Policy Research Working Paper 8377 Cash Transfers, Food Prices, and Nutrition Impacts on Nonbeneficiary Children Deon Filmer Jed Friedman Eeshani Kandpal Junko Onishi Development Research Group & Social Protection and Labor Global Practice March 2018 Policy Research Working Paper 8377 Abstract Based on a randomized evaluation, the paper shows that a in their experimental sample against price information from household-targeted Philippine cash transfer program sig- nationally-representative household expenditure surveys nificantly raised the local price of key foods relevant for across the 6-year rollout of the program. Failing to consider child nutritional status. This shift in prices increased stunt- such general equilibrium effects may overstate the net ben- ing among young nonbeneficiary children by 34 percent efits of targeted cash transfers. In areas where targeting of (11 percentage points). Price and stunting effects increase in social programs covers a large proportion of the population, program saturation; at median saturation, the village income offering the program on a universal basis may avoid such shock is 15 percent. These effects persist 2.5 years after pro- long-lasting negative impacts at moderate additional cost. gram introduction. The authors confirm the price patterns This paper is a product of the Development Research Group and the Social Protection and Labor Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The corresponding author may be contacted at ekandpal@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Cash Transfers, Food Prices, and Nutrition Impacts on Nonbeneficiary Children1 Deon Filmer, Jed Friedman, Eeshani Kandpal, Junko Onishi JEL Codes: H23; I38; O12; O15; O20 Keywords: cash transfers, general equilibrium effects, malnutrition, food prices, Philippines 1 All authors are with the World Bank. Corresponding author: ekandpal@worldbank.org. We thank Pablo Acosta, Richard Akresh, Harold Alderman, Jorge Avalos, Sarah Baird, Jishnu Das, Gabriel Demombynes, Brian Dillon, Quy-Toan Do, Emanuela Galasso, John Gibson, Francisco Ferreira, Jessica Goldberg, Seema Jayachandran, Yusuke Kuwayama, David McKenzie, Paul Niehaus, Owen O’Donnell, Owen Ozier, Berk Özler, Aleksandra Posarac, Martin Ravallion, Adam Wagstaff, Sarah West, and participants of the APPAM 2016, NEUDC 2016, AEA 2017, CSAE 2017, IHEA 2017, and SEA 2018 conferences, as well as seminar attendees at Georgetown University, the University of Tokyo, and Virginia Tech for their comments on this work. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. I. INTRODUCTION We leverage a randomized evaluation of the Philippine flagship anti-poverty program, a conditional cash transfer, to investigate impacts on prices of key food goods and implications for nonbeneficiary welfare. Conditional and unconditional cash transfer programs reach 380 million people in developing countries people around the world (World Bank 2014), and an extensive literature has considered the direct and indirect effects of such programs (Fiszbein and Schady 2009; Baird et al. 2013; Saavedra and Garcia 2012; Hanlon, Barrientos, and Hulme 2010). However, the study of possible general equilibrium effects has been limited to three potential channels. Evidence suggests significant general equilibrium effects can occur through the labor market (Muralidharan, Niehaus, and Sukhtankar 2017) as well as informal insurance and credit markets (Angelucci and De Giorgi 2009) but that there are no or limited food market responses (Attanasio and Pastorino 2019, Beegle, Galasso, Goldberg 2017; Cunha, De Giorgi and Jayachandran 2019; Haushofer et al. 2019).2 A central contribution of our paper is to identify meaningful local food price increases from a targeted cash transfer program, particularly for perishable and less-easily traded protein rich foods. This finding is consistent with a conceptual framework distinguishing between the relative price effects of cash transfers on easily traded and less easily traded goods.3 We then show that these price increases affect important outcomes: children of nonbeneficiary households have significantly worse nutritional status, an effect that we trace to a decline in protein intake. Child malnutrition, especially stunting, has lifelong economic consequences through its effects on human capital accumulation (Galasso and Wagstaff 2019; Vogl 2014; Currie and Vogl 2013; Hoddinott et al. 2013). 2 Attanasio and Pastorino (2019) show that while cash transfers from Mexico’s PROGRESA/Oportunidades program did not affect the average price of staples, they did steepen the slope of the price schedule for those goods. Beegle, Galasso and Goldberg (2017) find that a large-scale public-works program in Malawi significantly worsened food security for nonbeneficiary households in treated villages. The authors rule out prices, as well as crowding out of traditional risk-sharing mechanisms, as channels for the impact but are otherwise unable to pinpoint the exact channels. Cunha, De Giorgi, and Jayachandran (2019) study the impact of cash transfers on selected food prices in the context of a comparison of in-kind and cash transfers. They do not find that cash transfers led to significant food price effects, including in communities identified as less developed or that have less integrated markets. An earlier version of their paper (Cunha, De Giorgi, and Jayachandran 2015) documented a six percent increase in food prices in remote communities. Ongoing work in Kenya tests the effects of transfers on prices and wages in a context of low treatment saturation and find limited evidence of spillovers (Haushofer et al. 2019). 3 Broadly speaking, we use the terms tradable and non-tradable to distinguish goods that have relatively low versus high transportation and storage costs. We define the terms tradable and non-tradable precisely below. 2 Our paper thus ties local price effects of anti-poverty cash transfers to welfare outcomes among nonbeneficiaries.4 We do this using household-level data on health, nutrition, and food prices that we collected, and village-level data from the rollout of the Philippine cash transfer program, Pantawid Pamilya Pilipino Program. Stunting rates are high and there is substantial spatial variation in market integration and poverty levels in the Philippines, creating an environment conducive to these important—and pernicious— spillovers. That said, stunting rates remain high around the world (UNICEF 2019) and the fragmentation of local markets causing persistent price differences is found in many low-income countries.5 Our identification is as follows. First, we use village-level randomization of the program to estimate its impact on beneficiary households’ food budget shares, and child nutrition and food intake.6 After establishing a demand increase for nutritious foods among beneficiary households, we compare prices in treated and control villages to find that the program increased the prices of key protein rich foods. We then show that this price effect increases in saturation, where saturation is the proportion of potentially eligible households at baseline in both treated and control villages. Our identification thus leverages the randomized treatment assignment of the program in similarly 4 The unintended consequences of cash transfers have more broadly been studied at three levels: (1) the beneficiary households themselves; (2) participating schools, hospitals and other facilities; and (3) local markets. Within the household, evidence generally suggests the presence of positive spillovers through the information channel on nontargeted siblings’ education and work-for-pay, as well as adult health (Ferreira, Filmer, and Schady 2017; Contreras and Maitra 2013). Evidence also suggests that nontargeted outcomes for beneficiaries may be affected in unexpected ways: Bourguignon, Ferreira, and Leite (2003) and de Hoop et al. (2019) show that conditional cash transfers (CCTs) can increase schooling and work-for-pay by the same children. Within participating facilities, there is robust evidence of peer effects-driven increases in schooling enrollment of nontargeted populations, at least in the case of PROGRESA/Oportunidades (Bobba and Gignoux 2019; Bobonis and Finan 2009; Lalive and Cattaneo 2009). Within affected areas, evidence suggests a variety of positive externalities, including improvements in mental health in Kenya (Haushofer and Shapiro 2016) and reductions in crime and political violence in the Philippines (Crost, Felter, and Johnston 2014). Finally, there is some evidence that Pantawid increased clientelism, which in turn improved political stability (Labonne 2013). In contrast, there is also evidence in other settings of a complete absence of externalities: using data from a Nicaraguan CCT, Macours, Schady, and Vakis (2012) do not find any evidence of externalities on nonbeneficiaries. 5 Atkin and Donaldson (2015) find that the effect of log-distance on trade costs within Ethiopia or Nigeria is four to five times larger than in the US, suggesting that transportation costs significantly restrict spatial price convergence in poor countries. Related studies on the construction of new roads find similar market-level impacts in India (Aggarwal 2018) and Vietnam (Mu and van de Walle 2011). 6 As the program is targeted at the household level, eligibility is set through a Proxy Means Test that estimates household resources on the basis of a limited number of observed characteristics. We estimate direct effects on eligible households by comparing across households below the treatment threshold in treated and control villages. Spillovers on nonbeneficiaries are calculated by comparing households above the treatment threshold in treated and control villages. The extent to which a village is defined as a natural market is an open question. Markets in treated and control villages might interact, and we might be missing even more negative spillovers occurring outside the local area. Both channels would mean that the spillovers we capture, if anything, are a lower bound. 3 poor treated and control villages. We confirm that these patterns in the experimental sample are consistent with nationally representative data by comparing village-level unit values for the range of commonly consumed food items measured in national household expenditure surveys across the phased 6-year rollout of the program.7 The size of the aggregate income transfer at the village level, determined by program saturation, is an important factor in determining the extent to which local prices shift. The transfers amounted to an aggregate increase in village income of about 15 percent. (Median saturation in the study sample is 65 percent, but in some villages as many as 95 percent of the households received the transfer.) This increase in village income did not affect the prices of storable and easily traded food goods but raised the local prices of protein-rich perishable foods, such as eggs and fresh fish, by 6 to 8 percent. The rise in relative prices decreases nonbeneficiaries’ real income and leads them to substitute away from protein-rich foods thus creating significant negative effects on the nutritional status of children in nonbeneficiary households. We show that young nonbeneficiary children experienced a 0.4 standard deviation decrease in height-for-age z-scores (relative to a control mean of -1.1 standard deviations), which led to an 11-percentage point increase in stunting (relative to a control mean of 32 percent). Children in beneficiary households, which were compensated for the price change by the increase in household income, show a gain in nutritional status. The price increases for non-tradable goods and the reductions in child growth both increase in the proportion of households eligible for the transfers. This finding is consistent with our interpretation of the stunting increase as an unintended spillover effect of the program through local food markets. Perhaps surprising at first glance, these impacts on child growth translate to estimated elasticities very much in line with the broader literature and consistent with changes in overall consumption patterns. Our results are also consistent with a nascent literature that highlights the particularly important role of animal-sourced proteins in avoiding child stunting (Heady, Hirvonen, and Hoddinott 2018). Using estimates from the nutrition literature of the elasticity of child height to eggs (Puentes et al. 2016; Iannotti et al. 2017), we show that the observed reduction 7 Unit values are defined as the reported expenditure on a particular commodity or commodity group divided by the reported quantity consumed. While unit values combine price effects and quality effects, many of the key goods we examine, including regular commercially-milled rice and eggs, are fairly homogeneous with relatively little scope for quality differentiation. 4 in the consumption of eggs alone would explain much of the observed decline in height-for-age. Further, the observed household consumption patterns are consistent with increases in real income for beneficiaries and decreases for nonbeneficiaries. Beneficiary households spend more on nutrient-rich foods than do nonbeneficiary households (as typically found in CCT programs (Fiszbein and Schady 2009)) and nonbeneficiary children consume less of the expensive protein- rich items, compared with households of similar income levels in control villages. We consider and rule out several alternative explanations for the observed increases in prices and child stunting rates, including the availability of adult care-givers, changes in household composition, and changes in adult and older child productive activity, all of which can affect child health independently of food prices. We can also rule out changes in demand for health services by beneficiaries crowding out utilization by nonbeneficiaries.8 Our paper makes four main contributions. First, we identify local price increases from a cash transfer program on nutritionally important foods. Second, we discuss the conditions under which these price responses arise as they are not universal in our study setting. Third, we show that these price increases are restricted to a certain class of goods, namely perishable goods with high transport costs. These goods are largely animal-based protein-rich foods that recent literature indicates to be a critical input in the production of child height. Fourth, we link these price increases to deleterious and presumably long-lasting health impacts among young children. Since we are able to tie the increase in stunting to a change in the relative price of proteins, the fourth contribution highlights the role of this nutrient for early life growth. Our findings also contribute to the literature on the benefits and drawbacks of targeting cash transfers (Hanna and Olken 2018; Klasen and Lange 2016; Ravallion 2009). We show that general equilibrium effects can be important when a large proportion of the local population is eligible for targeted benefits and thus the aggregate increase in income is big enough to shift local prices. Of course, such effects may be concentrated in the poorest areas; across the Philippines as a whole in 2015, only 4.2 percent of villages had program coverage of 65 percent or more. Given the important long-term consequences we identify here, the choice of benefit targeting mechanism should consider whether local markets are integrated with surrounding 8 In our study, access to government services, including health clinics, electricity, running water does not vary by saturation, suggesting that our results are not driven by differential access to these services. 5 regions. In high saturation situations, particularly in remote areas, a geographic or community- based targeting rule may mitigate negative impacts (Elbers et al. 2007). II. CONCEPTUAL FRAMEWORK To illustrate how price effects might arise from a cash transfer program, we present a framework that begins with the basic conjecture that an exogenous increase in incomes in a local market (village) increases demand for normal goods, including many food goods. In a large integrated economy, a demand shift in either one or a handful of local areas should not significantly affect prices as the local demand increase is too small to influence aggregate demand. However, if markets are not integrated then the local market structure determines whether local demand changes affect local prices. The shape of the local supply curve and the size of the demand shock determine the resulting local equilibrium price level. The supply response to a demand increase may be constrained for a variety of reasons. First, poorer villages tend to suffer from high transport costs for imported goods as the villages tend to be more remote, and the transport cost wedge may offset any marginal gain in profit from importing units to sell at a marginally higher price. Second, if local production markets are oligopolistic (perhaps due to a fixed cost of entry) or competitive but with upward sloping marginal costs, then price increases will also likely arise from a demand shift. Central to our analysis is the observation that local market characteristics, including characteristics of the local production base and the degree of integration of the local producers with wider networks of producers and consumers, will vary with the type of traded good. Perishable protein rich foods have higher transport and storage costs than storable food goods and thus the local producers and consumers of these goods may be less integrated with national suppliers. Consequently, the price behavior of these goods can diverge from that of more easily traded goods when faced with a shift in local demand. This conceptual framework is closely related to the one in Cunha, De Giorgi and Jayachandran (2019). We add to their framework the notion of markets bifurcated by the nature of the good and posit that cash transfers are most likely to affect prices in goods markets where producers are not fully integrated with the broader national production base. This is especially true in poorer and more rural parts of developing countries that rely more on local sourcing. 6 Price Responses to Increases in Demand Income gains from cash transfers will increase demand for normal goods as, by definition, the income elasticities of demand for these goods are positive.9 Additionally, demand for nutritious foods may increase since anti-poverty programs such as conditional cash transfers (CCTs) typically broadcast messaging on recommended child feeding practices that can further shift demand. If supply does not fully respond to this demand shift, then prices will rise. Since any potential price change is related to the magnitude of the increase in demand, program saturation will be a relevant program feature. The higher the proportion of beneficiaries in the local economy, the greater the increase in aggregate demand and hence the greater potential for relative price change. When an increase in demand for a good occurs in a competitive market with constant marginal production costs then the increased quantity demanded will be accommodated without a change in price. However, even a competitive goods market will translate a demand shift into higher equilibrium prices if the marginal cost of production increases with quantity. Alternatively, if the goods market is supplied by a limited number of producers, then oligopolistic competition can translate a demand increase into higher consumer prices. Consider, for example, a Cournot model with N producers of a homogenous good. We model total quantity demanded, Q, as: = (, ) where p is the good’s price and X a demand shifter. If the demand function is additive in X, such that = () + , then the Cournot-Nash solution can be written as follows: 1 =− >0 () ( ) In other words, a demand shift will raise the price, while the magnitude of the increase depends on the shape of the demand curve and the number of producers in the market (which, in turn, is partly a function of the fixed cost of entry for potential competitors). This basic result holds for any normal good demand function with an additive shifter X. 9 Del Boca, Pronzato and Sorrenti (2018) show that a CCT in Italy increased beneficiaries’ demand for more nutritious foods. 7 Local Market Structure Interaction with Goods Characteristics We now apply the general framework above to the specific setting of an isolated local market. The same logic applies: if the local market is competitive with constant marginal costs, then prices will not change with an increase in local demand. If instead marginal costs increase in total quantity or the local market is not perfectly competitive, then prices can rise alongside a rise in local demand. The key difference for a local market is that the transport and storage costs of a good drive a wedge between the local and (presumed competitive) national price for that good. If the national competitive price of a good is pn, and the import cost into the local economy is δ, then any local good’s price, pl, as determined solely by local market forces, can be sustained as long as ≤ + The import cost and national price define both an upper and a lower bound on local prices. Which constraint binds depends on whether the local market is a net producer or consumer of the good. When the local price exceeds the expression to the right of the inequality above, then arbitrage opportunities arise that would compensate for the cost of importation into the village – the local price of the good stabilizes at the national price plus the import and storage cost. This good-specific import and storage cost determines the “tradability” of the good between the local and national markets. For goods whose δ is relatively low, the local market price will never substantially deviate from the national competitive price. However, there is far more scope for local market price deviations for perishable goods, such as fresh eggs, dairy, or fish, that would need relatively expensive dedicated technologies for transport and storage. For convenience, we term such goods, where δ >> 0, “non-tradables,” and conversely “tradables” are goods for which δ ≈ 0. Therefore, in addition to program saturation and market structure, the tradable nature of the specific good in question is another factor that should determine the price response. Imperfect competition is possibly a more realistic model for the local markets of many non- tradables, given the smaller number of local producers for certain goods and the fixed costs to entry in these markets (such as the purchase and maintenance of livestock or fishing boats). We apply the imperfect competition framework above, with a linear shifter in the demand function, to local markets bifurcated by the type of the good, where we index a good by T or NT depending on whether the homogenous good is a tradable or non-tradable good. In this case, the effect of 8 increases in demand (i.e., ∂XT > 0 and ∂XNT > 0) on the equilibrium price will depend on the nature of the good and the initial local price level relative to the national price and import cost. For tradables, we have = 0, ≅ + where we assume that import costs are low enough for these tradable goods to be bought from and sold to large integrated markets. For non-tradables that have a higher cost of import , 1 = − > 0, < + ( ) ( ) or = 0, ≅ + While arbitrage limits the magnitude of the change in the non-tradable good’s price from a demand shift, the price increase may still be enough to affect consumption choice if δ is large enough.10 In this framework, a price increase for non-tradables can be sustained indefinitely if the new equilibrium local price does not exceed the national price plus import cost. Our results are measured after 31 months of program exposure, suggesting that, even after a considerable amount of time, local supply responses to offset the demand increase from the cash transfer do not return the local price to the pre-program level. Since we do not observe the number of local producers for perishable food goods, the fixed cost of entry, or the price threshold that may entice traders to import the good into the village, we cannot incorporate information about the structure of local markets into our analysis. This framework is used only for illustrative purposes; our empirical analysis focuses on prices pre- and post-program introduction, investigates the degree of good-specific price change, and how the magnitude of change is correlated with factors likely to cause a relative price rise in isolated 10 For simplicity, we assume no cross-price elasticities between tradables and non-tradables. The basic intuition of differential price changes by goods type holds in a fuller model that allows for the two types of goods to be complements or, more likely in this context, substitutes. 9 markets. These factors include the degree of program saturation and the degree of tradability of the particular good. III. THE CASH TRANSFER PROGRAM AND EXPERIMENTAL DESIGN The Pantawid cash transfer program has been in place since 2008 and is targeted to individual households based on a proxy means test for household income with eligibility cutoffs determined by province-specific poverty lines. Starting with an initial pre-pilot of 6,000 households, the program reached approximately 4.5 million households by 2015 (DSWD 2015). For comparison, the Indonesian Program Keluarga Harapan covered 1.5 million households after five years (Nazara and Rahayu 2013) and the fully scaled-up Mexican PROGRESA/Oportunidades program covered 5.8 million households (World Bank 2014). The program gives eligible households cash transfers if the households enroll children in school and use maternal and child health services. Eligibility is granted if a household not only has a proxy means test score below the provincial poverty line but also contains children ages 0 to 14 years or a pregnant woman at the time of assessment. Eligible households receive a combination of health and education grants every two months, ranging from ₱500 to ₱1,400 (approximately US$11 to US$32) per household per month, depending on their number of eligible children and compliance with program conditions.11 The expected transfer size equaled approximately 23 percent of per capita beneficiary consumption. By way of comparison, the Mexican CCT PROGRESA transfers were about 22 percent of beneficiary consumption, while those from the Brazilian CCT, Bolsa Família, were about 12 percent (Fiszbein and Schady 2009). Our analysis is based on an experimental design implemented during the pilot phase of the program. A village-level randomized evaluation, stratified by eight purposively selected municipalities in four provinces (in some of the poorest areas of the country), was implemented 11 Poor households with children ages 0 to 14 years or pregnant women receive a lump sum health grant of ₱500 (about US$11) per household per month if (i) all children under age five attend growth monitoring visits at the local health center; (ii) pregnant women seek regular antenatal care; (iii) school-age children (6 to 14 years) accept school-based deworming; and (iv) a household member attends monthly health and nutrition workshops. Households can also receive a monthly education transfer of up to ₱300 (about US$6.50) per child (for three or fewer children) enrolled in and attending at least 85 percent of school days in primary or secondary school for the duration of the school year. 10 for the pilot phase of Pantawid to inform program scale-up (Chaudhury, Friedman, and Onishi 2013). The Household Assessment Form to estimate proxy means scores for beneficiary selection was fielded in these eight municipalities between October 2008 and January 2009; we use these data to calculate potential program saturation at baseline. A total of 130 villages in the eight municipalities were then randomly assigned to treatment or control status with equal probability.12 A follow-up detailed household survey was conducted in October and November 2011, thus allowing for a program exposure period of 30 to 31 months. IV. EMPIRICAL STRATEGY AND DATA Empirical Strategy Our analysis uses variation across villages in treatment assignment to identify program impact on local prices and child nutrition.13 To study the effects of these price changes on household behavior, we analyze the direct program impacts among eligible households and then estimate the spillover effects among nonbeneficiary households, defined as non-poor households that nonetheless have eligible children. Since potential program saturation is a key mediator of the cash transfer’s impact on prices of non-tradable goods, we allow the indirect effect to vary by program saturation. A secondary analysis exploits the staggered rollout of the program in a national sample of villages to estimate the covariation of food prices with village-level program exposure. Data 12 The experimental sample selection occurred in three stages. First, provinces that did not have the program as of October 2008 were selected. Three of the 11 available provinces were excluded due to security concerns. Four of the remaining eight provinces were chosen (Lanao Del Norte, Mountain Province, Negros Oriental, and Occidental Mindoro) to span the three macro areas of the country (North, Visayas, and Mindanao). Second, in each province, two municipalities were selected to represent the average poverty level of areas covered by the program in 2008. Third, villages in each municipality were randomly assigned to treatment or control by a computer based random number generator. 13 We thus identify impacts of the program by comparing similarly (non) poor households in treated and control villages. Our SUTVA thus relies on a lack of spillovers across villages. However, even if SUTVA is violated across villages, i.e. prices are higher in control villages due to connected markets or something, we might except higher stunting in those villages, leading our estimates to be a lower bound of the impact of Pantawid on prices and stunting. In addition, we have no evidence even of within village effects on wages. 11 We use four data sources in our analysis. Our primary data source is the specialized household survey data from the randomized evaluation of Pantawid that contains information on consumption, select food prices, child anthropometry, health care utilization, labor supply, demographic composition and other behavior. We also use village-level administrative data on program saturation, data on unit values of food prices from the national household budget survey, as well as the Household Assessment Form that determined program-eligible households in 2008. Household Data from the 2011 Pantawid Impact Evaluation 2,555 households were surveyed in the eight study municipalities during the follow-up survey.14 With an eye toward investigating potential spillovers on nonbeneficiary households, the study population included 1,418 households that were eligible for the transfers, as well as 1,137 households that had proxy means test scores above the provincial threshold but had age-eligible children or pregnant women. For anthropometric outcomes, we focus on children ages 6 to 36 months at the time of the follow-up survey, while for food intake, we consider children aged 6 to 60 months. Further details on sample sizes, program take up, anthropometric measures, and the robustness of our results to trimming cutoffs for the anthropometric measures are presented in Annex 1. Program Saturation We have comprehensive village-level data on the number of households enrolled in Pantawid each year from 2009 to 2015 as reported by the provincial office of the Department of Social Welfare. These data allow us to calculate annual village level program saturation, defined as the number of beneficiary households for that year divided by census bureau estimates for the total number of households in the village in 2015. We use this measure in conjunction with the food goods unit value information from the household budget survey. We also calculate potential program saturation for the villages in our experimental survey sample, defined as the number of households with a proxy-means test score below the province- specific poverty threshold divided by the village population. As the municipalities selected at this stage of the program were among the poorest in the Philippines, it is no surprise that in many of 14 Household-level attrition from the baseline sample was 11.4 percent (80 of 624 households) in treated villages and 11.2 percent (80 of 634 households) in control villages, with no evidence of a significant difference by treatment status (Chaudhury, Friedman, and Onishi 2013). 12 the study villages, a high proportion of the total population was eligible to receive program benefits. Figure 1 presents a histogram of this proportion of eligible households from among all households in the village for the treatment and control villages. Although there is a good degree of dispersion in this saturation measure, some villages have up to 90 percent of the household population eligible to receive benefits. The median village saturation level in the experimental sample is 65 percent, and the mean is 62 percent. This figure also suggests that program saturation was balanced across treatment and control villages: a two-sample Kolmogorov-Smirnov test has a p-value of 0.945, suggesting equality of saturation in the two distributions. Highly saturated villages are equally likely to be found in treatment and control suggesting that the saturation distribution is not significantly different. Household Budget Survey Data We use the 2009, 2012, and 2015 rounds of the Philippine national household budget survey, the Family Income and Expenditure Survey (FIES), which collected detailed consumption data in 8,500 villages, to calculate unit values of food items. There are 93 food items common across all three rounds of FIES; we construct a unit value for each food item, and then aggregate these into 15 food categories provided by FIES.15 Annex 2 details how we constructed these data. We relate the median village-level unit values for these 15 categories with the village’s Pantawid program saturation in 2009, 2012, and 2015 to study the relationship between Pantawid and food prices. 16 The Household Assessment Form and the Balance of Baseline Data The Household Assessment Form survey used to assess household eligibility for Pantawid also serves as a baseline for our analysis. This information, which is relatively limited in scope, is primarily used to determine the potential program saturation of each village in the RCT sample as well as assess balance across treatment and control villages for the sociodemographic and economic information collected. We explore baseline balance for 15 The 15 categories are: regular rice; other rice; roots and tubers; fresh eggs; processed eggs; fresh fish; processed fish; fresh meats; processed meats; fresh fruits; fresh vegetables; processed rice and grains; coffee, cocoa, and tea; sugar; and milk products. 16 An important distinction between the analysis of FIES data and our core analysis using the Pantawid survey is that the FIES analysis exploits the staggered rollout of Pantawid throughout the country and relates village level prices with saturation. Unlike our core analysis, which relies on the randomized assignment of villages to treatment, this analysis does not leverage experimental variation. 13 beneficiary and nonbeneficiary households respectively. Appendix Tables 1 and 2 present the control means, difference between treated and control means, the corresponding p-values for tests of equality, and standardized mean differences. Overall, the samples of beneficiaries and nonbeneficiaries appear balanced. Among beneficiary households, none of the 28 comparisons is imbalanced. For the nonbeneficiary households, balance is almost as comprehensive: only three of 28 comparisons between treated and control areas are significantly different at the 10% level or less. Nonbeneficiary households in treated villages are less likely to own video recorders or motorcycles, and more likely to have walls made of light materials. However, overall wealth, as measured by the logged proxy means test score, is highly balanced between treated and control areas. Furthermore, no standardized mean difference is greater than 0.25, which again suggests a significant degree of balance (Imbens and Wooldridge, 2009). Finally, since the program was randomized at the village level, we can also use the omnibus test of joint orthogonality. The omnibus test has an F-statistic of 2.03, suggesting that the baseline characteristics are jointly unrelated with treatment status. Because we carry out analyses with interactions between treatment and saturation, we also explore baseline characteristic balance among nonbeneficiary households for those in villages above and below the median saturation level (Appendix Tables 3 and 4), and those in the top quartile of saturation (Appendix Table 5). These comparisons confirm the overall balance of the experiment. For nonbeneficiary households in the above-median saturation areas, 22 of 28 comparisons, including the proxy means test score of wealth, household head’s education, and children’s school attendance, are balanced.17 Treated households are slightly smaller and slightly more likely to have light roof and wall materials. They are also slightly less likely to own a video recorder, telephone, or refrigerator than households in control areas. One standardized mean difference, for household size, exceeds the 25 percent threshold that may imply concern. In areas in the top quartile of saturation, 18 of 28 comparisons, again including proxy means test of wealth, are balanced. In this level of saturation, however, children aged 6 to 11 years in treated villages 17 To rule out differential access to services driving the observed effects on prices and child anthropometry, we compare above median saturated treated and control villages. We do not find evidence of such differential access to health clinics or electricity. Although none of the comparisons are significant, if anything, the point estimates suggest treated villages have slightly higher levels of access: 80 percent of all above-median saturated control villages have a health unit in the village while 90 percent of above-median saturated treated villages do. Similarly, 63 percent of all households in above median saturated control villages and 66 percent in treated villages have access to electricity. 14 are more likely to have been in school than those in control villages. As in above median saturated areas, treated households tend to be a little smaller, although this time the standardized mean difference is below the 0.25 threshold, at -0.19, and are less likely to have strong wall materials, or own a video recorder, stereo or refrigerator. However, treated households in these villages are also more likely to have their own toilet and are less likely to share a water source. We explore any potential effects of this imbalance in household size in further detail in the next section. In the below-median saturated areas, households are balanced along all 28 dimensions. V. RESULTS Impacts on Beneficiaries Direct Anthropometric Impacts Pantawid incentivized the health- and education-related behavior concerning children in beneficiary households. Kandpal et al. (2018) show that the program reduced severe stunting among young beneficiary children. Table 1A reproduces these main impacts among program beneficiaries on outcomes related to these targets. A range of nutrition indicators was investigated, as reducing childhood malnutrition is one of the main goals of Pantawid. The considered age group for these indicators is children ages 6 to 36 months, as these children transit a critical developmental period for physical growth—often referred to as “the first 1,000 days.” Children in this age range also are likely to have lived most or all their lives exposed to the program. The estimated impact on the height-for-age z-score suggests an improvement of 0.3 standard deviation (although this is somewhat imprecisely measured). The estimated impacts on the likelihood of stunting suggest a reduction of 2 percentage points, and on the likelihood of severe stunting a reduction of 9.3 percentage points (with the latter being significantly different from zero at the five percent level). Stunting is the most commonly used measure of chronic malnutrition, believed to reflect extended periods of inadequate nutritional intake and/or chronic infection. No program impacts were found on other measures of severe or acute malnutrition, such as wasting or severe 15 wasting.18 For the average beneficiary child there may only have been a limited improvement in nutrition status, but for the most disadvantaged there was a marked improvement.19 Direct Expenditure and Consumption Effects A channel through which the cash transfer may have improved child anthropometry is through an increase in the consumption of food goods associated with increases in child height- for-age. We look for evidence of this in two ways, first with respect to reported spending patterns of various food goods and, second, with the reported food intake of young children. Table 1B reports the program impacts on the household food budget share in column 1 and, in columns 3 to 5, the impacts on consumption of a selected number of individual food goods by children younger than age 60 months, based on parents’ recall over the week before the survey.20 In beneficiary households, the cash transfer increases available resources. Indeed, we find that the total food share of the household budget declines almost three percentage points, indicating that households are moving along the food Engle curve, as would be predicted after a gain in income. Among beneficiary children in Pantawid villages, there was an 8.2 percentage point increase in parents feeding their children (ages 6 to 60 months) eggs and a 6.9 percentage point increase in dairy as well as some indication of greater frequency of meat and fish consumption (although not precisely estimated) during the previous week compared with children in nonprogram villages. We show in Appendix Table 7 that saturation dampens the gains to beneficiary children’s anthropometry in treated villages. 18 Stunting is measured as height-for-age < -2 standard deviations (SD), and severe stunting is measured as height- for-age < -3 SD, applying the World Health Organization (WHO) Child Growth Standard (WHO 2006). Analogously, wasting is measured as weight-for-age < -2 SD and severe wasting as weight-for-age < -3 SD, applying the same WHO Child Growth Standard. 19 de Hoop et al. (2019) document education impacts of Pantawid and show that similar to findings from other CCT programs, the schooling behavior of age-appropriate children improves 4 percentage points for enrollment and 2 to 3 percentage points in attendance, depending on the age group analyzed. These improvements are apparent despite an already high level of enrollment and attendance in comparison communities. 20 Protein-rich foods such as eggs, meat, fish, and dairy are particularly important for the linear growth of young children (Baten and Blum 2014; Moradi 2010). Appendix Table 6 shows that the young Filipino children in our control communities eat a diet besides rice, which is almost universally eaten, consisting mainly of eggs, fish, green vegetables, and bananas. (We did not ask about rice consumption of young children, but the Philippine national nutrition survey confirms that virtually all Filipino children eat rice every day (Denney et al. 2018).) The chief sources of protein, then, are eggs, fish, and some meat with little dairy or legumes consumed. These patterns are similar to infant and young child feeding practices reported by Denney et al. (2018) using the Philippine national nutrition survey, suggesting that the children in the Pantawid survey are comparable to the average Filipino child in as far as eggs and fish are the most important protein sources for them. 16 Impacts on Local Prices The second step in the proposed causal chain from a local cash influx to nutrition-related impacts on nonbeneficiaries, after a rise in local aggregate demand, is the presence of higher food prices. We thus explore the covariation between the rollout of Pantawid and food prices. We begin by examining changes in the household reported prices of a few individual food goods recorded in the 2011 experimental sample. This information was recorded only for three individual foods of standardized quality – eggs, rice, and sugar.21 Eggs are therefore the key non-tradable protein in this portion of the analysis, while rice and sugar are storable and hence more easily traded foods. Figure 2 shows the relationship between height-for-age z-scores of children ages 0-3 years with respect to the price of eggs and that of rice. We see that height-for-age decreases in the price of eggs but not in the price of rice, an indication that the price of protein is a factor in determining child height-for-age. Figure 3 shows how the price of eggs co-varies with program saturation in treated and control villages. First, at low levels of saturation, there is no difference in the price of eggs between treated and control villages. In control villages, the relationship between price and (potential) program saturation is flat or, if anything, slightly negative, perhaps reflecting that higher saturation control villages are poorer (as higher saturation means more households below the provincial poverty threshold.) However, in treatment villages, as saturation increases, and particularly past the median level of 65 percent, we observe a positive relationship between prices and saturation. The price of eggs, a key signal protein-rich food, thus covaries with program saturation. Panel A in Table 2 explores through a regression framework how price levels vary at the time of the survey between program and control villages. None of the price differences are significantly different from zero and the point estimates for rice and sugar are close to zero as well. Although not precisely estimated, the point estimate for the price of eggs stands almost 2 percentage points higher, indicating some divergence in relative price difference between the storable goods, such as rice and sugar, and the perishable good, eggs. Price differences emerge more clearly when the program indicator is interacted with saturation. Panel B in Table 2 shows that egg prices increase by 16 percent as we go from zero to 21 Eggs refer to chicken eggs, rice to commercially milled rice, and sugar to commercially sold brown or white sugar. 17 maximum observed saturation of 94.7 percent. In panel C of Table 3, we see that the interaction effect indicates a price increase of approximately six percentage points in above-median saturated treated villages in comparison to the controls. Although not precisely estimated, the interaction effect of being in a top quartile saturated treated village (panel D in Table 3) is associated with a price increase of 5.4 percent. The price changes in high saturation villages for the “tradable” goods, rice and sugar, are small in magnitude and not precisely estimated. As eggs are the most perishable good in this three-good comparison, the price changes are consistent with the predictions discussed above—we observe a price rise in saturation in program villages, but only for the non-tradable good. We continue the exploration of price changes and program exposure with official price data. First, Appendix Table 8 explores the covariation in food prices and program saturation using a monthly time series of item-specific food prices collected in each of the Philippines’ 81 provinces by the Philippine Statistics Authority from 2006 to 2014.22 We present prices for three largely non- tradable goods – fresh eggs, fish, and chicken – and three tradable goods – rice, snacks, and sugar. The three largely “non-tradable” goods all exhibit price increases correlated with changes in program saturation at the provincial level, although the estimated coefficients don’t reach the standard levels of precision. If we take the point estimates for the coefficients at face value, maximum observed program saturation at the province level data is 0.40, suggesting that provincial prices for eggs can rise as much as 7.7 percent (0.192*0.40) because of price effects from the Pantawid program. Maximum price increases are on the order of 5 to 6 percent for fresh fish and chicken. In contrast, the coefficients for the three tradable goods show much smaller price co- variation with program saturation. This pattern of results is also apparent in the first differenced specifications, with covariation between price changes and program saturation changes for the 22 In each province, price enumerators visit six markets, four rural and two urban; however, the locations of these markets are unknown to the researchers. Since these price data serve as an input to the Consumer Price Index of the Philippines, rigorous field controls are used to ensure the quality and comparability of goods assessed (Philippine Statistics Authority 2015). In the analysis, the average annual provincial price for each good, Pipy, is related to program expansion according to the following specification: = 0 + 1 + + + , where S is the province- year-specific saturation measure and i, p, and y index good, province, and year. The specification also includes province and year fixed effects, Fp and Fy respectively. Standard errors are clustered at the province level. The coefficient of interest, 1 , captures the good-specific price deviation from its provincial mean level, net of common year effects, as a function of the (mean-differenced) changes in provincial program exposure. We also explore an alternative specification that regresses the year-on-year difference in a food good price on the contemporaneous change in program exposure, also controlling for common year fixed effects. 18 three protein-rich foods but not for other foods (and here some of these difference estimates reach standard levels of precision). We now turn from provincial level price data, where inferential power may be limited by the small number of provinces, to price variation at the village level. Specifically, we document the relationship between village-level unit values from FIES and village-level saturation of Pantawid. Figure 4 confirms the visual pattern that we saw in the household survey data now with the national expenditure data: the price (or unit value) of eggs increases in program saturation, with a steeper slope past approximately 65 percent saturation. Table 3 summarizes regression results not just for fresh eggs but for the prices of 6 categories of food goods from FIES. The food goods include several largely non-tradable goods—fresh eggs, fresh fish, fresh meat, fresh vegetables, and milk products—as well as one tradable ones—raw regular milled rice (the full set of results for 15 food goods are reported in Annex 2). All prices are converted into natural logs. Consistent with our experimental sample, we relate the log unit value, Pivy, for each good i in village v and year y (for the years 2009, 2012, and 2015) to program saturation according to the following specification: = 0 + 1 + 2 ∗ ∗ + 3 + + + where S is the village-year-specific saturation measure for v and y. The model allows for the relationship between prices and saturation to vary above and below a cutoff value (we estimate this via an indicator variable equal to 1 if saturation exceeds c which is included in the model directly as well as interacted with saturation). Alternative models change the cutoff c at which we allow the relationship with saturation to vary—setting it at 50, 65, and 75 percent saturation respectively. In addition, we also estimate a simple linear model where log prices are just a function of saturation. All specifications also include village and year fixed effects, Fv and Fy respectively. In each interactive model the mean effect of interest is 1 +2 which is reported below the model coefficient estimates—this captures the total effect of saturation in high-saturation areas on each relative price. The results show that the prices of all the largely non-tradable goods increase with program saturation above the 65 percent threshold (recall that 65 percent is the median saturation in the experimental sample). Moving from a saturation level of 65 to 100 percent between 2009 and 2015 increases the relative price of eggs by almost 3 percent. The magnitudes of the price changes for 19 the other less tradeable goods (except fresh meat) are similar in magnitude (with milk products higher in magnitude but imprecisely measured). In contrast, the price of rice, which is more easily traded is negatively related to saturation above this cutoff. Increasing the cutoff from 65 to 75 percent sharpens these patterns (although the effects for the more tradeable goods are more imprecisely measured). Reducing the cutoff value to 50 percent or estimating a linear model dampens the contrast between more and less tradeable goods somewhat. The results therefore support the notion that saturation leads to pressure on the prices on less tradeable goods, especially when it is high. Unit values have known interpretive difficulties; most importantly they reflect both quality choices made by the household as well as the prices faced by it. This longstanding problem has been recently explored by McKelvey (2011) and Gibson and Kim (2019) who demonstrate that the traditional method of correcting price elasticity estimates for quality adjustments (Deaton, 1988) does not in fact result in elasticity estimates free of quality choices. Therefore, instead of pursuing a related correction, we first note that the main foods in our study are relatively homogenous and should not exhibit significant quality-income elasticities. Second, we note that Attanasio et al. (2013) showed that in Mexico, changes in unit values reasonably track changes in local prices. We nevertheless carry out two robustness checks on these results that aim to minimize potential quality variation. First, we calculate median village unit values after trimming values that are smaller or greater than the village mean unit value plus or minus 3 standard deviations. Second, we calculate alternative median unit values only taking into account households in the top village- level quartile of the distribution of per capita consumption expenditures. Even in highly saturated villages, these households should largely be nonbeneficiaries and therefore their food quality choices unaffected by the income transfer from program participation. The results, reported in Annex 2, are consistent with our main findings. In fact, the point estimates for fresh fish, fresh meat, and even fruit are larger than the point estimates based on unit values from the entire sample. Finally, we estimate the same regression specification using the price data from the experimental sample. Exploiting the cross-sectional variation in saturation in the treatment arm of the experimental sample, we find a coefficient of 0.117 on the interaction 20 between saturation and an indicator for saturation above 0.65 for the price of eggs. 23 The corresponding coefficient in the national data (reported in Table 3) is 0.08. The estimated effect from the unit values is smaller than that obtained from the price analysis; if the quality upgrading were significant, we would expect bigger effects from the unit values. In sum, results from all three independent sources of price data suggest a modest but persistent increase in the relative price of protein rich foods.24 Are these relatively small magnitudes of uncompensated price changes large enough to shift demand choices of the nonbeneficiary households? We return to the survey data to investigate this question. Impacts on Nonbeneficiaries So far, the analysis has demonstrated that the program improved outcomes for children from beneficiary households while concomitantly raising their consumption of protein-rich foods. The analysis has also identified a rise in the price of selected non-tradable goods, at least for highly saturated areas, but not for more easily traded goods. We next explore the consequences of this rise in selected food prices for nonbeneficiary households and, especially, children. Indirect Expenditure and Consumption Effects Table 4 reports results that parallel those in Table 1, but now contrasts nonbeneficiary households in treated and control villages, and presents the program impacts on household food budget share as well as whether the household reported feeding eggs, meat, and fish to children ages 6 to 60 months. For nonbeneficiary households in treated villages, food expenditure as a share 23 The full results from this regression are not reported here but are available from the authors upon request. 24 While our model shows that a transportation cost-induced wedge may be sustained indefinitely in areas where markets are not perfectly integrated, chicken ownership does not at first glance appear to be difficult to scale up. The question thus arises as to why supply, particularly of eggs, does not respond to a greater degree to the price increase. Morris, Beesabathuni, and Headey (2018) discuss several aspects of egg production in developing countries that may shed light on this puzzle. First, they note that 96 percent of all egg production in the world happens “intensively”, i.e. in industrial settings and not “extensively”, i.e. by home producers, perhaps explaining why we do not observe more chicken ownership in our data. They also document that eggs tend to be unavailable and expensive, particularly in South Asia and sub-Saharan Africa but also in less wealthy parts of East Asia. Only a few countries, largely in sub-Saharan Africa, have high levels of poultry ownership. Then, they use data from South Asia and sub-Saharan African to calibrate a model and estimate scenarios of price impacts of changes in egg production. They find that egg prices (albeit nationally) largely respond to changes in industrial production rather than home production. Finally, they document that the Philippines has one of the lowest levels of per-person egg availability globally and that it has developed a “very large -scale industrial egg production industr y” in recent years (p. 8). 21 of household budget significantly increased by 3.6 percent. When disaggregated into spending on dairy and eggs, cereals, and other foods, the results suggest that the decline in real income through the rising local prices of perishable foods may lead nonbeneficiaries to substitute away from dairy and eggs and toward cereals (Appendix Table 9). It is difficult to infer too much from the disaggregated spending data, as the coefficients are not precisely estimated. Nevertheless, the change in patterns between treatment and control villages is consistent with a rise in demand for protein-rich foods (as well as greater spending on other child goods) among beneficiary households, and perhaps a substitution away from protein-rich foods for nonbeneficiaries.25 With the provision of cash coupled with parenting education provided during the program’s Family Development Sessions, the program was expected to impact feeding practices. Indeed, we find that the food intake of young beneficiary children has shifted in all treatment villages. But the estimates in Table 4 suggest that food intake among nonbeneficiary children does not change nearly as much as a result of the program (the point estimates for the intake of eggs and vegetables are positive although not precisely estimated, suggesting little change). However, as before, we find that saturation matters. We find that children’s consumption of proteins, particularly eggs and dairy, declines significantly in treated villages as program saturation increases. We note that in these models the direct impact of being in a program village on egg and dairy consumption is positive for this sample of nonbeneficiaries (i.e. the coefficient estimate for being in a program village is positive), perhaps due to informational spillovers of the program.26 However, the interaction term with saturation is strongly negative. The lower incidence of egg and dairy 25 Since many households in the Philippines produce food for own consumption as well as purchase food on the market, a natural question concerns whether food producers benefit from the price increase in protein-rich foods. We examine whether the typical nonbeneficiary is a net producer or consumer, using eggs as the signal good because the units are more easily measured and compared than, say, fish and fresh meat. In control villages, the average nonbeneficiary household owns 4.55 chickens (in treated villages, nonbeneficiaries own 4.26 chickens; 60 percent of nonbeneficiary households in both treatment arms own chickens). Assume the average small-flock chicken lays 20 eggs per year, yielding a total annual production of 91 eggs from a flock of 4.55 chickens (Sonaiya and Swan 2004). On average, children ages 6 to 36 months in these households consume 1.78 eggs per week, or 93 per year. Nonbeneficiary households in our sample have an average of 1.4 children under 5. S ince young children’s consumption alone exceeds the average production of nonbeneficiary households, even the average home producer is a net consumer of eggs. In addition to increasing egg consumption, the cash transfer may have induced beneficiaries to diversify out of subsistence farming. Such a switch is consistent with Singh, Squire and Strauss’s (1986) model of agricultural households that produce and consume certain goods. This in turn would reduce the total supply of eggs at the village level. 26 Egg consumption also appears greater in the highest saturation villages in general. This may be due to various unobserved differences at the village level since high saturation villages are poorer on average and may differ in other key characteristics that determine demand patterns. 22 consumption for these children when compared with children in highly saturated control villages (or compared with children in low-saturated but treated villages) is immediately apparent. The same holds for the number of eggs consumed, although the difference on this intensive margin is not precisely estimated. We find similar but imprecisely estimated effects for meat, but no clear pattern for fish. The results explored so far suggest that, even 31 months after the start of the program, treated villages at high saturation level have higher relative prices for select protein-rich foods, and lower consumption of at least the signal good of fresh eggs. Might there also be evidence of effects on longer run outcomes such as child growth? Indirect Anthropometric Impacts Table 5, panel A reports the same nutrition measures as presented in Table 1A and Appendix Table 7 but now contrasts nonbeneficiary households in program and non-program villages.27 Children ages 6 to 36 months in nonbeneficiary households are substantially shorter if they reside in program villages—0.4 points shorter in the height-for-age z-score—than their counterparts in villages without the program. They are also significantly more likely to be stunted. The post-program stunting rate is estimated at 32 percent in control villages compared with 43 percent in treated villages. Although the point estimate for weight-for-age is also negative, it is not precisely estimated, suggesting particularly pronounced effects among longer-term nutritional measures such as child height. This difference in impact on height-for-age relative to weight-for- age is consistent with the new public health literature on the importance of proteins for linear growth in the first 1000 days of life (Puentes et al., 2016). If increases in the prices of protein-rich foods, and the concomitant decrease in nonbeneficiary children’s consumption of these foods, are associated with the worsening nutritional outcomes, then we would also expect to see the strongest nutritional effects in the villages where the price increases are the biggest: villages with the highest rates of program 27 For brevity, we do not show schooling-related outcomes in this table. However, impact estimates on nonbeneficiary households suggest little change in enrollment or attendance. These levels are already near universal and substantially higher than the enrollment or attendance of beneficiary children residing in the targeted households in the village. As nonbeneficiary children are not enrolled in the program, it is perhaps not surprising that schooling-related indicators do not change after program introduction. However, it does suggest that there are few schooling-specific spillovers in terms of higher fees or increased crowding that may have deterred the attendance of these children. 23 saturation. Table 5, panel B presents the impact of saturation on children’s schooling and nutrition, panel C presents the impacts of living in an above-median saturated program village, and panel D presents the impacts of living in a village with a saturation rate in the top quartile. We find that weight-for-age is significantly lower and the likelihood of being underweight significantly higher in program villages that have high rates of saturation. Average height-for-age is also lower and stunting rates higher in highly saturated villages, but the coefficients are precisely estimated for only the villages in the fourth quartile of saturation. We therefore do indeed find that growth deficits in height-for-age and weight-for-age among nonbeneficiary children are greater when a greater proportion of the village participates in the program. As child growth is particularly sensitive to nutritional and health conditions in the first 1,000 days of life (Hoddinott et al. 2013), we can investigate the age patterns of child height differences among children who lived much of their first 1,000 days under the program compared with older children born and partially reared before program onset. If the nutritional impacts on nonbeneficiary children are attributed to program presence and not other unobserved factors, then we would not expect to see the same impact among older children. Figure 5 depicts the proportion of nonbeneficiary children stunted in treated versus control villages for three age ranges. The stunting prevalence for children ages 36 to 60 months, and hence only partially exposed to the program at critical ages for growth, is virtually identical. In contrast, among younger children, the stunting rate is substantially higher in treated villages, for children ages 6 to 24 months and those ages 24 to 36 months. The age differences suggested by Figure 5 are apparent in the regressions reported in Appendix Table 10, which investigate the nutrition impacts differentiated across child age categories (ages 6 to 23 months, 24 to 36 months, and 37 to 60 months). Height-for-age z-scores are significantly lower for children ages 6 to 23 months, that is, those children who have been exposed to the program in utero and for the entirety of their lives, with a deficit on the order of 0.70 standard deviation. Stunting rates are also higher (15 percentage points), but the impact is not as precisely estimated. Program impacts on weight-related nutrition measures, which capture shorter-run measures of health status, also emerge for this age group. The youngest nonbeneficiary children are significantly more likely to be underweight, on the order of 20 percentage points. For slightly older 24 nonbeneficiary children ages 24 to 36 months, the point estimates of impact also suggest a worsening of nutritional status but to a lesser degree—there is no difference in wasting, for example—and the difference is not as precisely estimated. If children in nonbeneficiary households suffer growth deficits because of the cash transfer, then we would expect to see a divergence in growth only for those children exposed to the program at critical ages when they are most vulnerable to a nutritional deficit. We do indeed see this for children younger than 36 months, and especially for those ages 6 to 24 months at the time of the survey. Relative Sizes of Direct and Indirect Effects of the Transfer on Child Growth The estimated indirect effect on nonbeneficiary children’s height-for-age (-0.397) is larger than the direct effect on beneficiary children’s height-for-age (0.257). This result may seem counterintuitive. To understand how a difference of this magnitude may arise, we first note that in highly saturated villages, there are many more beneficiaries than nonbeneficiaries, creating a relatively large price shock and concentrating its effects only on the smaller segment of the population without an income transfer. Regarding the observed changes in height-for-age per se, while perhaps surprising at first glance, these impacts on child growth translate to estimated elasticities in line with the broader literature. Unfortunately, we do not have the data to estimate price or income elasticities for individual food items at the household level, as we only observe food quantities consumed by young children. However, using observed household consumption expenditure for aggregated food groups, we estimate a price elasticity of demand for dairy and eggs to be -1.3 and an income elasticity of egg demand, estimated for beneficiaries in highly saturated areas, of 0.15. These point estimates are similar to those reported in the literature: Friedman and Levinsohn (2002) estimate a price elasticity of demand for eggs in a nationally representative survey from Indonesia to be -1. Estimates of income elasticity of egg demand in the United States range from 0.04 to 0.11, which are somewhat lower than our estimate of 0.15, although not surprising given the difference in income between the two populations (Okrent and Alston 2012). Finally, using a sample of poor rural households in Kenya, Almås, Haushofer and Shapiro (2019) estimate income elasticity for all protein to be 1.29, which is considerably larger than our estimate for eggs alone and might suggest even greater impacts on child growth than we observe in our setting. 25 Although eggs are a signal good with a price that we are able to observe, the results presented in Table 3 show that the price increases, and thus likely the shift in consumption patterns, apply more broadly for other key foods. Nevertheless, if we focus only on eggs, a shortfall in consumption of eggs of the observed magnitude (roughly half an egg a week) extended over the exposure of the program (31 months) would result in a height detriment of approximately 40 percent of the total negative spillover we observe. Informing this calculation are the estimated effects of egg consumption on linear growth from Puentes et al. (2016) and Iannotti et al. (2017). Puentes et al. (2016) estimate an additional 0.72 centimeter in linear growth from consumption of an additional egg per week as a child ages from 6 to 24 months. The converse is also true: the observed increase in egg consumption by beneficiaries over the duration of exposure to the program yields approximately 40 percent of the total height gain for this group. The observed changes in height for the beneficiaries and nonbeneficiaries are therefore consistent with the observed changes in our signal protein-rich food good, eggs.28 Iannotti et al. (2017) only report the impact on height-for-age z-scores but estimate that an egg a day for six months increased the height-for-age z-score of children ages 6 to 9 months by 0.63 standard deviation relative to a baseline of -1.90 standard deviation. Given the nonlinearity in the height-for-age z-score, these two estimates are not directly comparable; however, they are of the same order of magnitude, suggesting that the observed decline in egg consumption alone explains much of the impact on height-for-age z-scores. Furthermore, the consumption of other protein-rich food goods, such as fresh fish or meat, also plays a role in the production of child height, and price changes for these goods are also correlated with program expansion. The direct and indirect effects of the program 28 As reported in Denney et al. (2018), data from the National Nutrition Survey of 2013 show that median daily protein intake is 18.6 grams for 6-12 month olds, 23.3 grams for 12-24 month olds, 27.5 grams for 24-36 month olds, and 31.5 grams for 36-60 month olds. An egg contains 7 grams of protein, thus representing between 22 percent (for 36-60 month olds) to 37.6 percent (for 6-12 month olds) of median daily protein intake. If we were to assume a sustained decrease in egg consumption over the entire period (and a linear relation between growth and protein intake), our estimated reduction in consumption of half an egg per week by nonbeneficiary children would imply a deficit of 3.5 grams of protein a week over 30 months (from ages 6 to 36 months). This totals a reduction in protein intake of 420 grams over the 30 month period. The difference in child height in highly saturated treated (average height of 79.53 centimeters) and control areas (average height of 79.95 centimeters) is 0.42 centimeters. By comparison, the Puentes et al. study estimates that an egg a week over an 18-month period, representing a total of 504 additional grams of protein, contributes 0.72 centimeters in linear growth. While this exercise, like the elasticity estimates above, relies on strong assumptions this calibration exercise brackets what one might expect to see as a result of the general equilibrium effects documented here. 26 on stunting and height-for-age thus appear reasonable given the observed changes in consumption and spending patterns. Other Possible Mechanisms Until this point, the evidence presented is consistent with the hypothesis that this cash transfer program increased the prices of certain food goods that are important for the production of child height, with increases especially found in the highly saturated villages (as would be predicted in the presence of a general equilibrium spillover). These price changes in turn likely contributed to an increase in stunting rates among nonbeneficiary children, again especially the children residing in highly saturated villages. Furthermore, the estimated consumption and growth elasticities are in line with the wider literature, suggesting that the observed effects can be expected given the uncompensated price increases for key food goods. However, there are alternative transmission mechanisms, complementary to the price channel, that in principle might also result in increased rates of stunting for nonbeneficiary children. In this section, we investigate whether the evidence supports any of these additional channels. Changes in Health Care Access and Quality Child height is determined in early life not only by nutrition but also exposure to infections and other health shocks. Thus, another potential channel is through program impacts on access to early life health services and the quality of those services. This is especially important to investigate, since maternal and child care is directly incentivized by Pantawid. To the extent that the formal health care sector can improve child health, a degradation in access to or quality of health services could also in principle contribute to increased stunting. The Pantawid program may reduce access or quality of healthcare by crowding out available services due to an increase in service utilization by beneficiary households. This crowding-out mechanism can result in increased prices for care, reductions in the quality of available services, or both. In neighboring Indonesia, Triyana (2016) finds that a CCT conditioned on safe delivery practices resulted in a 10 percent increase in fees charged by midwives.29 29 There was also a 10 percent increase in the supply of local midwives in that setting, but the increase was not sufficient to prevent a price rise. 27 We thus turn to program impacts on a range of health care seeking behavior relevant to young children for beneficiary and nonbeneficiary households. Panel A of Appendix Table 11 indicates that, among beneficiaries, Pantawid increased the timely initiation (i.e. within the first trimester) of antenatal care, postnatal care within 24 hours, institutional or in-facility delivery, children with the mumps, measles, and rubella (MMR) vaccine30, growth monitoring, and general curative care seeking.31 We also examine the content of antenatal care using maternal recall and find that Pantawid significantly improved the completeness of antenatal care.32 The Pantawid program thus appears to have been successful at increasing utilization and quality of care. However, again, the interactions with saturation, presented in panels B-D of Appendix Table 11, suggest that increases in saturation dampened gains in growth monitoring and postnatal care within 24 hours. Table 6 examines care seeking by nonbeneficiary households and shows no significant impacts (panel A). A particular threat to our identification would arise from changes in health service utilization that covary in saturation; however, we find no such effects (panels B, C, and D). Indeed, for above median saturated villages, the interaction effect, although never significant, is generally positive (with the exception of growth monitoring and curative cake seeking for young children), if anything, suggesting positive spillovers. These findings thus do not suggest that a decrease in health care constituted a complementary channel to the price increase of protein-rich foods. Sub-group Imbalance in Baseline Characteristics Although the baseline characteristics were balanced between treatment and control communities, child height was not one of the measured indicators. Furthermore, as impacts are greatest in highly saturated villages, characteristic balance in this subgroup also needs to be investigated. While all the characteristics in low-saturated villages appear balanced, a handful of 30 We use the MMR vaccine as a marker for complete immunization as it is the last vaccine in the WHO/Unicef extended immunization schedule for the first year of life (WHO 2003). 31 The first three of these measures are based on children ages 6 to 36 months, since they refer to care around the birth, while the last two measures of more general care seeking are based on children ages 6 to 60 months. 32 Our measure of the content of ANC is the proportion of the ten following actions: measurement of weight, height, and blood pressure, urine and blood samples, administering the tetanus toxoid injection, iron tablets, and deworming pills, and discussing complications, including guidelines on where to go in the case of complications. The Pantawid survey elicited these data through maternal recall of ANC received during the last pregnancy (in the preceding two years). 28 characteristics in highly saturated villages are significantly different (Appendix Tables 3 and 5). Of particular note, nonbeneficiary households in above-median saturated treatment areas were significantly smaller (by 0.41 people—mostly adults— with a standardized mean difference of 0.269) than control nonbeneficiary households. This difference may in turn affect the number of caretakers available for young children and, consequently, factors such as household responsiveness to child illness or the available total household resources partially devoted to children. We address the possible influence of subgroup imbalance in three ways. First, we note that child height and stunting rates were appreciably the same for children ages 36 to 60 months at the time of the follow-up survey (Figure 5 and Appendix Table 10). As these children largely lived their first 1,000 days—a critical growth period in determining child height trajectories—before program introduction, the similar height measures in older nonbeneficiary children between treatment and control villages strongly suggest that this outcome would be balanced at baseline if measured (when these children were 31 months younger). Second, we use baseline data on household composition for a difference-in-differences approach to investigate potential differential changes in household composition across program and non-program villages among nonbeneficiary households with young children. We find no significant differential change in household size or composition in treated villages 31 months after rollout (Annex 3 Table 1). We find this lack of differential change in the full sample and, specifically, with respect to highly saturated villages. Therefore, any baseline difference in household size or composition persists over the study period. If household compositional differences partially determined the child height detriment in comparison villages, then we would also expect to see a persistent height difference among the older child sample, as these children were also raised in somewhat smaller households with somewhat fewer adults. But we do not observe this pattern. Third, we address possible baseline characteristic imbalance more generally by controlling for all baseline characteristics through propensity weighting (Hirano, Imbens, and Ridder 2003). The estimated impacts on child height and weight are appreciably the same as the results presented in Table 6 (Annex 3 Table 2). Adult Labor Supply Responses to Program Presence Adult labor market behavior may respond to the Pantawid program with knock-on effects for child development. For example, if real incomes decline for nonbeneficiary households 29 because of price changes, as predicted by the theoretical framework, adult household members may respond by increasing their labor supply to compensate for the fall in real income, thus reducing the availability of adult caregivers. We examine the labor force participation, work-for- pay, and full-time work (greater than 40 hours per week) for adult men and women in nonbeneficiary households (Appendix Table 12). Overall, there is little change in labor force participation or hours worked for men or women. There is also little evidence of change in male or female labor force participation in above-median saturated nonbeneficiary households. Therefore, it appears unlikely that the detrimental impacts among nonbeneficiary children are in part due to shifts in adult labor supply. Older child educational and labor supply responses to program presence In addition to adults, the activities of older children in the household, who may play a role in child care, might be affected by the presence of the Pantawid program. To determine if this is the case, we look at the education and child labor impacts of Pantawid on older nonbeneficiary children (Appendix Table 13). None of the estimated effects of Pantawid is significant, suggesting that the availability of older child caregivers was not affected by the program and thus is unlikely to be a channel that impacted nonbeneficiary child growth. VI. CONCLUSIONS Aid programs such as CCTs often introduce a large amount of resources into small village economies. Such infusions can bring substantial benefits but also risk raising local prices, thus generating negative spillovers on household welfare, particularly for nonbeneficiaries. This paper tests for such local general equilibrium effects on food prices and a range of outcomes for beneficiaries and nonbeneficiaries, using the randomized evaluation of a large CCT in the Philippines, Pantawid Pamilya Pilipino Program. Any price response to a demand increase will depend on the market structure of producers and/or suppliers. If the relevant market is principally local and not fully integrated into the wider economy, then the presence of oligopolistic producers or a rising marginal cost curve of local production if the local market is competitive will translate the demand increase from the cash transfer into higher prices. We study a combination of locally traded perishable food goods along 30 with staples and packaged goods traded in national markets. It is these locally traded foods that appear to exhibit price increases after the introduction of the program, especially when program exposure at the village level, which we term saturation, is high. In principal, any general equilibrium effect of a cash transfer should be greater as the proportion of beneficiaries increases. This is indeed what we find. Concomitant with these price changes, the CCT increased child stunting among nonbeneficiaries (while reducing it for beneficiaries). These impacts are observed only for children who were in the vulnerable first 1,000-day period of life when the program was introduced and not among older children. Moreover, where program saturation was higher, the detrimental impacts on nonbeneficiaries were larger. These are not short-run effects; the transfer program had been in place for 31 months at the time of the follow-up survey. Taken together, the findings suggest that anti-poverty cash transfer programs may have unanticipated effects operating through local markets. Although participating households are (likely more than) compensated for the price increases, this is not the case for nonbeneficiary households. In addition to the food price channel, a complementary channel that may contribute to worsening child growth is that of spillovers in the formal health care system. Access to key maternal and child health services, and indeed quality of care, significantly increased among beneficiary households, but the use of health services among nonbeneficiary mothers and children does not vary in treatment or saturation. We also investigate and rule out possible additional explanations for the nonbeneficiary child growth deficits, including adult and older child behavioral responses to the program that could result in the reduced availability of caregivers, as well as the possible influence of subgroup imbalance in baseline characteristics. We investigate and rule out possible additional explanations for the nonbeneficiary child growth deficits, including adult and older child behavioral responses to the program that could result in the reduced availability of caregivers, as well as the possible influence of subgroup imbalance in baseline characteristics. As suggested by the conceptual framework and confirmed by the analysis, we only detect local general equilibrium effects in the highly saturated villages. As saturation and remoteness are key mediators of this effect, we do not necessarily expect impacts in areas with many proximate 31 markets or where saturation is low.33 Fortunately, the conditions under which local price spillovers arise are not widespread. By 2015 only 4.2 percent of villages in a nationally representative expenditure survey had a saturation level of 65 percent or higher. We are able to identify such impacts as roughly half the experimental village sample attained such a level of saturation. In addition, the Philippines has one of the highest incidences of stunting (Bhutta et al. 2013); a similar program might not have such severe consequences in countries where poverty is perhaps less linked with poor nutritional status. Indeed, another reason such negative spillovers have not been detected in other studies may be that cash transfer experiments began in middle-income countries, like Mexico and Brazil, where stunting was less of an issue. As transfers are increasingly being introduced in poorer countries with higher levels of stunting, the possibility of adverse spillovers to young children in nonbeneficiary households through the local price mechanism merits more consideration. Might a different targeting scheme avoid such general equilibrium effects? The Pantawid program is targeted to individual households based on a proxy means test score. An alternative is a village-based targeting scheme for the subset of villages that are particularly poor or remote. In such a scheme all households would be offered the program which would compensate everyone for any rise in local prices, thereby averting negative nutrition impacts. Of course, area-based targeting or universal access, while averting spillovers, would also likely be more expensive. Using a limited number of assumptions, we consider the costs and benefits of such village-based or geographic targeting in poor areas (as detailed in Annex 5). Note that this is a conservative estimate of program benefit as we only consider the labor market impacts of stunting aversion on subsequent adult wages and not other possible detriments such as lowered intergenerational transmission of human capital. Given such parameter estimates as the average daily adult wages reported in our sample (US$6.3), the stunting differential attributable to Pantawid, and the impact of stunting on adult wages (Hoddinott et al. 2013), we estimate that the 33 In extensions of our empirical analysis, we consider remoteness indicators as an additional measure for the integration, or lack thereof, of the local market with wider regional ones. Unfortunately, the Pantawid evaluation was not designed to estimate differential impact by remoteness categories. Further, it is not clear what the relevant remoteness metric might be. Annex 4 explores the influence of two remoteness proxies on the main impacts and finds theoretically consistent but imprecise estimates. Given the unambiguous link between price changes, stunting, and program saturation, we believe that in this context the saturation metric is the most comprehensive available measure summarizing the channels through which a cash transfer can lead to general equilibrium impacts. 32 discounted lifetime benefits of the universal program’s impact on stunting, manifested in lifetime earnings, would equate the discounted program costs of expansion to universal coverage at a discount rate of 5 percent. At lower rates of discount, the benefit of local universal targeting exceeds the cost. While further work needs to be done to estimate more comprehensively the lifetime benefits of improvements in height, as well as the programmatic costs of different targeting mechanisms, this back-of-the-envelope estimate suggests that policy makers consider a hybrid targeting scheme for anti-poverty programs when faced with the possibility of local market price spillovers. For poorer or remote villages, offering the program to every household may be more cost-effective, particularly given the lower administrative costs of geographic targeting through reduced data collection and verification. Other areas of the country that likely will not experience local price spillovers could continue with offering the program only to the poorest households. 33 REFERENCES Almås, Ingvild, Johannes Haushofer and Jeremy P. Shapiro. 2019. “The Income Elasticity for Nutrition: Evidence from Unconditional Cash Transfers in Kenya.” No. w25711. National Bureau of Economic Research. Aggarwal, Shilpa. 2018. “Do Rural Roads Create Pathways Out of Poverty? Evidence from India.” Journal of Development Economics 133: 375-395. Angelucci, Manuela and Giacomo De Giorgi. 2009. “Indirect Effects of an Aid Program: How Do Cash Transfers Affect Ineligibles' Consumption?” The American Economic Review 99(1): 486-508. Atkin, David and Dave Donaldson. 2015. “Who's Getting Globalized? The Size and Implications of Intra-national Trade Costs.” No. w21439. National Bureau of Economic Research. Attanasio, Orazio, Vincenzo Di Maro, Valérie Lechene, and David Phillips. 2013. “Welfare Consequences of Food Prices Increases: Evidence from Rural Mexico.” Journal of Development Economics 104: 136-151. Attanasio, Orazio, and Elena Pastorino. 2019. “Nonlinear Pricing in Village Economies.” Econometrica (forthcoming). Baird, Sarah, Francisco Ferreira, Berk Özler and Woolcock, Michael. 2013. “Conditional, Unconditional and Everything in Between: A Systematic Review of the Effects of Cash Transfer Programmes on Schooling Outcomes.” Journal of Development Effectiveness 6(1): 1-43. Baten, Joerg and Matthias Blum. 2014. “Why Are You Tall While Others Are Short? Agricultural Production and Other Proximate Determinants of Global Heights.” European Review of Economic History, 18(2): 144-165. Beegle, Kathleen, Emanuela Galasso and Jessica Goldberg. 2017. “Direct and Indirect Effects of Malawi’s Public Works Program on Food Security.” Journal of Development Economics 128:1-23. Bhutta, Zulfiqar A., Jai K. Das, Arjumand Rizvi, Michelle F. Gaffey, Neff Walker, Susan Horton, Patrick Webb et al. 2013. “Evidence-based Interventions for Improvement of Maternal and Child Nutrition: What Can Be Done and At What Cost?” The Lancet 382 (9890): 452-477. Bobba, Matteo and Jeremie Gignoux. 2019. “Neighborhood Effects in Integrated Social Policies.” World Bank Economic Review 33(1): 116-139. Bobonis, Gustavo and Frederico Finan. 2009. “Neighborhood Peer Effects in Secondary School Enrollment Decisions.” Review of Economics and Statistics 91(4): 695–716. 34 Bourguignon, Francois, Francisco Ferreira and Phillippe Leite. 2003. “Conditional Cash Transfers, Schooling, and Child Labor: Micro-simulating Brazil's Bolsa Escola Program.” The World Bank Economic Review, 17(2): 229-254. Chaudhury, Nazmul, Jed Friedman and Junko Onishi. 2013. “Philippines Conditional Cash Transfer Program Impact Evaluation 2012.” World Bank. Contreras, Diana and Pushkar Maitra. 2013. “Health Spillover Effects of a Conditional Cash Transfer Program.” Working Paper No. 44-13. Monash University, Department of Economics. Crost, Benjamin, Joseph Felter and Patrick Johnston. 2014. “Aid Under Fire: Development Projects and Civil Conflict.” The American Economic Review 104(6): 1833-1856. Cunha, Jesse, Giacomo De Giorgi and Seema Jayachandran. 2015. “The Price Effects of Cash Versus In-kind Transfers.” Staff Reports 735, Federal Reserve Bank of New York. Cunha, Jesse, Giacomo De Giorgi and Seema Jayachandran. 2019. “The Price Effects of Cash Versus In-kind Transfers.” Review of Economic Studies 86(1): 240-281. Currie, Janet, and Tom Vogl. 2013. “Early-life Health and Adult Circumstance in Developing Countries.” Annual Review of Economics 5(1): 1-36. de Hoop, Jacobus, Jed Friedman, Eeshani Kandpal and Furio Rosati. 2019. “Child Schooling and Child Work in the Presence of a Partial Education Subsidy.” Journal of Human Resources 54(2): 503-551. Deaton, Angus. 1988. “Quality, Quantity, and Spatial Variation of Price.” The American Economic Review: 418-430. Del Boca, Daniela, Chiara Pronzato, and Giuseppe Sorrenti. 2018. “The Impact of a Conditional Cash Transfer Program on Households' Well-Being.” Working Papers 2018-093, Human Capital and Economic Opportunity Working Group. Denney, Liya, Imelda Angeles-Agdeppa, Mario Capanzana, Marvin Toledo, Juliana Donohue and Alicia Carriquiry. 2018. “Nutrient Intakes and Food Sources of Filipino Infants, Toddlers and Young Children are Inadequate: Findings from the National Nutrition Survey 2013.” Nutrients 10 (11): 1730. DSWD. 2015. Pantawid Pamilyang Pilipino Program. Government of the Philippines. Elbers, Chris, Tomoki Fujii, Peter Lanjouw, Berk Ozler, and Wesley Yin. 2007. "Poverty Alleviation Through Geographic Targeting: How Much Does Disaggregation Help?." Journal of Development Economics 83 (1): 198-213. Ferreira, Francisco H., Deon Filmer and Norbert Schady. 2017. “Own and Sibling Effects of Conditional Cash Transfer Programs: Theory and Evidence from Cambodia” in. S. Bandyopadhyay (ed.) Research on Economic Inequality 25: 259-298. 35 Fiszbein, Ariel and Norbert Schady. 2009. Conditional Cash Transfers: Reducing Present and Future Poverty. World Bank, Washington DC. Friedman, Jed and James Levinsohn. 2002. “The Distributional Impacts of Indonesia's Financial Crisis on Household Welfare: A “Rapid Response” Methodology.” The World Bank Economic Review, 16(3): 397-423. Galasso, Emanuela and Adam Wagstaff. 2019. “The Aggregate Income Losses from Childhood Stunting and the Returns to a Nutrition Intervention Aimed at Reducing Stunting.” Economics & Human Biology (forthcoming). Gibson, John, and Bonggeun Kim. 2019. “Quality, Quantity, and Spatial Variation of Price: Back to the Bog.” Journal of Development Economics 137: 66-77. Hanlon, Joseph, Armando Barrientos and David Hulme. 2010. Just Give Money to the Poor: The Development Revolution from the Global South. Kumarian Press. Hanna, Rema and Benjamin Olken. 2018. “Universal Basic Incomes Versus Targeted Transfers: Anti-poverty Programs in Developing Countries.” Journal of Economic Perspectives, 32(4): 201-26. Haushofer, Johannes and Jeremy Shapiro. 2016. “The Short-term Impact of Unconditional Cash Transfers to the Poor: Experimental Evidence from Kenya.” The Quarterly Journal of Economics 131(4): 1973-2042. Haushofer, Johannes et al. 2019. "General Equilibrium Effects of Cash Transfers in Kenya." AEA RCT Registry. February 18 2019. Headey, Derek, Kalle Hirvonen and John Hoddinott. 2018. "Animal Sourced Foods and Child Stunting." American Journal of Agricultural Economics 100 (5): 1302-1319. Hirano, Keisuke, Guido Imbens and Geert Ridder. 2003. “Efficient Estimation of Average Treatment Effects Using the Estimated Propensity Score”. Econometrica, 71(4): 1161- 1189. Hoddinott, John, Jere R. Behrman, John A. Maluccio, Paul Melgar, Agnes R. Quisumbing, Manuel Ramirez-Zea, Aryeh D. Stein, Kathryn M. Yount, and Reynaldo Martorell. 2013. "Adult Consequences of Growth Failure in Early Childhood." The American Journal of Clinical Nutrition 98(5): 1170-1178. Hoddinott, John, Harold Alderman, Jere R. Behrman, Lawrence Haddad and Susan Horton. 2013. “The Economic Rationale for Investing in Stunting Reduction.” Maternal and Child Nutrition 9(Suppl. 2): 69–82. Iannotti, Lora L., Chessa K. Lutter, Christine P. Stewart, Carlos Andres Gallegos Riofrío, Carla Malo, Gregory Reinhart, Ana Palacios et al. 2017. “Eggs in Early Complementary Feeding and Child Growth: A Randomized Controlled Trial.” Pediatrics 140(1). 36 ILO. 2014. “Wages in Asia and the Pacific: Dynamic But Uneven Progress” in Global Wage Report 2014/15 | Asia and the Pacific Supplement. Available from: http://www.ilo.org/wcmsp5/groups/public/---asia/---ro-bangkok/---sro- bangkok/documents/publication/wcms_325219.pdf. Bangkok: International Labour Organization Regional Office for Asia and the Pacific Regional Economic and Social Analysis Unit. Accessed 23 May, 2019. Imbens, Guido and Jeffrey Wooldridge. 2009. “Recent Developments in the Econometrics of Program Evaluation.” Journal of Economic Literature, 47, 5–86. Kandpal, Eeshani, Harold Alderman, Jed Friedman, Deon Filmer, Junko Onishi and Jorge Avalos. 2016. “A Conditional Cash Transfer Program in the Philippines Reduces Severe Stunting”. The Journal of Nutrition 146 (9), 1793-1800. Klasen, Stephan and Simon Lange. 2016. “How Narrowly Should Anti-poverty Programs be Targeted? Simulation Evidence from Bolivia and Indonesia.” No. 213. Courant Research Centre: Poverty, Equity and Growth-Discussion Papers. Labonne, Julien. 2013. “The Local Electoral Impacts of Conditional Cash Transfers: Evidence from a Field Experiment.” Journal of Development Economics 104: 73-88. Lalive, Rafael and Mattias Cattaneo. 2009. “Social Interactions and Schooling Decisions.” The Review of Economics and Statistics 91(3): 457–477. Macours, Karen, Norbert Schady and Renos Vakis. 2012. “Cash Transfers, Behavioral Changes, and Cognitive Development in Early Childhood: Evidence from a Randomized Experiment.” American Economic Journal: Applied Economics 4(2): 247-273. McKelvey, Christopher. 2011. “Price, Unit Value, and Quality Demanded.” Journal of Development Economics, 95(2):157-169. Moradi, Alexander. 2010. “Nutritional Status and Economic Development in Sub- Saharan Africa, 1950–1980.” Economics & Human Biology, 8(1): 16-29. Morris, Saul, Kalpana Beesabathuni and Derek Headey. 2018. An Egg for Everyone: Pathways to Universal Access to One of Nature's Most Nutritious Foods. Maternal & child nutrition, 14, e12679. Mu, Ren and Dominique van de Walle. 2011. “Rural Roads and Local Market Development in Vietnam,” Journal of Development Studies, 47(5): 709-734. Muralidharan, Karthik, Paul Niehaus and Sandip Sukhtankar. 2017. “General Equilibrium Effects of (Improving) Public Employment Programs: Experimental Evidence from India.” Working Paper. Nazara, Suahasil and Sri Kusumastuti Rahayu. 2013. “Program Kelyarga Harapan (PKH): Indonesian Conditional Cash Transfer Programme.” International Policy Center for Inclusive Growth: Policy Brief No. 42. 37 Okrent, Abigail and Julian Alston. 2012. The Demand for Disaggregated Food-Away- From-Home and Food-at-Home Products in the United States, ERR-139, U.S. Department of Agriculture, Economic Research Service. Philippine Statistics Authority. 2015. Price Indices. The Government of The Philippines. Web. 10 November 2015. Puentes, Esteban, Fan Wang, Jere R. Behrman, Flavio Cunha, John Hoddinott, John A. Maluccio, Linda S. Adair, Reynaldo Martorell and Aryeh D. Stein. 2016. “Early Life Height and Weight Production Functions with Endogenous Energy and Protein Inputs.” Economics & Human Biology 22: 65-81. Ravallion, Martin. 2009. “How Relevant is Targeting to the Success of an Antipoverty Program?” The World Bank Research Observer, 24(2), pp.205-231. Saavedra, Juan Esteban and Sandra Garcia. 2012. “Impacts of Conditional Cash Transfer Programs on Educational Outcomes in Developing Countries: A Meta-analysis.” RAND Labor and Population Working Paper Series, WR-921-1. Singh, Inderjit, Lyn Squire, and John Strauss. 1986. “A Survey of Agricultural Household Models: Recent Findings and Policy Implications.” The World Bank Economic Review 1.1: 149-179. Sonaiya, Emmanuel and S. E. J. Swan. 2004. “Small-Scale Poultry Production: Technical Guide.” Food and Agriculture Organization of the United Nations, Rome. Triyana, Margaret. 2016. “Do Health Care Providers Respond to Demand-Side Incentives? Evidence from Indonesia.” American Economic Journal: Economic Policy 8(4): 255- 88. UNICEF 2019. The State of the World’s Children 2019. Children, Food and Nutrition: Growing Well in a Changing World. UNICEF, New York. Vogl, Tom S. 2014. “Height, Skills, and Labor Market Outcomes in Mexico.” Journal of Development Economics 107: 84‐96 World Bank. 2014. A Model from Mexico for the World. http://www.worldbank.org/en/news/feature/2014/11/19/un-modelo-de-mexico-para-el-mundo. November 15, 2015. WHO. 2003. Selected Vaccine Introduction Status into Routine Immunization. Geneva: World Health Organization. Available from: www.who.int/immunization_monitoring/routine/schedule_analysis_2003.pdf. WHO Multicentre Growth Reference Study Group. 2006. WHO Child Growth Standards: Length/Height-for-age, Weight-for-age, Weight-for-length, Weight-for-height and Body Mass Index-for-age: Methods and Development. Geneva: World Health Organization. Available from: https://www.who.int/childgrowth/standards/technical_report/en/. 38 FIGURES AND TABLES Source: 2008 Household Assessment Form Figure 1: The across-village variation in program saturation (proportion of households with age appropriate children eligible for the Pantawid program) 39 Source: 2011 Pantawid impact evaluation survey Figure 2:Height-for-age Z scores (children aged 0-3 years) and the prices of eggs and rice 40 Source: 2011 Pantawid impact evaluation survey Figure 3:The price of eggs and program saturation in treated and control villages 41 Source: 2015 Family and Income Expenditure Survey Figure 4:The relative price of eggs and program saturation in treated and control villages 42 Source: 2011 Pantawid impact evaluation survey Figure 5: Stunting prevalence among nonbeneficiary children after 31 months of program exposure by age of child 43 Table 1A: Program impact on anthropometry of beneficiary children aged 6-36 months (1) (2) (3) (4) (5) Height- Weight- Severe for-Age Z Stunting for-Age Z Underweight stunting score score Panel A: Program impact Program village 0.256* -0.035 -0.093** 0.125 -0.023 (0.15) (0.05) (0.04) (0.15) (0.05) Control observations 162 162 162 186 186 Treated observations 181 181 181 201 201 Control mean -1.800 0.481 0.216 -1.207 0.280 Source: Kandpal et al. (2018) note: *** p<0.01, ** p<0.05, * p<0.1 All specifications include municipality fixed effects. Prices are in natural logs. Child-level specifications include quadratic terms for child age (in months). Standard errors are clustered by village; there are 130 villages. Source: 2011 Pantawid impact evaluation survey Table 1B: Program impact on the budget share of food and children’s food intake for beneficiary households with children aged 6-60 months (1) (2) (3) (4) (5) (6) Whether Whether Whether Number of Whether Budget Eggs Were Meat Was Dairy Was Eggs Fed Fish Was Share of Fed to Child Fed to Fed to to Child in Fed to Child Food in Past Child in Child in Past Week in Past Week Week Past Week Past Week Panel B: Program impact Program village -0.029** 0.082** 0.21 0.027 0.029 0.069* (0.02) (0.03) (0.17) (0.04) (0.03) (0.04) Control observations 328 405 402 406 406 399 Treated observations 335 437 434 437 436 433 Control mean 0.691 0.704 1.808 0.5 0.852 0.248 44 Table 2: Program impact on reported village-level food prices Egg Price Rice Price Sugar Price Reported by Reported by Reported by Household Household Household Panel A: Program Impact 0.017 -0.000 0.000 Program village (0.014) (0.006) (0.011) Panel B: Interaction with saturation Program village -0.083* -0.009 -0.020 (0.046) (0.020) (0.037) -0.007 0.017 0.004 Saturation (0.062) (0.027) (0.049) 0.164** 0.014 0.034 Program village* saturation (0.072) (0.031) (0.057) Panel C: Interaction with above median saturation Program village -0.011 0.003 -0.004 (0.019) (0.008) (0.015) 0.007 0.013 0.005 Above median saturation (0.022) (0.010) (0.017) Program village *Above median saturation 0.060** -0.007 0.009 (0.027) (0.012) (0.022) Panel D: Interaction with fourth quartile saturation Program village -0.011 -0.001 0.005 (0.027) (0.006) (-0.012) 0.030 0.014 0.027 Above median saturation (0.032) (0.015) (0.021) Program village *Above median saturation 0.054 -0.003 -0.018 (0.038) (0.017) (0.025) Control observations 65 65 65 Treated observations 65 65 65 Control Mean Price (in Philippine pesos) 6.015 33.392 38.977 note: *** p<0.01, ** p<0.05, * p<0.1 All specifications include municipality fixed effects. All prices are in natural logs. Source: 2011 Pantawid impact evaluation survey 45 Table 3: Selected food prices (ln) and program saturation in nationally representative data Rice, raw regular Eggs Fish Meats Vegetables Milk milled (fresh) (fresh) (fresh) (fresh) Products Panel A: Model with cutoff at saturation of 50 percent Saturation 0.061*** 0.083*** -0.016 0.002 0.030 0.140** (0.014) (0.018) (0.015) (0.013) (0.018) (0.063) Saturation*Above cutoff 0.088* 0.185*** 0.090* 0.053 0.180*** -0.008 (0.049) (0.062) (0.054) (0.044) (0.062) (0.237) Above cutoff -0.022 -0.094** -0.042 -0.007 -0.061 -0.095 (0.030) (0.038) (0.033) (0.027) (0.038) (0.144) Saturation + Saturation*Above cutoff 0.149*** 0.268*** 0.074 0.055 0.210*** 0.132 (0.047) (0.059) (0.051) (0.042) (0.058) (0.227) Panel B: Model with cutoff at saturation of 65 percent Saturation 0.078*** 0.089*** -0.009 0.019 0.054*** 0.071 (0.013) (0.016) (0.014) (0.012) (0.017) (0.058) Saturation*Above cutoff -0.252*** 0.154 0.332*** 0.029 0.160 0.527 (0.088) (0.110) (0.095) (0.080) (0.109) (0.429) Above cutoff 0.245*** -0.070 -0.249*** -0.004 -0.060 -0.456 (0.067) (0.085) (0.072) (0.062) (0.085) (0.328) Saturation + Saturation*Above cutoff -0.174** 0.242** 0.323*** 0.048 0.214** 0.598 (0.086) (0.107) (0.093) (0.079) (0.107) (0.422) Panel C: Model with cutoff at saturation of 75 percent Saturation 0.095*** 0.097*** -0.014 0.023** 0.064*** 0.055 (0.012) (0.016) (0.013) (0.011) (0.016) (0.056) Saturation*Above cutoff -0.489*** 0.220 0.194 0.070 0.154 0.309 (0.147) (0.185) (0.158) (0.132) (0.184) (0.743) Above cutoff 0.440*** -0.139 -0.116 -0.047 -0.058 -0.250 (0.127) (0.160) (0.137) (0.114) (0.161) (0.643) Saturation + Saturation*Above cutoff -0.394*** 0.317* 0.180 0.093 0.217 0.364 (0.146) (0.183) (0.157) (0.131) (0.183) (0.737) Panel D: Linear model Saturation 0.095*** 0.114*** 0.001 0.028*** 0.087*** 0.062 (0.011) (0.015) (0.012) (0.010) (0.015) (0.053) note: *** p<0.01, ** p<0.05, * p<.1. Standard errors in parenthesis. Each row corresponds to a separate regression. Observations are at the village-food item-year level and pooled across the years 2009, 2012, and 2015. All models include year and village fixed effects. Prices are calculated as unit values; see text for details. Source: 2009, 2012, and 2015 Family Income and Expenditure Surveys and Department of Social Welfare Administrative Data on Pantawid Program Enrollment 46 Table 4: Program impact on food budget share for nonbeneficiary households with 6-60-month- old children and children’s food intake (1) (2) (3) (4) (5) (6) Number Whether Whether Whether Whether of Eggs Meat Fish Was Budget Eggs Dairy Was Fed to Was Fed Fed to Share of Were Fed Fed to Child in to Child Child in Food to Child in Child in Past in Past Past Past Week Past Week Week Week Week Panel A: Program impact Program village 0.036* 0.037 -0.154 -0.058 0.002 0.011 (0.021) (0.038) (0.170) (0.046) (0.036) (0.047) Panel B: Interaction with saturation Program village -0.016 0.301** 0.743 0.102 -0.015 0.382*** (0.066) (0.127) (0.640) (0.131) (0.113) (0.135) Saturation 0.089 -0.071 -0.216 -0.385** -0.313** 0.171 (0.104) (0.212) (0.963) (0.173) (0.152) (0.200) Program village*Saturation 0.085 -0.439** -1.495 -0.253 0.045 -0.627*** (0.117) (0.215) (1.056) (0.223) (0.178) (0.226) Panel C: Interaction with above median saturation Program village 0.033 0.123** 0.063 -0.047 0.005 0.064 (0.024) (0.049) (0.221) (0.058) (0.048) (0.066) Above median saturation 0.032 0.100 0.184 -0.105 -0.114** 0.058 (0.041) (0.068) (0.311) (0.064) (0.052) (0.080) Program village*Above 0.007 -0.203*** -0.504 -0.024 -0.006 -0.125 median saturation (0.044) (0.077) (0.350) (0.095) (0.067) (0.097) Panel D: Interaction with top quartile of saturation Program village 0.031 0.053 -0.022 -0.026 0.002 0.054 (0.024) (0.044) (0.184) (0.051) (0.039) (0.054) Fourth quartile of saturation 0.005 0.033 0.516 0.042 -0.144** 0.012 (0.047) (0.085) (0.476) (0.073) (0.071) (0.092) Program village*Fourth 0.026 -0.092 -0.846 -0.176 0.055 -0.212** quartile of saturation (0.057) (0.099) (0.516) (0.120) (0.093) (0.108) Control Observations 214 269 268 269 269 269 Treated Observation 230 295 292 292 293 292 Control Mean 0.596 0.755 2.131 0.665 0.818 0.342 note: *** p<0.01, ** p<0.05, * p<0.1 All child-level specifications include quadratic terms for child age (in months), and all specifications include municipality fixed effects. Standard errors are clustered at village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 47 Table 5: Program impact on 6-36-month-old nonbeneficiary children's physical growth, overall and by program saturation Height-for-Age Weight-for- Stunting Severe stunting Underweight Z score Age Z score Panel A: Program impact Program village -0.431** 0.123** 0.044 -0.111 0.06 (0.17) (0.06) (0.03) (0.17) (0.04) Panel B: Interaction with saturation Program village 0.043 -0.137 -0.038 1.557*** -0.381** (0.63) (0.22) (0.14) (0.58) (0.15) Saturation -0.892 0.197 0.068 0.249 -0.291 (0.79) (0.24) (0.15) (0.63) (0.21) Program -0.773 0.428 0.136 -2.786*** 0.745*** village*saturation (1.00) (0.34) (0.23) (0.94) (0.25) Panel C: Interaction with above median saturation Program village -0.314 0.061 0.039 0.257 -0.046 (0.24) (0.08) (0.04) (0.22) (0.05) Above median -0.397 0.074 0.048 0.066 -0.075 saturation (0.29) (0.08) (0.04) (0.26) (0.07) Program village*Above -0.196 0.117 0.006 -0.798** 0.231*** median saturation (0.35) (0.11) (0.07) (0.34) (0.09) Panel D: Interaction with top quartile of saturation Program village -0.244 0.041 0.012 0.127 0.018 (0.20) (0.06) (0.03) (0.18) (0.05) Fourth quartile of 0.073 0.02 -0.009 0.498* -0.125 saturation (0.37) (0.09) (0.07) (0.27) (0.09) Program village*Fourth -0.740* 0.319** 0.139 -1.320*** 0.245** quartile of saturation (0.41) (0.13) (0.10) (0.37) (0.11) Control Observations 145 145 145 156 156 Treated Observations 158 158 158 175 175 Control Mean -1.124 0.317 0.069 -0.922 0.186 note: *** p<0.01, ** p<0.05, * p<0.1 All specifications include quadratic terms for child age (in months), and municipality fixed effects. Standard errors are clustered at village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 48 Table 6: Program impact on health seeking behavior of nonbeneficiary households Pregnant women/last pregnancy in the last 36 months Children younger than 60 months old Any Timely Postnatal care Content of Institutional Growth Treatment antenatal initiation within 24 MMR care in ANC Delivery Monitoring seeking care of ANC hours Panel A: Program impact Program village -0.039 0.004 0.000 0.000 -0.061 0.032 0.020 0.001 (0.029) (0.059) (0.202) (0.054) (0.060) (0.048) (0.032) (0.044) Panel B: Interaction with saturation Program village -0.074 0.022 -0.819 -0.087 -0.204 -0.170 0.144 0.117 (0.060) (0.229) (0.592) (0.207) (0.195) (0.143) (0.110) (0.161) Saturation -0.168 -0.227 -2.577*** -0.285 -0.631*** -0.125 -0.088 -0.168 (0.123) (0.301) (0.860) (0.262) (0.231) (0.220) (0.133) (0.224) Program 0.066 -0.018 1.498 0.159 0.267 0.345 -0.208 -0.191 village*saturation (0.110) (0.363) (0.994) (0.333) (0.312) (0.249) (0.172) (0.273) Panel C: Interaction with above median saturation Program village -0.043 -0.061 -0.154 -0.018 -0.068 -0.029 0.056 0.015 (0.031) (0.082) (0.228) (0.076) (0.080) (0.063) (0.043) (0.052) Above median saturation -0.075 -0.192** -0.791** -0.029 -0.164* -0.137* -0.014 -0.088 (0.049) (0.093) (0.338) (0.076) (0.090) (0.083) (0.046) (0.072) Program village*above 0.011 0.146 0.347 0.042 0.021 0.142 -0.083 -0.031 median saturation (0.044) (0.115) (0.381) (0.119) (0.115) (0.102) (0.063) (0.094) Panel D: Interaction with top quartile of saturation Program village -0.030 0.037 0.114 -0.010 -0.048 0.046 0.031 0.019 (0.024) (0.063) (0.216) (0.062) (0.069) (0.055) (0.037) (0.046) Fourth quartile of -0.107 0.032 -0.380 0.092 -0.079 0.022 -0.039 0.144 saturation (0.082) (0.143) (0.482) (0.118) (0.087) (0.079) (0.051) (0.108) Program village*fourth 0.019 -0.153 -0.329 -0.011 -0.012 -0.078 -0.038 -0.135 quartile of saturation (0.091) (0.160) (0.544) (0.142) (0.121) (0.108) (0.065) (0.137) Control observations 172 180 169 178 180 239 268 258 Treated observations 202 203 184 198 203 267 299 291 Control mean 0.965 0.472 5.533 0.371 0.439 0.473 0.209 0.508 note: *** p<0.01, ** p<0.05, * p<0.1 Child level outcomes include quadratic terms for child age (in months), and all specifications include municipality fixed effects. Standard errors are clustered at village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 49 Appendix Table 1: Characteristic balance at baseline: beneficiary households Difference Standardized Control Baseline survey variables (Treated- Mean p-value Mean Control) Difference LN Proxy Means Test Score 9.093 -0.014 0.072 0.494 Household composition: Household size 5.828 0.004 0.001 0.973 Children 5 years old and below 1.063 0.069 -0.065 0.244 Children between 6 and 14 years old 1.718 -0.022 0.018 0.742 Primary occupation: Farming and 0.685 0.040 0.335 livestock -0.096 Educational attainment of the household head: No grade completed 0.078 -0.002 0.002 0.913 Some elementary school 0.433 -0.006 0.000 0.856 Completed elementary school 0.214 0.014 -0.044 0.585 Some high school 0.132 -0.029 0.104 0.105 High school graduate 0.100 0.004 0.000 0.841 Some college 0.031 0.008 -0.036 0.444 College graduate 0.015 0.009 -0.057 0.206 School Attendance: Children between 6 to 11 years 0.857 0.002 0.000 0.925 Children between 12 to 14 years 0.767 0.038 -0.094 0.216 Housing Amenities: Strong roof materials 0.307 -0.041 0.111 0.164 Strong wall materials 0.191 -0.023 0.080 0.276 Light roof materials 0.478 0.050 -0.130 0.107 Light wall materials 0.445 0.017 -0.065 0.577 Owns a house and lot 0.345 -0.034 0.093 0.239 House has toilet 0.388 -0.018 0.013 0.632 Shares a water source 0.200 -0.023 0.066 0.463 Household Assets: Electricity in house 0.407 0.021 -0.016 0.542 Owns a television 0.203 -0.004 0.030 0.879 Owns a video recorder 0.090 -0.020 0.091 0.241 Owns a Stereo/CD player 0.100 -0.006 0.015 0.755 Owns a refrigerator 0.012 -0.004 0.044 0.430 Has a telephone/cellphone 0.072 -0.013 0.064 0.292 Owns a motorcycle 0.025 0.002 -0.002 0.822 note: *** p<0.01, ** p<0.05, * p<0.1. Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 130 villages. Source: 2008 Household Assessment Form 50 Appendix Table 2: Characteristic balance at baseline: nonbeneficiary households Difference Standardized Control Baseline survey variables (Treated- Mean p-value Mean Control) Difference 0.703 LN Proxy Means Test Score 9.871 0.008 0.026 Household composition: Household size 4.240 -0.146 0.108 0.149 Children 5 years old and below 0.566 -0.015 0.032 0.702 Children between 6 and 14 years old 0.952 -0.037 0.051 0.547 Primary occupation: Farming and 0.312 -0.000 0.990 livestock 0.007 Educational attainment of the household head: No grade completed 0.028 0.001 -0.005 0.899 Some elementary school 0.230 0.029 -0.083 0.320 Completed elementary school 0.171 -0.008 -0.004 0.751 Some high school 0.135 -0.001 0.016 0.944 High school graduate 0.195 -0.018 0.059 0.402 Some college 0.122 -0.002 0.012 0.923 College graduate 0.121 -0.001 0.018 0.951 School Attendance: Children between 6 to 11 years 0.947 -0.008 0.041 0.689 Children between 12 to 14 years 0.873 -0.009 -0.010 0.804 Housing Amenities: Strong roof materials 0.662 -0.046 0.125 0.160 Strong wall materials 0.552 -0.039 0.111 0.217 Light roof materials 0.232 0.027 -0.097 0.352 Light wall materials 0.220 0.047* -0.145 0.066 Owns a house and lot 0.537 -0.030 0.085 0.369 House has toilet 0.177 -0.036 0.080 0.219 Shares a water source 0.173 0.001 -0.007 0.962 Household Assets: Electricity in house 0.797 0.001 0.036 0.989 Owns a television 0.570 -0.017 0.061 0.687 Owns a video recorder 0.400 -0.088** 0.213 0.014 Owns a Stereo/CD player 0.349 -0.052 0.114 0.110 Owns a refrigerator 0.312 -0.012 0.052 0.682 Has a telephone/cellphone 0.273 -0.054 0.133 0.104 Owns a motorcycle 0.173 -0.042* 0.130 0.068 Note: *** p<0.01, ** p<0.05, * p<0.1. Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 130 villages. Source: 2008 Household Assessment Form 51 Appendix Table 3: Characteristic balance at baseline: nonbeneficiary households in above median saturated areas Difference Standardized Control Baseline survey variables (Treated- Mean p-value Mean Control) Difference LN Proxy Means Test Score 9.816 0.003 -0.070 0.910 Household composition: Household size 4.088 -0.429** 0.269 0.016 Children 5 years old and below 0.491 -0.047 0.099 0.344 Children between 6 and 14 years old 0.894 -0.078 0.127 0.397 Primary occupation: Farming and livestock 0.372 -0.036 0.143 0.566 Educational attainment of the household head: No grade completed 0.031 0.002 -0.080 0.931 Some elementary school 0.231 0.059 -0.110 0.236 Completed elementary school 0.200 -0.043 0.135 0.264 Some high school 0.124 -0.013 -0.003 0.653 High school graduate 0.213 -0.032 0.104 0.412 Some college 0.107 0.008 -0.030 0.794 College graduate 0.093 0.019 -0.060 0.499 School Attendance: Children between 6 to 11 years 0.931 0.022 -0.073 0.498 Children between 12 to 14 years 0.851 -0.037 0.057 0.579 Housing Amenities: Strong roof materials 0.611 -0.086 0.182 0.135 Strong wall materials 0.482 -0.079 0.164 0.149 Light roof materials 0.221 0.095* -0.148 0.087 Light wall materials 0.212 0.093** -0.139 0.032 Owns a house and lot 0.389 0.031 -0.061 0.573 House has no toilet 0.208 0.009 0.024 0.827 Shares a water source 0.195 0.032 -0.087 0.612 Household Assets: Electricity in house 0.735 0.036 -0.049 0.301 Owns a television 0.562 -0.051 0.110 0.334 Owns a video recorder 0.389 -0.099* 0.241 0.071 Owns a Stereo/CD player 0.367 -0.068 0.154 0.231 Owns a refrigerator 0.398 -0.091** 0.148 0.028 Has a telephone/cellphone 0.279 -0.109** 0.225 0.037 Owns a motorcycle 0.204 -0.051 0.147 0.235 *** p<0.01, ** p<0.05, * p<0.1. Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 65 villages. Source: 2008 Household Assessment Form 52 Appendix Table 4: Characteristic balance at baseline: nonbeneficiary households in below median saturated areas Difference Standardized Control Baseline survey variables (Treated- Mean p-value Mean Control) Difference LN Proxy Means Test Score 9.911 0.023 0.082 0.416 Household composition: Household size 4.348 0.064 -0.002 0.586 Children 5 years old and below 0.620 0.026 -0.005 0.621 Children between 6 and 14 years old 0.994 0.023 0.004 0.781 Primary occupation: Farming and 0.269 0.039 0.357 livestock -0.094 Educational attainment of household head: No grade completed 0.025 -0.007 0.065 0.556 Some elementary school 0.229 0.010 -0.063 0.780 Completed elementary school 0.150 0.008 -0.103 0.813 Some high school 0.143 0.004 0.030 0.874 High school graduate 0.182 -0.002 0.025 0.932 Some college 0.134 -0.008 0.041 0.785 College graduate 0.140 -0.004 0.069 0.892 School Attendance: Children between 6 to 11 years 0.956 -0.020 0.110 0.466 Children between 12 to 14 years 0.889 0.007 -0.048 0.877 Housing Amenities: Strong roof materials 0.699 0.019 0.086 0.595 Strong wall materials 0.601 0.021 0.078 0.514 Light roof materials 0.241 -0.041 -0.061 0.202 Light wall materials 0.225 0.000 -0.148 0.994 Owns a house and lot 0.642 -0.064 0.196 0.113 House has no toilet 0.155 -0.063 0.125 0.110 Shares a water source 0.158 -0.034 0.056 0.302 Household Assets: Electricity in house 0.842 -0.007 0.107 0.874 Owns a television 0.576 0.016 0.027 0.762 Owns a video recorder 0.408 -0.063 0.194 0.135 Owns a Stereo/CD player 0.335 -0.029 0.085 0.436 Owns a refrigerator 0.250 0.046 -0.024 0.186 Has a telephone/cellphone 0.269 -0.014 0.071 0.740 Owns a motorcycle 0.152 -0.039 0.116 0.357 *** p<0.01, ** p<0.05, * p<0.1. Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 65 villages. Source: 2008 Household Assessment Form 53 Appendix Table 5: Characteristic balance at baseline: nonbeneficiary households in top quartile of saturated areas Difference Standardized Control Baseline survey variables (Treated- Mean p-value Mean Control) Difference LN Proxy Means Test Score 9.797 -0.013 -0.037 0.642 Household composition: Household size 4.096 -0.742** -0.189 0.025 Children 5 years old and below 0.521 -0.143* -0.150 0.071 Children between 6 and 14 years old 0.877 -0.168 -0.106 0.196 Primary occupation: Farming and livestock 0.425 -0.037 -0.027 0.739 Educational attainment of the household head: No grade completed 0.055 0.021 0.040 0.618 Some elementary school 0.247 0.049 0.046 0.566 Completed elementary school 0.247 -0.109*** -0.210 0.013 Some high school 0.082 0.019 0.031 0.701 High school graduate 0.192 -0.054 -0.082 0.311 Some college 0.110 0.020 0.032 0.692 College graduate 0.068 0.055 0.113 0.168 School Attendance: Children between 6 to 11 years 0.882 0.090* 0.149 0.076 Children between 12 to 14 years 0.762 -0.081 -0.042 0.608 Housing Amenities: Strong roof materials 0.507 -0.021 -0.017 0.834 Strong wall materials 0.466 -0.148* -0.134 0.104 Light roof materials 0.233 -0.010 -0.009 0.908 Light wall materials 0.192 0.106 0.114 0.164 Owns a house and lot 0.315 0.063 0.069 0.392 House has toilet 0.192 0.167** 0.205 0.015 Shares a water source 0.288 -0.160** -0.172 0.040 Household Assets: Electricity in house 0.781 -0.072 -0.122 0.138 Owns a television 0.548 -0.084 -0.091 0.266 Owns a video recorder 0.466 -0.283*** -0.251 0.004 Owns a Stereo/CD player 0.411 -0.153* -0.138 0.094 Owns a refrigerator 0.397 -0.144** -0.185 0.027 Has a telephone/cellphone 0.274 -0.086 -0.077 0.345 Owns a motorcycle 0.192 -0.093 -0.098 0.227 *** p<0.01, ** p<0.05, * p<0.1. Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 65 villages. Source: 2008 Household Assessment Form 54 Appendix Table 6: Food intake patterns of young children in control villages Percentage of children who Food item consumed the food item in the past week Children aged Children aged 0-5 years 0-3 years Fish 83.58 82.37 Green Vegetables 86.86 85.02 Banana 81.75 81.12 Eggs 73.11 71.92 Meat 56.45 53.98 Sweet potatoes 49.15 47.27 Papaya 36.74 36.35 Mango 20.32 19.34 Dairy 27.86 28.55 Carrot 13.5 13.1 Source: 2011 Pantawid impact evaluation survey 55 Appendix Table 7: Program impact on 6-36-month-old beneficiary children's physical growth, overall and by program saturation Height-for-Age Weight-for-Age Z Stunted Severely stunted Underweight Z score score Panel A: Interaction with saturation Program village 0.324 -0.369** -0.099 -0.394 0.086 (0.442) (0.153) (0.113) (0.471) (0.146) Saturation -0.264 -0.231 0.363* -1.384** 0.243 (0.678) (0.249) (0.195) (0.595) (0.203) -0.115 0.532** 0.016 0.827 -0.175 Program village*Saturation (0.691) (0.248) (0.182) (0.744) (0.233) Panel B: Interaction with above median saturation Program village 0.098 -0.084 -0.095* 0.099 0.020 (0.206) (0.074) (0.055) (0.197) (0.068) Above median saturation -0.033 -0.127 0.019 -0.050 -0.016 (0.297) (0.096) (0.072) (0.244) (0.077) Program village*Above 0.313 0.097 0.004 0.052 -0.086 median saturation (0.295) (0.107) (0.082) (0.302) (0.094) Panel C: Interaction with top quartile of saturation Program village 0.335* -0.071 -0.120** 0.022 -0.014 (0.191) (0.062) (0.050) (0.163) (0.056) 0.225 -0.065 0.013 -0.104 0.018 Fourth quartile of saturation (0.265) (0.087) (0.081) (0.258) (0.073) Program village*Fourth -0.264 0.120 0.092 0.343 -0.032 quartile of saturation (0.332) (0.125) (0.093) (0.361) (0.107) Control observations 162 162 162 186 186 Treated observations 181 181 181 201 201 Control mean -1.800 0.481 0.216 -1.207 0.280 *** p<0.01, ** p<0.05, * p<0.1. All specifications include municipality fixed effects. All specifications include quadratic terms for child age (in months). Standard errors are clustered by village; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 56 Appendix Table 8: Log food price impacts of program saturation at the province level, 2006-2014 Egg Fish Chicken Rice Snacks Sugar Panel A: Covariation with Saturation Program saturation 0.192 0.152 0.126 0.022 0.083 0.023 (0.124) (0.112) (0.103) (0.060) (0.096) (0.077) Province Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes Number of Observations 640 608 608 648 640 648 Panel B: Covariation with First Differenced Saturation Change in Saturation > 0 0.020** 0.001 0.004 0.009 -0.007 -0.001 (0.008) (0.007) (0.008) (0.008) (0.009) (0.014) Change in Saturation > 0.05 0.015* 0.011 0.012* 0.010 -0.002 -0.016 (0.009) (0.010) (0.006) (0.010) (0.006) (0.010) Change in Saturation> 0.10 0.018 0.012 0.014 0.007 0.002 0.003 (0.013) (0.010) (0.008) (0.009) (0.008) (0.016) Province Fixed Effects No No No No No No Year Fixed Effects Yes Yes Yes Yes Yes Yes Number of Observations 640 608 608 648 640 648 Provinces 79 75 78 80 79 80 Mean Price (Philippine pesos, 2006) 5.388 118.949 133.175 24.911 8.161 38.092 note: *** p<0.01, ** p<0.05, * p<0.1 1 Egg: log price of one medium chicken egg 2 Fish: log price of one kilogram of milkfish (bangus) 3 Chicken: log price of one kilogram of mixed chicken parts 4 Rice: log price of one kilogram of standard white rice 5 Snacks: log price of one 60-gram foil pack of Pancit Canton 6 Sugar: log price of one kilogram of unbranded refined sugar Source: Provincial Consumer Price Index Data 2006-2014 57 Appendix Table 9: Disaggregated food consumption by beneficiary and nonbeneficiary households with 6-60-month-old children Beneficiary Households Nonbeneficiary Households Log Per Capita Log Per Capita Log Per Capita Log Per Capita Log Per Capita Log Per Capita Dairy and Egg Cereal Other Food Dairy and Egg Cereal Other Food Consumption Consumption Consumption Consumption Consumption Consumption Panel A: Program Impact Program village 0.019 -0.002 0.020 -0.042 0.018 0.032 (0.029) (0.066) (0.113) (0.046) (0.065) (0.125) Panel B: Interaction with Saturation Program village 0.152 0.141 0.295 0.003 0.160 0.139 (0.107) (0.215) (0.373) (0.207) (0.304) (0.450) Saturation -0.138 0.098 0.091 -0.533*** -0.348 -0.747 (0.138) (0.305) (0.535) (0.182) (0.293) (0.520) Program village*saturation -0.221 -0.234 -0.451 -0.062 -0.230 -0.160 (0.173) (0.354) (0.577) (0.322) (0.475) (0.701) Panel C: Interaction with above median saturation Program impact 0.056 0.051 0.119 -0.018 0.090 0.156 (0.036) (0.076) (0.145) (0.069) (0.089) (0.172) Above median saturation 0.027 0.087 0.184 -0.039 0.055 0.154 (0.048) (0.106) (0.193) (0.069) (0.099) (0.202) Program Village*Above -0.082 -0.112 -0.207 -0.058 -0.168 -0.290 median saturation (0.057) (0.134) (0.226) (0.089) (0.130) (0.240) Panel D: Interaction with top quartile of saturation Program impact 0.036 -0.003 0.037 -0.022 0.046 0.093 (0.029) (0.069) (0.129) (0.050) (0.076) (0.141) Fourth quartile of saturation 0.069 0.204 0.322 0.092 0.135 0.356* (0.061) (0.147) (0.247) (0.085) (0.100) (0.201) Program Village* fourth -0.059 0.024 -0.035 -0.105 -0.147 -0.320 quartile of saturation (0.075) (0.163) (0.231) (0.112) (0.141) (0.277) Control Observations 328 328 328 214 214 214 Treated Observations 334 335 334 230 230 230 Control Mean 0.214 1.007 5.227 0.333 0.886 5.209 note: *** p<0.01, ** p<0.05, * p<0.1. All specifications include municipality fixed effects. Standard errors are clustered by village; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 58 Appendix Table 10: Program impact on nonbeneficiary 6-60-month-old children's physical growth by age category Height-for- Severely Weight-for- Stunted Underweight Age Z score stunted Age Z score -0.016 0.012 0.025 0.058 -0.030 Program village (0.148) (0.071) (0.042) (0.125) (0.046) Dummy: 6-24 months 0.213 -0.049 -0.059 0.290 -0.038 old (0.445) (0.156) (0.083) (0.416) (0.115) Dummy: 24-36 months 0.159 -0.112 -0.006 0.262 0.039 old (0.301) (0.108) (0.055) (0.252) (0.081) Program village*6-24 -0.622** 0.131 0.062 -0.166 0.198** dummy (0.259) (0.098) (0.060) (0.247) (0.078) Program village*24-36 -0.165 0.097 -0.040 -0.157 -0.035 dummy (0.255) (0.095) (0.063) (0.247) (0.082) Control Observations 242 242 242 265 265 Treated Observations 265 265 265 287 287 Control Mean -1.305 0.335 0.070 -0.966 0.181 note: *** p<0.01, ** p<0.05, * p<0.1 Age measured to the day of birth, therefore age categories are not overlapping All specifications include quadratic terms for child age (in months), and municipality fixed effects. Standard errors are clustered at village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 59 Appendix Table 11: Program impact on health seeking behavior of beneficiary households Pregnant women/last pregnancy in the last 36 months Children younger than 60 months old Timely Content of Postnatal Any antenatal Institutional Growth Treatment initiation of care in care within MMR care Delivery Monitoring seeking ANC ANC 24 hours Panel A: Program impact Program village 0.039 0.107** 0.719*** 0.102** 0.018 0.056 0.152*** 0.134*** (0.031) (0.049) (0.201) (0.049) (0.048) (0.039) (0.032) (0.039) Panel B: Interaction with saturation Program village -0.052 0.440** 1.133 0.235 -0.125 -0.072 0.321*** 0.043 (0.104) (0.175) (0.694) (0.172) (0.143) (0.121) (0.110) (0.139) Saturation -0.450*** 0.376 -2.485*** -0.657*** -0.873*** -0.260 0.047 -0.401** (0.173) (0.232) (0.960) (0.198) (0.199) (0.167) (0.113) (0.179) Program 0.144 -0.526** -0.613 -0.214 0.226 0.204 -0.275* 0.142 village*saturation (0.193) (0.265) (1.132) (0.252) (0.222) (0.198) (0.165) (0.207) Panel C: Interaction with above median saturation Program village 0.014 0.166** 0.716*** 0.124* -0.020 0.034 0.179*** 0.070 (0.030) (0.071) (0.261) (0.067) (0.070) (0.050) (0.045) (0.057) Above median saturation -0.072 0.101 -0.783*** -0.072 -0.187** -0.001 0.051 -0.131** (0.052) (0.086) (0.296) (0.073) (0.074) (0.070) (0.053) (0.065) Program village*above 0.046 -0.110 0.004 -0.047 0.065 0.048 -0.053 0.127 median saturation (0.060) (0.102) (0.400) (0.094) (0.100) (0.080) (0.070) (0.078) Panel D: Interaction with top quartile of saturation Program village 0.029 0.169*** 0.762*** 0.154*** 0.029 0.044 0.195*** 0.115** (0.028) (0.057) (0.209) (0.055) (0.059) (0.044) (0.039) (0.048) Fourth quartile of -0.083 0.213** -0.771* -0.094 -0.148* -0.020 0.070 -0.036 saturation (0.063) (0.093) (0.400) (0.081) (0.080) (0.081) (0.051) (0.066) Program village*fourth 0.033 -0.201* -0.109 -0.161* -0.031 0.041 -0.153** 0.066 quartile of saturation (0.088) (0.105) (0.501) (0.093) (0.100) (0.097) (0.071) (0.083) Control observations 221 227 205 223 227 376 411 403 Treated observations 233 242 226 241 242 408 443 436 Control mean 0.910 0.410 4.932 0.296 0.291 0.436 0.175 0.409 note: *** p<0.01, ** p<0.05, * p<0.1 Child level outcomes include quadratic terms for child age (in months), and all specifications include municipality fixed effects. Standard errors are clustered at village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 60 Appendix Table 12: Labor force participation of nonbeneficiary households with 6-60-month-old children Males Females Participates in Works for Participates in Works for Participates labor force for pay more labor force for pay more Works Participates in Works in labor more than 40 than 40 more than 40 than 40 for Pay labor force for Pay force hours per hours per hours per hours per week week week week Panel A: Program Impact 0.015 0.027 0.055 -0.023 0.003 0.055 -0.019 0.116 Program village (0.026) (0.051) (0.060) (0.077) (0.022) (0.066) (0.075) (0.092) Panel B: Interaction with saturation Program village -0.003 0.049 -0.047 0.072 0.042 0.407** 0.386* 0.367 (0.073) (0.159) (0.204) (0.242) (0.073) (0.197) (0.221) (0.344) Saturation -0.166 -0.344 -0.763*** -0.177 -0.095 0.147 -0.311 0.105 (0.150) (0.210) (0.281) (0.288) (0.110) (0.275) (0.316) (0.538) Program village* saturation 0.036 -0.027 0.181 -0.162 -0.064 -0.629* -0.691* -0.451 (0.142) (0.268) (0.324) (0.418) (0.137) (0.365) (0.401) (0.599) Panel B: Interaction with above median saturation Program village -0.006 -0.012 0.016 -0.044 -0.007 0.112 0.062 0.197* (0.026) (0.064) (0.074) (0.100) (0.024) (0.077) (0.083) (0.105) Above median saturation -0.037 -0.128 -0.217*** -0.039 -0.040 0.060 -0.076 -0.005 (0.050) (0.082) (0.081) (0.110) (0.035) (0.107) (0.101) (0.145) Program village* Above 0.050 0.095 0.088 0.050 0.023 -0.169 -0.193 -0.233 median saturation (0.061) (0.099) (0.098) (0.146) (0.054) (0.144) (0.150) (0.196) Panel B: Interaction with top quartile of saturation Program village -0.009 0.007 0.065 0.007 -0.007 0.079 -0.004 0.172* (0.023) (0.057) (0.061) (0.085) (0.022) (0.072) (0.079) (0.094) Fourth Quartile of saturation -0.086 -0.118 -0.097 0.079 -0.067 0.063 -0.048 0.023 (0.070) (0.096) (0.090) (0.114) (0.052) (0.119) (0.123) (0.158) Program village* fourth 0.115 0.091 -0.079 -0.194 0.042 -0.130 -0.113 -0.361 quartile of saturation (0.079) (0.129) (0.138) (0.222) (0.079) (0.181) (0.177) (0.292) Control observations 292 226 136 92 293 112 71 53 Treated observations 300 238 146 100 310 117 82 62 Control Mean 0.928 0.624 0.691 0.685 0.935 0.607 0.775 0.660 note: *** p<0.01, ** p<0.05, * p<0.1. All specifications include municipality fixed effects. Standard errors are clustered by village; there are 130 villages. Source: 2011 Pantawid impact evaluation survey. 61 Appendix Table 13: Education and child labor program impacts of on nonbeneficiary children Children Children 10-14 years 10-14 years Enrollment Enrollment Attendance Attendance old who old who of 6-11- of 12-14- of 6-11- of 12-14- worked in worked for year-olds year-olds year-olds year-olds the last pay in the week last week Panel A: Program Impact Program village -0.004 0.007 -0.004 -0.001 0.031 0.010 (0.010) (0.018) (0.006) (0.010) (0.027) (0.018) Panel B: Interaction with saturation Program village 0.006 -0.062 -0.043** -0.015 0.006 -0.004 (0.028) (0.065) (0.019) (0.019) (0.109) (0.073) Saturation 0.006 -0.090 -0.034* -0.018 -0.055 -0.184** (0.037) (0.055) (0.020) (0.024) (0.118) (0.084) Program -0.017 0.129 0.069** 0.024 0.043 0.024 village*saturation (0.047) (0.112) (0.029) (0.034) (0.163) (0.105) Panel B: Interaction with above median saturation Program village -0.001 0.006 -0.010 -0.001 0.043 0.006 (0.013) (0.022) (0.008) (0.011) (0.041) (0.028) Above Median 0.010 -0.025 -0.008 -0.005 0.016 -0.033 Saturation (0.010) (0.029) (0.006) (0.012) (0.038) (0.027) Program village*Above -0.007 0.002 0.015 -0.001 -0.031 0.012 Median Saturation (0.018) (0.039) (0.011) (0.019) (0.049) (0.030) Panel B: Interaction with top quartile of saturation Program village -0.008 0.015 -0.009 -0.003 0.021 0.008 (0.011) (0.020) (0.007) (0.011) (0.031) (0.021) Fourth quartile of -0.007 -0.001 -0.017 0.007 -0.038 -0.017 saturation (0.019) (0.030) (0.012) (0.013) (0.026) (0.021) Program 0.028 -0.022 0.030** 0.011 0.065 0.011 village*fourth quartile of (0.024) (0.049) (0.013) (0.013) (0.052) (0.025) saturation Control 355 181 342 163 301 301 Observations Treated 341 185 332 176 295 297 Observation Control Mean 0.986 0.950 0.985 0.985 0.103 0.033 note: *** p<0.01, ** p<0.05, * p<0.1. All specifications include linear controls for child age (in months) and municipality fixed effects. Standard errors are clustered by village; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 62 Annex 1: Program take-up, anthropometric outcome construction, and the robustness of results to data trimming In terms of program saturation, the impact evaluation survey and program Management Information System (MIS) database yield slightly different estimates. 1,418 households surveyed were eligible to become Pantawid beneficiaries in 2008, yet only those in treated villages were offered the program in 2009 by design. Among the 704 beneficiary eligible households sampled in the Pantawid villages, 85 percent (581) reported being beneficiaries of the program, while 1 percent (7) in the control villages also reported being beneficiaries. According to the program Management Information System (MIS) database, however, the control villages did not have any beneficiary households, and 91 percent (647) of the 704 sampled beneficiary eligible households in the Pantawid villages were considered participating beneficiaries of the program.34 Five percent of the households not eligible for the program reported being Pantawid beneficiaries, even though none of these households were program beneficiaries according to the program MIS database. The survey data include complete height-for-age data on 172 out of 186 nonbeneficiary children 6 to 36 months of age in treated areas and 151 of 163 nonbeneficiary children of the same age range in control areas. Weight data were collected for 177 6-to-36-month-old nonbeneficiary children in treated areas and 156 nonbeneficiary children in control areas. Anthropometric z-scores were calculated based on the WHO (2006) growth standard. Scores of more than 5 standard deviations above or below the reference mean were dropped from the sample (WHO, 2006). This trimming resulted in 14 of the 172 treated children and 6 of the 151 34 The lower percentage of households in Pantawid villages that reported being program beneficiaries in the survey may be explained in part by the fact that program participation is voluntary. Some households that identified as potential beneficiaries may have waived their right to the program. Another possibility is that through the community validation process of NHTS-PR, these households may have been taken off the list of eligible households. It is also possible that a potential beneficiary household was unaware of the community assembly where attendance is required for potential beneficiaries to sign up for the program and confirm their basic household information collected for the PMT. Although very small in number, it is more difficult to explain why nonbeneficiary households according to the program MIS reported themselves to be Pantawid beneficiaries in the survey. There is no official way for a household that was not identified as poor by the NTHS-PR to be registered as a Pantawid beneficiary. It is possible that the respondent was thinking of some other program that they received rather than Pantawid. 63 control children being dropped from the height-for-age regressions, and 2 of the 177 treated and none of the 156 control children being dropped from the weight-for-age regressions. Given the relatively small number of children for whom we have anthropometric data, a concern may be that the estimated impacts of Pantawid on nonbeneficiary children’s height-for- age and weight-for-age are driven by outliers in the data. This annex explores this issue in further depth by considering varying thresholds at which to trim the data, using -3 and 3 SD, -4 and 4 SD, -6 and 6 SD, and -7 and 7SD as alternative thresholds instead of the WHO (2006) recommended -5 and 5 SD. These results, presented in Annex 2 Tables 1 through 3, show that the results presented above are highly robust for all three specifications discussed in Table 6 above—simple program impact, interaction with above median saturation, and the interaction with the fourth quartile of saturation. We also consider Winsorizing 5 and 10 percent of the anthropometry data. As shown in Annex 2 Tables 4 and 5, the results are robust to this alternative method of treating the outliers in the anthropometry data. 64 Annex 1 Table 1: Varying trimming threshold on anthropometry data for 6-36-month olds Height-for- Severe Weight-for- Stunting Underweight Age Z score stunting Age Z score Trimming >-3 and <3 SD Program village -0.262* 0.089 0.008 -0.169 0.066* (0.139) (0.058) (0.008) (0.145) (0.040) Control Observations 131 131 131 145 145 Treated Observations 142 142 142 151 151 Control Mean -1.271 0.341 0.001 -0.821 0.160 Trimming >-4 and <4 SD Program village -0.335** 0.113* 0.045 -0.152 0.065 (0.159) (0.060) (0.029) (0.159) (0.043) Control Observations 140 140 140 153 153 Treated Observations 156 156 156 172 172 Control Mean -1.371 0.382 0.166 -0.947 0.206 Trimming >-6 and <6 SD Program village -0.533*** 0.129** 0.067* -0.102 0.057 (0.188) (0.059) (0.037) (0.166) (0.043) Control Observations 147 147 147 156 156 Treated Observations 164 164 164 175 175 Control Mean -1.452 0.402 0.106 -0.986 0.217 Trimming >-7 and <7 SD Program village -0.396* 0.129** 0.067* -0.134 0.057 (0.204) (0.059) (0.037) (0.175) (0.043) Control Observations 149 149 149 156 156 Treated Observations 164 164 164 176 176 Control Mean -1.457 0.402 0.106 -1.003 0.216 *** p<0.01, ** p<0.05, * p<0.1. All specifications include linear controls for child age (in months). Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 130 villages Source: 2011 Pantawid impact evaluation survey 65 Annex 1 Table 2: Varying trimming threshold on anthropometry data for 6-36-month-olds with above median saturation interaction Height-for- Severe Weight-for- Stunting Underweight Age Z score stunting Age Z score Trimming >-3 and <3 SD Program village -0.077 0.022 0.001 0.042 -0.011 (0.178) (0.074) (0.002) (0.178) (0.049) -0.077 0.024 -0.004 0.042 -0.094 Above median saturation (0.225) (0.083) (0.004) (0.226) (0.061) Program village*Above -0.415 0.152 0.017 -0.488* 0.180** median saturation (0.273) (0.107) (0.017) (0.294) (0.079) Control Observations 131 131 131 145 145 Treated Observations 142 142 142 151 151 Control Mean -1.271 0.341 0.001 -0.821 0.160 Trimming >-4 and <4 SD Program village -0.226 0.056 0.043 0.128 -0.028 (0.211) (0.077) (0.031) (0.203) (0.052) -0.362 0.069 0.053 0.000 -0.062 Above median saturation (0.250) (0.084) (0.036) (0.251) (0.074) Program village*Above -0.253 0.129 0.007 -0.626* 0.208** median saturation (0.306) (0.110) (0.055) (0.330) (0.087) Control Observations 140 140 140 153 153 Treated Observations 156 156 156 172 172 Control Mean -1.371 0.382 0.166 -0.947 0.206 Trimming >-6 and <6 SD Program village -0.333 0.063 0.047 0.257 -0.046 (0.253) (0.075) (0.046) (0.216) (0.053) -0.214 0.054 0.020 0.066 -0.075 Above median saturation (0.306) (0.082) (0.050) (0.261) (0.074) Program village*Above -0.443 0.146 0.046 -0.798** 0.231*** median saturation (0.381) (0.110) (0.075) (0.344) (0.087) Control Observations 147 147 147 156 156 Treated Observations 164 164 164 175 175 Control Mean -1.452 0.402 0.106 -0.986 0.217 Trimming >-7 and <7 SD Program village -0.103 0.063 0.047 0.251 -0.046 (0.282) (0.075) (0.046) (0.216) (0.053) -0.076 0.054 0.020 0.088 -0.075 Above median saturation (0.315) (0.082) (0.050) (0.268) (0.074) Program village*Above -0.648 0.146 0.046 -0.853** 0.231*** median saturation (0.396) (0.110) (0.075) (0.359) (0.087) Control Observations 149 149 149 156 156 Treated Observations 165 165 165 176 175 Control Mean -1.457 0.402 0.106 -1.003 0.217 *** p<0.01, ** p<0.05, * p<0.1. All specifications include linear controls for child age (in months). Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 66 Annex 1 Table 3: Varying trimming threshold on anthropometry data for 6-36-month-olds with fourth quartile of saturation interaction Height-for- Severe Weight-for- Stunting Underweight Age Z score stunting Age Z score Trimming >-3 and <3 SD Program village -0.118 0.025 -0.000 -0.012 0.048 (0.150) (0.059) (0.002) (0.160) (0.044) 0.181 0.020 -0.002 0.348 -0.097 Fourth quartile of saturation (0.346) (0.089) (0.005) (0.259) (0.075) Program village*Fourth quartile -0.774** 0.302** 0.043 -0.939*** 0.136 of saturation (0.385) (0.127) (0.042) (0.308) (0.090) Control Observations 131 131 131 145 145 Treated Observations 142 142 142 151 151 Control Mean -1.271 0.341 0.001 -0.821 0.160 Trimming >-4 and <4 SD Program village -0.153 0.035 0.011 0.057 0.029 (0.174) (0.062) (0.028) (0.174) (0.047) 0.045 0.045 0.026 0.357 -0.098 Fourth quartile of saturation (0.364) (0.094) (0.062) (0.261) (0.085) Program village*Fourth quartile -0.836** 0.324** 0.139 -1.157*** 0.218** of saturation (0.410) (0.129) (0.096) (0.353) (0.109) Control Observations 140 140 140 153 153 Treated Observations 156 156 156 172 172 Control Mean -1.371 0.382 0.166 -0.947 0.206 Trimming >-6 and <6 SD Program village -0.349* 0.058 0.036 0.127 0.018 (0.208) (0.062) (0.040) (0.180) (0.047) 0.488 -0.005 -0.033 0.498* -0.125 Fourth quartile of saturation (0.552) (0.091) (0.068) (0.273) (0.087) Program village*Fourth quartile -1.079* 0.329** 0.159 -1.320*** 0.245** of saturation (0.573) (0.126) (0.102) (0.373) (0.111) Control Observations 147 147 147 156 156 Treated Observations 164 164 164 175 175 Control Mean -1.452 0.402 0.106 -0.986 0.217 Trimming >-7 and <7 SD Program village -0.194 0.058 0.036 0.071 0.018 (0.227) (0.062) (0.040) (0.199) (0.047) 0.583 -0.005 -0.033 0.543* -0.125 Fourth quartile of saturation (0.545) (0.091) (0.068) (0.294) (0.087) Program village*Fourth quartile -1.214** 0.329** 0.159 -1.237*** 0.245** of saturation (0.564) (0.126) (0.102) (0.397) (0.111) Control Observations 149 149 149 156 156 Treated Observations 164 164 164 176 176 Control Mean -1.457 0.402 0.106 -1.003 0.216 *** p<0.01, ** p<0.05, * p<0.1 All specifications include linear controls for child age (in months). Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 67 Annex 1 Table 4: Winsorizing five percent of anthropometry data on 6-36-month-olds Height-for- Severe Weight-for- Stunting Underweight Age Z score stunting Age Z score Panel A: Program Impact Program village -0.380* 0.129** 0.067* -0.163 0.057 (0.210) (0.059) (0.037) (0.169) (0.043) Panel B: Interaction with above median saturation Program village -0.036 0.063 0.047 0.205 -0.046 (0.249) (0.075) (0.046) (0.206) (0.053) -0.189 0.054 0.020 0.087 -0.075 Above median saturation (0.298) (0.082) (0.050) (0.261) (0.074) Program village*Above median -0.771** 0.146 0.046 -0.812** 0.231*** saturation (0.382) (0.110) (0.075) (0.349) (0.087) Panel C: Interaction with top quartile of saturation Program village -0.191 0.058 0.036 0.066 0.018 (0.229) (0.062) (0.040) (0.184) (0.047) -0.000 -0.005 -0.033 0.518* -0.125 Fourth quartile of saturation (0.525) (0.091) (0.068) (0.279) (0.087) Program village*Fourth quartile of -0.883 0.329** 0.159 -1.322*** 0.245** saturation (0.579) (0.126) (0.102) (0.396) (0.111) Control Observations 151 151 151 156 156 Treated Observations 172 172 172 177 177 Control Mean -1.593 0.402 0.106 -1.01 0.217 *** p<0.01, ** p<0.05, * p<0.1 All specifications include linear controls for child age (in months). Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 68 Annex 1 Table 5: Winsorizing ten percent of anthropometry data on 6-36-month-olds Weight-for- Height-for- Severe Stunting Age Z Underweight Age Z score stunting score Panel A: Program Impact Program village -0.380* 0.129** 0.067* -0.163 0.057 (0.210) (0.059) (0.037) (0.169) (0.043) Panel B: Interaction with above median saturation Program village -0.036 0.063 0.047 0.205 -0.046 (0.249) (0.075) (0.046) (0.206) (0.053) -0.189 0.054 0.020 0.087 -0.075 Above median saturation (0.298) (0.082) (0.050) (0.261) (0.074) Program village*Above -0.771** 0.146 0.046 -0.812** 0.231*** median saturation (0.382) (0.110) (0.075) (0.349) (0.087) Panel C: Interaction with top quartile of saturation Program village -0.191 0.058 0.036 0.066 0.018 (0.229) (0.062) (0.040) (0.184) (0.047) Fourth quartile of -0.000 -0.005 -0.033 0.518* -0.125 saturation (0.525) (0.091) (0.068) (0.279) (0.087) Program village*Fourth -0.883 0.329** 0.159 -1.322*** 0.245** quartile of saturation (0.579) (0.126) (0.102) (0.396) (0.111) Control Observations 151 151 151 156 156 Treated Observations 172 172 172 177 177 Control Mean -1.588 0.402 0.106 -1.010 0.218 *** p<0.01, ** p<0.05, * p<0.1 All specifications include linear controls for child age (in months). Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 69 Annex 2: Deriving village-level program saturation and unit values from national data and robustness checks on unit value results We explore whether the association between prices and program saturation is detectable beyond our experimental sample by combining a number of data sources to derive estimates of village- level program saturation and of unit values (which we use as proxies for prices) for various food categories. The following describes the steps we take to generate these variables. Beneficiary and Census data Program administrative data, reported by provincial offices of the Department of Social Welfare, provide village-level information on the number of beneficiary households enrolled in Pantawid in each year from 2009 to 2015 ( ). In order-to estimate village-level program saturation, we scale these by an estimate of the total number of households in each village. To generate this estimate we combine data from the Census for 2015 and from the Family Income and Expenditure Survey (FIES) for 2009, 2012, and 2015. The census provides an estimate of the number of individuals in each village for 2015 (2015 ); the FIES provides an estimate of the ̅ ); and the ratio of average number of individuals per household in each village in each year ( ̅ ). Dividing the program data on these is our estimate of households per village (2015 / beneficiary households per village by the estimated total number of households per village yields our estimate of village- and year-specific program saturation: = 2015 / ̅ Unit-values (or price proxies) We do not observe prices at the level of each village. Instead, we use unit values of food items derived from the 2009, 2012, and 2015 rounds of the Philippine national household budget survey (the FIES). There are 93 food items common across all three rounds of FIES. We derive the household-level unit value for each food item by dividing household expenditure on that item by the quantity consumed. The village-level unit value is defined as the median unit value observed for the village. We aggregate the 93 items into 15 broad food categories provided by FIES: regular rice; other rice; roots and tubers; fresh eggs; processed eggs; fresh fish; processed 70 fish; fresh meats; processed meats; fresh fruits; fresh vegetables; processed rice and grains; coffee, cocoa, and tea; sugar; milk products.35 We aggregate items into each of these categories using as weights the national average food expenditure share of each item in 2009:36 = ∑ 2009 × =1 Finally, we estimate the models using the natural log of these estimated unit values. Full set of results Table 5 of the paper presents selected results from various specifications relating unit values to saturations. The models allow for the relationship to vary above and below a cutoff point, with the cutoff set in turn at 50, 65, and 75 percent saturation. An additional model estimates a simple linear model. Annex 2 Tables 1 to 4 report the full set of results from these specifications (that is, the results for all the different foods). Robustness of unit values results As discussed in the main text of this paper, unit values have the potential drawback of possibly conflating quality effects with the price estimate. We implement two variants on our derivation of village-level unit values in order to minimize the potential variability in quality. First, we restrict the range of household-level values considered by only including those that are within three standard deviations of the village mean unit value. Second, we consider unit values only for households that are within the top quartile of the per capita household consumption expenditures distribution. The majority of these households are not potential program beneficiaries and hence any quality choice in food consumption decisions should not be affected by the increase in income from program participation. As documented in Annex 2 Tables 5 and 6 as compared to Annex 2 Tables 2 and 4 (that is comparing the interaction models with a cutoff at 65 percent, as well as the linear models) the results are only minimally affected. 35 Our analysis excludes the processed fruit, juice powders and concentrate, chocolate, and “other foods” categories, since these were not consistently defined over time, or were not homogenous enough to constitute a category. 36 Regular rice and sugar are reported as single unit values and don’t require this aggregation. 71 Annex 2 Table 1: Food prices (ln) and program saturation in nationally representative data; Model with cutoff at saturation of 50 percent Share of Saturation + food Saturation*Above Saturation*Above expenditures Saturation cutoff Above cutoff N cutoff Rice, raw regular milled 0.191 0.061 *** 0.088 * -0.022 8370 0.149 *** (0.014) (0.049) (0.030) (0.047) Other raw rice and grains 0.074 -0.030 -0.301 0.034 8555 -0.332 (0.111) (0.371) (0.228) (0.351) Roots and tubers 0.011 0.247 *** 0.357 * -0.196 8487 0.604 *** (0.055) (0.197) (0.120) (0.187) Eggs (fresh) 0.018 0.083 *** 0.185 *** -0.094 ** 8551 0.268 *** (0.018) (0.062) (0.038) (0.059) Eggs (processed) 0.001 0.097 *** 0.135 -0.076 6449 0.232 (0.026) (0.155) (0.091) (0.152) Fish (fresh) 0.058 -0.016 0.090 * -0.042 8530 0.074 (0.015) (0.054) (0.033) (0.051) Fish (processed) 0.023 -0.075 * 0.162 -0.096 8479 0.087 (0.042) (0.152) (0.092) (0.144) Meats (fresh) 0.035 0.002 0.053 -0.007 8510 0.055 (0.013) (0.044) (0.027) (0.042) Meats (processed) 0.011 1.779 *** -5.163 *** 2.781 *** 8242 -3.384 ** (0.359) (1.383) (0.836) (1.334) Fruits (fresh) 0.021 -0.211 *** 0.275 ** -0.106 8512 0.064 (0.034) (0.118) (0.072) (0.111) Vegetables (fresh) 0.046 0.030 0.180 *** -0.061 8561 0.210 *** (0.018) (0.062) (0.038) (0.058) Rice and grains processed 0.021 0.169 *** -0.509 *** 0.290 *** 8555 -0.340 ** (0.054) (0.180) (0.110) (0.170) Coffee, cocoa, tea 0.025 0.093 ** -0.023 0.040 8559 0.070 (0.046) (0.156) (0.096) (0.148) Sugar 0.016 -0.541 *** 0.390 *** -0.163 *** 8529 -0.151 * (0.029) (0.097) (0.060) (0.092) Milk Products 0.007 0.140 ** -0.008 -0.095 8117 0.132 (0.063) (0.237) (0.144) (0.227) note: *** p<0.01, ** p<0.05, * p<.1. Standard errors in parenthesis. Each row corresponds to a separate regression. Observations are at the village-food item-year level and pooled across the years 2009, 2012, and 2015. All models include year and village fixed effects. Prices are calculated as unit values, see text for details. Source: 2009, 2012, and 2015 Family Income and Expenditure Surveys and Department of Social Welfare Administrative Data on Pantawid Program Enrollment 72 Annex 2 Table 2: Food prices (ln) and program saturation in nationally representative data; Model with cutoff at saturation of 65 percent Saturation + Share of food Saturation*Above Saturation*Above expenditures Saturation cutoff Above cutoff N cutoff Rice, raw regular milled 0.191 0.078 *** -0.252 *** 0.245 *** N -0.174 ** (0.013) (0.088) (0.067) 8370 (0.086) Other raw rice and grains 0.074 -0.157 0.474 -0.456 0.318 (0.100) (0.654) (0.506) 8555 (0.640) Roots and tubers 0.011 0.241 *** 0.771 ** -0.519 * 1.012 *** (0.050) (0.351) (0.270) 8487 (0.344) Eggs (fresh) 0.018 0.089 *** 0.154 -0.070 0.242 ** (0.016) (0.110) (0.085) 8551 (0.107) Eggs (processed) 0.001 0.093 *** 0.424 -0.284 0.517 * (0.025) (0.284) (0.218) 6449 (0.283) Fish (fresh) 0.058 -0.009 0.332 *** -0.249 *** 0.323 *** (0.014) (0.095) (0.072) 8530 (0.093) Fish (processed) 0.023 -0.066 * 0.638 ** -0.512 ** 0.572 ** (0.038) (0.290) (0.223) 8479 (0.285) Meats (fresh) 0.035 0.019 0.029 -0.004 0.048 (0.012) (0.080) (0.062) 8510 (0.079) Meats (processed) 0.011 1.652 *** -4.552 * 2.523 -2.901 (0.330) (2.692) (2.064) 8242 (2.657) Fruits (fresh) 0.021 -0.178 *** 0.091 0.017 -0.087 (0.031) (0.207) (0.160) 8512 (0.203) Vegetables (fresh) 0.046 0.054 *** 0.160 -0.060 0.214 ** (0.017) (0.109) (0.085) 8561 (0.107) Rice and grains processed 0.021 0.176 *** -0.429 0.213 -0.253 (0.049) (0.316) (0.245) 8555 (0.310) Coffee, cocoa, tea 0.025 0.114 *** -0.262 0.219 -0.149 (0.042) (0.277) (0.214) 8559 (0.271) Sugar 0.016 -0.504 *** 0.590 *** -0.360 *** 0.086 (0.026) (0.174) (0.135) 8529 (0.170) Milk Products 0.007 0.071 0.527 -0.456 0.598 (0.058) (0.429) (0.328) 8117 (0.422) note: *** p<0.01, ** p<0.05, * p<.1. Standard errors in parenthesis. Each row corresponds to a separate regression. Observations are at the village-food item-year level and pooled across the years 2009, 2012, and 2015. All models include year and village fixed effects. Prices are calculated as unit values, see text for details. Source: 2009, 2012, and 2015 Family Income and Expenditure Surveys and Department of Social Welfare Administrative Data on Pantawid Program Enrollment 73 Annex 2 Table 3: Food prices (ln) and program saturation in nationally representative data; Model with cutoff at saturation of 75 percent Saturation + Share of food Saturation*Above Saturation*Above expenditures Saturation cutoff Above cutoff N cutoff Rice, raw regular milled 0.191 0.095 *** -0.489 *** 0.440 *** 8370 -0.394 *** (0.012) (0.147) (0.127) (0.146) Other raw rice and grains 0.074 -0.193 ** -1.071 1.003 8555 -1.264 (0.096) (1.103) (0.963) (1.093) Roots and tubers 0.011 0.243 *** 1.315 ** -1.005 ** 8487 1.558 *** (0.048) (0.579) (0.500) (0.574) Eggs (fresh) 0.018 0.097 *** 0.220 -0.139 8551 0.317 * (0.016) (0.185) (0.160) (0.183) Eggs (processed) 0.001 0.092 *** -0.134 0.229 6449 -0.042 (0.025) (0.509) (0.440) (0.508) Fish (fresh) 0.058 -0.014 0.194 -0.116 8530 0.180 (0.013) (0.158) (0.137) (0.157) Fish (processed) 0.023 -0.080 ** 0.939 * -0.773 * 8479 0.859 * (0.036) (0.486) (0.418) (0.481) Meats (fresh) 0.035 0.023 ** 0.070 -0.047 8510 0.093 (0.011) (0.132) (0.114) (0.131) Meats (processed) 0.011 1.476 *** -6.996 5.103 8242 -5.520 (0.321) (4.295) (3.667) (4.268) Fruits (fresh) 0.021 -0.154 *** 1.281 *** -1.099 *** 8512 1.127 *** (0.030) (0.349) (0.304) (0.346) Vegetables (fresh) 0.046 0.064 *** 0.154 -0.058 8561 0.217 (0.016) (0.184) (0.161) (0.183) Rice and grains processed 0.021 0.148 *** -1.073 ** 0.848 * 8555 -0.925 * (0.047) (0.534) (0.466) (0.530) Coffee, cocoa, tea 0.025 0.126 *** -0.381 0.306 8559 -0.256 (0.040) (0.466) (0.406) (0.462) Sugar 0.016 -0.489 *** 0.993 *** -0.750 *** 8529 0.505 * (0.025) (0.289) (0.252) (0.286) Milk Products 0.007 0.055 0.309 -0.250 8117 0.364 (0.056) (0.743) (0.643) (0.737) note: *** p<0.01, ** p<0.05, * p<.1. Standard errors in parenthesis. Each row corresponds to a separate regression. Observations are at the village-food item-year level and pooled across the years 2009, 2012, and 2015. All models include year and village fixed effects. Prices are calculated as unit values, see text for details. Source: 2009, 2012, and 2015 Family Income and Expenditure Surveys and Department of Social Welfare Administrative Data on Pantawid Program Enrollment 74 Annex 2 Table 4: Food prices (ln) and program saturation in nationally representative data; Linear model Share of food expenditures Saturation N Rice, raw regular milled 0.191 0.095 *** 8370 (0.011) Other raw rice and grains 0.074 -0.185 ** 8555 (0.089) Roots and tubers 0.011 0.290 *** 8487 (0.045) Eggs (fresh) 0.018 0.114 *** 8551 (0.015) Eggs (processed) 0.001 0.102 *** 6449 (0.024) Fish (fresh) 0.058 0.001 8530 (0.012) Fish (processed) 0.023 -0.062 * 8479 (0.034) Meats (fresh) 0.035 0.028 *** 8510 (0.010) Meats (processed) 0.011 1.238 *** 8242 (0.305) Fruits (fresh) 0.021 -0.136 *** 8512 (0.028) Vegetables (fresh) 0.046 0.087 *** 8561 (0.015) Rice and grains processed 0.021 0.111 *** 8555 (0.043) Coffee, cocoa, tea 0.025 0.114 *** 8559 (0.037) Sugar 0.016 -0.444 *** 8529 (0.023) Milk Products 0.007 0.062 8117 (0.053) note: *** p<0.01, ** p<0.05, * p<.1. Standard errors in parenthesis. Each row corresponds to a separate regression. Observations are at the village-food item-year level and pooled across the years 2009, 2012, and 2015. All models include year and village fixed effects. Prices are calculated as unit values, see text for details. Source: 2009, 2012, and 2015 Family Income and Expenditure Surveys and Department of Social Welfare Administrative Data on Pantawid Program Enrollment 75 Annex 2 Table 5: Alternative estimates of relationship between food prices (ln) and program saturation in nationally representative data-- unit values derived after trimming values +/- 3 SD from mean. Model with cutoff at saturation of 65 percent Linear Model Saturation + Saturation*Above Saturation*Above Saturation cutoff Above cutoff cutoff Saturation Rice, raw regular milled 0.078 *** -0.251 *** 0.245 *** -0.173 ** 0.095*** 0.013 0.088 0.067 0.086 0.011 Other raw rice and grains -0.157 0.475 -0.456 0.318 -0.185** 0.100 0.654 0.506 0.640 0.089 Roots and tubers 0.241 *** 0.767 ** -0.515 * 1.008 *** 0.290*** 0.050 0.351 0.270 0.344 0.045 Eggs (fresh) 0.089 *** 0.154 -0.070 0.242 ** 0.114*** 0.016 0.110 0.085 0.107 0.015 Eggs (processed) 0.093 *** 0.424 -0.284 0.517 * 0.102*** 0.025 0.284 0.218 0.283 0.024 Fish (fresh) -0.009 0.332 *** -0.249 *** 0.323 *** 0.001 0.014 0.095 0.072 0.093 0.012 Fish (processed) -0.066 * 0.638 ** -0.512 ** 0.572 ** -0.062* 0.038 0.290 0.223 0.285 0.034 Meats (fresh) 0.019 0.029 -0.004 0.048 0.028*** 0.012 0.080 0.062 0.079 0.010 Meats (processed) 1.652 *** -4.552 * 2.523 -2.901 1.238*** 0.330 2.692 2.064 2.657 0.305 Fruits (fresh) -0.178 *** 0.091 0.017 -0.087 -0.136*** 0.031 0.207 0.160 0.203 0.028 Vegetables (fresh) 0.054 *** 0.161 -0.061 0.214 ** 0.087*** 0.017 0.109 0.085 0.107 0.015 Rice and grains processed 0.176 *** -0.429 0.213 -0.253 0.111*** 0.049 0.317 0.245 0.310 0.043 Coffee, cocoa, tea 0.114 *** -0.262 0.219 -0.149 0.114*** 0.042 0.277 0.214 0.271 0.037 Sugar -0.504 *** 0.590 *** -0.360 *** 0.086 -0.444*** 0.026 0.174 0.135 0.170 0.023 Milk Products 0.071 0.527 -0.456 0.598 0.062 0.058 0.429 0.328 0.422 0.053 note: *** p<0.01, ** p<0.05, * p<.1. Standard errors in parenthesis. Each row corresponds to a separate regression. Observations are at the village-food item-year level and pooled across the years 2009, 2012, and 2015. All models include year and village fixed effects. Prices are calculated as unit values, see text for details. Source: 2009, 2012, and 2015 Family Income and Expenditure Surveys and Department of Social Welfare Administrative Data on Pantawid Program Enrollment 76 Annex 2 Table 6: Alternative estimates of relationship between food prices (ln) and program saturation in nationally representative data-- unit values derived from households in top quartile of per capita consumption-expenditures. Model with cutoff at saturation of 65 percent Linear Model Saturation + Saturation*Above Saturation*Above Saturation cutoff Above cutoff cutoff Saturation Rice, raw regular milled 0.081 *** -0.238 *** 0.231 *** -0.158 * 0.096*** 0.013 0.090 0.069 0.089 0.012 Other raw rice and grains -0.167 * 0.495 -0.462 0.328 -0.190** 0.100 0.652 0.505 0.639 0.088 Roots and tubers 0.248 *** 0.686 * -0.423 0.934 ** 0.308*** 0.058 0.405 0.311 0.397 0.052 Eggs (fresh) 0.088 *** 0.128 -0.051 0.216 * 0.112*** 0.017 0.113 0.087 0.111 0.015 Eggs (processed) 0.099 *** 0.413 -0.277 0.511 * 0.108*** 0.026 0.300 0.231 0.299 0.026 Fish (fresh) -0.012 0.271 *** -0.212 *** 0.259 *** -0.008 0.014 0.095 0.073 0.094 0.013 Fish (processed) -0.061 0.642 ** -0.507 ** 0.581 ** -0.054 0.038 0.293 0.226 0.288 0.034 Meats (fresh) 0.015 0.016 0.015 0.030 0.027*** 0.012 0.083 0.064 0.082 0.011 Meats (processed) 1.646 *** -4.532 * 2.511 -2.887 1.233*** 0.330 2.689 2.062 2.654 0.304 Fruits (fresh) -0.178 *** 0.117 -0.020 -0.061 -0.144*** 0.031 0.208 0.161 0.204 0.028 Vegetables (fresh) 0.053 *** 0.190 * -0.078 0.243 ** 0.089*** 0.017 0.110 0.085 0.107 0.015 Rice and grains processed 0.176 *** -0.326 0.146 -0.150 0.119*** 0.049 0.318 0.246 0.311 0.043 Coffee, cocoa, tea 0.128 *** -0.337 0.253 -0.209 0.116*** 0.042 0.276 0.214 0.270 0.037 Sugar -0.493 *** 0.564 *** -0.346 ** 0.071 -0.436*** 0.026 0.175 0.136 0.171 0.023 Milk Products 0.087 0.394 -0.362 0.481 0.072 0.059 0.434 0.331 0.427 0.054 note: *** p<0.01, ** p<0.05, * p<.1. Standard errors in parenthesis. Each row corresponds to a separate regression. Observations are at the village-food item-year level and pooled across the years 2009, 2012, and 2015. All models include year and village fixed effects. Prices are calculated as unit values, see text for details. Source: 2009, 2012, and 2015 Family Income and Expenditure Surveys and Department of Social Welfare Administrative Data on Pantawid Program Enrollment 77 Annex 3: Robustness of main effects to differences in observable baseline population characteristics Although our identification leverages the randomized treatment assignment of villages, one may still be concerned may be that differences in outcomes attributed to saturation interacted with treatment are correlated with other factors that affect nutritional outcomes. Regarding household characteristics, as shown in Appendix Table 3, at baseline, nonbeneficiary households in above median saturated treated areas had slightly smaller household sizes than nonbeneficiary households in control above median saturated areas as well as a handful of other differential household characteristics. Using differences-in-differences (Panel A) and triple differences (Panel B), we show in Annex 3 Table 1 that the household size and the dependency ratios do not significantly vary over time across saturation or treatment categories. Any difference among high saturation treatment villages is consistent over time. That older child heights show no difference between treatment and control villages (Table 7) even though these children were raised in the same household structure as the younger children with significantly different heights indicates the baseline difference in household size and composition likely does not cause the higher stunting rates for younger children in treatment villages. It may still be that the lack of baseline balance more generally is driving the effects on nonbeneficiary young children’s physical growth. This annex then further explores the robustness of the results presented in Table 6 after covariate balancing using the relatively rich set of observables from the baseline. To balance observable characteristics, we estimate propensity scores as a function of baseline population characteristics, and then use these propensity scores to approximate what the outcome would have been if the treated population had exhibited the same baseline characteristics as the control population (following the approach discussed in Hirano, Imbens and Ridder 2003). Results presented in Annex 3 Table 2 are nearly identical to those presented in Table 6. Taken together, the lack of differential stunting among older children and the robustness tests presented in Annex 3 Tables 1 and 2 suggest that small differences in household size and a few other characteristics are not driving the estimated effects of Pantawid on nonbeneficiary children’s growth. 78 Annex 3 Table 1: Program impact on the household size and composition of nonbeneficiary households with 6-60-month-old children Dependency Ratio Dependency Ratio- Dependency Ratio- Household (Children 0-15/Adults Female (Children 0- Male (Children 0- Size 15-59) 15/Females 15-59) 15/Males 15-59) Panel A: Difference-in-differences Program village -0.006 -0.026 -0.031 -0.013 (0.148) (0.045) (0.087) (0.09) After 1.096*** -0.062*** -0.115*** -0.140*** (0.092) (0.019) (0.032) (0.034) Program village*After 0.116 0.01 0.062 0.022 (0.137) (0.026) (0.043) (0.051) Panel B: Triple difference interaction with saturation Program village 0.942* 0.239 0.488 0.515 (0.499) (0.158) (0.304) (0.321) After 1.101*** -0.084 0.036 -0.093 (0.274) (0.056) (0.105) (0.119) Saturation -1.499** -0.154 -0.505 -0.538 (0.641) (0.185) (0.365) (0.367) Program village*after -0.246 -0.083 -0.209 -0.22 (0.39) (0.091) (0.153) (0.191) -1.558* -0.440* -0.862* -0.873* Program village*saturation (0.803) (0.262) (0.488) (0.51) 0.295 0.038 -0.261 -0.078 After*Saturation (0.47) (0.101) (0.166) (0.191) Program 0.683 0.156 0.464** 0.404 village*after*saturation (0.672) (0.148) (0.236) (0.287) Panel C: Triple difference interaction with above median saturation Program village 0.360* 0.033 0.088 0.128 (0.196) (0.06) (0.121) (0.128) After 1.077*** -0.081*** -0.106** -0.122** (0.124) (0.025) (0.048) (0.048) -0.263 -0.065 -0.196 -0.169 Above median saturation (0.227) (0.069) (0.127) (0.136) Program village*After -0.03 0.008 0.051 -0.04 (0.17) (0.035) (0.065) (0.077) Program village*Above -0.830*** -0.131 -0.264 -0.314* median saturation (0.297) (0.092) (0.178) (0.185) After*Above median 0.058 0.047 -0.014 -0.036 saturation (0.182) (0.042) (0.061) (0.068) Program 0.315 -0.001 0.01 0.121 village*After*Above median saturation (0.282) (0.051) (0.083) (0.096) Panel D: Triple difference interaction with top quartile of saturation Program village 0.142 0.007 0.051 0.097 (0.166) -0.049) -0.098) (0.102) After 1.325*** -0.070*** -0.119*** -0.131*** (0.102) (0.022) (0.037) (0.038) -0.473 -0.087 -0.183 -0.123 79 Fourth quartile of (0.306) (0.087) (0.171) (0.183) saturation Program village*After 0.051 0.006 0.062 -0.019 (0.156) (0.03) (0.052) (0.059) Program village*fourth -0.604* -0.136 -0.357* -0.498** quartile of saturation (0.36) -(.118) (0.212) (0.221) After*fourth quartile of -0.292 0.049 0.027 -0.037 saturation (0.247) (0.048) (0.055) (0.089) Program 0.547 0 -0.014 0.161 village*after*fourth quartile of saturation (0.355) (0.055) (0.071) (0.106) Control Observations 418 417 409 398 Treated Observations 451 447 445 433 Control Mean 4.488 0.625 1.208 1.189 note: *** p<0.01, ** p<0.05, * p<0.1 All specifications include municipality fixed effects. Standard errors are clustered at the village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 80 Annex 3 Table 2: Covariate balanced estimates of program impact on nonbeneficiary children's anthropometry Height-for- Weight-for- Stunting Severe stunting Underweight Age Z score Age Z score Panel A: Program Impact Program village -0.388** 0.104* 0.040 -0.064 0.044 (0.180) (0.061) (0.033) (0.174) (0.044) Panel B: Interaction with saturation Program village -0.076 -0.039 0.015 1.377** -0.437*** (0.242) (0.740) (0.240) (0.164) (0.619) Saturation -0.757 0.159 0.130 0.456 -0.437* (0.937) (0.273) (0.186) (0.677) (0.231) Program -0.638 0.279 0.056 -2.522** 0.813*** village*saturation (1.141) (0.374) (0.262) (0.994) (0.259) Panel B: Interaction with Above Median Saturation Program village -0.294 0.049 0.033 0.294 -0.072 (0.242) (0.247) (0.082) (0.042) (0.220) Above median -0.434 0.072 0.035 0.101 -0.088 saturation (0.303) (0.095) (0.050) (0.286) (0.074) Program village*Above -0.206 0.122 0.016 -0.785** 0.255*** median saturation (0.356) (0.117) (0.068) (0.359) (0.085) Panel C: Interaction with top quartile of saturation Program village -0.227 0.028 0.010 0.157 0.002 (0.198) (0.063) (0.034) (0.185) (0.048) Fourth quartile of -0.001 0.010 -0.020 0.604** -0.139 saturation (0.347) (0.096) (0.064) (0.280) (0.086) Program village*Fourth -0.730* 0.343*** 0.147 -1.338*** 0.264** quartile of saturation (0.392) (0.126) (0.092) (0.381) (0.116) Control Observations 145 145 145 156 156 Treated Observations 158 158 158 175 175 Control Mean -1.124 0.317 0.069 -0.922 0.186 *** p<0.01, ** p<0.05, * p<0.1. Estimates include municipality fixed effects and are propensity weighted following Hirano et al. (2003). Standard errors clustered at the village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 81 Annex 4: Remoteness and Prices This paper presents evidence suggesting that the increased cash flow from Pantawid in highly saturated villages led to the increase in prices of non-easily traded perishable foods, for which local markets are not as well integrated with wider regional markets as for tradables. If this hypothesis is correct, we may expect to see the greatest increase in reported prices for remote highly saturated villages if remoteness is a good proxy for local market structure and the degree of integration with wide regional markets. We can examine the relationship between distance and price increases using the prices of individual food goods reported by the survey respondents. Annex 4 Table 1 presents the estimates of Pantawid impact on unit values for treated and control villages that are above median saturation as well as above median travel time to the nearest market. As we only have 130 villages in the sample, the estimates are largely imprecise, but the estimated coefficients suggest that the price increases are indeed somewhat larger for treated villages that are above median saturated and have an above median travel time to the nearest market, at least for the signal protein-rich good, eggs. For more easily tradable goods such as rice and sugar, the triple interaction coefficient of treatment, above median saturation and above median travel time to the nearest market is close to zero. In theory, a local market is less likely to be integrated with wider regional markets if the village is remote. However, the Pantawid evaluation was not designed to estimate differential impact by remoteness categories. Indeed, in this context, it is not clear what the relevant remoteness metric might be. For example, we also conducted similar analysis to Annex 4 Table 1, but using above median travel time to the municipal center as an alternative metric of remoteness; the results are largely the same as those presented in the table. Given the clear linkage between price changes, stunting, and program saturation, we believe that the saturation metric is the most comprehensive available measure that encapsulates the potential for general equilibrium impacts in this context. 82 Annex 4 Table 1: Remoteness and food prices reported by households (in natural logs) Egg Price Rice Price Sugar Price Program village -0.076 -0.017 -0.033 (0.055) (0.025) (0.045) Above median saturation -0.092 0.008 -0.053 (0.071) (0.033) (0.059) Above Median Distance -0.172* -0.068 -0.091 to Nearest Market (0.093) (0.043) (0.077) Treated*Saturation 0.195** 0.025 0.063 (0.092) (0.042) (0.076) Treated*Above Median -0.053 0.038 0.032 Travel Time to Nearest Market (0.099) (0.045) (0.082) Saturation* 0.307** 0.089 0.165 Above Median Travel Time to Nearest Mkt (0.133) (0.061) (0.110) Treatment* Saturation * 0.004 -0.047 -0.067 Above Med. Travel Time to Nearest Market (0.150) (0.069) (0.124) Program village 0.014 -0.004 0.001 (0.020) (0.009) (0.017) Fourth quartile of saturation -0.037 0.021 -0.017 (0.060) (0.027) (0.050) Above Median Distance 0.013 -0.014 0.008 to Nearest Market (0.024) (0.011) (0.020) Treated*fourth quartile of saturation 0.103** 0.006 0.008 (0.042) (0.019) (0.034) Treated*Above Median -0.017 0.014 0.003 Travel Time to Nearest Market (0.031) (0.014) (0.025) Fourth quartile of saturation* 0.087*** 0.020 0.038 Above Median Travel Time to Nearest Mkt (0.034) (0.015) (0.028) Treatment*Fourth Quartile of Saturation * -0.101* -0.020 -0.036 Above Med. Travel Time to Nearest Market (0.058) (0.026) (0.048) Control Observations 65 65 65 Treated Observations 65 65 65 Control Mean 6.015 33.392 38.977 *** p<0.01, ** p<0.05, * p<0.1 Estimates include municipality fixed effects. Standard errors clustered at the village level; there are 130 villages. Source: 2011 Pantawid impact evaluation survey 83 Annex 5: Costs and benefits of area (or village) targeting in the poorest areas To provide an order of magnitude for the ratio of benefits to costs of extending program saturation in the poorest villages to the entire village, we carry out an exercise relating (i) the discounted value of labor market returns to averting stunting at age 36 months to (ii) the discounted costs of the family transfer associated with adding a household, with one 6-12 month old child, to the program roll. Note that the benefits in this exercise are narrowly defined to those associated with lifetime earnings. To estimate the impact of stunting on labor market returns, we use the parameter estimate from Hoddinott et al (2013) who find that hourly earnings among adults who were stunted at age 36 months are 0.58 times the hourly earnings of those who were not stunted, after controlling for a number of contextual factors. Based on the average daily adult wages reported in our sample (US$6.3), we assume that the annual earnings of an adult who was stunted as a child are 0.58 times those of an adult who was not stunted as a child in each year that they work. Since we find that Pantawid increased the prevalence of stunting by 12 percentage points among nonbeneficiary children, we further multiply this value by 0.12 to estimate only the value of the stunting differential attributable to the program. Using these parameters, we estimate that the discounted lifetime benefits of the program’s impact on stunting (manifested in lifetime earnings) equals the discounted program costs when (i) we assume that real wages will grow at a rate of 1.75 percent per year, which is close to the rate observed in 2012 and 201337, and (ii) apply a discount rate of 5 percent. At any lower discount rate (holding projected real wage growth constant) the benefit/cost ratio exceeds one; at any higher projected real wage growth (holding the discount rate constant) the benefit/cost ratio also exceeds one. To fix ideas, the above estimates assume an annual family transfer of US$132 for when the child is aged 1 to 14 (that is, the basic transfer amount); discounted back to age 0, this amounts to a total per-child value of the transfer (i.e. the cost) of US$1,636. Given the real wage growth estimate of 1.75 percent and a discount rate of 5 percent and attributing 12 percent of the gap in earnings between adults who were stunted and those who were not stunted to the program, this is also equal to lifetime earnings detriment associated with the program. 37 1.9 and 1.5 percent real wage growth reported for 2012 and 2013, respectively, in ILO (2014). 84 Of course, an exercise such as this is sensitive to several assumptions. In particular, the labor market penalty associated with having been stunted at age 36 months (drawn from Hoddinott et al. 2013) appears to be quite high. We estimate that the program would have a benefit/cost ratio of greater than one if hourly wages for those who were stunted are 0.75 or less than those who were not stunted. Further we have simplified the cost implications of switching from a household to a village targeting mechanism. Adding more households to the beneficiary rolls would undoubtedly increase total administrative cost to some degree, yet at the same time the adoption of a village-based targeting rule may require far less household information to be collected and so would likely provide savings in this dimension. Finally, other welfare benefits from stunting aversion, such as increased longevity and non-pecuniary benefits from increased educational attainment and higher cognitive ability, are not accounted for in this exercise but would serve to increase benefits without affecting the cost of program expansion. 85