选择贫困人口：秘鲁进行贫困人口地区定位的指标

POLICY RESEARCH WORKING PAPER 2477 Picking the Poor Geographic targeting of social programs to the poor has become increasingly Indicators for Geographic important in Peru. The Targeting in Peru potential payoffs of such targeting are large, and differences in outcomes with Norbert R. Schady different targeting indicators are small. The World Bank Latin America and the Caribbean Region Poverty Sector Unit November 2000  I POLIcY RESEARCH WORKING PAPER 2477 Summary findings Geographic targeting is perhaps the most popular He then conducts a series of simulations that estimate mechanism used to direct social programs to the poor in leakage rates; concentration curves; the impact of Latin America. transfers on poverty as measured by the headcount Schady empirically compares geographic targeting index, poverty gap, and PI measures of the Foster-Greer- indicators available in Peru. He combines household- Thorbecke family; and nonparametric (kernel) densities level information from the 1994 and 1997 Peru Living when transfers are based on alternative indicators. Standards Measurement Surveys and district-level He concludes that there is substantial potential for information from the 1993 Peru Population and Housing geographic targeting in Peru. The differences in Census. outcomes across geographic targeting indicators are small and not statistically significant. This paper-a product of the Poverty Sector Unit, Latin America and the Caribbean Region-is part of a larger effort in the region to explore the potential of geographic targeting. Copies of the paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contactTania Gomez, room 18-102, telephone 202-473-2127, fax 202- 522-0054, email address tgomez@worldbank.org. Policy Research Working Papers are also posted on the Web at www.worldbank.org/research/workingpapers. The author may be contacted at nschady@worldbank.org. November 2000. (23 pages) The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should he cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the view of the World Bank. its Executive Directors, or the comntries they represent. Produced by the Policy Research Dissemination Center  Picking the Poor: Indicators for Geographic Targeting in Peru Norbert R. Schady World Bank* I thank particularly Christina Paxson for much helpful advice. I have also benefited from comments made by Greg Felker, Francisco Ferreira, Jesko Hentschel, John Waterbury, and participants at a seminar held at the World Bank. I am indebted to Marcos Robles for help obtaining the necessary data from INEI, and to Vajeera Dorabawila for help with the consumption aggregates for the 1994 and 1997 LSMS. Please address correspondence to Norbert R. Schady, The World Bank, 1818 H Street N.W., Washington, DC 20433 (nschady@worldbank.org).   1. Introduction Lack of information is a serious constraint to targeting social programs effectively, especially in Less Developed Countries (LDCs) (Besley 1993; Ravallion 1993). Targeting social programs involves making distinctions between "deserving" (poor) and "undeserving" (non-poor) applicants. But this is no simple matter in countries where household characteristics such as income are rarely known. In such circumstances, policy-makers intent on targeting are forced to choose among imperfect solutions. They can rely on observable household characteristics, such as land ownership, the ratio of working age-adults to dependents, or ownership of durable goods that seem likely to separate poor from non-poor households. They can "self-target" programs by designing them so that they appeal mainly to the poor--perhaps by offering employment at below- market wages, or subsidizing foodstuffs consumed primarily by the poor. Or they can use "geographic targeting" to direct resources to areas in which, on average, poverty appears to be greatest (Akerlof 1978; Besley and Kanbur 1990; Grosh 1992). Geographic targeting is appealing because it is comparatively simple to administer. Different parts of a country--regions, provinces, districts, even city blocks--are ranked by some measure of deprivation. This measure could be income-based poverty or, more commonly, an indicator of health, educational or nutritional status, or access to basic services, such as electricity or running water. Resources are then allocated in inverse proportion to average welfare, so that poor regions receive higher per capita transfers than rich ones. Alternatively, rich areas can be excluded from the program altogether. The simplicity of geographic targeting is an important advantage when lack of information or administrative capacity is a serious concern. This paper compares a number of geographic targeting indicators that have been discussed by policy-makers in Peru. These include the infant mortality rate, which is used to target the Municipal Compensation Fund, the main block grant from the central government to local governments in Peru; a composite "poverty" index developed by the Peruvian Social Fund (FONCODES), which FONCODES uses to target its projects; and an estimate of "imputed" poverty which combines census and survey data in an attempt to approximate money-based measures of welfare. To test the potential impact on poverty of targeting with alternative indicators, I conduct a series of simulations which combine information on household expenditures from the 1994 and 1997 Peru Living Standards Measurement Surveys (LSMS) with district-level averages of various welfare measures. I use the results of the simulations to compare leakage rates, trace out concentration curves, estimate the impact on various poverty measures of the Foster-Greer- Thorbecke (FGT) family, and graph non-parametric (kernel) density estimates of the log of per capita expenditures (PCE) when transfers are made on the basis of alternative indicators. 2  The results show that geographic targeting with any of the indicators is a significant improvement over an untargeted regime in which resources are distributed equally across all districts. There is, therefore, substantial potential for geographic targeting in Peru. However, targeting outcomes are quite similar for all of the targeting indicators I consider, and differences in outcomes between indicators are not significant at conventional levels of significance. The rest of the paper proceeds as follows. Section 2 briefly describes alternative geographic targeting indicators available in Peru. Section 3 describes the methodology I use to compare indicators. It also describes altemative performance measures, identifies formulas that can be used to allocate resources to districts, describes data requirements, and states the assumptions made for the simulation exercise. Section 4 summarizes the results of the analysis. Section 5 draws conclusions. 2. Geographic targeting indicators in Peru In 1997 there were 13 administrative regions, 24 departments, 194 provinces, and 1812 districts in Peru (Webb and Fernandez Baca 1997, p. 112). Recent discussion about the geographic targeting of government programs has focused on the use of district-level averages. Districts can be quite small in Peru: According to the 1993 Population and Housing Census, the average district population was about 12,600 inhabitants, but some, predominantly rural districts had less than 200 inhabitants. Infant mortality The infant mortality rate measures the fraction of children ever born who do not reach the first year of age. In Peru, as in many other developing countries, one would be loathe to estimate infant mortality on the basis of seriously incomplete death registries. To address this problem, demographers have developed indirect methods to estimate mortality at early ages (Hill, Zlotnik and Trussell 1983; Brass and Macrae 1984 and 1985; Trussell and Menken 1984). The Peruvian National Statistical Institute (INEI) used one such method--the preceding births technique--to estimate infant mortality in Peru. The preceding births technique requires women to report the total number of children ever born and surviving at two different points in time. This information was available in Peru from the 1981 and 1993 population censuses. However, because the population of many districts is very small, applying the preceding births technique to individual districts would have produced highly uncertain results. INEI therefore followed a two-stage procedure. First, it estimated the infant mortality rate in every department, where the sample size was large enough. In addition, INEI regressed its estimate of the infant mortality rate in every department on some of its 3  correlates--indicators such as women's average education level, household characteristics, and place of residence on which inforrnation had been gathered in the 1993 census. In the second stage, the coefficients from these departmental regressions were applied to district-level data from the 1993 census to estimate district-level infant mortality rates (INEI 1997). The FONCODES index Peru has a long history of developing "poverty maps" based on composite indices of unmet basic needs. The first of these maps was constructed by Webb with information from the 1961 population census (Webb 1977). This poverty map was updated with new censuses conducted in 1972, 1981, and 1993 (Amat y Le6n n.d.; Banco Central de Reserva 1981; INEI 1994). FONCODES, in turn, has developed composite poverty maps since its creation in 1991, often with technical assistance from the Inter-American Development Bank, the World Bank, and the German development agency GTZ. The current district-level poverty map was developed by FONCODES and the Ministry of the Presidency. It is based on eight indicators--the rate of chronic malnutrition, illiteracy, school- aged children not in school, overcrowded housing, inadequate roofing, and the proportion of the population without access to water, sewerage, and electricity. All of these indicators except the rate of chronic malnutrition were estimated with data from the 1993 population census. The rate of chronic malnutrition was estimated from a census of height and weight among school-aged children, also conducted in 1993 (FONCODES 1995 and 1996). Composite indices invariably involve some arbitrary weighting of individual indicators. The 1996 index standardized each indicator by dividing it by its minimum value, multiplied the rate of chronic malnutrition by seven, and then added all of the individual indicators' For ease of interpretation, FONCODES then standardized the index by dividing all values by the lowest value. The resulting index ranges from a value of 1 to 36.38. Imputed poverty and income In Peru, there are no survey-based estimates of income or expenditures at a level more disaggregated than the department: for example, household surveys conducted by INEI, which generally have sample sizes of 15,000 to 20,000 households, can only be used to compare income i This procedure had the unintended consequence that the greatest weight was given to those indicators with the greatest variance. Thus, while the intended weights were 50% for the rate of chronic malnutrition, and 7.14% each for the seven other measures, the actual weights in the index turned out to be 15.3% for the rate of chronic malnutrition, and 3.4%, 2.2%, 3.0%, 38.3%, 8.8%, 7.4%, and 21.6% for the measures of illiteracy, school attendance, overcrowding, inadequate roofing, and access to water, sewerage and electricity, respectively (World Bank 1996, p. 7). 4  across "natural regions" and departments.2 But INEI has combined variables which are common to both the 1993 census and one such survey conducted in 1995 to develop imputed district-level measures of income and poverty (INEI 1996). This procedure is conceptually similar to that used to estimate infant mortality at the district level. INEI estimated income in 1995 on the basis of the household survey, and then regressed income in every department on its correlates--household composition, education levels, access to basic services such as water, sewerage and electricity, ownership of durable goods, such as television, radio and refrigerator, and other variables included in both the census and the survey. The coefficients from the 24 department-level regressions were then used to impute average income in every district, and the fraction of the population in each district below an income-based poverty line. 3. The analytic framework Measures ofperformance The simplest measure of targeting focuses on leakage and undercoverage rates (Grosh, pp. 16-17; Baker and Grosh 1994). A poverty line is chosen to separate "poor" from "non-poor". Leakage rates are then defined as the fraction of total program resources which go to the non- poor, and undercoverage rates as the fraction of the poor who do not benefit from the program. By this measure, better geographic targeting indicators result in lower leakage and lower undercoverage rates. A second approach simply ranks individuals by an indicator of welfare-say, per capita expenditures-and then cumulates the fraction of households and the fraction of resources transferred by different indicators. The results are often presented in terms of so-called "concentration curves" (see, for example, Milanovic 1995), and I follow this practice below. By this measure, the best targeting indicator is that whose concentration curve is above all others at every point. The concentration curve method does not require use of a poverty line, which may be an advantage given the fact that setting poverty lines can be contentious. Alternatively, one might want to compare the changes in various poverty measures which are likely to result when transfers are based on alternative targeting indicators (Chaudhuri and Ravallion 1994). This is an exercise in comparative statics: what is total poverty before and 2 These natural regions are Lima, and the urban and rural areas of the coast, sierra (highlands), and selva (3ungle), respectively. Natural regions do not, in general, correspond to the administrative regions mentioned before. 3 The methodology applied by INEI for these imputations is similar in spirit to that proposed in Hentschel et. al (2000). 5  immediately after the transfer? By this measure, the preferred geographic targeting indicator directs limited resources to areas where they would have the greatest short-term impact on poverty. More complex formulations, which might model the expected long-term returns from transfers to different districts, are beyond the scope of this paper (see the comments by Binswanger 1989, cited in Ravallion 1993). In what follows I use three poverty measures from the Foster-Greer-Thorbecke (FGT) family--the headcount index, the poverty gap, and the P2 measure (Foster, Greer, and Thorbecke 1984). The FGT family of poverty measures follows the general formulation below: N a (1) Pa= - (1i- yi /z) (forally, < z) where y is income, z the poverty line, and a is a parameter which represents the aversion to inequality. When a=O, Po corresponds to the headcount index--the number of people below the poverty line; when oa= 1, P1 corresponds to the poverty gap--a sum of the individual shortfalls in income for those below the poverty line, as a fraction of the poverty line itself. As a increases, the measure gives a greater weight to the poorest poor, and at very high values of a P,, approaches a "Rawlsian" measure of welfare which gives weight only to the poorest household. The P2 measure corresponds to a value of a=2. Finally, one could look at changes in the entire distribution of log PCE, rather than just at changes for those below the poverty line. I use non-parametric (kernel) density estimates for this purpose. Allocationformulas When there is no targeting, districts are simply allocated resources according to their share of the total population in the country, and everyone is assumed to receive the same per capita transfer. This no-targeting scenario serves as a benchmark to measure additional reductions in poverty that could be achieved when geographic targeting is conducted on the basis of some welfare indicator. To compare targeting indicators, one must develop a formula which allocates resources across districts. I consider one such formula, which is relevant for Peru because it has been the basis of targeting by FONCODES, the program which made the most significant early advances developing targeting indicators in Peru. FONCODES ranks all districts by its poverty index. It then allocates resources to each district according to the following formula: 6  (2) Allocation, (Index, * Population,) i (Index, * Population,) j=1 The "FONCODES method" thus makes all districts in the country eligible for benefits, but weights the population of each one by its poverty index. For example, Coronel Castaffeda, the district with the highest value of the FONCODES index (36.38), and a population of 607 inhabitants, would be allocated (36.38*607)/346,201,217 = .0064% of the total budget for that year. By contrast, Pacocha, the district with the lowest value of the FONCODES index (1.00), and a population of 6500, would be allocated (1*6,500)/346,201,217 = .0019% of the total budget for that year. Per capita allocations to inhabitants of Coronel Castafieda Castaiieda would therefore be almost 37 times per capita allocations to inhabitants of Pacocha4 Note that if the index is a poverty rate, allocations to district i simply correspond to the fraction of the poor who live in district i. I adapt the "FONCODES method" to other indicators by substituting the infant mortality rates and the imputed poverty measure for the FONCODES index in equation (2) above, and compute the corresponding district-level allocations. The simulations in this paper make a number of assumptions. The most important assumption is that there is no targeting of program resources within a given district. This is clearly unrealistic: Paxson and Schady (1999) use non-parametric regressions to estimate the probability of benefiting from social programs as a function of the number of standard deviations a given household's income is above or below mean district income. Their results show that the within-district distribution of investments made by FONCODES and the food distribution program PRONAA is hump-shaped, peaking at about one-standard deviation above mean district income, while the within-district distribution of investments made by the school construction program INFES is regressive, rising with household income. A more plausible assumption for the simulations in this paper is that the degree of within-district targeting is independent of the 4 In actual fact, FONCODES' allocation mechanism is a little more complicated than this. FONCODES first allocates 60% of resources to rural areas and 40% to urban areas. The final allocation to each district is then the sum of the rural and urban allocations--standardized to add up to 100%, Ad hoc adjustments are also made to privilege border areas, to coordinate investments with other public sector programs, and to ensure that each of FONCODES' regional offices (which correspond roughly to individual departments) has a minimum operating budget. I do not take these "refinements" into account in the simulations below. 5 Of course, this is only one of a potentially infinitely large number of fortnulas which could be used. For example, the Technical Team of the Ministry of the Presidency has proposed an allocation formula which makes a distinction between districts with a high proportion of poor people (as measured by the FONCODES index) and districts with a large number of poor people (as measured by the product of the FONCODES index and population). Proposed investments would then be directed to 262 districts with the highest proportion of the population in poverty and 232 districts with the highest number of people in poverty. The final count is 419 of Peru's 1812 districts, because some districts have a high value of the FONCODES index and a high value of the product of the index and population. 7  choice of welfare indicator which is used to distribute resources across districts. Under this assumption, which seems quite reasonable, the actual estimates of changes in poverty under different targeting regimes will be biased (up, if there is positive intra-district targeting, down if the converse is true), but the preferred rank-order of the indicators used to assign resources across districts should be unaffected. The simulations assume that benefits from program investments in a district accrue entirely to the residents of that district. This might not hold, say, if beneficiaries of a food distribution program implemented in one district are residents of a different district. But intra-district spill- overs are unlikely to be systematic--that is, they should not consistently favor residents of one kind of district over residents of another. Intra-district spill-overs should therefore not affect the rank order of indicators either. Some additional assumptions have to be made about the impact of transfers on various poverty measures. Poverty in Peru has generally been defined as an individual's inability to meet a specified level of expenditures-the poverty line-when individual expenditures are approximated by total household expenditures divided by the number of eligible household members.6 We must therefore translate expenditures by social programs into household expenditures--by first translating program expenditures into changes in household income, and then estimating the proportion of additional disposable income that is spent. As a matter of convenience, I have assumed that all program expenditures translate into additional household income and that all of this additional income is spent7 Finally, the simulations assume that the cost of administering programs is constant across regions, and ignore the effects of transfers on behavior such as migration towards districts which 6 The 1995 and 1996 INEI household surveys, which used income to measure poverty, are exceptions. 7Two points are worth noting here. First, many social programs in Peru are involved with the construction of small-scale infrastructure. The wages paid to laborers in these projects are only a fraction of the total cost, and other benefits--say, of having an additional classroom--are unlikely to have a short-term impact on household income. We can easily relax the assumption of a one-for-one equivalence between changes in program expenditures and changes in household income, however, if we simply model a smaller budget. For example, if only 50% of the expenditures on poverty alleviation programs translate into short-term increases in income, the "relevant" budget would be only half the actual budget. Second, households typically do not spend all of an increase in income. Assuming that all households spend a fixed fraction of additional income is not entirely satisfactory either, because the marginal propensity to save is likely to differ systematically across households. For example, if the fraction of income that is saved is higher in rich districts than in poor districts, simulations based on a constant marginal propensity to save might under-estimate the short-term impact of program investments in poor districts relative to rich districts, and under-estimate the relative performance of targeting indicators which assign a higher share of their resources to the poorest districts. One potential solution would be to estimate marginal propensities to save for households from the LSMS itself. The simplest way to do this would be to convert measures of income and expenditures into current prices, and then take the difference between them as a measure of savings (see especially Paxson 1992). By this measure, however, almost two-thirds (61%) of households in the 1994 LSMS dissaved. This seems unreasonable and suggests that income in these surveys is seriously underestimated vis-a-vis expenditures. 8  receive large per capita transfers, offsetting reductions in private intra-household transfers or employment, and the impact of taxes needed to finance poverty alleviation programs. The data set For all of the estimations below, I combine information from two sources: district-level averages of the infant mortality rate, the FONCODES index, and the measure of imputed poverty, and household-level data on expenditures. District-level averages are available from INEI, and household-level data can be estimated from the 1994 and 1997 LSMS, both of which were executed by the Peruvian think-tank Cuanto. District-level data can be used to estimate the proportion of total funds that would be allocated to every district under alternative targeting regimes. Further dividing this fraction by the total population of the district in question allows us to calculate the proportion of funds that would be allocated to every individual. Finally, multiplying this proportion by the total budget available for poverty alleviation programs, we can estimate per capita transfers. The LSMS can be used to estimate the expenditures of the households in the sample (3,558 for 1994, and 3,840 for 1997) and, dividing total household expenditures by household size, for household members (18,362 for 1994 and 19,562 for 1997). These estimates can be combined with information on poverty lines to calculate the headcount index, poverty gap, and P2 measure at a national level before any transfers take place.8 The 1994 and 1997 Peru LSMS drew households from 364 and 397 clusters, respectively, and the accompanying literature lists the districts from which each one of these clusters was drawn. Observations in the LSMS can be coded manually with district identifiers which match those used by INEI, and district-level and household-level data can then be merged. Having done this, we keep only those observations for which there are matching codes for place of residence in both data sets--in effect, discarding the district-level information for all but the 199 and 238 districts which were sampled in the 1994 and 1997 LSMS, respectively.9 The new, composite s Note that regional price deflators available from Cuanto and INEI are used throughout the paper to deflate both household expenditures and simulated transfers. 9 In theory, the first step in the FGT approach implemented in the simulations in this paper would divide a given budget amongst the 1812 districts in Peru. How much gets allocated to each district would then depend on the targeting indicator and the allocation formula in question. But the fact that households in the 1994 and 1997 LSMS were only drawn from 199 and 238 districts across the country, respectively, raises a potential problem: since every targeting indicator allocates a different amount to each district, the total budget for this sample of districts would not be constant across indicators. For example, the proportion of the total budget allocated to the 199 districts in the 1994 LSMS would be smaller by the FONCODES index than by the measure of imputed poverty. As a result, the total amount transferred to the 3,558 households in the survey would be smaller when we use the FONCODES index than when we use the imputed poverty measure, even after each household in the survey is weighted by its expansion factor. At the heart of the problem is the fact that the LSMS draws a nationally-representative sample of households 9  data set is representative in exactly the same way as the original LSMS data set, and can be used to make reasonable simulations about changes in expenditures and poverty at the national level. . 4. Results What is the effect of geographic targeting with alternative indicators in Peru? As a first step towards answering this question, I present estimates of poverty and allocations of funds by region. About two-thirds of the population of Peru lives in urban areas, and well over a third of these urban residents live in the capital city, Lima. Table 1 decomposes poverty measures and allocations by alternative geographic targeting indicators into three categories: Lima, other urban, and rural. The values in each cell correspond to the proportion of total poverty or the proportion of total allocations by region, so that every row sums to 100%. Here, and for all of the results presented in the paper, individuals in the household surveys are weighted with the appropriate expansion factors. Since the no targeting scenario makes an equal transfer to every Peruvian, regional allocations when there is no targeting correspond exactly to the fraction of the population living in each region. A comparison of these allocations with the various poverty measures shows that poverty in urban areas is below average: about 29% of the population lives in Lima, but only 14% to 22% of total poverty was found there in 1994, and 13% to 20% in 1997. Other urban areas account for about 36% of the population, and about one-third of poverty. Poverty in rural areas, by contrast, is well above average: just over one-third of the population lives in rural areas, but between 44% and 55% of poverty was found there in 1994, and as much as 47% to 57% in 1997. Table 1 suggests that geographic targeting using any of the indicators under consideration appears to approximate the distribution of poverty reasonably well. When any of these indicators is used for geographic targeting, about one-half of all resources are transferred to rural areas, and about two-thirds of the remaining resources to urban areas outside Lima. Comparing the three indicators, the FONCODES index transfers more to Lima than the measures of infant mortality and imputed poverty, while the measure of infant mortality makes the largest transfers to rural areas. irrespective of the district in which these households live. If a number of samples were drawn, on average, the total budget would be the same across indicators, but this does not hold for any one sample. I have corrected for this problem by normalizing the budget-in effect, summing equation (2) above only over the sample of districts in the LSMS. Note that this is not an issue with the concentration curve approach because concentration curves graph out the proportion of a given budget that is allocated to each household 10  Leakage rates Table 2 presents leakage rates by targeting indicator. The leakage rate when there is no targeting (46.50 in 1994, and 51.15 in 1997) corresponds exactly to the fraction of the population which is not in poverty. Table 2 shows that leakage rates would be minimized with geographic targeting by the infant mortality rate (according to the 1994 LSMS) or the FONCODES index (according to the 1997 LSMS). Bootstrapped standard errors (not reported, but available from the author upon request) suggest that geographic targeting with any of the indicators in question is a significant improvement on the no targeting scenario, while the differences in outcomes across indicators are not significant. Concentration curves Initial simulations show that the concentration curves for the FONCODES index, the infant mortality rate, and the imputed poverty rate are so close to each other as to be virtually indistinguishable from each other on a graph. For the sake of clarity, I therefore graph the difference between each of the concentration curves and the no targeting baseline case in Figures 1.1 and 1.2. (The "concentration curve" for the no targeting scenario is a straight, forty-five degree line: every individual receives the same transfer, so the cumulative fraction of the transfer equals the cumulative fraction of the population at every point). Figure 1.1 shows that this difference is positive throughout for all three curves: no matter where we take the cut-off between "poor" and "non-poor" to be, the poor receive a larger share of transfers when these are made on the basis of the infant mortality rate, the FONCODES index, or the measure of imputed poverty than when transfers are not targeted. Figures 1 .1 and 1.2 also show that no single concentration curve lies everywhere above all others, although the FONCODES index curve always transfers more resources to poor households than the imputed poverty measure. This is noteworthy given that much recent effort has gone into developing these kinds of measures, in Peru and elsewhere. Changes in poverty Figures 2.1 and 2.2 consider the impact of alternative geographic targeting regimes on poverty at various budget levels between 10 million soles and 5 billion soles (for a similar approach see Chaudhuri and Ravallion 1994, and Jalan and Ravallion 1998). As a point of reference, the Ministry of the Presidency, the single largest implementing agency of social in the survey-weighted, once again, by the appropriate expansion factors. Because concentration curves are mean-normalized in this way, the results are budget-independent. 11  programs in Peru, spent 2.2 billion soles on programs which could be targeted in 1995.10 For the sake of parsimony, I present graphs only for the poverty gap measure. Results for the headcount index and the P2 measure, which are available from the author upon request, are very similar." Once again, because targeting outcomes with different indicators are very similar to each other but are clearly superior to the no targeting regime, I graph the differences in poverty gaps. One way to understand the graphs is therefore as a double-difference: the first difference is the change in the poverty gap when a given budget is transferred-separately, for the no targeting scenario, and for geographic targeting using the FONCODES index, the infant mortality rate, and the measure of imputed poverty. The second difference is the difference between the change in the poverty gap under the no targeting scenario, on the one hand, and the changes in the poverty gap when there is targeting by a given indicator. The value of the y-axis at any given budget and for any given curve is therefore the additional increase in the poverty gap which could be attained from switching from no targeting to targeting with the indicator in question. Impact on povet: Like the concentration curve analysis, Figures 2.1 and 2.2 show that targeting with any one of the indicators in question is clearly preferable to no geographic targeting: all of the curves are above zero throughout. Figures 2.1 and 2.2 also show that none of the indicators clearly outperforms the others, although transfers based on imputed poverty appear to result in smaller decreases in poverty than the corresponding transfers made on the basis of infant mortality or the FONCODES index.12 Budget savings: Table 3 presents the same information from a different angle: it considers the impact on the headcount index, poverty gap, and P2 measures of spending the 2.2 billion sol reference budget without geographic targeting, and estimates the budget that would be necessary to achieve this same reduction in poverty when geographic targeting is conducted by the infant mortality rate, imputed poverty, or FONCODES index, respectively. The results show that the '