bennett hatzimasoura - poverty measurement with ordinal data
TRANSCRIPT
Poverty Measurement with Ordinal Data
Christopher J. Bennetta and Chrysanthi Hatzimasourab
First Version: September 2011; Revised: November 2012
Abstract
The Foster, Greer, Thorbecke (1984) class nests several of the most widely used indicesin theoretical and empirical work on economic poverty. Use of this general class of indices,however, presupposes a dimension of well-being that, like income, is cardinally measurable.Responding to recent interest in dimensions of well-being where achievements are recordedon an ordinal scale, this paper introduces a general methodology for constructing ordinalindices of poverty and, in particular, shows how this methodology may be applied toconstruct an ordinal analogue of the popular FGT class of indices. The resulting ordinalFGT indices retain the simplicity of the classical FGT indices and also many of theirdesirable features, including additive decomposability. To illustrate their use, we applythe ordinal FGT indices to self-reported data on health status in Canada and the UnitedStates.
JEL classification: I3, I32, D63, O1
Keywords: poverty measurement, ordinal data, FGT poverty indices
a(Corresponding author) Department of Economics, Vanderbilt University, VU Station B #351819,
2301 Vanderbilt Place, Nashville, TN 37235-1819, U.S.A. E-mail: [email protected] of Economics, George Washington University, 2115 G Street, NW Monroe Hall
340, Washington, DC 20052, U.S.A. E-mail: [email protected]∗We are deeply indebted to James Foster and John Weymark for their advice and supportthroughout our work on this project. We also wish to thank Michael Hoy and BuhongZheng for their insightful comments and suggestions. Finally, the second author grate-fully acknowledges the Institute for International Economic Policy of the Elliott Schoolof International Affairs for funding her visit to Vanderbilt University in order to conductthis research.
1 Introduction
In the twenty-five years since it was first introduced, the FGT (Foster, Greer, and
Thorbecke 1984) family of measures has become the most widely used class in empir-
ical work on the measurement of poverty. The attractiveness of the FGT measures stems
largely from their simple structure, their ease of interpretation, and their sound axiomatic
properties. Being defined by two parameters, namely the poverty line z and a scalar mea-
sure of poverty aversion α, each member of the FGT class is easily computed as an average
of the power function defined by α whose argument is the normalized income shortfall
from z. Specific members include the well-known poverty gap, squared poverty gap, and
headcount ratio (i.e., the proportion of the population identified as poor).
Use of the general FGT class of measures presupposes a dimension of well-being
that, like income, is cardinally measurable. Recently, however, considerable interest has
emerged in measures of aggregate deprivation in dimensions of well-being other than in-
come and, in particular, in dimensions of well-being—for example, health, education,
empowerment, and social inclusion—that are often recorded on an ordinal scale.1 Conse-
quently, “a crucial emerging issue is how to measure poverty when data do not have the
characteristics of income, which is typically taken to be cardinal and comparable across
persons ... Must we retreat to the headcount ratio [with ordinal data], or can we continue
to evaluate the depth or distribution of deprivations—key benefits provided by the higher
order FGT measures when the variable is cardinal?”(Foster, Greer, and Thorbecke 2010,
p. 516)
In order to address this issue, this paper introduces a methodology for constructing
ordinal poverty indices from cumulative distributions over the levels of achievement of the
poor. This general approach to the construction of ordinal poverty indices is motivated by
a thought experiment in which an individual is completely unaware of her relative position
in society and draws a level of achievement at random according to the actual distribution
in society and another level of achievement from a reference lottery over the poor states.
The extent of poverty in society is then recorded as the proportion of individuals who
would accept their realized level of poverty drawn from the reference lottery rather than
their draw from the actual distribution in society.
Poverty indices constructed in this manner are completely determined by the cumu-
1Problems surrounding the measurement of poverty with ordinal data are raised, for example, inFoster, Greer, and Thorbecke (2010) and Alkire and Foster (2011a,b). Allison and Foster (2004) werethe first to stress the problems raised by ordinal data in the related context of inequality measurement.See Zheng (2008), Abul Naga and Yalcin (2004), and Madden (2010) for more on the use of ordinal datain this latter context.
1
lative distribution associated with the reference lottery, thereby enabling entire classes
of indices to be constructed from parametric classes of distributions. In this sense, the
methodology is related to the Atkinson-Kolm-Sen (Atkinson 1970, Kolm 1969, Sen 1973)
methodology, where the specification of a social welfare function (or parametric class of
welfare functions) completely determines the inequality index (or class of indices). The
AKS methodology can also be motivated by a thought experiment, albeit one that asks
what percentage of total income can be discarded without affecting social welfare if income
is equally distributed.
A distinguishing feature shared by all poverty indices constructed from reference lot-
teries is that they are invariant to ordering-preserving transformations applied to the nu-
merical values representing the various levels of achievement. Consequently, a “retreat”
to the headcount ratio with ordinal data is entirely unnecessary since poverty indices con-
structed from reference lotteries not only include the headcount ratio as a special case,
but they can also be made sensitive to the ‘depth’ and ‘distribution’ of poverty that the
headcount ratio ignores.
As a concrete example, we apply our methodology to construct an analogue of the
FGT class of measures for use with ordinal data. In particular, we show that a simple
parametric class of distributions gives a counterpart of the classical FGT class of measures
that retains many of the attractive properties of the classical FGT measures (including,
for example, additive decomposability) and is without the obvious shortcomings inherent
in the application of conventional poverty measures to ordinal data. Furthermore, we
provide an axiomatization of the ordering induced by our ordinal analogue of the FGT
class of measures. This axiomatization is an ordinal counterpart to the axiomatization of
the classical FGT orderings developed by Ebert and Moyes (2002).
In the next section, we outline the construction of ordinal poverty indices using the
concept of a reference lottery, document the basic properties of poverty indices constructed
from reference lotteries, and introduce the parametric class of reference lotteries that
generate the ordinal analogue of the FGT class of indices. In Section 3, we present an
axiomatic characterization of the poverty orderings induced by the ordinal analogues of the
FGT class of indices. Then, in Section 4, we illustrate the application of the ordinal FGT
indices to self-reported health data from the United States and Canada. When applied
to this dataset, these indices suggest that there is unambiguously greater ill-health in the
United States than in Canada for the bottom 20% of their income distributions. Finally,
in Section 5, we present some concluding remarks.
2
2 Measuring Poverty with Ordinal Data
With ordinal data, there are K ordered categories or states of achievement represented
numerically by an ordered set Y = {y1, y2, . . . , yK} in such a way that yi > yj if, and only
if, state i is preferred to state j.2 The observed levels of achievement in a population of size
N are recorded in y ∈ YN and individuals within this population are identified as “poor”
if they fall into one of the k worst states, or equivalently if their level of achievement falls
at or below yk, where yk < yK .
2.1 Poverty Measures as Evaluations of Achievement Lotteries
In order to construct a meaningful measure of poverty for use with data recorded on
an ordinal scale, we consider a thought experiment in which one has the opportunity to
accept a realized level of achievement from the equiprobable lottery Y on y or to decline
this allocation in favor of an alternative allocation drawn independently from a reference
lottery Uα (indexed by α) over the k states of poverty y1, y2, . . . , yk.
When comparing the allocations from these two lotteries, one is certainly better off
accepting one’s realization of the equiprobable draw Y whenever it is above yk and, hence,
out of poverty. Conversely, one will choose to accept the state of poverty generated by
the reference lottery Uα whenever the realization of Y amounts to an even worse state of
poverty.
Ex ante, the probability that one will accept the state of poverty generated by the
reference lottery Uα is equal to the probability that the realization of Y is no larger than
the the realization of Uα, which is given by
πα(y, yk) = P[Y ≤ Uα]. (2.1)
The quantity πα(y, yk) thus tells us the probability that one will be better off facing the
lottery Uα rather than facing an equiprobable draw from the actual distribution in society.
In interpreting the right-hand side of (2.1), note that the statistical independence of Y
and Uα gives
P[Y ≤ Uα] = EY [EUα1(Y ≤ Uα)]
=1
N
N∑i=1
P(yi ≤ Uα),(2.2)
2The number of states can be countably infinite. Our focus on the case where the number of states isfinite is, however, without loss of generality.
3
where 1(·) is the indicator function and EY , for example, denotes the mathematical
expectation with respect to the probability distribution of Y . Because the first term in
the sum is the probability that Uα is no smaller than y1, the second term is the probability
that Uα is no smaller than y2, and the N th term is the probability that Uα is no smaller
than yN , we see that πα(y, yk) is merely the average of each individual’s probability
of receiving a higher level of achievement from the lottery Uα. The quantity πα(y, yk),
therefore, may also be interpreted as the proportion of individuals, each of whom in turn
faces the choice between their own realizations from the pair of lotteries Y and Uα, that
would accept their level of achievement generated from the reference lottery Uα.
Clearly, the quantity πα(y, yk) will be equal to zero for any distribution y in which
no individual in y is identified as poor. Indeed, when there are no poor individuals, the
realization of Y must be above yk and, hence, must be above any possible realization
of Uα. Conversely, πα(y, yk) will tend towards one as individual levels of achievement
fall towards the least desirable state y1. Consequently, we may regard the magnitude of
πα(y, yk) as an indicator of the extent of poverty in y (relative to Uα).
As a concrete example, consider the special case in which the reference lottery Uα is
degenerate at yk so that it yields the least deprived level of poverty yk with probability
one. In this case,
πα(y, yk) = P[Y ≤ Uα]
= P[Y ≤ yk],(2.3)
so that πα(y, yk) records the proportion of individuals who would prefer the guaranteed
state of poverty yk to their draw from the prevailing distribution of achievements y. As
a second example, consider the reference lottery Uα that assigns equal probability to the
states of the poor y1, y2, . . . , yk. In this case, πα(y, yk) records the proportion of individuals
that would prefer their random draw from the states of the poor rather than their realized
allocation drawn at random from y.
In general, the formulation in (2.1) provides us with a framework that is particularly
well suited for constructing meaningful poverty indices when the data are ordinal. This
is because the generic index πα(y, yk) (a) has both a simple and appealing interpretation
and (b) it is, by construction, invariant to order preserving transformations of the levels,
which is essential for any measure applied to ordinal data.3
Different reference lotteries over the states of the poor obviously produce different
3Indeed, P[Y ≤ Uα] = P[g(Y ) ≤ g(Uα)] for all strictly positive monotonic transformations g : R → R.
4
poverty indices, but it is not yet clear how the choice of reference lottery ultimately
shapes the index. The following proposition helps to shed light on this issue. Indeed, it
shows precisely how the reference lottery affects the properties of the resulting poverty
index. In our statement of the proposition we denote the collection of poor individuals
by q(y, yk) = {1 ≤ i ≤ N : yi ≤ yk}.
Proposition 2.1. For any reference lottery Uα over the states of the poor y1, y2, . . . , yk,
πα(y, yk) =1
N
∑i∈q(y,yk)
P[yi ≤ Uα]. (2.4)
Proof. It follows from (2.2) that
P[Y ≤ Uα] =1
N
N∑i=1
EUα1(yi ≤ Uα)1(yi ≤ yk) +1
N
N∑i=1
EUα1(yi ≤ Uα)1(yi > yk). (2.5)
Note that 1(yi ≤ Uα)1(yi > yk) = 0 because the support of Uα is restricted to the poor
states (i.e., y1, . . . , yk). Consequently, the second term in the last line of (2.5) is zero.
Hence, the desired result follows from the equivalence
1
N
N∑i=1
EUα1(yi ≤ Uα)1(yi ≤ yk) =1
N
∑i∈q(y,yk)
P[yi ≤ Uα].
The function πiα ≡ P[yi ≤ Uα] in (2.4) is the individual poverty function of the
ith individual. Proposition 2.1 shows that the poverty measure πα(y, yk) is always de-
composable (Foster and Shorrocks 1991, p. 691), with only the poverty functions πiα,
i = 1, . . . , N , influenced by the specification of the reference lottery. Consequently, the
cumulative distribution associated with the reference lottery determines the specification
of the individual poverty functions and, hence, ultimately determines if and how ‘depth’
and ‘distribution’ are accounted for by the aggregate measure.
In short, the proposed methodology amounts to constructing ordinal poverty indices
from cumulative distributions. This approach gives rise to a rather large set of choices, not
unlike the Atkinson-Kolm-Sen methodology that constructs inequality indices from social
welfare functions. In the next subsection, we examine a class of distributions that give rise
to a particularly simple and appealing class of indices that are ordinal analogues of the
classical FGT class. This new class of indices inherits many of the attractive properties
of the classical FGT class, including its simple structure and sound axiomatic properties
5
(Foster, Greer, and Thorbecke 2010). We also develop an axiomatic characterization of
the poverty ordering induced by this class of ordinal indices in Section 3.
2.2 A Parametric Class of Reference Lotteries
In this section we examine the parametric class Uα, α ≥ 0, of reference lotteries whose
corresponding probability distributions are given by
P[Uα ≥ yj] =
(k − j + 1
k
)α
, 1 ≤ j ≤ k, α > 0. (2.6)
When α = 0, the lottery Uα guarantees the least deprived state of poverty yk. Conse-
quently, this lottery when evaluated in (2.1) gives rise to the poverty index (2.3), which
is nothing other than the classical headcount ratio. When α = 1, the lottery is equally
weighted over the poor states. Hence, the index πα(y, yk) records the proportion of people
in society that would prefer an equiprobable draw from the states of the poor rather than
an allocation drawn at random from y. When α is chosen to be greater than 1, increased
probability is placed on the poorest of the poor states. The corresponding indices, there-
fore, become relatively more sensitive to the ‘depth’ of poverty experienced by individuals
in the population. In the limit, as α tends to ∞, the lottery is degenerate at y1, implying
that the corresponding poverty index will be sensitive only to changes in the proportion
of individuals experiencing the worst state of poverty.
It follows from Proposition 2.1 that substitution of this parametric class of lotteries
into (2.1) gives the class of ordinal poverty indices
πα(y, yk) =k∑
j=1
pj
(k − j + 1
k
)α
, α > 0, (2.7)
where pj is the proportion of the population y in the jth state. The class of indices
generated by (2.6) is an ordinal analogue to the classical FGT indices in that members
of this class are also given by average power functions of normalized gaps, albeit with
normalized gaps in levels replaced by normalized gaps in ranks.
To further elucidate this close connection to the classical FGT indices, let GY de-
note the cumulative distribution function that assigns equal probability to the potential
achievement levels in Y. The cumulative distribution GY(·) is a convenient mathematical
device that maps a given level of achievement yi ∈ Y to its corresponding (normalized)
achievement rank GY(yi) ∈ { 1K, 2K, . . . , 1}. Thus, for example, GY(yK) = 1 is the highest
6
achievement rank and GY(yj) = j/K is the achievement rank of an individual in the
jth state of achievement. With the distribution of (normalized) achievement ranks and
poverty rank cut-off computed as
x ≡(GY(y1), GY(y2), . . . , GY(yN)) ∈ [0, 1]N
and
z ≡ GY(yk+1) =k + 1
K,
respectively, the indices πα(y, yk), α > 0, which operate on the levels, are equivalent to
the indices
Π̃α(x; z) =1
N
N∑j=1
(z − xj
z −GY(y1)
)α
1(xj < z), α > 0, (2.8)
which operate on the (normalized) ranks.4 Furthermore, the expression in (2.8) is ordinally
equivalent to
Πα(x; z) =1
N
N∑j=1
(z − xj
z
)α
1(xj < z), α > 0. (2.9)
The alternative representation of πα in (2.9), which is formulated in terms of (nor-
malized) ranks, is identical to the computational formula for the classical FGT class of
indices. We exploit this alternative representation in the next section, where we provide
an axiomatic characterization of the ordinal FGT class of indices.
3 An Axiomatic Characterization of the Poverty Or-
dering induced by Πα
This section supplements our earlier construction of the ordinal FGT indices with an
axiomatic characterization of their induced poverty orderings. Our approach mirrors
Ebert and Moyes’s (2002) characterization of the poverty orderings induced by the classical
FGT class, albeit with their axioms suitably translated for when the data are ordinal.
Let <z (indexed by z) denote a complete, reflexive, and transitive binary relation
on the set of all possible distributions of (normalized) achievement ranks [0, 1]N . The
statement x <z y is interpreted as saying that x exhibits at least as much poverty as y.
4Note that we adopt the weak definition of the poor (Donaldson and Weymark 1986) here under whichthe poor consists of all individuals with endowments less than z = k+1
K . The presence of the term GY(y1)is due to the discreteness of the possible levels of achievement and would not appear with a continuumof levels as in the classical FGT class.
7
The corresponding asymmetric and symmetric factors of <z are ≻z and ∼z, respectively.
We define <z on all of [0, 1]N rather than on { 1K, 2K, . . . , K−1
K, 1}N for simplicity. Such
idealizations are standard in axiomatic analyzes with discrete variables, e.g., in consumer
theory where it is typically assumed that commodities are perfectly divisible even when
they are not.
The first two axioms are suitably reformulated statements of standard properties.
CONTinuity: <z is continuous on [0, 1]N .
FOCus: Let x ∈ [0, 1]N and suppose that xi ≥ z. Then,
x ∼z (x1, . . . ,xi−1,xi + c,xi+1, . . . ,xN)
for all constants c such that z < xi + c ≤ 1.
The focus axiom states that only the ranks of poor individuals play a role in deter-
mining the ordering of two distributions. Our next axiom is a separability axiom.
INDependence: Let x1,x2 ∈ [0, 1]N satisfy x1 ∼z x2 with x1
i = x2i for some 1 ≤ i ≤ N .
Then, for every γ ∈ [0, 1],
(x11, . . . ,x
1i−1, γ,x
1i+1 . . . ,x
1N) ∼z (x
21, . . . ,x
2i−1, γ,x
2i+1 . . . ,x
2N).
The independence axiom implies that the poverty ordering of achievement ranks for
any subgroup of individuals can be derived without reference to the ranks in which the
rest of the the population find themselves.
SYMMetry: For all x = (x1, . . . ,xN) ∈ [0, 1]N and any permutation π of {1, 2, . . . , N},(x1, . . . ,xN) ∼z (xπ(1), . . . ,xπ(N)).
MONotonicity: Let x ∈ [0, 1]N be such that xi < z ≤ 1. Then,
x ≻z (x1, . . . ,xi−1,xi + c,xi+1, . . . ,xN)
for all constants c > 0 satisfying xi + c ≤ 1.
Symmetry says that individual identities play no role in determining the intensity
of poverty, whereas Monotonicity says that an increase in a poor person’s rank should
decrease the overall poverty level; see, e.g., Zheng (1997). Lastly, we impose two invariance
8
axioms.
SCALE Invariance: For all x1,x2 ∈ [0, 1]N and all 0 < λ ≤ 1, x1 ∼z x2 implies
λx1 ∼λz λx2.
TRANSlation Invariance: For all x1,x2 ∈ [0, 1]N and all γ ∈ R such that x1 +
γ1N ,x2 + γ1N ∈ [0, 1]N and z + γ ∈ [0, 1],
x1 ∼z x2 implies x1 + γ1N ∼z+γ x2 + γ1N ,
where 1N is an n× 1 vector of ones.
The axioms stated above are ordinal analogues of the axioms for cardinally measurable
attributes used in Ebert and Moyes (2002). Taken together, they impose sufficient struc-
ture to characterize the representation of <z. Specifically, one can establish the following
result:
Proposition 3.1. The poverty ordering <z satisfies CONT, FOC, MON, IND, SYMM,
SCALE, and TRANS if and only if it is represented by
Πα(x; z) =1
N
N∑i=1
a(z)(z − xi
)α1(xi < z), for all x ∈ [0, 1]N and all α > 0. (3.1)
Proof. The proof is identical to the proof of Theorem 1 in Ebert and Moyes (2002) after
substituting levels for ranks.
As in the case of the classical FGT index, one can easily verify that α > 0 in Propo-
sition 3.1 above must be replaced by α > 1 if we impose the additional requirement
that <z satisfy the following transfer axiom, which states that the overall poverty level
should decrease when a poor individual’s shift upwards in rank if offset by a less deprived
individual’s equal downward shift in rank:
Transfer Suppose that 0 < xi < xj < z ≤ 1. Then,
x ≻z (x1, . . . ,xi−1,xi − c,xi+1, . . . ,xj−1,xj + c,xj+1, . . . ,xN)
for all c > 0 satisfying xi − c ≥ 0 and xj + c ≤ 1.
In summary, the ordinal FGT index is sensitive to ‘depth’ for α > 0, and sensitive to
both ‘depth’ and the ‘distribution’ of achievement ranks for α > 1.
9
4 Empirical Illustration
We now illustrate the ordinal FGT indices using self-reported health statuses in Canada
and the United States from the Joint Canada/United States Survey of Health (JCUSH).
In these surveys, approximately 3,500 Canadian and 5,200 U.S. residents rated their
individual health as either poor, fair, good, very good, or excellent.5 Due to the complex
sampling design and over-sampling of certain populations, sampling weights have been
appended to the survey data by the Centers for Disease Control and Prevention and
Statistics Canada to render the samples representative of their respective populations.
We use these sampling weights in our subsequent analysis.
We apply the ordinal FGT indices to examine health deprivation or health poverty, as
well as to examine health poverty when the population is decomposed by income quintiles.
We begin by considering the headcount ratios (α = 0) in each country and at various cut-
offs.6 As can be seen in Table 1, more U.S. residents as a proportion of the population
report their health as being less than or equal to poor, fair, or good, than is the case in
Canada. On the other hand, Canadians are less likely than U.S residents to rate their
health status as excellent rather than very good.
For α = 1, the ordinal FGT indices suggest that health status in the U.S. is worse
than in Canada for every cutoff.7 Perhaps more interestingly, the decomposition by
income quintiles demonstrates that the greatest contribution to the disparity between
the two countries occurs at the lowest income quintile. In other words, the disparity
in health statuses between the two countries is greatest at the bottom income quintile
where the self-reported health statuses of income poor U.S. residents are being compared
to self-reported health statuses of income poor Canadians. The α = 1 case provides
us with more insight into the distribution of the poor than the headcount ratios do by
themselves. Such insight may be helpful to policymakers when designing and targeting
their health-care policies.
These data can also be used to illustrate the simple interpretation of the ordinal FGT
indices provided above. For example, if we focus on the first income quintile and a cutoff
of 2, we observe that the FGT indices when α = 1 are 0.165 in the U.S. and 0.102 in
5The survey participants were asked “In general, would you say your health is: poor, fair, good, verygood, or excellent?” See Allison and Foster (2004), and references therein, for a discussion and a literaturereview of the role of self reported health data as a predictor of mortality and overall health.
6A similar analysis using the headcount ratio to look at poverty with self-reported health data wasperformed in Allison and Foster (2004).
7The ordinal FGT index is unchanged at the first cut-off y1 because πα(y, y1) is independent of αsince πα(y, y1) = P[Y ≤ y1].
10
Table 1: Health Poverty Estimates for Canada and the United States.
Index Country Income Quintile Cut-Off1 2 3 4 5
Headcount, α = 0
USA
All 0.037 0.136 0.398 0.732 1.0001 0.087 0.243 0.532 0.790 1.0002 0.052 0.202 0.497 0.783 1.0003 0.014 0.096 0.398 0.750 1.0004 0.012 0.057 0.288 0.684 1.0005 0.008 0.056 0.247 0.644 1.000
CAN
All 0.032 0.111 0.384 0.757 1.0001 0.051 0.154 0.495 0.820 1.0002 0.054 0.170 0.467 0.810 1.0003 0.032 0.104 0.362 0.756 1.0004 0.013 0.084 0.337 0.746 1.0005 0.005 0.042 0.249 0.651 1.000
Ordinal FGT α = 1
USA
All 0.037 0.087 0.191 0.326 0.4611 0.087 0.165 0.287 0.413 0.5302 0.052 0.127 0.251 0.384 0.5073 0.014 0.055 0.170 0.315 0.4524 0.012 0.034 0.119 0.260 0.4095 0.008 0.032 0.104 0.238 0.391
CAN
All 0.032 0.072 0.176 0.321 0.4561 0.051 0.102 0.233 0.380 0.5042 0.054 0.112 0.230 0.375 0.5003 0.032 0.068 0.166 0.314 0.4514 0.013 0.049 0.145 0.295 0.4365 0.005 0.023 0.098 0.236 0.389
Notes: Headcount (α = 0) and Ordinal FGT measure (α = 1) decomposed by incomequintiles. Estimates based on 2,960 and 3,815 Canadian and U.S. respondents from the2003 Joint Canada/United States Survey of Health.
Canada. Consequently, we have that 165 out of every 1,000 U.S. residents would prefer
an equiprobable lottery from the two lowest states of health rather than draw their health
status from the actual distribution of health in society. In contrast, only 102 out of every
1,000 Canadian residents would prefer the equiprobable two-state lottery over the random
draw from the societal distribution in Canada.
5 Concluding Remarks
This paper has developed a methodology for constructing poverty indices from cumulative
distributions over the states of the poor and has applied this methodology to construct an
ordinal analogue of the classical FGT class of poverty indices. This new class of ordinal
indices retains many of the attractive properties of the classical FGT class and yet is
without the obvious shortcomings inherent in the application of conventional poverty
measures to ordinal data.
11
References
Abul Naga, R. H., and T. Yalcin (2004): “Inequality Measurement for Ordered
Response Health Data,” Journal of Health Economics, 23, 1614–1625.
Allison, R. A., and J. E. Foster (2004): “Measuring Health Inequality using Qual-
itative Data,” Journal of Health Economics, 27, 505–524.
Atkinson, A. B. (1970): “On the Measurement of Inequality,” Journal of Economic
Theory, 2(3), 244–263.
Donaldson, D., and J. A. Weymark (1986): “Properties of Fixed-Population Poverty
Indices,” International Economic Review, 27, 667–688.
Ebert, U., and P. Moyes (2002): “A Simple Axiomatization of the Foster, Greer and
Thorbecke Poverty Orderings,” Journal of Public Economic Theory, 4(4), 455–473.
Foster, J., J. Greer, and E. Thorbecke (1984): “A Class of Decomposable Poverty
Measures,” Econometrica, 52, 761–766.
(2010): “The Foster-Greer-Thorbecke (FGT) Poverty Measures: 25 Years Later,”
Journal of Economic Inequality, 8, 491–524.
Foster, J. E., and A. F. Shorrocks (1991): “Subgroup Consistent Poverty Indices,”
Econometrica, 59(3), pp. 687–709.
Kolm, S. C. (1969): “The Optimal Production of Social Justice,” Public Economics:
An Analysis of Public Production and Consumption and their Relations to the Private
Sectors, pp. 145–200. Macmillan, London.
Madden, D. (2010): “Ordinal and Cardinal Measures of Health Inequality: An Empirical
Comparison,” Health Economics, 19, 243–250.
Sen, A. (1973): On Economic Inequality. Clarendon Press, Oxford.
Zheng, B. (1997): “Aggregate Poverty Measures,” Journal of Economic Surveys, 11,
123–162.
(2008): “Measuring Inequality with Ordinal Data: A Note,” Inequality and Op-
portunity: Papers from the Second ECINEQ Society Meeting (Research on Economic
Inequaltiy, Volume 16), pp. 177–188. Emerald Group Publishing Limited.
12
Technical Appendix (Not for publication)
Lemma 5.1. A poverty ordering <z on [0, 1]N satisfies the axioms CONT, FOC, IND,
SYMM, and MON if, and only if, it is represented by
Π(x, z) =∑xi<z
π(xi, z) +∑xi≥z
π(z, z) (5.1)
for some π : [0, 1]2 → R where π(x, z) is continuous and strictly increasing in x for all
x ∈ (0, 1).
Proof. The “if” part is obvious.
IND, SYMM, and FOC imply that the poverty ordering ≽z can be represented by an
additive function Π(x, z) of the form
Π(x, z) =∑xi<z
π(xi, z) +∑xi≥z
π(z, z)
The remaining axioms impose structure on the function π: CONT implies that π is
continuous; and MON implies that π(·, z) is strictly increasing.
The representation of <z in (5.1) allows us to associate each distribution x ∈ [0, 1]N
with an equivalent societal achievement (Ebert and Moyes, p. 462). That is, for a given
distribution of achievement percentiles x ∈ [0, 1]N and a given poverty threshold z ∈ [0, 1],
we define the equivalent societal achievement as
e(x, z) =
π−1(N−1Π(x, z), z) if xi < z for some i ∈ N
z if xi ≥ z for all i ∈ N(5.2)
Lemma 5.2. Suppose that <z satisfies the axioms CONT, FOC, MON, IND, SYMM,
SCALE, and TRANS. Then, for any given distribution of achievement x ∈ [0, 1]N and
any given poverty threshold z ∈ [0, 1],
i. x ∼z (e(x, z), e(x, z), . . . , e(x, z))
ii. e(λx, λz) = λe(x, z) for all λ ∈ (0, 1]
iii. e(x+γ1N , z+γ) = e(x, z)+γ for all γ such that x+γ1N ∈ [0, 1]N and γ+z ∈ (0, 1].
Proof. (i) Obvious from the definition; (ii) By definition,
e(x, z)1N ∼z x, (5.3)
13
and e(λx, λz)1N ∼λz λx. Because applying SCALE to (5.3) gives λe(x, z)1N ∼z λx, it
follows from the transitivity of ∼λz that
λe(x, z)1N ∼λz e(λx, λz)1n
The equality λe(x, z) = e(λx, λz) is then an immediate consequence of MON. (iii) Anal-
ogous to (ii).
Proof of Proposition 3.1. The “if” part is obvious.
To prove “only if”, let x ∈ [0, 1]N and z be given, and choose ϵ > 0 such that
0 ≤ z − ϵ ≤ xi ≤ z for all 1 ≤ i ≤ N . Parts (ii.) and (iii.) of Lemma 5.2 give
1
λπ−1
(1
N
N∑i=1
π(λxi, λz), λz
)= π−1
(1
N
N∑i=1
π(xi, z), z
)(5.4)
and
π−1
(1
N
N∑i=1
π(xi + γ, z + γ), z + γ
)− γ = π−1
(1
N
N∑i=1
π(xi, z), z
)(5.5)
for all λ ∈ (0, 1] and all γ such that x+ γ1N ∈ [0, 1]N and γ + z ∈ (0, 1].
Ebert and Moyes (2002) show that the only well-defined solution to the pair of func-
tional equations (5.4) and (5.5), which is also strictly increasing in x, is of the form
π(x, z) = a(z)(z − x)α(z) + b(z) (5.6)
with α(z) > 0, a(z) > 0, and b(z) ∈ R. Furthermore, it is demonstrated in Ebert
and Moyes (2002, p.471) that Part (ii.) of Lemma 5.2 implies that α(z) is functionally
independent of z so that α(z) = α and, hence,
π(x, z) = a(z)(z − x)α + b(z), α > 0, a(z) > 0, b(z) ∈ R (5.7)
The desired result, therefore, is obtained upon setting a(z) =(
1z
)α
and b(z) = 0.
14