77384 the world bank economic review, vol. 16, no. 2 297–319 Impact Evaluation of Social Funds The Allocation and Impact of Social Funds: Spending on School Infrastructure in Peru Christina Paxson and Norbert R. Schady Between 1992 and 1998 the Peruvian Social Fund (foncodes ) spent about US$570 million funding microprojects throughout the country. Many of these projects involved constructing and renovating school facilities. This article uses data from foncodes , the 1993 population census in Peru, and a 1996 household survey conducted by the Peruvian Statistical Institute to analyze the targeting and impact of foncodes invest- ments in education. A number of descriptive and econometric techniques are employed, including nonparametric regressions, differences in differences, and instrumental vari- ables estimators. Results show that foncodes investments in school infrastructure have reached poor districts and poor households within those districts. The investments also appear to have had positive effects on school attendance rates for young children. Since the creation of the Emergency Social Fund in Bolivia in late 1986, social funds have been established in dozens of countries, often with support from multilateral organizations and international donors. Social funds like Bolivia’s were originally put into place to mitigate the social costs of structural adjust- ment programs. Since then they have been proposed as a safety net for the poor- est people; as a means of generating employment and transferring income; as an efficient mechanism for constructing small-scale infrastructure, especially in outlying, traditionally neglected areas; and as a way of building on (or generat- ing) local social capital by involving communities in choosing, preparing, oper- ating, and maintaining projects (Rawlings and others 2002). This article analyzes the targeting and impact of investments by the Peru- vian Social Fund (Fondo Nacional de Compensación y Desarrollo Social, or foncodes ) between 1992 and 1998. Specifically, we look at the investments foncodes made in education, addressing two questions. First, who benefited Christina Paxson is Professor of Economics and Public Affairs and Director, Center for Health and Wellbeing at Princeton University. Norbert R. Schady is Senior Economist, Latin America and the Caribbean Region at the World Bank. Their e-mail addresses are cpaxson@princeton.edu and nschady@worldbank.org, respectively. For helpful comments and suggestions, the authors thank François Bourguignon, Angus Deaton, Olivier Deschenes, Rutheanne Deutsch, Esther Duflo, John Gallup, Doug Miller, Martin Ravallion, Julie van Domelen, and Juliana Weissman; participants in a World Bank seminar and in the Northeast Univer- sities Development Conference at Harvard University on October 8–9, 1999; and two anonymous referees. They are also grateful to staff at foncodes and infes and to Livia Benavides for providing data. An earlier version of the article was issued as World Bank Policy Research Working Paper 2229. © 2002 The International Bank for Reconstruction and Development / THE WORLD BANK 297 298 the world bank economic review, vol. 16, no. 2 from foncodes education investments? This is a question of targeting. foncodes aims to transfer resources, including investments in education, to poor areas and poor households within those areas. The article evaluates the extent to which it was successful in doing so. Second, did foncodes transfers improve education outcomes? This is a question about how investments in school facilities affected measures of school attendance. Although the article describes and evaluates a specific program, it adds to an ongoing debate about the relationship between education inputs and outcomes (for a summary see Hanushek 1995 and the response by Kremer 1995). A grow- ing body of literature suggests that expenditures on school facilities have high rates of return in many developing economies (for example, Duflo 2001, Glewwe and Jacoby 1994, Glewwe and others 1995, Hanushek 1995, Hanushek and Harbison 1992). The results here suggest that expenditures on school infrastruc- ture in Peru improved the attendance rate of young children. Because expendi- tures by foncodes on education were well targeted toward poor districts and (though less clearly) poor households, improvements in attendance rates were concentrated among the neediest. I. The Setting Peru made substantial economic progress between 1992 and 1998. After a brief recession following the adoption of stringent stabilization and adjustment mea- sures in 1990, growth was generally strong, inflation low, and poverty reduction sustained (World Bank 1996, 1999). Investments in the social sectors increased dramatically. The Peruvian government attempted to target these social invest- ments to the poor—though with only partial success (World Bank 1999). foncodes was created in 1991 with the stated objectives of generating em- ployment, helping to alleviate poverty, and improving access to social services (World Bank 1998). Between 1992 and 1998 foncodes funded almost 32,000 community-based projects for an aggregate outlay of about 760 million soles.1 These community-based projects included initiatives in health, education, agri- culture, community centers, rural electrification, and water and sanitation. Most of those in education entailed constructing and renovating classrooms (table 1). Before 1995, however, foncodes also had education projects focusing on constructing and renovating sports facilities and providing textbooks and other educational materials to students. In addition, foncodes executed a series of centrally designed special projects. Those in education included a school break- fast program and the distribution of uniforms for schoolchildren. Between 1992 and 1996 foncodes spent about 160 million soles on all special projects, in- cluding those in education. 1. All reported expenditures are in 1992 soles, unless otherwise noted. The December 1992 exchange rate was 1.63 soles to the U.S. dollar. Table 1. FONCODES Projects and Project Funding, 1992–98 All projects Projects to construct and renovate classrooms Other education projects Number Funding Number Funding Funding Number Funding Funding Year of projects (m 1992 soles) of projects (m 1992 soles) (percent total) of projects (m 1992 soles) (percent total) 1992 2,813 102.7 1,185 26.2 25.0 386 6.9 6.7 1993 5,238 144.9 2,327 49.4 34.1 430 8.0 4.0 1994 4,551 110.4 2,380 48.7 44.1 100 1.3 1.2 299 1995 3,056 79.3 1,037 24.7 31.0 42 0.7 0.9 1996 4,222 83.4 987 15.0 18.0 14 0.3 0.4 1997 5,807 114.8 607 11.0 9.6 9 0.2 0.2 1998 6,088 123.8 636 12.0 9.7 1 0.0 0.0 Total 31,775 759.2 9,160 187.1 24.6 981 13.0 2.0 Note: Includes only expenditures on community-based projects. Source: FONCODES. Paxson and Schady 299 300 the world bank economic review, vol. 16, no. 2 foncodes has much in common with other social funds in the region. Two features particularly important for this article are the demand-driven and tar- geted nature of its projects. foncodes projects are demand-driven in that communities themselves choose a project and prepare a proposal for funding. foncodes then functions as a financial intermediary: rather than execute projects itself, it approves proposals and releases funds to the nucleo ejecutor—a group of community members elected for that purpose. To target its investments, foncodes uses a poverty map to allocate resources (for details, see section on targeting). foncodes staff members also conduct an informal on-site assessment of the poverty of a community requesting a project. Since 1993 the demand for foncodes projects has far exceeded the program’s budget. As a result foncodes has had a backlog of project proposals and has had to ration its investments. Although decisions about which projects to fund within a district have often been ad hoc, an attempt is made to give preference to projects in communities that the foncodes evaluators deem to be poorer. How- ever, no attempt is made to target households within a community. II. The Data The evaluation of the targeting and impact of foncodes expenditures on school infrastructure uses several sources of data. These include district-level informa- tion on the geographic distribution of foncodes allocations and expenditures, kept by foncodes, and on district characteristics from a 1993 Population and Housing Census. The analysis also uses household-level information from a household survey conducted by the Peruvian Statistical Institute (Instituto Nacional de Estadística e Informática, or inei) in 1996 and from two Living Stan- dards Measurement Study (lsms) surveys, conducted in 1994 and 1997. District-Level Data Monthly records on the number of foncodes projects and amounts spent in each district are available for 1992 through 1998. Expenditures do not include administrative and overhead costs and are available only for the community-based projects, not the special projects. Similar information is available for expendi- tures by a second school infrastructure program in Peru, Instituto Nacional de Infraestructura Educativa y de Salud (infes), but only for 1995. Though both infes and foncodes are central government programs, an important difference is that infes has (mainly) built or renovated secondary schools in urban areas, whereas foncodes has (mainly) renovated primary schools in rural areas. infes has also spent considerably more on school infrastructure than foncodes: in 1995 it spent about 350 million soles on school infrastructure, whereas foncodes spent 25 million soles. An important district-level variable for the analysis is the foncodes index, a district-level poverty measure. This index forms the basis of the poverty map that foncodes has used to allocate resources. Specifically, since 1992 foncodes Paxson and Schady 301 has allocated resources to each of its 24 regional offices in a two-step process. First, foncodes makes a “referential allocation� to each district (before 1996, to each province) by weighting the population of that district by the foncodes index.2 The allocation to district i is given by: Indexi * Populationi (1) Allocationi = n ∑(Index * Population ) j =1 j j Second, it sums these referential allocations over the districts covered by a foncodes regional office. This determines the budget for each office. Regional offices are instructed to follow the original allocations across districts as closely as possible. Because these instructions require the regional offices to favor poorer districts in the allocation of funds, the foncodes index provides a useful mea- sure of the priority that projects in any given district should be given. The foncodes index is an ad hoc composite of different measures—includ- ing access to schooling, electricity, water, sanitation, and adequate housing and measures of illiteracy and chronic malnutrition. (All these are drawn from the Population and Housing Census conducted in 1993, except the rate of chronic malnutrition, which is based on a census of height and age among schoolchil- dren also conducted in 1993.) Composite indexes invariably involve some arbi- trary weighting of indicators. foncodes standardizes each indicator in its index by dividing it by the lowest value measured, multiplies the rate of chronic mal- nutrition by seven, and then adds all the indicators.3 For ease of interpretation, foncodes then standardizes the index by dividing all index values by the low- est value. The resulting index ranges from 1 to 36.38. Another district-level measure, used in the analysis of targeting, is imputed per capita income, constructed by inei. In Peru there are no survey-based esti- mates of income or expenditures at a level more disaggregated than the depart- ment: for example, household surveys conducted by inei, which generally have samples of 15,000–20,000 households, can be used only to compare income across “natural regions� and departments.4 inei has attempted to get around this prob- 2. Provinces and districts correspond to the two levels of local government in Peru. In 1997 there were 194 provinces and 1,812 districts in Peru (Webb and Fernández Baca 1997, p. 112). The median population of a district is about 4,000, but the population size varies considerably: rural districts can have fewer than 500 people, whereas urban districts can have more than 100,000. 3. This procedure had the unintended consequence of giving the greatest weight to the indicators with the greatest variance. Thus although the intended weights were 50 percent for the rate of chronic malnu- trition and 7.14 percent for each of the seven other measures, the actual weights in the index turned out to be 15.3 percent for chronic malnutrition, 3.4 percent for illiteracy, 2.2 percent for school attendance, 3.0 percent for overcrowding in homes, 38.3 percent for inadequate roofing on houses, 8.8 percent for access to water, 7.4 percent for access to sewerage, and 21.6 percent for access to electricity (World Bank 1996, p. 7). 4. These natural regions are Lima and the urban and rural areas of the coast, the sierra (highlands), and the selva (jungle). 302 the world bank economic review, vol. 16, no. 2 lem by combining variables common to both the 1993 census and one such sur- vey conducted in 1995 and imputing district-level measures of income and pov- erty (inei 1996).5 Although crude, the imputed income measure provides a useful measure of district-level welfare. Schady (2002) shows that in Peru various district-level measures of welfare (including the inei income measure and the foncodes index) are highly corre- lated with one another and do not differ significantly in their ability to separate poor from nonpoor districts. To make the analysis of geographic targeting more easily comparable to the analysis of household targeting, which is based on house- hold per capita income, this article uses this measure of district per capita income rather than the foncodes index for the targeting results. Household-Level Data The main source of household information used is the 1996 inei survey, a stan- dard multipurpose survey. It has a relatively large sample—more than 18,000 households in 403 districts. The inei survey collected information on household income, education levels, other household characteristics and benefits from vari- ous social programs. Households with at least one member attending public school were asked about recent improvements in school facilities and whether these had been carried out by (separately) foncodes, infes, or the local parents’ com- mittee (akin to the parent-teacher associations in the United States). These data are used to evaluate the household-level targeting of foncodes investments in school facilities. To test the robustness of the results of this evaluation, some of the results are reproduced using two lsms surveys, conducted in 1994 and 1997. The lsms sur- veys are both smaller than the inei survey—covering about 3,500 households each, in 199 districts (1994) or 228 districts (1997). But they offer an advantage in that a very similar questionnaire was applied in both years. The lsms data set includes a panel covering just over a quarter of the households in the two samples. Because the inei and lsms surveys were not designed specifically for an evalu- ation of foncodes, they have some shortcomings for the analysis. Three are worth noting. First, the surveys did not collect information on the quality of education as measured by, for example, scholastic achievement, pupil–teacher ratios, and the amount of time spent in school. Second, there appears to be a large amount of measurement error in the foncodes “treatment� variable in the 1996 inei survey, a point to which we return. Third, in the 1996 inei sur- 5. Specifically, 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 (water, sewerage, electricity), ownership of durable goods (radio, TV, refrigerator), and other variables included in both the census and the survey. The coefficients from the 24 department-level re- gressions 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. The methodology applied by inei for these imputa- tions is similar in spirit to that proposed in Hentschel and others (2000), although there are differences in how it was applied (Schady 2002). Paxson and Schady 303 vey, questions about benefits from foncodes programs were asked only of fami- lies with children in school. The survey results therefore cannot be used to deter- mine whether children not in school had access to a foncodes-improved school. III. The Targeting of foncodes Investments in Education Targeting education resources is important in Peru because there are large dif- ferences in measures of educational attainment across regions and income groups. At any given age, children in the poorest 25 percent of districts lag behind those in the richest 25 percent in years of schooling attained (figure 1). The differences across quartiles increase with age, so that by age 16 there is almost a 2-year dif- ference between children in the poorest and richest districts. A host of factors probably contribute to differences in the educational attain- ment of children, including differences in income, ethnicity, employment oppor- tunities, and the education of other household members. Many of these factors cannot be changed through public policies in the short run. But educational at- tainment is also likely to be a function of the quantity and quality of teachers, learning materials, and classrooms in a community. Poor districts and poor house- holds may therefore need additional resources, including resources spent on school facilities, to catch up with their better-off counterparts. Geographic Targeting Has foncodes effectively reached poor districts? Two aspects of the geographic targeting of foncodes investments in school infrastructure are considered: changes in targeting over time, and district-level expenditures by foncodes compared with those by infes. Regression results show that foncodes expenditures on school infrastruc- ture were targeted to poorer districts. A regression of per capita foncodes ex- penditures on education, summed over 1992–98, on the log of district per capita income indicates that a 10 percent increase in district per capita income is roughly associated with a 1 sol decrease in per capita expenditures. (The regression coef- ficient on log income is –10.69, with a standard error of 0.52.)6 Moreover, tar- geting appears to have improved over time. Nonparametric regressions of per capita foncodes expenditures on school infrastructure on log per capita income in three typical years—1992, 1995, and 1998—show that districts with lower per capita income clearly received more foncodes education expenditures, es- pecially in 1995 and 1998 (figure 2).7 6. All regression results reported are weighted by district population. Alternatively, per capita expen- ditures could have been regressed on the (imputed) poverty rate for each district. Ravallion (2000) shows that if there is no targeting within districts, so that poor households within a district are equally likely to receive a transfer whether they live in relatively “rich� or “poor� communities, the coefficient on such a regression can be interpreted as the difference between spending on the poor and spending on the nonpoor. 7. All nonparametric regressions are Fan regressions with a quartic kernel (see Fan 1992). 304 the world bank economic review, vol. 16, no. 2 Figure 1. Average Years of Schooling Attained, by Age, for Children in the Poorest and Richest 25 Percent of Districts in Peru, 1996 9 45-degree line 8 Richest districts 7 Average years of schooling attained Poorest districts 6 5 4 3 2 1 0 7 8 9 10 11 12 13 14 15 16 Age (years) Source: Authors’ calculation based on the 1996 inei survey. How well do foncodes targeting outcomes stand up to those of a compa- rable program? In regressions of per capita expenditures by foncodes and infes in 1995 on the log of district per capita income, the coefficient on foncodes expenditures is significantly negative (–1.60, with a standard error of 0.18), whereas the coefficient on infes expenditures is positive though insignificant (1.30, with a standard error of 0.80). Nonparametric regressions show that per capita expenditures on education infrastructure by infes in 1995 were much larger but much less well targeted. Per capita expenditures by foncodes de- creased monotonically with district per capita income, and those by infes were concentrated in the middle of the distribution and were lowest for the districts with the lowest per capita income (figure 3). Household-Level Targeting In Peru there is considerable heterogeneity in the distribution of welfare within districts. For example, a simple decomposition of the variance in per capita in- come in the 1996 inei survey into inter- and intradistrict components suggests that only 22 percent of the variance is explained by differences across districts. Reach- ing poor districts is therefore only a weak proxy for reaching poor households. To examine household-level targeting, the household-level incidence of foncodes benefits is calculated using information from the 1996 inei survey Paxson and Schady 305 Figure 2. Geographic Targeting of foncodes Education Projects, Various Years 25th percentile 75th percentile 3 Per capita education expenditures, 1992 soles 1995 2 1992 1 1998 0 4.5 5 5.5 6 6.5 7 Log of district per capita income Note: The top and bottom 1 percent of the distribution of log per capita income have been trimmed. Source: Authors’ calculations based on foncodes data. on access to education infrastructure and per capita income. Three separate vari- ables are defined that take the value of one for households that reported having benefited from spending by foncodes, infes, and the parents’ committees. These variables are regressed on the log of household per capita income. The weighted logit regression results (with the weights given by the expansion factors in the survey) suggest that poorer households are more likely than better-off households to benefit from foncodes investments: the estimated marginal effect of the log of household per capita income on the probability that the household benefits from foncodes is –0.010 (with a standard error of 0.001). Although poorer households are also more likely than better-off households to benefit from par- ents’ committees, the marginal effect of income on the probability of benefiting is –0.005 (standard error of 0.002), only half as large as the estimate for foncodes. Poorer households are less likely than better-off household to benefit from infes: the marginal effect of income is 0.009 (standard error of 0.001). Nonparametric regressions are used to capture possible nonlinearities in the relationship between investments and log income. The results confirm that house- holds with lower per capita income are more likely to benefit from foncodes 306 the world bank economic review, vol. 16, no. 2 Figure 3. Geographic Targeting of foncodes and infes Education Projects, 1995 25th percentile 75th percentile 12 Per capita education expenditures, 1992 soles 10 8 6 INFES 4 2 FONCODES 0 4.5 5 5.5 6 6.5 7 Log of district per capita income Note: The top and bottom 1 percent of the distribution of log per capita income have been trimmed. Source: Authors’ calculations based on data from infes and foncodes . spending than from infes spending (figure 4). To some extent this no doubt re- flects infes’s emphasis on secondary school infrastructure in urban areas and foncodes’s emphasis on primary school infrastructure in rural areas: in Peru, as in many other countries, the poor are less likely to send their children to sec- ondary school and more likely to live in rural areas.8 The nonparametric regres- sions also show that the foncodes distribution slopes upward at very low levels of (log) per capita income: the poorest 7 percent of households are less likely than their (slightly) better-off counterparts to benefit from foncodes invest- ments in education infrastructure. Measurement error is an important concern for the estimates of household targeting. Rural households in Peru are likely to have little choice of primary school. In the absence of measurement error, one would therefore expect a high 8. This also helps explain why the fraction of households that reported having benefited from infes (2.6 percent) is smaller than the corresponding fraction for foncodes (3.6 percent), despite the massive differences in the programs’ budgets. In Peru projects to repair primary schools have tended to be small and relatively inexpensive (with low-cost materials and community participation in construction), whereas projects to construct or repair secondary schools are more expensive because they are larger and more elaborate (with higher-end materials and payment of all labor costs for a contractor). Paxson and Schady 307 Figure 4. Household-Level Targeting of School Infrastructure Expenditures by foncodes, infes, and the Parents’ Committees, 1996 25th percentile 75th percentile .15 Estimated probability of benefiting .1 Parents’ committees .05 INFES 0 FONCODES 6 7 8 9 10 Log of household per capita income Note: The top and bottom 1 percent of the distribution of log per capita income have been trimmed. Source: Authors’ calculations based on the inei survey. degree of consistency in the answers given by households within a rural commu- nity to questions about the presence of foncodes-funded education projects. Unfortunately, this is not always the case. Consider households in rural areas that have only children attending primary school. In 4 rural communities in the sample all such households reported that foncodes had financed improvements to the local school, and in another 107 all such households reported that foncodes had not financed improvements. But in 46 rural communities different house- holds provided different responses, suggesting that households in the 1996 inei survey did not always report program benefits accurately. Measurement error of this sort will bias the estimates of program incidence if it is correlated with income so that richer (or poorer) households are more (or less) likely to report that they benefited from foncodes. To further explore issues related to measurement error, the analysis tests whether households that did not respond to questions about infrastructure im- provements in the 1996 inei survey differed systematically from responding households. The inei survey first asked households whether any member attended a public school and, for those answering affirmatively, whether they had “knowl- edge of any improvement to this public school in the last 12 months.� Next, the survey asked about the kind of improvement undertaken and finally about the 308 the world bank economic review, vol. 16, no. 2 agency that financed it. About 15 percent of households with a child in public school did not recall whether there had been a recent improvement, and about 10 percent with knowledge of school improvements did not know who financed them. The analysis finds that nonresponding households differed in some ways from responders, but the differences tended to be small. For example, the mean edu- cation of the household head was 7.1 years for households knowing of school improvements and knowing who financed them, 7.1 for households knowing of school improvements but not knowing who financed them, and 6.57 for house- holds not knowing whether there had been any school improvements. For these same groups of households, the log of household per capita income was 8.04, 7.94, and 8.04.9 Although these differences are sometimes significant, they are small and do not invalidate the analytical approach used. foncodes has placed a great deal of importance on geographic targeting and less on other forms of targeting, such as means testing (World Bank 1996). A comparison of figures 2 and 4 suggests that it has done better reaching the poor- est districts than it has reaching the poorest households. To explore this issue further, the estimated probability of benefiting from school investments by infes, the parents’ committees, and foncodes is graphed on the number of standard deviations that the income of household i in district j is above or below the mean income in district j when both household and district incomes are calculated using the 1996 inei survey. (Both the mean and the standard deviations are district- specific.) The nonparametric regression line for foncodes school infrastructure is humped, peaking at about 1.5 standard deviations above mean district income (figure 5). Within a given district households that are somewhat better off than their counterparts are more likely to benefit from foncodes investments in school infrastructure. This suggests that there was essentially no (positive) intradistrict targeting of foncodes resources in 1996. This finding adds to a debate about the relative importance of central and com- munity-level targeting and about the level at which targeting decisions should be made (for example, Alderman 1998 and Galasso and Ravallion 2000). foncodes is a central government program that has chosen how to allocate resources across districts from the center. Largely, decisions about which community projects to finance have been left to employees in foncodes’s regional offices—a much more aggregate level than the provinces and districts that form the basis of the poverty map. This targeting scheme has been effective at reaching poor districts but not at reaching the worst-off households within those districts. Without more information—such as a comparison with a similar small-scale infrastructure program using community-based targeting within districts—it is hard to know whether foncodes’s within-district targeting performance is bet- ter or worse than the alternatives. The analysis does suggest, however, that there 9. The authors thank an anonymous referee for this suggestion to analyze differences in the charac- teristics of responding and nonresponding households. Paxson and Schady 309 Figure 5. Probability of Benefiting from School Infrastructure Expenditures by Number of Standard Deviations of Household Income Above or Below Mean District Income Parents’ committees .15 Estimated probability of benefiting .1 FONCODES .05 INFES 0 -2 -1 0 1 2 3 4 Standard deviations above or below mean district income Note: The top and bottom 1 percent of the distribution of log per capita income have been trimmed. Source: Authors’ calculations based on the inei survey. are limits to the extent to which central government programs can reach poor households without such targeting mechanisms as indicator targeting or self- targeting through the provision of inferior infrastructure. IV. The Impact of foncodes Investments in Education Although it is too early to assess the long-term impact of foncodes education investments, the program has been in existence long enough to have had short- run effects, such as increasing school attendance rates. In this section district- level data are used to examine the relationship between school attendance rates and foncodes spending on school infrastructure. Using school attendance data from the 1993 census and the 1996 inei survey, the analysis shows that there is a positive association between foncodes education funding and gains in pri- mary education: districts that received the largest per capita allocations of foncodes funds for education experienced the largest increases in school at- tendance for children ages 6–11. The analysis begins by looking at the associations between district-level school attendance rates for children ages 6–11 and the foncodes index. As noted ear- 310 the world bank economic review, vol. 16, no. 2 lier, the foncodes index is higher for poorer districts, which were given prior- ity in foncodes funding decisions. To avoid a mechanical relationship between the foncodes index and school attendance, the analysis modifies the index to exclude one of its usual elements—the fraction of children ages 6–11 who are not attending school. This is done by replacing the district value of the fraction of children ages 6–11 who are not attending school with the countrywide aver- age. Otherwise the index is identical to that used by foncodes. Nonparametric (Fan) regressions of attendance rates in 1993 (based on the census data) and 1996 (based on the inei survey data) on the modified foncodes index, using observations on 349 districts, indicate that the relationship between attendance rates and the foncodes index changed during the period (figure 6). As might be expected, there is an obvious negative relationship between school Figure 6. Nonparametric Regressions of District School Attendance Rate on (Modified) foncodes Index 1.0 1993 .9 School attendance rate 1996 .8 .7 .6 0 10 20 30 40 FONCODES index Note: Figure is based on regressions that use a bandwidth of three and are weighted by the popula- tion of children in the district so that less-populated districts get less weight. The confidence intervals for the 1996 results (shown as dotted lines) were computed using a bootstrap procedure: drawing random samples (with replacement) from the original inei sample (with the probability of being drawn propor- tional to the sampling weight), estimating the nonparametric regressions 50 times using the micro-level data, and computing the standard deviation of the estimate at each value of the foncodes index on the x-axis. The confidence lines show the point estimate at each value of the foncodes index plus and minus two standard deviations. Source: Authors’ calculations based on 1993 census data and 1996 inei household survey data. Paxson and Schady 311 attendance and the foncodes index in 1993, so that poorer districts—with higher index values—have lower attendance rates. But this negative relationship is much less pronounced by 1996. In other words, worse-off districts had large gains in school attendance, but better-off districts did not. One puzzling feature of figure 6 is that attendance rates appear to have de- clined for children in well-off districts between 1993 and 1996. But this decline may at least in part reflect the timing of the surveys, coupled with the fact that both surveys explicitly asked about school attendance rather than school enroll- ment. In Peru the school year runs from April to December. The 1993 census data were collected in June, relatively early in the school year, whereas the 1996 inei survey data were collected in November. Attrition in attendance over the school year could account for the lower mean attendance rates in 1996. The lsms surveys provide some evidence that attrition does affect measures of district-level school attendance rates. The 1994 lsms survey, conducted be- tween June and August, shows a drop in attendance rates of primary-school-age children of two percentage points (from 96.6 percent to 94.6 percent) between June and July, after which attendance rates appeared to stabilize. The 1997 lsms survey, conducted between September and November, shows no systematic de- cline in attendance. The decline from June to July 1994 was concentrated among children in poorer districts with a higher foncodes index value. For example, for children in districts with a foncodes index greater than 14 (roughly the me- dian), the attendance rate declined from 96 percent to 92 percent between June and July. Because the 1993 census was conducted in June, part of the high atten- dance rate in 1993 (shown in figure 6) may reflect the higher attendance early in the school year. Moreover, because attrition after June is more likely for poorer children, the results may understate the gains made by children in poor districts relative to those in rich districts between 1993 and 1996.10 Were the districts that experienced the largest gains in school attendance also those that received the most foncodes funding for school infrastructure? Fig- ure 7 graphs both the change in the school attendance rate and the total per capita foncodes expenditures on school infrastructure in 1992–95 as a function of the foncodes index. (foncodes expenditures are summed over 1992–95 rather than 1993–96 because it is assumed that expenditures on school infrastructure cannot affect attendance until the year after they are made.) This figure shows that poorer districts that experienced greater gains in school attendance also received more funding for school improvements. The degree of comovement be- tween attendance gains and school funding is striking. A regression of the pre- dicted value of the attendance gain on the predicted value of school expenditure 10. To double-check the results, the analysis of figure 6 was repeated using the 1994 and 1997 lsms surveys, including and excluding observations from June. Although the results based on the lsms sur- veys are somewhat noisier, because there are fewer districts, they are similar to those based on the census and the inei survey. Excluding observations from June has almost no effect on the relationship between the gain in attendance and the foncodes index. 312 the world bank economic review, vol. 16, no. 2 Figure 7. Nonparametric Regressions of Change in District School Attendance Rate and Per Capita foncodes Education Expenditures in 1992–95 on (Modified) foncodes Index .10 15 Per capita FONCODES education expenditures Change in school attendance rate Change in school attendance rate .05 10 .00 Per capita FONCODES 5 education expenditures -.05 -.10 0 0 10 20 30 40 FONCODES index Source: Authors’ calculations based on 1993 census data, the 1996 inei household survey data, and foncodes data. yields a coefficient of 0.0151, which implies that a one-standard-deviation (9.6) increase in per capita foncodes spending on school infrastructure is associated with a gain in the attendance rate of 14.5 percentage points. Another interesting feature of figure 7 is that the relationships it shows are nonlinear and nonmonotonic. Districts with foncodes index values between 22 and 26 had greater attendance gains and higher school infrastructure spend- ing than did poorer districts with index values between 26 and 28. Both mea- sures increased again for even poorer districts with index values greater than 28. These nonlinear patterns are not the result of few observations at very high values of the foncodes index. About 15 percent of districts (accounting for 14 percent of children) had foncodes index values between 22 and 26, 6.3 percent of districts (4.7 percent of children) had index values between 26 and 28, and 9.6 percent of districts (8.4 percent of children) had index values greater than 28. Furthermore, these nonlinearities remain even when a wide range of bandwidths is used for the nonparametric regressions. Although these non- linearities are striking, it is not clear what drives them. The districts with index values between 26 and 28 were allocated more foncodes funds than the wealthier Paxson and Schady 313 districts with slightly lower index values but apparently did not apply or receive approval for greater funding for school infrastructure projects. Figures 6 and 7 show that there is a positive association between gains in school attendance and foncodes spending on school infrastructure. A key question is whether these gains in school attendance were caused by foncodes spending. Two features of the program make causality difficult to ascertain. First, foncodes is demand-driven, with community groups supplying proposals for specific projects. The unobserved characteristics that prompt community groups to apply for funds may be correlated with the outcomes of interest. For example, a poor district in which people begin to care more about education may have larger increases in school attendance rates and generate more proposals for foncodes school fund- ing than an equally poor district in which preferences for education do not change. In this example the positive association between gains in school attendance and foncodes education spending is driven by a third, unobserved factor—changes in district-level preferences about education—and there may be no causal rela- tionship between foncodes spending and gains in attendance. Second, foncodes funding was targeted to poorer districts. There was no explicit randomization of foncodes funds across districts, and no obvious natu- ral experiment that resulted in different funding levels across similar districts. It is therefore difficult to assess whether the gains in attendance in the poorer dis- tricts that received greater foncodes funding were driven by foncodes or by unobserved factors correlated with district-level poverty. For example, it is pos- sible that returns to education increased more in poorer districts than in wealthier districts over this period, prompting more parents in poorer districts to send their children to school. The analysis turns to instrumental variables techniques to deal with these two problems. The first problem—that changes in district-level tastes for education may be correlated with applications for foncodes funds for school infrastruc- ture—is the more easily handled. The district-level gain in school attendance is regressed on foncodes spending on school infrastructure, with the infrastruc- ture spending instrumented with the foncodes index and with a set of vari- ables reflecting the political preferences of the district population, as measured by the fraction voting for Alberto Fujimori in the 1990 and 1993 national elec- tions. Because the foncodes index was used by regional offices to prioritize allocations, it should be correlated with district-level spending on school infra- structure. Moreover, because the foncodes index is based on measures of a district’s unmet needs in 1993, it is plausibly uncorrelated with unobserved changes in tastes for education between 1993 and 1996. The use of the political variables as instruments is motivated by previous research indicating that dis- tricts that moved against Fujimori between 1990 and 1993 were subsequently treated more generously by foncodes, presumably in an attempt to regain votes (Schady 2000). Under the assumption that the political preferences of districts were not correlated with changes in preferences for education, the political mea- sures are valid instruments. 314 the world bank economic review, vol. 16, no. 2 The second problem—that district-level poverty may be correlated with un- observed factors that affect attendance rates—is more difficult to handle con- vincingly. The fact that foncodes spending was targeted to poorer districts— combined with the possibility that there could have been other, unobserved reasons that poorer districts experienced larger attendance gains—suggests that the foncodes index may not be an appropriate instrument for foncodes spend- ing on schools. Results are presented in which the foncodes index is included as an explanatory variable in the second-stage regressions. This specification allows initial district-level poverty to have an independent effect on the subse- quent gain in school attendance and relies solely on the political variables as in- struments. For this strategy to be valid, the political variables must not have af- fected other resource flows to districts that had an effect on school attendance. Although no conclusive evidence is available on this point, the analysis exam- ines whether school infrastructure spending by infes in 1995 is related to the political measures and to attendance gains. These results are discussed. Table 2 shows ordinary least squares and instrumental variables estimates of regressions of the district-level gain in school attendance on foncodes spend- ing on school infrastructure. Because the school attendance rate is bounded be- tween 0 and 1, the gain in attendance between 1993 and 1996 is measured as the (approximate) change in the log odds that a child ages 6–11 attends school. Specifically, the gain in school attendance is measured as  p96    p93   p96 − p93 (2) Gain = ln     − ln     ≈  1 − p96   1 − p93  p93(1 − p93) where p96 and p93 are the fractions of children ages 6–11 who attended school in the district in each year. Because the measure of p96 is derived from the inei survey rather than the census, there are some districts in which the observed frac- tion of 6–11-year-olds attending school equals 1. It is for this reason that the analysis uses the approximation of the change in the log odds, which does not require division by 1 – p96.11 In panel A of table 2 the assumption is maintained that the foncodes index affects gains in attendance only through its effect on foncodes spending. (This assumption is loosened in panel B.) The first column in panel A shows the ordi- nary least squares estimate from a regression of the gain in attendance on per capita spending on school infrastructure. The point estimate indicates that a 1-sol increase in per capita foncodes spending on school infrastructure increases the log odds that a child attends school by 0.059. This corresponds to a gain in the attendance rate of about 0.75 percentage point for districts with an initial atten- dance rate of 85 percent (about 25 percent of children were in districts with at- 11. All regressions shown were also estimated using p96 – p93 as the dependent variable. These es- timates are qualitatively similar to those using the log odds specification, although they are sometimes slightly less precise. Paxson and Schady 315 Table 2. Effects of foncodes Expenditures on District-Level Gain in School Attendance Variable (i) ols (ii) iv (iii) ols (iv) iv A Per capita foncodes school infrastructure 0.059 0.195 0.056 0.152 expenditures (0.011) (0.027) (0.010) (0.033) Per capita foncodes “other� expenditures 0.014 0.032 (0.004) (0.016) c2 test of overidentifying restrictions 7.62 (6) 4.16 (5) (degrees of freedom in parentheses, [0.267] [0.526] p-values in brackets) B Per capita foncodes school infrastructure 0.022 0.109 0.023 0.145 expenditures (0.011) (0.069) (0.011) (0.080) Per capita foncodes “other� expenditures 0.008 0.031 (0.004) (0.019) foncodes index 0.068 0.035 0.062 0.004 (0.009) (0.027) (0.009) (0.036) c2 test of overidentifying restrictions 3.69 (5) 3.48 (4) (degrees of freedom in parentheses, [0.595] [0.480] p-values in brackets) Note: Columns (i) and (iii) are based on ordinary least squares regressions, and columns (ii) and (iv) on instrumental variables regressions. Standard errors in parentheses. The depen- dent variable is the approximate gain in the log odds that a child ages 6–11 attends school, as given by Gain = (p96 – p93) / (p93 [1 – p93]), where p93 and p96 measure the fractions of children attending school in 1993 and 1996. Each regression is estimated using data from the 349 districts for which the political measures and 1996 school attendance measures were avail- able. The instruments for the instrumental variables regressions in panel A include the foncodes poverty index, the fraction of pro-Fujimori votes in the province in 1990 and in the district in 1993, and interactions of the two vote measures with the foncodes poverty index. The in- struments in panel B include the vote measures and interactions with the foncodes index but not the foncodes index itself (which is included as an explanatory variable in the second- stage regressions). All regressions are weighted by the number of children in the district in 1993. Source: Authors’ calculations. tendance rates equal to or less than 85 percent in 1993). This effect is large, given that median per capita spending on school infrastructure was 6.06 soles. The second column shows the instrumental variables estimate of the effect of school infrastructure spending on the attendance gain. The instruments in the first-stage regressions include the foncodes index, the fraction of the vote re- ceived by Fujimori in the province in 1990 and the district in 1993 (in logs), and interactions of the two vote variables with the foncodes index. (District-level voting data are not available for 1990.) The first-stage regressions indicate that districts with a smaller share of pro-Fujimori votes in 1993, holding the prov- ince-level vote in 1990 fixed, received more foncodes funding between 1993 and 1996. The positive association between the erosion in the pro-Fujimori vote and foncodes expenditures is stronger for wealthier districts. The instruments in the first-stage regressions are jointly significant (F[7,341] = 15.12, p = 0.0000), 316 the world bank economic review, vol. 16, no. 2 and the test for the validity of overidentifying restrictions is easily passed. The instrumental variables estimate of the effect of school infrastructure spending on the gain in school attendance is 0.195 and is significantly different from zero. This point estimate implies that a 1-sol increase in per capita foncodes spend- ing on school infrastructure would result in an increase in the attendance rate from 85 percent to 87.5 percent. In addition to spending on school infrastructure, foncodes also funded noneducation spending, and results in the third and fourth columns of panel A examine whether other spending affected school attendance. It is possible that “other� spending could have had a positive effect on school attendance: although most of this spending went to noneducation infrastructure, a small portion went to school supplies and other education inputs. Moreover, the income gains asso- ciated with general foncodes spending could have increased attendance. But other spending is unlikely to have had as large an effect on school attendance as school infrastructure spending. Consistent with expectations, the ordinary least squares results indicate that although other spending was associated with gains in school attendance, the effect was a quarter of the size of the effect of school infrastructure spending. The instrumental variables estimates also indicate that the effect of other expenditure was substantially smaller. Thus the results in panel A show that foncodes spending on school infra- structure had a substantial impact on school attendance for children ages 6–11 and a much larger effect than other foncodes spending. But the instrumental variables estimates do not account for the potential problem of unobserved fac- tors that may have resulted in increased attendance in the poorer districts that received more foncodes funds. This line of reasoning suggests that the foncodes index should be made an explanatory variable in the second-stage regressions, and this is done for the results in panel B. Removing the foncodes index from the list of instruments and including it in the second stage comes at a cost: the political variables that remain in the list of instruments, although related to foncodes expenditures, are fairly weak instruments, especially for school in- frastructure expenditure. The F-statistic for the political variables is F(4,343) = 2.60 (p = 0.036) in the school infrastructure equation and F(4,343) = 5.53 (p = 0.0003) in the other expenditure equation. The instrumental variables results in panel B should therefore be treated with caution. The results provide some evidence that, even controlling for poverty in 1993 (as measured by the foncodes index), districts that had higher school infra- structure spending had greater gains in school attendance. The effects of foncodes school infrastructure spending are generally smaller and less precisely estimated than those in panel A. For example, the ordinary least squares estimate of the effect of infrastructure spending is cut in half when the foncodes index is in- cluded. But the instrumental variables estimates in the last column of panel B (which includes both infrastructure and other spending) indicate that the effect of school infrastructure spending is nearly equal to that in the corresponding Paxson and Schady 317 column of panel A, though with a much larger standard error. In this specifi- cation neither the foncodes index nor other spending is significant. The instrumental variables results in panel B are identified through the effects of the political variables on foncodes spending. It is possible that voting pat- terns across districts affected flows of other resources—for example, education expenditure by infes. Unfortunately, no data are available on how infes spend- ing changed over time, so it is impossible to assess whether districts that moved away from Fujimori between 1990 and 1993 received increases in infes spend- ing. Data are available on district-level infes spending in 1995, however, and the models shown in table 2 were reestimated with a measure of per capita infes spending in that year. In no case was the coefficient on infes spending signifi- cant, and it had minimal effect on the size and significance levels of the other parameter estimates. The instrumental variables models were also reestimated with infes spending as an additional (although exogenous) regressor, with similar results. infes expenditure could not be instrumented, because the sets of instru- ments used in table 2 were not significantly related to infes spending. V. Conclusion This article analyzes the targeting and impact of foncodes projects in the educa- tion sector. Nonparametric regressions are used to evaluate the geographic and household incidence of foncodes investments. The findings show that foncodes reached poor districts and, to the extent that they lived in those districts, poor households. The targeting of foncodes projects in education compares favorably with the targeting of a comparable public-sector program. Geographic variation in expenditures and school outcomes is used to analyze the impact of foncodes spending on school attendance rates. The results show that districts with the high- est levels of foncodes spending on school infrastructure between 1992 and 1995 had the biggest improvements in attendance between 1993 and 1996. The results in the article are consistent with a causal relationship between spending on school facilities and improvements in attendance, especially of poor children. An earlier version of the article reached similar conclusions through analyses that used household-level data from the 1994 and 1997 lsms surveys.12 The results thus add to a growing literature that finds evidence of a positive as- sociation between school-based inputs and measures of educational attainment (for example, Angrist and Lavy 1999, Case and Deaton 1999, Glewwe and Jacoby 1994, Krueger 1999). Nonetheless, the analysis in this article was constrained by important limitations in the data. Three areas deserve attention. First, the usefulness of the instrumental variables results hinges on the valid- ity of the identifying assumptions—in this case that the political variables and 12. These results are available from the authors on request. 318 the world bank economic review, vol. 16, no. 2 (in panel A of table 2) the foncodes index are determinants of foncodes spend- ing in a district but are uncorrelated with unobserved factors that affected school attendance. Although these assumptions are plausible, an evaluation based on a randomized treatment and control strategy would have been preferable. Because applications for foncodes funds have exceeded the amounts available, an op- portunity exists for randomization in the allocation of funds (across equally poor districts). The successful use of randomization in the Bolivian Social Fund and in Programa de Educación, Salud y Alimentación (progresa) in Mexico high- lights the benefits of randomization for program evaluation. (On the Bolivian Social Fund see Newman and others 2002; on progresa see Behrman and Todd 1999, Schultz 2001.) Second, as a result of the absence of credible village-level measures of foncodes funding, all the estimates of the impact of education projects are based on dis- trict-level measures of foncodes expenditures. Further disaggregation could be important, especially in urban districts, which can be very large. Moreover, village-level measures of “treatment� would make it possible to use statistical matching as an estimation strategy. Third, lack of disaggregated information on such measures as scholastic achievement, pupil–teacher ratios, and the time children spend in school precluded analysis of the impact of foncodes education projects on school quality. Col- lecting such data and understanding the mechanisms whereby improvements in school infrastructure in Peru interact with other school-level changes to result in more learning should be priorities in future research. References Alderman, Harold. 1998. “Social Assistance in Albania: Decentralization and Targeted Transfers.� lsms Working Paper 134. World Bank, Washington, D.C. Angrist, Joshua, and Victor Lavy. 1999. “Using Maimonides’ Rule to Estimate the Effect of Class Size on Scholastic Achievement.� Quarterly Journal of Economics 114(2):533–75. Behrman, Jere, and Petra E. Todd. 1999. “Randomness in the Experimental Samples of progresa.� International Food Policy Research Institute, Washington, D.C. Case, Anne, and Angus Deaton. 1999. “School Inputs and Educational Outcomes in South Africa.� Quarterly Journal of Economics 114(3):1047–84. Duflo, Esther. 2001. “Schooling and Labor Market Consequences of School Construc- tion in Indonesia: Evidence from an Unusual Policy Experiment.� American Economic Review 91(4):795–813. Fan, Jianqing. 1992. “Design-Adaptive Nonparametric Regression.� Journal of the American Statistical Association 87(420):998–1004. Galasso, Emanuela, and Martin Ravallion. 2000. “Decentralized Targeting of an Anti- Poverty Program.� Policy Research Working Paper 2316. World Bank, Development Research Group, Washington, D.C. Glewwe, Paul, and Hanan Jacoby. 1994. “Student Achievement and Schooling Choice in Low-Income Countries.� Journal of Human Resources 29(3):843–64. Paxson and Schady 319 Glewwe, Paul, Margaret Grosh, Hanan Jacoby, and Marlaine Lockheed. 1995. “An Eclectic Approach to Estimating the Determinants of Achievement in Jamaican Pri- mary Education.� World Bank Economic Review 9(2):231–58. Hanushek, Eric A. 1995. “Interpreting Recent Research on Schooling in Developing Countries.� World Bank Research Observer 10(2):227–46. Hanushek, Eric A., and Ralph Harbison. 1992. Educational Performance of the Poor: Lessons from Rural Northeast Brazil. New York: Oxford University Press. Hentschel, Jesko, Jean Olson Lanjouw, Peter Lanjouw, and Javier Poggi. 2000. “Com- bining Census and Survey Data to Trace the Spatial Dimensions of Poverty: A Case Study of Ecuador.� World Bank Economic Review 14(1):147–65. inei (Instituto Nacional de Estadística e Informática). 1996. Metodología para determinar el ingreso y la proporción de hogares pobres. Lima. Kremer, Michael R. 1995. “Research on Schooling: What We Know and What We Don’t. A Comment on Hanushek.� World Bank Research Observer 10(2):247–54. Krueger, Alan B. 1999. ““Experimental Estimates of Education Production Functions.� Quarterly Journal of Economics 114(2):497–532. Newman, John, Menno Pradhan, Laura B. Rawlings, Geert Ridder, Ramiro Coa, and Jose Luis Evia. 2002. “An Impact Evaluation of Education, Health, and Water Sup- ply Investments by the Bolivian Social Investment Fund.� World Bank Economic Review 16(2):241–274. Ravallion, Martin. 2000. “Monitoring Targeting Performance When Decentralized Allo- cations to the Poor Are Unobserved.� World Bank Economic Review 14(2):331–45. Rawlings, Laura B., Lynne Sherburne-Benz, and Julie van Domelen. 2002. “Letting Communities Take the Lead: A Cross-Country Evaluation of Social Fund Perfor- mance.� World Bank, Social Protection Network, Washington, D.C. Schady, Norbert R. 2000. “The Political Economy of Expenditures by the Peruvian Social Fund (foncodes), 1991–1995.� American Political Science Review 94(2):289–304. ———. 2002. “Picking the Poor: Indicators for Geographic Targeting in Peru.� Review of Income and Wealth 48(3):417–33. Schultz, T. Paul. 2001. “School Subsidies for the Poor: Evaluating the Mexican progresa Poverty Program.� Economic Growth Center Discussion Paper No. 834. Yale Uni- versity, New Haven, Conn. Webb, Richard, and Gabriela Fernández Baca. 1997. Perú 1997 en números: anuario estadístico. Lima: Cuánto. World Bank. 1996. “Did the Ministry of the Presidency Reach the Poor in 1995?� Country Operations Division I, Country Department III, Latin America and the Caribbean Region, Washington, D.C. ———. 1998. “Implementation Completion Report: Peru, Social Development and Compensation Fund Project (foncodes).� Country Management Unit for Bolivia, Ecuador and Peru, Human Development Sector Management Unit, Latin America and the Caribbean Region, Washington, D.C. ———. 1999. Poverty and Social Developments in Peru, 1994–1997. World Bank Coun- try Study. Washington, D.C.