1
Targeting maps:
An asset-based approach to geographic targeting*
Corey Lang,** Christopher B. Barrett and Felix Naschold Cornell University
August 3, 2009
Abstract Proper targeting of policy interventions requires reasonable estimates
of the benefits of the various alternative interventions. In order to inform such decisions,
we develop an integrated approach that estimates the marginal returns to a range of assets
across geographically defined subpopulations allowing returns to vary by household and
by geography. We then create a series of maps illustrating the estimated marginal returns
to specific assets and the proportion of an area’s population that would benefit from
increased holdings of a specific asset. These maps can then be overlaid with traditional
poverty maps to identify areas that are strong candidates for a particular development
intervention. We develop a general method and demonstrate its potential with an
application using Ugandan data.
JEL classification: R12, O2, C15, I32 Keywords: geographic targeting, assets, poverty maps, spatial variation, Uganda
* We appreciate helpful discussions with Nancy Johnson, GIS data assistance from John Owuor and Ugandan data advice from Thomas Emwanu. Useful comments were received from seminar participants at Cornell University. This research was made possible through the support of the International Livestock Research Institute. ** Contact author. Address: Department of Economics, 404 Uris Hall, Ithaca, NY, 14853. Email addresses: [email protected] (Lang), [email protected] (Barrett), [email protected] (Naschold)
2
1 Introduction
Improved targeting of development interventions has long been recognized as
central to achieving greater impact from poverty reduction efforts. However, effective
targeting requires reasonable estimates of where the returns to various programs are
likely to be highest. Currently, no means exist for estimating and comparing expected
benefits across space and across alternative interventions. In this paper, we develop a
method that, first, estimates the marginal returns to a range of assets allowing returns to
vary by household and by geography and, second, maps the estimated marginal returns
creating a visual tool that can inform the targeting decisions of an in-kind transfer
scheme.
There are several methods of targeting, such as a means test, community-based
targeting, categorical or indicator targeting and self-targeting, each with its own
advantages and disadvantages.1 The empirical evidence suggests that geographic
targeting is particularly effective for poverty alleviation (Coady et al. 2004, Baker and
Grosh 1994) and is easier and less expensive to monitor and administer than other
methods (Bigman and Fofack 2000). The idea of geographic targeting is to determine a
subset of geographic regions most in need and then transfer benefits to individuals within
the chosen regions and exclude all others. The benefits to this method are intuitive as
there is ample evidence that individuals living in close geographic proximity tend to have
similar livelihoods and face the same constraints and risks (e.g., Bigman and Fofack
2000, Doss et al. 2008).
The major disadvantages to geographic targeting are that non-poor individuals
living in targeted regions receive benefits (leakage) and poor individuals not living in 1 Coady et al. (2004) discuss these targeting methods and more in detail.
3
targeted regions do not receive benefits (undercoverage). One remedy is to target more
finely partitioned regions. As regions become increasingly disaggregated, within region
heterogeneity decreases and targeting performance increases (Elbers et al. 2007, Baker
and Grosh 1994). A second solution is to combine geographic targeting with additional
targeting tools to limit leakage. Coady et al. (2004) surveyed 122 targeted transfer
programs and found the mean number of tools used is more than two – for example,
Mexico’s celebrated PROGRESA/Opportunidades program uses four (Coady 2006).
In this paper, we build on the proven successes of geographic targeting to propose
an enhanced, asset-based approach. In general, transfers can be monetary or in-kind,
where in-kind transfers usually come in the form of subsidies for food, education, or
health services. Here, we explore the possibility of transfers from an entire range of
private and public assets, such as livestock, mobile phones, means of transportation, and
access to roads or microfinance institutions. Our focus on assets stems from the
importance of a household’s asset portfolio in determining the nature and extent of
poverty and vulnerability (Moser 1998, Ellis and Freeman 2004, Adato et al. 2006).
Further, asset transfers may push households beyond an asset poverty threshold and allow
them to engineer their own escape from income poverty (Carter and Barrett 2006).
An obvious criticism of in-kind or asset transfers is that unlike with a cash
transfer, a household is constrained and cannot consume or invest in whatever they think
will best help them.2 While in-kind transfers can appear paternalistic, there are several
reasons why an asset-based approach could perform better than a monetary approach.
First, asset transfers can act as a natural self-selection mechanism to reduce leakage;
2 Currie and Gahvari (2008) review the debate over monetary versus in-kind transfers, though mainly from the perspective of developed countries.
4
whereas virtually everyone would accept a cash transfer, only those who benefit from a
given asset would accept it as a transfer. Second, in-kind transfers may stick to the
targeted households better than cash because of the well-established endowment effects
associated with physical goods but not with cash. The findings of Hoffman et al. (2009)
suggest that in-kind transfers of mosquito nets would result in greater use of the nets than
would equivalent cash transfers. Third, monetary transfers, due to their ready divisibility,
may also be subject to a high rate of social taxation compared to a lumpy asset, perhaps
undoing efforts to control leakage. Fourth, imperfect markets can make it difficult to
procure specific, desired assets; this is a common rationale for in-kind food or seed aid in
many remote or disaster-affected regions.
The targeting maps tool improves the information set informing geographic
targeting. Given substantial spatial heterogeneity in poverty incidence and its causes
(Emwanu et al. 2007, Okwi et al. 2007, Kam et al. 2005), there is little reason to believe
that any single poverty alleviation strategy is best suited for all places in a country.
Likewise, spatially heterogeneous asset valuation appears the norm, given the place-
specificity of many complementary inputs – e.g., agro-ecological conditions that affect
livestock value, urban proximity that affects the returns to land, etc. If poverty measures
and the returns to assets both vary markedly across space for a variety of geographic,
institutional, policy and technological reasons, then it seems desirable to exploit the
predictable component of such variation in targeting development interventions. Previous
research has found considerable intra-regional variation in expected returns to different
development investments in Africa and Asia (Fan and Hazell 2001, Fan and Chan-Kang
2004). By customizing asset-based interventions to specific geographic areas, significant
5
gains could be made in efficiently and cost-effectively addressing poverty. Our approach
integrates spatially-explicit estimates of the marginal benefits to multiple assets into a
single framework such that inter-asset comparisons of expected marginal benefits can be
made for each region. The output can then be used as one of several components
informing a targeted transfer plan.
Our method draws on the small-area estimation technique pioneered by Elbers et
al. (2003). Their method combines detailed, nationally representative household survey
data with national census data to estimate poverty rates at fine levels of disaggregation
for an entire country.3 First, they derive a relationship between household expenditure
and various demographic and asset variables using the survey data. Second, they predict
out-of-sample estimates of expenditure for the census data using the coefficient estimates
from the relationship derived with the survey data. By projecting expenditure estimates
onto the full population, the Elbers et al. (2003) method enables estimation of poverty
rates in places where no survey data exist and at finer levels of disaggregation than when
using household survey data alone, as these are typically statistically representative only
at relatively coarse scales of aggregation.
Once estimated, the poverty rates for the various regions of a country can be used
to create a poverty map – a visual illustration of the spatial distribution of poverty. This
simple tool is popular and widely used by governments, NGOs and donors in low-income
countries to guide poverty reduction efforts. Poverty maps can significantly bolster
3 All of the Foster-Greer-Thorbecke measures of poverty as well as inequality can be estimated using this method.
6
geographic targeting efforts because, as mentioned above, geographic targeting methods
are greatly improved as the geographic scale becomes finer.4
Although poverty maps illustrate problems well and can facilitate policy
discussions, they offer no explicit recommendation as to the best means of alleviating
poverty. If a government is trying to reach a specific welfare target such as the
Millennium Development Goals, poverty maps can at best guide the government to
regions with high poverty rates. What exactly the government should do in that region,
however, remains uncertain.
Targeting maps address these shortcomings by answering two general questions:
1) for a given region, which asset building activity will have the largest marginal gross
benefit? and 2) for a given type of asset building activity, in which regions are the
marginal gross benefits to such an investment highest? Both of these questions address
how to improve the efficacy of targeted, asset-based development programs. Answers to
the first question are paramount for those wishing to cut poverty by the most efficient
means possible. The second question appeals to groups interested in investments of a
specific type, such as Heifer International in building livestock holdings or The Nature
Conservancy in safeguarding natural resources. With scarce aid resources available,
targeting maps can help identify where the most bang-for-the-buck exists.
The construction of targeting maps involves several distinct steps similar to those
involved in creating a poverty map. Using detailed household survey data and spatially
explicit environmental and infrastructure data, we apply multivariate regression and
4 Small-area poverty estimates can additionally be used in subsequent regression analysis as either the key dependent variable to investigate the causes of poverty (Kam et al. 2005, Okwi et al. 2007) or as an explanatory variable to investigate its consequences (Demombynes and Ozler 2005).
7
bootstrapping to estimate the returns to various assets and how these estimated returns
vary across space. We then project the parameter estimates onto the broader national
census data and calculate the marginal returns as a function of projected estimates and
household asset holdings. Finally, we aggregate the estimated marginal returns across
households for small geographic areas and, using Geographic Information Systems (GIS),
generate maps that highlight both the magnitude and scope of benefits.
We illustrate our approach using Ugandan household survey and census data.
The results are encouraging; estimated and projected marginal benefits to asset transfers
seem reasonable and show remarkable variation across space. Our results clearly identify
promising areas to target as well as indications of key assets to use in a geographic
targeting scheme. These findings reinforce the value of geographic targeting and the
importance of spatial analysis in general.
The next section describes the methodology in detail, explains how it builds on
poverty mapping, and discusses concerns with the framework. Section three gives the
specifics of the Ugandan data. Section four reports the results including: several
examples of types of targeting maps, a simplified benefit-cost analysis for several assets,
and selection of areas that would be strong candidates for a hypothetical asset transfer.
Section five concludes and discusses ideas for future work.
2 Method
We estimate average expected marginal household-level returns to various assets
across geographically defined subpopulations. In the context of this paper, assets will be
taken as anything whose stock can affect a household’s income or expenditure. We
8
classify assets along two dimensions: private vs. public and targetable vs. non-targetable.
Private and public goods follow traditional definitions; public goods are non-rival and
non-excludable; private goods represent the rest. The distinction between targetable and
non-targetable concerns whether an asset’s quantity, quality or existence can be changed
by an intervention. This classification results in four categories: private targetable assets
(e.g., livestock holdings, literacy, land holdings), public targetable assets (e.g., source of
potable drinking water, access to health clinics, road access), private non-targetable assets
(e.g., education of household head, gender of household head) and public non-targetable
assets (e.g., rainfall, temperature). Our method estimates the returns to all types of assets,
but ultimately we are only interested in those that are targetable.
The minimum data necessary to create a targeting map are a nationally
representative household survey and a census taken at about the same time. Additional
environmental or public good variables can and should be added when available to
supplement both the survey and census data. In the first step of our analysis, we compare
the data available in the household survey and the census to generate a set of variables
that are common to both data sets, such as demographic variables, livestock and durable
goods. We restrict the data in this way because we must use a specification that is
replicable in the census for all independent variables.
The second step is to estimate the relationship between per capita equivalent
household expenditure and asset holdings, which include the variables selected in the first
stage as well as relevant environmental and public good variables. We assume that
household expenditure is a function of asset holdings and place-specific asset returns.5
5 This specification can be thought of as permanent or structural income (Carter and May 2001, Adato et al. 2006, Carter and Barrett 2006).
9
We remain agnostic about the functional form of the asset returns equation and model the
relationship between expenditure and asset holdings using a second order flexible
functional form. For household i in location c, letting icy = per adult male equivalent
household expenditure, icA = private, targetable assets, cA = place specific means of the
private, targetable assets, cB = public, targetable assets, icY = private, non-targetable
assets, cY = place specific means of the private, non-targetable assets, cZ = public, non-
targetable assets, and icX = additional household control variables, we can write the
assumed functional form as:6
(1) '),,,('),,,,(' ),,,('),,,,('ln
icicciccicZccciccicXic
ciccicBccicccicAicic
XZYBARZZYYBARYZYBARBZYBAARAy
)(jR is a vector of returns to asset type j = A, B, Y, Z, which is the object of estimation.
The functional form of asset returns implies that the expected returns to each asset can
depend on the stock of every other asset. For example, the returns to a head of cattle may
depend on the household head’s level of education, the average number of cattle owned
in that region, the existence of a nearby livestock market and/or local precipitation levels.
Place specific asset means are only interacted with household levels of the same variable
(i.e., average cattle holding is interacted with each household’s cattle holdings, but not
with each household’s pig holdings or mobile phone ownership). Further, we assume the
error term is composed of a location component and a household specific component.
(2) M')(M icciccc ic
where ],,,[ ccccc ZYBAM .
6 The place specific means, cA and cY , are derived from the census.
10
Our principal goal in this second step in constructing the targeting map is to
accurately estimate the coefficients in the expenditure asset relationship. We bootstrap
200 iterations of the regression, using weighted least squares (weighted by population
expansion factors) with errors clustered at the enumeration area level. We save the
coefficient estimates from each iteration of the bootstrapped regressions.
Having thus estimated the shape of asset returns (many times), in the third step we
project the estimated coefficients from the first stage regressions onto the census data.
Ultimately, however, we are not interested in the coefficient point estimates, but in the
expected marginal household-level return for a given targetable asset, k:
(3) )(ˆ
')(ˆ
')(ˆ
')(ˆ
'][ln ick
Zc
ick
Yic
ick
Bc
ick
Aic
ick
ic
ARZ
ARY
ARB
ARA
AyE
For each iteration of the bootstrap, we project the coefficient estimates onto the census
data and calculate the derivatives for all targetable assets. Aggregating all of the
estimates generates an empirical distribution of estimated marginal household-level
returns to specific assets. The mean estimated marginal return across bootstrapped
iterations yields our best estimate of a household’s expected marginal return for each
asset. We then aggregate households over geographically defined areas and calculate
statistics fundamental to the final product. First, we compute the mean and standard error
of the expected marginal returns for every geographic area and determine which areas
have returns that are statistically significantly greater than zero. The estimated average
marginal returns and their statistical significance inform essential questions about the
expected magnitude of average benefits associated with specific asset transfers in
particular areas. Second, we calculate the proportion of households with positive
11
expected marginal returns for every geographic area, which reflects the scope of benefits
from specific asset transfers in particular areas.
Finally, using GIS techniques, we display the results. Unlike with poverty
mapping, no one map can summarize all of the results; instead this targeting method
requires a series of maps. One map can display the most beneficial asset, as judged either
by the highest expected average marginal returns of any asset or the highest proportion of
positive expected marginal returns of any asset, for each geographic area. This map
would address question one above: for a given region, which asset building activity will
have the largest marginal gross benefit? Then, maps can be made for each asset, showing
either the expected average marginal returns or the proportion of households with
positive expected marginal returns to that asset for each geographic area. These maps
would address question two above: for a given type of asset building activity, in which
regions are the marginal gross benefits to such an investment highest? Two estimated
objects, two broad targeting questions, and many assets make for a large number of maps,
each catering to a different audience or targeting question.
2.1 Comparing our method with poverty mapping methods
No standard poverty mapping methodology exists, but there are common
practices from which our method deviates slightly, thus it is useful to contrast and justify
our approach. One common practice is to partition the data into the smallest regions for
which the survey data are statistically representative and run regressions for each of those
regions separately. For example, Okwi et al. (2006) and Emwanu et al. (2007) split
Ugandan data into nine strata and Demombynes and Ozler (2005) split South Africa into
12
nine provinces. The idea behind this step is to allow coefficient estimates to vary over
space. In contrast, we pool all survey data into a single regression. While our method
does not allow coefficient estimates to vary over space, asset returns can vary
dramatically over space via the large number of place-specific interaction terms. Our
motivation for this choice is to explicitly take into account the influence of place-specific
characteristics on asset returns. If the geographic scope of regressions is limited, the
variation in some variables, especially the place-specific variables such as climate, is
necessarily very limited. This constraint could lead to biased and inconsistent parameter
estimates.
Another common methodological step in poverty mapping is to use stepwise
regression to reduce the number of right hand side variables (Okwi et al. 2006, Emwanu
et al. 2007, Demombynes et al. 2007). When no underlying theory exists about which
variables belong on the right hand side, this method can be used to iteratively delete
variables based on a criterion such as adjusted-R2 or t-statistics. This approach would
likely exclude several asset variables, both limiting the scope of inter-asset comparisons
and potentially biasing the estimated returns of included assets (via omitted relevant
variable bias). Thus, we take a more structured approach and place priority on the
inclusion of all asset variables.
The most common way to estimate the error surrounding the poverty estimates is
to use parametric bootstrapping (Elbers et al. 2003, Demombynes et al. 2007).
Parametric bootstrapping projects coefficient estimates onto census households by taking
random draws from the distribution defined by a single set of regression coefficient
estimates and their associated covariance matrix. The poverty status of individual
13
households are then averaged by geographic areas. This process is repeated many times
to obtain a distribution of each area’s poverty. We choose instead to bootstrap the first
stage estimation in order to reduce bias in the estimates, since our method puts a greater
premium on the regression coefficient estimates themselves.
2.2 Endogeneity concerns
The major pitfall of our targeting maps methodology is the obvious endogeneity
of several asset variables, which can affect results in several ways. First, there is the
basic, natural correlation between expenditure and assets. Ideally, we could use an
accurate measure of income as the key left hand side variable of interest, but of course
such a measure rarely exists (and this would not fully assuage endogeneity concerns
anyway). As a result, there could be a natural positive correlation between asset holdings
and expenditure – in order to acquire assets, one must spend money. On the flip side,
there may be a negative correlation between assets and expenditure in a static setting; for
example, if a household has just sold lots of livestock, it may have increased consumption
but lower ex post livestock holdings. An additional confounding factor is that not all
asset ownership is motivated by current productive value. Some assets are acquired not
because they will produce more current expenditure, but because they enhance welfare in
some other way or at some future date. For example, some livestock may be held for risk
prevention or social status, and educational investments are aimed at increasing future,
not current, income.
Estimated returns to assets may be additionally biased due either to omitted
relevant variables or unobserved heterogeneity. Differences in preferences, ability and
14
various idiosyncratic features specific to households are all potential sources of bias. We
include in our specifications a rich set of place-specific covariates and a complete set of
interaction terms in order to pick up as much variation as possible and diminish the effect
of this kind of bias.
These are clearly serious concerns. They are inherent, however, to any analysis
that tries to answer the questions posed above. Short of running hundreds or thousands of
identical field experiments across vast spaces, there is no practical way to estimate
marginal returns to multiple assets across a large geographical space with ironclad
identification. While it is impossible to argue a purely causal relationship, knowing how
households change their asset portfolios as their welfare increases and how welfare is
related to the environment and infrastructure around them can nonetheless provide quite
useful insights to inform development policy. Given the considerable policy and
operational importance of the questions targeting maps aim to address, we think this
tradeoff is acceptable and hope and expect that future research can ameliorate this
problem somewhat.
3 Data
We apply our method to the 2002 Ugandan National Household Survey, the 2002
Ugandan Population and Housing Census and the 2002 Ugandan community survey, all
administered by the Ugandan Bureau of Statistics (UBOS). The household survey and
census are stratified by four regions (Central, East, North, and West) and an urban-rural
split. For the purposes of this paper, we restrict our attention to rural households only
(5,648 households in the survey and nearly 4.4 million in the census), although urban
15
households could be included easily. The hierarchy for Ugandan administrative units,
from largest to smallest, is nation, district, county, sub-county, and parish. Table 1 lists
how many administrative units of each type exist and the average and median number of
households in each unit. There are one or two enumeration areas (EA) per parish. The
household survey clustered observations at the EA level and randomly sampled
households within the EA.
The private asset variables all come from the household survey and the census.7
We use the census, the community survey and several GIS layers to create location
specific public asset variables. From the census, we calculate measures of ethnic and
religious diversity and population density, as well as means of all variables at the parish
level.8 The community survey includes information on drought, livestock and crop
extension services, markets, microfinance and violence against women. These variables
are aggregated to both the parish and sub-county level, necessary since not all parishes
contained a community that was surveyed. In addition, we use GIS to derive variables
such as average distance to secondary school, average distance to urban areas, average
distance to water, proportion of a region composed of various agro-ecological zones,
average annual rainfall, average variation in rainfall, average annual temperature and
average variation in temperature, among others.9 Data layers for school location, urban
areas, agro-ecological zones and water location were provided by the International
Livestock Research Institute (ILRI). Weather data were downloaded from
7 As stated above, we are constrained to only use variables that appear in both the census and the survey. There are several instances where a variable that would be fantastic to include (e.g., mosquito net coverage of all household members) is only contained in one data set. This underscores the importance of planning and coordinating household surveys and censuses. 8 Diversity is calculated (as in Easterly and Levine 1997) as the probability that two people of different ethnicity/religion meet. 9 Euclidean, or straight-line, distance is used.
16
www.worldclim.org at a resolution of 30 arc-seconds. These geographic variables are
aggregated at the sub-county level, due to limitations with the GIS software.10 Table 2
gives summary statistics for each asset variable for the survey and census. Once all
interaction and second order variables are added, there are a total of 1120 right hand side
variables in our specification of equation 1. Appendix 1 lists all variables used.
In addition to numerical comparability of the data, geographic comparability is
important. Table 3 gives the percentage of each administrative unit represented in the
survey and community census and Appendix 2 shows the geographic location of the
survey data. The survey data appear well dispersed and thus we have confidence that our
estimates are representative of many different geographies.
4 Results
Appendix 3 presents complete results from the bootstraped regressions.
As a first step in analyzing the results, we determine the appropriate level of
aggregation for the expected marginal returns. In standard poverty mapping exercises,
there is a tradeoff between geographic aggregation and precision (Elbers et al. 2003).
The goal is to aggregate households into the smallest possible geographic area without
sacrificing precision, which enables inter-regional comparison.
We aggregate derivatives and calculate means and standard errors of all targetable
assets at three different administrative levels: county, sub-county, and parish. Table 4
gives the estimated standard errors of four assets – bicycles, chickens, microfinance
10 Due to the small size (in terms of area) of some of the parishes and the relatively larger size of the weather raster data, the zonal statistics could not be calculated for all parishes.
17
access and road access11, which we use as examples throughout – as well as the average
across all targetable assets. Clearly, as the area of aggregation grows so does the
standard error. This finding contrasts with the standard inverse relationship found in
poverty mapping due to the difference in our method, which first estimates household
level marginal returns via simulation and then aggregates over geographic areas. Our
error estimates are a composite of ordinary imprecision plus inter-household variation.
As the geographic scale grows, more inter-household heterogeneity is introduced and the
standard errors increase. The empirical findings clearly indicate that parish is the
appropriate level of aggregation for our estimates, especially since the efficiency of
geographic targeting increases as the geographic area decreases in size (Baker and Grosh
1994, Elbers et al. 2007).
Figures 2a-2d plot the estimated average marginal returns that are significantly
greater than zero for chickens, bicycles, access to microfinance and road access,
respectively, at the parish level. The magnitude of the returns is not readily apparent; the
units are the expected additional natural log of monthly per adult male equivalent
expenditure associated with an additional unit of that asset. For purposes of comparison,
the average of the natural log of per adult male equivalent monthly expenditure across the
whole household survey sample is 10.15. The ordinality of the magnitudes of returns
seems reasonable with road access being most valuable, then microfinance access, then
bicycle ownership and finally chickens. There is also considerable spatial variation in the
estimates. We see pockets of high returns, like those in the northwestern Uganda for
11 Microfinance access is a binary variable indicating whether at least one community within a parish indicated having access to microfinance. Road access is measured as an index from 0 to 2, where 0 represents “no roads,” 1 is “seasonal roads” and 2 is “all weather roads”. Values for a parish are averaged responses from all communities surveyed within that parish.
18
chicken, and spatial patterns, such as the remarkable regularity in which the returns to
road access increase with proximity to urban areas. For each asset, a considerable portion
of the country does not exhibit statistically significant returns, reflecting both relatively
large standard errors and several negative point estimates. It is reasonable that some
returns are actually negative because we estimate marginal returns comprehensively,
including areas that are completely unsuitable for certain assets.12
Figures 2a-2d present the proportion of households with estimated marginal
returns greater than zero for chickens, bicycles, access to microfinance and road access,
respectively, at the parish level. The chicken map largely reinforces the information
contained in Figure 1b; areas with high returns, notably the northwest, also have a large
proportion of households with statistically significant positive expected returns. The road
access map offers little way to distinguish between areas as almost all areas have a
beneficiary proportion over 85%, signaling near-universal benefits from improved road
access. The bicycle map offers the greatest additional insight. Whereas coverage for
significantly greater than zero returns was only 25% for bicycles (Figure 1a), Figure 2a
clearly shows that the scope of benefits is wide ranging and over 90% in over half of
parishes.
Next, we derive which asset offers the largest benefits for each parish. In Figure
3a we map which asset, other than road access – which is nearly-universally the highest
return investment – is expected to generate the maximum marginal expected benefit.
Similarly, Figure 3b shows which asset is expected to generate the maximum proportion
of positive marginal expected returns, by parish. Consistent with the observed high
magnitude and near-universality of benefits to improved road access, private transport 12 Kam et al. (2005) also find that estimated returns to assets can be negative in some areas.
19
assets dominate these maps. Motor vehicle ownership and motorcycle ownership often
offer the maximum expected return. Motor vehicle ownership and bicycle ownership are
the most prominent assets on the map of maximum proportion of positive household-
level expected returns. These targeting maps make clear the high magnitude and spatial
extent of expected benefits to improved transport systems in rural Uganda.
It should not be the least bit surprising that motor vehicles offer the highest
average marginal return in a large number of parishes; motor vehicles are extremely
valuable and expensive. These targeting maps depict estimated marginal gross returns;
information about the costs of supplying different assets has been conspicuously absent
from our analysis thus far. In order to address this deficiency and enable explicit benefit-
cost comparisons (albeit simplistically and incompletely), we compare estimated benefits
with estimated costs based on the mean price of livestock purchased or sold, as reported
in the household survey (values of other assets are unavailable in the data).13 Cost data
do not include the marginal costs of maintaining stocks, thus total costs would be higher.
Table 5 presents the findings.
Because it is unclear what time horizon the stream of benefits would accrue, we
take an extremely conservative approach and report only the expected increase in
expenditure for a single month. Since durable assets typically affect monthly
expenditures over a period of many months, depending on their rate of depreciation, this
necessarily understates the benefits, in some cases by orders of magnitude. This
13 The expected household marginal benefit was calculated with the following formula,
)( )(ln)(ln yey eehsehmb where ehmb is the expected household marginal benefit, hs is the average household size (in adult male equivalents), y is the average household monthly expenditure, and e is the estimated average marginal return (like those displayed in Figure 1). Expected household marginal benefit is the expected increase in monthly expenditure per household that receives a one unit asset transfer.
20
extremely conservative approach underscores, however, the considerable marginal
returns to investment in rural Uganda. Three of the four livestock assets clearly pass a
benefit-cost test, and the other (cattle) passes for time horizons of three years or more,
even just one year in areas with expected returns on the high end of the distribution.
While detailed exploration of the behavioral and institutional reasons for these findings is
beyond the scope of this methodological paper, the results clearly underscore apparent
underinvestment in productive assets in rural Uganda. Targeting maps of this sort can
help development agencies identify best bet forms for asset transfers given such apparent
underinvestment, and especially preferred geographic locations for a specific asset
transfer program (e.g., livestock), since the costs of provision typically vary only
modestly across space for a given asset.
Beyond looking at estimated marginal benefits of an asset, we examine how those
benefits relate to existing holdings of that asset and to the poverty headcount rate by
parish.14 The correlation between benefits and holdings explores whether there are
positive or negative network externalities associated with each asset, i.e., are marginal
returns increasing or decreasing in total parish holdings. The correlation between the
marginal returns to an asset and the poverty rate reveals prospective tradeoffs or
synergies between efficiency and equity objectives.
The central column of Table 6 shows the correlations between estimated average
marginal returns and asset holdings for all targetable assets, while Table 7 presents
14 The poverty headcount rate is the percentage of the population that is poor. In Uganda, a household is deemed poor if their estimated monthly expenditure falls below the expenditure thresholds set by Emwanu et al. (2007). As a check on our method, we compare our poverty estimates to those previously estimated for Uganda using the same data from Emwanu et al. (2007), who estimated the poverty headcount rate at the sub-county level. The correlation between the two estimated poverty headcount rates is 0.64 and the rank correlation is 0.66. The poverty map created using our method is shown in Appendix 4.
21
analogous correlations with the estimated proportion of households with positive
marginal returns. Most correlations are qualitatively similar between the two measures of
estimated benefits. Significant positive network effects appear for literacy, mobile phone
ownership and road access, reflecting how these assets become more valuable when
others in the area already possess them. Conversely, congestion effects are evident with
respect to home and motorcycle ownership as well as for some livestock (chickens and
goats).
Tables 6 and 7 also display the correlations between estimated marginal benefits
to asset transfers and poverty. A positive (negative) correlation implies efficiency and
equity aims are mutually reinforcing (competing). The results point to chickens and
microfinance as good candidates for geographic targeting in the sense that poverty
headcount rates are positively correlated with estimates of both the magnitude and
breadth of benefits to asset transfers.
As the final step in illustrating the potential utility of targeting maps, we identify
parishes that might be especially strong candidates for receiving asset transfers. Explicit
decision criteria based on expected returns do not exist at present; targeting maps can fill
that void. We focus on a hypothetical chicken transfer program that a NGO might
implement, since chickens clearly pass a simple benefit-cost test and the expected
benefits are highest in areas of greatest poverty. We select parishes based on the
following three attributes: 1) expected average marginal returns to a chicken ≥0.13 and
statistically significantly greater than zero, 2) at least 90 percent of households have
positive expected marginal returns to chickens, and 3) a poverty headcount rate ≥0.51.15
A total of 58 parishes meet these criteria and are mapped in Figure 4. Of those, we 15 The numeric thresholds were chosen based on the mapping results; there were no a priori levels.
22
highlight two parishes that show particular promise for this sort of development
intervention. Bulumba parish (in the southeast) offers large expected marginal returns
and an intermediate amount of poverty. Moli parish (in the northwest), on the other hand,
has one of the highest poverty rates in the sample and above average expected marginal
returns. All households in both parishes have expected positive returns to chicken
transfers.
5 Conclusions
This paper presents a novel method that has the potential to greatly advance the
efficacy of asset-based, geographically targeted transfer schemes. We add to the
substantial literature of small-area estimation and go beyond estimating poverty and
begin to address the best means of alleviating it. Our method first estimates the marginal
returns to various assets and then creates a series of maps that can address a variety of
questions regarding the magnitude and scope of benefits and the efficient spatial
allocation of development programs. The results produced using Ugandan data are
promising; estimated and projected asset returns seem reasonable and show substantial
variation across space.
Continued work with additional inputs is needed to complement targeting maps.
First, even if a policy maker has a targeting map in hand, there are still unanswered
questions about net benefits to and final effects of various asset transfers. We addressed
some of these concerns with a limited benefit-cost analysis. A more thorough analysis
for all assets with more precise information on procurement costs is a natural and
23
straightforward exercise for agencies intending to implement a transfer scheme using
targeting maps as an input.
Second, while targeting maps estimate the best means to an end, policy makers
are most often interested in the end itself – i.e. poverty reduction. A natural extension of
the targeting maps method is to use panel data to determining the expected impacts of an
asset transfer program on poverty (or other outcome variables of interest).
The maps and other results produced in this paper serve mainly to demonstrate the
potential usefulness of this method. Our hope is that the method can be improved upon
and eventually implemented in development programming, complementing the well-
established use of poverty maps in less developed countries. The promise of these
methods might also help encourage organizers of household surveys and censuses to
better coordinate future questionnaires with poverty and targeting maps in mind.
24
References
Adato, M., Carter, M. R. and May, J., 2006. “Exploring poverty traps and social exclusion in South Africa using qualitative and quantitative data” Journal of Development Studies, 42(2), 226-247. Baker, J. L. and Grosh, M. E., 1994. “Poverty Reduction Through Geographic Targeting: how well does it work?” World Development, 22(7), 983-995. Bigman, D. and Fofack, H., 2000. “Geographical Targeting for Poverty Alleviation: An introduction to the special issue” World Bank Economic Review, 14(1), 129-145. Carter, M. R. and Barrett, C. B., 2006 “The economics of poverty traps and persistent poverty: an asset based approach” Journal of Development Studies, 42(2), 178-199. Coady D,2006 “The welfare returns to finer targeting: The case of the Progresa program in Mexico” International Tax and Public Finance, 13, 217-239. Coady D, Grosh, M and Hoddinott J., 2004 “Targeting Outcomes Redux” World Bank Research Observer, 19(1), 61-85. Currie, J and Gahvari, F, 2008 “Transfers in cash and in-kind: Theory meets the data” Journal of Economic Literature, 46(2), 333-383. Demombynes, G. and Ozler, B., 2005. “Crime and local inequality in South Africa” Journal of Development Economics, 76, 265-292. Demombynes, G., Elbers, C., Lanjouw, J. O., and Lanjouw, P., 2007. “How good a map? Putting small area estimation to the test” World Bank working paper 4155. Doss, C., McPeak, J. and Barrett, C. B., 2008. “Interpersonal, intertemporal and spatial variation in risk perceptions: Evidence from East Africa” World Development, 36(8), 1453-1468. Easterly, W. and Levine, R., 1997. “Africa’s growth tragedy: Policies and ethnic divisions” Quarterly Journal of Economics, 112(4), 1203-1250. Elbers, C., Fujii, T., Lanjouw, P., Ozler, B., and Yin, W., 2007. “Poverty alleviation through geographic targeting: how much does disaggregation help?” Journal of Development Economics, 88, 198-213. Elbers, C., Lanjouw, J. O., and Lanjouw, P., 2003. “Micro-level estimation of poverty and inequality” Econometrica, 71(1), 355-364. Elbers, C., Lanjouw, J. O., and Lanjouw, P., 2005. “Imputed welfare estimates in regression analysis” Journal of Economic Geography, 5, 101-118.
25
Ellis, F and Freeman, A, 2004 “Rural livelihoods and poverty reduction strategies in four African countries” Journal of Development Studies, 40(4), 1-30. Emwanu, T., Okwi, P. O., Hoogeveen, J. G., Kristjanson, P., and Henninger, N., 2007. Nature, distribution and evolution of poverty and inequality in Uganda 1992-2002. Uganda Bureau of Statistics and the International Livestock Research Institute. Fan, S and Hazell, P, 2001. “Returns to public investments in the less-favored areas of India and China” American Journal of Agricultural Economics, 83(5), 1217-1222. Fan, S and Chan-Kang, C, 2004. “Returns to investment in less-favored areas in developing countries: a synthesis of evidence and implications for Africa” Food Policy, 29, 431-444. Hoffman, V, Barrett, C and Just, D, 2009 “Do free goods stick to poor households? Experimental evidence on insecticide treated bednets” World Development, 37(3), 607-617. Kam, S., Hossain, M., Lal Bose, M., and Villano L.S., 2005. “Spatial patterns of rural poverty and their relationship with welfare-influencing factors in Bangladesh” Food Policy, 30, 551-567. Moser, C, 1998 “The asset vulnerability framework: Reassessing urban poverty reduction strategies” World Development, 26(1), 1-19. Okwi, P.O., Ndeng’e, G., Kristjanson, P. Arunga, M., Notenbaert, A. Omolo, A., Henninger, N., Bensen, T., Kariuki, P., and Owuor, J., 2007. “Spatial determinants of poverty in rural Kenya”. Proceedings of the National Academy of Sciences of the United States of America, 104 (43), 16769-16774. Tarozzi, A. and Deaton, A., 2008. “Using Census and Survey data to estimate poverty and inequality for small areas” forthcoming, Review of Economics and Statistics.
26
Tables Table 1: Hierarchy of Ugandan administrative units and associated number of households
Administrative unit Total units
Number of households per unitmean median
District 56 90,797 79,024 County 163 31,194 27,650 Sub-county 958 5,308 4,584 Parish 5,234 971 818
Notes: Data come from the census and include both rural and urban households.
27
Table 2: Summary statistics Survey Census Number of households 5,648 4,376,978 Monthly household expenditure ( in Ugandan Shillings) 144,345 ‐ Variable (assets) Mean St. dev. Mean St. dev. Household head male 0.80 0.40 0.78 0.42 Household head education (years) 5.19 3.69 4.55 3.85 Household head age 41.19 13.77 41.62 16.26 Household head married 0.84 0.36 0.74 0.44 Proportion of household literate 0.45 0.25 0.45 0.32 Cattle 2.15 14.34 1.19 12.49 Goats 0.41 4.56 1.00 7.22 Pigs 0.12 1.30 0.15 1.24 Chicken 2.15 26.33 2.37 16.93 Land ownership (1=yes) 0.32 0.47 0.16 0.36 House ownership (1=yes) 0.91 0.29 0.85 0.36 Motor vehicle ownership (1=yes) 0.01 0.11 0.01 0.10 Motorcycle ownership (1=yes) 0.04 0.19 0.02 0.15 Bicycle ownership (1=yes) 0.53 0.50 0.35 0.48 Mobile phone ownership (1=yes) 0.03 0.17 0.03 0.16 Derived from census Population density (people per sq. km) 292.10 463.80 396.86 875.49 Ethnic diversity of parish 0.28 0.26 0.29 0.27 Religious diversity of parish 0.55 0.14 0.56 0.15 Derived from Community Survey Existence of livestock market in parish (1=yes) 0.26 0.44 0.30 0.46 Existence of crop market in parish (1=yes) 0.51 0.50 0.51 0.50 Microfinance access (1=yes) 0.80 0.40 0.79 0.41 Road access index 1.10 0.27 1.12 0.42 Existence of cattle rustling in parish (1=yes) 0.23 0.42 0.22 0.41 Existence of rebel activity in parish (1=yes) 0.15 0.36 0.17 0.38 Existence of drought in parish (1=yes) 0.86 0.34 0.85 0.36 Existence of animal disease in parish (1=yes) 1.00 0.05 0.99 0.11 Existence of crop disease in parish (1=yes) 0.99 0.08 0.99 0.12 Existence of animal extension services in parish (1=yes) 0.92 0.26 0.91 0.28 Existence of crop extension services in parish (1=yes) 0.94 0.24 0.93 0.26 Existence of human epidemic in parish 0.96 0.20 0.95 0.21
28
(Table 2 continued) Derived from GIS Average distance to secondary school in parish (km) 4.32 4.03 4.73 4.89 Average distance to Kampala in parish (km) 185.30 96.89 187.29 99.35 Average distance to an urban area in parish (km) 15.64 10.63 15.98 11.40 Average distance to freshwater in parish (km) 1.92 3.48 1.85 3.16 Percentage of parish agro-ecological zone 1 0.08 0.26 0.08 0.26 Percentage of parish agro-ecological zone 2 0.10 0.30 0.11 0.30 Percentage of parish agro-ecological zone 3 0.38 0.47 0.38 0.46 Average annual temperature (°C) 21.83 2.02 21.86 2.02 Mean Diurnal Range Temperature 12.03 0.65 12.03 0.71 Temperature seasonality 65.79 26.68 66.49 27.54 Annual temperature range 14.65 1.26 14.66 1.34 Average annual total precipitation (mm) 1226.57 178.95 1224.53 182.93 Average precipitation in driest month (mm) 34.29 15.22 34.32 16.12 Precipitation seasonality 42.75 6.51 43.05 6.77
Table 3: Geographic coverage of data (percent) Administrative
unit Survey Community
census Census District 100 100 100 County 100 100 100 Sub-county 56.3 100 100 Parish 11.5 98.4 100
Table 4: Average standard deviation of estimated average marginal returns
Administrative unit Bicycle Chicken
Microfinance access
Road access index
Average over all assets
County 0.28 0.03 0.46 2.41 0.53 Sub-county 0.23 0.02 0.32 1.19 0.38 Parish 0.17 0.01 0.11 0.37 0.23
29
Table 5: Simplified benefit cost analysis (all numbers in Ugandan Shillings)
Asset Cost Expected marginal monthly
benefit Average 95th percentile
Cattle 343,502 10,470 29,448 Chicken 4,848 26,087 65,820 Goats 23,920 54,066 116,162 Pigs 24,600 333,269 1,049,289
Table 6: Correlation of estimated average marginal returns
Asset Correlation with
Average existing holding in parish
Poverty headcount
Cattle -0.02 -0.08 Goats -0.18 -0.08 Pigs -0.06 0.12 Chicken -0.05 0.03 Land ownership 0.03 0.17 House ownership -0.42 0.02 Motor vehicle ownership 0.15 0.13 Motorcycle ownership -0.17 -0.05 Bicycle ownership 0.17 -0.02 Mobile phone ownership 0.37 0.08 Proportion of household literate 0.67 0.05 Existence of livestock market in parish 0.07 0.01 Existence of crop market in parish 0.15 -0.07 Microfinance access -0.09 0.06 Road access index 0.31 -0.13 Household head education (years) 0.29 -0.19
30
Table 7: Correlation of estimated proportion of households with positive marginal returns
Asset Correlation with
Average existing holding in parish
Poverty headcount
Cattle 0.02 -0.10 Goats -0.26 -0.19 Pigs 0.16 -0.05 Chicken -0.09 0.09 Land ownership 0.07 -0.10 House ownership -0.26 -0.15 Motor vehicle ownership 0.05 -0.05 Motorcycle ownership -0.34 0.23 Bicycle ownership 0.19 0.19 Mobile phone ownership 0.23 -0.11 Proportion of household literate 0.73 -0.32 Existence of livestock market in parish 0.04 -0.01 Existence of crop market in parish 0.05 -0.14 Microfinance access -0.06 0.05 Road access index 0.15 -0.05 Household head education (years) 0.15 -0.14
31
Figures Figure 1: Examples of maps of estimated average marginal returns that are significantly greater than zero for the given asset.
Fig. 1a. A map of estimated average marginal returns to bicycles that are significantly greater than zero.
Fig. 1c. A map of estimated average marginal returns to microfinance access that are significantly greater than zero.
Fig. 1b. A map of estimated average marginal returns to chickens that are significantly greater than zero.
Fig. 1d. A map of estimated average marginal returns to road access that are significantly greater than zero.
32
Figure 2: Examples of maps of proportion of households with estimated positive marginal return for the given asset.
Fig. 2a. A map of the proportion of households with estimated positive marginal returns to bicycles.
Fig. 2c. A map of the proportion of households with estimated positive marginal returns to microfinance access.
Fig. 2b. A map of the proportion of households with estimated positive marginal returns to chickens.
Fig. 2d. A map of the proportion of households with estimated positive marginal returns to road access.
33
Figure 3: Maximum asset returns
Fig. 3a. A map of which asset offers the largest expected average marginal return that is significantly greater than zero. In the legend, the number in parentheses indicates the number of parishes for which that asset offers the maximum return that is significantly greater than zero. Note that road access has been excluded from this analysis.
Fig. 3b. A map of which asset offers a positive expected marginal return to the largest proportion of households. In the legend, the number in parentheses indicates the number of parishes for which that asset offers the maximum proportion of households with expected positive return. Note that road access has been excluded from this analysis.
34
Figure 4: Sample targeting exercise