fi
for all i 6 c, c^ 1 ; Fc Fc and Fc^ 1 Fc^ 1 . Likewise, we derive
e 0:1; 0:2; 0:1; 0:05; 0:05; 0:09; 0:11; 0:11; 0:19 from , that is, Fe
from F
through a transfer not below the median category. In this case, c 6 > m.
Formally,
1

Definition 6.3 Given any two distributions
F1 F1 ; F12 ; . . . ; F1q1 ; 1 ;

F2 F21 ; F22 ; . . . ; F2q1 ; 1 2
n with the same median category m, we say that
F2 is obtained from F1 by a transfer below the median category if and only if the
higher category c^ involved in the transfer does not exceed the median category, that
is, c^ m. Similarly, we say that the transfer is not below the median category if and
only if the lower category c involved in the transfer does not fall below the median
category, that is, c m.
102
For validity of this definition, we require the restriction that q 3.

To get the intuitive meaning underlying a transfer, note that a transfer takes place
through a shift of population mass from a high category to its immediate lower
category and from a low category to its immediate higher category such that the
lower category in the former is not below the higher category in the latter. This is
irrespective of whether the transfer takes place below or not below the median
category. Evidently, a transfer makes the distribution of population more withingroup homogeneous in terms of higher closeness of the individuals with respect to
their positions in the scales of the dimension. (The two subgroups considered here
are defined using the median category as the line of demarcation and the first
subgroup may consist of categories below the median category and the remaining
categories are in the second subgroup.) This increase in homogeneity should
increase polarization. Suppose the health status of a tiny population fraction
improves from good to very good and that of a population proportion of the
same size reduces from excellent to very good. Increase in homogeneity within
the subgroup, consisting of categories not below the median, should increase
polarization. The impact on polarization of the upward move from good to
very good should be more than the impact of the downward move from excellent to very good. In other words, the former shift should get more importance
than the latter one in the aggregation.
A variant of Apoueys (2007) increased bipolarity axiom can now be stated
formally as:
Increased Bipolarity (INB) For all n 2 N, 2 q , if F is transformed into F ,

where and have the same median category m; by at least one transfer below the
median category and/or at least one transfer not below the median category, then
PO F < PO F .
Figure 6.2 shows an increased bipolarity movement not above the median for a
hypothetical health status distribution of a population. An upward shift in the health
status of a population mass from the poor category to the fair category is
accompanied by a downward shift in the health condition of the same population
size from the good category to the fair category. The original and the new
distributions of the health status of the population are shown respectively in panels
(a) and (b) of the figure. INB requires that polarization is higher in the new
destruction than in the original one. (Note that here we have bipartitioned the set
of categories into the subgroups categories not above the median and categories
above the median. Evidently, the analysis applies equally well to the alternative
bipartitioning categories below the median and categories not below the
median.)
The next axiom suggested by Apouey (2007) is minimum bipolarization. It
specifies the minimum value of PO.
103
VG
G
F
Health categories
VG
Health categories
Fig. 6.2 Increased bipolarity for an ordinal variable
Minimum Bipolarization (MIB) For all

q 2 N,
PO 1, 1, 1, . . . 1, 1
PO 0; 1; 1; . . . ; 1; 1 PO 0; 0; 1; . . . ; 1; 1 PO 0; 0; 0; . . . ; 0; 1 0, the
minimum value of PO.
This axiom says that PO is bounded below by 0 and this bound is achieved if the
distribution of the population across categories is completely homogeneous in the
sense that all the individuals belong to a particular category. It treats the categories
symmetrically in the sense that irrespective of the status of the category (e.g., bad or
good); it assigns the value 0 to the bipolarization index whenever there is complete
concentration of population in the category. Thus, in the case of self-assessed health
dimension, the bipolarization index takes on the value 0 if all the individuals belong
to the category poor. Likewise, if the health status of all the persons is excellent, then also the level of bipolarization is 0.
Chakravarty and Maharaj (2014) proposed the following stronger version
of MIN.
Strong Homogeneity (STH) For all q 2 N, PO F takes on its minimum value
0 if and only if F is anyone of the forms (1, 1, 1, . . . 1,, 1) , (0, 1, 1, . . ., 1, 1) ,
(0, 0, 1, . . ., 1, 1) , . . ., (0, 0, 0, . . ., 0, 1).
The next axiom proposed by Apouey (2007) is Maximum Bipolarization.

Maximum Bipolarization (MAB) For all q 2 N, PO 12, 12, 12, . . . 12, 1 1, the
maximum value of PO.
This axiom demands that PO takes on the maximal value 1 under an equal
splitting of the entire population into two extreme categories, the lowest and the
104
highest, that is, the concentration of the entire population is the same in the two
extreme categories.
The following stronger version of MAB was suggested by Chakravarty and
Maharaj (2014).
Strong
Bipolarity
(STB) PO F takes on its maximal value 1 if and only if F
1

1
2; . . . ; 2; 1 .
Thus, while STH demands that any one of the cumulative population share
vectors (1, 1, 1, . . . 1,, 1) , (0, 1, 1, . . ., 1, 1) , (0, 0, 1, . . ., 1, 1) , . . ., (0, 0, 0, . . ., 0, 1) is
necessary and
sufficientfor PO to be 0, according to STB the cumulative population
share vector 12; . . . ; 12; 1 is necessary and sufficient for PO to take on the value 1. In
contrast, MIB and MAB specify only the sufficient conditions for PO to attain its
lower and upper bounds, respectively.
6.3
Some Indices of Bipolarization for a Dimension

of Well-being with Ordinal Information
In this section we analyze some bipolarization indices that are meant to be used for
dimensions with ordinal significance. We begin our discussion with Apoueys
index, which is defined as

q1
2 X
1
POA F 1
Fc ;
2
q 1 c1
6:1
where 2 0; 1 is a parameter that casts the weight ascribed to the median category.
As ! 0, the relative contribution of the median category increases, whereas the
relative contribution of the other categories decrease. The opposite happens as
! 1. POA can be interpreted as a measure of distance of the actual distribution
from the extreme bipolar position where 50 % of the population is in the lowest
category and the remaining 50 % is in the highest category. This index fulfills
the axioms INS, INB, MIB, and STB, the stronger version of MAB. The index
remains unchanged if the categories are arranged in the reverse order. For
instance, for self-evaluated health data, if we set excellent instead of poor
in category 1 and poor instead of excellent in category
5, and so on,
then the value of POA does not change. Formally, POA F1 , F2 , . . . , Fq1 , 1

POA 1 Fq1 , 1 Fq2 , . . . , 1 F1 , 1 .
While Apouyes (2007) index is concerned with the shift of the actual situation
from the situation of maximum bipolarization, an alternative possibility is to
consider the gap between the actual distribution and (0, 0, 0,..0, 1,. . . ., 1, 1), a
distribution that corresponds to the case of minimum bipolarization, where the
6.4 The IdentificationAlienation Approach to Polarization Measurement. . .
105
first 1 in (0, 0, 0,..0, 1,. . . ., 1, 1) appears at the mth place. Chakravarty and Maharaj
(2014) suggested the following functional form for a bipolarization index that takes
this into account:
POC F
m1
X
bi mFi
i1
q1
X
bi mj1 Fi j;
6:2
im
where the terms of the median category-dependent sequence {bi(m)} are

the weights designated to different categories. We now look at the
similarity between POC(F) and the Weymark (1981) generalized Gini index of
n
P
1
n
inequality I WG x 1
n
P n i1 ai x xi for an n person society,
x ai
i1
where the population-size-dependent positive weight ani is decreasing in i (see

Eq. (1.5)). Thus, while IWG looks at the deviation between the perfectly equal
distribution where everybody enjoys the mean income and the actual income
positions of the individuals, as we have noted, POC is also based on a deviation.
Furthermore, POC follows an aggregation similar to IWG. In view of this, POC may
be referred to as the generalized Gini index of polarization for an ordinal variable. A
sufficient condition for POC to verify INS and MIB is that bi m > 0 for all
1 i q 1. Under this condition POC verifies STH as well. Further, INB
holds if the weights are increasing below the median category and decreasing
from the median category onwards. POC becomes unambiguously bounded between
q1
P
bi m 2 and equality holds if it satisfies MAB (see Chakravarty and
0 and 1 if
i1
Maharaj 2014). Examples of the weights for which INS, INB, and the boundedness
condition are satisfied by the ChakravartyMaharaj (2014) index in (6.2) are bi m
2qi1
, i m.
2i1
q2 , i < m and bi m
q2
6.4
The IdentificationAlienation Approach to Polarization

Measurement for a Dimension with Ordinal
Representation
In Sects. 6.2 and 6.3, we have assumed that individuals in the entire population are
not distinguished by anything other than their association with the categories of the
dimension measured on ordinal scale. Often it may be necessary to look at polarization for an ordinal variable when the population is partitioned into two or more
subgroups by some characteristic. For instance, in a society consisting of Blacks
and Whites, the situation where all the Blacks have poor health and all the Whites
Fig. 6.3 Health

distributions across Blacks
and Whites-I
Population subgroups
106
Fig. 6.4 Health status

distribution across Blacks
and Whites-II
Population subgroups
G
VG
Health categories
G
Health categories
VG
have excellent health appears to be characterized by more polarization and more

tension than the situation where half the Whites and half the Blacks have poor
health and the other halves have excellent health. These two situations, which we
refer to as I and II, are depicted graphically in Figs. 6.3 and 6.4, respectively.
Therefore, for a deeper analysis we may be interested in knowing subgroup
polarizations and polarization of the entire population.
The initials in the two figures stand for poor (P), fair (F), good (G), very good
(VG), excellent (E), Blacks (B), and Whites (W).
mbrosio and Permanyer (2013) investigated this issue of measuring polariDA
zation for a variable with ordinal significance, given that the population has been
partitioned into k 2 social subgroups. They adopted the identificationalienation
approach to address the problem. As before, the ordinal variable is represented by
q categories. We write i c for the share of population in subgroup i for category c.
Identification of persons in a subgroup depends directly on the size of the subgroup
to which they belong. Alienation is assumed to be the same for all members
belonging to a particular subgroup.
Let
ij
k
X
c1

min i c ; j c
6:3
6.5 Bipolarization Orderings for an Ordinal Dimension
107
be the extent of overlap between subgroups i and j. This overlap coefficient is

bounded between 0 and 1, where the lower bound is achieved where the subgroups
are disjoint. On the other hand, ij becomes 1 if there is complete overlap. Similarity
of the subgroups is increasing with respect to the degree of overlap. Alienation is
defined by 1 ij , which takes on the values 1 and 0 when the subgroups are
completely disjoint and perfectly overlapping respectively. Thus, in this case for
capturing alienation, what matters is the clustering of subgroups in different
categories, not the distances between subgroups.
mbrosio and Permanyer (2013) axiomatically characterized the following
DA
form of the polarization index
k X
k

X
POA ,
i1 j 1 ij ;
6:4
i1 j1
where now i denotes the proportion of individuals in subgroup i, is the vector of

ij values and 2 0:71 , 1 is the polarization sensitivity parameter. The
mbrosioPermanyer index in (6.4) is a socioeconomic polarization index for a
DA
dimension with ordinal significance.
6.5
Bipolarization Orderings for an Ordinal Dimension
While different bipolarization indices can rank alternative distributions in opposite

directions, an ordering of the distributions may enable us to judge whether one
distribution turns out to be more polarized than another by all polarization indices
satisfying certain properties. From Definition 6.1 and the definition of first-order
stochastic dominance, it follows that for the distributions and , with the same
median category m, the statement that is obtained from by at least one median
preserving spread is equivalent to exactly one of the following conditions : (i)
first-order stochastic dominates below the median category and for the
median category and above (when the spread is only below the median category),
(ii) first-order stochastic dominates for the median category and above and
below the median category (when the spread is only not below the median category), and (iii) first-order stochastic dominates below the median category and
first-order stochastic dominates for the median category and above (when there
are spreads both below and not below the median category). In each of these
3 cases, PO F < PO F for all ordinal polarization indices PO that increase

under a median preserving spread, that is, for all PO that satisfy INS. The converse
is true as well, that is, if PO F < PO F for all ordinal polarization indices PO
that satisfy INS, then is obtained from by a sequence of median preserving
spreads. First-order stochastic dominance can be verified easily graphically by
108
plotting cumulative proportions of the populations for the two distributions, that is,
Fc and Fc values for different categories. As stated earlier, this relation is

incomplete, not all pairs of distributions can be compared using this criterion.
Since INS and INB are inherent to the idea of bipolarization, it becomes
worthwhile to look at the possibility of existence of a partial ordering that can
rank all distributions by indices satisfying these two postulates. Chakravarty and
Maharaj (2014) developed a partial ordering of this type using indices from the
generalized Gini family (6.2). They demonstrated that for two distributions F and
G, having the same median category (m), PF PG holds for all generalized
Gini bipolarization indices of the form (6.2) that satisfies INS, INB, and boundedness between 0 and 1 if and only if we have Fm1 Gm1 , Fm1 Fm2 Gm1
Gm2 , . . . ,
q1
P
im
m1
P
i1
Fi
m1
P
i1
Gi ; Fm Gm , Fm Fm1 Gm Gm1 , . . . ,
q1
P
Fi
im
Gi . The q 1 inequalities involving sums of cumulative population pro-
portions at different categories for the two distributions can be checked very easily
and hence their satisfaction enables us to conclude that one distribution has more
ordinal bipolarization than the other for a subfamily of the family (6.2). Note that
there is no restriction on the number of sequence of weights that can be used in (6.2)
such that INS, INB, and boundedness are verified. Therefore, the identified subfamily of indices is rather large. Since there may be violation of one or more of the
q 1 inequalities involving the distribution functions, the underlying ordering is
a quasi-ordering; it is transitive but not complete.
Chapter 7
Fractionalization, Polarization, and Conflict
7.1
Introduction
Loosely speaking, conflict refers to a situation of continued confrontation. Analysis

of conflict can provide a fruitful insight into the understanding of social and
economic relations. Class conflicts play an important role in the system of classical
political economy framed by Adam Smith (1933). A brief discussion on the theory
of class relations advocated by classical economists will give us an idea about the
nature of the underlying notion of class conflict.
Classical economists partition the society into three classes: capitalists who own
the stock, landowners, and laborers. They investigate the problem of class relations
from a static and also a dynamic perspective. The static approach regards the
problem as one of output sharing among the classes. Adam Smith developed a
theory of bargain that indicates a comparatively disadvantageous position for the
laborers in the context of sharing of their output. The capitalists, being the owners
of stocks, become successful in shooting down the wages to a subsistence level.
That is, the capitalists dominate the labor market. Likewise, land ownership provides landlord the privilege to rent. More rigorous demonstration of capitallabor
relationship was offered by Ricardo (1933). According to the Ricardian theory,
there is in general a decreasing relationship between wages and profits.
From the dynamic perspective, the problem involves sharing of growing output,
aftereffect of accumulation, and division of labor. Ricardos theory of distribution
concentrates largely on a scrutiny of this issue. In this analysis, there is a clear focus
on the conflicts of interest between classes. A similar theory was developed by
Adam Smith. But there are important differences between the two notions. In the
Ricardian setup, increased rent and reduced profits are a consequence of miserliness
of nature. This is intensified by accumulation and conflict between capitalists and
landlords increases. In this system the interest of the landlord and general interest of
the society are reverse in nature. In the framework considered by Smith, the interest
of the landlords is described as societys interest. Here the increased propensity for
Springer India 2015
109
110
7 Fractionalization, Polarization, and Conflict
reduced rate of profit is regarded as a consequence of higher competition. The two

systems vary on wages as well.
According to both John Stuart Mill (1920) and Karl Marx (1958), annexation
and clash among classes were instigated by the institution of private property.
Mills analysis of class relations was derived from the bargain theory of Smith,
and it also took into account the possibilities of labor as a social effort. Mills
conjecture on the possible long run implication of conflict was that ultimately the
employer class would give up.
The theory of class conflict was transformed into a tunnel of rebellion by Karl
Marx. He identified a positive rate of profit with a positive rate of delinquency. He
believed that the transition to a nonexistent employeremployee relation would be
underlined by a violent split, whereas John Stuart Mills belief was that the
changeover would be step-by-step. Overall, the classical economists were
concerned with the increase in wealth in the society and its distribution across
class subgroups (Robbins 1939).1
The conventional view that income or wealth inequality is a major source of
social conflict has been influenced substantially by the paradigm advocated by the
classical economists, particularly by Marx. Sen (1973) also argued that rebellion
has a close relation with inequality. Several empirical studies, including Nagel
(1974), Muller and Seligson (1989), Midlarski (1988), Muller et al. (1989), and
Brockett (1992), indicated a positive association between conflict and income or
wealth inequality. However, in a survey article, Lichbach (1989) concluded that the
reported results were generally not statistically significant. In their analysis Collier
and Hoeffer (1998) found that the role of income inequality in the explanation of
the outbreak of a civil war is statistically insignificant. Barron et al. (2004) analyzed
village-level conflict in Indonesia and concluded that poverty has very little correlation with conflict. On the other hand, changes in economic conditions and
unemployment were shown to be important factors. Using district-level data in
Nepal, Murshed and Gates (2005) found that inequality among subgroups, defined
along geographical concentration, was a significant explanatory factor for severity
of conflict (measured in terms of the number of deaths). (See also Macours 2011.)
In Chap. 2, we have argued that existence of a small middle-class and highincome inequality may be a causal factor of social unrest in a society. Factors like
abrupt changes in the price of an essential commodity (Montalvo and ReynalQuerol 2012) and distinct effects of shocks to the prices of goods in the local
segmented labor markets (Dube and Vargas 2008) have also been considered as
likely causes of conflicts. Dube and Vargas (2008) showed that in Colombia
increases in the price of oil increase conflict but increases in the price of coffee
do not. The reason is that in locations specializing in coffee production, increases in
coffee price raised the wages, and hence it becomes difficult to hire fighters. In
contrast, the opposite picture emerges in locations specializing in oil production.
For further discussion, see Dasgupta (1985), on which this brief presentation relies.
7.2 Fractionalization
111
There are also noneconomic factors that may generate social tensions and
conflicts. Examples are a natural disaster and the assassination of a popular political
leader. Kumar (2010) argued that hate and revenge can erode the chances of
peaceful settlement and lead to conflict. (See Amegashie and Runkel (2012) for a
different view.) For additional factors, see Chiozza and Goemans (2011), Chap. 1 of
Garfinkel and Skaperdas (2012), and Chap. 5 of this book.
Conflict can cripple economic performance, and bad economic performance can
build ground for commencement for civil war within nations (Justino 2012).
However, interest in the cost of conflict has been rather recent. In fact, an initiative
taken by the World Bank to investigate whether civil wars constitute an important
causal factor for poverty and underdevelopment is a demonstration of the fact that
in recent years there has been increased attention to the importance of conflict for
the economy (see the report by Collier et al. 2003). Some contributions have
estimated the cost of Iraq war (Stiglitz and Bilmes 2008) and conducted its cost
benefit analysis (Davis et al. 2009). Conflict and appropriation have been introduced into economic models, and their implications were investigated, among
others, by Hirshleifer (1988, 1994, 1995), Grossman (1991, 1994), and Grossman
and Kim (1995). Since conflict has economic consequences, it is always of some
interest to economists.
Since different ethnic subgroups in a society may have different preferences,
there may be conflicts of interests in economic decisions. People from different
culture may feel differently with respect to others depending on their similarity in
other dimensions. The index of fractionalization has been widely used in different
studies to link ethno-linguistic diversity to conflict, growth, or public good provision (see, e.g., Easterley and Levine 1997; Fearon and Laitin 2003b; Collier and
Hoeffler 2004; and Miguel et al. 2004).
In the next section of this chapter, we present an analytical discussion on the
index of fractionalization. Section 7.3 provides a rigorous analysis on the relationship between fractionalization, inequality, and polarization. The role of inequality,
polarization, and fractionalization in conflict is examined in Sect. 7.4 using a
behavioral model.
7.2
Fractionalization
Before Montalvo and Reynal-Querol (2002), a major part of the literature endeavored to explain conflict using an ethno-linguistic fractionalization index. This wellknown index is a measure of the extent to which a society is split into different
(predefined) ethnic subgroups. It is formally defined as
112

FRAC
1
k
X
!
2i
7:1
i1
where, as in Chap. 5, 1 ; 2 ; . . . ; k is the vector of population proportions in

k different ethnic subgroups of the population. The corresponding vector of absolute population sizes is n n1 ; n2 ; . . . ; nk . Assume that nis (hence is) are
non-increasingly ordered.
By definition, FRAC depends on population shares, not on absolute population
sizes. That
of degree zero in absolute population sizes so that
is, it is homogeneous

FRAC FRAC n . It has a simple interpretation of being the probability that
two persons selected at random from the population do not belong to the same
ethnic subgroup. To see this formally, note that the probability that an individual of
subgroup i is chosen is i . Hence the probability that the chosen person is matched
with someone from a different subgroup is i 1 i . Consequently, the probability
that any two individuals belong to two different subgroups is

k
k
P
P
i 1 i 1 2i . Alternatively, the probability that two randomly
i1
i1
selected individuals belong to the same subgroup is 2i and hence one minus sum
of these probabilities gives us FRAC.
This index is the complement of the well-known Herfindahl (1950)Hirschman
(1945) index of population concentration from unity. That is, the transformed
k
P
indicator 1 FRAC
2i is the HerfindahlHirschman index of population
i1
concentration. One common property of the two indices is that they remain
unchanged under deletion or addition of subgroups with zero population size. A
rank-preserving shift of population mass from one subgroup to a different subgroup
with a lower fraction of population increases FRAC. The reasoning underlying this
is that, given the number of subgroups, further splitting can take place through such
a shift. It remains invariant under a reordering of the population shares across
subgroups. Thus, in the context of measurement of ethno-linguistic fractionalization, we are concerned with only population proportions in different subgroups.
The concentration index indicates the dominance of population subgroups of
significant size. In contrast, FRAC satisfies a small-subgroup property, which says
that the addition of a subgroup with population size smaller than that of the smallest
of the existing subgroups, all other subgroups population sizes constant, will
increase the value of the fractionalization index. To state this formally, let (n1,
n2, . . ., nk) be the vector of population sizes in different subgroups and the
corresponding population share vector is 1 ; 2 ; . . . ; k . Now, suppose we
include a new subgroup k 1 to the existing population with the corresponding
population size nk1 , where 0 < nk1 < min fni g. We denote the subgroup
i1, 2, ::, k
population share vector for this enlarged population by t t1 ; t2 ; . . . ; tk ; tk1 ,
7.2 Fractionalization
113
where the associated populationsize vector is n1 ; n2 ; . . . ;nk; nk1 . Then the
small-subgroup property demands that FRAC t > FRAC . The following
theorem provides a rigorous demonstration of this.
Theorem 7.1 FRAC satisfies the small-subgroups property.
Proof The vector of population sizes across subgroups in the enlarged population
is n1 ; n2 ; . . . ; nk ; nk1 . We can transfer the new subgroups population mass nk1 to
all the other subgroups in proportion of their population sizes. The post transfer
distribution of population sizes is n1 c1 , n2 c2 , . . . , nk ck , 0, where
ci nkk1 ni , 1 i k. FRAC reduces under this sequence of transfers so that
P
nj
j1
FRACn1 c1 , n2 c2 , . . . , nk ck , 0 < FRACn1 ; n2 ; . . . ; nk ; nk1 . Since FRAC

remains unaltered under deletion of a subgroup with zero population size, we have
FRACn1 c1 , n2 c2 , . . . , nk ck , 0 FRACn1 c1 , n2 c2 , . . . , nk ck .
Now,
we 0can
rewrite
n1 c 1 , n2 c 2 , . . . , nk c k
as
1
C
B
C
B
n1 ; n2 ; . . . ; nk B1 nk k1 C. Hence
P A
@
nj
j1
0
1
FRACn1 c1 , n2 c2 , . . . , nk ck
B
C
B
C
FRACn1 ; n2 ; . . . ; nk B1 nk k1 C. By
P A
@
nj
j1
1
0
homogeneity
of
degree
C
B
C
B
FRACn1 ; n2 ; . . . ; nk B 1 knk1 C FRAC n1 ; n2 ; . . . ; nk :
P A
@
nj
zero,
Thus,
j1
FRACn1 ; n2 ; . . . ; nk < FRACn1 ; n2; . . . ; nk ; nk1

: Applying the homogeneity property again, we have FRAC t > FRAC , which is the desired
result.
The intuitive reasoning behind this result is that in the enlarged population the
share of population of each of the existing subgroups decreases, that is, i > ti for
all i 1, 2, k so that dominance of each of the large subgroups decreases which in
turn increases fractionalization.

The index FRAC is bounded between zero and 1 1k , where the lower bound is
achieved whenever all the individuals are concentrated in one subgroup. This
property seems intuitively reasonable. If the entire population is concentrated in
one subgroup, then there is minimum

fractionalization of the population. On the
other, it attains the upper bound 1 1k if the individuals are uniformly distributed
114
across subgroups, that is, the fraction of population in each subgroup is 1k. This is the
case of perfect fractionalization. This upper bound is increasing in the number of
subgroups k. The reason is that in the situation of perfect fractionalization, the
population becomes more fragmented as the number of subgroups increases.
Under a subgroup by subgroup replication, the population becomes more fractionalized and hence FRAC should increase. For any q-fold replication of the
population sizes in different subgroups, the fractionalization index for the repli

k
P
cated population, FRAC(q), becomes 1 1q 2i , where q > 1 is any integer. The
i1
relationship between FRAC(q) and FRAC is given by FRACq 1 1q1 FRAC,

which shows that as q increases, FRAC(q) increases. In fact, as the total population is
split into a higher number of subgroups, that is, as the value of k increases, FRAC
increases. For instance, if any subgroup with population size l is broken down into
two subgroups with population sizes l1 and l2 respectively, where l1 + l2 l, then
FRAC increases. Equivalently, a merger of two subgroups decreases FRAC.
If there are only two subgroups, then
FRAC up to a positive
PRQ equals

scalar.
To see this, note that for k 2, PRQ 4 21 1 1 22 1 2 4 1 2 .

2
P
On the other hand, for two subgroups FRAC becomes 1 2i 2 1 2 .
i1
Therefore, in the two-subgroup case, PRQ 2FRAC.

However, this relationship breaks down if we consider more than two subgroups.
To understand the difference between the two indices in the general case, note that
in FRAC the weight assigned to each of the k terms i 1 i , the probability of any
two individuals being members of two different subgroups, when one of them
belongs to subgroup i, is the same (1). But for PRQ this weight is i , the relative
size of subgroup i. Therefore, the two indices are subjected to two different types of
aggregation. For i 1k for all i, while PRQ decreases as the number of subgroups
k increases, the opposite happens for FRAC . In view of Property 1 of Chap. 5, we
know that in the three-subgroup case PRQ cannot decrease under a merger of the two
smallest subgroups. But FRAC decreases unambiguously for such a merger. Also
our discussion on FRAC and Property 1b of Chap. 5 shows that the two indices
behave differently under a splitting of subgroups. Further, PRQ is a violator of the
small-subgroup property. For instance, for the population size distribution (8, 6, 4),
the value of PRQ is 0.89. If we enlarge this population by adding a new subgroup
with a population size of 1, then the enlarged population is (8, 6, 4, 1) for which PRQ
becomes 0.83. This is a clear violation of the small-subgroup property by PRQ. But
this postulate is unambiguously satisfied by FRAC. Thus, the two indices are
fundamentally different. Bossert et al. (2011) axiomatically characterized a generalized index of fractionalization which is informationally richer and more flexible
than FRAC. This index takes the individuals as a primitive, as opposed to ethnic
7.3 Inequality and Fractionalization
115
subgroups, and uses information on the similarity among them. As in the case of
existing indices, this index does not require that the individuals are preassigned to
some exogenously given subgroups.
Using FRAC, Easterly and Levine (1997) were the first to find the evidence of a
negative correlation between ethnic diversity and economic growth. Later, Alesina
et al. (2003) used an updated dataset on ethnic fractionalization and found indication of negative influence of ethnic fractionalization on institutions and growth.
Glaeser et al. (1995) did not find affinity between ethnic fractionalization and
population growth for the US counties. Alesina and La Ferrara (2005) found
decreasing impact of ethnic fragmentation on redistributive policies in favor of
racial minorities and lower social capital, measured as trust. According to Collier
and Hoeffler (1998), FRAC is not statistically significant for probability of a civil
war but is weakly significant for duration of a civil war. Collier and Hoeffler (2004)
found that ethnic fractionalization and religious fractionalization are statistically
insignificant explanatory factors for outbreak of civil wars. Likewise, Fearon and
Laitin (2003b) did not find evidence for the power of ethnic fractionalization as an
explanatory factor for commencement of civil wars. Easterly et al. (2006) found
evidence of a relationship between mass killing and square of ethnic fractionalization. Lim et al. (2007) investigated the role of geographic distribution of population
as an explanatory factor of conflict. Highly mixed geographic regions and wellsegregated subgroups were not expected to generate violence. An implication of
this is that conflict has a low correlation with fractionalization but it may have a
high correlation with polarization. Monatalvo and Reynal-Querol (2005a) investigated the effect of polarization and fractionalization on the incidence of civil and
ethnic wars and, indirectly, on economic growth. No relationship between fractionalization and conflict was found for more than two subgroups. Fractionalization and
polarization were even found to be negatively correlated at higher level of fractionalization. Monatalvo and Reynal-Querol (2003) reported that religious polarization has more relevance to assess the impact of conflict on civil wars than
fractionalization. They found convincing negative impact of religious polarization
on development via its impact on investment, government expenditure, and frequency of civil wars. However, they did not find any effect of religious and ethnic
fractionalization on the incidence of ethnic and civil wars on growth or on investment (see also Monatalvo and Reynal-Querol 2002). In Montalvo and ReynalQuerol (2008), any significant effect of fractionalization on genocide was not
observed.
7.3
Inequality and Fractionalization
In the earlier section we have discussed the relationship between fractionalization

and polarization. In the present section our objective is to investigate the connection
between fractionalization and inequality. This also provides a second theoretical
support for FRAC.
116
For this purpose, the inequality index we consider is n12
k X
k
X

ni nj xi xj , twice
i1 j1
the absolute Gini index. As before, ni denotes the number of individuals in subgroup
k
P
i and xi is its income level, i 1, 2, . . . , k; ni n: We can rewrite this inequality
i1

i j xi xj .
i1 j1

When subgroups are identified by incomes, the distance xi xj used in the
inequality index considered above is an appropriate measure of income difference
between subgroups i and j. However, for distributions over political, ethnic, or
religious subgroups, the distance is nonmonetary. In such a case we need to
distinguish between two individuals only with respect to the subgroups to which
they belong.
A simple way to incorporate this is to assume that the Euclidean

metric xi xj takes on a positive value only when the corresponding subgroups
i and j are different; otherwise it is zero. For simplicity of exposition, this positive
value is taken as 1 (see Eq. 5.2). Then the value of the inequality index becomes
k P
k
k
k
k
P
P
P
P
i j . We can rewrite this as
i j
i 1 i , which is FRAC.
index in terms of subgroup population proportions i s as
i1 j6i
7.4
i1
j6i
k P
k
P
i1
Inequality, Polarization, and Fractionalization

as Indicators of Conflict: A Behavioral Model
In this section, following Esteban and Ray (2011), we present a behavioral model of
conflict that provides a link between conflict, inequality, and polarization. The
model allows for conflict over private goods and public goods. Also varying
degrees of within-group cohesion representing decisions to choices imposed by a
benevolent policy decision-making behavior are incorporated into the analysis.
It is assumed that the exogenously given k subgroups are engaged in a conflict.
They strive for a budget with per capita value normalized to unity. A fraction of
this budget is available for production of public goods. The remaining fraction
1 can be privately divided, that is, it can be seized by the members of the
winning subgroup.
Thus the winning camp enjoys two types of prizes, one is private, hence
excludable, and the other is public. Examples of private prizes can be an advantage
in access to rents from natural resources, bias in infrastructure facilities, and so
on. Since the private goods are divided among the winning subgroup, the size of the
subgroup is quite relevant (Olson 1971). It is assumed that the identity of the winner
is irrelevant to the losers.
7.4 Inequality, Polarization, and Fractionalization as Indicators of Conflict. . .
117
The enjoyment of the public prize is independent of the population size.

Examples can be political power, religious dominance, and control over policy.
Each subgroup i has its own ideal policy i. Other subgroups will derive benefits
from the public good depending on their distances from the winning subgroup. We
write uij to indicate the payoff of a member of subgroup i from the public services if
the ideal policy of subgroup j chosen, that is, when subgroup j is the winner. This
implicitly defines a distance or alienation between subgroups i and j as
dij uii uij . It is reasonable to assume that uii > uij for all i, j with i 6 j. If
subgroup i wins the conflict, it also enjoys the private services, and the payoff from
such services to a member of this subgroup is 1

i , where, as before, i is the
proportion of persons in subgroup i. Consequently, the payoff to a member of
subgroup i when it wins is uii 1

i . The parameter can, therefore, be treated as
an indicator of the importance of public service payoff in the overall payoff to a
member of the winning subgroup.
Conflict is treated as a situation in which there is no unambiguous rule aggregating the claims of different subgroups. The success of any subgroup is probabilistic. The probability depends on the subgroups expenditure to influence the final
outcome. Let be the resources sacrificed by a typical member of any subgroup to
control the ultimate resultant. The money equivalent of this expenditure is
represented by a cost C, where the cost function C verifies certain assumptions.
We write C0 for the derivative of C. An example of the desired type of cost function
is the iso-elastic function C , where 2 is a parameter.

We write i j for the contribution of resources
by member j of subgroup i. Then
P
subgroup is overall contribution is i j2Si i j, and the resource supplied by
k
P
the entire society is
i , where Si is the set of persons in subgroup i . The per
i1
capita conflict, the per capita value of the resource sacrificed in the contest, is n.
Let pi denote the probability that subgroup i will emerge as the winning camp. It
is assumed that pi i , where i 1, 2, . . . , k. That is, the probability that subgroup
i will win is exactly equal to the proportion of resource spent in support of the ideal
policy i. The overall expected payoff of individual j in subgroup i can now be
summarized as
i j
k
X
j1
pj uij pi
1
Ci j:
i
7:2
Since the model is probabilistic, the first term on the right hand side in this
expression is the expected value of payoff from public services. The second term
is the payoff to the concerned person from private services multiplied by the
probability of winning the contest by the subgroup of which the concerned person
is a member. The third term is the resource cost borne by the person under
consideration.
118
Each individual has an extended utility function (Sen 1966), which assigns
weight 1 to the own expected payoff defined in (7.2) and a weight on the
aggregate expected payoff of other members of the subgroup, where 0 1.
Formally,
X
U i j 1 i j
i l:
7:3
l2Si
Since j enters in the summation term also, the weight on js own expected payoff
i( j) is actually
P 1. For any positive value of the parameter , Ui( j) is an increasing
function of l2Si =fjgi l. Thus, the individual js utility function defined in (7.3) is
unselfishly devoted to the welfare of the other individuals in the subgroup. In other
words, Ui ( j) is an altruistic utility function. Therefore, represents altruism. If
0 the persons utility function is self-concerned, it depends on his own payoff. On
the other hand, if 1, utility depends on the subgroup aggregate payoff.2
Alternative interpretations of are possible. As Esteban and Ray (2011) argued,
suppose
a subgroup leader has the goal of maximizing the utilitarian objective
X
i l, whereas individual js selfish objective is i( j). Then a compromise
l2Si
solution achieved via bargaining is the convex mixture defined in (7.3). In such a
case represents the bargaining power of the leader (see also Grossman and
Helpman (1994) and Epstein and Nitzan (2002)).
A natural objective of the person with the utility function (7.3) is to choose i j
that maximizes Ui( j). Equilibrium in this context is a sequence of individual
contributions fi jg, where for every subgroup i and every member j, i j
maximizes Ui( j), given all other contributions. Esteban and Ray (2011) demonstrated that equilibrium exits and it is unique. In each subgroup all members make
the same resource contribution: i j i l for all j, l 2 Si .
It is now our objective to relate equilibrium conflict with distributional indices,
considered in the Chap. 3 and Sect. 7.2. For this we consider the EstebanRay
formula (3.2), expressed in terms of population shares, under the specifications that
1 and K 1 so that the formula reduces to the following form
P1
k X
k
X
2i j d ij ;
7:4
i1 j1

where the metric x^ i x^ j in (3.2) have been replaced by the public services utility
losses dij uii uij . A large weight is assigned to subgroup identification using the
2
There have been applications of altruistic behavior to many situations, including study of
cooperatives (Sen 1966), voting behavior (Fowler 2006; Edin et al. 2007; and Evren 2009),
intergenerational models (Barro and Becker 1989), joint decisions (Bell and Keeney 2009), and
deprivation (Chakravarty 2009a).
7.4 Inequality, Polarization, and Fractionalization as Indicators of Conflict. . .
squared term on population fraction. The inequality index
119
k X
k
X

i j xi xj ,
i1 j1
when rewritten in terms of public goods utility losses is
k X
k
X
i j d ij and we
i1 j1
denote this by A1G .

Esteban and Ray (2011) showed that the proxy ^ for equilibrium per capita
conflict can be determined as follows

0
^ C ^ 1 2 A1G P1 1 FRAC ;
7:5
k1
where 1 11
and 2 1
n
n . Equation (7.5) shows that equilibrium
per capita conflict is obtainable from a linear combination of three distributional
indices, twice the Gini index, the fractionalization index, and the EstebanRay
polarization index. Given that the cost function is increasing and strictly convex,
the left-hand side of (7.5) is an increasing function of ^ . For instance, in the
iso-elastic cost function if 1, the constant elasticity of the marginal cost,
takes on the value 1, then the left-hand side becomes ^ 2 .
The weights on the linear combination on the right-hand side provide information on the importance of each of these three indices as an interacting variable for
conflict. The weights depend explicitly on , a measure of the extent of publicness
of the service and on , an indicator of the intra-group cohesion.
As the population size becomes large, both 1 and 2 tend to zero. In such a case,
for any positive level of social cohesion, the equilibrium per capita conflict is
determined by a convex mix of fractionalization and polarization. Thus, in this
case, the existence of intragroup cohesion, not its value, is a relevant factor for per
capita conflict to be increasingly related to polarization and fractionalization. In
their study, Esteban et al. (2012) showed that ethnic polarization has a large and
significant impact on conflict. The same is true as well for fractionalization,
although to a somewhat lower extent.
The above notion of contest can be contrasted with the concept of P-power of
voting. Here the winning camp in a voting game receives some amount of
transferable utility, the prize of power, which the members of the camp can
freely divide among themselves without any efficiency loss. The payoff of the
members of the losing coalition is zero (see Chakravarty et al. 2015, for a recent
discussion). Thus, while in the EstebanRay framework, the public services are
divided among all the subgroups and the private services are owned by the
winning subgroup only, in the voting game, the prize is exclusively owned by
the winning camp and the question of sharing the prize with members of the losing
camp does not arise.
In the EstebanRay structure, the probability that a particular subgroup will win
is taken as the share of resources it spends to get its policy approved. Skaperdas
(1996) axiomatically characterized contest success functions which present each
120
players probability of winning contests as a function of efforts of all the players.3

These efforts can be arms in the context of military, a players skill in a tournament,
and amount of money spent by a player in a situation of rent-seeking and so
on. A players probability of success is an increasing function of his own effort
but a decreasing function of efforts of the other players. The only characteristic that
makes a distinction among the players is their efforts. All other characteristics like
their living conditions, names are irrelevant here. Skaperdas (1996) developed a
characterization of the contest success function given by P f ei , where ei is the
f ej
j2N
amount of effort provided by player i and N is the player set. The nonnegative
increasing function f, which we refer to as the effective
investment function,
P
f ej > 0.) It is unique up
is positive for positive efforts. (We assume that
j2N
to a positive scale transformation, that is, f and cf, where c > 0 is a scalar, give
us the same probabilities. This contest success function coincides with the
EstebanRay probability function pi if we assume that ei i and f is the identity
mapping (f t t for all t). (See also Clark and Riis 1998; Munster 2009; Rai and
Sarin 2009).
3
Contests have been used to analyze several issues, including rent seeking (Tullock 1980; Nitzan
1991; Baye et al. 1993), conflicts (Hirshleifer 1991; Skaperdas 1992), tournaments (Rosen 1986),
and political campaigns (Skaperdas and Grofman 1995). A general treatment of contests can be
found in Dixit (1986)
Glossary of Notation
N
1n
Rn
Dn
Dn
D
D
R
m
x^
x
xy
xTy
W
E(x)
L :
IAKS
ABDK
FW
RB
PER
PDER
PRQ
PO
set of positive integers

n-coordinated vector of ones
nonnegative orthant of the n-dimensional Euclidean space Rn
Rn with the origin deleted
positive part of Dn
[ Dn
n2N
[ Dn
n2N
[ Rn
n2N
mean income
median income
illfare-ranked permutation of the income distribution x, that is,
x^ 1 x^ 2 x^ n
welfare ranked permutation of the income distribution x, that is,
x1 x2 xn
xi yi for all i, with > for some i
x is obtained from y by a progressive transfer
social welfare function
AtkinsonKolmSen
equally
distributed
equivalent
income
corresponding to the income distribution x
Lorenz dominance, GL : generalized Lorenz dominance
AtkinsonKolmSen relative income inequality index
BlackorbyDonaldsonKolm absolute income inequality index
FosterWolfson income bipolarization ordering
relative bipolarization ordering, AB : absolute bipolarization ordering
EstebanRay income multipolar polarization index
DuclosEstebanRay income multipolar polarization index
Reynal-Querol social polarization index
ordinal polarization index
Springer India 2015

Inequality, Social Exclusion and Well-Being 12, DOI 10.1007/978-81-322-2166-1
121
Extended Bibliography
Aaberge R (2000) Characterizations of Lorenz curves and income distributions. Soc Choice
Welf 17
Aaberge R (2007) Ginis nuclear family. J Econ Inequal 5
Alesina A, La Ferrara E (2000) Participation in heterogeneous communities. Q J Econ 115
Alesina A, La Ferrara E (2005) Ethnic diversity and economic performance. J Econ Lit 43
Alesina A, Spolaore E (1997) On the number and size of nations. Q J Econ 113
Alesina A, Baqir R, Easterly W (1999) Public goods and ethnic divisions. Q J Econ 114
Alesina A, Glaeser E, Sacerdote B (2001) Why doesnt the United States have a European-style
welfare state? Brook Pap Econ Act 2
Alesina A, Devleeschauwer A, Easterly W, Kurlat S, Wacziarg R (2003) Fractionalization. J Econ
Growth 8
Alesina A, Easterly W, Matuszeski J (2011) Artificial states. J Eur Econ Assoc 9
Allison RA, Foster JE (2004) Measuring health inequality using qualitative data. J Health Econ 23
Amegashie JA, Runkel M (2012) The paradoxes of revenge in conflicts. J Confl Resolut 56
Amiel Y, Cowell FA (2003) Inequality, welfare and monotonicity. Res Econ Inequal 9
Amiel Y, Cowell FA, Ramos X (2007) On the measurement polarization: a questionnaire study.
Mimeographed
Anderson G (2004a) Making inferences about polarization, welfare and poverty of nations: a study
of 101 countries 19701995. J Appl Econ 19
Anderson G (2004b) Towards an empirical analysis of polarization. J Econ 122
Apouey B (2007) Measuring health polarization with self-assessed health data. Health Econ 16
Aristotle (-350) Politics, vol. 4, Part XI (trans Jowett B). The University of Adelaide,
eBooks@Adelaide
Atkinson AB (1970) On the measurement of inequality. J Econ Theory 2
Atkinson AB (2008) More on the measurement of inequality. J Econ Inequal 6
Atkinson AB, Bourguignon F (ed) (2014) Handbook of income distribution, vol II. North Holland,
Amsterdam (Forthcoming)
Azpitarte F, Alonso-Villar O ( 2012) A dominance criterion for measuring income inequality from
a centrist view: the case of Australia. Melbourne Institute working paper 3/12, University of
Melbourne, Melbourne
Barcena-Martin E, Silber J (2013) On the generalization and decomposition of the Bonferroni
index. Soc Choice Welf 41
Barro RJ, Becker GS (1989) Fertility choice in a model of economic growth. Econometrica 57
Barron P, Kaiser K, Pradhan M (2004) Local conflicts in Indonesia: measuring incidence and
identifying patterns. Working paper 3384. World Bank. Washington, DC
Springer India 2015

123
124
Baye M, Kovenock D, de Vries CG (1993) Rigging the lobbying process: an application to the
all-pay auction. Am Econ Rev 83
Beach C (1989) Dollars and dreams: a reduced middle class? Alternative explanations.
J Hum Res 24
Beach C, Chaykowski R, Slotsve G (1997) Inequality and polarization of male earnings in the
United States: 196890. North Am J Econ Finance 8
Bell DE, Keeney RL (2009) Altruistic utility functions for joint decisions. In: Brams SJ, Gehrlein
WV, Roberts FS (eds) The mathematics of preference, choice and order: essays in honor of
Peter C Fishburn. Springer, New York
Ben Porath E, Gilboa I (1994) Linear measures, the Gini index, and the income-equality tradeoff. J
Econ Theory 64
Berrebi ZM, Silber J (1989) Deprivation, the Gini index of inequality and the flatness of an income
distribution. Math Soc Sci 18(3):229237
Besley T, Preston I (1988) Invariance and the axiomatics of income tax progressivity: a comment.
Bull Econ Res 40
Bhattracharya N, Mahalanobis B (1967) Regional disparities in household consumption in India. J
Am Stat Assoc 62
Birdsall N (2007) Reflections on the macro foundations of the middle class in the developing
world. Centre for Global Development, Washington DC (www.cgdev.org)
Blackburn M (1994) International comparisons of poverty. American Economic Review 84
Blackburn M, Bloom DE (1995) What is happening to the middle class? American Demographics,
January
Blackorby C, Donaldson D (1978) Measures of relative equality and their meaning in terms of
social welfare. J Econ Theory 18
Blackorby C, Donaldson D (1980) A theoretical treatment of indices of absolute inequality. Int
Econ Rev 21
Blackorby C, Donaldson D (1984) Social criteria for evaluating population change. J Public Econ 25
Blackorby C, Donaldson D, Auersperg M (1981) A new procedure for the measurement of
inequality within and among population subgroups. Can J Econ 14
Blackorby C, Bossert W, Donaldson D (1999) Income inequality measurement: the normative
approach. In: Silber (1999)
Blackorby C, Bossert W, Donaldson D (2005) Population issues in social choice theory, welfare
economics and ethics. Cambridge University Press, Cambridge
Blackorby C, Primont D, Russell RR (1978) Duality, separability, and functional structure: theory
and economic applications. North Holland, Amsterdam
Blewett E (1982) Measuring lifecycle inequality. PhD dissertation. University of British Columbia, Vancouver
Bonferroni C (1930) Elemente di dtatistica generale. Libreria Seber, Firenze
Bosmans K, Cowell FA (2010) The class of absolute decomposable inequality measures. Econ
Lett 109
Bossert W (1990) An axiomatization of the single-series Ginis. J Econ Theory 50
Bossert W, Pfingsten A (1990) Intermediate inequality: concepts, indices and welfare implication.
Math Soc Sci 19
Bossert W, Schworm W (2008) Measures of polarization. J Pub Econ Theory 10
Bossert W, DAmbrosio C, La Ferrara E (2011) A generalized index of fractionalization.
Economica 78
Bourguignon F (1979) Decomposable income inequality measures. Econometrica 47
Braubaker R, Laitin DD (1998) Ethnic and nationalist violence. Annu Rev Sociol 24
Brockett CD (1992) Measuring political violence and land inequality in Central America. Am Polit
Sci Rev 86
Caselli F, Coleman II WJ ( 2006) On the theory of ethnic conflict. Working paper 12125. National
Bureau of Economic Research, Cambridge
Chakravarty SR (1988) Extended Gini indices of inequality. Int Econ Rev 29
125
Chakravarty SR (1989) On quasi-ordering of income profiles. Methods Oper Res 60

Chakravarty SR (1990) Ethical social index numbers. Springer, New York
Chakravarty SR (2007) A deprivation-based axiomatic characterization of the absolute Bonferroni
index of inequality. J Econ Inequal 5
Chakravarty SR (2009a) Inequality, polarization and poverty: advances in distributional analysis.
Springer, New York
Chakravarty SR (2009b) Equity and efficiency as components of a social welfare function. Int J
Econ Theory 5
Chakravarty SR (2013) An outline of financial economics. Anthem Press, New York
Chakravarty SR, DAmbrosio C (2006) The measurement of social exclusion. Rev Income
Wealth 52
mbrosio C (2010) Polarization ordering of income distributions. Rev Income
Chakravarty SR, DA
Wealth 56
Chakravarty SR, Dutta B (1987) A note on measures of distance between income distributions. J
Econ Theory 41
Chakravarty SR, Maharaj B (2011a) Subgroup decomposable inequality indices and reduced-form
indices of polarization. Keio Econ Stud 47
Chakravarty SR, Maharaj B (2011b) Measuring ethnic polarization. Soc Choice Welf 37
Chakravarty SR, Maharaj B (2012) Ethnic polarization orderings and indices. J Econ Interac
Coord 7
Chakravarty SR, Maharaj B (2014) Generalized Gini polarization indices for an ordinal dimension
of human well-being. Int J Econ Theory (Forthcoming).
Chakravarty SR, Majumder A (2001) Inequality, polarization and welfare: theory and applications. Aus Econ Pap 40
Chakravarty SR, Tyagarupananda S (2009) The subgroup decomposable intermediate indices of
inequality. Span Econ Rev 11
Chakravarty SR, Zoli C (2012) Stochastic dominance rules for integer variables. J Econ Theory 147
Chakravarty SR, Dutta B, Weymark JA (1985) Ethical indices of income mobility. Soc Choice
Welf 2
Chakravarty SR, Majumder A, Roy R (2007) A treatment of absolute indices of polarization. Japn
Econ Rev 58
Chakravarty SR, Chattopadhyay N, Maharaj B (2010) Inequality and polarization: an axiomatic
approach. In: Deutsch J, Silber J (2010)
Chakravarty SR, Mitra M, Sarkar P (2015) A course on cooperative game theory. Cambridge
University Press, New Delhi
Champernowne DG, Cowell FA (1998) Economic inequality and income redistribution. Cambridge University Press, Cambridge
Chateauneuf A, Moyes P (2006) A non-welfarist approach to inequality measurement. In:
McGillivray M (ed) Inequality, poverty and well-being. Palgrave-Macmillan, London
Chiozza G, Goemans HE (2011) Leaders and international conflict. Cambridge University Press,
Cambridge
Chipman JS (1974) The welfare ranking of Pareto distributions. J Econ Theory 9
Clark DJ, Riis C (1998) Contest success functions: an extension. Economic Theory 11
Collier P, Gunning W (2000) African trade liberalizations: alternative strategies for sustainable
reform. In: Bevan DL, Collier P, Gemmell N, Greenaway D (eds) Trade and fiscal adjustment
in Africa. Macmillan, London
Collier P, Hoeffler A (1998) On economic causes of civil war. Oxf Econ Pap 50
Collier P, Hoeffler A (2004) Greed and grievance in civil war. Oxf Econ Pap 56
Collier P, Elliott L, Hegre H, Hoeffler A, Reynal-Querol M, Sambanis N (2003) Breaking the
conflict trap: civil war and development policy. World Bank, Washington, DC
Cowell FA (1980) On the structure of additive inequality measures. Rev Econ Stud 47
Cowell FA, Kuga K (1981) Additivity and the entropy concept: an axiomatic approach to
inequality measurement. J Econ Theory 25
126
DAmbrosio C (2001) Household characteristics and the distribution of income in Italy: an

application of a social distance measure. Rev Income Wealth 47
Dagum C (1990) On the relationship between income inequality measures and social welfare
functions. J Econ 43
Dalton H (1920) The measurement of the inequality of incomes. Econ J 30
mbrosio C, Permanyer I (2013) Measuring social polarization with ordinal and categorical
DA
data. . J Pub Econ Theory (Forthcoming)
Dasgupta AK (1985) Epochs of economic theory. Oxford University Press, New Delhi
Dasgupta P, Sen AK, Starrett D (1973) Notes on the measurement of inequality. J Econ Theory 6
Davis JC, Huston JH (1992) The shrinking middle class: a multivariate analysis. East Econ J 18
Davis SJ, Murphy KM, Topel RH (2009) War in Iraq versus containment. In: Hess GD (ed) Guns
and butter: the economic causes and consequences of conflict. MIT Press, Cambridge
Del Rio C, Alonso-Villar O (2008) Rankings of income distributions: a case of intermediate
inequality indices. Res Econ Inequal 16
Del Rio C, Alonso-Villar O (2010) New unit- consistent intermediate inequality indices. Economic Theory 42
Del Rio C, Ruiz-Castillo J (2000) Intermediate inequality and welfare. Soc Choice Welf 17
Desmet K, Ortuno^a-Ortin I, Weber S (2008) Stability of nations and genetic diversity. Working
paper 004-08, International School of Economics at TSU, Tbilisi, Republic of Georgia
Desmet K, Ortuno^a-Ortin I, Wacziarg R (2012) The political economy of ethnolinguistic cleavages. J Dev Econ 97
Deutsch J, Silber J (2010a) Analyzing the impact of income sources on changes in bipolarization.
In: Deutsch, Silber (2010b)
Deutsch J, Silber J (eds) (2010b) Measurement of individual well-being and group inequalities:
essays in honor of ZM Berrebi. Routledge, London
Deutsch J, Hanoka M, Silber J (2007) On the link between the concepts of kurtosis and bipolarization. Econ Bull 4
Deutsch J, Fusco A, Silber J (2013) The BIP trilogy (bipolarization, inequality and polarization):
one saga but three different stories. Econ E-J 7
Dixit AK (1986) Strategic behavior in contests. Am Econ Rev 77
Dixit AK, Weibull JW (2007) Political polarization. doi:10.1073/pnas.0702071104
Do Q, Iyer L (2010) Geography, poverty and conflict in Nepal. J Peace Res 47
Donaldson D, Weymark JA (1980) A single parameter generalization of the Gini indices of
inequality. J Econ Theory 22
Donaldson D, Weymark JA (1983) Ethically flexible Gini indices of inequality for income
distributions in the continuum. J Econ Theory 29
Dube O, Vargas JF ( 2008) Commodity price shocks and civil conflict: evidence from Columbia.
Unpublished working paper. Harvard University
Duclos J-Y, Echevin D (2005) Bipolarization comparisons. Econ Lett 87
Duclos J-Y, Taptue A-M ( 2014) Polarization. In: Atkinson, Bourguignon (2014)
Duclos J-Y, Esteban J, Ray D (2004) Polarization: concepts, measurement, estimation.
Econometrica 72
Dutta B, Esteban JM (1992) Social welfare and equality. Soc Choice Welf 9
Easterly W (2001) The middle class consensus and economic development. J Econ Growth 6
Easterly W, Levine R (1997) Africas growth tragedy: policies and ethnic divisions. Q J Econ 112
Easterly W, Gatti R, Kurlat S (2006) Development, democracy and mass killings. J Econ Growth 11
Ebert U (1984) Measures of distance between income distributions. J Econ Theory 22
Ebert U (1987) Size and distributions of income as determinants of social welfare. J Econ Theory 41
Ebert U (1988a) On a family of aggregate compromise inequality measures. Int Econ Rev 29
Ebert U (1988b) Measurement of inequality: an attempt at unification and generalization. Soc
Choice Welf 5
Ebert U (1988c) On the decomposition of inequality: partition into non-overlapping subgroups. In:
Eichhorn W (1988)
127
Ebert U (1999) Dual decomposable inequality measures. Can J Econ 32

Ebert U (2010) The decomposition of inequality reconsidered: weakly decomposable measures.
Math Soc Sci 60
Edin A, Gelman A, Kaplan N (2007) Voting as a rational choice: why and how people vote to
improve the well-being of others. Ration Soc 19
Eichhorn W (ed) (1988) Measurement in economics: theory and applications of economic indices.
Springer, New York
Eichhorn W (ed) (1994) Models and measurement of welfare and inequality. Springer, New York
Epstein GS, Nitzan S (2002) Endogenous public policy, politicization and welfare. J Pub Econ
Theory 4
Esteban JM, Ray D (1994) On the measurement of polarization. Econometrica 62
Esteban JM, Ray D (1999) Conflict and distribution. J Econ Theory 87
Esteban JM, Ray D (2007) Polarization, fractionalization and conflict. J Peace Res 45
Esteban JM, Ray D (2008a) On the salience of ethnic conflict. Am Econ Rev 98
Esteban JM, Ray D (2008b) Polarization, fractionalization and conflict. J Peace Res 45
Esteban JM, Ray D (2011) Linking conflict to inequality and polarization. Am Econ Rev 101
Esteban JM, Ray D (2012) A comparison of polarization measures. In: Garifinvel, Skaperdas
(2012)
Esteban JM, Gradin C, Ray D (2007) Extensions of a measure of polarization, with an application
to income distribution of five OECD countries. J Econ Inequal 5
Esteban JM, Mayoral L, Ray D (2012) Ethnicity and conflict: an empirical study. Am Econ Rev 102
Evren O (2009) Altruism, voting turnout and strategic behavior. Department of Economics,
New York University
Fearon JD (2003) Ethnic and cultural diversity by country. J Econ Growth 8
Fearon JD, Laitin DD (2003a) Ethnicity, insurgency and civil war. Am Political Sci Rev 97:7590
Fearon JD, Laitin DD (2003b) Violence and the social construction of ethnic identity. Int Org 54
Fishburn P, Lavalle IH (1995) Stochastic dominance on unidimensional grids. Math Oper Res 20
Fleurbaey M, Maniquet F (2011) A theory of fairness and social choice. Cambridge University
Press, Cambridge
Fleurbaey M, Michal P (2001) Transfer principles and inequality aversion, with an application to
optimal growth. Math Soc Sci 42
Foster JE (1983) An axiomatic characterization of the Theil measure of inequality. J Econ Theory 31
Foster JE (1985) Inequality measurement. In: Young HP (ed) Fair allocation. American Mathematical Society, Providence
Foster JE, Ok E (1999) Lorenz dominance and variance of logarithms. Econometrica 67
Foster JE, Sen AK (1997) On economic inequality: after a quarter century, annex to enlarged
edition of on economic inequality by AK Sen. Clarendon Press, Oxford
Foster JE, Shneyerov A (1999) A general class of additively decomposable inequality measures.
Economic Theory 14
Foster JE, Shorrocks AF (1988) Poverty orderings and welfare dominance. Soc Choice Welf 5
Foster JE, Wolfson MC (2010) Polarization and the decline of the middle class: Canada and the
U.S. J Econ Inequal 8
Fowler JH (2006) Altruism and turnout. J Polit 68
Garifinvel MR, Skaperdas S (eds) (2012) The Oxford handbook of economics of peace and
conflict. Oxford University Press, New York
Gastwirth JL (1971) A general definition of the Lorenz curve. Econometrica 39
Glaeser E, Schleifer A, Scheinkman J (1995) Economic growth in a cross-section of cities. J Monet
Econ 36
Graaff JDV (1977) Equity and efficiency as components of the general welfare. South Afr J Econ 45
Gradin C (2000) Polarization by sub-populations in Spain, 197391. Rev Income Wealth 46
Gradin C (2002) Polarization and inequality in Spain, 197391. J Income Distrib 11
Gradin C (2006) Income distribution and income sources in Uruguay. J Appl Econ 9
128
Grew R (1997) On seeking the cultural context of fundamentalism. In: Marty ME, Appleby RS
(eds) Religion, ethnicity, and self identity: nations in turmoil. University Press of New
England, Hanover
Grier K, Maynard N (2010) Beyond two peaks: development and polarization in the world income
distribution. University of Oklahoma
Grossman HI (1991) A general equilibrium model of insurrections. Am Econ Rev 81
Grossman HI (1994) Production, appropriation and land reform. Am Econ Rev 84
Grossman GM, Helpman E (1994) Protection for sale. Am Econ Rev 84
Grossman HI, Kim M (1995) Swords or plowshares? A theory of the security claims to property. J
Polit Econ 103
Hadar J, Russell W (1969) Rules for ordering uncertain prospects. Am Econ Rev 59
Hardy GH, Littlewood J, Polya G (1934) Inequalities. Cambridge University Press, Cambridge
Harff B (2003) No lessons learned from holocaust? Assessing risks of genocide and political mass
murder since 1955. Am Polit Sci Rev 97
Hirshleifer J (1988) The analytics of continuing conflict. Synthese 76
Hirshleifer J (1991) The paradox of power. Econ Polit 3
Hirshleifer J (1994) The dark side of the force. Econ Inq 32
Hirshleifer J (1995) Anarchy and its breakdown. J Polit Econ 103
Horowitz D (1985) Ethnic groups in conflict. University of California Press, Berkeley
Horrigan MW, Haugen SE (1988) The declining middle class: a sensitivity analysis. Mon Labour
Rev 111
Juergensmeyer M (1993) The new cold war? Religious nationalism confronts the secular state.
University of California Press, Berkeley
Justino P (2012) War and poverty. In: Garifinvel, Skaperdas (2012)
Kakwani NC (1980a) Income inequality and poverty: methods of estimation and policy applications. Oxford University Press, Oxford
Kakwani NC (1980b) On a class of poverty measures. Econometrica 48
Kakwani NC (1985) Measurement of welfare with applications to Australia. J Dev Econ 18
Kats A (1972) On the social welfare function and the parameters of income distribution. J Econ
Theory 5
Knack S, Keefer P (2001) Polarization, politics and property rights: links between inequality and
growth. World Bank, Washington, DC
Kolm SC (1969) The optimal production of social justice. In: Margolis J, Guitton H (eds) Public
economics. Macmillan, London
Kolm SC (1976a) Unequal inequalities I. J Econ Theory 12
Kolm SC (1976b) Unequal inequalities II. J Econ Theory 13
Kondor Y (1975) Value judgment implied by the use of various measures of income inequality.
Rev Income Wealth 21
Kosters M, Ross M (1988) A shrinking middle class? Pub Interest 90
Kovacevic M, Binder D (1997) Variance estimation for measures of income inequality and
polarization: the estimating equations approach. J Off Stat 13
Krtscha M (1994) A new compromise measure of inequality. In: Eichhorn (1994)
Kumar R (2010) Essays on the economics of conflict. PhD dissertation. University of California,
Irvine
La Porta R, Lopez de Silanes F, Shleifer A, Vishny R (1999) The quality of government. J Law
Econ Organ 15
Lacina BA (2006) Explaining the severity of civil wars. J Confl Resolut 50
Lambert PJ (1985) Social welfare and the Gini coefficient revisited. Math Soc Sci 9
Lambert PJ (2001) The distribution and redistribution of income. Manchester University Press,
Manchester
Landes D (1998) The wealth and poverty of nations. Norton, New York
Lasso de la Vega MC, Urrutia AM (2006) An alternative formulation of the Esteban- Gradin-Ray
extended measure of polarization. J Income Distrib 15
129
Lasso de la Vega MC, Urrutia AM, Diez H (2010) Unit-consistency and polarization of income
distributions. Rev Income Wealth 56
Lawrence, RZ (1984) Sectoral shifts and the size of the middle class. Brook Rev:Fall
Leckie N (1988) The declining middle and technological change: trends in the distribution of
employment income in Canada, 197184. Economic Council of Canada, Ottawa
Levy F (1987) The middle class: is it really vanishing? Brook Rev 5
Levy H (2006) Stochastic dominance, investment decisions making under uncertainty, 2nd edn.
Springer, New York
Levy F, Murname RJ (1992) US earnings levels and earnings inequality: a review of recent trends
and proposed explanations. J Econ Lit 30
Lichbach MI (1989) An evaluation of Does economic inequality breed political conflict? studies.
World Polit 41
Lim M, Metzler R, Bar-Yam Y (2007) Global pattern formation and ethical/cultural violence.
Science 317
Macours K (2011) Increasing inequality and civil conflicts in Nepal. Oxf Econ Pap 63
Mansfield ED (1993) Concentration, polarity and the distribution of power. Int Stud Q 37
Marshall AW, Olkin I, Arnold BC (2011) Inequalities: theory of majorization and its applications,
2nd edn. Springer, New York
Marx K (1958) Capital: volume I. Foreign Languages Publishing House, Moscow
Matuszeski J, Schneider F (2006) Patterns of ethnic group segregation and civil conflict. Working
paper. Harvard University
Mauro P (1995) Corruption and growth. Q J Econ 110
McBride M, Richardson G (2012) Religion, conflict and cooperation. In: Garifinvel S (2012)
McBride M, Milnate G, Skaperdas S (2011) Peace and war with endogenous state capacity. J Confl
Resolut 55
McMahon PJ, Tsechetter JH (1986) The declining middle class: a further analysis. Mon Labour
Rev 109
McMlelland G, Rohrbaugh J (1978) Who accepts the Pareto axiom? The role of utility and equity
in arbitration decisions. Behav Sci 23
Mehran F (1976) Linear measures of income inequality. Econometrica 44
Midlarski MI (1988) Rulers and the ruled: patterned inequality and the onset of mass political
violence. Am Polit Sci Rev 82
Miguel E, Satyanath S, Sergenti S (2004) Economic shocks and civil conflicts: an instrumental
variable approach. J Polit Econ 112
Milanovic B (2000) A new polarization measure and some applications. World Bank, Washington,
DC
Mill JS (1920) Principle of political economy. Longmans, London
Montalvo JG, Reynal-Querol M (2002) Why ethnic fractionalization? Polarization, ethnic conflict
and growth. Mimeo
Montalvo JG, Reynal-Querol M (2003) Religious polarization and economic development. Econ
Lett 80
Montalvo JG, Reynal-Querol M (2005a) Ethnic polarization, potential conflict and civil wars. Am
Econ Rev 95
Montalvo JG, Reynal-Querol M (2005b) Ethnic diversity and economic development. J Dev
Econ 76
Montalvo JG, Reynal-Querol M (2008) Discrete polarization with an application to the determinants of genocides. Econ J 118
Montalvo JG, Reynal-Querol M (2012) Inequality, polarization and conflict. In: Garifinvel,
Skaperdas (2012)
Mookherjee D, Shorrocks AF (1982) A decomposition analysis of the trend in UK income
inequality. Econ J 92
Moyes P (1987) A new concept of Lorenz domination. Econ Lett 23
Moyes P (1999) Stochastic dominance and the Lorenz Curve. In: Silber (1999)
130
Muller EN, Seligson MA (1989) Inequality and insurgency. Am Polit Sci Rev 81
Muller EN, Seligson MA, Fu H, Midlarski MI (1989) Land inequality and political violence. Am
Polit Sci Rev 83
Munster GJ (2009) Group contest success functions. Economic Theory 41
Murshed SM, Gates S (2005) Spatial-horizontal inequality and Maoist insurgency in Nepal. Rev
Dev Econ 9
Nagel J (1974) Inequality and discontent: a nonlinear hypothesis. World Polit 26
Newbery DMG (1970) A theorem on the measurement of inequality. J Econ Theory 2
Nitzan S (1991) Collective rent dissipation. Econ J 101
Olson M (1971) The logic of collective action. Harvard University Press, Cambridge
Paap R, van Dijk HK (1998) Distribution and mobility of wealth of nations. Rev Income Wealth 42
Permanyer I (2012) The conceptualization and measurement of social polarization. J Econ
Inequal 10
Permanyer I (2014) Theoretical approaches to the measurement of income and social polarization.
In D Ambrosio C (ed) Handbook of research on economic and social well-being. Edward
Elgar Publishing, Northampton (Forthcoming)
Pigou AC (1912) Wealth and welfare. Macmillan, London
Pittau, MG, Zelli, R, Johnson PA ( 2010) Mixture models, convergence clubs and polarization.
Rev Income Wealth 56
Pollak RA (1971) Additive utility functions and linear Engel curves. Rev Econ Stud 38
Pressman S (2001) The decline of the middle class: an international perspective. Working paper
series, no. 280, Luxembourg Income Study
Quah D (1996) Two peaks: growth and convergence in models of distribution dynamics. Econ J
106
Quah D (1997) Empirics for growth and distribution: stratification, polarization and convergence
clubs. J Econ Growth 2
Rai BK, Sarin R (2009) Generalized contest success functions. Econ Theory 40
Rawls J (1971) A theory of justice. Harvard University Press, Cambridge
Reynal-Querol M (2002) Ethnicity, political systems and civil wars. J Confl Resolut 46
Ricardo D (1933) Principles of political economy and taxation. Everymans Library
Robbins L (1939) The economic basis of class conflict. Macmillan, London
Rodriguez J, Salas R (2003) Extended bipolarization and inequality measures. Res Econ Inequal 9
Rosen S (1986) Prizes and incentives in elimination tournaments. Am Econ Rev 76
Rothschild M, Stiglitz JE (1970) Increasing risk I: a definition. J Econ Theory 2
Rothschild M, Stiglitz JE (1973) Some further results on the measurement of inequality. J Econ
Theory 6
Saposnik R (1981) Rank dominance in income distribution. Public Choice 36
Saposnik R (1983) On evaluating income distributions: the SuppesSen grading principle of
justice and Pareto optimality. Public Choice 40
Seidl C, Pfingsten A (1997) Ray invariant inequality measures. Res Econ Inequal 7
Sen AK (1966) Labor allocation in a cooperative enterprise. Rev Econ Stud 33
Sen AK (1974) Informational basis of alternative welfare approaches: aggregation and income
distribution. J Public Econ 3
Seul JR (1999) Ours is the way of god: religion, identity, and intergroup conflict. J Peace Res 36
Shaked M, Shanthikumar G (2006) Stochastic orders. Springer, New York
Sheshinski E (1972) Relation between a social welfare function and the Gini index of inequality. J
Econ Theory 4
Shorrocks AF (1980) The class of additively decomposable inequality measures. Econometrica 48
Shorrocks AF (1982) On the distance between income distributions. Econometrica 50
Shorrocks AF (1983) Ranking income distributions. Economica 50
Shorrocks AF (1984) Inequality decomposition by population subgroups. Ecomometrica 52
Shorrocks AF (1988) Aggregation issues in inequality. In: Eichhorn (1988)
Silber J (ed) (1999) Handbook of income inequality measurement. Kluwer, Boston
131
Skaperdas S (1992) Cooperation, conflict and power in the absence of property rights. Am Econ
Rev 82
Skaperdas S (1996) Contest success functions. Economic Theory 7
Skaperdas S, Grofman B (1995) Modelling negative campaigning. Am Polit Sci Rev 89
Smith A (1933) Weal of nations: volume I. Everymans Library
Spolaore E, Wacziarg R (2013) War and relatedness. Nat Bureau Econ Res
Stiglitz JE, Bilmes LJ (2008) The three trillion dollar war: the true cost of the Iraq conflict. Norton,
New York
Theil H (1967) Economics and information theory. North Holland, Amsterdam
Theil H (1972) Statistical decomposition analysis. North Holland, Amsterdam
Thon D (1982) An axiomatization of the Gini coefficient. Math Soc Sci 2
Thurow L (1984) The disappearance of the middle class. New York Times, 5 February
Trannoy A (1986) On Thons axiomatization of the Gini index. Math Soc Sci 11
Tullock G (1980) Efficient rent seeking. In: Buchanan JM, Tollison RD, Tullock G (eds) Toward a
theory of rent seeking. Texas A & M University Press, College Station
Wang Y-Q, Tsui K-Y (2000) Polarization orderings and new classes of polarization indices. J Pub
Econ Theory 2
Weymark JA (1981) Generalized Gini inequality indices. Math Soc Sci 1
Weymark JA (2006) The normative approaches to the measurement of multidimensional inequality. In: Farina F, Savaglio G (eds) Inequality and economic integration. Routledge, London
Whitmore GA (1970) Third degree stochastic dominance. Am Econ Rev 24
Wolfson MC (1994) When inequalities diverge. Am Econ Rev 84
Wolfson MC (1997) Divergent inequalities: theory and empirical results. Rev Income Wealth 43
Yalonetzky G (2012) Measuring group disadvantage with inter-distributional inequality indices: a
critical review and some amendments to existing indices. Econ E-J 6
Yalonetzky G (2014) Relative bipolarization Lorenz curve. Econ Bull 34
Yitzhaki S (1983) On an extension of the Gini inequality index. Int Econ Rev 24
Yitzhaki S (1998) More than a dozen alternative ways of spelling Gini. Res Econ Inequal 8
Yoshida T (2005) Social welfare rankings of income distributions: a new parametric concept of
intermediate inequality. Soc Choice Welf 24
Zhang X, Kanbur R (2001) What differences do polarization measures make? An application to
China. J Dev Stud 37
Zheng B (2007a) Unit- consistent decomposable inequality measures. Economica 74
Zheng B (2007b) Inequality orderings and unit consistency. Soc Choice Welf 29
Zoli C (1999) A generalized version of the inequality invariance criterion: a surplus sharing
characterization, complete and partial orderings. In: De Swart H (ed) Logic, game theory
and social choice. Tilburg University Press, Tilburg
Index
A
Abbreviated\reduced form welfare function,
22, 63
Abbreviated polarization index, 63
Absolute bipolarization curve, 4850
Absolute Gini index, 13, 25, 30, 49, 60, 116
Absolute index, 4, 24, 25, 31, 48, 51, 60, 65, 66
Absolute Lorenz curve, 21
Absolutely continuous, 57
Absolute maximin index, 24
Accumulation, 109
Additive index, 84
Additive indices, 26
Additively decomposable, 26
Additively separable, 26
Algorithm, 76
Alienation, 5355, 57, 58, 6064, 67, 68, 70
75, 79, 8890, 98, 105107, 117
Altruism, 118
Altruistic utility, 118
Anonymity, 3
Artificiality, 78
Atkinson index, 22, 31
Atkinson-Kolm-Sen index, 31
B
Bargaining power, 118
Bargain theory, 110
Basic density function, 58
Behavioral models, 57, 111, 116120
Between-group Gini index, 67
Between-group inequality, 1, 26, 30, 31, 63,
65, 72
Between-group Lorenz curve, 67

Between-group radicalism, 90
Between-group term, 26, 63
Bipartitioning, 63, 70, 73, 102
Bipolarization dominance, 47
Bipolarization ordering, 34, 4048, 64, 70,
107108
Bipolarization ranking, 46
Bistochastic matrix, 2, 3, 19, 41
Blackorby-Donaldson-Kolm absolute index,
24, 25, 31
Bonferroni index, 12, 13, 28
Bonferroni welfare function, 23
Boundedness, 105, 108
Bunching of incomes, 38
C
Capitalists, 109
Capital-labor relationship, 109
Caste diversity, 78
Categorical variable, 97
Central mass, 55
Centrist concept, 4
Characteristic function, 66
Civil war, 34, 79, 110, 111, 115
Class conflict, 109, 110
Classical economists, 109, 110
Class relation, 109, 110
Clustering, 33, 38, 39, 51, 61, 64, 78, 107
Coefficient of variation, 12, 13, 23, 27, 44
Complete homogeneity, 82
Completely disjoint, 107
Compromise relative indices, 13
Springer India 2015

133
134
Compromise solution, 118
Concentration curve dominance, 93
Contest, 117, 119, 120
Contest success function, 119, 120
Continuity, 9, 21, 38, 99
Convex, 2, 5, 6, 13, 19, 41, 72, 74,
117, 119
Convex mixture, 5, 118
Cost of conflict, 111
Cost of polarization, 44
D
Daltonian principle of population (DPP), 79,
11, 12, 14, 28
Density function, 57, 58, 88
Deprivation, 56, 118
Dichotomous identification, 78
Difference form index, 75
Diminishing transfer principle, 11
Discrete metric, 7883
Discrete probability distribution, 83, 99
Distance function, 25, 57, 61, 79
Distributional judgments, 18
Distributionally homothetic, 25
Distributional ranking, 14
Distribution function, 14, 17, 35, 36, 57, 59,
60, 101, 108
Division of labor, 109
DPP. See Daltonian principle of
population (DPP)
E
Effective antagonism, 54, 57
Effective investment, 120
Efficiency, 15, 17, 19, 2224, 51, 63, 119
Efficiency considerations, 20
Effort, 110, 120
Elasticity of marginal cost, 119
Employment segregation, 84
Equally distributed equivalent, 21
Equilibrium conflict, 118
Equilibrium per capita conflict, 119
Equity, 15, 17, 19, 2225
Equity oriented, 20
Ethical decomposition, 31
Ethically indifferent, 21
Ethnic conflict, 77, 79
Ethnic fractionalization, 115
Ethno-linguistic diversity, 111
Ethno-linguistic fractionalization, 111
Euclidean metric, 80, 81
Index
F
Feasible index, 7072, 74
First order stochastic dominance, 18, 39,
98, 107
Fractionalization, 78, 109120
G
Generalized additive decomposability, 28
Generalized entropy family, 2729, 65
Generalized Gini index, 10
Generalized Gini inequality index, 105
Generalized Lorenz curve, 18, 20, 21
Generalized Lorenz dominance, 18, 21
Genetic distance, 79
Geographical concentration, 110
Geometric mean, 22, 29
Gini distance function, 25
Gini index, 911, 13, 15, 17, 25, 27, 28, 30,
42, 43, 49, 56, 60, 61, 63, 67, 73, 75,
105, 116, 119
Gini welfare function, 23, 51
H
Harmonic mean, 22
Health bipolarization, 98
Health inequality, 98
Herfindahl-Hirschman index, 112
Homothetic, 21, 25, 51
Hybrid indices, 64
I
Ideal policy, 117
Identification, 5355, 57, 58, 60, 61, 63,
64, 67, 68, 7075, 78, 79, 87, 88,
98, 105107, 118
Illfare ranked permutation, 23
Income bipolarization, 3740
Income multi-polar polarization, 33, 5362
Income space, 35
Incomplete ordering, 94
Increased bipolarity, 3739, 70, 72, 98100,
102104
Increased spread, 3739, 59, 72, 99, 100
Increasing function, 22, 44, 54, 63, 72, 88,
119, 120
Incremental improvement condition, 25
Independence property, 44, 75
Index of dominance, 79
Index of equality, 22, 61
Inequality sensitive welfare, 45
Index
Influence function, 84, 86
Intensity of conflict, 92
Intergroup distance, 79
Intermediate bipolarization curve, 49, 50
Intermediate invariance, 5, 6
International relations, 53
Intra-group cohesion, 119
Inverse distribution function, 14
Isoelastic, 117, 119
Iso-polarization contour, 74
K
Kolmogorov distance, 61
Kolm-Pollak index, 31, 66
Kurtosis, 75
L
Laborers, 109
Landowners, 109
Language divisions, 79
Lateral mass, 55
Leftist concept, 4
Linear homogenous, 21, 22, 50
Line of equality, 13, 67
Line of symmetry, 36
Linguistic distance, 79
Lorenz curve, 1315, 1821, 36, 67
Lorenz dominance, 14, 15, 18, 21, 42
Lorenz ratio, 13
M
Marginal polarization, 84
Mass killing, 115
Maximal inequality, 13
Maximin distance function, 25
Maximin index, 11, 12, 22, 44
Maximum bipolarization, 103104
Mean deviation about the mean, 44
Mean deviation about the median, 44, 48
Median category, 98102, 104, 105, 107, 108
Median income, 33, 35, 4042, 44, 45, 47, 63,
66, 98
Median preserving spread, 39, 100, 107
Middle class, 3337, 56, 110
Minimal increasingness, 23
Minimum bipolarization, 102104
Minimum concentration, 94
Minimum fractionalization, 113
Multidimensional indices, 64
135
N
Normalization, 9, 21, 40, 44, 50, 56, 72
Normative approach, 50
O
Onto function, 65
Ordinal distribution, 99
Ordinal polarization, 107
Ordinal variable, 97, 98, 101, 103,
105, 106
Output sharing, 109
Overlap, 62, 107
P
Pareto absolute dominance, 16, 21
Pareto rank dominance, 16, 17, 21, 39
Partially merged profiles, 94
Partial means, 12
Per capita conflict, 117, 119
Perfect bipolarity, 82
Perfect fractionalization, 114
Perfectly bimodal, 42
Perfectly homogenous, 82, 87
Perfectly overlapping, 107
Piecewise linear, 67
Pigou-Dalton matrix, 8, 9, 20
Pigou-Dalton transfer principle, 8, 70
Polarization ordering, 42, 64, 70, 71, 79, 80,
9195
Polarization proper, 33
Polarization sensitivity, 57, 90, 107
Pole, 53, 58, 59
Political polarization, 77
Population concentration, 112
Population concentration curve, 9395
Population homogeneity, 88, 89, 91
Population replication invariant, 3, 7, 16,
1820, 29, 30
Population space, 35, 36
Population-wise ethically ideal
distribution, 31
Positional transfer sensitivity, 12
Power disparity, 53
P-power, 119
Private good, 116
Prize of power, 119
Progressive transfer, 3, 8, 9, 12, 19, 20, 3841,
43, 51, 65, 67, 70, 98
Public good, 1, 111, 116, 117, 119
Publicness of service, 119
136
Q
Quantile function, 14, 36
Quasi-ordering, 14, 18, 41, 70, 98, 108
Quasi-relation, 46
R
Radicalism, 79, 8791
Rank preserving increment, 72
Rank preserving transfer, 8
Ratio form indices, 64, 75
Ratio scale variable, 97, 98
Rawlsian maximin welfare function, 23
Reduced-form index, 65, 66, 68
Reduced-form welfare function, 22, 63
Reference income, 34
Regressive transfer, 3, 74
Regularity conditions, 21
Relative bipolarization curve, 42, 4549
Relative index, 4, 21, 22, 31, 34, 42, 43, 49, 50
Religious beliefs, 77
Religious divisions, 79
Resemblance factor, 78
Residual term, 28
Resource cost, 117
Reverse welfare, 45
Ricardian theory, 109
Rightist concept, 4, 6
Risk of genocide, 79
S
Scale improvement condition, 23
Scale invariant index, 4
S-concave, 2, 3, 15, 1921, 42, 45, 50, 51
S-convex, 3, 22, 36, 42
Second order stochastic dominance,
19, 39, 98
Self assessed health status, 98
S-Gini index, 11, 63
S-Gini polarization index, 73
Siginificantly sized groups, 92
Simple increment, 3, 16, 19
Single-series Gini index, 11
Small-subgroups property, 113
Smoothed distribution, 28
Social bitterness, 90
Social capital, 78, 115
Social characteristic, 77, 83, 95
Social cohesion, 97, 119
Social conflict, 33, 57, 79, 110
Social distribution, 82, 84, 85, 9194
Social heterogeneity, 87
Social modernization, 34
Index
Social polarization, 7795
Social tension, 1, 34, 77, 111
Social welfare function, 15, 16, 19, 20, 2225,
50, 51, 75, 76
Socioeconomic indices, 64
Socioeconomic polarization, 64, 107
Spread of the distribution, 38
Squeeze, 58, 59
STH. See Strong homogeneity (STH)
Stochastic dominance, 1, 1719, 98, 107
Strictly concave, 2, 19, 20, 31, 4143, 48, 74
Strong bipolarity (STB), 104
Strong homogeneity (STH), 103105
Strong Pareto principle, 16, 21, 25
Structural conditions, 76
Subgroup, 1, 2631, 33, 34, 3739, 42, 43,
45, 53, 54, 5970, 7493, 95, 98, 102,
105107, 110119
Subgroup decomposability, 26, 28, 29
Subgroup decomposable index, 1
Subgroup decomposable inequality index,
26, 63, 65, 76
Subgroup replication, 114
Subgroup-wise ethically ideal distribution,
30, 31
Sub-indices, 31
Submajorized, 41
Symmetric distribution, 36, 37, 44, 58, 59
Symmetric function, 54
Symmetric mean, 22, 28, 29, 43, 51, 65
Symmetry, 3, 8, 36, 38, 44, 45, 59, 83
T
Theil entropy index, 27
Theil mean logarithmic deviation index,
65, 66
Theory of distribution, 109
Third order stochastic dominance, 19
Three-component densities, 34
Tolerance limit, 71
Trade-off, 6, 22, 63
Transitive ordering, 14, 41, 108
Translatable, 4, 24, 25, 51
Translation invariant, 69
Twin-peaks hypothesis, 34
Two-components hypothesis, 34
U
Unit consistency, 7, 21, 30, 50
Unit translatable, 24, 51
Utilitarian rule, 16, 1820
Utility loss, 119
Index
V
Variance, 25, 29, 30, 58, 65, 66, 70
Variance of logarithm, 29
Voting game, 119
W
Weak Pareto principle, 16
Welfare ranked permutation, 10, 23
Winning camp, 116, 117, 119
Within-group alienation, 61, 90
137
Within-group Gini index, 67
Within-group inequality, 1, 26, 31, 63, 65,
67, 70
Within-group Lorenz curve, 67
Within-group radicalism, 89
Within-group term, 26, 63
Z
Zero-frequency independence, 82, 83

Bok:978 81 322 2166 1 PDF

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Bok:978 81 322 2166 1 PDF

Загружено:

Авторское право:

Доступные форматы

Economic Studies in Inequality, Social Exclusion

Economic Studies in Inequality, Social Exclusion

More information about this series at http://www.springer.com/series/7140

It is now well-known that human well-being is a multidimensional phenomenon.

Measuring Income Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . .

On the Measurement of Income Bipolarization . . . . . . . . . . . . . . . .

Measurement of Income Multipolar Polarization . . . . . . . . . . . . .

Reduced-Form Indices of Income Polarization . . . . . . . . . . . . . . .

Measuring Polarization for a Dimension of Human

Fractionalization, Polarization, and Conflict . . . . . . . . . . . . . . . . . .

Glossary of Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

About the Author

Satya R. Chakravarty is a Professor of Economics at the Indian Statistical

Measuring Income Inequality

Inequality in an income distribution in a society delineates disparities of incomes

Springer India 2015

Measuring Income Inequality

For a population of size n, an income distribution is represented by a vector x

be the illfare-ranked or nondecreasingly ordered permutation of x, that is, x^ 1 x^ 2

Measuring Income Inequality

The population replication invariance property enables us to view the functional

Postulates for an Index of Inequality

In this section, we present alternative postulates for an index of inequality and

where c is a scalar such that x c1n 2 Dn . Such an index is called translation

1.3 Postulates for an Index of Inequality

where , 0 1, is a parameter which reflects

Fig. 1.1 Different inequality concepts

Measuring Income Inequality

1.3 Postulates for an Index of Inequality

1, it represents the relative notion of inequality. The Krtscha (1994)

Measuring Income Inequality

1.3 Postulates for an Index of Inequality

order 3. (In the literature, it is also known as a T-transformation; see Marshall

Measuring Income Inequality

1997; Chakravarty 2009a). Equivalence between 1 n2 1x

1.3 Postulates for an Index of Inequality

The inequality restriction > 1 is necessary and sufficient for I DW

2, I DW coincides with the Gini index. As increases, more weight is assigned to

Measuring Income Inequality

It satisfies PDT and a stronger redistributive principle known as the principle of

and, for all n 2 N, x 2 Dn ,

1.3 Postulates for an Index of Inequality

the share of the total income

population proportion, is the ordinate of the Lorenz

Measuring Income Inequality

cumulative income distribution function F : 0; 1 ! 0; 1. For any v 2 0; 1,

total income enjoyed by the poorest t proportion of the population:

1.4 Ethical Approaches to the Measurement of Inequality

Cumulative population shares

Fig. 1.2 Lorenz curve and Lorenz dominance

Ethical Approaches to the Measurement of Inequality

and , respectively. In fact, x Le

Measuring Income Inequality

strong Pareto principle if W x > W y where x is obtained from y 2 Dn by a simple

In contrast, W : D ! R1 satisfies the weak Pareto principle if W x > W y, whenever each

1.4 Ethical Approaches to the Measurement of Inequality

Fr1 vdv for all s 2 0; 1;

where r 2 is an integer and F1 s Fs. For all n 2 N, x 2 Dn , the discrete

F1 s 23, F2 s s23 s3 2s5

Measuring Income Inequality

for all utility functions U : 0; 1 ! R1 such that 1i1 Ui > 0, where

Fig. 1.3 Generalized Lorenz curve and generalized Lorenz dominance

1.4 Ethical Approaches to the Measurement of Inequality

cumulative income distribution function F : 0; 1 ! 0; 1. For any v 2 0; 1,