Professional Documents
Culture Documents
A Unified Approach to
Measuring Poverty and
Inequality
Theory and Practice
James Foster
Suman Seth
Michael Lokshin
Zurab Sajaia
© 2013 International Bank for Reconstruction and Development / The World Bank
1818 H Street NW
Washington, DC 20433
Telephone: 202-473-1000
Internet: www.worldbank.org
1 2 3 4 16 15 14 13
This work is a product of the staff of The World Bank with external contributions. Note that The World Bank does not necessarily own
each component of the content included in the work. The World Bank therefore does not warrant that the use of the content contained in
the work will not infringe on the rights of third parties. The risk of claims resulting from such infringement rests solely with you.
The findings, interpretations, and conclusions expressed in this work do not necessarily reflect the views of The World Bank, its Board
of Executive Directors, or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this
work. The boundaries, colors, denominations, and other information shown on any map in this work do not imply any judgment on the
part of The World Bank concerning the legal status of any territory or the endorsement or acceptance of such boundaries.
Nothing herein shall constitute or be considered to be a limitation upon or waiver of the privileges and immunities of The
World Bank, all of which are specifically reserved.
This work is available under the Creative Commons Attribution 3.0 Unported license (CC BY 3.0) http://creativecommons.org/licenses/
by/3.0. Under the Creative Commons Attribution license, you are free to copy, distribute, transmit, and adapt this work, including for
commercial purposes, under the following conditions:
Attribution—Please cite the work as follows: Foster, James, Suman Seth, Michael Lokshin, and Zurab Sajaia. 2013. A Unified Approach to
Measuring Poverty and Inequality: Theory and Practice. Washington, DC: World Bank. doi: 10.1596/978-0-8213-8461-9 License: Creative
Commons Attribution CC BY 3.0
Translations—If you create a translation of this work, please add the following disclaimer along with the attribution: This translation was
not created by The World Bank and should not be considered an official World Bank translation. The World Bank shall not be liable for any content
or error in this translation.
All queries on rights and licenses should be addressed to the Office of the Publisher, The World Bank, 1818 H Street NW, Washington,
DC 20433, USA; fax: 202-522-2625; e-mail: pubrights@worldbank.org.
Foreword .................................................................................................... xi
Preface ....................................................................................................... xv
Chapter 1
Introduction ................................................................................................ 1
The Income Variable ..........................................................................................4
The Data..............................................................................................................4
Income Standards and Size .................................................................................5
Inequality Measures and Spread .......................................................................13
Poverty Measures and the Base of the Distribution .........................................26
Note ...................................................................................................................44
References..........................................................................................................44
Chapter 2
Income Standards, Inequality, and Poverty ......................................... 45
Basic Concepts ..................................................................................................49
Income Standards ..............................................................................................54
Inequality Measures ...........................................................................................81
Poverty Measures .............................................................................................105
Exercises ..........................................................................................................144
Notes................................................................................................................149
References........................................................................................................151
v
Contents
Chapter 3
How to Interpret ADePT Results ......................................................... 155
Analysis at the National Level and Rural/Urban Decomposition .................157
Analysis at the Subnational Level ..................................................................170
Poverty Analysis across Other Population Subgroups....................................183
Sensitivity Analyses ........................................................................................199
Dominance Analyses .......................................................................................207
Advanced Analysis..........................................................................................216
Note .................................................................................................................223
Reference .........................................................................................................223
Chapter 4
Frontiers of Poverty Measurement ...................................................... 225
Ultra-Poverty ...................................................................................................225
Hybrid Poverty Lines....................................................................................... 226
Categorical and Ordinal Variables .................................................................228
Chronic Poverty ..............................................................................................229
Multidimensional Poverty ...............................................................................230
Multidimensional Standards ...........................................................................234
Inequality of Opportunity ...............................................................................238
Polarization ......................................................................................................240
References........................................................................................................241
Chapter 5
Getting Started with ADePT ................................................................. 245
Conventions Used in This Chapter ...............................................................246
Installing ADePT ............................................................................................246
Launching ADePT ..........................................................................................247
Overview of the Analysis Procedure............................................................... 248
Specify Datasets ...............................................................................................249
Map Variables..................................................................................................252
Select Tables and Graphs ...............................................................................254
Generate the Report .......................................................................................257
Examine the Output........................................................................................258
Working with Variables ..................................................................................258
Setting Parameters ..........................................................................................264
Working with Projects ....................................................................................264
Adding Standard Errors or Frequencies to Outputs .......................................265
vi
Contents
Figures
vii
Contents
2.16: Poverty Severity Curve and the Squared Gap Measure .......................139
3.1: Probability Density Function of Urban Georgia ..................................157
3.2: Age-Gender Poverty Pyramid ...............................................................198
3.3: Poverty Incidence Curves in Urban Georgia, 2003 and 2006 .............208
3.4: Poverty Deficit Curves in Urban Georgia, 2003 and 2006 ..................209
3.5: Poverty Severity Curves in Rural Georgia, 2003 and 2006 .................211
3.6: Growth Incidence Curve of Georgia between 2003 and 2006 ............212
3.7: Lorenz Curves of Urban Georgia, 2003 and 2006 ................................214
3.8: Standardized General Mean Curves of Georgia, 2003 and 2006 .........216
A.1: The Quantile Functions of Urban Per Capita Expenditure,
Georgia ..................................................................................................287
A.2: Generalized Lorenz Curve of Urban Per Capita Expenditure,
Georgia ..................................................................................................288
A.3: Generalized Mean Curve of Urban Per Capita Expenditure,
Georgia ..................................................................................................290
A.4: Generalized Lorenz Growth Curve for Urban Per Capita
Expenditure, Georgia ............................................................................291
A.5: General Mean Growth Curve of Urban Per Capita Expenditure,
Georgia ..................................................................................................292
Tables
viii
Contents
ix
Foreword
xi
Foreword
can construct macrodata series suitable for empirical analyses. Students can
replicate and check the robustness of published results.
Several recent initiatives have lowered the cost of accessing household
datasets. The goal of this book, then, is to reduce the cost of analyzing data
and sharing findings with interested parties.
This book has two unique aspects. First, the theoretical discussion is
based on a highly accessible, unified treatment of inequality and poverty
in terms of income standards or basic indicators of the overall size of the
income distribution. Examples include the mean, median, and other tradi-
tional ways of summarizing a distribution with one or several representative
indicators. The literature on the measurement of inequality has proliferated
since the 1960s. This book provides an excellent overview of that extensive
literature.
Most poverty measures are built on two pillars. First, the “poverty line”
delineates the income levels that define a poor person, and second, various
measures capture the depths of the incomes of those below the poverty line.
The approach here considers income standards as the basic measurement
building blocks and uses them to construct inequality and poverty measures.
This unified approach provides advantages in interpreting and contrasting
the measures and in understanding the way measures vary over time and
space.
Second, the theoretical presentation is complemented by empirical
examples that ground the discussion, and it provides a practical guide to the
inequality and poverty modules of the ADePT software developed at the
World Bank. By immediately applying the measurement tools, the reader
develops a deeper understanding of what is being measured. A battery of
exercises in chapter 2 also aids the learning process.
The ADePT software enables users to analyze microdata—from sources
such as household surveys—and generate print-ready, standardized tables
and charts. It can also be used to simulate the effect of economic shocks,
farm subsidies, cash transfers, and other policy instruments on poverty,
inequality, and labor. The software automates the analysis, helps minimize
human error, and encourages development of new methods of economic
analysis.
For each run, ADePT produces one output file—containing your selec-
tion of tables and graphs, an optional original data summary, and errors and
notifications—in Microsoft Excel® format. Tables of standard errors and
frequencies can be added to a report, if desired.
xii
Foreword
Kaushik Basu
Senior Vice President and Chief Economist
The World Bank
xiii
Preface
This book is made possible by financial support from the Research Support
Budget of the World Bank, the Knowledge for Change Program (KCP), and
the Rapid Social Response (RSR) Program. The KCP is designed to pro-
mote high-quality, cutting-edge research that creates knowledge to support
policies for poverty reduction and sustainable development. KCP is funded
by the generous contributions of Australia, Canada, China, Denmark, the
European Commission, Finland, France, Japan, the Netherlands, Norway,
Singapore, Sweden, Switzerland, the United Kingdom, ABN AMRO
Bank, and the International Fund for Agricultural Development. RSR is
a multidonor endeavor to help the world’s poorest countries build effec-
tive social protection and labor systems that safeguard poor and vulnerable
people against severe shocks and crises. RSR has been generously supported
by Australia, Norway, the Russian Federation, Sweden, and the United
Kingdom.
James Foster is grateful to the Elliott School of International Affairs
and Dean Michael Brown for facilitating research on global poverty and
international development. The Ultra-poverty Initiative of its Institute
for International Economic Policy (IIEP), spearheaded by its former direc-
tor, Stephen Smith, has been a focal point of these efforts. A major gift to
the Elliott School from an anonymous donor significantly enhanced the
research capacity of IIEP and helped make the present project a reality.
xv
Preface
xvi
Chapter 1
Introduction
What is poverty? At its most general level, poverty is the absence of accept-
able choices across a broad range of important life decisions—a severe lack of
freedom to be or to do what one wants. The inevitable outcome of poverty
is insufficiency and deprivation across many of the facets of a fulfilling life:
1
A Unified Approach to Measuring Poverty and Inequality
2
Chapter 1: Introduction
3
A Unified Approach to Measuring Poverty and Inequality
the appendix includes tables and figures that may be useful in understanding
some of the concepts and examples in the first two chapters.
Our discussion begins with the variable income, which may also represent
consumption expenditure or some other single dimensional outcome vari-
able. Data are typically collected at the household level. So to construct an
income variable at the individual level, one must make certain assumptions
about its allocation within the household. Using these assumptions, house-
hold data are converted into individual data that indicate the equivalent
income level an individual commands, thereby taking account of household
structure and other characteristics.
One simplification is to assume that overall income is spread evenly
across each person in the household. However, many other equivalence scales
can be used. This adjustment enables comparisons to be made symmetri-
cally across people irrespective of household or other characteristics. This
simplification justifies the assumption of symmetry invoked when evaluating
income distributions—whereby switching the (equivalent) income levels
of two people leaves the evaluation unchanged. Additionally, it is assumed
that the resulting variable can be measured with a cardinal scale that allows
comparison of income differences across people.
The Data
4
Chapter 1: Introduction
Given an income distribution, three separate but related aspects are of inter-
est: the distribution’s size, the distribution’s spread, and the distribution’s
base. We will discuss the size issue here. Subsequent sections deal with the
spread and base concepts.
Distribution size is most often indicated by the mean or per capita income.
For the vector representation, the mean is obtained by dividing total income
5
A Unified Approach to Measuring Poverty and Inequality
by the total number of people in the distribution. The mean can also be
viewed as the average height (or, in mathematical terms, the integral) of
the quantile function. It is the income level that all people would achieve if
they were given an equal share of overall resources.
Another size indicator, median income, is the income at the midway point
of the quantile function, with half the incomes below and half above. Most
empirical income distributions are skewed so that the mean (which includes
the largest incomes in the averaging process) exceeds the median income
(which is unaffected by the values of the largest incomes). Still another
measure of size is given by the mean income of the lowest fifth of the popula-
tion, which focuses exclusively on the lower incomes in a distribution. Each
of these indicators is an example of an income standard, which reduces the
overall income distribution to a single income level indicating some aspect
of the distribution’s size.
6
Chapter 1: Introduction
Common Examples
• First are the quantile incomes, such as the income at the 10th per-
centile, the income at the 90th percentile, and the median. Each is
informative about a specific point in the distribution but ignores the
values of the remaining points.
• Next are the (relative) partial means obtained by finding the mean of
the incomes below a specific percentile cutoff (the lower partial means)
or above the cutoff (the upper partial means), such as the mean of the
lowest 20 percent and the mean of the highest 10 percent. Each of
these income standards indicates the size of distribution by focusing
on one or the other side of a given percentile and by measuring the
average income of this range while ignoring the rest. As the cutoff
varies between 0 percent and 100 percent, the lower partial mean
varies between the lowest income and the mean income, whereas the
upper partial mean varies between the mean income and the highest
income.
By focusing on a specific income or a range of incomes, the quantile
incomes and the partial means ignore income changes outside that
range. The remaining two forms of income standard, by contrast, are
monotonic so that the increase in income causes the income standard
to strictly rise.
• The general means take into account all incomes in the distribution,
but emphasize lower or higher incomes depending on the value of
parameter a (that can be any real number). When a is nonzero, the
general mean is found by raising all incomes to the power a, then
by averaging, and finally by taking the result to the power 1/a. This
process of transforming incomes and then untransforming the aver-
age ensures that the income standard is, in fact, measured by income
(or, in income space, as we might say).
In the remaining case of a = 0, the general mean is defined to be
the geometric mean. It is obtained by raising all incomes to the power
1/N, then taking the product. For a < 1, incomes are effectively trans-
formed by a concave function, thus placing greater emphasis on lower
incomes. For a > 1, the transformation is convex, and the general
mean emphasizes higher incomes.
7
A Unified Approach to Measuring Poverty and Inequality
As a varies across all possible values, the general mean rises from
minimum income (as a approaches −∞), to the harmonic mean
(a = −1), the geometric mean (a = 0), the mean (a = 1), the Euclidean
mean (a = 2), up to the maximum income (as a approaches ∞).
General means with a < 0 are only defined for positive incomes.
• The final income standard is a step in the direction of a “maximin”
approach, which evaluates a situation by the condition of the least
advantaged person. The usual mean can be reinterpreted as the
expected value of a single income drawn randomly from the popula-
tion. Suppose that instead of a single income, we were to draw two
incomes randomly from the population (with replacement). If we
then evaluated the pair by the lower of the two incomes, this would
lead to the Sen mean, defined as the expectation of the minimum of
two randomly drawn incomes.
Because we are using the minimum of the two, this number can be no
higher than the mean and is generally lower. Consequently, the Sen mean
also emphasizes lower incomes but in a different way to the general means
with a < 1, the lower partial means, or the quantile incomes below the
median.
Calculating the Sen mean for an income vector is straightforward.
Create an N × N matrix that has a cell for every possible pair of incomes,
and place the lower value of the two incomes in the cell. Add up all the
entries and divide by the number of entries (N2) to obtain their mean,
which is the expected value of the lower income. This mean has close ties
to the well-known Gini coefficient measure of inequality.
Welfare
The general means for a < 1 and the Sen mean are also commonly inter-
preted as measures of welfare. The key additional property that allows this
interpretation is the transfer principle, which requires an income transfer
from one person to another who is richer (or equally rich) to lower the
income standard. In other words, a regressive transfer that does not change
the mean income should lower the income standard.
One way to justify this property begins with a utilitarian symmetric
welfare function that views welfare derived from an income distribution
to be the average level of (indirect) utility in society, where it is assumed
8
Chapter 1: Introduction
Applications
9
A Unified Approach to Measuring Poverty and Inequality
Subgroup Consistency
One motivation for examining several income standards at the same time is
robustness: Do conclusions about the direction of change in the distribution
size using one income standard (say, the mean) hold for others (say, the
nearby generalized means)? A second reason might be focus or an identified
10
Chapter 1: Introduction
concern with different parts of the distribution: Has rapid growth at the top
(say, the 90th percentile income) been matched by growth at the middle
(say, the median) or the bottom (say, the 10th percentile income)?
We can answer questions of this sort by plotting an entire class of income
standards against percentiles of income distribution. We can then use the
curve to determine if a given comparison is unambiguous (one curve is
above the other) or if it is contingent (the curves cross).
A first curve is given by the quantile function itself, which simultane-
ously depicts incomes from lowest to highest. As income standards, quan-
tiles are somewhat partial and insensitive—yet when they all agree that
one distribution is larger than another, this ensures that every other income
standard must follow their collective judgment.
The quantile function represents first-order stochastic dominance, which
also ensures higher welfare according to every utilitarian welfare function
with identical, increasing utility functions. Thus, on the one hand, the
robustness implied by an unambiguous comparison of quantile functions
extends to all income standards and all symmetric welfare functions for
which “more is better.” On the other hand, if some quantiles rise and others
fall, then the resulting curves will cross and the final judgment is contingent
on which income standard is selected. In this case, the quantile function can
be helpful in identifying winning and losing portions of the distribution.
A second curve of this sort is given by the generalized Lorenz curve, which
graphs the area under the quantile function up to each percent p of the
population. It can be shown that the height of the generalized Lorenz curve
at any p is the lower partial mean times p itself. For example, if the lowest
income of a four-person vector were 280, then the generalized Lorenz curve
value (ordinate) at p = 25 percent would be 70.
A comparison of generalized Lorenz curves conveys information on
lower partial means, with a higher generalized Lorenz curve indicating
agreement among all lower partial means. As income standards, the lower
partial means are insensitive to certain increments and income transfers.
Yet when all these income standards are in agreement, it follows that every
monotonic income standard satisfying the transfer principle would abide by
their judgment.
Indeed, the generalized Lorenz curve represents second-order stochas-
tic dominance, which signals higher welfare according to every utilitarian
welfare function with identical and increasing utility function exhibiting
diminishing marginal utility (Atkinson’s general class of welfare functions).
11
A Unified Approach to Measuring Poverty and Inequality
Growth Curves
12
Chapter 1: Introduction
For the quantile function, the resulting growth curve is called the growth
incidence curve. The height of the curve at p = 50 percent gives the growth
rate of the median income. Varying p allows us to examine whether this
growth rate is robust to the choice of income standard or whether the lower
income standards grew at a different rate than the rest.
The generalized Lorenz growth curve indicates how the lower partial means
are changing over time, so that the height of this curve at p = 20 percent is
the rate at which the mean income of the lowest 20 percent of the popula-
tion changed over time. Finally, the general means growth curve plots the
rate of growth of each general mean against the parameter a. When a = 1,
the height of the curve is the usual growth rate of the mean income; a = 0
yields the rate of growth for the geometric mean, and so forth. As we will
see below, each of these growth curves can be of help in understanding the
link between growth and the evolution of inequality over time.
There are two main ways to understand what an income inequality measure
actually gauges. The first way is through the properties it satisfies. The
second makes use of a fundamental link between inequality measures and
income standards. We begin with the first approach.
13
A Unified Approach to Measuring Poverty and Inequality
• The first two are symmetry and population invariance properties, which
are analogous to those defined for income standards. They ensure that
inequality depends entirely on income distribution and not on names
or numbers of income recipients.
• The third is scale invariance (or homogeneity of degree zero), which
requires the inequality measure to be unchanged if all incomes are
scaled up or down by a common factor. This ensures that the inequal-
ity being measured is a purely relative concept and is independent of
the distribution size. In contrast, doubling all incomes will double
distribution size as measured by any income standard, thereby reflect-
ing its respective property of linear homogeneity.
• The final property is the weak transfer principle, which in this context
requires income transfer from one person to another who is richer
(or equally rich) to raise inequality or leave it unchanged. In other
words, a regressive transfer cannot decrease inequality. This is an
intuitive property for inequality measures. It is often presented in a
stronger form, known as the transfer principle, which requires a regres-
sive transfer to (strictly) increase inequality.
The Gini coefficient and the Kuznets ratio satisfy all four basic properties
for inequality measures. The 90/10 ratio satisfies the first three but violates the
weak transfer principle: a regressive transfer between people at the 5th percen-
tile and the 10th percentile can raise the 10th percentile income, thus lowering
inequality as measured by the 90/10 ratio. Although this result does not rule
out the use of the intuitive 90/10 ratio as an inequality measure, it does suggest
that conclusions obtained with this measure should be scrutinized.
The four basic properties define the general requirements for inequality
measures. Additional properties help to discern between acceptable mea-
sures. For example, decomposability and subgroup consistency (discussed in a
later section) are helpful in certain applications. Transfer sensitivity ensures
that an inequality measure is more sensitive to changes in the income dis-
tribution at the lower end of the distribution.
A second way of understanding inequality measures relies on an intui-
tive link between inequality measures and pairs of income standards. The
basic structure is perhaps easiest to see in the extreme case where there are
only two people and, hence, only two incomes. Letting a denote the smaller
14
Chapter 1: Introduction
income of the two, and b denote the larger income, it is natural to measure
inequality by the relative distance between a and b, such as I = (b − a)/b,
or some other increasing function of the ratio b/a. Indeed, scale invariance
and the weak transfer principle essentially require this form for the measure.
Suppose that instead of evaluating the inequality between two people, we
want to measure the inequality between two equal-sized groups. A natural
way of proceeding is to represent each group’s income distribution using an
income standard. This yields a pair of representative incomes—one for each
group—that can then be compared. Where a denotes the smaller of these
two incomes and b the larger, it is natural to measure inequality between the
two groups as I = (b − a)/b, or some other increasing function of the ratio
b/a. For example, if the distributions are the earnings of men and women and
the income standard is the mean, then b/a would be the ratio of the aver-
age income for men to the average income for women—a common indica-
tor of inequality between the two groups. As will be discussed below, this
“between-group” approach is useful in decompositions of inequality by popu-
lation subgroup and also in the measurement of inequality of opportunities.
The general idea that inequality depends on two income standards is also
relevant when evaluating the overall inequality in a population’s distribu-
tion of income. But instead of applying the same income standard to two
distributions, we now apply two income standards to the same distribution.
One of the income standards (the upper standard) places greater weight
on higher incomes, and the second (the lower standard) emphasizes lower
incomes; so for any given income distribution, the lower-income standard’s
value is never larger than the upper-income standard’s value.
This is true, for example, when the lower standard is the geometric mean
and the upper is the arithmetic mean or, alternatively, when the lower is
the 10th percentile income and the upper is the 90th percentile income.
Inequality is then seen as the extent to which the two income standards are
spread apart: where a denotes the lower-income standard and b the upper-
income standard, overall inequality is I = (b − a)/b, or some other increasing
function of the ratio b/a.
Common Examples
Virtually all inequality measures in common use are based on twin income
standards. This is easily seen in the case of the 90/10 ratio, and generalizes to
any quantile ratio b/a, where a corresponds to the income at a percentile p of
15
A Unified Approach to Measuring Poverty and Inequality
The Gini coefficient has as its twin income standards the mean and the
Sen mean and can be written as I = (b − a)/b, where b is the mean and a is
16
Chapter 1: Introduction
the Sen mean. The expected (absolute) difference between two incomes
can be written as (a′ − a), where a′ is the expectation of their maximum and
a is the expectation of their minimum. Because the mean b can be written
as (a′ + a)/2, the difference (b − a) is half of the expected absolute difference
between incomes, which confirms that (b − a)/b is an equivalent definition
of the Gini coefficient. In other words, the Gini coefficient is the extent to
which the Sen mean falls below the mean as a percentage of the mean.
Atkinson’s class of inequality measures also takes the form I = (b − a)/b,
where the upper-income standard b is also the mean, but now the lower-
income standard a is a general mean with parameter a < 1. This income
standard focuses on lower incomes by raising each income to the a power,
averaging across all the transformed incomes, then converting back to
income space by raising the result to the power 1/a. A lower value of the
parameter a yields an income standard that is more sensitive to lower
incomes and is lower in value. This will be reflected in a higher value for
(b − a)/b, so the percentage loss from the mean is seen to be higher.
The final example is the family of generalized entropy measures, whose
definition and properties vary with a parameter a. There are three distinct
ranges for the parameter: a lower range where a < 1, an upper range where
a > 1, and a limiting case where a = 1.
When a < 1, the generalized entropy measures evaluate inequality in
the same way as the Atkinson class of inequality measures (and, in fact, are
monotonic transformations). For example, when a = 0, the measure is the
mean log deviation or Theil’s second measure given by ln(b/a), where b is the
arithmetic mean and a is the geometric mean. Atkinson’s version is (b − a)/b.
Over the second range where a > 1, the general mean places greater
weight on higher incomes and yields a representative income that is typi-
cally higher than the mean. If all incomes were equal, the general mean
and the mean would be equal. However, when incomes are unequal, the
general mean will rise above the mean. The extent to which this occurs
is used by the measure to evaluate inequality. For example, the inequality
measure obtained when a = 2 is (half) the squared coefficient of variation, that
is, one-half of the variance over the squared mean. The general mean in this
case is the Euclidean mean, which first squares all incomes, then averages
the transformed incomes, and finally returns to income space by taking the
square root. The Euclidean mean and the mean of the two-income distribu-
tion (4, 4) are both 4. Altering the distribution to (1, 7) raises the Euclidean
mean to 5 but leaves its mean at 4.
17
A Unified Approach to Measuring Poverty and Inequality
The final case where a = 1 leads to Theil’s first measure, which is one of
the few inequality measures without a natural twin standards representation,
but is, in fact, a limit of such measures.
Applications
Inequality measures are used to assess the extent to which incomes are
spread apart in a country or region and the way this level changes over time
and space. Of particular interest is the interplay between a population’s aver-
age prosperity, as represented by the mean income, and the income distribu-
tion, as represented by an inequality measure. The positive achievement of
18
Chapter 1: Introduction
a high per capita income can be viewed less favorably if inequality is high,
too. The combined effect on welfare can be evaluated using an inequality-
adjusted mean.
The Kuznets hypothesis postulates that growth in per capita income ini-
tially comes at a cost of a higher level of inequality, but eventually inequal-
ity falls with growth. The resulting Kuznets curve, which depicts per capita
income on the horizontal axis and inequality on the vertical axis, has the
shape of an inverted U. If the hypothesis were true, then a rapidly grow-
ing developing country could have only moderate welfare improvements,
whereas a moderately growing developed country could experience rapid
improvements in welfare, all because of the changing levels of inequality.
An alternative view takes the initial level of inequality as one of the
determinants of income growth. For example, greater inequality might lead
to a higher average savings rate if the richer groups have a greater propen-
sity to save, and this can positively influence long-term growth. Conversely,
high inequality might create political pressure to raise the marginal tax rate
on the rich, which could diminish incentives to invest and grow. These
applications of inequality measures view inequality as a valuable macro
indicator of the health of a country’s economy that influences and is affected
by other macro variables.
Other applications try to assess the origins of inequality in the micro
economy. Could inequality in earned incomes be due to (a) a high return
to education, (b) a decline in union power, (c) increased competition from
abroad, (d) discrimination, or (e) demographic changes such as increased
female labor force participation? Mincer (1974) equations can help trace
earnings inequality to the underlying characteristics of the labor force,
including the level and distribution of human capital. Oaxaca decomposi-
tions (1973) test for the presence of discrimination by sex, race, or other
characteristics and have been adapted to evaluate the contribution of demo-
graphic changes to observed earnings inequality.
Depending on the policy question, it may make sense to move from
an overall inequality measure (that evaluates the spread across the entire
distribution) to a group-based inequality measure (that compares the mean
or other income standard across several groups). The latter, more limited,
notion of inequality can often have greater significance, particularly if
the underlying groups are easy to understand and have social or political
salience. Examples include racial, sex, and ethnic inequality, or the inequal-
ity between urban and rural areas.
19
A Unified Approach to Measuring Poverty and Inequality
20
Chapter 1: Introduction
incomes of the giver and the receiver). Transfer sensitive measures include
the Atkinson family of measures, Theil’s two measures, and the “lower half”
of the generalized entropy measures with a < 2.
Note that the coefficient of variation (a monotonic transformation of
the generalized entropy measure with a = 2) is transfer neutral in that a
given transfer has the same equalizing effect up and down the distribution:
a one-unit transfer of income between two rich people has the same effect
on inequality as does a one-unit transfer of income between two poor people
the same initial income distance apart. The upper half of the generalized
entropy measures with a > 2 focuses on inequality among upper incomes.
The Gini coefficient is often considered to be most sensitive to changes
involving incomes at the middle, but this is not entirely accurate. The effect
of a given-sized transfer on the Gini coefficient depends on the number of
people between giver and receiver, not on their respective income levels.
Because, empirically, there tend to be more observations bunched together
in the middle of the distribution, the effect of a transfer near the middle
tends to be larger.
Although the variance is not itself a measure of relative inequality (it vio-
lates scale invariance and focuses on absolute differences), the analysis of
variance (ANOVA) provides a natural model for decomposition of inequal-
ity measures into a within-group and a between-group term. The motivating
question here is given a collection of population subgroups, how much of
the overall inequality can be attributed to inequality within the subgroups,
and how much can be attributed to inequality across the subgroups.
Answers to this type of question become feasible when an inequality mea-
sure is additively decomposable, in which case the within-group inequality term
is expressible as a weighted sum of the inequality levels within the groups, the
between-group term is the inequality measure applied to the smoothed distri-
bution, and the overall inequality level is just the sum of the within-group and
between-group terms. The contributions of within-group and between-group
inequality (within-group inequality divided by total inequality and between-
group inequality divided by total inequality, respectively) will sum to one.
Decomposition analysis can help clarify the structure of income inequal-
ity across a population. It can identify regions or sectors of the economy
that disproportionally contribute to inequality. And when the subgroups are
21
A Unified Approach to Measuring Poverty and Inequality
22
Chapter 1: Introduction
23
A Unified Approach to Measuring Poverty and Inequality
24
Chapter 1: Introduction
25
A Unified Approach to Measuring Poverty and Inequality
The final aspect examined here is the base or the bottom of the income
distribution and the main topic of this book: poverty. Evaluation of poverty
begins with an identification step in which the people considered poor are
specified and continues with an aggregation step in which the data of the
poor are combined to obtain a numerical measure. These two steps make up
a methodology for measuring poverty in an income distribution.
The identification step is usually accomplished by selecting a level of
income, called the poverty line, below which a person in a given distribution
is considered poor. In its most general formulation, a poverty line is specified
for every possible income distribution, so that the set of poor people in a pop-
ulation depends on the prevailing living conditions. Finding a proper func-
tional relation between poverty line and income distribution is, of course,
a challenging problem, and one that is subject to much controversy.
Most evaluations of poverty have settled on two very simple approaches:
(a) an absolute approach that takes the poverty line to be a constant and
26
Chapter 1: Introduction
(b) a relative approach that takes the poverty line to be a constant fraction
of an income standard.
An absolute poverty line is a fixed cutoff that does not change as the distribu-
tion being evaluated changes. Examples include the following:
27
A Unified Approach to Measuring Poverty and Inequality
However, there are some practical challenges associated with the con-
struction of absolute poverty lines:
28
Chapter 1: Introduction
29
A Unified Approach to Measuring Poverty and Inequality
• The first two are the symmetry and population invariance properties given
above for income standards and inequality measures. They are impor-
tant for ensuring that the measure is based on the anonymous distribu-
tion and not on the income recipients’ names or the population size.
• The third basic property is the focus axiom, which requires the pov-
erty measure to ignore changes in the distribution involving nonpoor
incomes. This approach ensures that the measure focuses on poor
incomes in evaluating poverty.
30
Chapter 1: Introduction
The first four properties are invariance properties, which indicate how
various changes in the distribution should not be taken into account by the
measure. The next two properties are dominance properties that require the
measure to be consistent with certain basic changes in the distribution.
Notice that both the monotonicity axiom and the transfer principle
allow the number of poor to be altered in the process, whereas the weaker
versions do not.
The headcount ratio, the poverty gap measure, and the FGT index satisfy
all six basic axioms. The headcount ratio satisfies weak monotonicity and
the weak transfer principle (because it is unaffected by the distributional
changes specified in the two properties), but it violates the two stronger
versions. The poverty gap measure satisfies the monotonicity axiom, but it
violates the transfer principle (because it is unaffected by a small regressive
transfer). The FGT index satisfies both stronger axioms.
Some additional properties can also be helpful in evaluating poverty
measures. Transfer sensitivity requires a decrement in the income of a poor
person, when combined with an equal-sized increment in the income of a
richer poor person, to raise poverty. It ensures that a given-sized transfer has
a larger poverty-reducing effect at lower poor incomes. Decomposability and
31
A Unified Approach to Measuring Poverty and Inequality
Income Standards
32
Chapter 1: Introduction
discuss several other poverty measures that share this structure but use dif-
ferent income standards in constructing the gap standard.
Common Examples
The first general form of poverty measures uses an income standard applied
to the censored distribution. An income standard that puts progressively
greater weight on lower incomes will yield a poverty measure that is sensi-
tive to the distribution of income among the poor. The Sen-Shorrocks-Thon
(SST) index is given by (b − a)/b, where a is the Sen mean applied to x*and
b is the poverty line. This measure inherits its characteristics from the Sen
mean: it satisfies all six basic properties and monotonicity and the transfer
property. Increments and progressive transfers among the poor are reflected
in a strictly higher poor income standard a, and hence a lower poverty level.
The next measure is based on another income standard that emphasizes
lower incomes. The Watts index is defined as ln(b/a), where a is the geomet-
ric mean applied to the censored distribution and b is the poverty line z. It
likewise satisfies the six basic axioms and the strict forms of monotonicity
and the transfer principle. Additionally, the geometric mean has the prop-
erty that a given-sized transfer among the poor has a greater effect at lower
income levels, so the poverty measure satisfies transfer sensitivity.
The Watts index can be expanded to an entire class of measures, each
of which uses a general mean to evaluate the censored distribution. The
Clark-Hemming-Ulph-Chakravarty (CHUC) family of indices compares the
poor income standard a = ma (x*) for a ≤ 1 and the poverty line b = z. There
are two forms of the measure: the original form (b − a)/b and a decompos-
able form obtained by a simple transformation. The measure becomes the
poverty gap at a = 1 and the Watts index (or a transformation) at a = 0.
The properties of the general means ensure that the CHUC measures satisfy
all six basic properties for poverty measures, for monotonicity, and for a < 1
the transfer principle as well as transfer sensitivity.
The second general form of poverty measures uses an income stan-
dard applied to the gap distribution. The key family of measures has a
traditional decomposable version and an alternative version that is only
subgroup consistent.
The FGT family of decomposable poverty indices was defined above
as the mean of the a-gap distribution and includes the headcount ratio for
33
A Unified Approach to Measuring Poverty and Inequality
a = 0, the poverty gap measure for a =1, and the FGT or squared gap mea-
sure for a = 2. Alternatively, we can transform each of the measures in the
range a > 0 by raising it to the power 1/a. This yields a subgroup-consistent
measure that compares a gap standard a = ma (g*) to the poverty line b = z
via the formula P = a/b.
The properties for the FGT measures in this range follow from the prop-
erties of the associated general means. The first five properties and mono-
tonicity are immediately satisfied for all a > 0. For the transfer principles,
note that the general means with a < 1 emphasize the smaller entries, those
with a > 1 emphasize the larger entries, and a = 1 ignores the distribution
altogether. Thus, the FGT measures satisfy the weak transfer principle for
a ≥ 1 and the transfer principle for a > 1. In an analogous way, the FGT
index for a = 2 is transfer neutral in that a given-sized progressive transfer
among the poor has the same effect at lower incomes, whereas the FGT
measures with a > 2 satisfy transfer sensitivity.
The above discussion excludes the case a = 0, which corresponds to
the headcount ratio. The simple structure of this poverty measure does not
admit an interpretation of an income standard applied to the censored or
gap distribution. Instead, a second censoring must be applied to obtain a
distribution in which all nonpoor incomes are replaced by z and all poor
incomes are replaced by 0. Let x** denote the resulting doubly censored dis-
tribution. The headcount ratio can be represented as (b − a)/b, where a =
μ(x**) and b = z. In other words, it is the poverty gap of the doubly censored
distribution that converts nonpoor incomes to z and poor incomes to 0.
The first censoring ensures that the measure focuses on incomes of the
poor. The second censoring forces the headcount ratio to ignore the actual
income levels of poor people and violate monotonicity. The headcount
ratio suppresses information that is relevant to poverty (the actual incomes
of the poor) in order to capture one key aspect of poverty (the prevalence
of poverty). Replacing x** with x* in this representation would recover this
information and yield the poverty gap measure.
Poverty measures satisfying the transfer principle are called distribution sensi-
tive because they account for the inequality of poor incomes in ways that
the headcount ratio or the poverty gap cannot. In fact, each of the above
distribution-sensitive poverty indices is built on a specific income or gap
34
Chapter 1: Introduction
standard that is closely linked to an inequality measure. For the SST index,
it is the Gini coefficient. For the CHUC indices, the Atkinson measures are
used. For the Watts index, the mean log deviation is the inequality measure.
In each case, the inequality measure is applied to the censored distribution
x* with greater censored inequality being reflected in a higher level of poverty
(for a given poverty gap level).
The FGT measures (for a > 1) use generalized entropy measures applied
to the gap distribution g* with greater gap inequality leading to a higher
level of poverty (for a given level of the poverty gap). The focused inequal-
ity measures underlying these distribution-sensitive poverty indices ignore
variations in incomes above the poverty line. Trends in focused inequality
may well be very different from trends in overall inequality.
Certain income standards can be viewed as welfare functions, and this
link can provide yet another lens for interpreting poverty measures. The Sen
mean underlying the SST index and the general means for a ≤ 1 that are
behind the CHUC indices can be interpreted as welfare functions. In each
case, the welfare function is applied to the censored distribution to obtain
the poor income standard a, which is now seen to be a censored welfare func-
tion that takes into account the incomes of the poor and only part of the
incomes of the nonpoor (up to the poverty line).
For these measures, poverty and censored welfare are inversely related.
Every increase in poverty is seen as a decrease in censored welfare. Of
course, the trends in censored welfare may be very different from the trends
in overall welfare, as the latter take into account the actual incomes of the
nonpoor. We will see below another link between welfare and poverty when
we consider poverty comparisons over a range of lines.
Applications
• Who are the poor and how do they differ from the nonpoor? A range
of characteristics can be examined—including location, household
size, ethnicity, education indicators, health indicators, housing, and
35
A Unified Approach to Measuring Poverty and Inequality
36
Chapter 1: Introduction
Many programs designed to address the needs of the poor are implemented
at the local level. Suppose we are evaluating such a program in a country
with two equal-sized regions. We find that poverty has fallen significantly
in each region, yet when poverty is measured at the country level, it has
increased. This possibility could present significant challenges to the analyst
and could prove rather difficult to explain to policy makers. It turns out that
the inconsistency between regional and national poverty outcomes may be
due entirely to the way poverty is measured.
To ensure that this possibility does not arise, one can require the poverty
measure to satisfy subgroup consistency. This property requires that if poverty
falls in one subgroup and is unchanged in another and both have fixed popu-
lation sizes, then the overall poverty level must likewise fall. The SST index
is not subgroup consistent because of its use of the Sen mean. The FGT and
CHUC measures, which depend on general means, are subgroup consistent
and thus would not be subject to the regional-national dilemma.
Subgroup consistency requires overall poverty to move in the same
direction as an unambiguous change in subgroup poverty levels. A stronger
property provides an explicit formula that makes the link between overall
and subgroup poverty. A poverty measure is said to be (additively) decompos-
able if overall poverty is a population-share weighted average of subgroup
poverty levels. Unlike the case of inequality measures, there is no between-
group term in this decomposition. The reason is that the standard against
which subgroup poverty is evaluated is a fixed poverty line. In contrast, an
inequality measure typically evaluates subgroup inequality relative to sub-
group means, then takes the variation of subgroup means into account as
another source of inequality.
Additively decomposable poverty measures transparently link subgroup
poverty to overall poverty. This approach can be particularly useful in
generating a coherent poverty profile in which a broad array of population
subgroups and their poverty levels can be broken down or reassembled as
needed. Consider these questions:
37
A Unified Approach to Measuring Poverty and Inequality
38
Chapter 1: Introduction
curve for the headcount ratio traces the cumulative distribution function
associated with the distribution (except that it takes its limits from the left
rather than the right when it has jumps), so the poverty ordering is first-
order stochastic dominance.
If one recalls the above discussion of stochastic dominance, this poverty
ordering is equivalent to having a higher quantile function and also to
having greater welfare according to every utilitarian welfare function with
identical, increasing utility functions. The poverty curve associated with the
headcount ratio is often called the poverty incidence curve.
The poverty curve for the poverty gap measure is closely linked to the
area beneath (or the integral of) the poverty incidence curve (or the cdf),
which is another way of representing second-order stochastic dominance.
Hence, the poverty ordering for the poverty gap measure is simply second-
order stochastic dominance. By the previous discussion, this means that the
poverty ordering can also be represented by the generalized Lorenz curve,
with a higher generalized Lorenz curve indicating unambiguously lower (or
no higher) poverty according to the poverty gap measure.
In addition, there is a useful welfare interpretation of this poverty order-
ing: it indicates higher welfare according to every utilitarian welfare func-
tion with identical and increasing utility function exhibiting diminishing
marginal utility (Atkinson’s general class of welfare functions). The curve
found by plotting the area beneath the poverty incidence curve for each
income level z is often called the poverty deficit curve.
The FGT index has a poverty curve that is closely linked with the area
beneath the poverty deficit curve (or the double integral of the cdf), and
hence its poverty ordering is linked to a refinement of second-order stochas-
tic dominance called third-order stochastic dominance. This poverty ordering
also has a welfare interpretation: higher welfare according to every utilitar-
ian welfare function with identical and increasing utility function exhibit-
ing diminishing and convex marginal utility.
The final condition on the convexity of marginal utility ensures that
the welfare function is more sensitive to transfers at the lower end of the
distribution—a welfare version of the transfer sensitivity axiom. The curve
found by plotting the area beneath the poverty deficit curve for each income
level z is often called the poverty severity curve.
Notice that the poverty orderings for the three FGT measures are nested
in that if the headcount ratio’s ordering ranks two distributions, then the
poverty gap’s ordering also ranks the distributions in the same way (but not
39
A Unified Approach to Measuring Poverty and Inequality
vice versa). Further, the poverty gap’s ordering implies (but is not implied
by) the FGT index’s ordering. Because the poverty deficit curve is found by
taking the area under the poverty incidence curve, a higher poverty inci-
dence curve leads to a higher poverty deficit curve. The same is true for the
poverty deficit and poverty severity curves.
The poverty orderings of the Watts and CHUC indices can also be eas-
ily constructed and lead to another nested set starting with second-order
dominance for the poverty gap measure. The poverty ordering for the Watts
index, for example, is simply generalized Lorenz (or second-order stochastic)
dominance applied to the distributions of log incomes. Each CHUC poverty
ordering likewise applies generalized Lorenz dominance to distributions of
transformed incomes (see Foster and Jin 1998).
Placing an upper limit z* on the range of poverty lines is equivalent
to comparing poverty curves (or the poverty incidence, deficit, or severity
curves) over this limited range or to using censored distributions associated
with z*. For example, the limited range poverty ordering for the poverty gap
is equivalent to comparing the generalized Lorenz curves of the censored
distributions or to comparing censored welfare levels across all utilitarian
welfare functions with identical and increasing utility functions that have
diminishing marginal utility.
In the above example, we varied the poverty line while holding the
poverty measure fixed. We can also vary the poverty measure for a given
poverty line to examine robustness to the choice of measure. For example,
using a five-dimensional vector, one can depict the poverty levels of the
FGT measures for a = 0, 1, and 2; the Watts index; and the SST index.
Vector dominance would then be interpreted as a variable measure poverty
ordering that ranks distributions when all five measures unanimously agree.
An analogous approach using poverty curves can be employed when
using poverty measures indexed by a parameter. Consider a poverty curve
that depicts the CHUC indices (z − ma (x*))/z for a ≤ 1 and the FGT indices
ma (g*))/z for a ≥ 1. We are using the income standard version of each mea-
sure (rather than the decomposable version) because of its nice interpreta-
tion as a normalized average gap. The poverty measure at a = 1 is the usual
poverty gap measure. As a rises, the FGT values progressively rise because
the measures with higher a use a general mean that focuses on the higher
gaps in the gap vector g*.
The extent to which poverty rises as a > 1 rises depends on the gen-
eralized entropy inequality in g* for a. To the left, the CHUC values
40
Chapter 1: Introduction
progressively rise as the measures with lower a use a general mean that
focuses on lower incomes in the censored vector x*. The extent to which
poverty rises as a < 1 falls depends on the generalized entropy inequality in x*
for a. A higher curve would then be interpreted as the variable measure pov-
erty ordering that ranks distributions when all these poverty measures agree.
The above approaches to varying the poverty line and the poverty mea-
sure can be combined to examine the robustness of comparisons to changing
both simultaneously. Interestingly, though, in certain cases it is enough to
examine a variable line poverty ordering. For example, if two distributions
can be ranked by the poverty ordering of the headcount ratio, then they will
also be ranked in the same way by the poverty ordering associated with any
given poverty measure satisfying the basic axioms and monotonicity. This is
also true for certain limited range poverty orderings.
Even in cases lacking a clear ranking for the relevant set of poverty lines
(or measures), a poverty curve can be very useful in identifying ranges of
poverty lines (and measures) where the ranking is unchanged and where the
ranking reverses. This general methodology for checking the robustness of
poverty comparisons is quite powerful.
41
A Unified Approach to Measuring Poverty and Inequality
42
Chapter 1: Introduction
43
A Unified Approach to Measuring Poverty and Inequality
Note
1. The third step may have two substeps, depending on the type of poverty
measure selected: (a) evaluation of individual poverty and (b) selection
of a method to aggregate individual poverty to obtain overall poverty.
References
44
Chapter 2
45
A Unified Approach to Measuring Poverty and Inequality
In this book, we define the space for evaluating poverty as money metric
and single dimensional. The welfare indicator is either consumption expen-
diture or income:
46
Chapter 2: Income Standards, Inequality, and Poverty
which may be a month or any longer period of time. If the reference period
is short (for example, one month), seasonality concerns may be overcome,
but a shorter reference period may also lead to more noise in the expenditure
data. Noise can be avoided by using a longer reference period, but difficulties
in recollection may bias expenditures downward.1
A person may consume many private and public goods from the long
list of commodities in a consumer expenditure survey. For a private good,
total expenditure is the amount of commodity consumed times that com-
modity’s price. Consumption expenditure for two individuals having the
same consumption patterns and requirements, therefore, should be twice the
consumption expenditure for either of the two.
This straightforward expenditure computation may not be possible when
the consumed commodities are, instead, public goods. Given that public
goods are nonrival and nonexcludable, the same amount of public goods
may be consumed by multiple individuals without additional cost. Multiple
individuals living together and sharing public goods enjoy economies of scale.
Examples of public goods include a radio, a water pump, bulk purchase dis-
counts of food items, and food preparation efficiencies (which may lower the
cost of fuel and time).
Although the goal is to construct a money-metric wealth indicator for
each person, fulfilling that goal may not be straightforward. Most of the
time, data for commodities and services consumed are collected at the
household level. A household typically consists of members with different
characteristics, such as age, sex, and employment status. Usually, an individ-
ual’s welfare indicator is calculated by dividing total household expenditures
by the number of people residing in that household. The result is called the
per capita expenditure.
Analyzing poverty on the basis of per capita expenditure, however,
ignores the fact that different individuals may have different needs.
The cost per person to reach a certain welfare level may be lower in
large households, because large households enjoy certain economies of
scale. For example, a child may not need the same share of income as
an adult member, or the food consumption expenditure may not be the
same across men and women within a household. The minimum income
needed to meet the subsistence needs of a household with four adults
may be much more than the subsistence income needed for a household
with two adults and two children. This intrahousehold allocation can be
adjusted using an equivalence scale tool.
47
A Unified Approach to Measuring Poverty and Inequality
48
Chapter 2: Income Standards, Inequality, and Poverty
are positive but not larger than one. When ϱ = 1, then the cost of a child
is equal to the cost of an adult. The lower the value of ϱ, the lower the cost
of each child compared to an adult. Similarly, when ϑ = 1, no economy of
scale is assumed. The lower the value of ϑ, the larger the economy of scale
is assumed to be.
For example, suppose there are five members in a household: three adults
and two children. If a child is assumed to be half as costly as an adult, then
ϱ = 0.5 and ϑ = 0.5. Then AELSMS = (3 + 0.5 × 2)0.5 = 2. Therefore, if the
actual total income of the household is Rs 20,000, then the real per capita
income of the household is equivalent to Rs 10,000. However, if no econ-
omy of scale is assumed and each child is considered as equally expensive as
an adult, then the household’s per capita income is only Rs 4,000.
In the subsequent analysis in this chapter, we assume that we are using
a dataset having all the information required for constructing a welfare
indicator either at the individual level or at the household level. The
dataset may cover the entire population or may just be a collection of
samples from the population. There are other important issues one should
take into account regarding a dataset (such as its survey design, sample
coverage, sample variability, and so on), which are not covered in this
chapter.2
To keep explanations and mathematical formulas simple, we make two
fundamental assumptions. First, we use income as the welfare indicator
and assume that information on income is available for every person in our
dataset. Second, we assume that every household contains only one adult
member. As a result of the second assumption, we do not need to make
any adjustment for the economy of scale and equivalent scale because each
member is an adult and lives in a single-member household. However, the
tools and techniques introduced in this chapter can be easily extended to
situations when the welfare indicator is consumption expenditure and more
than one person lives in a household.
Basic Concepts
49
A Unified Approach to Measuring Poverty and Inequality
Density Function
• The median is the income in the distribution that divides the entire
population into two equal shares. In the figure, xM is the median of
50
Chapter 2: Income Standards, Inequality, and Poverty
Density
fx
x1 xMo xM b ′ b ″ xN
Income
51
A Unified Approach to Measuring Poverty and Inequality
Fx(xN) = 100%
Fx(b ″) Fx
Fx(b ′)
Fx(bM) = 50%
Mean
bM b ′ b ″ xN
Income
52
Chapter 2: Income Standards, Inequality, and Poverty
Quantile Function
xN
Qx
Income
bM
x1 Mean
0 50 100
Population share (percent)
53
A Unified Approach to Measuring Poverty and Inequality
Income Standards
Desirable Properties
An income standard can satisfy several basic properties. We refer to the first
two properties—symmetry and population invariance—as invariance properties
because they describe changes in the distribution that leave the income
standard unaltered. The second pair of properties—weak monotonicity and
the weak transfer principle—are called dominance properties because they
require the income standard to rise (or not fall) when the income distribution
changes in a particular way. Finally, normalization and linear homogeneity are
calibration properties that ensure the income standard is measured by income.
The additional property of subgroup consistency is not a part of the basic prop-
erties, but it is desirable when evaluating income standards of subpopulations.
Symmetry requires that switching two people’s incomes leaves the
income standard evaluation unchanged. In other words, a person should
not be given priority on the basis of his or her identity when calculating a
society’s income standard. Thus, symmetry is also known as anonymity. In
technical terms, symmetry requires the income standard of distribution x to
be equal to the income standard of distribution x', if x' is obtained from x by
a permutation of incomes.
54
Chapter 2: Income Standards, Inequality, and Poverty
The third property requires that if the income of any person in a society
increases, then the income standard should register an increase, or at least
should not fall. Implicitly, this property assumes that increasing someone’s
income is not harmful to the entire society.
There are two versions of this property. One is weak monotonicity, which
requires that the income standard not fall because of an increase in any-
one’s income. The other version is monotonicity, the stronger version, which
requires that the income standard register an increase if anyone’s income in
the society increases.
55
A Unified Approach to Measuring Poverty and Inequality
For vectors x and x', the notation x' > x implies that at least one element
in x' is strictly greater than that in x, and all other elements in x' are no less
than the corresponding elements in x. For example, if x' = ($20k, $10k, $30k)
and x = ($25k, $10k, $30k), then x' > x. However, if x' = ($20k, $10k, $30k)
and x = ($25k, $10k, $25k), then x' ⬎ x because the income of the third
person is lower in x than that in x'.
56
Chapter 2: Income Standards, Inequality, and Poverty
57
A Unified Approach to Measuring Poverty and Inequality
Quantile Income
58
Chapter 2: Income Standards, Inequality, and Poverty
The quantile income at the pth percentile is the income below which the
incomes of p percent of the population fall. For the income distribution x
with N people, the quantile income at the pth percentile is the income that
is larger than the incomes of the poorest pN/100 people.
We denote the quantile income at the pth percentile of distribu-
tion x by WQI (x; p). For example, if p = 50 percent, then the quantile
income at the pth percentile of distribution x is denoted by WQI (x; 50).
If WQI (x; 50) = $200, then it should be read as 50 percent of the population
in society X earns less than $200. Similarly, if WQI (x; 90) = $1,000, then
90 percent of its population earns less than $1,000.
Commonly reported quantile incomes used when gauging societies’
standard of living are the quantile incomes at the 10th percentile, 20th per-
centile, 50th percentile, 80th percentile, and 90th percentile. A close look
at the quantile income at the 50th percentile reveals that this is the income
below which half of the population of a distribution lies. Therefore, the
quantile at the pth percentile income is just the median of a distribution. For
a particular income distribution where each and every person earns equal
income, the quantile incomes at all percentiles are equal to each other,
ensuring that the quantile incomes satisfy the normalization property.
A quantile function is the most helpful tool for visualizing quantile
incomes. Figure 2.4 shows the quantile function for income distribution x.
Qx(100) WQI(x;100)
Corresponding value of
a quantile function Qx
Qx(90) WQI(x;90)
Quantile income
Qx
Qx(75) WQI(x;75)
bM WQI(x;50)
Qx(25) WQI(x;25)
Qx(10) WQI(x;10)
0 10 25 50 75 90 100
Population share or percentiles
59
A Unified Approach to Measuring Poverty and Inequality
As in figure 2.3, the horizontal axis in figure 2.4 denotes the population share
in percentage, which lies between 0 and 100. The left-hand vertical axis
denotes the corresponding value of a quantile function Qx and the right-hand
vertical axis reports the quantile incomes.
By definition, the quantile income for a certain percentile is the value of
the quantile function at that percentile, so WQI (x; p) = Qx(p). In the figure,
WQI (x; 50) = bM is the median of distribution x. Likewise, WQI (x; 25) and
WQI (x; 75) are the first and the third quartiles of distribution x. The well-
known 10th and 90th percentiles of distribution x are WQI (x; 10) = Qx(10)
and WQI (x; 90) = Qx(90), respectively. Given that a cdf is an inverse of a
quantile function, quantile incomes can also be graphically portrayed and
calculated using a cdf.
What properties do quantile incomes satisfy? It is straightforward to verify
that any quantile income satisfies symmetry, normalization, population invari-
ance, linear homogeneity, and weak monotonicity. However, no quantile income
satisfies the other dominance properties: monotonicity, transfer principle,
and subgroup consistency. Quantile incomes do not satisfy monotonicity
because a person’s income may increase, but as long as it does not surpass a
certain quantile, that quantile income remains unaltered. Similarly, quantile
incomes do not satisfy the transfer principle because they do not change to a
transfer that takes place at a nonrelevant part of the distribution.
The income standards are not subgroup consistent because the quantile
incomes of the subregions may increase, but the overall quantile income may
fall. Consider the following example. Suppose the income vector of society
X is x = ($10k, $20k, $30k, $50k, $60k, $80k) and the income vector of two
subgroups is x' = ($10k, $20k, $30k) and x" = ($50k, $60k, $80k). The 67th
quantile of the three distributions is WQI (x'; 67) = $20k, WQI (x"; 67) = $60k,
and WQI (x; 67) = $50k. Now, suppose the subgroup income vectors over time
become y' = ($10k, $20k, $30k) and y" = ($45k, $65k, $80k). Apparently, the
quantile income at the 67th percentile of the first group does not change, but
that of the second does. In fact, WQI (x'; 67) = WQI (y'; 67) but WQI (y"; 67) >
WQI (x"; 67). What happens to the quantile income at the 67th percentile
of the overall distribution? It turns out that WQI (y; 67) = 45 < WQI (x; 67).
Partial Mean
The next set of commonly used means is the partial means. There are two
types of partial means: lower partial means and upper partial means. A lower
60
Chapter 2: Income Standards, Inequality, and Poverty
partial mean is obtained by finding the mean of the incomes below a specific
percentile cutoff. An upper partial mean is obtained by finding the mean of
incomes above a specific percentile cutoff. Lower partial means are more
commonly used than upper partial means.
The lower partial mean of the pth percentile is the average or mean
income of the bottom p percent of the population. The upper partial mean
of the pth percentile, in contrast, is the average or mean income of the
top (1 – p) percent of the population. We denote the lower partial mean
and upper partial mean of distribution x for percentile p by WLPM(x; p) and
WUPM(x; p), respectively. For example, if p = 50 percent, then the lower par-
tial mean of the pth percentile of distribution x is denoted by WLPM(x; 50).
If WLPM(x; 50) = $100 and WUPM(x; 50) = $10,000, then together they
should be read as the mean income of the bottom 50 percent of the population
is $100, and the mean income of the top 50 percent of the population is $10,000
(see example 2.1).
Example 2.1: Consider the income vector x = ($2k, $4k, $8k, $10k).
The lower partial mean of the 50th percentile of the distribution is
($2k + $4k)/2 = $3k, and that of the 75th percentile of the distribution
is ($2k + $4k + $8k)/3 = $4.7k. In contrast, the upper partial mean
of the 50th percentile of the distribution is ($8k + $10k)/2 = $9k and
that of the 75th percentile of the distribution is $10k.
61
A Unified Approach to Measuring Poverty and Inequality
xN
Qx(p″)
Income
Qx
Qx(p′)
0 p′ p″ 100
Population share
consistency. Like the quantile incomes, one can easily show using a simple
example that partial means do not satisfy subgroup consistency.
Quantile incomes and partial means are crude income standards because
they do not depend on the entire income distribution. Yet they are highly
informative and easy to understand. Especially when income data are miss-
ing for certain parts of the income distribution, these crude income stan-
dards are useful tools for understanding a society’s performance.
In contrast, when rich datasets are available, a study based on quan-
tile incomes and partial means may be limited because they do not reflect
changes in every part of the distribution. For example, if the income
of a person below the median increases—but not by enough to surpass
the median income—then the distribution median does not reflect any
change.
The following income standards are designed to consider the entire
distribution. These income standards will, in most cases, reflect a change in
any part of the distribution.
General Mean
62
Chapter 2: Income Standards, Inequality, and Poverty
Although a may take any value between − ∞ and + ∞, four means in this
family are more well known than others: arithmetic mean, geometric mean,
harmonic mean, and Euclidean mean.
ln x1 + ln x 2 + L + ln x N
WL (x) = ln WG (x) = . (2.6)
N
63
A Unified Approach to Measuring Poverty and Inequality
1
⎛ x12 + x 22 + L + x 2N ⎞ 2
WE (x) = ⎜ ⎟⎠ . (2.8)
⎝ N
Example 2.2 shows the results of calculating these means for a given
income vector.
Example 2.2: Consider the income vector x = ($2k, $4k, $8k, $10k).
• The arithmetic mean of x is ($2k + $4k + $8k + $10k)/4 = $6k.
• The geometric mean of x is ($2k × $4k × $8k × $10k)1/4
= $5.03k.
• The harmonic mean of x is [($2k−1 + $4k−1 + $8k−1 +
$10k−1)/4]−1 = $4.10k.
• The Euclidean mean of x is [($2k2 + $4k2 + $8k2 + $10k2)/4]1/2
= $6.78k.
Having been introduced to the family, one can now understand the
properties of general means and the way they depend on parameter a. All
means in this family satisfy symmetry, normalization, population invariance,
linear homogeneity, monotonicity, and subgroup consistency. Furthermore, for
a < 1, general means satisfy the transfer principle. Thus, the general means
satisfy all the dominance properties introduced earlier. One reason is that,
unlike the quantile means and the partial means, general means consider all
incomes in the distribution.
It is straightforward to show that general means satisfy symmetry, nor-
malization, population invariance, linear homogeneity, and monotonicity.
That general means satisfy subgroup consistency may be verified as follows:
if vector x is divided into subgroup vectors x' and x", then the general mean
of x can be expressed as
WGM(x; a) = WGM((WGM(x'; a), WGM(x"; a)); a). (2.9)
In other words, the general mean of x is the general mean of the general
means of x' and x". Then the monotonicity property ensures that subgroup
consistency is satisfied.
64
Chapter 2: Income Standards, Inequality, and Poverty
65
A Unified Approach to Measuring Poverty and Inequality
WE(x) WGM(x; α)
WA(x)
WG(x)
WH(x)
x1
–∞ –2 –1 0 1 2 ∞
Parameter
Sen Mean
The usual mean can be reinterpreted as the expected value of a single income
drawn randomly from the population. Now, suppose that instead of a single
income, we were to draw two incomes randomly from the population (with
replacement). If we then evaluated the pair in terms of the lower of the two
incomes, this would lead to the Sen mean, which is defined as the expecta-
tion of the minimum of two randomly drawn incomes.6 These two random
incomes are drawn with replacement, which means that these two incomes
may belong to the same person in a society. If every income in distribution x
66
Chapter 2: Income Standards, Inequality, and Poverty
is compared with every other income in x with replacement, then there are
N2 possible comparisons. Thus, the Sen mean can be defined as
1 N N
WS (x) = ∑ ∑ min{x n′ x n′}.
N 2 n =1 n ′ =1
(2.10)
Because we are using the minimum of the two incomes, this number can be
no higher than the mean, and is generally lower. The Sen mean also empha-
sizes the lower incomes but in a way that differs from the general means with
α < 1, the lower partial means, or the quantile incomes below the median.
There is a straightforward way of calculating the Sen mean for an income
vector—by creating an N × N matrix that has a cell for every possible pair of
incomes and placing the lower value of the two incomes in the cell. Adding
all the entries and dividing by the number of entries (N2) to obtain their
mean provides the Sen mean. Consider example 2.3 to better understand
this way of calculating the Sen mean.
Example 2.3: Consider the income vector x = ($2k, $4k, $8k, $10k).
First, we construct the following matrix:
67
A Unified Approach to Measuring Poverty and Inequality
person has the highest income and the first person has the lowest income.
Thus, element xN receives the highest rank and element x1 receives the low-
est rank. The Sen mean attaches the highest weight to the lowest income,
the second-highest weight to the second-lowest income, and the lowest
weight to the highest income.
For distribution x, the Sen mean can be expressed as WS(x) = a1x1 + … +
aNxN, where aN = (2(N − n) + 1)/N2 for all n. Thus, the weight attached
to the highest income xN is aN = 1/N2; the weight attached to the second-
highest income xN–1 is aN–1 = 3/N2; and the weight attached to the lowest
income x1 is a1 = (2N − 1)/N2. The weight attached to the richest income
in the example above ($10k) is 1/16, whereas the weight attached to the
poorest income ($2k) is 7/16. Notice that the weights sum to one, that is,
1 N2
a1 + a 2 + ... + a N = (1 + 3 + 5 + ... + (2 N − 1) = = 1. (2.11)
N2 N2
1
WS (x) = ((2N − 1)x1 + (2N − 3)x 2 + L + 3x N −1 + x N ). (2.12)
N2
68
Chapter 2: Income Standards, Inequality, and Poverty
mean goes up. This feature of the Sen mean is inherited by the inequality
and poverty measures that are based on the Sen mean—the famous Gini
coefficient and the Sen-Shorrocks-Thon index of poverty.
Finally, unlike Atkinson, Sen suggested going beyond the utilitarian
form. His key nonutilitarian example, the Sen mean, can be viewed as both
an ede and a general welfare function, because it satisfies the transfer prin-
ciple. If we denote the Sen ede as EDES(x), then EDES(x) = WS(x).
During our subsequent discussion in this chapter, we will see that these
five means (arithmetic, geometric, harmonic, Euclidean, and Sen) and their
various functional forms are often used in the measurement of welfare,
inequality, and poverty.
69
A Unified Approach to Measuring Poverty and Inequality
70
Chapter 2: Income Standards, Inequality, and Poverty
Figure 2.7: First-Order Stochastic Dominance Using Quantile Functions and Cumulative
Distribution Functions
100
Fy
Income
Qx Fx
Qy
0 100 0
Population share Income
71
A Unified Approach to Measuring Poverty and Inequality
30
25
Q
x
20 20 V
WA(x)
Income
Income
E
15 14 IV
D
C
10 III GLx
B 9
A II
5
I
2
0 20 40 60 80 100 0 20 40 60 80 100
Population share Population share
72
Chapter 2: Income Standards, Inequality, and Poverty
the quantile function divided by the percentile itself. Thus, the height of
the generalized Lorenz curve at any percentile of population p is the lower
partial mean times p itself, because the height of the generalized mean is the
area underneath the quantile function at corresponding percentile p, that is,
GLx(p) = pWLPM(x; p). If income distribution x has a large enough sample
size, the generalized Lorenz curve takes a form similar to the one described
in figure 2.9.
The horizontal axis in figure 2.9 shows the population share, and the ver-
tical axis denotes the height of the generalized Lorenz curve by income. The
generalized Lorenz curve for distribution x is denoted by GLx. The maximum
height of GLx is WA(x). The height of GLx for the 50th percentile of the
population is GLx(0.5).
If the total income in distribution x is distributed equally across all
people in the society and distribution y is obtained, then the generalized
Lorenz curve GLy becomes a straight line. The maximum height of GLy is
also WA(x), because redistribution of incomes does not change the mean
income. Notice that the height of GLy is higher than the height of GLx
for every percentile p. This implies that every partial mean of distribution
y is larger than the corresponding partial mean of distribution x. Thus, two
generalized Lorenz curves of this sort show a dominance relation between two
distributions in terms of partial means.
GLx(100) WA(x)
GLy
Income
GLy(50) WA(x)/2
GLx
WA(x ′)
GLx(50)
GLx ′ (50) GLx ′
0 50 100
Population share
73
A Unified Approach to Measuring Poverty and Inequality
All partial means agree that distribution y has higher welfare than dis-
tribution x. Similarly, if there is another distribution x' whose generalized
Lorenz curve, GLx', lies completely below GLx (also shown in figure 2.9),
then all partial means agree that distribution x has higher welfare than dis-
tribution x'. The heights of the generalized Lorenz curves for distributions
y, x, and x' at the 50th percentile are GLy(50), GLx(50), and GLy'(50),
respectively. The generalized Lorenz curve represents second-order stochas-
tic dominance, which signals higher welfare according to every utilitarian
welfare function with identical and increasing utility function exhibiting
diminishing marginal utility. Example 2.4 provides a practical illustration of
generalized Lorenz calculations. The generalized Lorenz curve is also closely
related to the Sen mean. For distribution x, the Sen mean, WS(x), is twice
the area underneath GLx.
74
Chapter 2: Income Standards, Inequality, and Poverty
Finally, a third curve depicts the general mean levels as parameter a varies.
We call this curve a general mean curve. This curve has already been outlined
in figure 2.6, where it is increasing in α; tends to the minimum income for
very low a ; rises through the harmonic, geometric, arithmetic, and Euclidean
means; and tends toward the maximum income as α becomes very large.
Why is this curve useful? At the beginning of this subsection, an example
showed that different generalized means may rank an income distribution
differently. So the general mean curve is useful for determining (a) whether
a given comparison of general means is robust to the choice of any income
standard from the entire class of general means, and, if not, (b) which of the
income standards is higher or lower.
General mean curves are also related to the quantile function and the
generalized Lorenz curve. A higher quantile function will always yield a
higher general mean curve, and a higher generalized Lorenz curve will raise
the general mean curve for a < 1, or the general means that favor the low
incomes. The general mean curve concept will be particularly relevant to
our later discussions of Atkinson’s inequality measure.
Growth Curves
75
A Unified Approach to Measuring Poverty and Inequality
We start with the growth incidence curve. Consider two income distributions,
x and y, at two different periods of time, where x is the initial income distri-
bution. The quantile incomes of distribution x and distribution y at percen-
tile p are denoted by WQI(x; p) and WQI(y; p), respectively. The growth of
quantile income at percentile p is denoted by
If every quantile registers an increase over time, then gQI(x, y; p) > for
all p. The curve’s height at p = 50 percent gives the median income’s growth
rate. Note that no part of this growth curve provides any information about
the growth of mean income. Varying p allows us to examine whether this
growth rate is robust to the choice of income standard, or whether the low-
income standards grew at a different rate than the rest.
Figure 2.10 depicts the growth curves of quantile incomes. The vertical
axis denotes the growth rate of quantile income and the horizontal axis denotes
the cumulative population share. Suppose there are two societies, X and X'.
The income distributions of society X at two different points in time are
x and y, while those of society X' are x' and y'. The dashed growth curve
gQI(x, y) denotes the quantile income growth rates of society X over time,
whereas the dotted growth curve gQI(x', y') denotes the quantile income
growth rates of society X' over time.
Suppose the growth rates of mean income across these two distributions
are the same and are denoted by –g > 0. Thus, the solid horizontal line at
–g denotes the growth rate if the growth rate had been the same for all per-
centiles or the cumulative population share.
76
Chapter 2: Income Standards, Inequality, and Poverty
A
B
g
gQI(x,y)
A′ B′ gQI(x ′,y ′)
0 20 40 60 80 100
Cumulative population share
The next growth curve is the generalized Lorenz growth curve. Consider the
two income distributions, x and y, used previously. The lower partial means
of distribution x and distribution y at percentile p are denoted by WLPM
(x; p) and WLPM(y; p), respectively. The growth of partial means at percen-
tile p is denoted by
77
A Unified Approach to Measuring Poverty and Inequality
C
D gLPM(x,y)
g
C′ D′
gLPM(x ′,y ′)
0 20 40 60 80 100
Cumulative population share
78
Chapter 2: Income Standards, Inequality, and Poverty
The final of the three growth curves is the general mean growth curve.
Considering the income distributions x and y discussed previously, we
denote the general mean of order a of distribution x and distribution y by
WGM(x; a) and WGM(y; a), respectively. The growth of general mean of
order a is denoted by
79
A Unified Approach to Measuring Poverty and Inequality
When every general mean registers an increase over time, gGM(x,y; a) > 0.
When a = 1, the curve’s height is the usual mean income growth rate. This
rate is equal to the growth of the generalized Lorenz growth curve at p = 100
percent. At a = 0 the curve shows the growth rate for the geometric mean,
and so forth. As we will see later, each of these growth curves can help
in understanding the link between growth and change in inequality
over time.
Figure 2.12 shows the growth curves of general mean incomes. The verti-
cal axis denotes the growth rate of general mean income, and the horizontal
axis denotes the values of parameter a. Following the same notations as the
previous two growth incidence curves, suppose that there are two societies,
X and X'. Income distributions of society X at two different points in time
are x and y, whereas those of society X' are x' and y'. The dashed growth
curve gGM(x, y) denotes the growth rates of general mean income of soci-
ety X over time, whereas the dotted growth curve gGM(x', y') denotes the
growth rates of general mean income of society X' over time.
Suppose the growth rates of mean income across these two distributions
–
are the same and are denoted by g > 0. Thus, the solid horizontal line at
–
g denotes the growth rate if the growth rate had been the same for all a.
What information do these two growth curves provide? Growth between
x and y is pro-poor in the sense that general means for lower a, which focus
more on the lower end of the distribution, have positive growth, whereas
gG ( ′)
M x,y) ′,y
(x
M
gG
g
–∞ –2 –1 0 1 2 ∞
Parameter
80
Chapter 2: Income Standards, Inequality, and Poverty
general means for larger a have negative growth. Growth between x' and y', in
contrast, is not pro-poor because the general means for lower a have negative
income growth whereas the general means for larger a have positive growth.
The mean income growth rates are the heights of the two growth curves at
a = 1, which are equal by assumption for this example. Heights at a = 0 and
a = –1 are growth rates of the geometric and harmonic means, respectively.
Inequality Measures
Desirable Properties
81
A Unified Approach to Measuring Poverty and Inequality
82
Chapter 2: Income Standards, Inequality, and Poverty
remains unchanged when compared to the income standard, even after all
incomes are scaled up or down by the same factor. This similarity supports
the idea that the relative inequality level remains unchanged.7
The fourth property, normalization, requires that if incomes are the same
across all people in a society, then no inequality exists within the society
and the inequality measure should be zero. Normalization is a natural
property. For example, if the income vector of a three-person society is
($20k, $20k, $20k), then the inequality measure should be zero. Even if
everyone’s income increases 10-fold and the new income vector is ($200k,
$200k, $200k), the inequality measure should still be zero.
The fifth property is the transfer principle, which requires that a regressive
transfer between two people in a society should increase inequality and a
progressive transfer between two people should reduce inequality. Regressive
and progressive transfers were defined earlier for income standards.
83
A Unified Approach to Measuring Poverty and Inequality
84
Chapter 2: Income Standards, Inequality, and Poverty
85
A Unified Approach to Measuring Poverty and Inequality
urban areas, states, provinces, and other geographic regions; across ethnic
and religious groups; across genders; or across age groups. One may want to
evaluate the source of inequality, such as whether overall income inequality
is due to unequal income distribution within sex or unequal income distri-
bution across sex.
The eighth property is additive decomposability, which requires overall
inequality to be expressed as a sum of within-group inequality and between-
group inequality. Within-group inequality is a weighted sum of subgroup
inequalities. Between-group inequality is the inequality level obtained when
every person within each subgroup receives the subgroup’s mean income.
Kanbur (2006) discussed the policy significance of this type of inequality
decomposition. It is often found that the contribution of the between-group
term is much lower than the within-group term, and, thus, policy priority
is directed toward ameliorating within-group rather than between-group
inequality. These types of policy conclusions should be carefully drawn,
because the lower between-group term may receive much larger social
weight than its within-group counterpart. Also, the between-group term’s
share of overall inequality may increase as the number of groups increases.
How to incorporate these issues into inequality measurement requires fur-
ther research, and solving these issues is beyond the scope of this book.
However, if the policy interest is in understanding how the between-group
inequality as a share of total inequality has changed over time for a fixed
number of groups, then the decomposability property is very useful.
To formally outline the additive decomposability property, we will use
two groups to simplify the interpretation, but the definition can be extended
to any number of groups. Suppose the income vector x with population size
N is divided into two subgroup vectors: x' with population size N' and x"
with population size N" such that N' + N" = N. Let us denote the means of
these three vectors by x̄, x̄', and x̄". The additive decomposability property
can be stated as follows:
86
Chapter 2: Income Standards, Inequality, and Poverty
Commonly used inequality measures are mostly related to the five kinds
of income standards we discussed earlier. The inequality measures that we
discuss in this section are quantile ratios, partial mean ratios, Gini coefficient,
Atkinson’s class of inequality measures, and generalized entropy measures.
Quantile Ratio
87
A Unified Approach to Measuring Poverty and Inequality
For income distribution x, let the quantile income at the pth percentile
be denoted by WQI(x; p), and let the quantile income at the p'th percentile
be denoted by WQI(x; p'), such that p > p'. A quantile ratio is commonly
reported as a ratio of the larger quantile income to the smaller quantile
income. However, this view leads the values of inequality measures to range
from one to ∞. This range is not comparable to other inequality measures,
which commonly range from zero to one. In this book, we formulate the
quantile ratio in such a way that it ranges from zero to one. The p/p' quantile
ratio is represented by the following formula:
• The number may be directly read as the gap between the lowest
income of the richest 10 percent and the highest income of the poorest
10 percent of the population, being 90 percent of the lowest income of
the richest 10 percent of the population.
• The number may be seen as the highest income of the poorest
10 percent of the population, being 10 percent (1 − 0.9 = 0.1) of the
lowest income of the richest 10 percent of the population.
88
Chapter 2: Income Standards, Inequality, and Poverty
Quantile ratios may be classified into three categories: upper end quantile
ratio, lower end quantile ratio, and mixed quantile ratio. The first two categories
capture inequality within any one side of the median, and the third category
captures inequality in one side of the median versus that of the other side of
the median. For example, IQR(x; 90/50) is an upper end quantile ratio, and
IQR(x; 50/10) is a lower end quantile ratio, whereas IQR(x; 90/10) is a mixed
quantile ratio.
What properties does a quantile ratio satisfy? A quantile ratio, as defined
earlier, satisfies symmetry, normalization, population invariance, and scale
invariance. Thus, a quantile ratio satisfies all four invariance properties.
What about the dominance properties? It turns out that a quantile ratio
satisfies none of the dominance properties.
The following example shows that a quantile ratio does not satisfy the weak
transfer principle. Suppose the highest income of the poorest 10 percent of
the population is $100 and the lowest income of the richest 10 percent of the
population is $2,000. Then IQR(x; 90/10) = ($2,000 − $100)/$2,000 = 0.95.
Now, suppose that a regressive transfer takes place between the poorest
person in the society and the richest person among the poorest 10 percent
of the population such that the highest income in that group increases to
$120. Then the post-transfer quantile ratio is IQR(x; 90/10) = ($2,000 −
$120)/$2,000 = 0.94.
Therefore, the quantile ratio shows a decrease in inequality even when
a regressive transfer has taken place. If a quantile ratio does not satisfy the
weak transfer principle, then it cannot satisfy its stronger version—the
transfer principle, or transfer sensitivity. The quantile ratios are not addi-
tively decomposable and also do not satisfy subgroup consistency.
89
A Unified Approach to Measuring Poverty and Inequality
90
Chapter 2: Income Standards, Inequality, and Poverty
What properties does a partial mean ratio satisfy? A partial mean ratio,
as defined in equation (2.17), satisfies symmetry, normalization, population
invariance, and scale invariance. Thus, a partial mean ratio satisfies all four
invariance properties. What about the dominance properties? A quantile
ratio satisfies the weak transfer principle but does not satisfy the transfer
principle, transfer sensitivity, and subgroup consistency. It does not satisfy
the transfer principle because some regressive and progressive transfers may
leave the inequality measure unchanged, since a partial mean ratio does not
consider the entire income distribution.
91
A Unified Approach to Measuring Poverty and Inequality
We already know from our discussion of income standards that the value
of general means falls as a falls and vice versa. As a decreases, the distance
between WA(x) and WGM(x; a) increases, implying that IA increases as a
falls for a particular income distribution. Among the entire class of mea-
sures, three are used more frequently: a = 0, a = –1, and a = –2. For a = 0,
the general mean takes the form of the geometric mean. The corresponding
Atkinson’s inequality measure for distribution x is expressed as
For a = –1, the general mean is known as the harmonic mean. The cor-
responding Atkinson’s inequality measure for distribution x is expressed as
For a = –2, the general mean has no such name, and we will call it
simply WGM(X; –2). The corresponding Atkinson’s inequality measure for
distribution x is expressed as
Example 2.6: Consider the income vector x = ($2k, $4k, $8k, $10k)
used previously in the general means example. The arithmetic mean
is WA(x) = $6k, the geometric mean is WG(x) = $5.03k, the harmonic
mean is WH(x) = $4.10k, and WGM(x; –2) = $3.44k.
Thus,
IA(x; 0) = ($6k − $5.03k)/$6k = 0.162.
IA(x; −1) = ($6k − $4.10k)/$6k = 0.317.
IA(x; −2) = ($6k − $3.44k)/$6k = 0.427.
92
Chapter 2: Income Standards, Inequality, and Poverty
and the geometric mean of x. The arithmetic mean represents the level
of welfare obtained when the overall income is distributed equally across
everyone in the society. This is an ideal situation when there is no inequality
in the society.
The geometric mean, in contrast, is the equally distributed equivalent
(ede) income, which, if received by everyone in the society, would yield
the same welfare level as in x for the degree of inequality aversion a = 0. So
IA(x; 0) = 0.162 implies that the loss of welfare because of inequality in dis-
tribution x is 16.2 percent of what the welfare level would be if the overall
income had been equally distributed.
Suppose the society becomes more averse to inequality and a is reduced
to −1. In this case, the equally distributed equivalent income is the har-
monic mean of x. The loss of total welfare because of unequal distribution
increases from 16.2 percent to 31.7 percent. Likewise, the percentage loss
of welfare would increase to 42.7 percent if the society became even more
averse to inequality and a fell to −2.
What properties does any index in this family satisfy? Any measure in
this family satisfies all four invariance properties: symmetry, population invari-
ance, scale invariance, and normalization. In addition, unlike the quantile
ratios and the partial mean ratios, measures in this class satisfy the transfer
principle, transfer sensitivity, and subgroup consistency.
If distribution x' is obtained from distribution x by at least one regres-
sive transfer, then the level of inequality in x' is strictly higher than that
in x. Furthermore, if transfers take place between poor people, then the
inequality measure changes more than if the same amounts of transfers take
place among rich people. Finally, because these measures satisfy subgroup
consistency, they do not lead to any inconsistent results while decomposing
across subgroups. If inequality in certain subgroups increases while inequal-
ity in the others does not fall, then overall inequality increases. However,
measures in this class are not additively decomposable.
Gini Coefficient
93
A Unified Approach to Measuring Poverty and Inequality
N N
1
IGini (x) = ∑ ∑ x n − x n′ .
2N 2 × WA (x) n =1 n’=1
(2.22)
Note that equation (2.22) may be broken into two components: WA(x)
(the mean of the distribution) and (∑N n=1∑ n'=1|xn – xn'|)/2N (the average
N 2
1 N N 1 N N
∑ ∑ n n′ A
2 N 2 n =1 n ′ =1
x − x = W (x) − ∑ ∑ min {x n , x n′}= WA (x) − Ws (x), (2.23)
N 2 n = 1 n ′= 1
where WS(x) is the Sen mean of distribution x. Therefore, the Gini coefficient
may be simply formulated by using the arithmetic mean and the Sen mean.
Like any measure in Atkinson’s class, the Gini coefficient can be expressed as
94
Chapter 2: Income Standards, Inequality, and Poverty
Example 2.7: Consider the income vector x = ($2k, $4k, $8k, $10k)
that we used previously. First, we calculate the Gini coefficient using
the formulation in equation (2.22). It can be easily verified that
WA(x) = $6k. The second component is
1
( 2 − 2 + 2 − 4 + 2 − 8 + 2 − 10 + 4 − 2 + 4 − 4 + 4 − 8 + 4 − 10
2 × 42
+ 8 − 2 + 8 − 4 + 8 − 8 + 8 − 10 + 10 − 2 + 10 − 4 + 10 − 8 + 10 − 10 )
1
= (0 + 2 + 4 + 6 + 8 + 2 + 0 + 4 + 6 + 6 + 4 + 0 + 2 + 8 + 6 + 2 + 0)
32
56
= = 1.75.
32
the welfare level if overall income had been equally distributed. We will see
later that the Gini coefficient has an interesting relationship with the well-
known Lorenz curve.
The Gini coefficient satisfies all invariance properties: symmetry, population
invariance, scale invariance, and normalization. In addition, it satisfies the transfer
principle. If distribution x' is obtained from distribution x by at least one regres-
sive transfer, then the level of inequality in x' is strictly higher than that in x.
However, the Gini coefficient is neither transfer sensitive nor subgroup con-
sistent. It is not transfer sensitive because the Gini coefficient changes by the
same amount whether transfers take place between poor people or between
rich people. That the Gini coefficient is not subgroup consistent means that if
the inequality in some subgroups increases while inequality in other subgroups
does not fall, then the overall inequality may register a decrease.
The following is an example showing that the Gini coefficient is neither
transfer sensitive nor subgroup consistent. Consider the vector x = ($4k,
$5k, $6k, $7k, $14k, $16k). If a progressive transfer of $0.5k takes place
between the first person and the second person, then x' = ($4.5k, $4.5k, $6k,
$7k, $14k, $16k). If a progressive transfer of the same amount takes place
95
A Unified Approach to Measuring Poverty and Inequality
between the two richer people, then x" = ($4k, $5k, $6.5k, $6.5k, $14k,
$16k). As a result, IGini(x') = IGini(x") = 0.279. Thus, the Gini coefficient
cannot distinguish between these two transfers.
The next example shows that the Gini coefficient is not subgroup consis-
tent. We use the same example that we used to show that the Sen mean does
not satisfy subgroup consistency. The original income vector x = ($4k, $5k,
$6k, $7k, $14k, $16k) becomes, over time, y = ($3.4k, $6.1k, $6k, $6.5k,
$14k, $16k). The income vector of the first subgroup x' = ($4k, $5k, $7k)
becomes y' = ($3.4k, $6.1k, $6.5k), whereas the income vector of the sec-
ond subgroup remains unaltered. The Sen mean of the first group falls from
WS(x') = $4.67k to WS(y') = $4.64k, whereas the mean income remains
unchanged at WA(x') = WA(y') = $5.33k. So the inequality of the first
group increases from IGini(x') = 0.125 to IGini(y') = 0.129. What happens
to the overall inequality? It turns out that the overall Sen mean increases
from WS(x) = $6.22k to WS(y) = $6.24k, whereas the overall mean income
remains unchanged at WA(x) = WA(y) = $8.67k. The overall inequality
decreases from IGini(x) = 0.282 to IGini(y) = 0.280.
However, unlike the Atkinson class of measures, the Gini coefficient is
additively decomposable, but with an added residual term. If distribution x is
divided into population subgroups x' with population size N' and x" with
population size N", then the decomposition formula of the Gini coefficient is
IGini(x) = w'IGini(x')+w" IGini(x") + IGini(x–', x– ") – residual, (2.25)
where the weights are w' = (N'/N)2(x̄'/x̄) and w" = (N"/N)2(x̄"/x̄). Note,
however, that the weights may not sum to one. The residual term is not zero
if and only if the groups’ income ranges overlap. If we consider the example
above, where the income vector x = ($4k, $5k, $6k, $7k, $14k, $16k) is
divided into two subgroup vectors: x' = ($4k, $5k, $7k) and x" = ($6k, $14k,
$16k). These vectors overlap as $7k > $6k. Thus, the residual term will
not vanish. However, if the two subgroups were x' = ($4k, $5k, $6k) and
x" = ($7k, $14k, $16k), then the residual term would be zero.9
96
Chapter 2: Income Standards, Inequality, and Poverty
⎧ 1 ⎡1 N ⎛ xn ⎞
a
⎤
⎪ ⎢ ∑ n =1 ⎜ ⎟ − 1⎥ if a ≠ 0,1
⎪ a(a − 1) ⎢⎣ N ⎝ x⎠ ⎦⎥
⎪
⎪1 N x ⎛x ⎞
IGE (x; a) = ⎨ ∑ n =1 n ln ⎜ n ⎟ if a = 1. (2.26)
⎪N x ⎝ x⎠
⎪1 ⎛ x⎞
⎪ ∑ n =1 ln ⎜ ⎟
N
if a = 0
⎪⎩ N ⎝ xn ⎠
where we replace the term x̄ by WA(x) (the arithmetic mean), and where
WGM(x; a) denotes the general mean of order a. Thus, a generalized
entropy measure for any a ≠ 0,1 may be easily calculated once we know the
arithmetic mean and the general mean of order a.
For a = 1, the generalized entropy is Theil’s first measure of inequality
and can be written as
1 N xn ⎛ xn ⎞
IT1 (x) = ∑ ln ⎜
N n =1 WA (x) ⎝ WA (x) ⎟⎠
. (2.28)
This is the only measure in this class that cannot be expressed as a function
of general means and does not have a natural twin-standards representation.
For a = 0, the generalized entropy index is Theil’s second measure of
inequality, which is also known as the mean logarithmic deviation and can be
expressed as a function of the arithmetic mean, WA(x), and the geometric
mean, WG(x), as follows:
WA (x)
IT 2 (x) = ln WA (x) − ln WG (x) = ln . (2.29)
WG (x)
Besides the two Theil measures, the other commonly used measure in
the entropy class is the index for a = 2, which is closely related to the coef-
ficient of variation (CV). The CV is the ratio of the standard deviation and
97
A Unified Approach to Measuring Poverty and Inequality
mean. For a = 2, the general entropy measure is half the CV squared and
can be expressed as
2 2
1 ⎢⎣WE (x)⎢⎣ − ⎢⎣WA (x)⎢⎣ 1 Var(x) CV 2
IGE (x; 2) = 2
= 2
= , (2.30)
2 2 2
⎣⎢ WA (x)⎣⎢ ⎣⎢ WA (x)⎢⎣
where Var(x) is the variance of the distribution x, which is the square
of its standard deviation. In equation (2.29), WE(x) is the Euclidean
1
mean, as in equation (2.8) and [WE (x)]2 = ∑ n =1 x 2n . Clearly,
N
1 N
⎢⎣WE (x)⎢⎣ − ⎢⎣WA (x)⎢⎣ = ∑ n =1 x n − x is the variance of x (see example 2.8).
2 2 N 2 2
Example 2.8: Consider the income vector x = ($2k, $4k, $8k, $10k)
that we used in the general means example. The arithmetic mean
is WA(x) = $6k, the geometric mean is WG(x) = $5.03k, and the
Euclidean mean is WE(x) = $6.78k.
We now calculate the two Theil inequality measures and the
squared coefficient of variation:
IGE(x; 2) = ([WE(x)]2 − [WA(x)]2)/(2[WA(x)2] = (6.782 − 62)/(2 × 62)
= 0.279.
IT2(x) = ln[WA(x)/WG(x)] = ln [$6k/$5.03k] = 0.176.
The calculation of Theil’s first measure is not as straightforward
as that of the previous two measures. However, it can be calculated
using the following steps. First, create a new vector from vector x by
dividing every element by the mean of x as (2/6, 4/6, 8/6, 10/6). Then
1 ⎡ 2 ⎛ 2 ⎞ 4 ⎛ 4 ⎞ 8 ⎛ 8 ⎞ 10 ⎛ 10 ⎞ ⎤
IT1 (x) = ln ⎜ ⎟ + ln ⎜ ⎟ + ln ⎜ ⎟ + ln ⎜ ⎟ = 0.15.
4 ⎢⎣ 6 ⎝ 6 ⎠ 6 ⎝ 6 ⎠ 6 ⎝ 6 ⎠ 6 ⎝ 6 ⎠ ⎥⎦ .
98
Chapter 2: Income Standards, Inequality, and Poverty
and a limiting case where a = 1. For the lower range, a < 1, measures in this
class are monotonic transformations of the Atkinson’s class of measures and
can be written as
⎧ ⎢⎣1 − IA (x ;a − 1)⎥⎦a
⎪ if a ≠ 0, a < 1
⎪ a (a − 1)
IGE (x; a ) = ⎨ , (2.31)
⎪ ln 1
if a = 0
⎪⎩ 1 − IA (x; 0)
99
A Unified Approach to Measuring Poverty and Inequality
The Gini coefficient and the inequality measures in Atkinson’s family share
a social welfare interpretation. As we have already discussed, they can be
expressed as I = (x̄ − a)/x̄, where x̄ is the mean income of the distribution
x and a is an income standard that can be viewed as a welfare function
(satisfying the weak transfer principle). Note that the distribution in which
everyone has the mean income has the highest level of welfare among all
distributions with the same total income, and the distribution’s measured
welfare level is just the mean itself. This finding results from the normaliza-
tion property of income standards.
Thus, the mean WA(x) = x̄ is the maximum value that the welfare func-
tion can take over all income distributions of the same total income. When
incomes are all equal, a = WA(x) and inequality is zero. When the actual
welfare level a falls below the maximum welfare level WA(x), the percentage
welfare loss I = (WA(x) − a)/WA(x) is used as a measure of inequality. This
is the welfare interpretation of both the Gini coefficient and the Atkinson’s
class of measures.
The simple structure of these measures allows us to express the welfare
function in terms of the mean income and the inequality measure. A quick
rearrangement leads to a = WA(x)(1 – I), which can be reinterpreted as
saying that the welfare function a can be viewed as an inequality-adjusted
mean. If there is no inequality in the distribution, then (1 – I) = 1 and
a = WA(x). If the inequality level is I > 0, then the welfare level is obtained
by discounting the mean income by (1 – I) < 0.
For example, if we take I to be the Gini coefficient, IGini(x), then the Sen
mean (or Sen welfare function) can be obtained by multiplying the mean by
[1 – IGini(x)], that is, WS(x) = WA(x)[1 – IGini(x)]. Similarly, if we take I to
be the Atkinson’s measure with parameter a = 0, IA(x; 0), then the welfare
function is the geometric mean, and the geometric mean can be obtained by
multiplying the mean by [1 – IA(x; 0)], that is, WG(x) = WA(x)[1 – IA(x; 0)].
100
Chapter 2: Income Standards, Inequality, and Poverty
101
A Unified Approach to Measuring Poverty and Inequality
Lx(100) 100
Ly Lx ′
Income share
Lx
Ly(20) A 20
Lx ′(20) B 14
Lx(20) C 4
0
0 20 100
Population share
102
Chapter 2: Income Standards, Inequality, and Poverty
The twin-standard view of inequality offers fresh insights into the relation-
ship between growth and inequality. Almost all inequality measures are
constructed in terms of a larger income standard b and a smaller income
standard a, and these income standards are expressed as 1 – a/b or b/a – 1.
Suppose income standard a changes to a' over time with growth rate g– a,
103
A Unified Approach to Measuring Poverty and Inequality
that is, a' = (1 + g– a)a, and income standard b changes to b' over time with
growth rate g– b, that is, b' = (1 + g– b)b. The inequality measure then changes
from I = 1 – a/b to I' = 1 – a'/b'. To have a fall in inequality, we require I' <
I or 1 – a'/b' < 1 – a/b, which occurs when g– a > g– b. Therefore, for a reduction
in inequality, the smaller income standard a needs to grow faster than the
larger income standard b.
Consider the example of the Gini coefficient, which is constructed from
two income standards. The larger income standard is the arithmetic mean
WA, and the smaller income standard is the Sen mean WS. Let us denote
the growth rate of the mean income by g– and the growth rate of the Sen
mean by g–S. The Gini coefficient will register a fall in inequality when the
growth rate of the Sen mean is larger than the growth rate of the arithmetic
mean, that is, g–S > g–. Similarly, inequality over time, in terms of the Gini
coefficient, increases when g–S < g–.
What about the Atkinson’s measures and the generalized entropy mea-
sures? Measures in these classes, including Theil’s second measure, are based
on the arithmetic mean and on any income standard from the class of gen-
eral means. For a < 1, the arithmetic mean is the larger income standard,
and the other general mean–based income standard is the smaller income
standard. In this case, if the growth rate of the smaller income standard of
order a is denoted by g–GM(a), then inequality decreases when g–GM(a) > g–.
If inequality is evaluated by Theil’s second index, then inequality falls when
the growth of geometric mean g–GM(0) is larger than that of the arithmetic
mean, that is, g–GM(0) > g–. For a > 1 in the generalized entropy measure,
the arithmetic mean is the smaller income standard, and the other general
mean–based income standard is the larger one. Inequality falls, according
to these indices, when the growth rate of the arithmetic mean g– is higher.
Is there any way to tell if all inequality measures in the Atkinson family
and the generalized entropy family have fallen? Yes, it is possible to do so
just by looking at the general mean growth curve, as described in figure 2.12.
A generalized mean growth curve is the loci of the growth rates of all
income standards in the class of general means. Comparing distributions x
and y for the general mean growth curve gGM(x,y) in figure 2.12 shows that
all inequality measures in Atkinson’s class and the generalized entropy class
agree that the inequality has fallen because the growth rates of the lower
income standards are higher than ḡ. The growth rates of the larger income
standards are lower than ḡ. However, for the general mean growth curve
gGM(x',y') in the same figure, all inequality measures in Atkinson’s class and
104
Chapter 2: Income Standards, Inequality, and Poverty
the generalized entropy class agree that the inequality has risen because the
growth rates of the lower income standards are lower than ḡ, whereas the
growth rates of the larger income standards are higher than ḡ.
In a similar manner, the growth incidence curve may be used to under-
stand the change in inequality using quantile ratios. If the growth rate of
the upper quantile income is larger than the growth rate of a lower quantile
income, then inequality has risen over time. In contrast, if the growth rate
of a lower quantile income is larger than the growth rate of the higher quan-
tile income, then inequality has fallen. For example, consider the growth
incidence curve gQI(x,y) in figure 2.10. If inequality is measured by the
90/10 measure IQR(x; 90/10), then inequality has fallen. Furthermore, for
growth incidence curve gQI(x',y'), the level of inequality has increased for
the same inequality measure.
Poverty Measures
105
A Unified Approach to Measuring Poverty and Inequality
nonpoor. In this example, society X has two poor people and two nonpoor
people. We summarize the incomes of the poor in vector x by the vector xq.
Poverty analysis is concerned only with the poor or the distribution’s
base, which should be the group targeted for public assistance. It naturally
ignores the incomes of nonpoor people in a society. In this way, the identifi-
cation step allows us to construct a censored distribution or censored vector
of incomes for society X, which we denote by x* = (x*1,x*2, …,x*N) such that
x*n = xn if income xn is less than the poverty line z and xn* = z if income xn is
greater than or equal to the poverty line z.
For the four-person income vector x = ($1k, $2k, $50k, $70k) in the
previous example, the censored vector is denoted by x* = ($1k, $2k, $10k,
$10k). Notice that incomes of the two nonpoor people are replaced by
the poverty line, and portions of their income above the poverty line are
ignored. A policy maker’s objective should be to include poor people at or
above the poverty line. Including all poor people at or above the poverty
line results in a nonpoverty censored distribution of income. We denote the
nonpoverty censored distribution of society X corresponding to poverty line
z by x– z* such that x– z* = (z,z,…,z).
The second step for constructing a poverty measure is aggregation. In this
step, incomes of individuals who are identified as poor using the poverty
line in the identification stage are aggregated to obtain a poverty measure.
Therefore, a poverty measure depends on both the incomes of the poor and
the criterion that is used for identifying the poor—that is, the poverty line.
In fact, it turns out that any poverty measure is obtained by aggregating ele-
ments in the censored distribution x∗.
In this section, we denote a poverty measure by P, where specific indi-
ces are denoted using corresponding subscripts. We denote the poverty
measure of distribution x for poverty line z by P(x; z). Alternatively, it may
be denoted by P(x∗). There are two different ways to understand a poverty
measure: one is based on the properties it satisfies and the other is through
its link with income standards. First, we discuss the properties that a poverty
measure should satisfy.
Desirable Properties
106
Chapter 2: Income Standards, Inequality, and Poverty
107
A Unified Approach to Measuring Poverty and Inequality
The fifth and final axiom in the invariance properties is focus, which
requires that if the income of a nonpoor person in a society changes but
does not fall below the poverty line, then the level of poverty should not
108
Chapter 2: Income Standards, Inequality, and Poverty
change. This property ensures that the measure focuses on the poor incomes
in evaluating poverty. In fact, focus ensures that the income distribution
is censored at the poverty line before evaluating a society’s poverty. For
example, suppose the initial income vector is x = ($1k, $2k, $50k, $70k)
and the poverty line income is $6k. Thus, the third person and the fourth
person are nonpoor. If the income of either the third or the fourth person
increases, but the poverty line remains unaltered at $6k, then the society’s
poverty level does not change.
The next group of properties are dominance properties. The first of these
properties requires that if the income of a poor person in a society increases,
then the poverty level should register a fall, or at least it should not increase.
There are two versions of this property. One is weak monotonicity, which
requires that poverty should not increase because of an increase in a poor
person’s income. The other is monotonicity, the stronger version, which
requires that poverty should fall if a poor person’s income in the society
increases.
These two properties are the same as the two corresponding properties
of income standards, except the ones introduced here are solely concerned
with incomes of the poor. For example, suppose the initial income vector
is x = ($1k, $2k, $50k, $70k) and the poverty line income is $6k so that
the first two people are identified as poor. If a new vector x' is obtained by
increasing the income of either the first or the second person, while the
poverty line remains unchanged, then according to the weak monotonicity
property, poverty should not be higher in x', and, according to the monoto-
nicity property, poverty should be lower in x'.
109
A Unified Approach to Measuring Poverty and Inequality
110
Chapter 2: Income Standards, Inequality, and Poverty
poverty line, or between two moderately poor people, who are much closer
to the poverty line.
Should the effect of transfer, no matter where it takes place, have
the same effect on the poverty level? We elaborate this situation with an
example. Consider the five-person income vector x = ($80, $100, $800,
$50, 000, $70,000). Let the poverty line be set at $1,050. Then the first four
people are identified as poor because their incomes are below the poverty
line. First, suppose $10 is transferred from the second person to the first per-
son. Then the post-transfer income vector is x' = ($90, $90, $800, $1,000,
$50,000, $70,000). Transferring 10 percent of the second person’s income
has increased the first person’s income by 12.5 percent.
Suppose, instead, that the same $10 transfer takes place between the
third and the fourth persons, who are also poor. The post-transfer income
vector is x'' = ($80, $100, $810, $990, $50,000, $70,000), where transfer-
ring 1 percent of the fourth person’s income increases the third person’s
income by 1.25 percent. This transfer makes hardly any difference in the
large pool of income of the two richer poor people. Therefore, one might
feel that a transfer of the same amount between two extreme poor and
two richer poor should not have the same effect on the society’s overall
poverty.
The third dominance property, transfer sensitivity, requires a poverty
measure to be more sensitive to a transfer between poor people at the lower
end of the income distribution of the poor. In other words, this property
requires that a poverty measure should change more when a transfer takes
place between two extremely poor people than between two richer poor
people. In terms of the example above, the level of deprivation should be
lower in x' than in x''.
Suppose the initial income distribution is x and distribution x" is obtained
from distribution x by a progressive (or regressive) transfer between two
extremely poor people. Suppose further that distribution x" is obtained from dis-
tribution x by a progressive (or regressive) transfer of the same amount between
two richer poor people. The following is the transfer sensitivity property:
111
A Unified Approach to Measuring Poverty and Inequality
112
Chapter 2: Income Standards, Inequality, and Poverty
113
A Unified Approach to Measuring Poverty and Inequality
poverty measures, and the second category lists advanced poverty measures.
There are two basic poverty measures in common use: headcount ratio and
poverty gap measure.
Headcount Ratio
The headcount ratio (PH) is a crude measure of poverty that simply counts
the number of people whose incomes are below the poverty line z and
divides that number by the total number of people in the society. In society
X with population size N, if there are q poor people, then the headcount
ratio is simply q/N. It is obvious that the headcount ratio lies between zero
and one. If all people are poor in a society, then the headcount ratio is one.
When there are no poor, it is zero.
The headcount ratio can also be understood using income standards
applied to the nonpoverty censored distribution and a doubly censored dis-
tribution. What is a doubly censored distribution, and how do we obtain
it? A doubly censored distribution x** is obtained from an original income
distribution x by replacing nonpoor incomes with the poverty line income z
and by replacing the poor incomes with zero. Therefore, income distribution
x is censored upward at poverty line z for nonpoor and again censored at zero
for the poor. The term doubly censored comes from the fact that distribution
x*z* is obtained by censoring distribution x twice.
The arithmetic mean is the income standard used to understand head-
count ratio. The arithmetic mean of the nonpoverty censored distribution is
poverty line z, and the arithmetic mean of the doubly censored distribution
is called the dichotomous mean. If there are q poor people, or N − q nonpoor
people, in society X, then the dichotomous mean of the society is
N− q N− q
WA (x ** ) = q × 0 + z= z. (2.35)
N N
N−q
WA (xz* ) − WA (x** ) z− z
N q
PH (x; z) = *
z
= = . (2.36)
WD (xz ) z N
114
Chapter 2: Income Standards, Inequality, and Poverty
The headcount ratio is the most well-known and most widely used
poverty measure because its interpretation is highly intuitive and simple.
However, the effectiveness of the headcount ratio depends on which prop-
erties the headcount ratio satisfies. It satisfies all invariance properties:
symmetry, normalization, population invariance, scale invariance, and focus.
However, it does not satisfy any dominance property except subgroup consis-
tency. The headcount ratio is not sensitive to changes in the income level
of the poor as long as incomes do not cross the poverty line. This is why
the headcount ratio does not satisfy the other dominance properties and
monotonicity, which require poverty measures to change as the incomes of
the poor change. The headcount ratio satisfies subgroup consistency because
the headcount ratio is additively decomposable, as shown by example 2.11.
The second basic poverty measure is the poverty gap measure. Like headcount
ratio, it is also widely used. The poverty gap measure (PG) is the average
normalized shortfall with respect to the poverty line across the poor. In
society X, the normalized income shortfall of a person, say, n, is calculated as
(z − x*n)/z, which means that the normalized income shortfall of a nonpoor
person is zero. The average normalized income shortfall is the average of all
normalized income shortfalls within a society. We denote the normalized gap
vector of x by g* = ((z − x*1)/z,…,(z − x*N)/z). Then the poverty gap measure is
115
A Unified Approach to Measuring Poverty and Inequality
1 N z − x *n
PG (x; z) = WA (g * ) = ∑
N n =1 z
. (2.37)
The poverty gap measure lies between zero and one. Zero is obtained
when there are no poor in the society. A value of one is obtained when
116
Chapter 2: Income Standards, Inequality, and Poverty
everyone in the society is poor and has zero income. When everyone in
a society is poor, then the poverty gap measure is the average normalized
income shortfall among the poor, PIG, because the headcount ratio is one in
this situation, that is, PH = 1 (see example 2.12).
What properties does the poverty gap measure satisfy? It satisfies all
invariance properties: symmetry, normalization, population invariance, scale
invariance, and focus. Among dominance properties, it satisfies only mono-
tonicity and subgroup consistency and does not satisfy the transfer principle
and transfer sensitivity. Although it does not satisfy the transfer principle, it
satisfies the weak transfer principle, which means that the poverty gap mea-
sure does not increase (or decrease) because of a regressive (or progressive)
transfer but also does not fall (or increase). The poverty gap measure satis-
fies the monotonicity property, meaning that if the income of a poor person
increases, then (unlike the headcount ratio) the poverty gap increases. The
poverty gap measure satisfies subgroup consistency because, like the head-
count ratio, it is additively decomposable.
117
A Unified Approach to Measuring Poverty and Inequality
Watts Index
The Watts index was proposed by Watts (1968), and it is the average dif-
ference between the logarithm of the poverty line and the logarithm of
incomes. For income distribution x with population size N and poverty line
z, the Watts index can be written as
1 N
PW (x; z) = ∑ (ln z − ln x *n).
N n =1
(2.40)
The lowest value the Watts index can take is zero, which is obtained
when no one is poor in the society. However, unlike the headcount ratio
and the poverty gap measure, the Watts index has no maximum value.
Like the two basic measures, the Watts index can also be expressed as a
difference between two income standards. The income standard used for the
headcount ratio and the poverty gap measure is the arithmetic mean, where-
as the income standard for the Watts index is the geometric mean. The
higher income standard is obtained by applying the geometric mean to the
nonpoverty censored distribution x̄*z. Because the geometric mean satisfies
normalization, the higher income standard is equal to the common ele-
ment in x*, which is the poverty line z itself. The lower income standard is
obtained by applying the geometric mean to the censored income distribu-
tion x*. The Watts index is the logarithm of the ratio of the higher and the
lower income standards.
The other way of interpreting the measure is by calculating the differ-
ence of their logarithms (see example 2.13). The formulation of the Watts
index in terms of income standards is
⎡ W (x * ) ⎤ ⎡ z ⎤
PW (x; z) = ln ⎢ G z ⎥ = ln ⎢ ⎥ = ln z − ln ⎡⎣WG (x * )⎤⎦ . (2.41)
⎣ WG (x* ) ⎦ ⎣ WG (x * ) ⎦
118
Chapter 2: Income Standards, Inequality, and Poverty
Sen-Shorrocks-Thon Index
119
A Unified Approach to Measuring Poverty and Inequality
among the poor, then PSST(x;z) reaches its minimum. As inequality increases,
the values of PSST(x;z) increase because 1 − PG(x;z) > 0, which results from
the fact that PG(x;z) lies between zero and one. The Gini coefficient lies
between zero and one. When there are no poor in a society, the SST index is
zero. The maximum value of one is obtained when everyone in the society is
poor and has zero income.
The SST index has an interesting relationship with the average normal-
ized income shortfall among the poor, PIG. When everyone is poor in a
society, but has equal income, then the SST index is equal to the average
normalized income shortfall among the poor, that is, PSST(x;z) = PIG(x;z).
This is because in this situation IGini(x*) is zero and PH = 1. When the
inequality level among the poor increases while the average normalized
income shortfall remains the same, the SST index becomes larger than the
average normalized income shortfall.
The SST index can also be interpreted by an income standard. The
income standard in this case would be the Sen mean. The SST index is the
normalized difference between the Sen mean of the nonpoverty censored
distribution and the Sen mean of the censored distribution. The Sen mean
satisfies the normalization property of income standards. Thus, the Sen
mean of the nonpoverty censored distribution is the poverty line itself, that
is, WS(x̄*) = z. The Sen mean of the censored distribution x* is denoted by
WS(x*). The SST index16 can be presented as
WS (xz* ) − WS (x* ) z − WS (x * )
PSST (x; z) = = . (2.43)
WS (xz* ) z
Given a censored distribution, once the Sen mean is calculated using the
procedure discussed in the income standard section, the SST index can eas-
ily be obtained by applying equation (2.43). How do equations (2.42) and
(2.43) give the same result? That question can easily be answered as
z − WS (x * ) z − WA (x * ) WA (x * )
= + IGini (x * ) = PG + (1 − PG )IGini (x * ). (2.44)
z z z
In the previous section, when discussing dominance and ambiguity
results for income standards, we mentioned that the Sen mean is related to
the generalized Lorenz curve. The SST index is based on the Sen mean and
thus is naturally related to the generalized Lorenz curve, which has been
graphically depicted in Zheng (2000). Example 2.14 shows how to calculate
the SST index.
120
Chapter 2: Income Standards, Inequality, and Poverty
What properties does the SST index satisfy? It satisfies all invariance
properties: symmetry, normalization, population invariance, scale invariance,
and focus. However, it does not satisfy all dominance properties because it
is based on the poverty gap measure and the Gini coefficient. It inherits
the monotonicity property from the poverty gap measure, and it inherits
the transfer principle from the Gini coefficient. However, neither the Gini
coefficient nor the poverty gap ratio satisfies transfer sensitivity; conse-
quently, the SST index does not satisfy transfer sensitivity. Furthermore,
the Gini coefficient is neither subgroup consistent nor additively decom-
posable in the usual way. This shortcoming is also inherited by the SST
index.
Despite these shortcomings, the SST index is useful because it can be
broken down into the poverty gap measure and the Gini coefficient. In fact,
the poverty gap measure can be further broken down into the headcount
ratio (PH) and the average income gap of the poor (PIG).
The next poverty measure in the advanced measures category is the squared
gap measure. This measure is calculated by averaging the square of the nor-
malized income shortfalls and is denoted by
2
1 N ⎛ z − x *n ⎞
PSG (x; z) = ∑
N n =1 ⎜⎝ z ⎟⎠
. (2.45)
One way of interpreting the squared gap measure is as the weighted aver-
age of normalized income shortfalls, where each normalized income shortfall
121
A Unified Approach to Measuring Poverty and Inequality
PSG(x;z) = PH[PIG
2 + 2(1 − P )2I (xq;2)].
IG GE (2.46)
The squared gap measure lies between zero and one (see example 2.15).
A zero value is obtained when there are no poor people in the society
because the headcount ratio is zero. The maximum value of one is reached
when everyone in the society is poor and has zero income.
What properties does the squared gap measure satisfy? It satisfies all
invariance properties: symmetry, normalization, population invariance, scale
invariance, and focus. However, among the dominance properties, it satisfies
monotonicity, the transfer principle, and subgroup consistency, but it does not
satisfy transfer sensitivity because the headcount ratio, the income gap ratio,
and the generalized entropy of order 2 do not satisfy this property. Hence,
like the basic poverty measures and the SST index, the squared income gap
122
Chapter 2: Income Standards, Inequality, and Poverty
measure is transfer neutral. However, unlike the SST index, it satisfies sub-
group consistency because it is additively decomposable.
123
A Unified Approach to Measuring Poverty and Inequality
Note that this formula is not valid for the headcount ratio when a = 0.
For the example above, P'FGT(x;z,a) = 1/2 for all a > 0 because the general
mean satisfies the normalization property of income standards.
The mean gap measure of poverty can be obtained by taking the Euclidean
mean (WE) of the normalized income shortfalls. This is a monotonic trans-
formation of the squared gap measure. More specifically, the mean gap mea-
sure is the square root of the squared gap measure. The mean gap measure
can be expressed as
124
Chapter 2: Income Standards, Inequality, and Poverty
1
1 ⎛ 1 N ⎛ z − x* ⎞ 2 ⎞ 2
PMG (x; z) = WE (g ) = P = ⎜ ∑ ⎜
* 2
SG
n
⎟⎠ ⎟ . (2.50)
N
⎝ n =1 ⎝ z ⎠
PMG = PSG = PH ⎡⎣P12G + z (1 − P1G )2 IGE (xa; z)⎤⎦ = P12G = P1G . (2.51)
125
A Unified Approach to Measuring Poverty and Inequality
and is the normalized shortfall of the generalized mean of the observed cen-
sored income distribution x* from the generalized mean of the nonpoverty
censored income distribution x̄*. Again, the generalized mean satisfies the
normalization property of income standards; thus, the generalized mean of
the nonpoverty censored income distribution is the poverty line itself. The
CHUC index for a ≤ 1 can be expressed as
The CHUC index lies between zero and one. The minimum value of
zero is obtained when there are no poor people in a society. However, the
maximum value of the CHUC index cannot be larger than one. When
everyone in a society is poor, having equal income, this measure is equal
to the average normalized income shortfall. It satisfies all invariance and
dominance properties. However, not all measures in this class are addi-
tively decomposable. For a = 1, the CHUC index is the poverty gap mea-
sure, and for a = 0, the CHUC index is a monotonic transformation of the
Watts index.
We have shown that the two basic measures—the headcount ratio and the
poverty gap measure—do not satisfy transfer-related properties and so are
not sensitive to inequality across the poor. Besides not being sensitive to
inequality, the headcount ratio does not satisfy monotonicity, which, if it is
used as a target for public policy, may cause inefficiency in public spending.
All of the subsequent advanced poverty measures, in contrast, are sensitive
to inequality across the poor. The SST index and the mean gap measure are
both equal to the poverty gap measure when everyone in a society is poor
and no inequality exists among them. These two measures become larger
than the poverty gap measure when the income gap remains the same, but
inequality among the poor increases.
Each advanced measure, however, has its own pros and cons. Let us
begin with the SST measure. We know from our previous discussion that
this measure is not subgroup consistent, which means that it may lead to
inconsistent outcomes when group-level analysis is of interest. This measure
is also not transfer sensitive, which means that if a similar amount of transfer
126
Chapter 2: Income Standards, Inequality, and Poverty
takes place between two poorer poor people and two richer poor people,
then this measure cannot distinguish between the two situations.
What, then, are the SST index’s advantages? The first is that it can
be neatly broken down into the headcount ratio, the average normalized
income shortfall among the poor, and the well-known Gini coefficient. If
one is not interested in group-level analysis, then this measure can be bro-
ken down into these three components to understand the source of change
in poverty. In fact, the Gini coefficient can be broken down further into a
within-group and a between-group component using the Gini decomposi-
tion formula introduced earlier. The within-group component assesses
inequality among the poor, and the between-group component measures
inequality between the average income of the poor and the poverty line.
This decomposition reveals whether the change in the measure’s inequal-
ity component is caused by the change in inequality among the poor or due
to a change in the average income of the poor compared to the poverty line.
Note that there is no within-group inequality among the nonpoor because they
all have the same income equal to the poverty line. Furthermore, there is no
residual term, which is commonly seen in the Gini decomposition, because
there is no income overlap between the poor and the nonpoor.
Second, consider the squared gap measure. This measure has many posi-
tive features, such as it is additively decomposable and subgroup consistent.
Furthermore, like the SST index, it can be broken down into the head-
count ratio, the average normalized income shortfall among the poor, and
the generalized entropy measure order of 2 among the poor to understand
the poverty composition. However, like the SST index, this measure is not
transfer sensitive, which means that if a similar amount of transfer takes
place between two poorer poor people and two richer poor people, then this
measure cannot distinguish between these two situations.
Also, the generalized entropy measure order of 2 may be a bit unintuitive
in the sense that it may range from zero to infinity, unlike the Gini coefficient
that ranges from zero to one. The same pros and cons apply to the mean gap
measure, which is just a monotonic transformation of the squared gap measure.
Third, consider the Watts index. This measure appears to be a perfect
measure of poverty in the sense that it satisfies all the properties that we dis-
cussed earlier: it is additively decomposable, is transfer sensitive, and satisfies
the transfer principle and all other properties. However, this measure has two
shortcomings. One is that it is not applicable when there are zero incomes
because the logarithm of zero is undefined. The second shortcoming is that
127
A Unified Approach to Measuring Poverty and Inequality
it does not have an intuitive interpretation like the two basic measures, the
SST index and the squared gap measure and its monotonic transformation
(the mean gap measure). Also, like these other measures, it does not have an
upper bound of one. Finally, the CHUC class of indices is a generalization of
the Watts index. Like the Watts index, its members satisfy all the properties
discussed earlier and also lie between zero and one. However, measures in
this class are not defined for zero incomes when α ≤ 0.
128
Chapter 2: Income Standards, Inequality, and Poverty
poor person (with an income of $1,000). The next $300 would be spent on
the second-richest poor person (with an income of $800).
After bringing these two poor people out of poverty, the policy maker
still has $600 in his or her budget that remains unused. How and whom
should this amount assist? Given that the headcount ratio does not satisfy
the monotonicity property, because even if this entire amount is transferred
to either of the two remaining poor people, the poorest people still remain
under the poverty line and do not add to the headcount ratio. The policy
maker in this situation would have no incentive to spend the remaining
budget. This lack of incentive creates inefficiency in public spending.
Although poverty is reduced by 50 percent, the poverty status of the two
severely deprived people remains unchanged.
What if the society’s poverty is assessed by the poverty gap measure? Recall
that, unlike the headcount ratio, the poverty gap measure satisfies monotonicity;
but, like the headcount ratio, it does not satisfy the transfer principle or transfer
sensitivity. Thus, it is not sensitive to inequality among the poor. What implica-
tion does it have on the targeting exercise? In this case, the policy maker will
be inclined to spend his or her entire budget because the poverty gap measure
satisfies monotonicity. An increase in a poor person’s income, even when he or
she is not driven out of poverty, reduces the poverty gap measure. Therefore,
unlike the headcount ratio, inefficiency in public spending does not arise.
Then how should the budget of $1,000 be allocated among the poor? The
straightforward way is to spend the budget on any of the four poor people as
long as they do not surpass the poverty line income. Given that the poverty
gap measure is not sensitive to inequality among the poor, it does not matter
who among the poor receives the assistance. For example, in one case, out
of the budget of $1,000, the richest poor person, with an income of $1,000,
may receive $100; the second-richest poor person may receive $300; and the
third-richest poor person may receive the rest, or, in another case, the poor-
est person, with an income of $80, may receive the entire amount. In both
cases, the improvements in the poverty gap measure are the same. Thus, the
poverty gap measure is insensitive to whoever receives the assistance. The
poorest section of a society may perpetually remain poor in spite of showing
decent progress in terms of the poverty gap measure.
How would this policy exercise be affected when the society’s poverty is
gauged by a distribution-sensitive poverty measure? A distribution-sensitive
measure requires that assistance should go to the poorest of the poor first.
Thus, out of the $1,000 budget allotted for the poor, the first $20 should go
129
A Unified Approach to Measuring Poverty and Inequality
to the poorest person whose income is $80 so that the incomes of the two
poorest poor people are made equal. Then the rest of the budget should be
equally divided between the two poorest people so that, after allocating
the entire budget, the income distribution becomes x' = ($590, $590, $800,
$1,000, $50,000, $70,000).
What if, instead of $1,000, there was $1,600 allotted to the welfare of
the poor? Then the first $20 would be transferred to the poorest person.
Next, out of $1,580, $1,400 would be divided equally between the two poor-
est people so that the incomes of all three of the poorest people would be
equalized at $800. Finally, the rest of the budget of $180 is equally divided
among the three poorest poor so that the post-allocation income vector
is x" = ($860, $860, $860, $1,000, $50,000, $70,000). All distribution-
sensitive poverty measures support this type of targeting. However, not all
measures reflect similar amounts of decrease in poverty, which depends on
how these measures weight different people.
Like the targeting exercise, can a poverty measure influence public policy?
Consider an example of a developing country where the major staple food is
rice. As with other agricultural producers, rice producers are poor and their
incomes are scattered around the country’s poverty line income. Some rice
producers earn enough income to live just above the poverty line, but many
rice producers are unfortunate enough to live below the poverty line.
There are other poor people in the country, such as those whose major
occupation is agricultural labor, plantation labor, or other unskilled jobs.
These poor people are the poorest in the country, and their major source
of energy and nutrition is the staple food, rice. Rice is, in fact, a necessary
commodity in that country, and the government controls its price.
Being benevolent, the government wants to see a reduction in poverty
by adjusting the price of rice. Which of the following two policy options
would reduce poverty?
130
Chapter 2: Income Standards, Inequality, and Poverty
producers would be adversely affected because their income would fall, and
rice consumers would benefit because their real incomes would increase.
Given that most rice consumers are poorer than rice producers, one does
not know whether more or fewer people would become poor. Thus, the
impact on the headcount ratio is uncertain.
However, if the price of rice increases, then producers gain, but the
poorer consumers lose because their real incomes fall. Given that the
already poor consumers become poorer, this is not taken into account by
the headcount ratio because it does not satisfy monotonicity. Therefore,
the number of poor people would most likely fall, thereby leading to a fall
in the country’s headcount ratio. Thus, the potential assessment of poverty
using the headcount ratio would incline the government to choose option
2 and increase the price because poverty, according to the headcount ratio,
would fall.
Note, however, that the decrease in the headcount ratio has ignored the
change in inequality among the poor. The marginally poor producers would
become better off because of the price increase, but the severely poor people
would be worse off for the same reason. This occurrence is very similar to the
idea of regressive transfer. The higher price paid by the poorer consumers is
obtained by the lesser poor producers as profit.
Any inequality-sensitive poverty measure, such as the squared gap,
the Watts index, or the SST index, would be sensitive to such inequality
among the poor. Suppose the poverty level in that country is now assessed
with one such measure that is sensitive to inequality among the poor. If
the government now chooses option 1 and reduces the price of rice, then
the poorer consumers benefit at the cost of a reduction in the producers’
income. The result is uncertain. If some producers become poorer than
some consumers, then the poverty measure may increase. But if the pro-
ducers remain less poor than the consumers, then the poverty measure
may fall.
However, if option 2 is chosen and the rice price rises, then inequality
among the poor increases and, most certainly, the poverty measure would
increase. Hence, the potential assessment of poverty using any inequality-
sensitive poverty measure would incline the government to not raise the
price because poverty, according to any inequality-sensitive measure, would
increase. The conclusion is that different poverty measures would incline the
government to choose different policies.
131
A Unified Approach to Measuring Poverty and Inequality
132
Chapter 2: Income Standards, Inequality, and Poverty
Poverty measures that satisfy the transfer principle are called distribution-
sensitive poverty measures. The distribution-sensitive poverty measures
133
A Unified Approach to Measuring Poverty and Inequality
introduced earlier were the Watts index, the SST index, the FGT family of
measures for α > 1, and the CHUC family of indices. Each of these distribu-
tion-sensitive poverty measures is built on a specific income or gap standard
that is closely linked to an inequality measure. For example, the Watts index
is closely linked with Theil’s second measure of inequality, the SST index
is closely linked with the Gini coefficient, the FGT family of indices for
α > 1 is linked with the generalized entropy measures, and the CHUC fam-
ily of indices is linked with Atkinson’s family of measures.
For the Watts index, SST index, and CHUC family of indices, the
inequality measure is applied to the censored distribution x*, with greater
censored inequality being reflected in a higher level of poverty for a given
poverty gap level. The FGT indices for α > 1, however, use generalized
entropy measures applied to the gap distribution g*, with greater gap inequal-
ity leading to a higher level of poverty for a given poverty gap level.
Recall from our earlier discussion in the income standard section that
certain income standards can be viewed as welfare functions, and this link
provides yet another lens for interpreting poverty measures. The Sen mean
used in the SST index and the general means for α ≤ 1 that are behind the
CHUC indices can be interpreted as welfare functions. In each poverty
measure, the welfare function is applied to the censored distribution to
obtain the censored income standard, which is now seen to be a censored
welfare function that takes into account poor incomes and only part of non-
poor incomes up to the poverty line. For these measures, poverty and cen-
sored welfare are inversely related—every increase in poverty can be seen as
a decrease in censored welfare.
134
Chapter 2: Income Standards, Inequality, and Poverty
The horizontal axis of figure 2.14 denotes income, and the vertical axis
denotes the values of a cumulative distribution function. If the poverty line
is set at z, then the headcount ratio is PH(x; z), which is the percentage of
people in distribution x who have incomes less than z. Similarly, PH(x; z')
and PH(x; z") are the headcount ratios of distribution x corresponding to
poverty lines z' and z'', respectively.
135
A Unified Approach to Measuring Poverty and Inequality
PH(x; z ′)
PH(x ′; z)
PH(x; z)
z z′ z″ xN
Income
Suppose there is another distribution x'. One can see in figure 2.14 that
the headcount ratios corresponding to poverty lines z, z', and z" lie above
the respective headcount ratios for distribution x. Is there any other poverty
line that reflects a higher headcount ratio in x than in x'? The answer is no.
The cdf of x lies to the right of the cdf of x', which means that the headcount
ratio for x' for no poverty line can be lower than the headcount ratio for x.
When a cdf lies to the right of another cdf, first-order stochastic dominance
(introduced earlier) occurs. When such dominance relation holds between
two cdfs, not only do the headcount ratios agree for all poverty lines, but the
poverty gap measure, the squared gap measure, the mean gap measure, the
Watts index, and the CHUC indices also agree for all poverty lines.
This approach also answers the second question, which asks when all
poverty measures agree. Therefore, if the first-order stochastic dominance
holds, then there is no need to compare any two distributions by any poverty
measure introduced earlier with respect to varying the poverty line. The
choice of poverty measure and the choice of poverty line simply do not mat-
ter when the first-order dominance condition holds. The cdf in the context
of poverty measurement is also known as the poverty incidence curve.
136
Chapter 2: Income Standards, Inequality, and Poverty
Figure 2.15: Poverty Deficit Curve and the Poverty Gap Measure
C
Deficit
Fx
Dx
Fx(z)
Dx ′
A B
B
z xN z xN
Income Income
137
A Unified Approach to Measuring Poverty and Inequality
This can be easily verified from the quantile function as described earlier
in figure 2.5.
Recall that an income distribution’s cdf is just the inverse of the relevant
distribution’s quantile function. Thus, A is WA(xA)(xq)q/N = (x1 + … + xq )/N.
Another interpretation of area A is that it is the per capita income of an aver-
age poor person in the society. The combined area A + B denotes the society’s
per capita income, which, if held by each poor person, means that the poor will
not be poor anymore.
This per capita income is qz/N. Thus, area B, which is also the height of
the poverty deficit curve Dx at poverty line z, is the difference between the
area A + B and the area A, or the average income shortfall or the deficit,
that is, [z − WA(xq)]q/N. This deficit is the minimum per capita income of
the society, which, if transferred to the poor, will lift the poor out of poverty.
Area B is also zPG(x; z). The maximum height of the poverty deficit curve
is denoted by C, which is xN − WA(x).
Note that the larger height of the poverty deficit curve Dx compared
to the poverty deficit curve Dx' at z reflects a larger poverty gap measure
in distribution x than in distribution x' at poverty line z. It is evident from
figure 2.15 that the poverty deficit curve Dx lies above the poverty deficit
curve Dx' for all poverty lines. Hence, distribution x has higher poverty than
distribution x' for all poverty lines in terms of the poverty gap measure.
This type of unanimity result, however, fails to hold when two poverty
deficit curves cross each other. We should then check the poverty severity
curve of these two distributions. If the poverty severity curve of a distribu-
tion lies above the poverty severity curve of another distribution, then the
former distribution has higher poverty than the latter in terms of the squared
gap measure or the mean gap measure for all poverty lines.
138
Chapter 2: Income Standards, Inequality, and Poverty
Panel a of figure 2.16 displays the poverty deficit curve that we will use to
show how a poverty severity curve is constructed. As explained earlier, the
height B of a poverty deficit curve is proportional to the poverty gap measure
and is the poverty gap measure times the poverty line. As shown in panel b,
the height of the poverty severity curve Sx at poverty line z is D, which is the
area underneath the poverty deficit curve Dx. Area D is proportional to the
squared gap measure. Therefore, the larger the height of the poverty sever-
ity curve Sx than the poverty severity curve Sx at z, the larger the squared
gap measure in distribution x than in distribution x' at poverty line z. It
turns out that the poverty severity curve Sx lies above the poverty severity
curve Sx' for all poverty lines. Hence, distribution x has higher poverty than
distribution x' for all poverty lines.
Note that the dominance by the poverty deficit curve is equivalent to
the second-order stochastic dominance, and the dominance by the poverty
severity curve is equivalent to the third-order stochastic dominance.19
When there is dominance in terms of poverty incidence curves, all pov-
erty measures satisfying the invariance properties and monotonicity agree
with each other when ordering distributions according to the level of pov-
erty for any poverty line. Such dominance relationships do not always hold.
When two poverty incidence curves cross, one distribution has higher or
lower poverty only for a part of the entire range of incomes. In fact, different
poverty measures may order two distributions differently.
Figure 2.16: Poverty Severity Curve and the Squared Gap Measure
E
C
Severity
Deficit
Dx Sx
Sx ′
B D
D
z xN z xN
Income Income
139
A Unified Approach to Measuring Poverty and Inequality
The dominance analysis discussed earlier helps us understand whether one dis-
tribution has more or less poverty than another distribution. It is not concerned
about the level of poverty, which is often of particular policy interest. The num-
ber of poor people in a country or the fact that many poor people have been
moved out of poverty over a particular time period are always matters of great
concern. These data, of course, depend on the particular poverty line chosen.
As discussed in the introductory chapter, there are three different types
of poverty lines:
• An absolute poverty line may be adjusted with the rate of inflation over
time, but it is not adjusted with income growth over time.
• A relative poverty line is not fixed over time, and it changes with income
growth. For example, if a poverty line is set at 50 percent of the median
income, then the poverty line changes as the median income changes
over time. Or the poverty line may be set at 50 percent of mean
income. In this case, the growth rate of the poverty line over time is
the same as the growth rate of per capita income over time.
• A hybrid poverty line is created by taking a weighted average of an
absolute poverty line and a relative poverty line.
140
Chapter 2: Income Standards, Inequality, and Poverty
No matter how a poverty line is chosen, one can argue that it is arbitrary.
It is possible to propose a feasible alternative, which may change the
perspective of poverty significantly. Thus, one must examine the sensitiv-
ity of poverty with respect to the poverty line. One way of conducting the
sensitivity analysis is to change the poverty by certain percentages, then
estimate how much the poverty level has changed.
For example, suppose the headcount ratio of society x is 25 percent for
poverty line z = $10,000. Let this figure increase to 30 percent when the
poverty line is increased to $10,200. This means that a 2 percent increase in
the poverty line increases the headcount ratio by 5 percent. The lower the
change in the poverty estimate because of change in the poverty line, the
more reliable the point estimate based on a particular poverty line. If there is
too much variation, then the poverty estimate may not be considered reliable.
When a country is rapidly growing, one must evaluate the quality of the
growth. By growth, we generally mean a country or society’s growth in mean
income, and, by merely looking at the growth, there is no way of knowing
who has benefited from this growth. This growth may result from a rise in
incomes of the richer part of the distribution or from a rise in incomes of the
poorer part of the distribution.
There are various ways of understanding if the growth is pro-poor or
anti-poor. First, we may be interested in knowing directly if poverty has
increased or decreased because of the growth. Second, we may want to
know if the growth has relatively benefited or hurt the population with lower
incomes. In this case, it is not enough just to understand if poverty has
increased or decreased; it is also important to understand whether the situ-
ation of the poor has changed in comparison to others in the distribution.
Third, we may be interested in knowing if the growth has lowered poverty
more than a counterfactual-balanced growth path would. In this case, one
may be interested in knowing how much of the change in poverty is due to
growth and how much is due to the redistribution.
Consider some examples to clarify these various ways of understand-
ing pro-poor growth. Suppose the society consists of four people and the
income vector is x = ($80, $100, $200, $260). The society’s mean income
is $160. First, if the poverty line income is $120, then two people are
poor. Suppose that, over time, incomes of these four people change to
141
A Unified Approach to Measuring Poverty and Inequality
x' = ($100, $125, $160, $575). The society’s mean income has grown by
50 percent to $240. If the poverty line remains unchanged at $120, then
the headcount ratio goes down. In fact, poverty goes down for any poverty
measure that satisfies the monotonicity property. Thus, if one is merely
interested in knowing if poverty has decreased because of growth, then the
growth has been pro-poor for a fixed poverty line. If, instead of $120, the
poverty line is set at $180, then the change in poverty may not appear to
be pro-poor by all measures. For example, despite growth of 50 percent, the
headcount ratio deteriorates. Thus, in terms of the headcount ratio, the
growth in the distribution appears to be anti-poor.20
Given that a fixed poverty line is difficult to defend, we must understand
the change in poverty for a variable poverty line. The approach is analogous
to the dominance analysis. If one poverty curve (incidence, deficit, or sever-
ity) dominates another poverty curve, then poverty has improved unambigu-
ously in the dominant distribution because of growth. Besides merely knowing
the direction of change in poverty, we may be interested in the magnitude of
the reduction in poverty relative to the growth in mean—the growth elastic-
ity of poverty. The growth elasticity of poverty is defined as the percentage
change in poverty resulting from a 1 percent change in the mean income.
If the elasticity is greater than one, then the percentage change in poverty
has been larger than the percentage change in mean income, or the growth
of mean income. For an application of the growth elasticity of poverty using
the headcount ratio, see Bourguignon (2003). To understand the change in
the growth or elasticity of poverty for a variable poverty line, various poverty
growth curves can be constructed (similar to the various growth curves dis-
cussed in the income standard section).
A second way of understanding a change in poverty as pro-poor is by look-
ing at the gain of the poor relative to the gain in the mean. Reconsider the two
income vectors in the previous example. The growth rate of the mean was
50 percent. Have the incomes of individuals at the bottom of the distribution
improved enough to catch up with the growth in mean? The answer is no. The
growth of the poorest person’s income was 25 percent. The income growth of
the two poorest people also totaled 25.0 percent, and the growth of the three
poorest people totaled 1.3 percent. Then how was the 50 percent growth
achieved? It was achieved because the richest person’s income grew by about
121 percent. Thus, this second way understands the relationship between
poverty and growth from an inequality perspective and may be referred to as
an inequality-based approach, as discussed in chapter 1.
142
Chapter 2: Income Standards, Inequality, and Poverty
143
A Unified Approach to Measuring Poverty and Inequality
Exercises
There are five income categories (Xi) in the economy. Each category
contains a certain number of people (ni).
a. What is the total number of people (n) in the economy?
b. Let pi denote the proportion of people in each category. Fill in the
column corresponding to pi for each i. The probability mass function is
defined as a function that gives the probability of a discrete variable
taking the same value. Now draw the probability mass function.
Hint: Draw a diagram with x on the horizontal axis and p on the
vertical axis.
c. Let F(xi) denote the proportion of people who have an income no
higher than xi. Fill in the column corresponding to F(xi) for each i.
Now draw the cdf.
Hint: Draw a diagram with x on the horizontal axis and F(x) on
the vertical axis.
d. What is the relationship between pi and F(xi)?
e. Calculate the proportion of people having an income less than
$14,100. What is the proportion of people having an income more
than $14,900?
f. What is the average income for the economy?
g. Fill in the last column, and find the sum of all cells in that column.
What does the sum give you?
144
Chapter 2: Income Standards, Inequality, and Poverty
h. Use the cdf to calculate the area to the left of the cdf bounded by
x = 0 and F(x) = 1. What do you get?
i. Calculate the median, the 95th percentile, and the 20th percentile
using the cdf that you drew in 1c.
2. The Gini coefficient is probably the most commonly used index of
relative inequality. What are some of the advantages and disadvan-
tages of this measure?
3. The variance of logarithm (VL) is an inequality measure that is com-
puted as
1 N
VL (x) = ∑[ln x n − WL (x)]2,
N n =1
145
A Unified Approach to Measuring Poverty and Inequality
146
Chapter 2: Income Standards, Inequality, and Poverty
147
A Unified Approach to Measuring Poverty and Inequality
comparisons not only for a single poverty line but also for a range
of poverty lines. Consider the three distributions from the previous
example: x = (4,2,10), y = (4,1,10), u = (5,1,10).
a. If z = 6 is the poverty line, does x or y have more poverty accord-
ing to the headcount ratio? Will this determination be reversed at
some other poverty line? Explain. Does x or y have more poverty
according to the poverty gap measure? Will this determination be
reversed at some other poverty line? Explain.
b. If z = 6 is the poverty line, does x or u have more poverty accord-
ing to the headcount ratio? Will this determination be reversed at
some other poverty line? Explain. Does x or u have more poverty
according to the poverty gap measure? Will this determination be
reversed at some other poverty line? Explain.
c. Do you think unambiguous comparisons with variable poverty
lines might be made in practice? If not, why not? If so, why?
19. Which inequality measure is the Sen-Shorrocks-Thon (SST) poverty
index based on?
a. Explain why the SST index is not subgroup consistent and provide
a counterexample to illustrate your point.
b. Which inequality measure is the Foster-Greer-Thorbecke (FGT)
index PSG(x; z) based on? Show that the measure is subgroup con-
sistent.
20. Why should a measure of poverty satisfy scale invariance (homoge-
neity of degree 0 in incomes and the poverty line)? Which poverty
measures satisfy scale invariance?
21. Suppose instead of the PSG(x; z) measure one were to use the
PMG(x; z) measure.
a. What is the main constructive difference between these two
measures?
b. What would be the advantages and disadvantages of using the
PMG(x; z) measure?
22. Why do inequality decompositions have a between-group term but
poverty decompositions do not?
23. Suppose inequality decreases without growth of mean income. What
may likely happen to poverty? Suppose growth of mean income
occurs without a change in inequality. What may likely happen to
poverty? Explain.
148
Chapter 2: Income Standards, Inequality, and Poverty
24. Suppose that the per capita poverty gap measure is used with a rela-
tive poverty line that sets z = αμ for some α > 0. When does one
distribution have a lower level of relative poverty for all α > 0? (Hint:
Think Lorenz.)
25. We have already shown that the poverty measures are different from
each other and differ in their sensitivity to a distribution. Please pro-
vide certain examples with illustrative distributions and poverty lines
such that
a. The SST index rises, but the three FGT indices fall.
b. The headcount ratio rises, but the SST index, poverty gap mea-
sure, and squared gap measure fall.
c. The poverty gap measure rises, but the headcount ratio, SST, and
squared gap measures fall.
d. The squared gap measure rises, but the headcount ratio, poverty
gap measure, and SST measure fall.
Notes
149
A Unified Approach to Measuring Poverty and Inequality
150
Chapter 2: Income Standards, Inequality, and Poverty
15. Previously, Sen (1976b) proposed the index PS(x; z) = PH[PIG + (1 − PIG)-
IGini(xq)], where xq is the income distribution of the poor only. This mea-
sure was modified later by Thon (1979) and Shorrocks (1995).
16. For a more elaborated discussion on various formulations of the SST
index, see Xu and Osberg (2003).
17. Rawls’s welfare function maximizes the welfare of society’s worse-off
member. “Social and economic inequalities are to be arranged ... to the
greatest benefit of the least advantaged...” (Rawls 1971, 302).
18. For an in-depth discussion on poverty ordering, see Atkinson (1987),
Foster and Shorrocks (1988), and Ravallion (1994).
19. Note that the poverty deficit curve and the generalized Lorenz curve
have an interesting relationship. They are based on the area under-
neath the cdf and the quantile function, where a quantile function is
an inverse of a cdf. See figure 2.7.
20. For various approaches to measuring pro-poor growth for a fixed poverty
line, see Kakwani and Son (2008).
21. For a discussion on the poverty-growth-inequality triangle, see
Bourguignon (2003).
22. The growth-redistribution decomposition becomes a bit more compli-
cated when there is interregional migration. For such decomposition
with change in population, see Huppi and Ravallion (1991). An appli-
cation of their method can be found in table 30 of chapter 3.
References
151
A Unified Approach to Measuring Poverty and Inequality
152
Chapter 2: Income Standards, Inequality, and Poverty
153
Chapter 3
155
A Unified Approach to Measuring Poverty and Inequality
156
Chapter 3: How to Interpret ADePT Results
While preparing a report on poverty and inequality, one would first be inter-
ested in results at the national level. This part of the chapter contains seven
tables with results at the national level. We then decompose the results
across urban and rural areas.
Urban
0.008
Probability density function
0.006
0.004
0.002
Median
0
0 200 400 600 800
Welfare aggregate
2003 2006
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
157
A Unified Approach to Measuring Poverty and Inequality
In figure 3.1, the solid curve is urban Georgia’s density function for 2003,
and the dotted curve is the density function of urban consumption expendi-
ture distribution for 2006. The median is an important income standard that
can be found in the diagram. It is indicated by the corresponding vertical
lines: solid line for 2003 and dotted line for 2006.
A density function can also be useful for understanding a distribution’s
skewness. As can be seen from figure 3.1, the density functions for both years
are positively skewed. However, an important change from 2003 to 2006 is
that more people mass around the distribution’s median in 2006. We can
also see that the density functions for both years are unimodal. When more
than one mode exists, a society is considered to be polarized by consumption
expenditure or income.
Table 3.1 reports the mean and median per capita consumption expenditure
and their growth over time, and the inequality across the population using
the Gini coefficient. It also decomposes them across rural and urban areas
and across two years: 2003 and 2006. Table rows denote three geographical
regions: urban area, rural area, and all of Georgia (row 3). Per capita con-
sumption expenditure is measured in lari per month.
Columns A and B report the mean per capita consumption expenditure
for 2003 and 2006, respectively. Column C reports the percentage change
or growth in per capita expenditure over the course of these three years. The
average per capita expenditure of the urban area in 2003 is GEL 128.9 [1,A],
which is larger than the average rural per capita expenditure of GEL 123.5
[2,A]. The mean urban per capita expenditure in 2006 is GEL 127.3 [1,B],
Table 3.1: Mean and Median Per Capita Consumption Expenditure, Growth, and the Gini
Coefficient
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
158
Chapter 3: How to Interpret ADePT Results
which fell by 1.2 percent [1,C]. The mean rural per capita expenditure, in
contrast, increased by 1.0 percent to GEL 124.8 in 2006 [2,B]. Georgia’s
overall per capita consumption expenditure in 2003 is GEL 126.1 [3,A],
which fell by 0.1 percent to GEL 126.0 in 2006 [3,B].
Columns D, E, and F report the median per capita expenditures for
2003 and 2006 and their growth rates. The percentage changes in medians
or median growths are much larger than the mean per capita expenditure
growth. The rural median growth is 3.7 percent [2,F], whereas the urban
median “growth” is –6.8 percent [1,F]. The overall change in median is
–1.4 percent [3,F].
Columns G, H, and I use the Gini coefficient to capture inequality in
the distribution. The rural Gini coefficient has marginally fallen from 35.3
[2,G] to 35.1 [2,H], while the urban Gini coefficient over these three years
increased from 33.5 in 2003 [1,G] to 35.6 in 2006 [1,H]. The overall Gini
coefficient changed by 0.9 from 34.4 [3,G] to 35.4 [3,H]. (Gini coefficient
is reported on a scale from 0 to 100 in this chapter, rather than from 0 to 1.)
Note that the mean and the median, two different measures of standard of
living, are differently sensitive to the distribution of per capita consumption
expenditure. Mean is more sensitive to extreme values, whereas median is
more robust to extreme values. For example, if the only change in the dis-
tribution of per capita expenditure is at the highest quintile or the lowest
quintile, the change would be reflected by the mean, but the median would
not change. In contrast, in certain situations, when changes occur in the
middle of the distribution, mean per capita expenditures may remain unal-
tered, but the median may reflect the change.
It is important to analyze and understand the growth in both these
measures of central tendency. However, changes in different measures of
central tendency do not provide enough information about the change
in the overall distribution. They do not tell us how the spread or inequal-
ity within the distribution changes over time, which can be captured by
an inequality measure. In the above exercise, rural mean and median per
capita expenditure increased, but rural inequality marginally fell. On the
contrary, the urban inequality has increased over these three years from
33.5 in 2003 [1,G] to 35.6 in 2006 [1,H], while the mean and median
have fallen.
159
A Unified Approach to Measuring Poverty and Inequality
Overall Poverty
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
160
Chapter 3: How to Interpret ADePT Results
GEL 45.2 per month, Georgia’s headcount ratio increased from 10.2 percent
in 2003 [6,A] to 10.7 percent in 2006 [6,B]. The rural headcount ratio in this
case increased from 11.4 percent [5,A] to 12.1 percent [5,B]. This change
implies that the proportion of extreme poor (per capita expenditure below
GEL 45.2) in the rural area increased, but the proportion of nonextreme
poor (per capita expenditure between GEL 45.2 and GEL 75.4) decreased.
Columns D, E, and F analyze the poverty gap measure in 2003 and 2006.
The poverty gap measure lies between a minimum of 0 and a maximum of
100, where the minimum is when no one in a region is poor and the maxi-
mum is when everyone has zero consumption expenditure and the poverty
line is positive. When the poverty line is GEL 75.4, the urban area’s poverty
gap measure is 8.6 in 2003 [1,D], which increases by 0.7 to 9.3 in 2006 [1,E].
Likewise, the rural area’s poverty gap measure increases by 0.2 from 10.7 in
2003 [2,D] to 10.9 in 2006 [2,E]. The total increase in poverty gap measure
is 0.4 from 9.7 [3,D] to 10.1 [3,E]. When the poverty line is GEL 45.2, the
overall poverty gap measure increases by 0.2 from 3.0 in 2003 [6,D] to 3.2
in 2006 [6,E].
Columns G, H, and I analyze the squared gap measure. The squared gap
measure also lies between a minimum of 0 and a maximum of 100, where
the minimum is when no one in a region is poor and the maximum is when
everyone has zero consumption expenditure and the poverty line is positive.
This measure is sensitive to inequality across the poor. Column I shows
that the rural area’s squared gap measure when the poverty line is GEL 75.4
increased by 0.3 from 5.2 in 2003 [2,G] to 5.5 in 2006 [2,H]. For the rural
area it increased by 0.1 point from 3.9 [1,G] to 4.0 [1,H]. A similar pattern
of changes is visible for the lower poverty line.
Consider the situation when the poverty line is GEL 75.4. From column C,
one can see that the headcount ratio increased in the urban area by 2.7 per-
centage points and it decreased in the rural area by 0.5 percentage point. In
other words, the rural area performed better than the urban area in reducing
the proportion of poor people.
However, when we look at the poverty gap numbers, we see a different
scenario. It turns out, in fact, from column F that the poverty gaps for both
regions have registered increases, with the urban area registering a larger
increase (0.7 point increase in the urban area compared with 0.2 point
161
A Unified Approach to Measuring Poverty and Inequality
increase in the rural area). Thus, although the number of poor in the rural
area decreased, the same is not true when deprivation is measured in terms
of the average relative shortfall. Column F still reflects that the increase in
the rural poverty gap is lower than that of its urban counterpart. But col-
umn I shows that the increase in the squared gap measure is larger in the
rural area (0.3) than in the urban area (0.2), which implies that inequality
among the rural poor has been sufficiently high that despite a fall in the
headcount ratio, the increase in the squared gap measure is larger than that
in the urban area.
The change in the rural area’s headcount ratio is quite different when
the poverty line is GEL 45.2 per month. The increase in rural poverty is
much higher than the increase in urban poverty by all three measures. In
fact, the squared gap measure slightly decreases for the urban area. We con-
clude from this result that the situation for the rural area’s extreme poor has
actually worsened in 2006 compared with 2003.
Table 3.3 analyzes the distribution of population and poor people across
rural and urban areas. Table rows denote three geographic regions: urban,
rural, and all of Georgia (rows 3 and 6). The variable is per capita consump-
tion expenditure in l per month. There are two poverty lines: GEL 75.4 per
month and GEL 45.2 per month.
Columns A, B, and C analyze the headcount ratio, that is, the popula-
tion percentage that is poor. Columns A and B report the headcount ratio
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
162
Chapter 3: How to Interpret ADePT Results
for the years 2003 and 2006, respectively, while column C reports the differ-
ence across these two years. Columns D, E, and F report the distribution of
poor people across rural and urban areas, with the number in the cell being
the proportion of poor people located in that region. Another way of seeing
this is as the region’s percentage contribution to poverty, or the headcount
ratio times the share of the region’s overall population divided by the overall
headcount ratio. Columns G, H, and I provide the population distribution
across rural and urban areas, or the percentage of the overall population
residing in that region.
The headcount ratio for the urban area’s population in 2003 is 28.1
percent [1,A]. In other words, 28.1 percent of the urban area popula-
tion is poor. The headcount ratio increased for urban Georgia in 2006 to
30.8 percent [1,B].
Of all poor people in Georgia in 2003, 45.6 percent [1,D] reside in
urban areas. The share of all poor people living in urban areas increases to
48.6 percent in 2006 [1,E]. This represents an increase of 3.0 percentage
points [1,F]. The shares of rural and urban area population do not change
much over the course of the three years. But when the poverty line is GEL
75.4 per month, the share of poor in urban areas increases in 2006 because
of the increase in headcount ratio.
This exercise has a very useful policy implication because the headcount
ratio does not provide any information about where most poor people live.
A region may have a lower headcount ratio, but if that region is highly
populated, then the number of poor may be high. Thus, policies should focus
on regions with high headcount ratios as well as regions with larger shares
of poor.
Table 3.4 analyzes the composition of poverty figures reported in table 3.2.
Table rows denote three geographic regions: urban, rural, and all of Georgia
(rows 3 and 6). The variable is per capita consumption expenditure in lari
per month. There are two poverty lines: GEL 75.4 Lari per month and GEL
45.2 Lari per month.
163
A Unified Approach to Measuring Poverty and Inequality
Headcount ratio Income gap Poverty gap GE(2) among Squared gap
(%) ratio measure the poor measure
Region A B C D E
Poverty line = GEL 75.4
2003
1 Urban 28.1 30.5 8.6 4.6 3.9
2 Rural 31.6 33.7 10.7 5.9 5.2
3 Total 29.9 32.3 9.7 5.3 4.6
2006
4 Urban 30.8 30.1 9.3 4.1 4.0
5 Rural 31.1 34.9 10.9 6.4 5.5
6 Total 31.0 32.6 10.1 5.3 4.8
Poverty line = GEL 45.2
2003
7 Urban 8.9 26.8 2.4 4.0 1.0
8 Rural 11.4 31.8 3.6 5.3 1.7
9 Total 10.2 29.7 3.0 4.7 1.4
2006
10 Urban 9.3 25.7 2.4 3.3 1.0
11 Rural 12.1 32.7 4.0 5.7 1.9
12 Total 10.7 29.7 3.2 4.7 1.4
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
The squared gap (PSG) can be decomposed into three factors: headcount
ratio (PH), income gap ratio (PIG), and generalized entropy measure (GE)
for α = 2 among the poor, such that PSG = PH [P2IG + 2(1 − PIG)2 IGE (x; 2)].
These measures make possible a richer set of information for policy
analysis. An improvement in the poverty gap measure may result from a
reduction in the number of poor or a reduction in the average normalized
gap among the poor. Similarly, an improvement in the squared coefficient of
variation may result from a decrease in the number of poor, a decrease in the
164
Chapter 3: How to Interpret ADePT Results
Besides analyzing poverty, one must understand the situation of the rela-
tively poor population compared to the rest of the population. Table 3.5
reports five quantile per capita expenditures (PCEs) and certain quantile
ratios of per capita consumption expenditure for Georgia and its rural
and urban areas. It compares two different periods: 2003 and 2006. Table
rows denote three geographic regions: urban, rural, and all of Georgia
165
A Unified Approach to Measuring Poverty and Inequality
Table 3.5: Quantile PCEs and Quantile Ratios of Per Capita Consumption Expenditure
Percentile
Quantile ratio
10th 20th 50th (median, 80th 90th
(GEL) (GEL) GEL) (GEL) (GEL) 90-10 80-20 90-50 50-10
Region A B C D E F G H I
2003
1 Urban 47.4 64.1 108.4 182.1 229.6 79.3 64.8 52.8 56.3
2 Rural 42.2 58.8 101.5 173.1 230.0 81.6 66.0 55.9 58.4
3 Total 44.8 61.4 104.7 177.0 229.8 80.5 65.3 54.4 57.3
2006
4 Urban 46.7 61.2 101.1 174.0 231.3 79.8 64.8 56.3 53.8
5 Rural 41.0 58.5 105.3 175.9 229.1 82.1 66.8 54.0 61.1
6 Total 43.8 59.8 103.3 175.0 230.5 81.0 65.8 55.2 57.6
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: PCE = per capita expenditure.
166
Chapter 3: How to Interpret ADePT Results
the 90th percentile per capita expenditure and the quantile PCE at the
10th percentile is larger in the rural area (81.6 [2,F]) than in the urban area
(79.3 [1,F]) in 2003. The 90/10 measure increases for Georgia and both its
urban and rural areas in 2006 [4,F] and [5,A].
Table 3.6 reports two lower partial means, two upper partial means, and two
partial mean ratios, based on the partial means between two periods: 2003
and 2006. Table rows denote three geographic regions: urban, rural, and all
of Georgia (rows 3 and 6). Per capita consumption expenditure is measured
in lari per month.
Columns A and B report two lower partial means (LPM), columns C and
D report two upper partial means (UPM), and columns E and F report partial
mean ratios. The first partial mean ratio, for example, reports the 90/10
partial mean ratio, computed as (90th percentile UPM – 10th percentile
167
A Unified Approach to Measuring Poverty and Inequality
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
LPM) / 90th percentile UPM). The larger the 90/10 ratio, the larger is the
gap between these two partial means.
A lower partial mean is the average per capita expenditure of all people
below a specific percentile cutoff. An upper partial mean is the mean per
capita expenditure above a specific percentile. A partial mean ratio captures
inequality between a lower partial mean and an upper partial mean.
It is evident from the table that the average per capita expenditure of the
urban Georgian population’s poorest 20 percent is only GEL 45.2 in 2003
[1,B], whereas the average income of the population’s richest 20 percent is
GEL 261.8 [1,D]. The corresponding 80/20 partial mean ratio is 82.7 [1,F],
which means that the gap between the two partial means is 82.7 percent
of the 80th upper partial mean. Stated another way, the mean per capita
expenditure of the population’s richest 20 percent is 100 / (100 – 82.7) =
5.8 times larger than the mean per capita expenditure of the population’s
poorest 20 percent. Likewise, in rural areas, the mean per capita expendi-
ture of the population’s richest 20 percent (GEL 259.1 [2,D]) is 6.5 times
larger than the mean per capita expenditure of the population’s poorest
20 percent (GEL 39.9 [2,B]) in 2003. The corresponding 80/20 partial mean
ratio is 84.6 [2,F].
168
Chapter 3: How to Interpret ADePT Results
Table 3.7 analyzes the population distribution in Georgia and its rural and
urban areas across five quintiles of per capita consumption expenditure.
It compares two time periods: 2003 and 2006. Table rows denote three
geographic regions: urban, rural, and all of Georgia (row 1). Per capita
consumption expenditure is measured in lari per month. Each of the five
columns denotes a quintile. Column A denotes the lowest, or first, quintile,
column B denotes the second quintile, and so forth.
All cells in row 1 have a value of 20, obtained by dividing Georgia’s
entire population into five equal groups in terms of per capita expenditure.
Each group contains 20 percent of the population. The fifth quintile con-
tains the richest 20 percent of the population, the fourth quintile consists
of the second-richest 20 percent of the population, and so on, and the first
quintile consists of the poorest 20 percent of the population.
Quintile
First Second Third Fourth Fifth
Region A B C D E
1 Total 20.0 20.0 20.0 20.0 20.0
2003
2 Urban 18.1 19.6 20.4 20.8 21.1
3 Rural 21.8 20.4 19.6 19.2 19.0
2006
4 Urban 19.0 21.6 20.6 19.2 19.7
5 Rural 21.0 18.5 19.4 20.8 20.3
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
169
A Unified Approach to Measuring Poverty and Inequality
Rows 2 and 3 report the population distribution in urban and rural areas
for 2003 using the national quintiles. Consider the value 18.1 [2,A] in the
urban row. This value implies that 18.1 percent of the total urban popula-
tion falls in the first quintile. The next cell is 19.6 [2,B], meaning that
19.6 percent of the total urban population falls in the second quintile. Similarly,
21.1 percent [2,E] of the total urban population falls in the fifth quintile.
The picture is slightly different for the rural area, where 19.0 percent
[3,E] of the total rural population falls in the fifth quintile and 21.8 per-
cent [3,A] falls in the lowest quintile. In 2006, the urban population share
in the first two quintiles increased to 19.0 percent [4,A] and 21.6 percent
[4,B], respectively, but the rural population share in the same two quintiles
decreased to 21.0 percent [5,A] and 18.5 percent [5,B], respectively. In
contrast, the rural population share in the two highest quintiles increased,
[3,D] and [3,E] compared with [5,D] and [5,E], but the urban population
share in the two highest quintiles decreased, [2,D] and [2,E] compared with
[4,D] and [4,E].
170
Chapter 3: How to Interpret ADePT Results
Table 3.8 results from calculating the mean and median per capita consump-
tion expenditure, and the Gini coefficient, for Georgia’s subnational regions.
Columns A and B report the mean per capita consumption expenditure for
years 2003 and 2006, respectively. Column C reports the percentage change
or growth in per capita expenditure over the course of these three years.
The mean per capita expenditure decreases for some regions (such as
Kakheti [1,C], Tbilisi [2,C], and Imereti [9,C]) and increases for others
(such as Shida Kartli [3,C], Kvemo Kartli [4,C], and Samtskhe-Javakheti
[5,C]). Imereti registers the steepest fall (7.0 percent [9,C]) in mean per
capita consumption expenditure, from GEL 150.3 in 2003 [9,A] to GEL
139.9 in 2006 [9,B]. In contrast, Kvemo Kartli reflects the highest increase
in mean per capita expenditure, 16.1 percent [4,C]. It increased from GEL
93.5 in 2003 [4,A] to GEL 108.5 in 2006 [4,B].
Columns D, E, and F report median per capita expenditures and their
growth. Although the change in overall median is −1.4 percent [11,F] (much
larger than the change in overall mean), changes in subnational regions are
mixed. For Kvemo Kartli, the growths of mean and median are almost the
same [4,C] and [4,F]. For Samtskhe-Javakheti, the growth in mean [5,C] is
three times larger than the growth of median [5,F]. In contrast, the growth
Table 3.8: Mean and Median Per Capita Income, Growth, and the Gini Coefficient across
Subnational Regions
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
171
A Unified Approach to Measuring Poverty and Inequality
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: n.a. = not applicable.
172
Chapter 3: How to Interpret ADePT Results
173
A Unified Approach to Measuring Poverty and Inequality
Table 3.10 analyzes Georgia’s poverty gap measure across subnational regions.
The poverty line is GEL 75.4 per month. Table rows list subnational regions.
Columns A, B, and C analyze poverty gap measures for 2003, 2006, and the
changes over time. Columns D, E, and F report the percentage contribution
of the subnational regions to the overall poverty gap measure. Columns G,
H, and I depict the population distribution of the subnational regions, or the
percentage of the overall population that resides in each region.
The overall poverty gap measure increases from 9.7 in 2003 [11,A] to
10.1 in 2006 [11,B], reflecting a 0.4 point increase [11,C]. For Kakheti,
the poverty gap measure in 2003 is 13.4 [1,A]. The poverty gap measure
for the same population subgroup in 2006 is 17.8 [1,B]. Thus, the poverty
gap measure increased by 4.4 points [1,C] over three years. The poverty gap
measure in Kvemo Kartli decreased by 3.5 points, from 15.4 in 2003 [4,A] to
11.9 in 2006 [4,B]. The poverty gap measure also fell between 2003 and 2006
in other regions, such as Samegrelo [8,C] and Mtskheta-Mtianeti [10,C].
Kakheti’s contribution to the overall poverty gap measure is 13.4 percent
[1,D]. Its contribution increased to 16.3 percent in 2006 [1,E], an increase
of 2.9 percentage points [1,F].
Contribution to Distribution of
Poverty gap measure overall poverty (%) population (%)
2003 2006 Change 2003 2006 Change 2003 2006 Change
Region A B C D E F G H I
Poverty line = GEL 75.4
1 Kakheti 13.4 17.8 4.4 13.4 16.3 2.9 9.7 9.3 −0.4
2 Tbilisi 5.5 7.3 1.8 14.0 18.2 4.2 24.6 25.2 0.6
3 Shida Kartli 11.7 10.9 −0.8 8.5 7.8 −0.7 7.0 7.2 0.2
4 Kvemo Kartli 15.4 11.9 −3.5 18.1 12.8 −5.3 11.3 10.8 −0.5
5 Samtskhe-Javakheti 10.0 6.6 −3.4 4.7 3.2 −1.6 4.6 4.8 0.2
6 Ajara 12.8 14.6 1.8 11.5 13.7 2.2 8.7 9.5 0.8
7 Guria 8.3 10.6 2.3 2.8 3.3 0.5 3.2 3.1 −0.1
8 Samegrelo 11.0 8.8 −2.2 12.0 8.2 −3.8 10.5 9.5 −1.1
9 Imereti 6.1 7.5 1.4 11.1 13.4 2.4 17.5 18.0 0.5
10 Mtskheta-Mtianeti 13.1 11.7 −1.4 3.9 3.1 −0.8 2.9 2.7 −0.2
11 Total 9.7 10.1 0.4 100.0 100.0 n.a. 100.0 100.0 n.a.
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: n.a. = not applicable.
174
Chapter 3: How to Interpret ADePT Results
Now compare Guria and Imereti. Clearly, Guria’s poverty gap measure
(8.3 [7,A]) is larger than Imereti’s poverty gap measure (6.1 [9,A]) in 2003.
But Guria’s contribution is only 2.8 percent [7,D], whereas Imereti’s contri-
bution is 11.1 percent [9,D]. The contribution of subnational regions to the
overall poverty gap and the share of poor in each region are quite different.
The share of poor in each of Kakheti and Ajara is almost identical in 2006
(9.3 percent for Kakheti [1,H], compared with 9.5 percent in Ajara [6,H]),
but their contributions to the total poverty gap measure are quite different
(16.3 percent in Kakheti [1,E], compared with 13.7 percent in Ajara [6,E]).
Thus, the average normalized shortfall of per capita expenditure from the
poverty line is much higher in Kakheti, and that is not captured by the
headcount ratio analysis.
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: n.a. = not applicable.
175
A Unified Approach to Measuring Poverty and Inequality
176
Chapter 3: How to Interpret ADePT Results
Table 3.12: Quantile PCE and Quantile Ratio of Per Capita Consumption Expenditure, 2003
Quantile PCE
50th
Quantile ratio
10th 20th percentile 80th 90th
percentile percentile (median, percentile percentile 90-10 80-20 90-50 50-10
(GEL) (GEL) GEL) (GEL) (GEL) (%) (%) (%) (%)
Region A B C D E F G H I
1 Kakheti 37.8 52.6 92.7 150.4 191.1 80.2 65.0 51.5 59.2
2 Tbilisi 56.0 74.3 122.2 202.8 252.9 77.9 63.3 51.7 54.2
3 Shida Kartli 38.6 55.9 98.7 169.8 228.4 83.1 67.1 56.8 60.9
4 Kvemo Kartli 34.3 48.3 81.0 126.5 165.1 79.2 61.8 51.0 57.7
5 Samtskhe-Javakheti 43.0 61.2 98.8 160.5 190.2 77.4 61.9 48.0 56.5
6 Ajara 37.8 53.1 91.6 146.5 203.3 81.4 63.7 54.9 58.7
7 Guria 47.7 64.0 113.9 189.1 241.9 80.3 66.1 52.9 58.1
8 Samegrelo 41.2 56.2 97.0 160.7 208.5 80.2 65.0 53.5 57.5
9 Imereti 54.0 74.1 128.6 211.6 267.0 79.8 65.0 51.8 58.0
10 Mtskheta-Mtianeti 33.9 52.5 103.7 162.0 200.1 83.1 67.6 48.2 67.3
11 Total 44.8 61.4 104.7 177.0 229.8 80.5 65.3 54.4 57.3
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: PCE = per capita expenditure.
B denotes the quantile PCE at the 20th percentile, column C denotes the
median, column D denotes the quantile PCE at the 80th percentile, and
column E denotes the quantile PCE at the 90th percentile.
Columns F through I report the quantile ratios based on the quantiles
reported in the first five columns. Column G, for example, reports the 80/20
ratio, computed as (quantile PCE at the 80th percentile – quantile PCE at
the 20th percentile) / quantile PCE at the 80th percentile. The larger the
80/20 ratio, the larger is the gap between these two percentiles.
In 2003, the quantile PCE at the 10th percentile of Kakheti is 37.8 [1,A],
which implies that 10 percent of the population in Kakheti lives with per
capita consumption expenditure less than GEL 37.8. Similarly, 20 percent
of Kakheti’s population lives with per capita consumption expenditure less
than GEL 52.6 [1,B]. In contrast, 10 percent of people in Kakheti live with
per capita expenditure more than GEL 191.1 [1,E], the quantile PCE at the
90th percentile. The corresponding 90/10 measure using these two quantile
PCEs is 80.2 [1,F], meaning that the gap between the two quantile PCEs is
80.2 percent of the quantile PCE at the 90th percentile. Described another
way, the quantile PCE at the 90th percentile is 100 / (100 – 80.2) = 5.1
times larger than the quantile PCE at the 10th percentile.
Likewise, the quantile PCE at the 80th percentile of Kakheti is
GEL 150.4 [1,D], nearly three times larger than the quantile PCE at the
177
A Unified Approach to Measuring Poverty and Inequality
20th percentile per capita expenditure [1,B]. It is evident that Shida Kartli
has a lower quantile PCE at the 10th percentile than Samegrelo but a larger
quantile PCE at the 90th percentile. As a result, the 90/10 quantile ratio of
Shida Kartli [3,F] is higher than the 90/10 quantile ratio of Samegrelo [8,F].
Table 3.13 reports two lower partial means, two upper partial means, and
two partial mean ratios for Georgia’s subnational regions in 2003. Columns
A and B report the two lower partial means, columns C and D report the two
upper partial means, and columns E and F report the partial mean ratios.
The first of the partial mean ratios, for example, reports the 90/10 partial
Table 3.13: Partial Means and Partial Mean Ratios for Subnational Regions,
2003
Lower partial mean Upper partial mean Partial mean ratio (%)
p10 p20 p20 p10 90-10 80-20
Region A B C D E F
1 Kakheti 25.6 35.9 222.3 276.0 90.7 83.8
2 Tbilisi 44.1 54.9 286.7 348.8 87.3 80.9
3 Shida Kartli 26.2 37.1 263.3 331.0 92.1 85.9
4 Kvemo Kartli 23.9 32.3 186.7 230.9 89.6 82.7
5 Samtskhe-Javakheti 30.5 41.5 234.4 294.8 89.6 82.3
6 Ajara 26.2 36.5 222.4 273.4 90.4 83.6
7 Guria 35.9 45.8 275.3 337.2 89.4 83.4
8 Samegrelo 30.8 40.1 241.1 302.9 89.8 83.4
9 Imereti 39.8 52.4 299.1 362.1 89.0 82.5
10 Mtskheta-Mtianeti 25.0 34.7 222.7 265.7 90.6 84.4
11 Total 31.5 42.3 260.5 320.4 90.2 83.8
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
178
Chapter 3: How to Interpret ADePT Results
A larger inequality in terms of the quantile ratio does not necessarily trans-
late into higher inequality in terms of the partial mean ratio. In table 3.12,
we found that the 80/20 quantile ratio for Imereti (65.0) was larger than
that of Ajara (63.7), but in table 3.13 Ajara’s 80/20 partial mean ratio (83.6
[3,F]) is slightly larger than Imereti’s (82.5 [9,F]).
179
A Unified Approach to Measuring Poverty and Inequality
Quintile
First Second Third Fourth Fifth
Region A B C D E
1 Total 20.0 20.0 20.0 20.0 20.0
2 Kakheti 27.6 20.9 20.8 18.3 12.5
3 Tbilisi 12.4 17.9 19.9 22.5 27.2
4 Shida Kartli 23.0 21.7 17.0 19.7 18.5
5 Kvemo Kartli 30.0 27.5 21.2 13.5 7.9
6 Samtskhe-Javakheti 20.1 24.0 21.6 20.2 14.1
7 Ajara 25.9 22.9 21.7 15.6 13.8
8 Guria 17.4 17.7 20.7 20.4 23.8
9 Samegrelo 23.5 19.8 20.6 21.3 14.7
10 Imereti 13.3 15.3 18.2 22.9 30.3
11 Mtskheta-Mtianeti 25.5 17.5 20.2 21.2 15.6
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
180
Chapter 3: How to Interpret ADePT Results
quintiles, respectively, and 30.3 percent [10,E] of its population falls in the
fifth or richest quintile. Kvemo Kartli appears to be the poorest among all
subnational regions because 30.0 percent [5,A] of its population falls in the
poorest quintile and only 7.9 percent [5,E] of its population falls in the rich-
est quintile.
Table 3.15 decomposes poverty to explore the factors that caused a change
in headcount ratio. Table rows are divided into two categories. Rows 1
through 4 report the change in the overall poverty and three factors affect-
ing this change: total intrasectoral effect, population-shift effect, and inter-
action effect. Rows 5 through 14 report the intrasectoral effects for various
regions in Georgia.1 Column A reports the absolute change in headcount
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
181
A Unified Approach to Measuring Poverty and Inequality
poverty and the size of the factors contributing to this change. Column B
shows how these factors change the headcount ratio.
The change in overall headcount ratio between 2003 and 2006 is 1.04
[1,A]. This overall change of 1.04 percentage points is divided into three
different effects. The first is the total intrasectoral effect, 1.09 [2,A]. The
total population-shift effect is negative and amounts to –0.18 [3,A]. The
interaction between the intrasectoral factor and the population shift fac-
tor is 0.13 [4,A]. If we sum these three effects, we get the overall absolute
change in headcount ratio poverty: (1.09 – 0.18 + 0.13) = 1.04 [1,A].
The next column reports the proportion these effects have relative to the
overall change. The proportion of the total intrasectoral effect on the over-
all change in poverty is 104.98 percent [2,B]. This number is calculated by
dividing the total intrasectoral effect by the change in poverty: (100 × 1.09)
/ 1.04 = 104.98. The corresponding entries for the population-shift effect
and the interaction effect are calculated by the same method. For example,
(100 × –0.18) / 1.04 = –17.38 and (100 × 0.13) / 1.04 = 12.40 [4,B].
The next set of results decomposes the total intrasectoral effect across
Georgia’s regions. Column A reports the size of the intrasectoral effect, and
column B reports the intrasectoral effect as a proportion of the total change
in the overall headcount ratio. For example, the intrasectoral effect for
Kakheti is 0.70 [5,A], and its proportion of the overall poverty change is
67.93 percent [5,B], calculated as (100 × 0.70) / 1.04 = 67.93.
The intrasectoral effect of Kakheti is calculated as the change in
headcount ratio between 2003 and 2006, which is 7.3 percentage points
(reported in column C of table 3.9), multiplied by its population share in
2003 (reported in column G of table 3.9). The intrasectoral effects are nega-
tive for regions such as Shida Kartli, Kvemo Kartli, Samtskhe-Javakheti, and
Samegrelo, because the poverty headcount ratio fell in these regions. For the
rest of the regions, the intrasectoral effects are positive. The contribution of
this effect is highest for Tbilisi and lowest for Kvemo Kartli.
The total intrasectoral effect is even higher than the total change in the
overall headcount ratio. Thus, if the region-wise population shares are
kept constant, then the change in poverty is 1.09 percentage points [2,A].
However, if we keep the regional headcount ratios constant and consider
182
Chapter 3: How to Interpret ADePT Results
only the changes in regional population shares, then the poverty rate would
have fallen by 0.18 percentage point [3,A]. Thus, the intrasectoral effect
dominates and the overall headcount ratio increases. Finally, the second set
of results gives us an idea about the headcount ratio’s regional contribution
in terms of intrasectoral effect.
In this section, we discuss the results when the population is divided in various
ways: household head’s characteristics, household member’s employment sta-
tus, education level, age group, demographic composition, and landownership.
Table 3.16 reports the mean and median per capita consumption expendi-
ture and their growth over time and inequality across the population using
the Gini coefficient across various household characteristics. Table rows
denote various household characteristics. Columns A and B report the
mean per capita consumption expenditure for 2003 and 2006, respectively.
Column C reports the percentage change in per capita expenditure over
these three years. It is evident from rows 1 and 2 that the mean per capita
expenditure goes up by 1.1 percent [2,C] for female household heads but
decreases by 0.5 percent [1,C] for male household heads.
Columns D, E, and F report the median per capita expenditures for 2003
and 2006 and the growth rates between these years. Although the overall
change in median is –1.4 percent [20,F] (much larger than the change in
overall mean of –0.1 percent [20,C]), the changes in the groups with vari-
ous household characteristics are mixed. For female household heads, the
median increases by 1.5 percent [2,F], but it falls by 2.2 percent [1,F] for
male household heads. We find a mixed picture for the other household
characteristics.
Columns G, H, and I report inequality by household head’s characteris-
tics using the Gini coefficient, which lies between 0 and 100. Although the
overall Gini coefficient increases by 0.9 [20,I] in 2006, changes for different
household characteristics vary over a broad range.
183
A Unified Approach to Measuring Poverty and Inequality
Table 3.16: Mean and Median Per Capita Consumption Expenditure, Growth, and Gini
Coefficient, by Household Characteristics
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
184
Chapter 3: How to Interpret ADePT Results
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: n.a. = not applicable.
185
A Unified Approach to Measuring Poverty and Inequality
186
Chapter 3: How to Interpret ADePT Results
Quintile
First Second Third Fourth Fifth
Characteristic of household head A B C D E
1 Total 20.0 20.0 20.0 20.0 20.0
Gender
2 Male 18.6 20.2 20.1 20.7 20.3
3 Female 23.8 19.3 19.7 18.0 19.1
Age (years)
4 15–19 0.0 27.1 51.1 17.1 4.8
5 20–24 10.5 19.9 12.1 19.1 38.4
6 25–29 23.4 15.5 13.4 26.1 21.6
7 30–34 16.5 21.5 16.8 23.4 21.8
8 35–39 21.6 19.9 20.0 17.7 20.8
9 40–44 19.5 17.5 21.1 20.9 21.0
10 45–49 21.7 19.9 19.2 19.2 19.9
11 50–54 21.4 20.5 19.9 20.0 18.2
12 55–59 16.2 18.6 18.2 24.1 22.9
13 60–64 15.5 19.5 22.4 20.9 21.8
14 65+ 21.5 21.1 20.3 18.6 18.5
Education
15 Elementary or less 32.5 23.8 16.5 15.9 11.3
16 Incomplete secondary 25.5 23.7 20.1 16.9 13.9
17 Secondary 22.3 22.4 20.3 18.5 16.5
18 Vocational-technical 20.1 18.9 20.1 21.0 19.9
19 Special secondary 17.4 18.1 19.9 22.7 21.8
20 Higher education 10.7 14.4 20.6 23.6 30.7
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
187
A Unified Approach to Measuring Poverty and Inequality
188
Chapter 3: How to Interpret ADePT Results
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: n.a. = not applicable.
189
A Unified Approach to Measuring Poverty and Inequality
[13,A] is larger than the poverty headcount ratio in 2003 among the sub-
group employed in the agricultural sector [1,A]. However, if we consider
the share of all poor people who are found in these two subgroups in 2003,
this number is nearly twice as large in the agricultural sector as that among
the unemployed group. This is because of the different population shares of
the two subgroups as given in the final columns. The population share in the
agriculture subgroup in 2003 is 23.6 percent [1,G], while the share in the
unemployed subgroup is only 7.2 percent [13,G].
In row 1, the agricultural subgroup’s poverty headcount ratio falls 1.3
percentage points [1,C], while the share of poor in this subgroup falls by
3.0 percentage points [1,F]. For the other services subgroup, the headcount
ratio increased 7.2 percentage points [5,C] between 2003 and 2006, while
the share of poor in this subgroup increased by only 0.2 percentage point,
from 0.7 percent [5,D] to 0.9 percent [5,E].
One might wonder why these two ways of evaluating changes are so
different. Look at columns G and H. Notice that the population share
employed in the agricultural sector is more than 20 percent of the total
population in both 2003 [1,G] and 2006 [1,H]. In comparison, the popu-
lation share engaged in other services is only 1.0 percent in 2003 [5,G]
and 2006 [5,H]. Consequently, a change of smaller magnitude in the
headcount ratio in the agricultural sector has a larger impact on its share
of the poor and vice versa.
Table 3.20 analyzes poverty by education levels. The poverty line is set at
GEL 75.4 per month. Columns A, B, and C analyze poverty headcount
ratios for 2003, 2006, and the difference over time. Columns D, E, and F
outline the distribution of poor people across the subgroups, with the num-
ber in the cell being the proportion of all poor people in the country located
in that subgroup. This is the subgroup’s percentage contribution to overall
poverty, or the headcount ratio times the population share in that group.
Columns G, H, and I depict subgroup population distribution, or the popula-
tion percentage in that subgroup. Row 7 shows that the overall headcount
190
Chapter 3: How to Interpret ADePT Results
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: n.a. = not applicable.
ratio increases from 29.9 percent in 2003 [7,A] to 31.0 percent in 2006 [7,B],
reflecting a 1.0 percentage point (rounded) increase in the headcount ratio.
We find that 40.4 percent [1,A] of the population who have elementary-
level education or less are poor. In other words, the headcount ratio for this
population subgroup is 40.4 percent. The headcount ratio for the same popu-
lation subgroup fell to 35.9 percent in 2006 [1,B]. Thus, the headcount ratio
fell by 4.6 percentage points [1,C] between these three years. At the other
extreme, the headcount ratio for the subgroup with higher education increased
by 3.4 percentage points, from 17.6 percent [6,A] to 20.9 percent [6,B].
Of all people who are poor in Georgia in 2003, 6.5 percent [1,D]
have elementary education or less. The share of all poor with elementary
education or less decreased to 5.7 percent in 2006 [1,E], a decrease of 0.7
percentage point [1,F].
Clearly, the poverty headcount ratio among the population with incom-
plete secondary education in 2003 [2,A] is larger than the poverty head-
count ratio in 2003 among the higher education subgroup [6,A]. However,
if we consider the share of all poor people who are found in these two
subgroups in 2003, the number is larger for the population with higher edu-
cation because of the two subgroups’ different population shares, as given
in the table’s final columns. The population share with higher education
in 2003 is 24.1 percent [6,G], whereas the population share with incom-
plete secondary education is only 11.5 percent [2,G]. The headcount ratios
increased for the population with incomplete secondary education from
191
A Unified Approach to Measuring Poverty and Inequality
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: n.a. = not applicable.
192
Chapter 3: How to Interpret ADePT Results
from 29.9 percent in 2003 [12,A] to 31.0 percent in 2006 [12,B], reflecting
a 1.0 percentage point (rounded) increase in the headcount ratio.
First, consider the results based on the number of children in households.
We find that 28.8 percent [1,A] of the population with no child in the
household is poor in 2003. In other words, the headcount ratio for this popu-
lation subgroup is 28.8 percent. The headcount ratio for the same population
subgroup decreased to 28.5 percent in 2006 [1,B]. Thus, the headcount ratio
decreased by 0.4 percentage point [1,C] over the course of these three years.
Headcount ratios also decreased for the population with three or more
children in the household by 3.1 percentage points from 43.7 percent
[4,A] in 2003 to 40.6 percent [4,B] in 2006. Similarly, consider the set of
results corresponding to the household size. The headcount ratio among the
population with only one member in the household in 2003 is 25.8 percent
[5,A], which falls by 1.7 percentage points to 24.1 percent in 2006 [5,B]. At
the other extreme, the headcount ratio among the people living in house-
holds with seven or more members increased by 4.0 percentage points from
39.3 percent [11,A] to 43.2 percent [11,B].
The next cell in row 1 is 69.6 [1,D], meaning that of all people who are
poor in Georgia in 2003, 69.6 percent of the population live in households
with no child. In the next column, we find that the share of poor with no
child decreased to 66.1 percent in 2006 [1,E], a decrease of 3.5 percentage
points [1,F].
Compare those results with the subgroup having three or more children.
It is evident that the headcount ratio among the subgroup with no child in
both years (28.8 percent in 2003 [1,A] and 28.5 percent in 2006 [1,B]) is
lower than the headcount ratio for the subgroup with three or more children
(43.7 percent in 2003 [4,A] and 40.6 percent for 2006 [4,B]). Note that the
share of the former subgroup to total poverty is 69.6 percent in 2003 [1,D],
which fell by 3.5 percentage points to 66.1 percent in 2006 [1,E]. The share
of the latter to total poverty is 1.5 percent in 2003 [4,D], which fell by
0.2 percentage point to 1.3 percent in 2006 [4,E]. However, in both years,
the share of poor in the former subgroup is more than 40 times higher than
that in the latter subgroup.
Note that the poverty rate among the subgroup with three or more children
is higher than the subgroup with no child. However, the population share
193
A Unified Approach to Measuring Poverty and Inequality
in the subgroup with no child is so large (72.2 percent in 2003 [1,G] and
72 percent in 2006 [1,H]), compared to the subgroup with three or more
children (only 1.0 percent in both years [4,G] and [4,H]), that the share of
the subgroup with no child in total poverty is high. The analysis in table
3.21 enables a policy maker to understand the origin of poverty at a more
disaggregated level. A policy maker should also focus on households with no
child, even though the headcount ratio is lowest in this subgroup. Similar
intuition should hold for the next set of results where the subgroups are
based on household size.
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: n.a. = not applicable.
194
Chapter 3: How to Interpret ADePT Results
each subgroup. Row 6 shows that the overall headcount ratio increases from
29.9 percent in 2003 [6,A] to 31.0 percent in 2006 [6,B], reflecting a
1.0 percentage point (rounded) increase in the headcount ratio.
We find that 29.4 percent [1,A] of people who belong to households
with no landownership are poor in 2003. In other words, the headcount
ratio for this population subgroup is 29.4 percent. The headcount ratio for
the same population subgroup increases to 32.7 percent in 2006 [1,B]. Thus,
the headcount ratio increased by 3.3 percentage points [1,C] over these
three years. We see that the headcount ratio for the population in house-
holds with landownership of 0.5–1.0 hectare decreased by 0.8 percentage
point, from 25.1 percent [4,A] to 24.3 percent [4,B].
Of all poor people in Georgia in 2003, 39 percent [1,D] lived in house-
holds with no landownership. The share of poor with no landownership
increased to 46.4 percent in 2006 [1,E]. The headcount ratio among the
subgroup with landownership of 0.5–1.0 hectare (25.1 percent in 2003
[4,A] and 24.3 percent in 2006 [4,B]) is lower than the headcount for the
subgroup with a landownership of less than 0.2 hectare (39.4 percent in
2003 [2,A] and 36.2 percent for 2006 [2,B]). Note that the share of the
former subgroup to total poverty is 19.5 percent in 2003 [4,D], which fell
by 4.1 percentage points to 15.4 percent in 2006 [4,E]. The share of the
latter to total poverty is 12.7 percent in 2003 [2,D], which fell by only 0.7
percentage point to 11.9 percent in 2006 [2,E]. Note that despite a larger
fall in the poverty rate of 3.1 percentage points [2,C] for the subgroup with
landownership of less than 0.2 hectare, the share of poor in that subgroup
fell by only 0.7 percentage point [2,F]. One might wonder about the reason
behind this phenomenon.
The answer can be found if we look at columns G and H. Notice that the
population share with landownership of less than 0.2 hectare is 9.6 percent
in 2003 [2,G], and it increased by 0.6 percentage point to 10.2 percent in
2006 [2,H]. In contrast, the population share with landownership of 0.5–1.0
hectare fell by 3.6 percentage points, from 23.2 percent [4,G] in 2003 to
19.6 percent [4,H] in 2006. Moreover, the population share in the latter
subgroup is almost twice as high as that in the former subgroup in both
years. Thus, despite a larger fall in headcount ratio for the subgroup with
landownership of less than 0.2 hectare, its share in total number of poor did
not decrease significantly compared to the subgroup with landownership of
0.5–1 hectare.
195
A Unified Approach to Measuring Poverty and Inequality
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: n.a. = not applicable.
196
Chapter 3: How to Interpret ADePT Results
Between 2003 and 2006, the headcount ratios decreased for age group
50–54 years by 1.1 percentage points, from 28.3 percent [10,A] to 27.1 per-
cent [10,B], and for the age group 65+ years by 0.6 percentage point from
29.3 percent [13,A] to 28.8 percent [13,B]. In contrast, headcount ratios
increased for all other groups by 0.4 to 3.8 percentage points. For example,
the headcount ratio for age group 30–34 years increased by 2.4 percentage
points from 30.2 percent [6,A] in 2003 to 32.6 percent [6,B] in 2006.
Of all poor people in Georgia in 2003, 5.9 percent are in the age group
of 0–5 years [1,D]. The share of all poor in age group 0–5 years increased to
6.2 percent in 2006 [1,E], an increase of 0.2 percentage point. Now consider
age groups 6–14 and 65+ years. The headcount ratio among the population
in age group 6–14 years increased by 1.2 percentage points from 33.3 percent
in 2003 [2,A] to 34.5 percent in 2006 [2,B], but the headcount fell by 0.6
percentage point for age group 65+ years [13,C]. However, if we consider
the change in share of all poor people found in these two subgroups in 2003
(column F), this number went up for age group 65+ (0.8 [13,F]) and fell for
age group 6–14 years (–1.8 [2,F]).
One might ask why the share of the poor has fallen in spite of an increase in
headcount ratios. The answer can be found in columns G and H. Note that
the share of people in the age group 6–14 years decreased by 1.7 percent-
age points from 12.9 percent in 2003 [2,G] to 11.3 percent in 2006 [2,H].
In contrast, the population share in age group 65+ years increased by 1.6
percentage points from 15.0 percent in 2003 [13,G] to 16.6 percent in 2006
[13,H]. Thus, despite a decrease in headcount ratio for age group 65+ years,
its share of poor increased. A policy maker, therefore, should notice that a
decrease in headcount among the 65+ years age group did not necessarily
decrease the number of total poor in that age group.
197
A Unified Approach to Measuring Poverty and Inequality
2003
95 33 30 95
90 31 38 90
85 36 31 85
80 31 27 80
75 31 29 75
30 25
Age in years
60 23 21 60
55 25 29 55
50 28 28 50
45 26 28 45
40 31 29 40
35 29 30 35
30 29 30 30
25 31 29 25
20 34 30 20
15 35 34 15
10 32 31 10
5 31 34 5
6 5 4 2 1 0 1 2 4 5 6
Share in total population, %
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
198
Chapter 3: How to Interpret ADePT Results
poor households. The headcount ratio among females in that age group is
32 percent and among males it is 31 percent. The headcount ratio is highest
among male members in the 85–90 years age group: 38 percent of the males
in that age group reside in poor households. The largest headcount ratio
among females is seen in the 80–85 years age group.
Sensitivity Analyses
Table 3.24 provides a tool for checking the sensitivity of the three poverty
measures to consumption expenditure. The table shows the result of increas-
ing everyone’s consumption expenditure by 1.0 percent and compares those
values across two years, 2003 and 2006. There are two poverty lines: GEL
75.4 and GEL 45.2 per month.
The percentage change in poverty caused by a 1.0 percent change in the
mean or average per capita consumption expenditure is referred to as the
elasticity of poverty with respect to per capita consumption. The particular way
Table 3.24: Elasticity of FGT Poverty Indices to Per Capita Consumption Expenditure
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: FGT = Foster-Greer-Thorbecke.
199
A Unified Approach to Measuring Poverty and Inequality
200
Chapter 3: How to Interpret ADePT Results
the squared gap measure did not change much between these two years for
either poverty line.
Note that poverty lines are set normatively, which is difficult to justify
exclusively. A slight change in per capita consumption expenditure may or
may not change the poverty rates by significant margins. If the distribution is
highly polarized or, in other words, if the society has two groups of people—
one group consisting of rich people and the other group consisting of extreme
poor—then a slight change in everyone’s income by the same proportion
may not affect the headcount ratio.
In contrast, if marginal poor are concentrated around the poverty line,
then a slight change in everyone’s income by the same proportion would
have a huge impact on the poverty measures. For example, in the table
the poverty measures are more sensitive to the lower GEL 45.2 per month
poverty line than the higher GEL 75.4 per month poverty line. This is
because the concentration of poor around the lower poverty line is much
larger than that around the higher poverty line. Hence, this type of analysis
may tell us about the impact of any policy on the poverty rate used by the
policy maker.
Table 3.25 presents a tool for checking the sensitivity of the headcount ratio
with respect to the chosen poverty line. This exercise is similar to the exer-
cise for checking the elasticity of poverty measures to per capita consump-
tion expenditure, but it is more rigorous. It is always possible to find a certain
percentage of decrease in the poverty line that matches the increase in the
consumption expenditure for everyone by 1.0 percent. In this exercise,
we check the sensitivity of the poverty measure by changing the poverty
line in more than one direction. Thus, in the table, we ask how the actual
headcount ratio changes as the poverty line changes from its initial value,
whether it is GEL 75.4 per month or GEL 45.2 per month.
Rows denote the change in poverty line, both upward and downward.
Columns report the change in three poverty measures: the headcount ratio,
the poverty gap measure, and the squared gap measure, and their change
from actual. The variable is per capita consumption expenditure, measured
201
A Unified Approach to Measuring Poverty and Inequality
Table 3.25: Sensitivity of Poverty Measures to the Choice of Poverty Line, 2003
Headcount Change from Poverty gap Change from Squared gap Change from
ratio actual (%) measure actual (%) measure actual (%)
A B C D E F
Poverty line = GEL 75.4
1 Actual 29.9 0.0 9.7 0.0 4.6 0.0
2 +5 percent 32.6 9.0 10.7 10.7 5.1 11.4
3 +10 percent 35.3 18.0 11.7 21.7 5.6 23.3
4 +20 percent 40.5 35.2 13.9 44.3 6.8 48.5
5 −5 percent 26.9 −10.0 8.7 −10.2 4.1 −10.8
6 −10 percent 24.2 −19.1 7.7 −19.9 3.6 −21.1
7 −20 percent 19.4 −35.3 6.0 −38.1 2.7 −40.0
Poverty line = GEL 45.2
8 Actual 10.2 0.0 3.0 0.0 1.4 0.0
9 +5 percent 11.4 11.8 3.4 12.2 1.5 12.4
10 +10 percent 12.7 24.1 3.8 25.2 1.7 25.6
11 +20 percent 15.8 54.5 4.7 53.8 2.1 54.6
12 −5 percent 9.2 −9.9 2.7 −11.6 1.2 −11.6
13 −10 percent 8.0 −21.4 2.4 −22.4 1.1 −22.4
14 −20 percent 6.0 −40.9 1.8 −41.6 0.8 −41.4
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
in lari. In this table, we report the results only for 2003, but this analysis can
be conducted for any year.
Column A reports the headcount ratios for different poverty lines, and
column B reports the change in the headcount ratios from the actual pov-
erty line, which can be either GEL 75.4 per month or GEL 45.2 per month.
Rows 2 and 9, corresponding to +5 percent, denote the increase in poverty
line by 5 percent. Thus, when the poverty line is GEL 75.4, then a 5 percent
increase means the poverty line becomes GEL 79.2 and the headcount ratio
increases by 3.7 percentage points from 29.9 percent [1,A] to 32.6 percent
[2,A], or the headcount ratio increases by 9.0 percent [2,B] from its actual
level of 29.9 percent.
Similarly, if the poverty line is decreased by 10 percent (–10 percent)
from GEL 75.4, then the poverty headcount rate falls by 5.7 percentage
points from 29.9 percent [1,A] to 24.2 percent [6,A], or the headcount ratio
decreases by 19.1 percent from the actual level of 29.9 percent. The head-
count ratio is more sensitive to the change in poverty line when the actual
poverty line is GEL 45.2 than when the poverty line is GEL 75.4. In fact,
the poverty gap measure and the squared gap measure are also more sensitive
to change in poverty line when the actual poverty line is GEL 45.2 rather
than GEL 75.4.
202
Chapter 3: How to Interpret ADePT Results
Table 3.26 analyzes the overall poverty for Georgia and decomposes it across
rural and urban areas using three other poverty measures not in the FGT class.
The table reports three different sets of poverty measures: the Watts index,
Sen-Shorrocks-Thon (SST) index, and Clark-Hemming-Ulph-Chakravarty
(CHUC) index (these measures are defined in chapter 2). This is a type of
sensitivity analysis, but of the poverty measurement methodology. There are
two poverty lines: GEL 75.4 per month and GEL 45.2 per month.
Columns A and B report the Watts index for both years. The Watts
index is the mean log deviation relative to the poverty line. It is evident
from row 1 that the urban Watts index increases from 12.0 in 2003 [1,A]
to 12.7 in 2006 [1,B] when the poverty line is GEL 75.4 but falls slightly
between the same years when the poverty line is GEL 45.2 [4,A] and [4,B].
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: CHUC = Clark-Hemming-Ulph-Chakravarty.
203
A Unified Approach to Measuring Poverty and Inequality
Columns D and E report the SST index, which is also based on the
headcount ratio, the income gap ratio, and the Gini coefficient across the
censored distribution of consumption expenditure. The last is obtained
by replacing consumption expenditure of all nonpoor people by the pov-
erty line. We see that when the poverty line is GEL 75.4, the SST index
for the urban region in 2003 is 15.7 [1,D], and it increases by 1.1 to 16.8
in 2006 [1,E]. Likewise, the rural region’s SST index increased by 0.4,
from 19.2 in 2003 [2,D] to 19.6 in 2006 [2,E], for the same poverty line.
The total increase in SST index is 0.7, from 17.5 in 2003 [3,D] to 18.3
in 2006 [3,E].
The final three columns report the CHUC index and its changes across
time. Unlike the SST index, the CHUC index does not reflect an increase
in poverty across all regions. In fact, urban poverty falls marginally between
2003 [1,G] and 2006 [1,H] when the poverty line is GEL 75.4. Furthermore,
when the poverty line is set at GEL 45.2, the CHUC index shows a fall in
Georgia’s overall poverty [6,I].
Table 3.27 reports the Atkinson inequality measures and generalized entropy
measures for 2003, then decomposes them across different regions. This is
a type of sensitivity analysis for inequality measurement methodology. We
report the Gini coefficient only in the last two sections of this chapter.
However, the Gini coefficient may not be subgroup consistent (subgroup
consistency is defined in chapter 2). Rows denote results for urban and rural
population subgroups and for different geographic regions, such as Kakheti,
Tbilisi, and Shida Kartli.
204
Chapter 3: How to Interpret ADePT Results
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
Note: GE = generalized entropy.
Columns A, B, and C report the Atkinson measures for a = 1/2, 0, and –1,
respectively, and columns D, E, and F report the generalized entropy measures
for a = 0, 1, and 2, respectively. (For a theoretical discussion on the Atkinson
inequality measure and generalized entropy measures, please refer to chapter
2.) Intuitively, an Atkinson inequality measure of order a is the gap between
the mean achievement and the generalized mean of achievements of order
a divided by the mean achievement. Generalized mean is sensitive to inequal-
ity across the distribution, where a lower value of a reflects higher sensitivity
to inequality across the distribution. In other words, a lower value of a reflects
higher aversion toward inequality and, thus, it is also known as the inequality
aversion parameter. When everyone has identical achievement, then it does
not matter how sensitive one is toward inequality, so the generalized mean
is equal to the arithmetic mean for all a. For the analysis in table 3.27, the
inequality measures put more emphasis on the lower end of the distribution
and thus assume a < 1. The Atkinson measure lies between 0 and 1. Similarly,
if a household has equal per capita expenditure in a region, then the general-
ized entropy measure is also 0 for all a.
The Atkinson measure for a = 1/2, or A(1/2), for the urban area is
9.1 [1,A]. Intuitively, the number implies that the generalized mean of
205
A Unified Approach to Measuring Poverty and Inequality
order 0.5 for urban Georgia is 9.1 percent lower than Georgia’s mean per
capita expenditure in 2003. The next two cells to the right report A(0) and
A(–1) for urban Georgia, where A(0) = 17.7 [1,B] and A(–1) = 34.3 [1,C].
Therefore, A(0) is 17.7 percent lower than the mean per capita expenditure
and A(–1) is 34.3 percent lower than the mean per capita expenditure.
Columns D, E, and F report three generalized entropy measures for a = 0, 1,
and 2, denoted by GE(0), GE(1), and GE(2), respectively.
Row 2 reports the three Atkinson measures and three generalized
entropy measures for rural Georgia. Each of these six measures shows that
rural Georgia is more unequal than urban Georgia. For example, the A(1/2)
for the rural area is 10.1 [2,A], compared with 9.1 in the urban area [1,A],
and A(0) for the rural area is 19.8 [2,B], compared with 17.7 for the urban
area [1,B]. However, the difference is much larger when the two regions are
compared with respect to A(–1): 38.9 for the rural area [2,C] and 34.3 for
the urban area [1,C].
Next, we consider the results across regions. The level of inequality
of Ajara according to A(1/2) is 9.4 [8,A], which is higher than that of
Samtskhe-Javakheti at 9.0 [7,A]. This means that Ajara has larger income
inequality than Samtskhe-Javakheti. Even according to A(0), A(–1),
GE(0), and GE(–1), Ajara has higher income inequality than Samtskhe-
Javakheti. However, in terms of GE(2), which gives more weight to larger
incomes across the population, Samtskhe-Javakheti [7,F] has higher income
inequality than Ajara [8,F].
206
Chapter 3: How to Interpret ADePT Results
Dominance Analyses
207
A Unified Approach to Measuring Poverty and Inequality
Figure 3.3: Poverty Incidence Curves in Urban Georgia, 2003 and 2006
Urban
1.0
0.8
Cumulative distribution
0.6
0.4
0.2
0
0 160 320 480 640 800
Welfare indicator
2003 2006
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
Note: The red vertical line is the poverty line.
Figure 3.3 graphs the poverty incidence curves for urban Georgia in
2003 and 2006. The vertical axis reports the headcount ratio. The solid
line denotes the poverty incidence curve for 2003, while the dashed line
denotes the poverty incidence curve for 2006. We saw earlier that the urban
headcount ratio is higher in 2006 for both poverty lines: GEL 75.4 and GEL
45.2. What about other poverty lines? Can we say that poverty has unam-
biguously fallen for any poverty line? Figure 3.3 suggests that we may not
be able to. If we set the hypothetical poverty line somewhere between GEL
320 and GEL 480, then the headcount ratio would have been lower in 2006
than that in 2003.
Although such a poverty line seems very high and unlikely to be set at that
value, the main point of the exercise is clear. When two poverty incidence
curves cross, then an unambiguous judgment cannot be made. The crossing
208
Chapter 3: How to Interpret ADePT Results
may take place at a much lower level, as happened in the rural area. We
have already seen that the headcount ratio showed an increase in 2006
when the poverty line is set at GEL 75.4 but showed a decrease when the
poverty line is set at GEL 45.2. Given the infinite number of possible pov-
erty lines, it would be cumbersome to check them all one by one. Instead,
the poverty incidence curve is a convenient way of checking for dominance
(if two poverty incidence curves never cross). If dominance does not hold,
then the graph can tell us which part is responsible for the ambiguity.
Figure 3.4: Poverty Deficit Curves in Urban Georgia, 2003 and 2006
Urban
300
200
Total deficit
100
0
0 160 320 480 640 800
Welfare indicator
2003 2006
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
Note: The red vertical line is the poverty line.
209
A Unified Approach to Measuring Poverty and Inequality
Although such a poverty line seems very high and unlikely to be set at that
value, the main point of the exercise is clear. When two poverty deficit
curves cross, then an unambiguous judgment cannot be made based on the
poverty gap measure. Given the infinite number of possible poverty lines, it
would be cumbersome to check them all one by one. Instead, the poverty
deficit curve is a convenient way of checking for dominance (if two poverty
deficit curves never cross). If dominance does not hold, then the graph can
tell us which part is responsible for the ambiguity.
210
Chapter 3: How to Interpret ADePT Results
Figure 3.5: Poverty Severity Curves in Rural Georgia, 2003 and 2006
Rural
100
80
Total severity, '000
60
40
20
0
0 160 320 480 640 800
Welfare indicator
2003 2006
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
Note: The red vertical line is the poverty line.
lower squared gap, or the former distribution has a lower squared gap for all
poverty lines.
Figure 3.5 graphs the poverty severity curves of rural Georgia for 2003 and
2006. The figure’s vertical axis reports total severity, which is directly propor-
tional to the squared gap measure of the corresponding poverty line. The solid
line denotes the poverty severity curve for 2003, while the dashed line denotes
the poverty severity curve for 2006. We saw earlier that the rural squared gap
measure is higher in 2006 for both poverty lines: GEL 75.4 and GEL 45.2.
What about the other poverty lines? Can we say that poverty has unambigu-
ously fallen for any poverty line? The figure suggests that we may not be able
to. One of the poverty severity curves does not lie below another poverty
severity curve for all poverty lines. When two poverty severity curves cross,
then an unambiguous judgment cannot be made based on the squared gap
211
A Unified Approach to Measuring Poverty and Inequality
Figure 3.6 graphs the growth incidence curve of Georgia’s per capita con-
sumption expenditure. The vertical axis reports the growth rate of consump-
tion expenditure between 2003 and 2006, and the horizontal axis reports
the per capita consumption expenditure percentiles. We earlier reported the
growth rate of mean per capita consumption expenditure and found that the
overall growth rate was slightly negative. We also compared the median and
four other quantile incomes.
Figure 3.6: Growth Incidence Curve of Georgia between 2003 and 2006
Urban
3
2
Annual growth rate %
–1
–2
–3
1 10 20 30 40 50 60 70 80 90 100
Expenditure percentiles
Growth-incidence 95% confidence bounds
Growth at median Growth in mean
Mean growth rate
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
212
Chapter 3: How to Interpret ADePT Results
However, that analysis does not give us the entire picture, so we perform
this dominance analysis through a growth incidence curve that graphs the
growth rate of per capita consumption expenditure for each percentile of the
population. The height of a growth incidence curve for a particular percentile
of population is the per capita consumption expenditure growth of that per-
centile. In fact, a growth incidence curve assesses how the quantile incomes
change over time. If the growth rates of the lower quintiles are larger than
the growth rates of the upper quintiles, then the growth is said to be pro-poor.
The dotted-dashed straight line denotes the growth in mean per capita
expenditure, which is negative in this case. It is not necessary that the entire
population received an equal share of this growth. It is evident from the fig-
ure that the per capita expenditure growth rate for the population’s higher
percentiles between 2003 and 2006 is much larger and more positive than
that for their lower percentile counterparts. Even though growth has been
mixed throughout the quantile incomes, the lowest quantile income has a
large negative growth. Given that the growth rate was negative, this means
that the population’s poorer section had a proportionally larger decrease in
its per capita expenditure.
What we can state by looking at the figure is that the quantile ratios—
such as 90/10, 80/20, or 90/50—increased between 2003 and 2006. The
shaded area around the growth incidence curve reports the 95 percent con-
fidence bounds. Can we say something about poverty? Yes, we can. For an
absolute poverty line, the headcount ratio between 2003 and 2006 should
not fall because the per capita expenditures of the population’s lower per-
centile decreased. Thus, growth in Georgia between 2003 and 2006 was not
pro-poor.
Lorenz Curve
Figure 3.7 graphs the Lorenz curve of urban Georgia’s per capita expenditure
for 2003 and 2006. The vertical axis reports the share of total consumption
expenditure, and the horizontal axis reports the percentile of per capita
expenditure. A Lorenz curve graphs the share of total consumption expendi-
ture spent by each percentile of the population. Thus, the height of a Lorenz
curve for a particular percentile is the share of total consumption expenditure
spent by that percentile out of the region’s total consumption expenditure.
The Lorenz curve’s height is 1 when the percentile is 1. In other words, the
share of the total consumption expenditure spent by the entire population is
213
A Unified Approach to Measuring Poverty and Inequality
Urban
1.0
0.8
Lorenz curve
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1.0
Cumulative population proportion
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
100 percent. The diagonal straight line denotes the situation of perfect
equality: each person has the same per capita expenditure.
As inequality increases, the Lorenz curve bows out, and the area between
the Lorenz curve and the line of perfect equality increases. The area between
a Lorenz curve and the line of perfect equality is proportionally related to
the Gini coefficient: it is twice the Gini coefficient. If a distribution’s Lorenz
curve lies completely to the right of another Lorenz curve, then the former
distribution has unambiguously lower inequality, and any Lorenz-consistent
measure—such as the Gini coefficient, the Atkinson class of indices, and
the generalized entropy measures—ranks the former distribution as less
unequal. If the Lorenz curves of two distributions cross, we cannot unam-
biguously rank those two distributions, even when one is ranked as more
unequal than another by all the Lorenz-consistent measures we discussed
earlier. Therefore, the Lorenz curve provides an opportunity to conduct a
sensitivity analysis for the reported inequality measures.
214
Chapter 3: How to Interpret ADePT Results
The solid line represents the Lorenz curve for 2003, while the dotted
line corresponds to 2006. It is evident that the dotted curve lies nowhere to
the left of the solid curve. This implies that the inequality in urban Georgia
unambiguously increased in 2006 compared with 2003. If these two curves
had crossed, then the reported inequality measures would not have neces-
sarily agreed with each other.
215
A Unified Approach to Measuring Poverty and Inequality
Figure 3.8: Standardized General Mean Curves of Georgia, 2003 and 2006
Total
3.0
2.4
Generalized mean
1.8
1.2
0.6
0
–5 –4 –3 –2 –1 0 1 2 3 4 5
Alpha
2003 2006
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
Advanced Analysis
216
Chapter 3: How to Interpret ADePT Results
Consumption Regression
Table 3.28 analyzes determinants of the variable used for measuring welfare
(the per capita consumption expenditure in this case). Rows denote the set
of regressors (such as logarithm of household size, share of children in the
age group of 0–6 years, share of male adults, share of elderly) and a set of
2003 2006
Urban Rural Urban Rural
Coef SE Coef SE Coef SE Coef SE
Factors A B C D E F G H
Household characteristics
1 Log of household size −0.093 0.06 −0.010 0.05 −0.001 0.06 0.051 0.05
2 Log of household size squared −0.020 0.03 −0.078*** 0.02 −0.102*** 0.03 −0.114*** 0.02
3 Share of children age 0–6 years (dropped) (dropped) (dropped) (dropped)
4 Share of children age 7–16 years −0.252*** 0.09 0.223** 0.09 0.249** 0.10 0.076 0.09
5 Share of male adults −0.064 0.10 0.254*** 0.09 0.477*** 0.11 0.251*** 0.10
6 Share of female adults −0.004 0.10 0.453*** 0.10 0.592*** 0.11 0.435*** 0.10
7 Share of elderly (age ≥60 years) −0.124 0.11 0.462*** 0.10 0.488*** 0.12 0.355*** 0.10
Characteristics of household head
8 Log of household head’s age −0.063 0.05 0.076 0.05 −0.318*** 0.05 0.210*** 0.05
Regions
9 Kakheti (dropped) (dropped) (dropped) (dropped)
10 Tbilisi 0.446*** 0.05 (dropped) 0.258*** 0.05 (dropped)
11 Shida Kartli 0.182*** 0.06 0.147*** 0.03 −0.050 0.06 0.182*** 0.03
12 Kvemo Kartli 0.061 0.06 0.075** 0.03 −0.023 0.06 0.183*** 0.03
13 Samtskhe-Javakheti −0.115* 0.06 0.185*** 0.03 0.231*** 0.07 0.163*** 0.04
14 Ajara 0.226*** 0.06 −0.035 0.04 0.103* 0.06 0.067* 0.04
15 Guria −0.077 0.08 0.250*** 0.03 0.030 0.08 0.131*** 0.04
16 Samegrelo 0.112** 0.06 0.194*** 0.03 −0.007 0.06 0.238*** 0.03
17 Imereti 0.270*** 0.05 0.529*** 0.03 0.208*** 0.05 0.381*** 0.03
18 Mtskheta-Mtianeti −0.060 0.07 0.164*** 0.03 0.020 0.08 0.144*** 0.04
sland
19 0 ha (dropped) (dropped) (dropped) (dropped)
20 Less than 0.2 ha 0.121*** 0.03 0.162*** 0.05 0.104*** 0.03 0.166*** 0.04
21 0.2–0.5 ha 0.180*** 0.04 0.356*** 0.04 0.138*** 0.05 0.193*** 0.03
22 0.5–1.0 ha 0.255*** 0.05 0.478*** 0.04 0.125* 0.07 0.365*** 0.03
23 More than 1.0 ha 0.021 0.09 0.565*** 0.05 0.192** 0.08 0.484*** 0.04
Gender of household head
24 Male (dropped) (dropped) (dropped) (dropped)
25 Female −0.073*** 0.02 −0.002 0.02 −0.101*** 0.02 −0.027 0.02
Education of household head
26 Elementary or less (dropped) (dropped) (dropped) (dropped)
27 Incomplete secondary −0.067 0.07 0.034 0.03 0.226*** 0.07 0.086*** 0.03
28 Secondary 0.021 0.06 0.105*** 0.03 0.179*** 0.06 0.196*** 0.03
29 Vocational-technical 0.118* 0.06 0.147*** 0.04 0.225*** 0.07 0.255*** 0.04
30 Special secondary 0.156*** 0.06 0.217*** 0.03 0.269*** 0.06 0.322*** 0.04
31 Higher education 0.289*** 0.06 0.274*** 0.03 0.441*** 0.06 0.477*** 0.04
(continued)
217
A Unified Approach to Measuring Poverty and Inequality
2003 2006
Urban Rural Urban Rural
Coef SE Coef SE Coef SE Coef SE
Factors A B C D E F G H
Employment status of household head
Self-employed
32 Agriculture (dropped) (dropped) (dropped) (dropped)
33 Industry −0.028 0.09 0.430*** 0.09 −0.122 0.11 0.208* 0.11
34 Trade 0.082 0.05 0.275*** 0.05 0.056 0.06 0.193*** 0.06
35 Transport 0.026 0.08 0.311*** 0.07 −0.039 0.08 0.311*** 0.07
36 Other services 0.072 0.07 0.340*** 0.08 −0.099 0.07 0.033 0.09
Employed
37 Industry −0.043 0.06 0.127** 0.06 −0.036 0.06 0.140** 0.06
38 Trade −0.094 0.06 0.144 0.09 −0.024 0.07 0.115 0.10
39 Transport −0.021 0.06 0.212*** 0.08 −0.174** 0.07 0.282*** 0.08
40 Government −0.041 0.06 0.227*** 0.06 0.012 0.07 0.277*** 0.08
41 Education −0.037 0.06 0.054 0.06 −0.029 0.07 0.045 0.07
42 Health care −0.041 0.08 0.085 0.15 −0.039 0.08 0.279* 0.15
43 Other −0.120** 0.05 0.150*** 0.04 −0.022 0.06 0.005 0.05
44 Unemployed −0.376*** 0.05 −0.138** 0.06 −0.325*** 0.05 −0.066 0.06
45 Inactive −0.219*** 0.04 −0.117*** 0.02 −0.169*** 0.05 −0.067*** 0.02
Other
46 Constant 4.851*** 0.21 3.425*** 0.19 5.328*** 0.20 2.976*** 0.20
47 Number of observations 4,525 7,106 4,112 6,773
48 Adjusted R2 0.18 0.20 0.17 0.16
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: Coef = coefficient. ha = hectare. SE = standard error, sland = area of land owned.
*** p < 0.01, ** p < 0.05, * p < 0.1.
218
Chapter 3: How to Interpret ADePT Results
Rows 46, 47, and 48 report the intercept term, number of observations,
and adjusted R-squares (R2), respectively. The intercept term, or constant
term, denotes the level of the consumption expenditure logarithm not
explained or determined by any regressors, or adjusted R-square denoted
power of prediction of all regressors, or the model’s goodness-of-fit. If the
adjusted R-square is 1, then the regressors predict the regressand with com-
plete accuracy. If a regressor’s p-value is less than 0.01, then *** is added
to the coefficient. If the p-value is less than 0.05, then ** is added to the
coefficient. Finally, if the p-value is less than 0.1, then * is added to the
coefficient. P-values denote regressors’ significance level.
Note that all variables in the regions, sland, gender of household head, edu-
cation of household head, and employment status of household head categories
are dummy variables or binary variables. They take a value of only 0 or 1.
A binary variable coefficient denotes the change in regressand when
the dummy variable’s value changes from 0 to 1. For example, consider the
coefficient of the regressor Female in the household head gender category for
urban regression in 2003. The coefficient is –0.073 [25,A], implying that the
per capita expenditure logarithm for a member in a female-headed household
is 0.073 units lower than that of a male-headed household. The regressor’s
standard error is 0.02 [25,B] with a p-value less than 0.01 (indicated by ***
after the regressor), and thus the coefficient is highly significant. The coef-
ficient of the same regressor for urban regression in 2006 is –0.101 [25,E]
with a p-value of less than 0.01, implying that the per capita consumption
expenditure gap between female- and male-headed households increased
over the three-year period.
219
A Unified Approach to Measuring Poverty and Inequality
2003 2006
Urban Rural Urban Rural
Variables A B C D
Demographic event, child born in the family
1 Change from having no children 0–6 years old to having 1 child 2.0 18.0 31.5 17.8
2 Change from having no children 0–6 years old to having 2 children 4.7 33.1 57.9 33.5
Land acquisition event
3 Change from “0 ha” to “less than 0.2 ha” −18.9 −15.3 −16.6 −18.0
4 Change from “0 ha” to “0.2–0.5 ha” −27.4 −33.6 −21.8 −20.9
5 Change from “0 ha” to “0.5–1.0 ha” −37.4 −44.5 −19.8 −38.5
6 Change from “0 ha” to “over 1.0 ha” −3.5 −51.9 −29.5 −49.6
Change of household head (following divorce, migration, and so forth)
7 Change from “Male” to “Female” 13.0 0.2 18.4 3.9
Education event: change in household head’s education
8 Change from “Elementary or less” to “Incomplete 10.5 −4.5 −28.0 −10.3
secondary”
9 Change from “Elementary or less” to “Secondary” −3.3 −13.6 −22.4 −23.0
10 Change from “Elementary or less” to “Vocational-technical” −17.6 −18.8 −27.8 −29.5
11 Change from “Elementary or less” to “Special secondary” −23.0 −27.2 −33.0 −36.7
12 Change from “Elementary or less” to “Higher education” −40.3 −33.7 −51.4 −51.8
Sector of employment event: household head’s sector of employment
13 Change from “Agriculture” to “Industry” 5.7 −53.1 25.3 −28.1
14 Change from “Agriculture” to “Trade” −15.6 −36.5 −10.6 −26.2
15 Change from “Agriculture” to “Transport” −5.2 −40.5 7.7 −40.3
16 Change from “Agriculture” to “Other Services” −13.8 −43.8 20.3 −4.7
17 Change from “Agriculture” to “Industry” 8.8 −17.7 7.1 −19.4
18 Change from “Agriculture” to “Trade” 19.6 −20.0 4.8 −16.0
19 Change from “Agriculture” to “Transport” 4.2 −28.7 36.8 −37.0
20 Change from “Agriculture” to “Government” 8.4 −30.7 −2.3 −36.4
21 Change from “Agriculture” to “Education” 7.5 −7.8 5.7 −6.4
22 Change from “Agriculture” to “Health Care” 8.3 −12.0 7.8 −36.7
23 Change from “Agriculture” to “Other” 25.3 −20.9 4.3 −0.7
24 Change from “Agriculture” to “Unemployed” 87.4 20.7 72.1 9.7
25 Change from “Agriculture” to “Inactive” 48.2 17.4 35.6 9.8
220
Chapter 3: How to Interpret ADePT Results
from owning 0 hectare of land to > 1 hectare of land, and change from
male-headed household to female-headed household. Columns report the
percentage changes in the probability of being in poverty for rural and urban
areas and across 2003 and 2006.
Recall from our discussion about table 3.28 that the interpretation of
dummy or binary variables is different from that of continuous variables.
A dummy variable, unlike a continuous variable, may take only a value of
either 0 or 1. Table 3.28 described how the probability of being in poverty
changes as values of certain variables change.
The probability of being in poverty in 2003 increased by 2.0 percent
[1,A] if an individual moved from an urban household with no children in
the 0–6 years age group to an urban household with one child in the same
age group, all other factors being identical. The probability of being in pov-
erty in 2003 is increased by 18.0 percent [1,B] if an individual moved from
a rural household with no children in the 0–6 years age group to a rural
household with one child in the same age group, all else being identical. In
the urban area, the increase in the probability of being in poverty in 2006
for the same reason is 31.5 percent [1,C].
Similarly, in 2003 if an individual moved from a male-headed urban
household to a female-headed urban household, all else being identical,
then the probability of being in poverty increased by 13.0 percent [7,A].
If an individual moved from a male-headed rural household to a female-
headed rural household, all else being identical, then the probability of
being in poverty increased by only 0.2 percent [7,B]. The largest increase in
the probability of being in poverty in 2003 in the urban area occurred when
an individual moved from a household where the head is employed in the
agricultural sector to a household where the head is unemployed [24,A], all
else being identical.
The table provides a detailed analysis of how the probability of being in pov-
erty changes when some of the crucial determinants of poverty are adjusted.
Note that if the household head’s education in the urban area in 2006
increased from elementary education or less to secondary education, all else
remaining equal, then the probability of a member in that household being
in poverty fell by 22.4 percent [9,C]. Similarly, in rural Georgia for both
years, if the household head transferred from the agricultural sector to any
221
A Unified Approach to Measuring Poverty and Inequality
other employment sector, all else being equal, then the probability of being
in poverty fell. Hence, this analysis provides a tool to better understand the
source of poverty and what type of policy would be more efficient in terms
of eradicating poverty.
Table 3.30 decomposes the change in poverty into a change in the mean per
capita consumption expenditure and a change in distribution of consump-
tion expenditure around the mean, following Huppi and Ravallion (1991).
Table rows denote three regions—urban, rural, and all of Georgia—for two
different poverty lines. The per capita consumption expenditure is measured
in lari per month. Poverty lines are set at GEL 75.4 (poor) and GEL 45.2
(extremely poor) for each household and household member. For simplicity
in this table, we present the decomposition for headcount ratio only, but the
technique is equally applicable to other poverty measures in the FGT class.
Columns A and B report the headcount ratio of the three regions for
years 2003 and 2006, respectively, and column C reports the changes over
time. Columns D, E, and F decompose the change in the headcount ratio
between 2003 and 2006 into three different terms. Column D reports the
effect of growth on poverty, referred to as the growth effect. Column E reports
the effect of redistribution on poverty and is called the redistribution effect.
Column F reports the interaction term and is referred to as the interaction
effect, following Huppi and Ravallion (1991).
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
222
Chapter 3: How to Interpret ADePT Results
It is evident from the table that the overall headcount ratio in 2003 is
29.9 percent [3,A], which increased to 31.0 percent in 2006 [3,B]. These
numbers can be verified from table 3.2. The actual change in the overall
headcount ratio is 1.0 percentage point (rounded) [3,C]. The actual change
is broken down into three components: growth effect, redistribution effect,
and interaction effect. By looking at the corresponding figures in columns D,
E, and F, we see that the change is caused mainly by redistribution rather
than growth. We can verify from table 3.1 that growth in mean is negligible
compared to change in inequality in terms of the Gini coefficient.
The picture is slightly different for the urban and rural areas. The urban
headcount ratio rose by 2.7 percentage points from 28.1 percent [1,A] to
30.8 percent [1,B], with both growth effect and redistribution effect being
positive. The urban redistribution effect [1,E] is more than three times
larger than the urban growth effect [1,D]. For the rural area, the headcount
ratio fell from 31.6 percent [2,A] to 31.1 percent [2,B]. In this case, both the
growth effect [2,D] and the redistribution effect [2,E] are negative.
The appendix contains additional tables and figures that may be helpful
in understanding concepts and results in terms of the data for Georgia in
2003 and 2006.
Note
Reference
Huppi, M., and M. Ravallion. 1991. “The Sectoral Structure of Poverty dur-
ing an Adjustment Period: Evidence for Indonesia in the Mid-1980s.”
World Development 19 (12): 1653–78.
223
Chapter 4
Ultra-Poverty
Our first enhancement builds on a theme that originally led to the con-
struction of poverty measures beyond the headcount ratio, namely, that
within the poor population important differences exist in the nature of
poverty. The headcount ratio P0 ignores these differences by valuing each
poor person equally without regard to the depth of poverty. Measures like
the poverty gap P1 reflect the depth of poverty among the poor, while oth-
ers like the FGT (Foster-Greer-Thorbecke) index P2 take into account its
distribution by emphasizing those with the largest gaps. The measurement of
ultra-poverty carries this differentiation one step further by focusing on the
poorest of the poor.
People who are most impoverished according to some well-defined cri-
terion are often the subject of special concern. The poverty experienced
by this group is often called “extreme” or “acute.” Here we use the term
ultra-poverty to describe the condition of poorest poor. Who are the ultra-
poor and how can their poverty be measured? The answer depends on the
225
A Unified Approach to Measuring Poverty and Inequality
It was argued above that an absolute poverty line zu may not be sustainable
when a large change occurs in the size of the income distribution. A similar
observation applies when comparing two countries at very different levels of
development using an absolute line. The problem is that when the income
226
Chapter 4: Frontiers of Poverty Measurement
standard varies a great deal, it seems reasonable that the poverty line should
reflect this change, at least to some extent. Yet an absolute poverty line, by
definition, is fixed and independent of any changes in the income standard.
Stated differently, when an income standard (such as the mean) changes by
1 percent, an absolute poverty line changes by 0 percent, so that the income
elasticity of the poverty line is zero.
An alternative approach uses a relative poverty line zr, defined as a fixed
proportion of a given income standard. For example, 60 percent of median
income is a relative poverty line used in the European Union. For relative
poverty lines, if a country’s income standard changes by 1 percent, then
the poverty line also changes by 1 percent, implying that the poverty line’s
income elasticity is one. An argument against this approach is that it makes
the poverty line too sensitive to changes in the income standard.
Several approaches have been explored to negotiate the landscape
between the extremes of absolute and relative poverty lines. Foster (1998)
suggests a hybrid poverty line that is a weighted geometric mean of rela-
tive and absolute poverty lines. In symbols, the poverty line is z = zrrza1-r,
where 0 ≤ r ≤ 1. Note that r can be interpreted as the income elasticity of
the hybrid poverty line, because when zr’s income standard rises by 1 percent,
the relative component zr rises by 1 percent, which, in turn, increases the hybrid
poverty line z by r percent.
On the one hand, if parameter r is set to zero so the entire weight is
given to the absolute component, then the hybrid poverty line becomes the
absolute poverty line where the elasticity is zero. On the other hand, if r is
one so the full weight is on the relative component, then the hybrid poverty
line becomes the relative poverty line and the elasticity is one. If 0 < r < 1,
then the hybrid poverty line will lie between the absolute and relative lines
and have an elasticity between zero and one.
How is the elasticity to be set? One approach is to estimate the param-
eter using data on existing poverty lines and income levels. Foster and
Székely (2006) regress poverty lines on private consumption per capita for
92 household surveys across 18 countries and find an elasticity of 0.36. A
second approach is to select “reasonable” values and check for robustness.
Madden (2000), for example, analyzed Irish poverty using the 1987 and
1994 Irish Household Budget Surveys for two intermediate values of the
parameter, 0.5 and 0.7, and found that results were similar for the two.
Finally, by interpreting r as the extent to which society believes the
poor should share in growth, we can view the selection of r as a normative
227
A Unified Approach to Measuring Poverty and Inequality
The previous analysis applies to any cardinal welfare indicator, where cardi-
nality requires that values convey more information than just more or less.
Nonmonetary examples of cardinal variables might include calories, years of
schooling, or hectares of land. Many other variables are more appropriately
interpreted as ordinal, because their values are only indicators of order.
Others might be categorical and have no values or underlying ordering at
all. Examples of ordinal variables include self-reported health and subjec-
tive well-being. Categorical variables include sanitation facilities or the
228
Chapter 4: Frontiers of Poverty Measurement
floor materials in a house. What can be done if we want to evaluate the size,
spread, or base of such a welfare indicator?
Allison and Foster (2004) describe ways of comparing distributions of
self-reported health in terms of spread and, in the process, provide new
approaches to evaluating size and base for this ordinal variable. The main
tools are dominance rankings. Changes in size and poverty are evaluated
using first-order stochastic dominance. Changes in spread are twin first-
order dominance movements away from the median category. To calculate a
mean, an inequality measure, or an FGT poverty index for a > 0, one must
cardinalize the ordinal variable, and hence the comparisons obtained are
not generally meaningful (because a different cardinalization could reverse
the ranking).
The headcount ratio, however, is identical for all cardinalizations and
thus is an appropriate tool for measuring poverty when the variable is
ordinal or even categorical. Of course, the headcount ratio provides no
information at all about depth. Bennett and Hatzimasoura (2011) provide
one approach to evaluating depth with an ordinal variable, based on a
reinterpretation of the poverty gap as “average headcount ratios” across
different poverty lines.
Chronic Poverty
Returning to the case of income, we saw how poor people can differ
from one another in policy-relevant ways. For example, poor people with
deeper income shortfalls are distinct from those just below the poverty
line. Time is a second dimension for differentiating among the poor:
persistent poverty is different from temporary poverty. Persistent poverty is
usually termed chronic poverty, and there are two main ways of identifying
and measuring it:
229
A Unified Approach to Measuring Poverty and Inequality
Multidimensional Poverty
230
Chapter 4: Frontiers of Poverty Measurement
used to identify the poor and measure poverty. Several reasons exist for
this interest:
231
A Unified Approach to Measuring Poverty and Inequality
232
Chapter 4: Frontiers of Poverty Measurement
233
A Unified Approach to Measuring Poverty and Inequality
Multidimensional Standards
234
Chapter 4: Frontiers of Poverty Measurement
income and other welfare indicators are often positively correlated, for
both individual and country-level data, which may suggest that the
nonincome indicators are not needed. But as emphasized by Sen (1999),
notable exceptions to these regularities exist. A proper measure of progress
should convey empirical realities in all eventualities, including excep-
tional cases. Correlation does not justify dropping important dimensions
in assessing progress.
The Human Development Index (HDI) of the United Nations
Development Programme was designed as an alternative to income per
capita that includes health and education achievements in a country
(thus addressing the second critique). The underlying structure of the
traditional HDI is straightforward, even if the details of its construction
are not. A country’s achievements in income, health, and education are
summarized in three indicators that are normalized to lie between zero
and one. The traditional HDI is a simple mean of these components. The
precise construction of the indicators—including the choice of “goalposts”
for normalizing a variable and its specific transformation—can affect the
HDI’s picture of development across countries. As an example, the income
indicator used in the HDI is based on a logarithm of income per capita; if
the untransformed variable were used instead, the ranking at the upper end
would more closely follow the income ranking of these countries.
This lack of robustness is indicative of the challenge of constructing
component indicators that can be meaningfully combined into a composite
indicator. One alternative to combining dimensions into an overall indica-
tor is to provide a dashboard of dimensional indicators. If indicators were not
being combined, normalizing goalposts and special transformations would
not be needed; the variables could be presented in their original, more
comprehensible forms. In particular, one could dispense with the log trans-
formation, because average income itself would rank countries the same
way—within this dimension.
However, many good reasons exist for using a composite indicator rather
than a vector of components. A single numerical indicator is more salient
and easier to track. A comprehensive measure emphasizes the point that we
are more interested in overall progress than progress in any given dimension.
Moreover, given that the aggregation formula is decomposable, it invites
further analysis to identify which components are driving the overall results.
The success of the HDI would have been unlikely if only a dashboard of
dimensional indicators had been provided.
235
A Unified Approach to Measuring Poverty and Inequality
236
Chapter 4: Frontiers of Poverty Measurement
237
A Unified Approach to Measuring Poverty and Inequality
Inequality of Opportunity
The previous section examined the general case where several welfare indi-
cators contribute to a person’s well-being. We now return to the simpler
case of a single welfare indicator, but where other variables provide infor-
mation on relevant characteristics or “identities” of the individuals. Recent
238
Chapter 4: Frontiers of Poverty Measurement
work has moved from evaluations of inequality across all people to measures
of inequality across groups of people, with the goal of isolating forms of
inequality that are particularly objectionable or policy relevant.
Roemer (1998) divides identity variables into circumstances, which
are unrelated to actions taken by the person and hence the person is not
accountable for such circumstances, and efforts, which are under the per-
son’s control. He argues that inequality across groups of people defined by
circumstances is particularly objectionable. For example, income inequal-
ity across racial groups or across groups defined by the education levels of
one’s parents should be of special concern because it reflects an underlying
inequality of opportunity. Ferreira and Gignoux (2008) implement this
approach by applying Theil’s second inequality measure, or the mean log
deviation, to a smoothed distribution defined by replacing each income in
a group with the group mean. In other words, inequality of opportunity is
measured as a between-group inequality term. This general approach can
be applied for different circumstance variables, and hence ways of defining
groups, to obtain different inequality of opportunity measures conditional
on that choice.
Stewart (2002) contends that group inequalities, which she calls horizon-
tal inequalities, can be more important than overall or vertical inequalities.
But rather than invoking a normative notion of equal opportunity, she uses
an empirical argument: horizontal inequalities, such as those across ethnic
groups, tend to be more closely linked to conflict than are vertical inequali-
ties. Stewart emphasizes that many possible dimensions of achievements
could be evaluated. The horizontal inequalities in a given dimension for a
configuration of groups can be measured and monitored using the associated
between-group inequality term.
The World Bank’s Human Opportunity Index (HOI) is another group
inequality measure that uses an opportunity interpretation of group inequal-
ities. Here the focus is on the provision of social services, so the underlying
distribution is taken to be dichotomous, with a zero being posted for all
people without access to the service and a one for those having access. The
overall mean of this variable then corresponds to the coverage rate for the
social service. The aim is to go beyond the mean coverage to account for
differential coverage rates across population subgroups, where the groups
are defined using circumstantial variables. An inequality measure is applied
to the smoothed distribution (which replaces a person’s actual value with
the group’s coverage rate) to obtain a measure of inequality of opportunity.
239
A Unified Approach to Measuring Poverty and Inequality
The HOI is the overall coverage rate discounted by the inequality of oppor-
tunity or, equivalently, a distribution-sensitive income standard applied
directly to the smoothed distribution.
The inequality measure used in the original HOI was the relative mean
deviation, a rather crude inequality measure that ignores transfers on either
side of the mean (see de Barros and others 2009). However, it is easy to
consider other inequality measures that generate between-group inequality
measures that are sensitive to differential coverage across subgroups on the
same side of the mean. For example, if we use the Atkinson inequality mea-
sure based on the geometric and arithmetic means, the resulting HOI will
evaluate the smoothed distribution using the geometric mean. Note that
every different social service can lead to a different picture of a population’s
opportunity to access social services. An overall view may require aggregat-
ing access to services at the individual level or aggregating HOIs into an
overall index. In addition, the measure is dependent on the particular cir-
cumstances selected to define population subgroups. These implementation
challenges are worthwhile because the measures can help reveal inequalities
that are especially salient and unjust.
Polarization
240
Chapter 4: Frontiers of Poverty Measurement
and the within-group inequality (as measured by the Gini coefficient) times
the ratio of mean to median (where the ratio of mean to median is a measure
of skewness of the distribution).
Foster and Wolfson (1992, 2010) also propose dominance orderings
based on polarization curves that can determine whether unambiguous
increases in polarization have taken place. First-order polarization occurs
when there are first-order stochastic dominant movements away from the
median. Second-order polarization occurs when there are second-order
dominant movements away from the median. The Foster-Wolfson polar-
ization measure is related to the area below the second-order polarization
curve. This approach has been extended by Zhang and Kanbur (2001) and
Chakravarty and D’Ambrosio (2010).
In contrast to the Foster-Wolfson approach, in which two groups of
observations are endogenously defined using the median as the dividing
line, Esteban and Ray (1994) assume that several groups of observations are
exogenously given, each around its own pole. Their polarization measure
rises when the groups pull apart from one another, or when observations
within a group become more tightly clustered together. The measure is
challenging to implement in practice because no clear way is given for
dividing an overall distribution into relevant clusters. These and other
practical problems of implementation are addressed in Duclos, Esteban,
and Ray (2004).
References
241
A Unified Approach to Measuring Poverty and Inequality
242
Chapter 4: Frontiers of Poverty Measurement
243
A Unified Approach to Measuring Poverty and Inequality
244
Chapter 5
This chapter provides basic information about installing and using ADePT.
The instructions are sufficient to perform a simple analysis. More informa-
tion is available:
245
A Unified Approach to Measuring Poverty and Inequality
• Windows, buttons, tabs, dialogs, and other features you see on screen
are shown in bold. For example, the Save As dialog has a Save button
and a Cancel button.
• Keystrokes are shown in small caps. For example, you may be
instructed to press the enter key.
• Menu commands use a shorthand notation. For example, Project
Exit means “open the Project menu and click the Exit command.”
Installing ADePT
System Requirements
Installation
246
Chapter 5: Getting Started with ADePT
Launching ADePT
3. You now see the ADePT main window. (The example below shows
ADePT configured with the Poverty module. The lower left-hand
and upper right-hand panels will be different when another module
is loaded.)
247
A Unified Approach to Measuring Poverty and Inequality
248
Chapter 5: Getting Started with ADePT
The next sections in this chapter provide detailed instructions for the
four steps.
Specify Datasets
249
A Unified Approach to Measuring Poverty and Inequality
Operations in this section take place in the upper left-hand corner of the
ADePT main window.
250
Chapter 5: Getting Started with ADePT
• To remove a dataset: Click the dataset, then click the Remove button.
Three datasets have been specified in this example.
251
A Unified Approach to Measuring Poverty and Inequality
Map Variables
252
Chapter 5: Getting Started with ADePT
Method 2: In the upper dataset Variables tab, drag the variable name and
drop it in the corresponding field in the lower input Variables tab.
Note: You can also type dataset variable names in the input variable
fields. The above methods are preferred, however, because typing
may introduce spelling errors. A spelling error is indicated by the red
exclamation point next to the input variable field.
253
A Unified Approach to Measuring Poverty and Inequality
In some input variable fields, you can specify multiple dataset variables.
For example, in the ADePT Poverty module, you can specify two poverty
lines (variables or numeric constants) instead of one, and the program will
replicate all tables for each of the specified poverty lines.
In this example, the pline_u and pline_l dataset variables have been
mapped to the Poverty line(s) input variable.
The italic variable name indicates that this input variable field accepts
multiple dataset variables. When you select or drag a new input variable
to one of these fields, it is appended to the previous value rather than
replacing it.
Tip: Open the example project (Project Open Example Project) to
see the result of mapping dataset variables to input variables.
After mapping variables, you are ready to select the tables and graphs you
want ADePT to generate.
Operations described in this section take place in the right-hand side of
the ADePT main window.
254
Chapter 5: Getting Started with ADePT
In the upper right-hand (outputs) panel, select the tables and graphs you
want to generate.
Note: If a name is gray, it cannot be selected. These tables and graphs cannot be gener-
ated because required variables have not been specified.
255
A Unified Approach to Measuring Poverty and Inequality
256
Chapter 5: Getting Started with ADePT
257
A Unified Approach to Measuring Poverty and Inequality
258
Chapter 5: Getting Started with ADePT
259
A Unified Approach to Measuring Poverty and Inequality
260
Chapter 5: Getting Started with ADePT
• To hide or show variable columns in the Data View tab: In the Variable View tab, click
the checkbox next to the variable name.
Tip: The ADePT User’s Guide describes other functions available in the Data
Browser.
Generating Variables
You can create new variables that are based on variables present in a dataset. This might be
useful for simulating the effects of changes in parameters on various economic outcomes. For
example, in the Poverty module you can model the effect of income transfers on some popula-
tion groups on the basis of poverty and inequality.
1. In the Datasets tab in the main window, click the dataset that you want to modify.
2. Click the Variables | [dataset label] tab.
3. Right-click in the table, then click Add or replace variable... in the pop-up menu.
261
A Unified Approach to Measuring Poverty and Inequality
The new variable will be listed in the Variables | [dataset name] tab
and in the Data Browser. If the variable was generated for all loaded data-
sets, it will appear in the drop-down lists in the input Variables tab.
When you save a project, variable expressions are saved with the project,
and the variables are regenerated when you open that project. Generating
new variables does not change original datasets.
Replacing Variables
262
Chapter 5: Getting Started with ADePT
Variable Expressions
Variable expressions can include constants, and strings can be used for
variables that are of type string.
Expression examples are as follows:
Expression Description
x=1 sets all variable x observations to 1
x=y+z sets variable x observations to y observation plus z observation
x=y=1 sets variable x observations to 1 (true) if y is 1; otherwise, sets variable x
observations to 0 (false)
x = 23 if z == . sets variable x observations to 23 if z is missing ( . ); otherwise, sets to.
x = Log(y) if z = 1 sets variable x observations to log of y observations if z is 1; otherwise, sets to.
s = “test” sets all variable x observations to the string “test”
Note: The periods ( . ) in the table above represent system-missing values. This symbol is defined in
SPSS® and is used to indicate missing data in datasets.
Deleting Variables
You can remove variables from the working copy of a dataset that ADePT
uses for its calculations. This operation does not change the original data-
set. Native variables, as well as generated and replaced variables, can be
deleted.
263
A Unified Approach to Measuring Poverty and Inequality
Setting Parameters
Some modules have a Parameters tab next to the input Variables tab. In
the Parameters tab, you can set ranges, weightings, and other module-
specific factors that ADePT will apply during its processing. A Parameters
tab may also have input variable fields for mapping dataset variables, as
shown in the drop-down list below.
After specifying datasets and mapping variables, you can save the con-
figuration for future use. A saved project stores links to datasets, variable
264
Chapter 5: Getting Started with ADePT
• To save a project:
a. Project Save Project or Project Save As...
b. In the Save As dialog, select a location and name for the project,
then click the Save button.
265
A Unified Approach to Measuring Poverty and Inequality
longer. A good approach is to obtain the result you want without stan-
dard errors, then generate final results with standard errors.
266
Chapter 5: Getting Started with ADePT
1. In the list of tables and graphs, click the table or graph name.
2. Enter the if-condition at the bottom of the Table description and
if-condition tab (see list of operators below).
267
A Unified Approach to Measuring Poverty and Inequality
268
Chapter 5: Getting Started with ADePT
269
Appendix
This appendix provides additional tables and figures that may be useful in
understanding the concepts and results discussed in chapters 1–3. We use
the same Integrated Household Survey dataset of Georgia for 2003 and
2006 that we used in chapter 3. Results in this appendix are reported at the
national level, with rural and urban breakdown, and at the subnational level
for 2003 only. Figures for a particular region cover both 2003 and 2006.
271
A Unified Approach to Measuring Poverty and Inequality
General mean
Sen
a=1 a=2 a=0 a = –1 a = 0.5 mean
Region A B C D E F
1 Urban 128.9 155.5 106.1 84.7 117.2 85.7
2 Rural 123.5 151.8 99.1 75.4 111.0 79.9
Subnational
3 Kakheti 107.9 131.6 87.2 65.7 97.4 70.7
4 Tbilisi 144.5 171.8 121.5 101.4 132.5 98.0
5 Shida Kartli 122.9 153.6 96.3 67.8 109.3 77.9
6 Kvemo Kartli 93.5 113.7 77.3 61.8 85.2 63.0
7 Samtskhe-Javakheti 116.5 142.3 96.2 76.8 106.0 78.2
8 Ajara 107.8 129.9 87.9 68.5 97.7 71.2
9 Guria 134.3 166.9 109.9 86.4 121.6 88.7
10 Samegrelo 117.2 142.3 95.7 75.8 106.1 77.2
11 Imereti 150.3 178.9 124.3 99.6 137.1 100.7
12 Mtskheta-Mtianeti 113.0 134.0 92.0 71.5 102.6 75.2
13 Total 126.1 153.6 102.4 79.7 113.9 82.7
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
Columns A through E show the general means for five different values
of the inequality aversion parameter a : a = 1 for the arithmetic mean,
a = 2 for the Euclidean mean, a = 0 for the geometric mean, a = –1 for
the harmonic mean, and a = 0.5. From our discussions of general means
in chapter 2, we know that a distribution’s general mean decreases as a
increases. Column F lists the Sen mean.
Column A reports the mean per capita consumption expenditure when
a = 1. The other income standards, with the mean, can be used to construct
a particular inequality measure. For example, the mean can be combined
with the Euclidean mean to construct the generalized entropy measure for
a = 2. The mean and the geometric mean can be used to construct the
Atkinson inequality measure A(0) and the generalized entropy measure
GE(0). The mean and the harmonic mean are used together to compute the
Atkinson measure of inequality A(–1). The mean and the general mean for
a = 0.5 are combined to compute A(0.5). Finally, the mean and the Sen
mean can be used to compute the Gini coefficient.
For example, the mean per capita expenditure in Kakheti is GEL 107.9
[3,A], whereas the Sen mean is GEL 70.7 [3,F]. Thus, the Gini coefficient
is easily computed as 100 × (107.9 – 70.7)/107.9 = 34.4, which can be veri-
fied from table 3.8. Similarly, the mean for Tbilisi is GEL 144.5 [4,A] and
272
Appendix
the geometric mean is GEL 121.5 [4,C], so the Atkinson measure A(0) is
computed as 100 × (144.5 – 121.5)/144.5 = 15.9, which can be verified from
table 3.27.
Doubly
censored General mean
mean a =1 a =0 a = –1 Sen mean
Region A B C D E
Poverty line = 75.4
1 Urban 54.2 68.9 66.9 62.9 63.6
2 Rural 51.6 67.4 64.5 58.6 60.9
Subnational region
3 Kakheti 46.1 65.3 61.8 54.4 57.7
4 Tbilisi 59.7 71.3 70.2 68.4 67.7
5 Shida Kartli 48.8 66.6 63.0 54.2 59.6
6 Kvemo Kartli 41.9 63.8 60.0 53.7 55.3
7 Samtskhe-Javakheti 52.7 67.9 65.3 60.4 61.8
8 Ajara 47.4 65.8 62.4 56.4 58.3
9 Guria 56.4 69.1 67.1 62.9 63.9
10 Samegrelo 50.2 67.1 64.5 59.7 60.6
11 Imereti 59.8 70.8 69.3 66.5 66.7
12 Mtskheta-Mtianeti 49.6 65.5 62.1 56.8 57.9
13 Total 52.8 68.1 65.6 60.6 62.2
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
273
A Unified Approach to Measuring Poverty and Inequality
censored distribution, people whose per capita expenditure is not less than
the poverty line are assumed to have poverty-line income, and people whose
per capita expenditure is less than the poverty line are assumed to have zero
per capita expenditure. The doubly censored mean is the mean of the doubly
censored distribution. The rest of the columns report income standards for
distributions that are censored once at the poverty line. Columns B, C, and
D show the arithmetic mean, the geometric mean, and the harmonic mean,
respectively. Column E reports the censored distribution’s Sen mean.
Those five censored income standards are related to five different poverty
measures, as explained in chapter 3. If the poverty line is denoted by z and a
censored income standard is denoted by a, then a poverty measure can be com-
puted by combining each of those five income standards and the poverty line.
The poverty line in this exercise is z = GEL 75.4. If the censored
income standard a is the doubly censored mean, then the headcount ratio is
(z – a)/z. Similarly, if the censored income standard a is the censored arith-
metic mean and the censored Sen mean, then (z – a)/z would be the poverty
gap measure and the Sen-Shorrocks-Thon (SST) index, respectively. If the
censored income standard a is the censored geometric mean, then the corre-
sponding poverty measure is the Watts index, computed as lnz – lna. Finally,
if the censored income standard a is the censored harmonic mean, then the
corresponding poverty measure is the Clark-Hemming-Ulph-Chakravarty
(CHUC) index, computed as (z – a)/z. Thus, a mere comparison of the
censored income standards for the same poverty line can provide a good
understanding for poverty comparisons.
Here is how different poverty measures can be obtained using each of
these censored income standards.
• In table 3.2, Georgia’s headcount ratio in 2003 for poverty line GEL
75.4 is 29.9. This can be obtained from table A.2 using the national
doubly censored mean of GEL 52.8 [13,A]: 100 × (75.4 – 52.8)/
75.4 = 29.9.
• In table 3.2, the national poverty gap measure is 9.7. This can be
obtained from table A.2 using the poverty line and the national
censored arithmetic mean of GEL 68.1 [13,B]: 100 × (75.4 – 68.1)/
75.4 = 9.7.
• In table 3.26, the national Watts index is 13.9. This can be obtained
from table A.2 using the poverty line and the national censored geo-
metric mean GEL 65.6 [13,C]: 100 × (ln75.4 – ln65.6) = 13.9.
274
Appendix
• In table 3.26, the national CHUC index is 19.6. This can be obtained
from table A.2 using the poverty line and the national censored har-
monic mean of GEL 60.6 [13,D]: 100 × (75.4 – 60.6)/75.4 = 19.6.
• In table 3.26, the national SST index is 17.5. This can be obtained
from table A.2 using the poverty line and the national censored Sen
mean of GEL 62.2 [13,E]: 100 × (75.4 – 62.2)/75.4 = 17.5.
Table A.3 presents a tool for checking the sensitivity of three poverty mea-
sures to consumption expenditure: the Watts index, the SST index, and the
CHUC index. In the table, we ask what the percentage change in poverty
would be if everyone’s consumption expenditure increased by 1 percent.
Results are compared across 2003 and 2006.
The percentage change in poverty caused by a 1 percent change in the
mean or average per capita consumption expenditure is called the elasticity
of poverty with respect to per capita consumption. The particular way in which
we consider an increase in the average per capita consumption expenditure
is by increasing everyone’s consumption expenditure by the same percent-
age. This type of change is distribution neutral, because the relative inequal-
ity does not change.
Table A.3: Elasticity of Watts Index, SST Index, and CHUC Index to Per Capita Consumption
Expenditure
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: Change is shown between years 2003 and 2006. CHUC = Clark-Hemming-Ulph-Chakravarty; SST = Sen-Shorrocks-Thon.
275
A Unified Approach to Measuring Poverty and Inequality
Because poverty lines are set normatively, they are difficult to justify exclu-
sively. A slight change in per capita consumption expenditure may or may
not change the poverty measures by significant amounts. If the distribution is
highly polarized or, in other words, there are two groups in the society—one
group of rich people and the other group of extremely poor people—then a
slight change in everyone’s income by the same proportion may not have any
impact on headcount ratio. In contrast, if there is a concentration of mar-
ginal poor around the poverty line, then a slight change in everyone’s income
by the same proportion would have a huge impact on poverty rates. Hence,
this type of analysis may tell us how policy changes impact the poverty rate.
276
Appendix
The exercise in table A.4 is analogous to the exercise for checking the elas-
ticity of poverty measures to per capita consumption expenditure, but it is
more rigorous. It is always possible to find a certain percentage of decrease
in the poverty line that matches the increase in the consumption expendi-
ture for everyone by 1 percent. In this exercise, we check the sensitivity of
poverty measures by changing the poverty line in more than one direction.
The table shows how the actual headcount ratio changes as the poverty
line changes from its initial level, whether GEL 75.4 per month or GEL
45.2 per month. Rows denote the change in poverty line in both upward
and downward directions. Columns report the change in three poverty mea-
sures: Watts index, SST index, and CHUC index. The variable is per capita
consumption expenditure measured in lari. This table shows results for 2003
only, but this analysis can be conducted for any year.
Columns A and B report the national Watts index for different pov-
erty lines, and column C shows the change in the index from the actual
poverty line. The rows corresponding to +5 percent denote the results for a
Table A.4: Sensitivity of Watts Index, SST Index, and CHUC Index to the Choice of Poverty Line,
2003
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003.
Note: CHUC = Clark-Hemming-Ulph-Chakravarty; SST = Sen-Shorrocks-Thon.
277
A Unified Approach to Measuring Poverty and Inequality
5 percent increase in the poverty line. Thus, when the poverty line is GEL
75.4, a 5 percent increase moves the poverty line to GEL 79.2. The Watts
index increases by 1.5 points from 13.9 [1,A] to 15.4 [2,A], or by 11 percent
from its actual level of 13.9.
Similarly, if the poverty line changes by –10 percent from GEL 75.4,
then the poverty Watts index falls by 2.8 from 13.9 [1,A] to 11.1 [6,A], or
by 20.4 percent from the actual level of 13.9. This index is more sensitive
to change in the poverty line when the actual poverty line is lower at GEL
45.2. In fact, the SST index and the CHUC index are also more sensitive
to change in poverty line when the actual poverty line is GEL 45.2 rather
than GEL 75.4.
278
Appendix
2003 2006
A B
1 Total 34.4 35.4
Urban and rural
2 Within-group inequality 17.2 17.7
3 Between-group inequality 1.1 0.5
4 Overlap term 16.2 17.2
Geographic regions
5 Within-group inequality 4.9 5.2
6 Between-group inequality 8.7 7.0
7 Overlap 20.8 23.2
Row 1 reports the overall Gini coefficients. Subsequent rows report Gini
coefficient decompositions for two different population subgroups: rural
and urban regions and geographic regions. The first row of each set reports
the within-group inequality and the second and the third rows report the
between-group inequality and the overlap term, respectively. The overall
Gini coefficient in 2003 is 34.4 [1,A], which increases to 35.4 in 2006 [1,B].
Thus, in terms of the Gini coefficient, inequality increased in 2006.
The first set decomposes the population into rural and urban areas. The
total within-group Gini coefficient is 17.2 in 2003 and increases to 17.7 in
2006 [row 2]. However, the between-group inequality decreased from 1.1
in 2003 to 0.5 in 2006 [row 3]. The overlap term registers an increase from
16.2 to 17.2 [row 4].
The decomposition of population by geographic regions has a similar
story. The total within-group inequality increases from 4.9 in 2003 to 5.2 in
2006 [row 5], but the between-group inequality decreases from 8.7 in 2003 to
7.0 in 2006 [row 6], and the overlap term increases from 20.8 in 2003 to 23.2
in 2006 [row 7]. Note that the overlap term is larger for the decomposition
across geographic regions [row 7] than across rural and urban areas [row 4].
A possible reason could be the number of groups: as the number of groups
increases, the possibility of overlap increases.
This type of analysis is important for policy purposes and may affect policy
recommendations. Both the overall inequality and the intergroup inequality
279
A Unified Approach to Measuring Poverty and Inequality
The Gini coefficient is not decomposable in the usual way because it has
an overlap term. Thus, it is important to look at the usual decomposition
(within-group and between-group inequalities) using additively decompos-
able measures. With this objective, table A.6 analyzes the decomposition
of inequality across urban and rural areas and across geographic regions.
The analysis is based on three different types of generalized entropy (GE)
measures: the first Theil measure denoted by GE(1), the second Theil
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: GE = generalized entropy.
280
Appendix
281
A Unified Approach to Measuring Poverty and Inequality
where x– is the mean of the income vector x and N is the total population
size.
Suppose the overall population is divided into K > 1 population
subgroups. These population subgroups may be different geographic regions,
ethnic groups, or rural and urban regions. For rural and urban decom-
position, K = 2. We denote the income vector of subgroup k by xk, the
population size of subgroup k by Nk, and the mean income of subgroup k by
x– k. Let us denote the population share of subgroup k by vk = Nk/N and the
income share of subgroup k by m k = x– k / x–. The second Theil measure can
then be decomposed as
IT 2 (x) = ∑ k =1 v k IT 2 (x k ) + ∑ k =1 v k ln 1 .
K k
mk (A.2)
282
Appendix
(A.3)
where Δ represents the change in the variables from time t0 to t1. The four
components can be interpreted as (a) the intertemporal change in within-
group inequality, (b) the change in the population shares of the groups in
the within-group component, (c) the change in population shares of the
groups in the between-group component, and (d) the change in the relative
incomes of the subgroups.
Table A.7 provides a dynamic decomposition of the overall Georgian
income inequality using the second Theil measure. Results in the table cor-
respond to changes across years 2003 and 2006. The variable for our analysis
is consumption expenditure in lari per month. Row 1 reports the change in
overall inequality. Rows 2 through 5 decompose this change into four factors,
as explained in the previous paragraph. Row 2 reports the change in overall
within-group inequality. Rows 3 and 4 report the effect of changes in popula-
tion shares on the within-group inequality and the between-group inequality,
respectively. Row 5 reports the change in relative subgroup incomes.
GE(0)
A
1 Change in aggregate inequality −0.011
2 Within-group inequality −0.015
3 Population shares of within-group inequality 0.000
4 Population shares of between-group inequality 0.000
5 Mean group incomes 0.004
283
A Unified Approach to Measuring Poverty and Inequality
The decrease in the overall inequality between 2003 and 2006 is −0.011
[1,A]. Row 2 indicates that this decline is mostly attributed to the decrease
in the within-group inequality because it is evident from row 5 that the
relative income share does not change in the same direction. The effect of
change in population share on the within-group inequality [row 3] and the
between-group inequality [row 4] is negligible.
In table A.8, we first break down the single variable into several compo-
nents, then we decompose the overall inequality across that variable into
the inequality of its components. For example, the total disposable income
of a household has several components such as male earnings, female earn-
ings, benefits, and income taxes. Analyzing inequality across disposable
income may not reveal inequality across these various components. This
type of inequality decomposition into factor components was studied in
detail by Shorrocks (1982), but only for a single period. Jenkins (1995)
conducted a dynamic intertemporal decomposition analysis across the popu-
lation. Following Jenkins, we use the generalized entropy measure of order
Relative Proportionate
Mean Correlation Absolute factor
mean GE(2) factor contribution
(GEL) with total contribution
(%) (%)
A B C D E F
2003
1 Food consumption 76.9 61.0 80.8 27.2 12.7 52.3
2 Expenditures on nonfood goods 15.2 12.0 62.5 57.2 3.0 12.4
3 Utilities 8.4 6.7 35.5 140.0 1.4 5.9
4 Expenditures on services 17.4 13.8 55.4 140.5 4.8 19.6
5 Other expenditures 8.2 6.5 48.5 179.6 2.4 9.8
6 Per capita consumption expenditure 126.1 100.0 24.2 24.2 100.0
2006
7 Food consumption 72.8 57.8 72.3 26.2 11.3 40.5
8 Expenditures on nonfood goods 13.2 10.5 56.3 74.8 2.9 10.5
9 Utilities 10.4 8.3 40.2 161.3 2.3 8.2
10 Expenditures on services 20.2 16.1 62.9 221.9 8.4 30.2
11 Other expenditures 9.3 7.4 50.0 186.7 2.9 10.6
12 Per capita consumption expenditure 126.0 100.0 27.8 27.8 100.0
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of Georgia 2003 and 2006.
Note: GE = generalized entropy.
284
Appendix
two for our analysis in this table, mainly because some components may be
zero and the measure is additively decomposable, as discussed in chapter 2.
Before discussing the results, let us provide a brief theoretical back-
ground. Interested readers can refer to Shorrocks (1982) for a further theo-
retical discussion. The following theoretical brief was heavily drawn from
Shorrocks (1982) and Jenkins (1995). Suppose the variable for our analysis
is income and is denoted by vector x. Income has K components, and the
distribution of the kth component across the population is denoted by xk.
The mean of incomes is denoted by x–, and the mean of the kth component
is denoted by x– k. Inequality across incomes is denoted by IGE(x; 2), and
inequality across the kth component is denoted by IGE(xk; 2). The overall
inequality can be expressed as
where rk is the correlation between x and xk, and Xk is the share of that
component in the overall income. Thus, Sk is the absolute contribution of
component k to the overall income. It turns out that the relative contribu-
tion of component k is
∑
K
sk = Sk / IGE (x; 2) = rk c k IGE (x k ; 2)/ IGE (x; 2) and S = 1. (A.5)
k =1 k
k
t +1
ΔIGE (x; 2) = IGE (x; 2) − IGE
t
(x; 2) = ∑ ΔSk
k =1
k
= ∑ Δrk c k IGE (x k ; 2)/ IGE (x; 2). (A.6)
k =1
t +1
IGE (x; 2) − IGE
t
(x; 2) K ΔSk K
ΔS K
d IGE (x; 2) = t
=∑ = ∑ sk k = ∑ skd Sk . (A.7)
IGE (x; 2) k =1 Sk / sk k =1 Sk k =1
Table A.8 presents the results using the Georgian dataset for 2003 and
2006. Rows denote different categories of consumption expenditure on food
285
A Unified Approach to Measuring Poverty and Inequality
items, nonfood items, utilities, services, and other expenditures for two
years. Column A reports the mean consumption expenditure and the mean
expenditure in each category. Georgia’s mean per capita expenditure in
2003 is GEL 126.1 [6,A], which changes marginally to GEL 126.0 in 2006
[12,A]. The mean per capita expenditure on food in 2003 is GEL 76.9 [1,A],
which decreases to GEL 72.8 in 2006 [7,A]. Mean expenditure on nonfood
also decreases over three years. However, mean expenditures for the other
three categories increase.
Column B reports the mean expenditure of each category as a percent-
age of overall per capita expenditure. The food category accounts for 61.0
percent of per capita expenditure in 2003 [1,B], which falls to 57.8 percent
in 2006 [7,B]. Per capita expenditure on foods is highly correlated with the
overall per capita expenditure—the correlation in 2003 is 80.8 [1,C] (the
upper bound and the lower bound of correlation is 0), which falls to 72.3 in
2006 [7,C], while the correlation between per capita expenditure on utili-
ties and the overall expenditure increases. Inequality of GE(2) for Georgia
increases from 24.2 [6,D] to 27.8 [12,D]. Inequality in per capita food con-
sumption expenditure does not change much, but inequalities in utilities
and expenditures on services drastically increase.
Finally, we look at the contribution of each component to over-
all inequality. As expected, the food category contributes the most to over-
all inequality. This category’s contribution is more than half of the overall
inequality. Its proportionate contribution, however, falls to 40.5 percent in
2006. The proportionate contribution of expenditure on services increases
from 19.6 percent in 2003 [4,F] to 30.2 percent in 2006 [10,F].
Table A.8 is helpful for understanding the source of inequality. This table
can identify components responsible for changes in inequality across two
time periods and the contributory factor to the overall inequality in a single
period of time.
Quantile Function
Figure A.1 graphs the quantile function of per capita expenditure for urban
Georgia. The vertical axis reports per capita expenditure, and the horizontal
286
Appendix
Figure A.1: The Quantile Functions of Urban Per Capita Expenditure, Georgia
600
480
360
Quantile
240
120
0
0 0.2 0.4 0.6 0.8 1.0
Expenditure percentile
2003 2006
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
axis reports percentiles. A quantile function reports the level below which
per capita expenditure falls for a given population percentage, when the
population is ranked by per capita expenditure. The solid line represents
the quantile function for 2003, and the dotted line corresponds to the urban
distribution of consumption expenditure for 2006. The horizontal lines are
poverty lines for 2003 and 2006.
If a distribution’s quantile function lies completely above that of
another distribution, then the situation is called first-order stochastic
dominance. When a distribution first-order stochastically dominates
another distribution, then every income standard reported ranks the former
distribution better than the latter distribution. If two quantile functions
cross each other, then a dominance relationship may not hold and rank-
ing distributions would depend on the particular per capita expenditure
standards used.
The curve with the solid line represents Georgia’s urban quantile func-
tion in 2003, and the quantile function with the dotted line corresponds
to Georgia in 2006. If a quantile function lies completely above another
287
A Unified Approach to Measuring Poverty and Inequality
quantile function, then every lower partial mean of the former distribution
is larger than the corresponding lower partial mean of the latter distribu-
tion. However, in the case of urban Georgia, the two quantile functions
cross each other, which prevents an unambiguous ranking. As evident from
the figure, the 90th percentile in 2006 is larger than the 90th percentile
in 2003, whereas the 40th percentile in 2006 is smaller than that in 2003.
Given that a quantile function is an inverse of the cumulative distribution
function, the example implies that first-order stochastic dominance does not
hold between these two time periods.
Figure A.2 graphs the generalized Lorenz curve of Georgia’s urban per capita
expenditure for 2003 and 2006. The vertical axis reports the cumulative
mean per capita expenditure and the horizontal axis reports the percentile
of per capita expenditure. A generalized Lorenz curve graphs the share of
mean per capita consumption expenditure spent by each percentile of the
150
Cumulative mean per capita
120
expenditure (lari)
90
60
30
0
0 0.2 0.4 0.6 0.8 1.0
Percentile of per capita expenditure
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
288
Appendix
population. The curve graphs the area under the quantile function up to
each percentile of population, or the height of the Lorenz curve times the
mean per capita expenditure. Thus, the height of the generalized Lorenz
curve is equal to the mean consumption expenditure when the percentile is
one. In other words, the share of the total consumption expenditure spent
by the entire population is 100 percent.
The curve with the solid line represents the generalized Lorenz curve
for urban Georgia in 2003. The generalized Lorenz curve with the dotted
line corresponds to urban Georgia in 2006. If a generalized Lorenz curve
lies completely above another generalized Lorenz curve, then every lower
partial mean of the former distribution is larger than the corresponding
lower partial mean of the latter distribution, and the former distribution
has a larger Sen mean than the latter distribution. Also, when one gener-
alized Lorenz curve lies above another, the distribution corresponding to
the former generalized Lorenz curve is said to second-order stochastically
dominate the distribution corresponding to the latter. In this particular
example, the distribution of per capita expenditure in 2003 second-
order stochastically dominates the distribution of per capita expenditure
in 2006.
Figure A.3 graphs the general mean curve of urban Georgia’s per capita
expenditure for two years. The vertical axis reports per capita expenditure,
and the horizontal axis reports parameter α, also known as a society’s degree
of aversion toward inequality. A general mean curve plots the value of
general means of a distribution corresponding to parameter α. The general
mean of a distribution tends toward the maximum and the minimum per
capita expenditures in the distribution when α tends to ∞ and – ∞, respec-
tively.
Given that the largest per capita expenditure in any distribution is usu-
ally several times larger than the minimum per capita expenditure, allowing
α to be very large would prevent any meaningful graphic analysis. So we
restrict α = 1 to be between –5 and 5, which we consider large enough.
The height of the curve at α = 1 denotes the arithmetic mean. Similarly,
the heights at α = 0, α = –1, and α = 2 denote the geometric mean, the
harmonic mean, and the Euclidean mean, respectively.
289
A Unified Approach to Measuring Poverty and Inequality
400
320
160
80
0
–5 –4 –3 –2 –1 0 1 2 3 4 5
Parameter α
2003 2006
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
The solid line represents urban Georgia’s general mean curve for 2003.
The general mean curve with the dotted line corresponds to urban Georgia
for 2006. If a general mean curve of a distribution lies completely above
the general mean curve of another distribution, then every general mean
of the former distribution is larger than the corresponding general mean of
the latter. Then, for example, the former distribution would have a higher
arithmetic mean, higher geometric mean, higher harmonic mean, and
higher Euclidean mean than the latter distribution. Note that the standard-
ized general mean curve can be obtained from the general mean curve by
dividing the curve throughout by the arithmetic mean.
Figure A.4 graphs the generalized Lorenz growth curve for Georgia’s per
capita expenditure. The vertical axis reports the annual growth rate of
290
Appendix
Figure A.4: Generalized Lorenz Growth Curve for Urban Per Capita
Expenditure, Georgia
0.2
means consumption expenditure (%)
Annual growth rate of lower partial
0.15
0.1
0.05
–0.05
0 0.2 0.4 0.6 0.8 1.0
Percentile of per capita expenditure
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
the lower partial mean consumption expenditures and the horizontal axis
reports the cumulative population share. A generalized Lorenz growth curve
graphs the growth of lower partial mean per capita consumption expendi-
ture for each population percentile. Thus, a generalized Lorenz growth curve
indicates how every lower partial mean is changing over time.
Figure A.5 graphs the general mean growth curve for Georgia’s per capita
expenditure. The vertical axis reports the annual growth rate of the general
mean consumption expenditures and the horizontal axis reports parameter
α, also known as a society’s degree of aversion toward inequality. A general
mean growth curve graphs the growth of different general means and thus
indicates how the general means are changing over time. The growth rate
in mean per capita expenditure is the same as the growth rate of general
mean at α = 1.
291
A Unified Approach to Measuring Poverty and Inequality
Figure A.5: General Mean Growth Curve of Urban Per Capita Expenditure,
Georgia
0.4
–0.2
–5 –4 –3 –2 –1 0 1 2 3 4 5
Parameter α
Source: Based on ADePT Poverty and Inequality modules using Integrated Household Survey of
Georgia 2003 and 2006.
References
292
Index
293
Index
294
Index
295
Index
296
Index
297
Index
298
Index
299
Index
300
Index
301
Index
302
Index
303
Index
304
ECO-AUDIT
Environmental Benefits Statement
The World Bank is committed to preserving Saved:
endangered forests and natural resources. The • 13 trees
Office of the Publisher has chosen to print A • 6 million BTUs
Unified Approach to Measuring Poverty and of total energy
Inequality: Theory and Practice on recycled • 1,111 pounds of net
paper with 50 percent postconsumer fiber in greenhouse gases
accordance with the recommended stand- • 6,030 gallons of waste
ards for paper usage set by the Green Press water
Initiative, a nonprofit program supporting
• 404 pounds of solid
publishers in using fiber that is not sourced
waste
from endangered forests. For more informa-
tion, visit www.greenpressinitiative.org.