Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Review of Development Economics, 15(2), 199–211, 2011

DOI:10.1111/j.1467-9361.2011.00602.x

Equity in Poverty Program Evaluations rode_602 199..211

Rocio Garcia-Diaz*

Abstract
Poverty-reducing programs often use indicators other than household income to transfer resources to the
poor which has important implications in terms of equity of the program. This paper offers a decomposition
methodology to analyze the horizontal and vertical equity components of a redistribution policy for sub-
groups defined according to policy-relevant characteristics. Our measures are derived by comparing the cost
of inequality in terms of equally distributed equivalent poverty gaps in the pre-transfer income and post-
transfer income distributions. An application to the poverty-reducing program Oportunidades in Mexico in
2006 reveals that the use of indicator variables more strongly correlated with pre-transfer incomes in rural
households can improve the redistributive effect while it also points out the difficulties of the program in
targeting different levels of support among the poor, both within and between groups.

1. Introduction
In several developing countries, social policy directed towards poverty relief consists
(partly) of transfers programs, in which poor households receive direct money trans-
fers. The main difficulty faced by the government implementing such programs is the
unobservability of income, which is supposed to be the basis for eligibility for support.
The use of other indicators than household’s incomes (such as geographical location,
family composition, etc.) to target support may be well justified in terms of the attain-
ment of some objectives of the program, most notably the reduction of poverty
estimates or purely in terms of administrative convenience. But one likely consequence
is the violation of the principle of horizontal equity, which states that equals should be
treated equally. In particular, horizontal inequities may arise from systematic differ-
ences in the level of support received by households with identical levels of pre-transfer
income. Moreover, even after controlling for levels of pre-transfer income in different
household types, inequities may arise from the heterogeneity of the households. On the
other hand, vertical equity requires that differences in people’s circumstances be
appropriately taken into account in both the formulation and the implementation of
public policy. That is, vertical equity accounts for differences between people at differ-
ent pre-transfers income levels. Concerns for vertical equity also raise questions about
the desirability of targeting public resources at groups with the highest needs.
The body of literature concerned with targeting is continuously growing, but in the
past little emphasis has been placed on the equity aspects of poverty reduction policies.
Keen (1992) pointed out that it may be that optimal targeting requires that an increase
in needs in some groups be met by the reduction of resources allocated to them. This is
because needy households are expensive to help, which results in targeting resources in
favor of groups whose poverty is relatively cheap to alleviate. Hence, optimal targeting
may reduce the vertical equity of the program if a substantial part of the support goes

* Garcia-Diaz: Tecnológico de Monterrey, Monterrey, N.L., 64849, Mexico. Tel: +52 (81) 8358-2000 ext. 4305;
Fax: ext. 4351; E-mail: rociogarcia@itesm.mx. I am most grateful to one anonymous referee for all his valuable
comments and suggestions.

© 2011 Blackwell Publishing Ltd


200 Rocio Garcia-Diaz

to the least poor households.1 On the other hand, exclusion errors in targeting schemes
that use other indicators than income arise when eligible poor or poor families are not
awarded with a transfer. This in consequence, will lead to horizontal inefficiency of the
policy, since these errors discriminate among the poor (Bibi and Duclos, 2007).2 The
main contribution of this paper is to provide a characterization of the overall redistri-
bution effect of a poverty reduction policy among the poor that yields consistent
measures of vertical equity (VE) and of classical horizontal inequity (HI) of the policy.
As one referee pointed out, the two concepts are tightly related, a VE measure tells us
how far the society has gotten towards poverty reduction for the most needed, and we
compare this measure with HI, the magnitude of exclusion failures that are common in
policy design of poverty-reducing programs. We also illustrate how judgments about
the importance of some policy-relevant characteristics can contribute to inequity in the
redistribution by decomposing measures of vertical equity and horizontal inequity into
relevant component distributions.
The seminal article on the redistributive decomposition in terms of classical hori-
zontal inequity, vertical inequity and re-ranking effects is the one by Aronson and
Johnson (1994), in which they decompose a Gini index of redistributive effect in terms
of vertical inequity, horizontal inequity, and re-ranking components. Lambert and
Ramos (1997) found that the re-ranking term in this decomposition is a consequence of
using a nondecomposable inequality measure. Duclos and Lambert (2000) proposed an
alternative measure of horizontal inequity by capturing the dispersion of post-transfer
income among pre-transfer equals using the cost of inequality approach. The cost of
inequality approach allows measuring the redistributive gain that would come from
eliminating horizontal inequity with no loss of social welfare. A particular interest of
this approach is that it permits measuring the local cost of horizontal inequity in terms
of poverty gaps as proposed by Bibi and Duclos (2007). They decompose the policy
effectiveness of a program in terms of horizontal inequity, vertical equity, and a leakage
component. The leakage component refers to the transfers that raise the pre-policy
poor above the poverty line and are thus classified as inefficient, as referred to in
Atkinson (1993) and Creedy (1996).
The approach proposed here measures the overall redistributive effect of a poverty
reduction policy that results from change in the cost of inequality in terms of poverty
gaps. The decomposition yields horizontal inequity and vertical equity contributions to
the total redistributive effect. The chosen poverty measure is a Foster et al. (FGT, 1984)
one and is expressed in terms of equally distributed equivalent poverty gaps. Its
decomposability property serves to provide a further breakdown across policy-relevant
characteristic subgroups. The approach applies the techniques developed for subgroup
decomposition. Typically, this involves partitioning a household sample data using a set
of policy-relevant characteristics and then calculating two components of each aggre-
gate redistributive component: a weighted average policy-relevant characteristic
inequity value known as the within-group inequity component; and a between-group
inequity term which captures the inequity due to variations in the average across
subgroups. The resultant characterization of redistributive effects is more incisive since
it identifies additional sources of horizontal inequity and vertical equity. This allows
distinguishing the systematic discrimination effects of the policy and the implications of
the analysis for policy design are more clearly identifiable in consequence.
The procedure is applied to the analysis of the program Oportunidades in Mexico in
2006. The results show a substantial loss of total redistribution attributable to horizon-
tal inequity of the program, especially among the households with children in rural
areas. Data show that factors other than groups’ specific characteristics were the

© 2011 Blackwell Publishing Ltd


EQUITY IN POVERTY PROGRAM EVALUATIONS 201

principal cause of horizontal inequity and source of vertical equity in the incidence of
the transfers. However, we notice that the differences between groups are higher in the
vertical equity component than in the horizontal one.These differences point to the fact
that a considerable amount of the success of the policy tends to be concentrated on
rural households, both with and without children. This result can be explained reason-
ably well by the recognized difficulty in the targeting mechanism of the Oportunidades
program in allocating different amounts of resources to the poorest households in
urban areas (Fay et al., 2005). Finally, armed with consistent measures of equity aspects,
the paper investigates changes in bandwidth and inequality aversion parameters and
concludes that similar results are reached.
The structure of this paper is as follows. The poverty framework used in the analysis
is supplied in section 2. In section 3, we define our decomposition of overall redistribu-
tive effect in terms of horizontal inequity and vertical equity for different policy-
relevant subgroups. Section 4 contains the application to the poverty-reducing program
Oportunidades in Mexico. Section 5 concludes.

2. Poverty Framework
We are interested in how a poverty-reducing policy performs in different socio-
economic groups in the population. For that purpose we divide the population into K
mutually exclusive groups, k = 1, 2, . . . , K, which are homogeneous according to certain
policy-relevant characteristics. For any income vector x broken down into subgroup
income vectors x1, x2, . . . , xk and given a poverty line z, we define as poor all households
with incomes xik < z, where xik is the income of household i in group k. The population
sets corresponding to x and xk are n(x) and n(xk) and their sizes are #n(x) and #n(xk),
respectively. The number of poor households is n(xk; z) = {i ∈ n(xk) : xk < z} which com-
prises all households that belong to groups k whose incomes do not reach the poverty
level.3
Let gi(xk; z) be the poverty gap associated with household i with income xik and let the
poverty line z be defined as
gi ( x k ; z) = max {z − xik , 0} .

Given the value of the poverty line z, many familiar poverty measures may be defined
as functions of the vector gi(xk; z). We choose the FGT decomposable poverty measure
defined for group k as
1 α
Pα ,k ( g ( x k , z)) = ∑ g i ( x k ; z) , 0 ≤ α , (1)
# n ( x ) n( xk ; z)
k

and overall as
K
Pα ( g ( x, z)) = ∑ θ k Pα ,k ( g ( x k ; z)), 0 ≤ α , (2)
k =1

where q k = #n(xk)/#n(x).4
The advantage of this measure is that a large number of indices can be defined simply
by varying one key parameter. For instance, when a = 0, we get the head-count ratio
which is the ratio between the number of poor and the population size. If a = 1, we get
the average poverty gap which is the ratio of the average income shortfall of the poor
to the poverty line. These two measures are not sensitive to the distribution of income

© 2011 Blackwell Publishing Ltd


202 Rocio Garcia-Diaz

among the poor. When a = 2, the poverty measure is distribution sensitive in that a
transfer of income from a poor person to someone poorer decreases poverty. When
a → •, Pa(g(x, z)) approaches the Rawlsian max-min criterion of poverty and the
wellbeing of the poorest person dictates the overall picture of poverty.
We measure the redistributive impact of transfers by means of the equally distributed
equivalent poverty gap concept, computed on the basis of the poverty function defined
in (1) and (2). We therefore have
n( x k ; z)
1 α
Pα (Γα ( g ( x k ; z))) = ∑ g i ( x k ; z) , (3)
n ( xk ) i =1

k
Pα ( Γα ( g ( x; z))) = ∑ θ k Pα ( g ( x k ; z)). (4)
k =1

Formula (3) implicitly represents the equally distributed equivalent (EDE) poverty gap
for subpopulation k, Ga(g(xk; z)). According to Atkinson (1970), it measures the level of
poverty gap which, which if suffered by every household in subgroup k, would produce
the same poverty measure as that generated by the actual distribution of poverty gaps
in subgroup k, while Ga(g(x; z)) in formula (4) does the same but with respect to the
entire population.
This money-metric representation of poverty allows us to quantify the cost of
inequality as the amount the social decision-maker would give up to have inequality
among the poor eliminated without enhancement of total poverty. For every household
this is
Cα ( g ( z)) = Γα ( g ( x; z)) − Γ 1 ( g ( x; z)) for α > 1, (5)

where G1(g(z)) is the average poverty gap and is not distribution sensitive. For a > 1,
Ga(g(z)) becomes an EDE poverty gap that is distribution-sensitive and is able to
capture the level of inequality among the poor. The size of the cost of inequality among
the poor will depend on the parameter a and our view of inequality among the poor.
We might say that Pt( · ) is more inequality-averse than Pu( · ) if Gt(g(x; z)) > Gu(g(x; z)),
which implies that Ct(g(z)) > Cu(g(z)) for all x (Lambert, 2001).

3. The Redistribution of Benefits in Population Subgroups


We assume that the redistributive objectives of poverty program designers are framed
in terms of equivalent income. Given the distribution of incomes in the population
vector x, the government is considered to transfer an amount t (x) of benefits to poor
households according to a certain targeting mechanism of the program. Let xi and yi
define the pre-transfer and post-transfer equivalent incomes of household i in the
population, and similarly, denote xik and yik as the pre-transfer and post-transfer
equivalent incomes of household i belonging to group k.
We suppose that households do not receive the same benefits, and that the transfer
they receive is subject to an error of targeting. Thus we write
ti = τ ( xi ) + ε i , (6)
where t (xi) is the amount that all households should have received according to their
level of income xi, and ei is the household deviation from this amount of benefits. The
presence of ei in (6) means that households with the same level of income can end up

© 2011 Blackwell Publishing Ltd


EQUITY IN POVERTY PROGRAM EVALUATIONS 203

receiving different amounts of benefits or no benefits at all. This is the classical notion
of horizontal inequity. Only if ei is equal to zero for all households having the same level
of income is the transfer system horizontally equitable.
The first step when measuring horizontal inequity is to identify the equals group for
a given subpopulation k. For this purpose we define as equals those households having
the same pre-transfer equivalent income xi and belonging to a subgroup k—that is,
n(xi, k) = n(xi) 艚 n(xk)—and the size of this group as #n(xi, k). We want to measure
the unequal transfer benefits among these groups of equals and the horizontal inequity
that results when people in these groups obtain different values of ei.
Following Bibi and Duclos (2007) we conceptualize horizontal inequity at xik by the
inequality introduced by the policy at xik. That is, the local measure of horizontal
inequity is given by

Hxk = Γα , xi ( g ( yk ; z)) − Γ 1, xi ( g ( yk ; z)) ,


i

that is the cost of inequality for the post-transfer distribution within the group of equals
n(xi, k). Therefore, only if there is nonzero cost of inequality within n(xi, k), does the
transfer program display horizontal inequity at the income level xik.
The next step is to aggregate the Hxk into a global index of horizontal inequity in
i
subgroup k, using a weighting scheme that depends on population shares. Thus we have

Hk = ∑ θ x ,k H x
i k , (7)
( )
i
n xk

where θ xi,k is the proportion of the overall population in subgroup k located at point xi
on the pre-transfer income scale; that is, θ xi,k = # n ( xi , k ) # n ( x k ). This is the pure
weighting scheme proposed by Lambert and Ramos (1997) of which one advantage is
that a local violation of horizontal equity does not depend upon the income level
at which is experienced and in consequence Hk is not polluted with vertical
considerations.
Vertical equity would occur if equals had been treated equally and there was no
horizontal inequity. In that case, every household in the equals group n(xi, k) would
have received Γα , xi ( g ( yk ; z)). This will generate a vector of post-transfer EDE poverty
gaps according to every pre-transfer living standard xik.We can write this distribution as
Γα , x ( g ( yk ; z)) = {Γα , x1 ( g ( yk ; z)) , Γα , x2 ( g ( yk ; z)) , . . . , Γα , xN ( g ( yk ; z))} , (8)

where the cost of inequality of this distribution is given by


Cα (Γα , x ( g ( yk ; z))) = Γα (Γα , x ( g ( yk ; z))) − Γ 1 (Γα , x ( g ( yk ; z))) .

The vertical equity of a given policy can be measured by the difference between the
cost of inequality of this distribution in which equals have been treated equally and the
initial cost of inequality; that is, by

Vα ,k = Cα (Γα ( g ( x k ; z))) − Cα (Γα , x ( g ( yk ; z))) . (9)

In this way Va,k represents the decrease in inequality generated by a poverty program
which treats equals equally. Thus, the index measures the underlying vertical equity or
progressivity of a poverty program with respect to the initial inequality distribution
among poor households. This agrees with Musgrave (1990), who also proposes to assess
the vertical performance of a policy reform in terms of one hypothetical income

© 2011 Blackwell Publishing Ltd


204 Rocio Garcia-Diaz

distribution generated by the actual distribution (in this case, among the poor) but
assumes equal division of the benefits within each group of equals.
We can write REa,k as the redistribution effect of a poverty-reducing policy among
poor households in subgroup k, with respect to a given parameter a, in the following
way:
REα ,k = Cα (Γα ( g ( x k ; z))) − Cα (Γα ( g ( yk ; z))) . (10)

This is a measure in terms of the reduction of the cost of inequality among poor
households due to the transfer, and the size of this reduction will depend on the
parameter a and our view of the inequality among the poor. The cost of inequality as
defined in equation (5) are based on poverty gaps. Hence, to define the change in the
cost of inequality before and after the transfer we follow Creedy (1996), in that
transfers received by a pre-transfer poor are regarded as unnecessary if individuals
after the transfer are raised in excess above the poverty line. The transfers received by
pre-transfers poor will certainly reduce the total amount of poverty, but this is not fully
recognized in the measurement of REa,k since what it concerns is how the benefits are
redistributed among households that were poor before the policy has taken place.
We can decompose the cost of inequality of post-transfer income in subpopulation k
into its within-equals groups and between-equals groups components:

Cα (Γα ( g ( yk ; z))) = ∑ θ k , x [Γα , x ( g ( yk ; z)) − Γ 1, x ( g ( yk ; z))]


i i i
( )
n xk

+ [ Γα , x ( g ( yk ; z)) − Γ 1, x ( g ( yk ; z))]
= ∑ θ k , x H x ,k + Cα (Γα , x ( g ( yk ; z))) .
i i (11)
( )
n xk

Using (10) and (7) we have that

REα ,k = Cα (Γα ( g ( x k ; z))) − ∑ θ k , x H x ,k − Cα (Γα , x ( g ( yk ; z))) ,


i i
( )
n xk

which can be written as

REα ,k = Vα ,k − Hα ,k . (12)
This is the total amount of redistribution among the poor due to the implementation of
a poverty-reducing policy. The horizontal inequity component, Ha,k, accounts for the
loss in redistributive performance of a given policy due to different treatments of
equally poor households in subpopulation k. The specification of separate groups
according to some policy-relevant characteristics in equation (12) allows for the pos-
sibility to track down these effects to their sources to better understand the
determinants of inequity and the policies which could be implemented to reduce it.
The result in (12) is similar to the one found in Lambert and Ramos (1997) for the case
of a redistribution decomposition of a tax system when using a decomposable inequality
measure such as the logarithmic mean deviation. For the case of a nondecomposable
poverty measure, such as the Gini coefficient of the censored distribution proposed by
Takayama (1979), the last term in (12) would have resulted in a classical horizontal
inequity component and a re-ranking component like the one found in Aronson and
Johnson (1994). Bibi and Duclos (2007) get a related but different result: instead of using
the amount of redistribution REk they use a measure of poverty effectiveness of a policy
given by the fall in poverty that results from moving from the pre-transfer distribution of

© 2011 Blackwell Publishing Ltd


EQUITY IN POVERTY PROGRAM EVALUATIONS 205

poverty gap to the post-transfer distribution of EDE poverty gaps. They also derive a
measure of leakage of benefits as a component in the decomposition of poverty effec-
tiveness.The measure of leakage of benefits includes the value of the transfers that raise
the pre-policy poor above the poverty line, keeping “equal poverty” comparisons as
defined by Atkinson (1993) and in Creedy (1996).
The overall redistributive effect of the poverty-reducing policy is given by

REα = ∑ θ k REα ,k + REα*,


n( x )

where q k is k’s population share (the number of households in group k, divided by the
total number of households), and REα* is the amount of redistribution between groups:

REα = ∑ θ k Vα ,k + Vα* − ∑ θ k ∑ θ k , x H x ,α ,k − ∑ Hα* ,


i i
(13)
n( x ) n( x ) ( )
n xk ( )
n xk

where Hα* and Vα* are the between-group contributions in terms of horizontal inequity
and vertical equity, respectively.
The first and third components in (13) are the vertical equity and horizontal inequity
in subpopulation k, respectively.The specification of separate groups according to some
policy-relevant characteristics in equation (13) allows the possibility to measure hori-
zontal and vertical discrimination of the program; this is captured in the between-
groups components. Between-group horizontal inequity measures the amount of
redistribution that results when different types of households with identical pre-
transfer incomes do not receive the same transfer because of systematic differences due
to the targeting mechanism of the program. Between-group vertical equity can be used
to approximate the amount of vertical discrimination of the program. The amount of
vertical discrimination, as proposed by Kakwani and Lambert (1999) for the case of
income taxation, is caused by differences in the effective average transfer for different
population subgroups k with respect to the average transfer received by the total
population.
In our analysis we draw on the ideas of Kakwani and Lambert (1999) on income
taxation, to identify both vertical equity and horizontal inequity between group com-
ponents of a poverty-reducing policy as approximations of the total amount of
discrimination between groups. The idea is that the between-groups components
capture the deviations of post-transfer income schedules from a nondiscriminatory
schedule that is based on the idea that different policy-relevant subgroups should be
treated in the same way. However, we refer to it as an approximation because typically
a government transfer depends not only on many attributes, some of which can be
modeled as group-specific socio-economic characteristics, but also on behavior. This is
the case in some poverty-alleviation programs that impose costs (psychological and
pecuniary) on the poor households who have to claim, which may deter some of them
from claiming. Nevertheless, the methodology gives important insights into policy
performance and, for most applications, the candidate groupings for the investigation
of potential discrimination will be defined in pursuit of some socio-economic agenda.

4. Application to the Poverty-Reducing Program Oportunidades in Mexico


An important policy aim of the Mexican government is to direct more spending to
social programs with a substantial share targeted to the benefit of the poor. One of the
most important poverty-reducing programs in Mexico is Oportunidades (or Progresa,

© 2011 Blackwell Publishing Ltd


206 Rocio Garcia-Diaz

as it was known before 2000), which started in 1998 as a rural program and was
extended to reach urban areas in 2000. The program Oportunidades has 2.6 million
families, equivalent to 40% of all rural families, and one-ninth of all families were
receiving program benefits. The total annual budget of the program in 2006 was just
under 20% of the federal poverty alleviation budget or 0.2% of GDP. The program
includes rural and urban rural poor areas in Mexico.Although these transfers represent
less than 5% of total public expenditure, their allocation to the extreme poor makes
them highly effective as a redistributive instrument, being equivalent to up to 50% of
the autonomous income of the poorest households (Mexico report, World Bank,
2004).
This program is one of the most successful in Latin America and is a prototype for
other programs in the region because of the targeting methods it uses. This makes it a
particularly interesting program to evaluate from the perspective of equity consider-
ations. The program follows a two-phase procedure in order to reach the poorest
households. First, eligible communities are identified by means of a marginality index,
with the most marginal localities being eligible to receive the program. Second, within
this eligible community survey information of the households is used to calculate
proxy-mean scores that depends on the household composition in order to identify the
eligible poor households. Once a household is deemed entitled to the benefits the
members decide whether to take the benefits or not.
Data used in this analysis come from the Encuesta Nacional de Ingresos y Gastos de
los Hogares (ENIGH) of Mexico for 2006. This survey contains information on socio-
demographic characteristics of individuals and households and on their main sources of
income, including state transfers. We use gross per capita household current income as
the welfare measure. The gross current income of household is the sum of all monetary
income comprising: earnings, income from business, capital income, Oportunidades
transfers, income from cooperatives, and other income.5 Pre-transfer income x is
defined as gross income before Oportunidades transfers are given and post-transfer
income, y, is total income after transfers.
The statistical data are monthly average figures in August of the corresponding year’s
prices and are adjusted to the consumer price index. In order to adjust the data for
heterogeneity in the size and composition of families, we first equivalized total monthly
income and transfer variables using the Engel equivalence scale constructed for
Mexico (Teruel et al., 2005). This scale gives 0.82 units for each adult between 19 and 65
years old; 0.63 units for each child less than 5 years old; 0.66 units for each child
between 6 and 12 years old; and 0.61 for each teenager between 13 and 18 years old.
The proposal is based on the idea that it is possible to compare the transfer treatment
of households with the same policy-relevant characteristics apart from the equivalence
scale used in the analysis. Getting a consensus about which are the policy-relevant
characteristics may well be difficult. Our personal view is that important distinctions of
the program are between rural and urban areas and between household with and
without dependent children.Therefore, to illustrate the approach and the informational
advantages of the equity decomposition, the population has been divided into four
groups: urban households with children (41.21%), rural households with children
(25.83%), urban households without children (10.47%), and rural households without
children (22.46%).6
The poverty line we use is the official food poverty line, which is an estimate of the
income required to purchase a food basket to satisfy minimum nutritional require-
ments for rural areas.7 The real poverty line z is of 598.70 Mexican pesos (MXP) and is
close to standardized poverty lines of approximately two dollars a day, at 1993

© 2011 Blackwell Publishing Ltd


EQUITY IN POVERTY PROGRAM EVALUATIONS 207

Table 1. EDE Poverty Gaps of Market Incomes

Groups

a Total 1 2 3 4

1 36.62 77.91 11.10 89.43 6.43


(0.80) (3.57) (1.03) (2.27) (0.46)
2 121.45 183.20 70.12 189.33 43.76
(1.63) (5.12) (4.00) (2.94) (2.26)
3 189.99 254.47 136.88 253.67 90.50
(2.00) (5.58) (5.97) (3.09) (4.27)
4 242.56 305.67 195.03 299.08 135.76
(2.19) (5.74) (7.01) (3.12) (6.10)

Notes: Group 1 is rural households without children, group 2 is urban


households without children, group 3 is rural households with children,
and group 4 is urban households with children. Standard errors in
parentheses.

purchasing power parities (Chen and Ravallion, 2004) despite following different
methodologies.
In Table 1 we first analyze the trends in poverty for the different groups in the
population using incomes before transfers. Under each column we find the EDE
poverty gap estimates in terms of the pre-transfer income for every group and the total
population, as well as the standard deviation (in parentheses). Consistent with many
analysts, we note that at different values of the inequality-aversion parameter (a),
monetary poverty is considerably higher in rural areas than in urban areas, especially
for the group of households with children in rural areas, which faces the highest
monetary poverty among the groups. This result is the same for some relevant values of
the inequality aversion parameter, but changes when a = 4 when households without
children in rural areas become the group with the highest monetary poverty. On the
other hand, households with children in urban areas have the lowest monetary poverty
at every level of the inequality-aversion parameter.
One problem with the empirical estimation of horizontal inequity is that there are
typically few exact equals in sample micro data that are drawn from a continuous
distribution of household pre-transfer income x. It is said that in this case, pure hori-
zontal inequity will be underestimated if computed from sample micro data and in
consequence estimates in (7) and (9) will be biased. This is known as the identification
problem and refers to the loss of information about the presence of exact equals which
comes from the act of sampling. However, there are several ways in which the analyst
can cope with this problem. The approach we use here consists in identifying groups of
close equals by means a discrete banding with respect to levels of pre-transfer income;
see Aronson and Johnson (1994), Lambert and Ramos (1997), and van de Ven et al.
(2001). For this discrete banding we define a window that is passed across the pre-
transfer income range of a given fixed width and nonoverlapping windows. In order to
identify the optimal bandwidth we use income intervals ranging (on a monthly basis)
from 14 MXP to 222 MXP.
Table 2 shows the relevant values for total redistribution (REa), total horizontal
inequity (Ha), and total vertical equity (Va), for various choices of bandwidth using the
total population. As the bandwidth is increased, the number and disparity of

© 2011 Blackwell Publishing Ltd


208 Rocio Garcia-Diaz

Table 2. The Choice of Bandwidth

Bandwidth No. of groups REa Ha Va

14.21 1484 12.1456 1.6823 13.8279


20.41 1113 12.1456 1.6823 13.8278
30.67 742 12.1456 1.7165 13.8621
51.18 445 12.1456 1.7657 13.9113
77.02 296 12.1456 1.8560 14.0016
102.78 222 12.1456 2.0119 14.1575

households allocated to each close-equals group increases. Increasing the number of


households identified in each close-equals group necessarily results in an increase of
the estimate H. To analyze this, we successively selected smaller and smaller band-
widths and computed Ha and Va. The results are that a bandwidth of approximately
20 MXP would enjoy the same orders of magnitude as for even smaller bandwidths.
Using this bandwidth size the horizontal inequity of the whole population corresponds
to approximately 13% of REa.
An alternative approach to the identification problem is nonparametric kernel esti-
mation which uses exogenous definition of intervals following nonparametric
techniques as in Duclos and Lambert (2000), Ramos and Lambert (2003) and
Rodríguez et al. (2005). The nonparametric procedure follows a kernel estimation
procedure with the Gaussian kernel and bandwidth chosen to minimize the square
error. This procedure uses a moving window along the pre-transfer income distribution
giving greater weight to the information near the middle of the window. Given that the
group of equals is determined by the total population and not just for the population of
poor households, choosing the bandwidths that minimize the square error results in
larger bandwidths (fewer groups of equals) than the discrete banding procedure pro-
posed above. This is a consequence of the chosen poverty line which is intended to
capture the extreme poor in the population. For this reason we choose the discrete
banding approach in preference to the nonparametric procedure, in order to allow for
a higher number of groups of equals among the poor and generate a more detailed
equity analysis in that part of the distribution.
We now move to our main results in formula (13). Table 3 shows the estimates from
our 2006 sample. For a given inequality-aversion parameter, we show estimates for each
policy-relevant group, between policy-relevant groups and for the total population. We
note first that most of the redistribution takes place in the rural areas, especially among
the households without children. The group with the lowest impact is urban households
with children. The amount of redistribution for the policy-relevant groups follows
closely the structure of poverty in the pre-transfer income we saw in Table 1 so
the program is giving more benefits to those households in the groups with higher
monetary poverty for relevant levels of a.
There are differences in the policy-relevant groups that help us to identify the
sources of inequity in the program. First, it is important to note that it is the sum of
horizontal inequity within groups that accounts most for the total horizontal inequity of
the program, since the between-components are very small. The group with the highest
horizontal inequity is rural households with children, where redistribution can be
improved up to 49% if horizontal inequity within this group is eliminated. In the urban
sector, despite having the lowest redistributive effect, households with children face

© 2011 Blackwell Publishing Ltd


EQUITY IN POVERTY PROGRAM EVALUATIONS 209

Table 3. Redistribution Effects for Policy-Relevant Subgroups

Ga(g(x; z)) Ga(g(y; z)) REa,k Ha,k Va,k

a=2
Group 1 183.196 147.734 15.921 5.530 21.451
Group 2 70.121 67.236 2.104 0.530 2.634
Group 3 189.330 154.826 12.200 6.034 18.234
Group 4 43.764 38.340 0.903 0.903 1.806
Between groups 6.456 0.070 6.526
Total 121.448 100.733 12.121 2.700 14.821
a=3
Group 1 254.466 212.451 22.530 9.363 31.893
Group 2 136.879 132.633 3.467 0.881 4.348
Group 3 253.668 214.531 16.841 10.260 27.101
Group 4 90.503 82.475 6.751 1.535 8.286
Between groups 13.230 0.107 7.527
Total 189.994 162.951 17.693 4.570 22.263

Notes: Bandwidth of 20 Mexican pesos. Group 1 is rural households without children, group 2 is urban
households without children, group 3 is rural households with children, and group 4 is urban households with
children.

more horizontal inequity than those without children. This reflects the difficulty of the
program to target benefits to households with children in the urban and rural sectors,
which are also the more representative in the total population. The between-group
component of horizontal inequity, that measures the differences in benefits of house-
holds with the same pre-transfer income among policy-relevant subgroups, is very low
reflecting almost no horizontal discrimination in the program.
The last column in Table 3 shows the vertical equity component of the policy. For the
total population, vertical redistribution accounted for 134% of the actual redistributive
effect in 2006, with a 34% loss due to horizontal inequity. It is again the within
component that contributes the most to total vertical equity for both inequality-
aversion parameters. Within-group vertical equity varies considerably among the
policy-relevant groups, where the rural households without children experience more
vertical equity than the other socio-economic subpopulations. This is probably due to
the fact that the Oportunidades program means-testing benefits are the same for wide
range of income levels and are primarily focused on households in rural areas. We also
find that in terms of vertical equity of the policy the between-group component is much
more relevant, in contrast with the case of horizontal inequity, since it accounts for
almost 48% of total redistribution. This suggests a considerable amount of discrimina-
tion among policy-relevant groups due to differences in the average benefits received
by households at different levels of pre-transfer income, poor households in urban
areas being the ones that face more discrimination.

5. Conclusion
The central purpose on this paper has been to propose a coherent framework for the
characterization and measurement of the effects of a poverty-reducing policy on the
distribution of incomes among poor households and according to some policy-relevant
characteristics defining groups. The basis of the methodology is the measurement of the

© 2011 Blackwell Publishing Ltd


210 Rocio Garcia-Diaz

overall redistributive effect of a poverty-reducing policy as the resultant change in the


cost of inequality among pre-transfer income-poor households. We then provide a
general decomposition into components that allow identifying between-group and
within-group horizontal inequity and vertical equity aspects. This serves not only to
quantify but also to characterize the redistributive performance of the policy.
We have applied the methodology to the poverty-reducing program Oportunidades
in Mexico. It shows that, according to the policy-relevant characteristic groups selected,
the majority of the horizontal inequity takes place within these groups showing almost
no horizontal discrimination among them. However, the within-group horizontal in-
equity varies considerably, the households in rural areas being the ones that have a
greater scope for improvement by reducing the within-group horizontal inequity. On
the other hand, the between-group vertical component explains almost half of the
vertical equity of the program reflecting an important discrimination among policy-
relevant groups due to differences in the average benefits received for households at
different levels of pre-transfer income. These results are subject to the implementation
difficulties that arise from the identification problem, as well as changes in equivalence
scales and inequality-aversion parameters. For that we have done a sensitivity analysis
to changes in the bandwidth and inequality-aversion parameters and found no quali-
tative changes in the results. The methodology has wide applicability to evaluate equity
of a poverty-reducing policy and can well complement traditional methodologies that
measure efficiency aspects of welfare impacts such as leakage and undercover rates.

References
Aronson, R. L. and P. Johnson, “Redistributive Effect and Unequal Income Tax Treatment,”
Economic Journal 104 (1994):262–70.
Atkinson, A. B., “On the Measurement of Inequality,” Journal of Economic Theory 2 (1970):244–
63.
———, “On Targeting and Family Benefits,” in A. B. Atkinson (ed.), Income and the Welfare
State, Cambridge: Cambridge University Press (1993):223–61.
Bibi, S. and J.-Y. Duclos, “Equity and Policy Effectiveness with Imperfect Targeting,” Journal of
Development Economics 83 (2007):109–40.
Chen, S. and M. Ravallion, “How Have the World’s Poorest Fared since the Early 1980s?” World
Bank Research Observer 19 (2004):141–69.
Coady, D., “The Welfare Returns to Finer Targeting: The Case of the Progresa Program in
Mexico,” International Tax and Public Finance 13 (2006):217–39.
Cornia, G. A. and F. Stewart, “Two Errors of Targeting,” in D. van de Walle (ed.), Public Spending
and the Poor, Theory and Evidence, Washington, DC: Johns Hopkins University Press
(1995):459–96.
Creedy, J., “Comparing Tax and Transfer Systems: Poverty, Inequality and Target Efficiency,”
Economica 63 (1996):S163–74.
Duclos, J.-Y., “On Equity Aspects of Imperfect Income Redistribution,” Review of Income and
Wealth 41 (1995):177–90.
Duclos, J.-Y. and P. J. Lambert, “A Normative and Statistical Approach to Measuring Classical
Horizontal Inequity,” Canadian Journal of Economics 33 (2000):87–113.
Ebert, U.,“Optimal Anti-Poverty Programmes: Horizontal Equity and the Paradox of Targeting,”
Economica 72 (2005):453–68.
Fay, M., L. Cohan, and K. McEvoy, “Public Social Safety Nets and the Urban Poor,” in M. Fay
(ed.), The Urban Poor in Latin America, Washington: The World Bank (2005):239–66.
Foster, J. E., J. Greer, and E. Thorbecke, “A Class of Decomposable Poverty Measures,” Econo-
metrica 52 (1984):761–66.
Kakwani, N. and P. J. Lambert, “Measuring Income Tax Discrimination,” Review of Economics
and Statistics 81 (1999):27–31.

© 2011 Blackwell Publishing Ltd


EQUITY IN POVERTY PROGRAM EVALUATIONS 211

Keen, M., “Needs and Targeting,” Economic Journal 102 (1992):67–79.


Lambert, P. J., The Distribution and Redistribution of Income, Manchester: Manchester Univer-
sity Press (2001).
Lambert, P. J. and X. Ramos,“Horizontal Inequity and Vertical Redistribution,” International Tax
and Public Finance 4 (1997):25–37.
Musgrave, R. A., “Horizontal Equity, Once More,” National Tax Journal 43 (1990):113–22.
Ramos, X. and P. J. Lambert, “Horizontal Equity and Differences in Income Tax Treatment: A
Reconciliation,” Research on Economic Inequality 10 (2003):45–63.
Rodríguez, J., R. Salas, and I. Perrote, “Partial Horizontal Inequity Orderings: A Non-Parametric
Approach,” Oxford Bulletin of Economics and Statistics 67 (2005):347–68.
Schady, N., “Picking the Poor Indicators for Geographic Targeting in Peru,” Review of Income
and Wealth 48 (2002):417–34.
Skoufias, E. and D. Coady, “Are the Welfare Losses from Imperfect Targeting Important?”
Economica 74 (2007):756–76.
Takayama, N., “Poverty Income Inequality and their Measures: Professor Sen’s Axiomatic
Approach Reconsidered,” Econometrica 47 (1979):747–59.
Teruel, G., L. Rubalcava, and A. Santana, “Escalas de equivalencia para México,” Serie de
Documentos de Investigación de SEDESOL 23 (2005).
van de Ven, J., J. Creedy, and P. Lambert, “Close Equals and Calculation of the Vertical,
Horizontal and Reranking Effects of Taxation,” Oxford Bulletin of Economics and Statistics 63
(2001):381–94.
World Bank, Mexico Public Expenditure Review, Washington, DC: World Bank (2004).
Zheng, B., “Aggregate Poverty Measures,” Journal of Economic Surveys 11 (1997):123–62.
———, “On the Power of Poverty Orderings,” Social Choice and Welfare 16 (1999):349–71.

Notes
1. Ebert (2005) found out that Keen’s (1992) paradox can be avoided for two alternative
strategies of targeting: the universal provision and the finance-guaranteed minimum incomes. In
doing so, he proposed a relative equivalence scale as indicator of needs where the weight
attached to each household type equals the respective equivalence scale value.
2. Exclusion errors are well recognized in the literature of policy targeting; they are termed F
mistakes, that is, failing to reach the targeted population; they are also known as Type I
mistakes—see, for instance, Cornia and Stewart (1995), Duclos (1995), Schady (2002), Coady
(2006), and Skoufias and Coady (2007).
3. The results of this analysis do not change with the use of the strong definition of the poor.
However, we use the weak definition since it is more consistent with the idea that the goal of any
anti-poverty program is to bring all poor to the poverty line (Zheng, 1997).
4. For other examples of poverty measures see Foster et al. (1984) and Zheng (1999).
5. Before calculating the monthly income of each household the category “other income” was
eliminated.
6. Dependent children are those persons under the age of 18. Rural households are those that
belong to localities with fewer than 2500 inhabitants and urban households are those that belong
to localities with at least 2500 inhabitants.
7. The bundle is based upon the food-spending patterns of households who just satisfy minimum
nutrient requirements, assuming all spending was on food. Then, the food poverty line is deflated
using a food-specific consumer price index.

© 2011 Blackwell Publishing Ltd

You might also like