Targeting maps: An asset-based approach to geographic targeting *

Transcription

1 Targeting maps: An asset-based approah to geograph targeting * Corey Lang, ** Christopher B. Barrett and Felix Nashold Cornell University August 3, 2009 Abstrat Proper targeting of poly interventions requires reasonable estimates of the benefits of the various alternative interventions. In order to inform suh deisions, we develop an integrated approah that estimates the marginal returns to a range of assets aross geographally defined subpopulations allowing returns to vary by household and by geography. We then reate a series of maps illustrating the estimated marginal returns to speif assets and the proportion of an area s population that would benefit from inreased holdings of a speif asset. These maps an then be overlaid with traditional poverty maps to identify areas that are strong andidates for a partular development intervention. We develop a general method and demonstrate its potential with an applation using Ugandan data. JEL lassifation: R12, O2, C15, I32 Keywords: geograph targeting, assets, poverty maps, spatial variation, Uganda * We appreiate helpful disussions with Nany Johnson, GIS data assistane from John Owuor and Ugandan data adve from Thomas Emwanu. Useful omments were reeived from seminar partipants at Cornell University. This researh was made possible through the support of the International Livestok Researh Institute. ** Contat author. Address: Department of Eonoms, 404 Uris Hall, Ithaa, NY, addresses: CL395@ornell.edu (Lang), CBB2@ornell.edu (Barrett), FN23@ornell.edu (Nashold) 1

2 1 Introdution Improved targeting of development interventions has long been reognized as entral to ahieving greater impat from poverty redution efforts. However, effetive targeting requires reasonable estimates of where the returns to various programs are likely to be highest. Currently, no means exist for estimating and omparing expeted benefits aross spae and aross 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, seond, maps the estimated marginal returns reating a visual tool that an inform the targeting deisions of an in-kind transfer sheme. There are several methods of targeting, suh as a means test, ommunity-based targeting, ategoral or indator targeting and self-targeting, eah with its own advantages and disadvantages. 1 The empiral evidene suggests that geograph targeting is partularly effetive 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 Fofak 2000). The idea of geograph targeting is to determine a subset of geograph regions most in need and then transfer benefits to individuals within the hosen regions and exlude all others. The benefits to this method are intuitive as there is ample evidene that individuals living in lose geograph proximity tend to have similar livelihoods and fae the same onstraints and risks (e.g., Bigman and Fofak 2000, Doss et al. 2008). The major disadvantages to geograph targeting are that non-poor individuals living in targeted regions reeive benefits (leakage) and poor individuals not living in 1 Coady et al. (2004) disuss these targeting methods and more in detail. 2

3 targeted regions do not reeive benefits (underoverage). One remedy is to target more finely partitioned regions. As regions beome inreasingly disaggregated, within region heterogeneity dereases and targeting performane inreases (Elbers et al. 2007, Baker and Grosh 1994). A seond solution is to ombine geograph 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, Mexo s elebrated PROGRESA/Opportunidades program uses four (Coady 2006). In this paper, we build on the proven suesses of geograph targeting to propose an enhaned, asset-based approah. In general, transfers an be monetary or in-kind, where in-kind transfers usually ome in the form of subsidies for food, eduation, or health serves. Here, we explore the possibility of transfers from an entire range of private and publ assets, suh as livestok, mobile phones, means of transportation, and aess to roads or mrofinane institutions. Our fous on assets stems from the importane 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 esape from inome poverty (Carter and Barrett 2006). An obvious ritism of in-kind or asset transfers is that unlike with a ash transfer, a household is onstrained and annot onsume or invest in whatever they think will best help them. 2 While in-kind transfers an appear paternalist, there are several reasons why an asset-based approah ould perform better than a monetary approah. First, asset transfers an at as a natural self-seletion mehanism to redue leakage; 2 Currie and Gahvari (2008) review the debate over monetary versus in-kind transfers, though mainly from the perspetive of developed ountries. 3

4 whereas virtually everyone would aept a ash transfer, only those who benefit from a given asset would aept it as a transfer. Seond, in-kind transfers may stk to the targeted households better than ash beause of the well-established endowment effets assoiated with physal goods but not with ash. 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 ash transfers. Third, monetary transfers, due to their ready divisibility, may also be subjet to a high rate of soial taxation ompared to a lumpy asset, perhaps undoing efforts to ontrol leakage. Fourth, imperfet markets an make it diffult to proure speif, desired assets; this is a ommon rationale for in-kind food or seed aid in many remote or disaster-affeted regions. The targeting maps tool improves the information set informing geograph targeting. Given substantial spatial heterogeneity in poverty inidene and its auses (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 plaes in a ountry. Likewise, spatially heterogeneous asset valuation appears the norm, given the plaespeifity of many omplementary inputs e.g., agro-eologal onditions that affet livestok value, urban proximity that affets the returns to land, et. If poverty measures and the returns to assets both vary markedly aross spae for a variety of geograph, institutional, poly and tehnologal reasons, then it seems desirable to exploit the predtable omponent of suh variation in targeting development interventions. Previous researh has found onsiderable intra-regional variation in expeted returns to different development investments in Afra and Asia (Fan and Hazell 2001, Fan and Chan-Kang 2004). By ustomizing asset-based interventions to speif geograph areas, signifant 4

5 gains ould be made in effiently and ost-effetively addressing poverty. Our approah integrates spatially-explit estimates of the marginal benefits to multiple assets into a single framework suh that inter-asset omparisons of expeted marginal benefits an be made for eah region. The output an then be used as one of several omponents informing a targeted transfer plan. Our method draws on the small-area estimation tehnique pioneered by Elbers et al. (2003). Their method ombines detailed, nationally representative household survey data with national ensus data to estimate poverty rates at fine levels of disaggregation for an entire ountry. 3 First, they derive a relationship between household expenditure and various demograph and asset variables using the survey data. Seond, they predt out-of-sample estimates of expenditure for the ensus data using the oeffient estimates from the relationship derived with the survey data. By projeting expenditure estimates onto the full population, the Elbers et al. (2003) method enables estimation of poverty rates in plaes where no survey data exist and at finer levels of disaggregation than when using household survey data alone, as these are typally statistally representative only at relatively oarse sales of aggregation. One estimated, the poverty rates for the various regions of a ountry an be used to reate 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-inome ountries to guide poverty redution efforts. Poverty maps an signifantly bolster 3 All of the Foster-Greer-Thorbeke measures of poverty as well as inequality an be estimated using this method. 5

6 geograph targeting efforts beause, as mentioned above, geograph targeting methods are greatly improved as the geograph sale beomes finer. 4 Although poverty maps illustrate problems well and an failitate poly disussions, they offer no explit reommendation as to the best means of alleviating poverty. If a government is trying to reah a speif welfare target suh as the Millennium Development Goals, poverty maps an at best guide the government to regions with high poverty rates. What exatly the government should do in that region, however, remains unertain. Targeting maps address these shortomings by answering two general questions: 1) for a given region, whh asset building ativity will have the largest marginal gross benefit? and 2) for a given type of asset building ativity, in whh regions are the marginal gross benefits to suh an investment highest? Both of these questions address how to improve the effay of targeted, asset-based development programs. Answers to the first question are paramount for those wishing to ut poverty by the most effient means possible. The seond question appeals to groups interested in investments of a speif type, suh as Heifer International in building livestok holdings or The Nature Conservany in safeguarding natural resoures. With sare aid resoures available, targeting maps an help identify where the most bang-for-the-buk exists. The onstrution of targeting maps involves several distint steps similar to those involved in reating a poverty map. Using detailed household survey data and spatially explit environmental and infrastruture data, we apply multivariate regression and 4 Small-area poverty estimates an additionally be used in subsequent regression analysis as either the key dependent variable to investigate the auses of poverty (Kam et al. 2005, Okwi et al. 2007) or as an explanatory variable to investigate its onsequenes (Demombynes and Ozler 2005). 6

7 bootstrapping to estimate the returns to various assets and how these estimated returns vary aross spae. We then projet the parameter estimates onto the broader national ensus data and alulate the marginal returns as a funtion of projeted estimates and household asset holdings. Finally, we aggregate the estimated marginal returns aross households for small geograph areas and, using Geograph Information Systems (GIS), generate maps that highlight both the magnitude and sope of benefits. We illustrate our approah using Ugandan household survey and ensus data. The results are enouraging; estimated and projeted marginal benefits to asset transfers seem reasonable and show remarkable variation aross spae. Our results learly identify promising areas to target as well as indations of key assets to use in a geograph targeting sheme. These findings reinfore the value of geograph targeting and the importane of spatial analysis in general. The next setion desribes the methodology in detail, explains how it builds on poverty mapping, and disusses onerns with the framework. Setion three gives the speifs of the Ugandan data. Setion four reports the results inluding: several examples of types of targeting maps, a simplified benefit-ost analysis for several assets, and seletion of areas that would be strong andidates for a hypothetal asset transfer. Setion five onludes and disusses ideas for future work. 2 Method We estimate average expeted marginal household-level returns to various assets aross geographally defined subpopulations. In the ontext of this paper, assets will be taken as anything whose stok an affet a household s inome or expenditure. We 7

8 lassify assets along two dimensions: private vs. publ and targetable vs. non-targetable. Private and publ goods follow traditional definitions; publ goods are non-rival and non-exludable; private goods represent the rest. The distintion between targetable and non-targetable onerns whether an asset s quantity, quality or existene an be hanged by an intervention. This lassifation results in four ategories: private targetable assets (e.g., livestok holdings, literay, land holdings), publ targetable assets (e.g., soure of potable drinking water, aess to health lins, road aess), private non-targetable assets (e.g., eduation of household head, gender of household head) and publ 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 neessary to reate a targeting map are a nationally representative household survey and a ensus taken at about the same time. Additional environmental or publ good variables an and should be added when available to supplement both the survey and ensus data. In the first step of our analysis, we ompare the data available in the household survey and the ensus to generate a set of variables that are ommon to both data sets, suh as demograph variables, livestok and durable goods. We restrt the data in this way beause we must use a speifation that is replable in the ensus for all independent variables. The seond step is to estimate the relationship between per apita equivalent household expenditure and asset holdings, whh inlude the variables seleted in the first stage as well as relevant environmental and publ good variables. We assume that household expenditure is a funtion of asset holdings and plae-speif asset returns. 5 5 This speifation an be thought of as permanent or strutural inome (Carter and May 2001, Adato et al. 2006, Carter and Barrett 2006). 8

9 We remain agnost about the funtional form of the asset returns equation and model the relationship between expenditure and asset holdings using a seond order flexible funtional form. For household i in loation, letting y = per adult male equivalent household expenditure, A = private, targetable assets, A = plae speif means of the private, targetable assets, B = publ, targetable assets, Y = private, non-targetable assets, Y = plae speif means of the private, non-targetable assets, Z = publ, nontargetable assets, and assumed funtional form as: 6 X = additional household ontrol variables, we an write the ln y A ' R A Y ' R ( A X ( A, A, B, Y, B, Y, Z, Y, Z ) B ' R ) Z B ' R ( A Z ( A, B, Y, B, Y, Z, Z ) ) ' X (1) R j ( ) is a vetor of returns to asset type j = A, B, Y, Z, whh is the objet of estimation. The funtional form of asset returns implies that the expeted returns to eah asset an depend on the stok of every other asset. For example, the returns to a head of attle may depend on the household head s level of eduation, the average number of attle owned in that region, the existene of a nearby livestok market and/or loal preipitation levels. Plae speif asset means are only interated with household levels of the same variable (i.e., average attle holding is interated with eah household s attle holdings, but not with eah household s pig holdings or mobile phone ownership). Further, we assume the error term is omposed of a loation omponent and a household speif omponent. (M) 'M (2) where M A, B, Y, Z ]. [ 6 The plae speif means, A and Y, are derived from the ensus. 9

10 Our prinipal goal in this seond step in onstruting the targeting map is to aurately estimate the oeffients in the expenditure asset relationship. We bootstrap 200 iterations of the regression, using weighted least squares (weighted by population expansion fators) with errors lustered at the enumeration area level. We save the oeffient estimates from eah iteration of the bootstrapped regressions. Having thus estimated the shape of asset returns (many times), in the third step we projet the estimated oeffients from the first stage regressions onto the ensus data. Ultimately, however, we are not interested in the oeffient point estimates, but in the expeted marginal household-level return for a given targetable asset, k: E[ln y A k ] Rˆ ( ) ˆ A RB ( ) A ' B ' Y A A k k Rˆ Y ( ) ' Z A k Rˆ Z ( ) ' A k (3) For eah iteration of the bootstrap, we projet the oeffient estimates onto the ensus data and alulate the derivatives for all targetable assets. Aggregating all of the estimates generates an empiral distribution of estimated marginal household-level returns to speif assets. The mean estimated marginal return aross bootstrapped iterations yields our best estimate of a household s expeted marginal return for eah asset. We then aggregate households over geographally defined areas and alulate statists fundamental to the final produt. First, we ompute the mean and standard error of the expeted marginal returns for every geograph area and determine whh areas have returns that are statistally signifantly greater than zero. The estimated average marginal returns and their statistal signifane inform essential questions about the expeted magnitude of average benefits assoiated with speif asset transfers in partular areas. Seond, we alulate the proportion of households with positive 10

11 expeted marginal returns for every geograph area, whh reflets the sope of benefits from speif asset transfers in partular areas. Finally, using GIS tehniques, we display the results. Unlike with poverty mapping, no one map an summarize all of the results; instead this targeting method requires a series of maps. One map an display the most benefial asset, as judged either by the highest expeted average marginal returns of any asset or the highest proportion of positive expeted marginal returns of any asset, for eah geograph area. This map would address question one above: for a given region, whh asset building ativity will have the largest marginal gross benefit? Then, maps an be made for eah asset, showing either the expeted average marginal returns or the proportion of households with positive expeted marginal returns to that asset for eah geograph area. These maps would address question two above: for a given type of asset building ativity, in whh regions are the marginal gross benefits to suh an investment highest? Two estimated objets, two broad targeting questions, and many assets make for a large number of maps, eah atering to a different audiene or targeting question. 2.1 Comparing our method with poverty mapping methods No standard poverty mapping methodology exists, but there are ommon prates from whh our method deviates slightly, thus it is useful to ontrast and justify our approah. One ommon prate is to partition the data into the smallest regions for whh the survey data are statistally representative and run regressions for eah 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 Afra into 11

12 nine provines. The idea behind this step is to allow oeffient estimates to vary over spae. In ontrast, we pool all survey data into a single regression. While our method does not allow oeffient estimates to vary over spae, asset returns an vary dramatally over spae via the large number of plae-speif interation terms. Our motivation for this hoe is to explitly take into aount the influene of plae-speif haraterists on asset returns. If the geograph sope of regressions is limited, the variation in some variables, espeially the plae-speif variables suh as limate, is neessarily very limited. This onstraint ould lead to biased and inonsistent parameter estimates. Another ommon methodologal step in poverty mapping is to use stepwise regression to redue 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 whh variables belong on the right hand side, this method an be used to iteratively delete variables based on a riterion suh as adjusted-r 2 or t-statists. This approah would likely exlude several asset variables, both limiting the sope of inter-asset omparisons and potentially biasing the estimated returns of inluded assets (via omitted relevant variable bias). Thus, we take a more strutured approah and plae priority on the inlusion of all asset variables. The most ommon way to estimate the error surrounding the poverty estimates is to use parametr bootstrapping (Elbers et al. 2003, Demombynes et al. 2007). Parametr bootstrapping projets oeffient estimates onto ensus households by taking random draws from the distribution defined by a single set of regression oeffient estimates and their assoiated ovariane matrix. The poverty status of individual 12

13 households are then averaged by geograph areas. This proess is repeated many times to obtain a distribution of eah area s poverty. We hoose instead to bootstrap the first stage estimation in order to redue bias in the estimates, sine our method puts a greater premium on the regression oeffient estimates themselves. 2.2 Endogeneity onerns The major pitfall of our targeting maps methodology is the obvious endogeneity of several asset variables, whh an affet results in several ways. First, there is the bas, natural orrelation between expenditure and assets. Ideally, we ould use an aurate measure of inome as the key left hand side variable of interest, but of ourse suh a measure rarely exists (and this would not fully assuage endogeneity onerns anyway). As a result, there ould be a natural positive orrelation between asset holdings and expenditure in order to aquire assets, one must spend money. On the flip side, there may be a negative orrelation between assets and expenditure in a stat setting; for example, if a household has just sold lots of livestok, it may have inreased onsumption but lower ex post livestok holdings. An additional onfounding fator is that not all asset ownership is motivated by urrent produtive value. Some assets are aquired not beause they will produe more urrent expenditure, but beause they enhane welfare in some other way or at some future date. For example, some livestok may be held for risk prevention or soial status, and eduational investments are aimed at inreasing future, not urrent, inome. Estimated returns to assets may be additionally biased due either to omitted relevant variables or unobserved heterogeneity. Differenes in preferenes, ability and 13

14 various idiosynrat features speif to households are all potential soures of bias. We inlude in our speifations a rh set of plae-speif ovariates and a omplete set of interation terms in order to pk up as muh variation as possible and diminish the effet of this kind of bias. These are learly serious onerns. They are inherent, however, to any analysis that tries to answer the questions posed above. Short of running hundreds or thousands of idental field experiments aross vast spaes, there is no pratal way to estimate marginal returns to multiple assets aross a large geographal spae with ironlad identifation. While it is impossible to argue a purely ausal relationship, knowing how households hange their asset portfolios as their welfare inreases and how welfare is related to the environment and infrastruture around them an nonetheless provide quite useful insights to inform development poly. Given the onsiderable poly and operational importane of the questions targeting maps aim to address, we think this tradeoff is aeptable and hope and expet that future researh an 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 ommunity survey, all administered by the Ugandan Bureau of Statists (UBOS). The household survey and ensus are stratified by four regions (Central, East, North, and West) and an urban-rural split. For the purposes of this paper, we restrt our attention to rural households only (5,648 households in the survey and nearly 4.4 million in the ensus), although urban 14

15 households ould be inluded easily. The hierarhy for Ugandan administrative units, from largest to smallest, is nation, distrt, ounty, sub-ounty, and parish. Table 1 lists how many administrative units of eah type exist and the average and median number of households in eah unit. There are one or two enumeration areas (EA) per parish. The household survey lustered observations at the EA level and randomly sampled households within the EA. The private asset variables all ome from the household survey and the ensus. 7 We use the ensus, the ommunity survey and several GIS layers to reate loation speif publ asset variables. From the ensus, we alulate measures of ethn and religious diversity and population density, as well as means of all variables at the parish level. 8 The ommunity survey inludes information on drought, livestok and rop extension serves, markets, mrofinane and violene against women. These variables are aggregated to both the parish and sub-ounty level, neessary sine not all parishes ontained a ommunity that was surveyed. In addition, we use GIS to derive variables suh as average distane to seondary shool, average distane to urban areas, average distane to water, proportion of a region omposed of various agro-eologal zones, average annual rainfall, average variation in rainfall, average annual temperature and average variation in temperature, among others. 9 Data layers for shool loation, urban areas, agro-eologal zones and water loation were provided by the International Livestok Researh Institute (ILRI). Weather data were downloaded from 7 As stated above, we are onstrained to only use variables that appear in both the ensus and the survey. There are several instanes where a variable that would be fantast to inlude (e.g., mosquito net overage of all household members) is only ontained in one data set. This undersores the importane of planning and oordinating household surveys and ensuses. 8 Diversity is alulated (as in Easterly and Levine 1997) as the probability that two people of different ethnity/religion meet. 9 Eulidean, or straight-line, distane is used. 15

16 at a resolution of 30 ar-seonds. These geograph variables are aggregated at the sub-ounty level, due to limitations with the GIS software. 10 Table 2 gives summary statists for eah asset variable for the survey and ensus. One all interation and seond order variables are added, there are a total of 1120 right hand side variables in our speifation of equation 1. Appendix 1 lists all variables used. In addition to numeral omparability of the data, geograph omparability is important. Table 3 gives the perentage of eah administrative unit represented in the survey and ommunity ensus and Appendix 2 shows the geograph loation of the survey data. The survey data appear well dispersed and thus we have onfidene that our estimates are representative of many different geographies. 4 Results Appendix 3 presents omplete results from the bootstraped regressions. As a first step in analyzing the results, we determine the appropriate level of aggregation for the expeted marginal returns. In standard poverty mapping exerises, there is a tradeoff between geograph aggregation and preision (Elbers et al. 2003). The goal is to aggregate households into the smallest possible geograph area without sarifing preision, whh enables inter-regional omparison. We aggregate derivatives and alulate means and standard errors of all targetable assets at three different administrative levels: ounty, sub-ounty, and parish. Table 4 gives the estimated standard errors of four assets byles, hkens, mrofinane 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 statists ould not be alulated for all parishes. 16

17 aess and road aess 11, whh we use as examples throughout as well as the average aross all targetable assets. Clearly, as the area of aggregation grows so does the standard error. This finding ontrasts with the standard inverse relationship found in poverty mapping due to the differene in our method, whh first estimates household level marginal returns via simulation and then aggregates over geograph areas. Our error estimates are a omposite of ordinary impreision plus inter-household variation. As the geograph sale grows, more inter-household heterogeneity is introdued and the standard errors inrease. The empiral findings learly indate that parish is the appropriate level of aggregation for our estimates, espeially sine the effieny of geograph targeting inreases as the geograph area dereases in size (Baker and Grosh 1994, Elbers et al. 2007). Figures 2a-2d plot the estimated average marginal returns that are signifantly greater than zero for hkens, byles, aess to mrofinane and road aess, respetively, at the parish level. The magnitude of the returns is not readily apparent; the units are the expeted additional natural log of monthly per adult male equivalent expenditure assoiated with an additional unit of that asset. For purposes of omparison, the average of the natural log of per adult male equivalent monthly expenditure aross the whole household survey sample is The ordinality of the magnitudes of returns seems reasonable with road aess being most valuable, then mrofinane aess, then byle ownership and finally hkens. There is also onsiderable spatial variation in the estimates. We see pokets of high returns, like those in the northwestern Uganda for 11 Mrofinane aess is a binary variable indating whether at least one ommunity within a parish indated having aess to mrofinane. Road aess 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 ommunities surveyed within that parish. 17

18 hken, and spatial patterns, suh as the remarkable regularity in whh the returns to road aess inrease with proximity to urban areas. For eah asset, a onsiderable portion of the ountry does not exhibit statistally signifant returns, refleting both relatively large standard errors and several negative point estimates. It is reasonable that some returns are atually negative beause we estimate marginal returns omprehensively, inluding areas that are ompletely unsuitable for ertain assets. 12 Figures 2a-2d present the proportion of households with estimated marginal returns greater than zero for hkens, byles, aess to mrofinane and road aess, respetively, at the parish level. The hken map largely reinfores the information ontained in Figure 1b; areas with high returns, notably the northwest, also have a large proportion of households with statistally signifant positive expeted returns. The road aess map offers little way to distinguish between areas as almost all areas have a benefiary proportion over 85%, signaling near-universal benefits from improved road aess. The byle map offers the greatest additional insight. Whereas overage for signifantly greater than zero returns was only 25% for byles (Figure 1a), Figure 2a learly shows that the sope of benefits is wide ranging and over 90% in over half of parishes. Next, we derive whh asset offers the largest benefits for eah parish. In Figure 3a we map whh asset, other than road aess whh is nearly-universally the highest return investment is expeted to generate the maximum marginal expeted benefit. Similarly, Figure 3b shows whh asset is expeted to generate the maximum proportion of positive marginal expeted returns, by parish. Consistent with the observed high magnitude and near-universality of benefits to improved road aess, private transport 12 Kam et al. (2005) also find that estimated returns to assets an be negative in some areas. 18

19 assets dominate these maps. Motor vehle ownership and motoryle ownership often offer the maximum expeted return. Motor vehle ownership and byle ownership are the most prominent assets on the map of maximum proportion of positive householdlevel expeted returns. These targeting maps make lear the high magnitude and spatial extent of expeted benefits to improved transport systems in rural Uganda. It should not be the least bit surprising that motor vehles offer the highest average marginal return in a large number of parishes; motor vehles are extremely valuable and expensive. These targeting maps dept estimated marginal gross returns; information about the osts of supplying different assets has been onspuously absent from our analysis thus far. In order to address this defieny and enable explit benefitost omparisons (albeit simplistally and inompletely), we ompare estimated benefits with estimated osts based on the mean pre of livestok purhased or sold, as reported in the household survey (values of other assets are unavailable in the data). 13 Cost data do not inlude the marginal osts of maintaining stoks, thus total osts would be higher. Table 5 presents the findings. Beause it is unlear what time horizon the stream of benefits would arue, we take an extremely onservative approah and report only the expeted inrease in expenditure for a single month. Sine durable assets typally affet monthly expenditures over a period of many months, depending on their rate of depreiation, this neessarily understates the benefits, in some ases by orders of magnitude. This 13 The expeted household marginal benefit was alulated with the following formula, (ln y e) (ln y) ehmb hs ( e e ) where ehmb is the expeted 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). Expeted household marginal benefit is the expeted inrease in monthly expenditure per household that reeives a one unit asset transfer. 19

20 extremely onservative approah undersores, however, the onsiderable marginal returns to investment in rural Uganda. Three of the four livestok assets learly pass a benefit-ost test, and the other (attle) passes for time horizons of three years or more, even just one year in areas with expeted returns on the high end of the distribution. While detailed exploration of the behavioral and institutional reasons for these findings is beyond the sope of this methodologal paper, the results learly undersore apparent underinvestment in produtive assets in rural Uganda. Targeting maps of this sort an help development agenies identify best bet forms for asset transfers given suh apparent underinvestment, and espeially preferred geograph loations for a speif asset transfer program (e.g., livestok), sine the osts of provision typally vary only modestly aross spae 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 headount rate by parish. 14 The orrelation between benefits and holdings explores whether there are positive or negative network externalities assoiated with eah asset, i.e., are marginal returns inreasing or dereasing in total parish holdings. The orrelation between the marginal returns to an asset and the poverty rate reveals prospetive tradeoffs or synergies between effieny and equity objetives. The entral olumn of Table 6 shows the orrelations between estimated average marginal returns and asset holdings for all targetable assets, while Table 7 presents 14 The poverty headount rate is the perentage 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 hek on our method, we ompare our poverty estimates to those previously estimated for Uganda using the same data from Emwanu et al. (2007), who estimated the poverty headount rate at the sub-ounty level. The orrelation between the two estimated poverty headount rates is 0.64 and the rank orrelation is The poverty map reated using our method is shown in Appendix 4. 20

21 analogous orrelations with the estimated proportion of households with positive marginal returns. Most orrelations are qualitatively similar between the two measures of estimated benefits. Signifant positive network effets appear for literay, mobile phone ownership and road aess, refleting how these assets beome more valuable when others in the area already possess them. Conversely, ongestion effets are evident with respet to home and motoryle ownership as well as for some livestok (hkens and goats). Tables 6 and 7 also display the orrelations between estimated marginal benefits to asset transfers and poverty. A positive (negative) orrelation implies effieny and equity aims are mutually reinforing (ompeting). The results point to hkens and mrofinane as good andidates for geograph targeting in the sense that poverty headount rates are positively orrelated 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 espeially strong andidates for reeiving asset transfers. Explit deision riteria based on expeted returns do not exist at present; targeting maps an fill that void. We fous on a hypothetal hken transfer program that a NGO might implement, sine hkens learly pass a simple benefit-ost test and the expeted benefits are highest in areas of greatest poverty. We selet parishes based on the following three attributes: 1) expeted average marginal returns to a hken 0.13 and statistally signifantly greater than zero, 2) at least 90 perent of households have positive expeted marginal returns to hkens, and 3) a poverty headount rate A total of 58 parishes meet these riteria and are mapped in Figure 4. Of those, we 15 The numer thresholds were hosen based on the mapping results; there were no a priori levels. 21

22 highlight two parishes that show partular promise for this sort of development intervention. Bulumba parish (in the southeast) offers large expeted 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 expeted marginal returns. All households in both parishes have expeted positive returns to hken transfers. 5 Conlusions This paper presents a novel method that has the potential to greatly advane the effay of asset-based, geographally targeted transfer shemes. 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 reates a series of maps that an address a variety of questions regarding the magnitude and sope of benefits and the effient spatial alloation of development programs. The results produed using Ugandan data are promising; estimated and projeted asset returns seem reasonable and show substantial variation aross spae. Continued work with additional inputs is needed to omplement targeting maps. First, even if a poly maker has a targeting map in hand, there are still unanswered questions about net benefits to and final effets of various asset transfers. We addressed some of these onerns with a limited benefit-ost analysis. A more thorough analysis for all assets with more preise information on prourement osts is a natural and 22

23 straightforward exerise for agenies intending to implement a transfer sheme using targeting maps as an input. Seond, while targeting maps estimate the best means to an end, poly makers are most often interested in the end itself i.e. poverty redution. A natural extension of the targeting maps method is to use panel data to determining the expeted impats of an asset transfer program on poverty (or other outome variables of interest). The maps and other results produed in this paper serve mainly to demonstrate the potential usefulness of this method. Our hope is that the method an be improved upon and eventually implemented in development programming, omplementing the wellestablished use of poverty maps in less developed ountries. The promise of these methods might also help enourage organizers of household surveys and ensuses to better oordinate future questionnaires with poverty and targeting maps in mind. 23

24 Referenes Adato, M., Carter, M. R. and May, J., Exploring poverty traps and soial exlusion in South Afra using qualitative and quantitative data Journal of Development Studies, 42(2), Baker, J. L. and Grosh, M. E., Poverty Redution Through Geograph Targeting: how well does it work? World Development, 22(7), Bigman, D. and Fofak, H., Geographal Targeting for Poverty Alleviation: An introdution to the speial issue World Bank Eonom Review, 14(1), Carter, M. R. and Barrett, C. B., 2006 The eonoms of poverty traps and persistent poverty: an asset based approah Journal of Development Studies, 42(2), Coady D,2006 The welfare returns to finer targeting: The ase of the Progresa program in Mexo International Tax and Publ Finane, 13, Coady D, Grosh, M and Hoddinott J., 2004 Targeting Outomes Redux World Bank Researh Observer, 19(1), Currie, J and Gahvari, F, 2008 Transfers in ash and in-kind: Theory meets the data Journal of Eonom Literature, 46(2), Demombynes, G. and Ozler, B., Crime and loal inequality in South Afra Journal of Development Eonoms, 76, Demombynes, G., Elbers, C., Lanjouw, J. O., and Lanjouw, P., How good a map? Putting small area estimation to the test World Bank working paper Doss, C., MPeak, J. and Barrett, C. B., Interpersonal, intertemporal and spatial variation in risk pereptions: Evidene from East Afra World Development, 36(8), Easterly, W. and Levine, R., Afra s growth tragedy: Polies and ethn divisions Quarterly Journal of Eonoms, 112(4), Elbers, C., Fujii, T., Lanjouw, P., Ozler, B., and Yin, W., Poverty alleviation through geograph targeting: how muh does disaggregation help? Journal of Development Eonoms, 88, Elbers, C., Lanjouw, J. O., and Lanjouw, P., Mro-level estimation of poverty and inequality Eonometra, 71(1), Elbers, C., Lanjouw, J. O., and Lanjouw, P., Imputed welfare estimates in regression analysis Journal of Eonom Geography, 5,

25 Ellis, F and Freeman, A, 2004 Rural livelihoods and poverty redution strategies in four Afran ountries Journal of Development Studies, 40(4), Emwanu, T., Okwi, P. O., Hoogeveen, J. G., Kristjanson, P., and Henninger, N., Nature, distribution and evolution of poverty and inequality in Uganda Uganda Bureau of Statists and the International Livestok Researh Institute. Fan, S and Hazell, P, Returns to publ investments in the less-favored areas of India and China Ameran Journal of Agrultural Eonoms, 83(5), Fan, S and Chan-Kang, C, Returns to investment in less-favored areas in developing ountries: a synthesis of evidene and implations for Afra Food Poly, 29, Hoffman, V, Barrett, C and Just, D, 2009 Do free goods stk to poor households? Experimental evidene on insetide treated bednets World Development, 37(3), Kam, S., Hossain, M., Lal Bose, M., and Villano L.S., Spatial patterns of rural poverty and their relationship with welfare-influening fators in Bangladesh Food Poly, 30, Moser, C, 1998 The asset vulnerability framework: Reassessing urban poverty redution strategies World Development, 26(1), Okwi, P.O., Ndeng e, G., Kristjanson, P. Arunga, M., Notenbaert, A. Omolo, A., Henninger, N., Bensen, T., Kariuki, P., and Owuor, J., Spatial determinants of poverty in rural Kenya. Proeedings of the National Aademy of Sienes of the United States of Amera, 104 (43), Tarozzi, A. and Deaton, A., Using Census and Survey data to estimate poverty and inequality for small areas forthoming, Review of Eonoms and Statists. 25

26 Tables Table 1: Hierarhy of Ugandan administrative units and assoiated number of households Administrative Total units Number of households per unit unit mean median Distrt 56 90,797 79,024 County ,194 27,650 Sub-ounty 958 5,308 4,584 Parish 5, Notes: Data ome from the ensus and inlude both rural and urban households. 26

27 Table 2: Summary statists 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 Household head eduation (years) Household head age Household head married Proportion of household literate Cattle Goats Pigs Chken Land ownership (1=yes) House ownership (1=yes) Motor vehle ownership (1=yes) Motoryle ownership (1=yes) Byle ownership (1=yes) Mobile phone ownership (1=yes) Derived from ensus Population density (people per sq. km) Ethn diversity of parish Religious diversity of parish Derived from Community Survey Existene of livestok market in parish (1=yes) Existene of rop market in parish (1=yes) Mrofinane aess (1=yes) Road aess index Existene of attle rustling in parish (1=yes) Existene of rebel ativity in parish (1=yes) Existene of drought in parish (1=yes) Existene of animal disease in parish (1=yes) Existene of rop disease in parish (1=yes) Existene of animal extension serves in parish (1=yes) Existene of rop extension serves in parish (1=yes) Existene of human epidem in parish

28 (Table 2 ontinued) Derived from GIS Average distane to seondary shool in parish (km) Average distane to Kampala in parish (km) Average distane to an urban area in parish (km) Average distane to freshwater in parish (km) Perentage of parish agro-eologal zone Perentage of parish agro-eologal zone Perentage of parish agro-eologal zone Average annual temperature ( C) Mean Diurnal Range Temperature Temperature seasonality Annual temperature range Average annual total preipitation (mm) Average preipitation in driest month (mm) Preipitation seasonality Table 3: Geograph overage of data (perent) Administrative unit Survey Community ensus Census Distrt County Sub-ounty Parish Table 4: Average standard deviation of estimated average marginal returns Administrative unit Byle Chken Mrofinane aess Road aess index Average over all assets County Sub-ounty Parish

29 Table 5: Simplified benefit ost analysis (all numbers in Ugandan Shillings) Expeted marginal monthly Asset Cost benefit Average 95th perentile Cattle 343,502 10,470 29,448 Chken 4,848 26,087 65,820 Goats 23,920 54, ,162 Pigs 24, ,269 1,049,289 Table 6: Correlation of estimated average marginal returns Correlation with Asset Average existing holding in parish Poverty headount Cattle Goats Pigs Chken Land ownership House ownership Motor vehle ownership Motoryle ownership Byle ownership Mobile phone ownership Proportion of household literate Existene of livestok market in parish Existene of rop market in parish Mrofinane aess Road aess index Household head eduation (years)

30 Table 7: Correlation of estimated proportion of households with positive marginal returns Correlation with Asset Average existing holding in parish Poverty headount Cattle Goats Pigs Chken Land ownership House ownership Motor vehle ownership Motoryle ownership Byle ownership Mobile phone ownership Proportion of household literate Existene of livestok market in parish Existene of rop market in parish Mrofinane aess Road aess index Household head eduation (years)

31 Figures Figure 1: Examples of maps of estimated average marginal returns that are signifantly greater than zero for the given asset. Fig. 1a. A map of estimated average marginal returns to byles that are signifantly greater than zero. Fig. 1b. A map of estimated average marginal returns to hkens that are signifantly greater than zero. Fig. 1. A map of estimated average marginal returns to mrofinane aess that are signifantly greater than zero. Fig. 1d. A map of estimated average marginal returns to road aess that are signifantly greater than zero. 31

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 byles. Fig. 2b. A map of the proportion of households with estimated positive marginal returns to hkens. Fig. 2. A map of the proportion of households with estimated positive marginal returns to mrofinane aess. Fig. 2d. A map of the proportion of households with estimated positive marginal returns to road aess. 32

33 Figure 3: Maximum asset returns Fig. 3a. A map of whh asset offers the largest expeted average marginal return that is signifantly greater than zero. In the legend, the number in parentheses indates the number of parishes for whh that asset offers the maximum return that is signifantly greater than zero. Note that road aess has been exluded from this analysis. Fig. 3b. A map of whh asset offers a positive expeted marginal return to the largest proportion of households. In the legend, the number in parentheses indates the number of parishes for whh that asset offers the maximum proportion of households with expeted positive return. Note that road aess has been exluded from this analysis. 33

34 Figure 4: Sample targeting exerise 34