77522 The ‘‘Glass of Milk’’ Subsidy Program and Malnutrition in Peru David Stifel and Harold Alderman This study of the Vaso de Leche (‘‘Glass of Milk’’) feeding program in Peru looks for evidence that this in-kind transfer program aimed at young children furthers nutritional objectives. The study links public expenditure data with household survey data to substantiate the targeting and to model the determinants of nutritional outcomes. It confirms that the social transfer program targets poor households and households with low nutritional status. Nevertheless, the study fails to find econometric evidence that the nutritional objectives are being achieved. In designing transfer programs, governments are motivated by equity or efficiency objectives or both. Das, Do, and O ¨ zler (2005) discuss these objectives for condi- tional cash transfer programs, but their analysis also applies to in-kind transfers and commodity price subsidies. While such subsidies or transfers may be politi- cally pragmatic or administratively more feasible where markets or banks are rudimentary, the choice of in-kind or conditional transfers over direct uncondi- tional cash transfers is generally based on the assumed presence of a market failure. For example, a food price subsidy or commodity transfer may be designed to improve the nutritional status of vulnerable groups—as well as to augment the real incomes of constituents—based on the possibility that intrahousehold alloca- tions do not reflect the rates of return to investments in children. Food subsides may also be motivated by the view that past underinvestments in education led to current inefficiencies in the allocation of inputs into the production of health. No direct measure of behavior is necessary for assessing equity-driven transfers to households. The equity-improving objective can be assessed in terms of its effect on the distribution of household incomes or on poverty reduction. In contrast, as Das, Do, and O ¨ zler (2005) point out, the evaluation of transfers differs if the main motivation is increasing efficiency rather than addressing equity. If a conditional David Stifel is an assistant professor at Lafayette College; his email address is stifeld@lafayette.edu. Harold Alderman is lead human development economist at the World Bank; his email address is halderman@worldbank.org. The authors thank Emanuela Galasso, Stephen Younger, and three anon- ymous referees for extensive comments on an earlier draft. They also thank Jose Roberto Lopez-Calix and Norbert Schady for support with the data and Erik Wachtenheim (Instituto Apoyo) and Jose ´ Carlos Arca for their assistance. THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3, pp. 421–448 doi:10.1093/wber/lhl002 Advance Access publication July 10, 2006 Ó The Author 2006. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org. 421 422 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 transfer is designed to increase consumption of a commodity or use of a service, one approach is to look at the net increase (after any substitution) of consumption of that good. An alternative or additional approach is to look at the outcome that the conditionality is designed to affect. Nutritional objectives of in-kind transfers are often expressed as incremental consumption of one or more goods or one or more nutrients. However, since the transferred or subsidized good can be substi- tuted for other items in the diet, it is preferable to focus the evaluation on the impact on child growth, as, for example, in evaluations of the impact of condi- tional transfers on nutrition in Mexico (Behrman and Hoddinott 2005; Rivera, Sotres-Alvarez, and others 2003). These individual-specific measures are behavior- induced outcomes that are distinct from the standard welfarist measure of total consumption. When it comes to meeting nutritional objectives through in-kind transfers, milk is often believed to be a particularly effective commodity.1 While exclusive breast-feeding is widely advocated for children under six months old, the value of supplementation with other milk at a later age is less clear. There is some clinical evidence that milk supplementation contributes to child growth but mainly in communities where the diet is based almost entirely on root crops or when milk supplements are combined with specific interventions to shift beha- vior (Rivera, Hotz, and others 2003). Thus, nutritionists generally do not advocate milk as a candidate for subsidies because of its high nutrient costs and low energy density (Kennedy and Alderman 1987). Milk subsidy programs are nonetheless prevalent, and so it is important to assess their ability to achieve their nutritional objectives. It is surprising therefore that while there are several published studies on the distributional incidence of milk or milk product subsidies, evaluations of the nutritional impact of subsidy programs are hard to find. The literature generally takes three forms. First, there are studies of milk programs that do not include evidence on nutritional impacts. These include Tuck and Lindert’s (1996) study of milk consumption in Tunisia’s subsidy program (accounting for 10 percent of overall subsidies at their peak), Esanu and Lindert’s (1996) analysis of Romania’s milk program, and the World Bank’s (2003) report on the distribution of fluid milk in the Brazilian state of Rio Grande do Sul. Similarly, while Alderman and del Ninno (1999) estimate that exempting milk from South Africa’s value added tax—similar to a consumer subsidy costing more than $150 million a year—leads to a 0.18 percent increase in protein consumption (0.03 percent for the poorest 40 percent of the popula- tion), they do not provide evidence for its effect on malnutrition rates. Second, there are analyses of programs that include milk as one of many subsidized foods for which nutritional impacts are documented but for which the effect cannot be singled out. For example, Rush and others (1988) and Carlson and Senauer 1. For example, as reported in the December 13, 2003, issue of The Economist, China recently instituted a school milk program after noting the comparatively small stature of its citizens. Stifel and Alderman 423 (2003) find that the Women, Infants, and Children Feeding Program in the United States is clearly beneficial. That program is not confined to either milk or milk substitutes, however, and these studies do not single out the role of milk subsidies. Nor is milk generally distinguished in the literature on school feeding, even though it is often included in such programs. Powell and others (1998) report the impact of one such successful program that included milk. However, the nutritional experi- ence of such programs is generally mixed, in part because of irregular implementa- tion (Levinger 1986). Third, there are a handful of studies of milk programs that do present nutri- tional evidence—for example, studies of Mexico’s Liconsa fluid milk distribu- tion program by Gundersen and others (2000), Kennedy and Alderman (1987), and Grosh (1994). Although Grosh (1994) finds the subsidies to be distributed progressively, none of these studies shows an impact on child growth. Simply stated, it is not known whether policies to subsidize milk are effective at achieving nutritional objectives, and without knowing this it is hard to fully understand the motivation for the in-kind transfer. This article addresses this question by studying Peru’s Vaso de Leche (‘‘Glass of Milk’’) program, which provides primarily milk and milk substitutes to low-income households and is motivated by nutritional objectives. The program is well suited for this analysis because its benefits are distributed progressively (Stifel and Alderman 2005), thus eliminating one common reason for a commodity distribution program to have a limited nutritional impact. This permits focusing on the nutritional outcomes that might have motivated the subsidy program. Addressing the question of nutritional impact is not straightforward, however, since randomized evaluations of full-scale interventions are often hard to implement in politically popular transfer programs. Therefore, the approach applied here links public expenditure data with house- hold survey information to assess the program’s impact on nutritional outcomes. I. DESCRIPTION OF THE VASO DE LECHE PROGRAM At a cost of $97 million in 2001, Vaso de Leche is the largest social transfer in Peru and the second largest component of transfers from the central government to municipa- lities (Instituto Apoyo and World Bank 2002). Introduced as a pilot in Lima in 1984, the program expanded nationally during the economic crises in the late 1980s and early part of the 1990s. By 1998 the program had expanded to reach 44 percent of households with children aged from 3 to 11 through earmarked monthly transfers to municipalities (Younger 2002). By law, these municipalities are required to have an administrative committee composed of elected representatives of beneficiaries, the mayor, another local official, and a representative from the ministry of health. In addition to this administrative committee, each community has an elected Vaso de Leche mothers committee. This committee, which has a fair degree of discretionary decision-making (Instituto Apoyo and World Bank 2002), identifies the beneficiaries, the timing of deliveries, and, within limits, the commodities to be distributed. 424 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 Despite its name, the Vaso de Leche program distributes more than milk and milk substitutes. In some cases, cereals or a combination of commodities are distributed instead of or in addition to milk products. For example, 46 percent of recipient households receive one product (67 percent of them receive milk or milk substitutes), while 51 percent receive two products (88 percent of them receive milk or milk substitutes).2 Nonetheless, according to calculations of this study using data from the Vaso de Leche Public Expenditure Tracking Survey, milk and milk substitutes (such as powdered milk and soymilk) account for an average of 77.5 percent of the value of total transfers. Furthermore, for house- holds in the two poorest quintiles, milk accounts for 93.3 percent of the value of the transfers and milk substitutes for 80.4 percent. Priority is given to households with children six years old or younger or with pregnant or lactating women. Once these first-tier beneficiaries are attended to, households with children aged from 7 to 13 and people with tuberculosis may participate. Within both categories, priority is based on need.3 There have been many excellent recent studies on the distribution of social expenditures in Peru and of Vaso de Leche in particular. For example, Younger (2002) finds a pattern of progressive distribution of Vaso de Leche benefits, with improved targeting between 1994 and 1997 as coverage increased. Using a different methodology and one of the data sources employed in the current analysis (a 1997 household survey), Ruggeri Laderchi (2001) also examines the overall distribution of food transfers and their impact on food consumption and nutrition. She finds that the transfers are slightly progressive, although the poorest 40 percent of households received only 46 percent of total transfers. She also finds that while the total share of income from food-related transfers had no impact on the height of children, the income share from participation in the Vaso de Leche program had a significant impact on standardized child height (Ruggeri Laderchi 2001, p. 36). Her specifica- tion, which treats participation as exogenous, yields a positive effect only when income is instrumented and when district fixed effects are included at the same time. The impact appears negative and, in some specifications, statistically significant in the absence of these recommended econometric procedures. A recent Public Expenditure Tracking Survey followed the budget trail from the central government to the Vaso de Leche beneficiaries (Instituto Apoyo and World Bank 2002; World Bank and Inter-American Development Bank 2002). The study finds an appreciable variation between communities in the timing of delivery, the commodities chosen, and the administrative fees charged. Virtually all the funds released by the center were transferred to municipal Vaso de Leche 2. Powdered milk is considered a milk substitute in this context. Although the law states that the distributed products should be in prepared form, this occurs in only 39 percent of the committees outside of Lima, and only 7 percent of the recipients report consuming the products at the point of pickup (Instituto Apoyo and World Bank 2002). 3. While the laws on the Vaso de Leche indicate that malnourished individuals are to receive priority, nutritional measures (such as anthropometric indicators) are not used for targeting purposes. See, for example, Law 27470 in El Peruano (2001). Stifel and Alderman 425 administrative budgets and further down to the mothers committees, with only some documented small-scale leakage in the allocations. The study finds more substantial discrepancies between the commodity allocations reported by the committees and by the household, however. The study could not account for a quarter of the product transferred. Most of the unexplained gap was in urban districts (particularly provincial capitals).4 II. RESEARCH STRATEGY This section describes the approach to modeling the determinants of nutritional status5 and discusses the estimation strategy in the presence of endogenous program placement. Modeling the Determinants of Nutritional Status To determine the impact of the Vaso de Leche program on nutrition, the determi- nants of child nutritional status are estimated using program expenditures as an explanatory variable. The approach is to estimate the intention to treat rather than the effect of the treatment on the treated. The intention to treat can be concep- tualized as the effect of the Vaso de Leche transfers being offered regardless of actual participation or dropout. In this analysis, the counterfactual of interest is the state of the world if the program had not existed, which is compared with the state of the world in the presence of the program (Heckman, Lalonde, and Smith 1999). This is distinct from the counterfactual for the effect of the treatment on the treated, which is the state of the treated if the program had not existed compared with the state of the treated in the presence of the program. Evaluation of the intention to treat looks at the difference between outcomes among the eligible population where the treatment is available compared with the same population where it has not been made available, preferably controlling for site selection. Evaluation of the treatment on the treated looks at differences in the expected outcomes, conditional on participation. It is generally not possible to go directly from one form of evaluation to the other without additional assumptions since it is not usually possible to ascertain the participation of members of the control group had they had the same opportunities as the treatment group. Both types of comparisons convey useful information. But, as Heckman, Lalonde, and Smith (1999) observe in their review of methodologies for 4. The study also found leakage or dilution in the sense that children did not always receive the milk that was obtained by the household. However, this is not only a difficult topic to quantify, but the welfare interpretations of this so-called leakage also differ from those of leakages in the public expenditure allocation chain. As argued in Alderman and others (1995), expecting a transferred good to be consumed entirely by one targeted individual within a household unit is not easily reconciled with any standard household model. Nor is the intrahousehold allocation as likely to be influenced by program administra- tion as are errors of inclusion and exclusion in targeting on poverty. 5. The model is fairly standard in the literature (Strauss and Thomas 1995). This exposition draws on Sahn and Stifel (2002). 426 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 evaluation, it is often evaluation of the intention to treat that is of policy relevance (see also Rouse 1998). So, while this analysis does not measure the marginal contribution of milk consumption itself to nutritional status, the impact that is measured allows one to assess whether government expenditures on milk subsidies improve nutrition. This focus, then, differs from that of other studies on feeding programs and in-kind transfers, such as Ruggeri Laderchi (2001), which attempt to measure the effect of the treatment on the treated. The theoretical framework for the estimation is derived from a household model in the tradition of Becker (1981). Assume that the household maximizes a quasi-concave utility function that takes as its arguments consumption of milk, xm, all other commodities and services, xo, leisure, l, and the health status, y (of which a child’s anthropometric measurement, h, is one dimension) of each household member. The household solves the following problem ð1Þ max uðxm ; xo ; l; ; A; ZÞ xm ;xo ;l; where A and Z, respectively, represent household and community characteris- tics, some of which are not observed. Allocation choices are made conditional on the budget constraint: ð2Þ pm xm þ po xo ¼ wðT À lÞ þ y where pm is the price of milk, po a vector of prices, w a vector of household members’ wages, T a vector of the household members’ maximum number of work hours, and y the household nonwage income. The nutritional status of children, h, is determined by a biological health production technology: ð3Þ hi ¼ HðI;A; Z; i Þ where I is a vector of health inputs and mi represents the unobservable individual, family, and community characteristics that affect the child’s nutritional out- comes. Household characteristics (such as demographics and educational levels), A, can have an impact on health by affecting household allocation decisions. Community characteristics (such as access to clean water), Z, can also have direct impacts on nutritional outcomes. Note that the input vector I includes consumption goods (such as milk) that contribute positively to household wel- fare both directly through xm and xo and indirectly through h. This represents the simultaneous choice of consumption goods and health inputs. Given this simultaneity, the household’s optimization problem can be solved to get a set of demand equations for goods and services (x), leisure (l), and health (y). A subset of the health demand equations is the reduced-form demand equation for child nutrition, represented as follows: ð4Þ ~ðA; Z; "i Þ hi ¼ h Stifel and Alderman 427 where "i is the child-specific random disturbance term, which is assumed to be uncorrelated with the other elements of the demand function. The dependent variable is the standardized anthropometric height-for-age z- score (HAZ) for children under five years of age. HAZ is defined as (h – hr)/sr, where h is the observed height of a child of a specified sex and age group, hr the median height in the reference population of children of that sex and age group, and sr the standard deviation of height measurement for the reference popula- tion of that sex and age group. The standard reference population recommended by the World Health Organization is that of the United States National Center for Health Statistics. As several studies have indicated that less than 10 percent of the worldwide variance in height can be ascribed to genetic or racial differ- ences (Martorell and Habicht 1986), this reference population is appropriate. Children with a HAZ score of less than –2 are usually classified as stunted. The set of predictors consists of characteristics of the child (such as age, sex, and birth order), household demographic variables (such as household size and age and sex composition), characteristics of the parents (such as educational attainment and mother’s age and height), access to public services (such as piped drinking water), and a dummy variable for living in an urban area. Given the propoor targeting of the Vaso de Leche program, predicted log per capita household expenditure was also included in the estimated model to control for household wealth. Endogenous Program Placement and Explanatory Variables The primary purpose of this exercise was not to model the overall determinants of nutrition but to see whether the Vaso de Leche program has an impact on nutrition. In terms of the model described in the previous section, this effect can be trans- mitted in two ways: by increasing household income by the value of the milk transfer (if this is the entire effect, the transfer is said to be inframarginal) and by directly increasing the level of milk consumption above what would have been consumed had the transfer been made in the form of cash by influencing the marginal price. Thus, the reduced-form health demand function is adapted to include the Vaso de Leche transfers as an explanatory variable to pick up the direct effect of the program on child health independent of its role as an income transfer: ð5Þ hi ¼ hðVL; A; Z; "i Þ: As noted, the program is evaluated based on the intention to treat—in this case, conditional on the funding for the Vaso de Leche at the local level—and not on the household choice to take up this opportunity. Modeling the impact on self-selecting participants would require making a set of additional assumptions to determine the impact on the random eligible participant. While conditioning on the Vaso de Leche allocation to the community in lieu of participation means not having to solve the issue of endogenous household choice, the problem of potential bias from endogenous program placement remains (Rosenzweig and Wolpin 1986). Not even the sign of any potential bias can be established since an 428 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 estimated impact may be overestimated if programs are placed where the antici- pated return is higher than average or be underestimated (or even negative) if programs go to favored but more developed communities. This issue is addressed by using observations of Vaso de Leche expenditures from two different rounds of Demographic and Household Surveys (1996 and 2000). Thus, fixed effect estimations can control for the initial conditions in the communities. The general form of the models to be estimated is as follows: X ð6Þ hi;d;t ¼ þ xi;d;t þ Ai;d;t þ Zi;d;t þ VLd;t þ Dd þ "i;d;t d where i is the index for individual children, d the indicator for the district in which the child resides, and t the year (1996 or 2000). VL is the district level per capita Vaso de Leche expenditure. The fixed effect version also includes, D, the set of district dummy variables. The inclusion of these dummy variables removes the influence of any time-invariant district effects, including any that might correlate with the allocation of Vaso de Leche funds. This approach compares the differences in the changes in health status when Vaso de Leche transfers change, controlling for other community characteristics. In effect, g is the difference in differences estimator (Moffitt 1991) of the effect of Vaso de Leche transfers on child health. Per capita expenditure on the program increased between survey years from 29.4 soles in 1996 to 37 soles in 2000, or more than 25 percent. The coefficient of variation for the change in expenditures is 0.47, indicating substantial variation in the rates by which coverage increased to identify a first difference at the community level.6) The expenditure data are based on total expenditures in each district and not a sample and are thus analogous to a census of expenditures. As an additional precaution for site selection bias, instrumental variable methods were also employed with the fixed effects models to account for the possibility of any remaining unobserved factors affecting malnutrition that vary over time and are also correlated with the change in Vaso de Leche expenditures. This was approached in two ways. First, standard two-stage least squares models were used in which the identifying instrument is the district-level Peruvian Social Fund Fondo Nacional de Compensa- cio´n y Desarollo (FONCODES) index of unmet basic needs, a composite of various measures—including access to schooling, electricity, water, sanitation, adequate housing, and measures of illiteracy—based on the 1993 census (Schady 1998). As shown in the results, the FONCODES index is correlated with district-level Vaso de 6. There were 315 districts in the study. While the relatively small number of observations per district implies that the district dummy variable will not be measured with precision, the estimates are unbiased. If the aim were to make a statement about the level of malnutrition in any given district, then the sample size in that district would be critical. However, for making a statement about the nutritional status of children at a given level of per capita program expenditures, it is the overall sample, adjusted for cluster sampling, as well as the variance of the regressor, the district means and the covariance between them, that determines this precision. Stifel and Alderman 429 Leche expenditures, satisfying one condition for valid instruments. The other condition, which is uncorrelated with the error term, is plausible given that the index was formulated based on the 1993 census, three years before the 1996 Demographic and Health Survey (DHS). The FONCODES index may be correlated with the levels of the unobserved factors, but since the analysis also includes fixed effects estimates that are, in effect, based on the change in Vaso de Leche expen- ditures, the properties of the instrument in these estimates are based on the assumption that the index is uncorrelated with changes in unobserved factors. If the parameter of interest—the impact of Vaso de Leche expenditures—points to the same conclusion over the set of estimates, there can be reasonable confidence that the conclusion is robust. Although the surveys are pooled, implicitly restricting the parameters of individual and household characteristics to be constant over time, the instrumenting equations are allowed to vary between periods. This is done in two ways: by including a time dummy variable as a shifter and by allowing all of the parameters to vary over time (in which case province-level fixed effects models are estimated). Moreover, because the basis for the FONCODES index will remain problematic, a second instrumental variable-type method is also employed using a different means of identification. In this method, proposed by Lewbel (2004) (see also Rigobon 2003), the identification of g comes from exploiting the heteroskedas- ticity of the first-stage equation (Vaso de Leche expenditures). To illustrate, begin by defining the first-stage equation as ð7Þ VL ¼ 2 þ  X þ v where X can include all or a subset of the explanatory variables in the main equation and can include instruments such as the FONCODES index. If cov(X,v2) is nonzero (i.e., if the data are heteroskedastic), then g and the other parameters in the main equation can be estimated consistently without external instruments by an ordinary linear two-stage least squares regression in which all the exogenous right-side variables and ðX À X  Þv ^2 are used as instruments for Vaso de Leche expenditures. The requirement that cov(X,v2) „ 0 is tested by applying a Breusch and Pagan (1979) test for heteroskedasticity to the first-stage equation. The district-level Vaso de Leche expenditure data are merged with the DHS data for 1996 and 2000 to create a data set with 19,053 observations on child heights, which is used to estimate the model. The per capita Vaso de Leche district expenditure variable is the district average amount spent in the two years before and including the 1996 and 2000 surveys. Thus, five variations of the model are estimated using individual child nutri- tional status as the dependent variable. First, ordinary least squares (OLS) model is used to estimate the basic nonfixed effects model. Second, a series of fixed effects models are estimated, starting with an OLS model. Third, this is followed by a time-varying instrumental variable model that is run with province-level (not district-level) dummy variables, since time-varying district dummies in the 430 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 instrumenting equation would perfectly predict the district-level Vaso de Leche expenditure values. The fourth is a fixed effects model using instrumental vari- able methods in which a time dummy variable is included in the instrumenting equation. Lastly, Lewbel’s (2004) method of taking advantage of the hetero- skedastic nature of the data for identification is used to verify that the results are robust. The particular advantage of this fifth model is that it does not use the FONCODES index classifications at all and thus is free of any possible problems associated with that instrument. In summary, the following models are esti- mated: (a) basic OLS; (b) district fixed effects, OLS; (c) province fixed effects, time-varying instrumental variable; (d) district fixed effects, time dummy vari- able in instrumental variable equation; and (e) district fixed effects, heteroske- dasticity identification. In all of the estimates, Huber–White standard errors are estimated to correct for homogeneity among observations in the 1,364 primary sampling units (clusters).7 Finally, while selective migration into high Vaso de Leche districts is theore- tically a possibility, it is unlikely that the small transfer (1.8 percent of the income of the poor on average in 1997) is a major determinant of migration. To get an indication of whether this is a major concern, survey data were used to examine the probability that an individual migrated to the district of current residence from another district. While the Vaso de Leche allocation in the district of origin is not known, the marginal effect of current Vaso de Leche expendi- tures on the probability of migrating in the past ten years is known to be –0.0004 (z = –0.73). Thus, the results are unlikely to be biased if current residence is taken as exogenously determined. III. DESCRIPTION OF DATA SOURCES This analysis of the Vaso de Leche program benefits from a wealth of data sources available in Peru. The data come from four main sources: information on the geographic allocation of Vaso de Leche program expenditures, national household living standard surveys, national DHS, and the Public Expenditure Tracking Survey. While having multiple data sources is preferred, for evaluating nutritional impacts as illustrated here analysts need only program expenditure and household survey data with child anthropometrics. Vaso de Leche Expenditures The Vaso de Leche program has maintained monthly records of expenditures allocated to each administrative (department, province, and district) region in Peru since 1994. This information, along with district population sizes from the 1993 census and the 2000 pre-census, is used to determine real 7. There is slight change in standard errors if the Huber-White standard errors correction is based on districts rather than the less aggregated sample units. Stifel and Alderman 431 annual total and per capita program expenditures in each of the recipient districts for 1994–2000. Allocations to the district program committees do not translate fully into benefits to recipients, but considering the small scale of the leakages found at the committee level by the Public Expenditure Tracking Survey, they likely represent a reasonably accurate proxy for the value of benefits available to district residents. Living Standard Surveys Two sources of household living standard surveys were available for this study. The first is the National Household Survey (Encuesta Nacional de Hogares, ENAHO), collected by the National Institute of Statistics and Informa- tion (INEI) in 1998, 1999, and 2000. These nationally representative surveys of more than 6,500 households (2,000 for the 2000 survey) were carried out quarterly, with each quarter’s survey focusing on a different theme. This analysis concentrates on data from the second quarter module, which focuses on social services and includes information on participation in the Vaso de Leche program. Household income information is also available for each module. The second source is the 1994 and 1997 National Living Standards Survey (Encuesta Nacional de Hogares sobre Medicion de Niveles de Vida, ENNIV), collected by Instituto Cuanto. These nationally representative surveys of more than 3,500 households collect multiple indicators of household and individual well-being (e.g., education, housing, health, economic activity, consumption, and assets). The 1994 ENNIV includes information on Vaso de Leche participation by household, and the 1997 data also include estimates of the values of the transfers made to the household. Anthropometric measurements of heights and weights of young children were also recorded. Demographic and Health Surveys DHS were carried out in Peru in 1996 and 2000. These nationally representative surveys of more than 28,000 households each are part of a program funded by the United States Agency for International Development and implemented by Macro International Inc., which has included more than 70 nationally represent- ative household surveys in more than 50 countries. The surveys are conducted in single rounds with two main instruments: an individual questionnaire for women of reproductive age (15–49 years old) and a household schedule. Child anthropometric measurements are recorded in the individual module. The household schedule collects information on household members, assets, and access to public services. Since income or expenditure data are not collected, the asset data in the survey were used to predict household per capita expendi- tures. This was done by estimating a model of log household per capita expen- ditures in the 1997 ENNIV data, including the value of in-kind transfers. The explanatory variables in this model are assets in the 1997 ENNIV data that are also 432 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 available in the DHS data. This model was then used to predict the values of household expenditures in the DHS data.8 Public Expenditure Tracking Survey The Public Expenditure Tracking Survey for the Vaso de Leche program was conducted by Instituto Apoyo at the end of 2001 and early in 2002, to quantify leakages and delays in public expenditure disbursements and to assess the effects of deficiencies in the system on the quality of the services provided. Thus, interviews were conducted at three levels: the municipality, the mothers com- mittee, and the household. One hundred municipalities were sampled, and four mothers committees were randomly selected from each. Lastly, four beneficiary households were selected randomly from each mothers committee in the sample. Because there are fewer than four committees in some municipalities, only 393 committees and 1,587 beneficiary households were interviewed. The household survey includes information on household demographics, assets, and participa- tion in the Vaso de Leche program, including the values of transfers, products received, and additional purchases made. As with the DHS, neither income nor expenditure data were collected. None- theless, the share of the total program received, by wealth quintiles, was estab- lished by constructing a wealth index from the households’ asset information using a factor-analysis methodology that is regularly applied to the DHS data sets (Filmer and Pritchett 2001; Sahn and Stifel 2003). Because information on households that do not participate in the Vaso de Leche program is not included in the Public Expenditure Tracking Survey data, asset weights are derived from the nationally representative 2000 DHS and applied to the tracking survey data. This permits determining how households sampled in the tracking survey rank relative to the overall national population. The purposes for which the various data sets are used are summarized in appendix Table A.1. Vaso de Leche targeting is evaluated using the DHS, ENNIV, and ENAHO data. The Public Expenditure Tracking Survey data are used to examine the degree to which Vaso de Leche transfers are inframarginal, and the DHS9 and the district-level Vaso de Leche expenditure data are used in the child nutrition models. 8. The results of the first-stage regressions estimated with the ENNIV 1997 data are available on request from the authors. 9. Only the DHS data are employed for the nutrition models for two main reasons. First, two comparable data sets with anthropometric measurements are needed, so that district dummy variables as well as Vaso de Leche subsidies can be included in the models (see World Bank 1999 for a discussion of some of the comparability issues related to the ENNIV surveys). Second, although both the 1994 and 1997 ENNIV data sets have anthropometric data, the earliest year for which Vaso de Leche expenditure data are available is 1994. As explained in the text, the models use the average subsidies for the two years prior to and including the survey as explanatory variables, which would not be available for the 1994 ENNIV. Stifel and Alderman 433 IV. RESULTS Before examining the impact of Vaso de Leche expenditures on nutritional outcomes, this section clarifies earlier statements regarding the distribution of Vaso de Leche transfers and targeting. Distribution The results confirm that the Vaso de Leche program is reasonably well targeted in terms of both household incomes and child nutritional status, though there have been some leakages. For incomes, this is done by comparing the coverage rates of house- holds by their per capita income levels10 for five household surveys (Table 1). The percentage of households with children aged six and under (tier I target group) who receive Vaso de Leche transfers declines sharply with the level of income. For example, in 1994, coverage rates declined from 38 percent ([39.3 + 37.0]/2) of the households in the two poorest quintiles to less than 8 percent in the richest. As the coverage for all households with children increased from 28 percent in 1994 to 48 percent in 2000, coverage in the two poorest quintiles rose from 38 to 68 percent ([68.2 + 66.9]/2). While there was a concurrent increase for the more well-off people in the population, the poorest 40 percent of eligible households received more than three times as much as the richest 20 percent on average. These coverage rates compare favorably to the experiences in other Latin American countries (Grosh 1994) and other developing countries (Coady, Grosh, and Hoddinott 2004). The Public Expenditure Tracking Survey data show that the mean transfer to households in the poorest national asset index quintile is 23 percent larger than to households in the richest quintile. Notably, the bulk of this comes in the form of milk products. The mean value of milk products transferred to the poorest quintile is 135 soles and 18 soles for milk substitutes and other products. Conversely, the mean values of other products received in the other quintiles are between 52 and 100 percent of the mean value of the milk products they receive (see Stifel and Alderman 2005 for more details). Therefore, milk product transfers are generally progressive in the values received by beneficiaries, while transfers of nonmilk products are not. In nutrition-based targeting, the Vaso de Leche program also is concentrated on households with children of low nutritional status, as illustrated by coverage rates of all children under five years of age by quintile of HAZ for the three household surveys with information on both Vaso de Leche participation and anthropo- metric status of children (Table 2). To give a sense of program leakage to nonmalnourished children, the percentage of the children in each of the quintiles who are stunted is also shown (HAZ below –2). In 1997, for example, 64 percent of children in the least well-nourished quintile (those who are all stunted) lived in households that received Vaso de Leche food transfers, while just over 30 percent 10. Household per capita consumption is used for the 1994 and 1997 ENNIV data. 434 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 T A B L E 1 . Vaso de Leche Coverage Rates by Quintiles of Per Capita Income (Percent) Annual transfers per a a ENNIV ENNIV ENAHO ENAHO ENAHO capita (1997 soles) Quintile 1994 1997 1998 1999 2000 ENNIV 1997 1 (poorest) 39.3 60.5 65.5 59.4 68.2 26 2 37.0 52.4 61.5 50.0 66.9 30 3 34.3 44.6 48.2 39.4 49.4 19 4 20.1 30.7 36.0 29.3 37.3 22 5 (richest) 7.8 15.8 20.2 15.8 15.2 7 Total 27.7 40.8 46.3 38.8 47.5 21 a Expenditure per capita rather than income quintiles. Note: Domain is the set of households with at least one child aged six or younger. Encuesta Nacional de Hogares sobre Medicion de Niveles de Vida (ENNIV) is a National Living Standards Survey; Encuesta Nacional de Hogares (ENAHO) is a National Household Survey. Source: Authors’ analysis based on data described in the text; see also appendix table A.1. T A B L E 2 . Vaso de Leche Coverage Rates and Child Malnutrition by Quintiles of Height for Age z-scores (HAZ) (Percent) Quintile ENNIV 1994 DHS 1996 ENNIV 1997 Share of children in program 1 42.8 41.6 64.0 2 33.8 33.2 49.2 3 28.4 26.4 41.8 4 25.4 21.7 34.8 5 20.0 20.8 30.5 Total 30.1 28.7 44.1 Share of children who are stunted 1 100.0 100.0 100.0 2 39.1 29.5 19.0 3 0.0 0.0 0.0 4 0.0 0.0 0.0 5 0.0 0.0 0.0 Note: Domain is the set of children with HAZ. ENNIV (Encuesta Nacional de Hogares sobre Medicion de Niveles de Vida) is National Living Standards Survey; DHS, Demographic and House- hold Survey. Source: Authors’ analysis based on data described in the text; see also appendix table A.1. in the most nourished quintile lived in households that received transfers. None- theless, despite the fact that the primary stated objective of the Vaso de Leche program is to reduce the levels of malnutrition in Peru, over a third of the intended beneficiaries in the most malnourished quintile were missed. It is possible that targeting of children based on ex ante nutritional needs would have resulted in improved ex post outcomes. This could explain the low Stifel and Alderman 435 T A B L E 3 . Inframarginality of Vaso de Leche Transfers by Quintile of Per Capita Expenditure (Percent of Beneficiary Households) Total 1 (Poorest) 2 3 4 5 (Richest) Share that receive Fluid milk/dairy products 29.4 18.4 20.1 23.0 32.9 41.3 Milk substitutesa 53.3 74.8 48.8 43.2 51.5 48.7 Milk and milk substitutesa 79.5 89.6 67.3 63.9 81.7 85.4 Other products 58.3 22.3 58.7 62.6 62.9 74.9 Share that purchase additionalb Fluid milk/dairy products 42.5 20.7 14.6 30.4 52.1 52.6 Milk substitutesa 2.6 2.2 3.0 4.1 1.5 3.0 Milk and milk substitutesa 48.6 36.1 38.9 36.4 57.2 58.7 Other products 26.5 9.7 19.7 15.0 26.6 37.1 a Includes powdered milk. b Share of beneficiary households that receive the product. Source: Authors’ analysis based on data described in the text; see also appendix table A.1. levels of coverage of malnourished children. However, if targeting based on ex ante needs is persistently effective, then as the nutritional status of participants improves over time, deterioration in the degree of targeting on malnutrition should be observed. This appears not to be the case; coverage rates for malnourished children rose from 42.5 percent in 1994 to 63.5 percent in 1997. The Public Expenditure Tracking Survey data offer further indication of whether the quantities of milk provided to households by the program are extramarginal.11 If so, they are expected to have a larger impact on milk con- sumption than an inframarginal program might have. While inframarginal transfers and extramarginal transfers have the same income effect, extramargi- nal transfers have a price effect as well. Nearly half of recipients consume additional amounts of the products distrib- uted to them through the Vaso de Leche program (Table 3). For example, for the 80 percent of recipients who receive milk and milk substitutes from the program, 49 percent purchase additional milk and milk substitutes. For 29 percent of households that receive milk and dairy products, the program is inframarginal for 43 percent of them with respect to these products. While only 3 percent of households that receive milk substitutes (53 percent of recipient households) purchase additional milk substitutes, most of these households also purchase milk and dairy products. For half of these households, the Vaso de Leche program 11. The transfer is extramarginal if ex post consumption (what is observed) is exactly equal to the transfer - the recipient would consume less of the product if the transfer were in the form of cash (the recipient is consuming at the kink in the budget constraint). Alternatively, if the recipient purchases additional amounts of the product, the transfer is inframarginal. 436 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 is inframarginal over the more broadly defined category of milk and milk sub- stitutes (hence the 49 percent figure above) but not for milk substitutes alone. Thus, although the Vaso de Leche program is found to be reasonably well targeted to the expected beneficiary groups, it is unclear ex ante what effect the program has had on reducing child malnutrition. The econometric analysis, dis- cussed below, sheds some light on this issue. Impact of Vaso de Leche Transfers on Nutritional Outcomes This section assesses the impact of the Vaso de Leche food transfer program on nutrition by examining how the transfers affect child nutritional outcomes. This is done by estimating reduced-form models with standardized HAZ of children less than five years of age as the dependent variable. The summary statistics of the variables used in the model are shown in Table 4. The stunting rate dropped only marginally from 26.0 percent in 1996 to 25.8 percent in 2000. This difference is not statistically significant. Because of the many confounding influences, however, the lack of progress in reducing child malnutrition is not sufficient in itself to assess the impact of the Vaso de Leche program. So models are also used. These reduced-form models are conditioned on predicted log per capita household expenditures as a proxy for the potentially endogenous actual household expendi- tures.12 In all of the models, the parameter estimates on expenditures are statistically significant at the 99 percent level of confidence (Table 5), confirming that household wealth has a positive impact on child nutritional status, a finding consistent with results on the impact of instrumented expenditure for Peru (Haddad and others 2003). For the basic models (without fixed effects), the parameter estimate is approxi- mately 0.43; in the fixed effects models, these parameter estimates drop to between 0.16 and 0.19. Using the income response in the fixed effects models and an average income growth of 3.5 percent per capita, neutrally distributed, a counterfactual can be constructed using the 1996 DHS data. They indicate that the 26 percent malnutri- tion rate in that year would have declined to 25.2 percent in 2000, which is somewhat below the observed level. Moreover, the small transfer embodied in the Vaso de Leche would by itself have a negligible impact of roughly 0.003 on average z-scores. However, as discussed in Das, Do, and O ¨ zler (2005), in-kind or conditional transfers are expected to have a greater impact on behavior than indicated by an income transfer alone. If so, an additional direct impact from the Vaso de Leche expenditures would be expected above any impact on the level of expenditures. But the direct effect of program expenditures is negative in all of the models (Table 5), although it is not statistically significant in any. Moreover, the parameter estimates are substantively small. Thus, overall there is no evidence that expendi- tures on the Vaso de Leche program have a direct positive impact on the 12. These models were also estimated using an asset index constructed using factor analysis (Sahn and Stifel 2002) as a control for wealth. The results, which are qualitatively the same, are available on request from the authors. T A B L E 4 . Means of Variables Used in HAZ Models, DHS 1996 and 2000 Pooled sample 1996 2000 Standard Standard Standard Variable Mean deviation Mean deviation Mean deviation Percent stunted 25.9 26.0 25.8 HAZa À1.11 1.34 À1.07 1.34 À1.18 1.32 Per capita Vaso de 31.90 7.63 29.43 5.90 37.00 8.23 Leche district expenditures Log per capita household 8.22 0.39 8.23 0.40 8.21 0.38 expenditures (predicted) Male dummy variable 0.50 0.50 0.50 0.50 0.50 0.50 Multiple birth dummy 0.01 0.11 0.01 0.12 0.01 0.11 variable Birth order, second child 0.22 0.42 0.22 0.42 0.23 0.42 Birth order, third child 0.16 0.37 0.16 0.37 0.17 0.37 Birth order, fourth child 0.10 0.30 0.11 0.31 0.10 0.29 437 Birth order, fifth child 0.07 0.25 0.07 0.25 0.07 0.25 Birth order, sixth child 0.14 0.35 0.15 0.35 0.14 0.34 and above Age 0–6 months 0.10 0.30 0.10 0.31 0.09 0.29 Age 7–12 months 0.10 0.31 0.11 0.31 0.10 0.29 Age 13–18 months 0.10 0.30 0.10 0.30 0.10 0.30 Age 25–35 months 0.19 0.39 0.19 0.39 0.18 0.39 Age 36–59 months 0.41 0.49 0.41 0.49 0.42 0.49 Share household 0.30 0.13 0.30 0.13 0.29 0.12 members age 0–5 (%) Share household girls 0.10 0.12 0.10 0.12 0.10 0.12 age 6–15 (%) Share household boys 0.10 0.12 0.10 0.12 0.10 0.12 age 6–15 (%) (Continued) T A B L E 4 . Continued Pooled sample 1996 2000 Standard Standard Standard Variable Mean deviation Mean deviation Mean deviation Share household 0.10 0.12 0.10 0.12 0.10 0.12 women 16–25 (%) Share household 0.16 0.11 0.16 0.11 0.16 0.11 women 26–65 (%) Share household 0.07 0.11 0.07 0.11 0.06 0.11 men 16–25 (%) Share household 0.17 0.10 0.16 0.10 0.17 0.09 men 26–65 (%) Number of household 6.38 2.56 6.42 2.60 6.29 2.49 members 438 Head is male 0.88 0.32 0.88 0.32 0.88 0.33 Head is indigenous 0.12 0.33 0.12 0.32 0.14 0.34 Mother’s age 29.24 6.77 29.21 6.75 29.29 6.82 Mother’s height 150.33 5.67 150.33 5.80 150.33 5.40 (centimeters) Mother’s education, 0.36 0.48 0.36 0.48 0.38 0.48 primary Mother’s education, 0.39 0.49 0.40 0.49 0.38 0.48 secondary Mother’s education, 0.16 0.37 0.15 0.36 0.17 0.38 postsecondary Father’s education, 0.30 0.46 0.30 0.46 0.30 0.46 primary Father’s education, 0.44 0.50 0.45 0.50 0.42 0.49 secondary (Continued) T A B L E 4 . Continued Pooled sample 1996 2000 Standard Standard Standard Variable Mean deviation Mean deviation Mean deviation Father’s education, 0.20 0.40 0.20 0.40 0.20 0.40 postsecondary House floor dirt 0.51 0.50 0.51 0.50 0.51 0.50 dummy variable Piped drinking water 0.55 0.50 0.54 0.50 0.58 0.49 439 dummy variable Flush toilet dummy 0.40 0.49 0.41 0.49 0.39 0.49 variable Urban dummy 0.68 0.47 0.71 0.45 0.61 0.49 variable Number of 19,053 12,045 7,008 observations a Standardized anthropometric height for age z-score (HAZ). Source: Authors’ analysis based on data described in the text; see also appendix table A.1. T A B L E 5 . Reduced-Form Models of Height for Age z-score (HAZ) (Ages 0–59), Peru Demographic and Household Surveys 1996 and 2000 District fixed effects Time varying Time dummy Heteroskedasticity OLS (1) OLS (2) IV (3)a IV (4) Identification (5) Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Per capita Vaso de Leche À0.0015 À0.90 À0.0022 À0.90 À0.0019 À0.79 À0.0018 À0.75 À0.0025 À0.89 district expenditures Log per capita household 0.422 7.61*** 0.164 2.90*** 0.192 3.36*** 0.164 2.90*** 0.164 2.90*** expenditures (predicted) Male dummy variable À0.083 À4.27*** À0.075 À3.96*** À0.080 À4.23*** À0.075 À3.96*** À0.075 À3.96*** Multiple birth dummy variable À0.415 À4.03*** À0.459 À4.80*** À0.461 À4.86*** À0.459 À4.80*** À0.459 À4.81*** Birth order, second child À0.069 À2.06*** À0.062 À1.86* À0.067 À2.01** À0.062 À1.86* À0.062 À1.86* Birth order, third child À0.160 À3.99*** À0.140 À3.53*** À0.151 À3.82*** À0.140 À3.52*** À0.140 À3.53*** Birth order,fourth child À0.274 À5.53*** À0.256 À5.29*** À0.267 À5.57*** À0.256 À5.29*** À0.256 À5.29*** Birth order,fifth child À0.382 À5.97*** À0.333 À5.42*** À0.351 À5.68*** À0.333 À5.41*** À0.333 À5.42*** Birth order, sixthchild À0.434 À6.69*** À0.364 À5.82*** À0.378 À6.02*** À0.364 À5.81*** À0.364 À5.82*** 440 and above Age 0–6 months 1.163 25.12*** 1.138 24.47*** 1.143 24.79*** 1.138 24.47*** 1.138 24.47*** Age 7–12 months 0.662 14.02*** 0.638 13.76*** 0.644 13.96*** 0.638 13.76*** 0.638 13.75*** Age 13–18 months 0.148 3.24*** 0.139 3.13*** 0.146 3.25*** 0.140 3.13*** 0.139 3.13*** Age 25–35 months 0.283 6.90*** 0.265 6.72*** 0.270 6.89*** 0.265 6.72*** 0.265 6.72*** Age 36–59 months 0.054 1.51 0.044 1.29 0.046 1.34 0.044 1.29 0.044 1.29 Share householdmembers À0.670 À3.04*** À0.599 À2.74*** À0.600 À2.78*** À0.598 À2.73*** À0.600 À2.74*** age 0–5 (%) Share householdgirls À0.3067 À1.38 À0.211 À0.95 À0.198 À0.90 À0.210 À0.94 À0.212 À0.95 age 6–15 (%) Share householdboys À0.373 À1.62 À0.300 À1.33 À0.297 À1.33 À0.299 À1.33 À0.301 À1.33 age 6–15 (%) Share householdwomen À0.029 À0.12 0.019 0.08 0.022 0.09 0.020 0.08 0.019 0.08 16–25 (%) Share householdwomen 0.107 0.43 0.225 0.94 0.231 0.97 0.226 0.95 0.224 0.94 26–65 (%) Share householdmen 16–25 (%) 0.214 0.91 0.290 1.25 0.278 1.21 0.291 1.25 0.290 1.24 (Continued) T A B L E 5 . Continued District fixed effects Time varying Time dummy Heteroskedasticity OLS (1) OLS (2) IV (3)a IV (4) Identification (5) Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Share householdmen 26–65 (%) 0.348 1.46 0.424 1.79* 0.428 1.83* 0.424 1.79* 0.424 1.78* Number of householdmembers À0.026 À5.24*** À0.023 À4.64*** À0.023 À4.74*** À0.023 À4.63*** À0.023 À4.64*** Head is male À0.133 À3.41*** À0.095 À2.46** À0.097 À2.53** À0.095 À2.45** À0.096 À2.46** Head is indigenous À0.226 À4.68*** À0.272 À4.91*** À0.254 À4.77*** À0.272 À4.91*** À0.272 À4.91*** Mother’s age 0.063 3.75* 0.057 3.43*** 0.058 3.51*** 0.057 3.43*** 0.057 3.43*** Mother’s agesquared À0.001 À2.67*** À0.001 À2.52** À0.001 À2.54** À0.001 À2.52** À0.001 À2.52** Mother’sheight (centimeters) 0.054 23.07*** 0.051 22.65*** 0.052 23.10*** 0.051 22.65*** 0.051 22.65*** Mother’s education, primary 0.013 0.25 À0.021 À0.44 À0.017 À0.36 À0.022 À0.45 À0.021 À0.44 441 Mother’s education, secondary 0.072 1.19 0.020 0.35 0.028 0.49 0.020 0.34 0.020 0.35 Mother’s education, 0.116 1.70* 0.113 1.75* 0.120 1.89* 0.112 1.74* 0.113 1.75* postsecondary Father’s education, primary À0.002 À0.04 0.013 0.26 0.017 0.36 0.013 0.27 0.012 0.26 Father’s education, 0.108 2.38** 0.113 2.52** 0.113 2.53** 0.113 2.53** 0.113 2.52** secondary Father’s education, 0.012 0.24 0.104 2.00** 0.099 1.90* 0.104 2.00** 0.104 1.99** postsecondary House floor dirtdummy À0.052 À1.85* À0.088 À2.99*** À0.080 À2.82*** À0.089 À2.99*** À0.088 À2.99*** variable Piped drinking waterdummy À0.029 À0.97 0.028 0.95 0.031 1.05 0.028 0.95 0.028 0.96 variable Flush toilet dummy variable 0.107 2.93*** 0.111 3.08*** 0.096 2.7*** 0.111 3.08*** 0.110 3.07*** Urban dummyvariable 0.211 5.30*** District fixed effect Fixed effect Fixed effect Fixed effect (Continued) T A B L E 5 . Continued District fixed effects Time varying Time dummy Heteroskedasticity OLS (1) OLS (2) IV (3)a IV (4) Identification (5) Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic Constant À13.75 À22.12*** dummies dummies dummies dummies omitted omitted omitted omitted FONCODES index (t = 0) in IV equation 0.222 49.21*** FONCODES index (t = 1)in IV equation 0.311 53.63*** FONCODES index inIV equation 0.150 2.43** F-statistic (H0:instruments jointly 0) 153.6** 4329.0** w2 (H0: OLS and IVestimates same) 5.56* 6.27* 442 w2 (Breusch–Pagan test) 699.90** Number of observations 19,053 19,053 19,053 19,053 19,053 R2 0.300 0.356 0.341 0.355 0.355 OLS,ordinary least squares. *Significant at the 90 percent level of confidence. **Significant at the 95 percent level of confidence. ***Significant at the 99 percent level of confidence. a Province-level fixed effects. Note: Instrument in IV models is district-level FONCODES index. Source: Authors’ analysis based on data described in the text; see also appendix table A.1. Stifel and Alderman 443 nutritional outcomes of young children—the group to whom the program is directed—using either the preferred approach (controlling for the initial condi- tions in communities with district fixed effects) or other models. In both of the standard instrumental variable models, the identifying instrument (FONCODES index of unmet needs) is significantly correlated with per capita district expenditures, and the instruments overall (including the time dummy variable) are jointly significant at the 1 percent level.13 The first-stage parameter estimates for the FONCODES index for both 1996 and 2000 in the time-varying instrumental variable model (model 3) are positive and strongly significant, with a larger effect in 2000. While this confirms that the instrument is valid in terms of its correlation with Vaso de Leche expenditures, it also implies that marginal program targeting is propoor (Lanjouw and Ravallion 1999). The positive parameter estimates for the instruments suggest that the incidence of inframarginal Vaso de Leche spending benefits districts with higher FONCODES indices—the poor benefit more from marginal increases in program spending that may not be distributed homogeneously across all districts. In model 5, the chi-squared statistic for the Breusch–Pagan test of heteroskedas- ticity is 699.9, implying that the data in the first-stage equation are heteroskedastic. Thus, following Lewbel (2004), the condition is met for consistent estimation of the impact of Vaso de Leche expenditures on child health. Although the methodology differs from that of models 3 and 4, the parameter estimate is similar. A few additional specifications were also tried (these are not shown here but are available from the authors). For example, while no average impact is observed in these regressions, it is possible that the impact is greater among the poor, where malnutrition rates are higher. Thus, the regressions were rerun for only the poorest 40 percent. The point estimates for the coefficient of per capita Vaso de Leche expenditures remained negative and were greater in absolute value for all of the models than the results reported in Table 5.14 Thus, there is no indication that the impact on the poor was masked by aggregation. Similarly, the results are unchanged when children under six months of age are excluded, to rule out the possibility that children who were being breast-feed would respond less to an in-kind subsidy15and when the 13. Durbin-Wu-Hausman chi-square tests that the OLS and instrumental variable estimates are the same are all rejected at the 5 percent level (table 5). 14. While probit analysis does not use all of the information available in the data since the variation in the continuous dependent variable is ignored when it is converted to a binary variable, the model was also tested with a probit analysis as the threshold to keep the focus on the malnourished. The coefficient of Vaso de Leche expenditures remained nonsignificant. 15. In addition, the models were reestimated using other measures of nutritional status. The results were similar to the estimates presented here. More specifically, two measures were used: weight-for-height z-scores (WHZ), a measure of short-term nutritional status, and weight-for-age z-scores (WAZ), a composite of weight for height and height for age. A further test of robustness involved splittig the sample into urban and rural samples and estimating separate models. Since program leakage is higher in urban areas than in rural areas, a program effect might have been expected in rural areas, but the separate estimates did not differ substantially from those presented in this article. Therefore, an aggregation bias from pooling the urban and rural samples appears not to be driving the results. 444 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 parameters are allowed to differ across urban and rural areas in the fixed effects models. Nonsignificant coefficients in a population survey may mask a response if beneficiaries are a small share of the population. However, the program covers nearly half the population and two-thirds of the poor. Thus, more than 8,000 individuals in the sample are beneficiaries of the program. Nonsignificant coef- ficients may also reflect imprecision in measurement. In this case, the confidence intervals—while narrow—do cross over into positive values. Still, at the largest positive value for the 95 percent confidence interval for the impact of 37 soles ($11.50) of program expenditures per capita in, say, model 2, the program would increase z-scores by 0.09. V. CONCLUSION This article looked for evidence that the politically popular focus on the provi- sion of milk to young children can further nutritional objectives. It studied expenditures on the Vaso de Leche feeding program in Peru, a program that is reasonably progressive in its distributional impact relative to international experience (Grosh 1994; Coady, Grosh, and Hoddinott 2004). The article was motivated by the paucity of evidence on the nutritional impact of similar programs worldwide despite the millions of dollars spent on them each year. It illustrates a methodology for linking public expenditure data with household survey data first to substantiate the targeting and then to model the determinants of nutritional outcomes of children to see whether Vaso de Leche program interventions have an impact on nutrition. In the models of standardized child heights, the magnitude of program expenditures provided to the community rather than household participation is used as an explanatory variable, solving the issue of endogenous household choice. Even when accounting for endogenous program placement with fixed effects models, and further with alternative approaches to instrumental vari- ables, Vaso de Leche program expenditures are found to have no impact on the nutritional outcomes of young children—the group to whom the program is targeted. The results do confirm that the Vaso de Leche program is reasonably well targeted to poor households and to households with low nutritional status: some 50 percent of the poor received program benefits, while less than 20 percent of the nonpoor did. In value terms, more than 60 percent (possibly up to 75 percent) of the allocated Vaso de Leche budget goes to the poor. Therefore, the absence of a measurable impact on child growth is not likely explained by mistargeting. Indeed, given that the program expenditure has no observed impact on nutritional status other than that through any increase in household expenditures, further improvements in targeting are not by themselves likely to affect nutritional outcomes. To gauge this, the Stifel and Alderman 445 effect of redistributing all of the Vaso de Leche benefits received by the nonpoor evenly among the poor was simulated using the household expen- diture coefficient from the district fixed effects models. Malnutrition rates decreased only 0.28 percentage points. One possible reason why the impact of food subsidies beyond their value as income transfers is limited may be the degree to which the commodity transfers are inframarginal. The 2002 Public Expenditure Tracking Survey data show that transfers of milk and milk substitutes from the Vaso de Leche program are inframarginal for approximately half the households that receive them. This can be only a partial explanation, however, since the results do not change when the estimates are confined to the poorest 40 percent, a group for which milk is less likely to be inframarginal. There are, however, other means by which a subsidy to milk might achieve an improvement in child nutrition. For example, milk fortification may be a promising way to address anemia; programs in Argentina, Chile, and Mexico currently fortify milk with iron to achieve this objective. Randomization of fortification in two dozen communities in Mexico verified the efficacy of this approach (J. Rivera, pers. comm.). However, this was not an objective of the Vaso de Leche program in Peru during the period being studied. Another possible approach to milk subsidies is to include them as part of a broader program for improving nutrition, as with the Women, Infants, and Children program in the United States. The review by Rivera, Hotz, and others (2003) and an earlier article by Garcia and Pinstrup-Andersen (1987) indicate the efficacy of supplementation programs that include nutrition education. Again, the Vaso de Leche program is not directly embedded in other interven- tions aimed at improving nutrition, so the current study cannot assess this possibility. Still, as this study (and the few similar studies in the literature) fails to find an impact on child growth from a subsidy on commercial milk products, the results reinforce the view that without additional measures dairy subsidies do not address the efficiency objectives of a transfer program in addressing malnu- trition. Despite being reasonably well targeted to the poor and malnourished, the Vaso de Leche program fails to improve the nutritional status of young children. The implication for Peru and for other countries is that where costly in-kind transfers are also largely inframarginal and not fortified with micronutrients, they offer little if any efficiency gains as measured by improved child nutritional outcomes. Under such circumstances, cash transfers could be less costly and potentially an equally effective means of achieving nutritional and other distri- butional objectives. A P P E N D I X T A B L E A . 1 . Data Sources and Their Use in the Analysis Information on Vaso Anthropo Welfarist de Leche Use in Data source Type Year metrics metric participation analysis National Household Survey Household survey 1998 No Income Indicators Income targeting (Encuesta Nacional 1999 No Income Indicators Income targeting de Hogares, ENAHO) 2000 No Income Indicators Income targeting National Living Standards Household survey 1994 Yes Consumption Indicators Income and nutrition Survey (Encuesta Nacional targeting de Hogares sobre 1997a Yes Consumption Values Income and nutrition 446 Medicion de Niveles targeting de Vida, ENNIV) Demographic and Household survey 1996 Yes Asset index Indicators Nutrition targeting Household Survey and child nutrition models 2000 Yes Asset index No Child nutrition models Public Expenditure Expenditure 2001/02 No Asset index All households Inframarginality analysis Tracking Survey tracking/household surveyed are survey participants Vaso de Leche program District-level records 1994–2000 Nutrition impact expenditure data regressions a Data set used by Ruggeri Laderchi (2001). Stifel and Alderman 447 REFERENCES Alderman, H., P. Chiappori, L. Haddad, J. Hoddinott, and R. Kanbur. 1995. ‘‘Unitary Versus Collective Models of the Household: Time to Shift the Burden of Proof?’’ World Bank Research Observer 10(1):1–20. Alderman, H., and C. del Ninno. 1999. ‘‘Poverty Issues for Zero Rating VAT in South Africa.’’ Journal of African Economies 8(2):182–208. Becker, G. 1981. A Treatise on the Family. Cambridge, Mass.: Harvard University Press. Behrman, J. R., and J. Hoddinott. 2005. ‘‘Program Evaluation with Unobserved Heterogeneity and Selective Implementation: The Mexican Progresa Impact on Child Nutrition.’’ Oxford Bulletin of Economics and Statistics 67(4):547–69 Breusch, T., and A. Pagan. 1979. ‘‘A Simple Test for Heteroskedasticity and Random Coefficient Varia- tion.’’ Econometrica 47(5):1287–94. Carlson, A., and B. Senauer. 2003. ‘‘The Impact of the Special Supplemental Nutrition Program for Women, Infants, and Children on Child Health.’’ American Journal of Agricultural Economics 85(2):479–91. Coady, D. , M. Grosh, and J. Hoddinott. 2004. Targeting of Transfers in Developing Countries: Review of Lessons and Experience. Washington, D.C.: World Bank. ¨ zler. 2005. ‘‘Reassessing Conditional Cash Transfer Programs.’’ World Bank Das, J., Q.-T. Do, and B. O Research Observer 20(1):57–80. El Peruano. 2001. ‘‘Normas Legales.’’ Diario Oficial El Peruano 7650:203875–76. Esanu, C., and K. Lindert. 1996. An Analysis of Consumer Food Price and Subsidy Policy in Romania. Washington, D.C.: World Bank. Filmer, D., and L. Pritchett. 2001. ‘‘Estimating Wealth Effects without Expenditure Data—or Tears: An Application of Educational Enrollment in States of India.’’ Demography 38(1):115–32. Garcia, M. , and P. Pinstrup-Andersen. 1987. ‘‘The Pilot Food Price Subsidy Scheme in the Philippines: Its Impact on Income, Food Consumption, and Nutritional Status.’’ Research Report 61. International Food Policy Research Institute, Washington, D.C. Grosh, M. 1994. Administering Targeted Social Programs in Latin America: From Platitudes to Practice. Washington, D.C.: World Bank. ˜ ez, C. Valdez, and B. Kuhn. 2000. ‘‘A Comparison of Food Assistance Programs in Gundersen, C., M. Yan Mexico and the United States.’’ Economic Research Service Food Assistance and Nutrition Research Report 6. United States Department of Agriculture, Washington, D.C. Haddad, L., H. Alderman, S. Appleton, L. Song, and Y. Yohannes. 2003. ‘‘Reducing Child Malnutrition: How Far Does Income Growth Take Us?’’ World Bank Economic Review 17(1):107–31. Heckman, J., R. Lalonde, and J. Smith. 1999. ‘‘The Economics and Econometrics of Active Labor Market Programs.’’ In O. Ashenfelter and D. Card, eds., Handbook of Labor Economics, Vol. 3A. Amster- dam: Elsevier Science. Instituto Apoyo and World Bank. 2002. ‘‘Central Government Transfers to Municipalities in Peru: A Detailed Look at the Vaso de Leche Program.’’ Lima. Kennedy, E., and H. Alderman. 1987. ‘‘Comparative Analyses of Nutritional Effectiveness of Food Subsidies and Other Food-Related Interventions.’’ International Food Policy Research Institute, Washington, D.C. and World Health Organization–United Nations Children’s Fund, Geneva, Switzerland. Lanjouw, P., and M. Ravallion. 1999. ‘‘Benefit Incidence, Public Spending Reforms, and the Timing of Program Capture.’’ World Bank Economic Review 13(2):257–73. Levinger, B. 1986. ‘‘School Feeding Programmes in Developing Countries: An Analysis of Actual and Potential Impact.’’ AID Evaluation Special Study 30. United States Agency for International Develop- ment, Washington, D.C. 448 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 3 Lewbel, A. 2004. ‘‘Identification of Heteroskedastic Endogenous or Mismeasured Regressor Models.’’ Unpublished paper. Boston College, Boston, Mass. Martorell, R., and J. Habicht. 1986. ‘‘Growth in Early Childhood in Developing Countries.’’ In F. Falkner and J. Tanner, eds., Human Growth: A Comprehensive Treatise, Vol. 3. New York: Plenum Press. Moffitt, R. 1991. ‘‘Program Evaluation with Nonexperimental Data.’’ Evaluation Review 15(3):291– 314. Powell, C., S. Grantham-McGregor, S. Walker, and S. Chang. 1998. ‘‘School Breakfast Benefits Children’s Nutritional Status and School Performance.’’ American Journal of Clinical Nutrition 68(4):873–79. Rigobon, R. 2003. ‘‘Identification through Heteroskedasticity.’’ Review of Economics and Statistics 85(4):777–92. Rivera, J., C. Hotz, T. Gonza ´o, L. Neufeld, and A. Garcı ´lez-Cossı ´a-Guerra. 2003. ‘‘The Effect of Micronutrient Deficiencies on Child Growth: A Review of Results from Community-Based Supple- mentation Trials.’’ Journal of Nutrition 133(November):4010S–20S. Rivera, J., D. Sotres-Alvarez, J.-P. Habicht, T. Shamah, and S. Villalpando. 2003. ‘‘Impact of the Mexican Program for Education, Health, and Nutrition (Progresa) on Rates of Growth and Anemia in Infants and Young Children: A Randomized Effectiveness Study.’’ Journal of the American Medical Associa- tion 291(4):2563–70. Rosenzweig, M., and K. Wolpin. 1986. ‘‘Evaluation of the Effects of Optimally Distributed Programs: Child Health and Family Planning Initiatives.’’ American Economic Review 76(3):470–82. Rouse, C. 1998. ‘‘Private School Vouchers and Student Achievement: An Evaluation of the Milwaukee Parental Choice Program.’’ The Quarterly Journal of Economics 113(2):553–602. Ruggeri Laderchi, C. 2001. ‘‘Killing Two Birds with the Same Stone? The Effectiveness of Food Transfers on Nutrition and Monetary Policy.’’ Paper presented at the Latin American Economic Association Conference, October 18–20, Montevideo, Uruguay. Rush, D., J. Leighton, N. Sloan, J. Alvir, D. Horvitz, W. Seaver, G. Garbowski, S. Johnson, R. Kulka, and J. Devore. 1988. ‘‘The National WIC Evaluation: Evaluation of the Special Supplemental Food Program for Women, Infants, and Children. VI. Study of Infants and Children.’’ American Journal of Clinical Nutrition 48(2 Suppl.):484–511. Sahn, D., and D. Stifel. 2002. ‘‘Parental Preferences for Nutrition of Boys and Girls: Evidence from Africa.’’ Journal of Development Studies 39(1):21–45. ———. 2003. ‘‘Exploring Alternative Measures of Welfare in the Absence of Expenditure Data.’’ Review of Income and Wealth 49(4):463–89. Schady, N. 1998. Picking the Poor: Indicators for Geographic Targeting in Peru. Washington, D.C.: World Bank. Stifel, D., and H. Alderman. 2005. ‘‘Targeting at the Margin: The ‘Glass of Milk’ Subsidy Program in Peru.’’ Journal of Development Studies 41(5):839–64. Strauss, J., and D. Thomas. 1995. ‘‘Human Resources: Empirical Modeling of Household and Family Decisions.’’ In J. Behrman and T. Srinivasan, eds., Handbook of Development Economics, Vol. 3A. Amsterdam: Elsevier Science. Tuck, L., and K. Lindert. 1996. ‘‘From Universal Food Subsidies to a Self-Targeted Program: A Case Study of Tunisian Reform.’’ Discussion Paper 351. World Bank, Washington, D.C. World Bank and Inter-American Development Bank. 2002. ‘‘Peru, Restoring Fiscal Discipline for Poverty Reduction: Public Expenditure Review.’’ Report 24286-PE. Washington, D.C.: World Bank. World Bank. 1999. Poverty and Social Development in Peru, 1994–1997. Washington, D.C.: World Bank. World Bank. 2003. ‘‘Brazil, Growth and Poverty Reduction in Rio Grande do Norte: A State Economic Memorandum.’’ Report 24891-BR. World Bank, Washington, D.C. Younger, S. 2002. ‘‘Public Social Sector Expenditures and Poverty in Peru.’’ In C. Morrison, ed., Education, Health Expenditure, and Development: The Cases of Indonesia and Peru. Paris: OECD Development Centre.