Informal Monitoring and Enforcement Mechanisms in Public Service Delivery: Evidence from the Public Distribution System in India

Transcription

1 Informal Monitoring and Enforcement Mechanisms in Public Service Delivery: Evidence from the Public Distribution System in India Sriniketh Nagavarapu and Sheetal Sekhri February 14, 2014 Abstract This paper shows that social networks provide informal monitoring and enforcement services that can curb leakages and improve the efficacy of public service delivery. We examine India s Targeted Public Distribution System, under which poor households are entitled to subsidized grains and fuel, and show that Scheduled Castes (SC) are more likely to buy grains when facing SC delivery agents. Using our theoretical and empirical framework, we show that increased informal monitoring drives our findings. Our structural estimates indicate that welfare gains from caste-based monitoring are approximately one-fifth of the average subsidy. Welfare gains from expanding program generosity, as envisioned in the recent National Food Security Bill, are lower than the subsidy cost due to an increase in delivery agents incentives to black market goods. Keywords: Food Security, Informal Networks, Service Delivery JEL Codes: I38, O17, Q18 and Z13 We thank Florencia Borrescio-Higa, Paul Christian, Brian Cross, Sisir Debnath, and David Glick for excellent research assistance. In addition, we have received valuable comments from Anna Aizer, Ken Chay, Marcel Fafchamps, Leora Friedberg, Andrew Foster, Vernon Henderson, Yongsuk Lee, Thomas MaCurdy, Aprajit Mahajan, Kaivan Munshi, Karthik Muralidharan, Holger Sieg, Steve Stern, Sandip Sukhtankar, Nick Wilson, Frank Wolak, and seminar participants at Brown s Race and Inequality Seminar, the WBCA Colloquium, the Stanford Institute for Theoretical Economics, the Freeman Spogli Institute, the 21st BREAD Conference, the IGC-ISI Development Conference, University of Virginia, Tufts University, and SAIS Johns Hopkins. Funding from International Growth Centre (IGC) under CPR-INC-HUM-2012-CPP is acknowledged and greatly appreciated. Brown University, ssn@brown.edu University of Virginia, ssekhri@virginia.edu 1

2 1 Introduction The efficacy of public service delivery is undermined by large scale leakages in developing countries. Quality of service delivery may be compromised by elite capture (Bardhan and Mookherjee (2000)) or self-enrichment by authorities (Reinikka and Svensson (2004), Niehaus and Sukhtankar (2013)). As a result, services often do not reach intended beneficiaries. Accordingly, finding ways of improving service delivery is a first order research goal. In this paper, we examine the role of social networks in improving public service delivery. We illustrate that social networks provide informal monitoring and enforcement services which improve outcomes by curbing leakage. We also asses whether improving monitoring and enforcement through informal means can yield quantitatively important welfare gains. We study the distribution of subsidized goods in India under the Targeted Public Distribution System (TPDS) by a network of Fair Price (FP) Shops and explore whether the historically disadvantaged Scheduled Castes (SCs) 1 have a lower take-up of the program when facing non-sc shopkeepers. India is a pertinent setting for examining the diversion of resources and consequences of this diversion for vulnerable populations. The incidence of malnourishment among children in India is very high. India recently accounted for approximately 40% of malnourished children in the world (von Braun, Ruel, and Gulati (2008)). As one of the government programs aimed at addressing malnutrition, the TPDS distributes grains and other goods through over 450,000 FP Shops at below market prices. The subsidy for grain is primarily targeted toward below-poverty-line (BPL) households, where BPL status is assigned by local elected officials. The scale of beneficiaries is massive, with 65.2 million BPL households entitled to benefits from the TPDS (GoI (2011)). The financial scale of the program is equally impressive, with the food subsidy accounting for over Rs. 582 billion in FY (approximately $12.6 billion at December 2009 exchange rates), or 1.3% of GDP (GoI (2011)). However, take-up rates vary greatly geographically and are quite low in many states, leaving the promise of an entitlement to food unfulfilled. 2 In part, this is because of massive leakage from the delivery system. Recent estimates for place the fraction of grain diverted at an astounding 43.9% across India, which is actually an improvement over previous years (Khera (2011); PEO (2005)). 3 Thus, examining take-up of different castes under TPDS presents 1 The poverty rate among SCs in was 37%, as compared to a population-wide rate of 28% (Ahluwalia (2011)). 2 These differences arise due to both differential take-up by households and intended supply. 3 The leakage comes in two forms, both of which are facilitated by the absence of effective monitoring and enforcement. The first involves inclusion and exclusion errors in the allocation of BPL cards by local officials. (see 2

3 a unique opportunity to understand the role of informal monitoring and enforcement in improving public service delivery. Our analysis progresses in four steps. First, we use data from the state of Uttar Pradesh to examine whether SC households are disproportionately hurt by a breakdown in local service delivery. Our key empirical result appears in Figure 1. For each of three PDS goods sold in the state grain, sugar, and kerosene we graph the proportion of households that purchase the good in the prior month. The top panel refers to SC households, and the bottom panel refers to non-sc households. It is clear that the take-up of grain is higher among SCs when they face an SC shopkeeper, increasing from about 49% with a non-sc shopkeeper to about 76% with a SC shopkeeper. By contrast, we do not see the same pattern for non-sc households, who actually have lower take-up when facing a SC shopkeeper. SCs differential benefit from SC shopkeepers is positive but smaller for sugar, and less than 4 percentage points for kerosene. Of course, these raw differences may be contaminated by confounding factors. Therefore, to allay concerns about omitted variables bias and more formally quantify the reduced-form impact of shopkeeper caste on household take-up, we use a large array of specifications that control for socio-economic and political characteristics that may co-vary with household caste and shopkeeper caste. We gain further confidence that omitted variables are not affecting our results from the fact that, at the time of our survey in Uttar Pradesh, village council seats and additional positions including the FP shopkeeper were reserved for SCs on a rotational basis based on SC population under the provisions of the 72nd Amendment to the Constitution of India, and this resulted in potentially exogenous variation in the caste of the shopkeepers. 4 Second, we develop a model that explains why the caste of the shopkeeper and the commodity type matter for SC households. We posit that informal social networks may serve as a conduit of information and lower the cost of monitoring and enforcement. Only those with a BPL card are entitled to the grain subsidy. By contrast, all households are entitled to the sugar and kerosene subsidy. Thus information may flow more easily for these goods and collective punishment of a Niehaus, Atanassova, Bertrand, and Mullainathan (2013) for details). The second form of leakage comes from the fact that shopkeepers and agents in other portions of the supply chain have a strong incentive to sell goods on the black market (a specific example from Delhi is discussed in (Parivartan (2004)). Because shopkeepers have privileged access to the supply chain, it is difficult for citizens or elected officials to verify shopkeepers claims of receiving only limited supplies for distribution. 4 This reservation policy has been used as an identification strategy in many recent studies. For example, Chattopadhyay and Duflo (2004) have used the 1992 constitutional amendment creating the reservation system for women to examine the influence of female leaders on village-level outcomes. 3

4 shopkeeper may be forthcoming. Therefore, these goods will have a lower initial cost of monitoring and enforcement than grain will. We hypothesize that when SCs face SC shopkeepers (or non-scs face non-sc shopkeepers), it lowers the cost of monitoring and enforcement. Moreover, the value of this effect matters most when the cost of monitoring and enforcement is high to begin with. Thus, the empirical patterns across different goods could emerge because of the different costs of monitoring and enforcement across goods and castes. We provide additional evidence supporting this interpretation, including data we recently collected in the same villages covered in our original sample. Our reduced-form empirical analysis shows that our findings do not support a series of potential alternative explanations, most notably taste-based discrimination against SCs and elite capture. Third, we estimate our model and quantify the welfare gains to SCs and non-scs of improving monitoring, as envisioned by the National Food Security Bill that was enacted into law on September 10, 2013 (NAC (2011)). 5 Specifically, we examine the welfare gains of moving from the higher monitoring costs of facing a different caste shopkeeper to the lower monitoring costs of facing an own caste shopkeeper. We find that the average welfare gains are quantitatively important, approximately one-fifth the size of the average grain subsidy. The gains are higher for SCs than non-scs. Finally, we use the model to assess the welfare impacts of alternative policies. Specifically, we examine the short-term welfare impact of increasing the effective subsidy rate as proposed by the National Food Security Bill. We find that, at the baseline level of monitoring and enforcement, a more generous subsidy rate leads to lower welfare gains than the cost of the policy. One reason for this is that the increased generosity of the program stimulates greater black marketing of goods and perversely lowers take-up for a fraction of beneficiaries. Improving monitoring through the caste mechanisms significantly increases the welfare gains from this policy, though the average welfare gain is still lower than the cost of the program. To the best of our knowledge, our study presents the first careful cost-benefit analysis of the food subsidy component of the National Food Security Bill. Our paper contributes to four strands of literature. The first strand explores the role of caste in transmitting information and enabling enforcement of agreements. Fisman, Paravasini, and Vig (2011) demonstrate that caste allows bank officials to better screen loan applicants, thereby reducing 5 This bill aims to provide subsidized food grains to around 67 percent of India s population of 1.6 billion people and will cost around 1.1 percent of Indian GDP. 4

5 defaults. Munshi and Rosenzweig (2006) examine the implications of caste-based occupational networks that channel members of the same caste into particular jobs. And in the context of local government, Munshi and Rosenzweig (2010) find evidence that caste can serve as a disciplinary mechanism for politicians. Our work complements the previous papers by showing that not only the caste identity of the agents in a transaction matters, but the type and scale of the item being transacted does as well; moreover, our structural estimates allow us to quantify the welfare effects that result from differential access to the government subsidy in question. Second, an important line of literature provides empirical evidence for the possibility that few, if any, nutritional gains may result from programs that subsidize staple goods. Recently, Jensen and Miller (2008) and Jensen and Miller (2011) have stressed demand-side reasons for this in different settings. Kochar (2005) and Tarozzi (2005) examine the nutritional impact of PDS, specifically. Both studies find only limited effects of PDS program changes. Our framework provides an explanation for limited impacts of the PDS that focuses on supply-side incentives and households need to find monitoring and enforcement mechanisms that can overcome shopkeepers incentives to sell goods on the black market. Third, our results expand the literature on improving public service delivery more generally. Electoral mechanisms involving either formal reservation of political seats or political strength can ensure that economically disadvantaged households enjoy public benefits (Pande (2003), Foster and Rosenzweig (2004)). National-level monitoring and enforcement and simple technological solutions, coupled with strictly enforced changes in incentive structures, have also been effective (Olken (2007), Duflo, Hanna, and Ryan (2012)). However, the success of top-down approaches depends on local monitoring, as implementing the technology or enforcing the incentive structures may be too expensive and logistically difficult for national governments. By contrast, we posit that informal, local monitoring and enforcement can help deter leakages. Fourth, recent studies have found mixed success of community participation in increasing local accountability. While Bjorkman and Svensson (2009) demonstrate that community-level disciplinary actions encouraged by NGOs improved health care in Uganda, Banerjee, Banerji, Duflo, Glennerster, and Khemani (2010) show that participatory education interventions in the state of Uttar Pradesh, predicated purely on improving monitoring among beneficiaries, had no appreciable impact on child educational outcomes. Our study contributes to the understanding of how informal monitoring and enforcement mechanisms operate, what the resulting benefit is, and what considerations determine the size of this benefit. We show that, in the absence of strong formal 5

6 checks and balances, the economically vulnerable may rely on informal social networks to strengthen accountability of public service delivery agents and ensure access to public services. The rest of the paper is organized as follows. In Section 2, we briefly discuss background on the TPDS. In Section 3, we describe the data and provide summary statistics. In Section 4, we conduct empirical analysis of PDS take-up. In Section 5, we develop a model of household demand and shopkeeper supply of PDS goods, test the implications of our model and structurally estimate it. The structural estimates are then used to understand the welfare implications of informal monitoring and enforcement. In Section 6, we use our structural estimates to examine policy counterfactuals. In Section 7, we explore alternative explanations for our empirical results. Finally, Section 8 concludes. 2 The Public Distribution System The Public Distribution System (PDS) is the government of India s flagship program to combat hunger. Through the network of Fair Price Shops run by government licensed shop-keepers, the program delivers subsidized grain (wheat and rice), sugar and kerosene to entitled households. In rural areas, each household is designated to a FP Shop (also called PDS shop) in or near its village, and it is entitled to purchase a certain amount of each good at a below market price. While all households were entitled to purchase sugar and kerosene at the time of our survey data in 1997, only below-poverty-line (BPL) households were entitled to purchase grain at the steeply discounted rates. Households receive their BPL designation through a BPL Census process that occurs every five years and involves significant involvement by local-level officials. Appendix A provides further details about the PDS program and concerns about its effectiveness. The program design, in combination with extremely low commissions paid to shopkeepers for selling goods properly and limited formal monitoring and enforcement, is conducive to the black marketing of goods by shopkeepers and others in the supply chain. Investigations by government authorities have uncovered large amounts of leakage in the system across India (see PEO (2005)). Prior work suggests that up to one-third of grains are diverted from the legal PDS channels into the black market (Mane (2006)). Recent work suggests that, while there have been substantial improvements in some states, nationally the rate of leakage is still estimated to be 43.9% (Khera (2011)). While it is often difficult to find evidence for leakage and theft, the PDS requires record keeping that can be used to reveal substantial irregularities. For example, according to a recent report involving shops in Delhi, a number of systematic discrepancies have been found in these 6

7 records that suggest shopkeepers siphon off goods from these shops (Parivartan (2004); details appear in Appendix A). Leakage could occur anywhere in the supply chain. However, we focus on leakage at the level of the Fair Price Shops. In these areas, the most common form of leakage involves the use of fake stock-outs. That is, an entitled household arrives at the shop and requests its goods. The shopkeeper refuses to make a sale, claiming that he did not receive enough stocks from the state and is already sold out of what he did receive. It is difficult for the household to determine whether it has actually been cheated because the shopkeeper s claim is difficult to verify. Basic patterns in our data are consistent with this sort of illicit behavior. 3 Data and Descriptive Statistics 3.1 Data We use the Uttar Pradesh-Bihar Survey of Living Conditions (SLC) collected by the World Bank. These data consist of household surveys and village surveys undertaken between December 1997 and March 1998, shortly after the initiation of the targeting in the TPDS. The SLC covers 120 villages in UP and Bihar, with 15 or 30 households sampled in each village. We restrict our analysis to the UP portion of the data. The primary analysis is restricted to Hindu households, for whom caste serves as a vital form of kinship network. Moreover, we focus on BPL households who are eligible for a subsidy on all goods distributed at the FP Shops at the time. These goods include grains, sugar, and kerosene. Our main sample comprises of 229 households in 37 villages. 6 The household and village surveys provide a rich set of information. The household data contain basic demographic characteristics of the household members, including caste, age and education. The data include extensive information on income, food and non-food expenditures, and farming and non-farming assets. The data also contain information on whether the household is included in the village BPL list, the take-up of each of the four PDS goods (rice, wheat, kerosene, and sugar), prices paid at PDS shops, the distance to PDS shops, knowledge of entitlements, and crude self-reported measures of quality relative to market quality. 7 6 Some villages do not have a PDS shop. We do not include these in the main sample. 7 The process of handing out cards was in progress during the time of the survey. Around three-quarters of households in Uttar Pradesh who were on the list already had their card at the time of the survey.kriesel and Zaidi (1999) conducted detailed qualitative work in Uttar Pradesh villages at the time and they say In all villages we visited, the fact that a household s name was included on the BPL list entitled it to make purchases from the PDS 7

8 From the village-level data, we use information on caste composition, total number of households, the location of the PDS shop, the caste of the village council president (Pradhan), and the caste of the PDS shopkeeper. In villages without a PDS shop, households must acquire PDS goods from a neighboring village s shop. Unfortunately, we do not have any information on the PDS shopkeeper or the PDS shop used by residents in such villages. 8 No villages have more than one PDS shop in our sample. 3.2 Selection into Sample Naturally, the restrictions mentioned above might lead to concerns about sample selection. We discuss our results relaxing these restrictions with alternative samples in Section 4.2, but we discuss the reasons for our sample choices here. We restrict the sample to Uttar Pradesh because at the time of the survey, Bihar was lagging in implementation of the TPDS. While 72.5 percent of the households identified as BPL were allotted BPL identification cards by the time of the survey in Uttar Pradesh, only 47 percent had been allotted such cards in Bihar. As mentioned before, in Uttar Pradesh, field validations of the process of determining BPL status and thus eligibility were carried out, while this was not done in Bihar. Additionally, reservations for the Pradhan (head of the village council) and the shop keeper had not been implemented completely in Bihar. These reservations provide a useful source of identification in our main results. Moreover, the data do not provide the caste of the Pradhan for Bihar villages, which is an important control variable to address politically-driven explanations for our empirical results. For these reasons, we chose to restrict the sample to Uttar Pradesh, where such elections had been held, the program was implemented more widely, and data on caste of the Pradhan are available. Nevertheless, below we show that the basic specification of our main results (removing controls for which we do not have data in Bihar) are not sensitive to using the full sample. We do not include Muslim households or households in villages with Muslim shopkeepers in the main analysis. We do this because caste-based informal networks are not crucial for Muslims, as caste is a feature of the Hindu religion. In a previous version of this paper, we had extensively shop, regardless of whether or not it had actually received the card. Therefore, we treat all households on the BPL list as BPL households. Importantly, being on the BPL list may not mean the same thing in Bihar, since the card process was lagging in Bihar (only 47 percent had a card) and Kriesel and Zaidi (1999) did not conduct qualitative verification in Bihar villages. 8 If every village in a district were sampled, we could use the names and location of villages to infer information about the PDS shop frequented by every household in villages without a shop. However, due to the random sampling of villages, this is not necessarily possible in our case. 8

9 shown that our results are not sensitive to including Muslims for every specification. In the current version, the complete results with Muslims are not shown for the sake of brevity, but are available on request. Below we do show, however, that our basic results are unchanged if we include these households. The third sample restriction is more important. We focus only on households who have BPL status because APL households are not eligible for grain subsidy and consistent with the rules of the program, take-up of grain is close to zero among APL households. One concern might be that cultural proximity might affect the attribution of BPL status itself, causing a sample selection problem (Alatas, Banerjee, Hanna, Olken, and Tobias (2012)). In a later section on alternative explanations, we show there is no evidence that SC households who gain the BPL designation are poorer in villages with SC shopkeepers than in villages with non-sc shopkeepers. Moreover, supplementary evidence is not consistent with the idea that shopkeepers themselves are screening out BPL households that are not truly poor. Nevertheless, our basic results are qualitatively the same when we include both APL and BPL households in specifications; that is, SC households see a statistically significant and important gain in grain take-up with SC shopkeepers relative to non-sc shopkeepers, but not for sugar and kerosene. These results are reported in Appendix Table 5 and discussed in Section Summary Statistics Summary statistics for the SLC are reported in Tables 1, 2, and Appendix Tables A1 and A2. Table 1 provides summary statistics for basic household characteristics used in the analysis. Approximately 60% of the sample consists of SC households, while there are very few Upper or Middle Caste (UMC) BPL cardholders. In results not shown here, we verify the results of Kriesel and Zaidi (1999), who show that even after conditioning on BPL criteria, SCs are much more likely to obtain BPL cards than other households. Almost 70% of the sample has a PDS shop in their village, and of those responding, only 42% have the shop in their tola (neighborhood). On average, households live approximately 21 minutes away from their PDS shop. The largest share of households face a Backward Caste (BC) shopkeeper, but about 25% faces a SC shopkeeper. A quarter of households face a shopkeeper of their own (broad) caste category. 9 In Appendix Table A1, we report the prices, quantities, and quality of PDS goods purchased. 9 We define caste categories as Scheduled Castes (SC), Backward Castes (BC), and Upper and Middle Castes (UMC). 9

10 Four points are worth highlighting. First, grain purchases (and, to a lesser extent, sugar purchases) are less common than kerosene purchases. Second, grain and sugar are often deemed worse than market quality. 10 Third, there is very little variation in prices charged, regardless of good. 11 Fourth, we also observe that the vast majority of households purchase the entire entitled quantities of 3 kg of rice and 7 kg of wheat. 12 Appendix Table A2 compares the characteristics of villages with non-sc shopkeepers, those with SC shopkeepers, and those where there is no PDS shop (so that information on the shopkeeper is not available from the village survey). The third column of the table shows the P-value for a test of equality between a covariate s mean in non-sc shopkeeper villages and SC shopkeeper villages (columns 1 and 2). The fifth column shows the P-value for a test of equality between a covariate s mean in non-sc shopkeeper villages and villages without shopkeeper information (columns 1 and 4). Villages with SC shopkeepers tend to have a higher proportion of SCs in the village, are more likely to have a SC Pradhan, and are much more likely to have the largest share of households be SC. However, these differences are imprecisely estimated. Villages with SC shopkeepers do have more BPL card holders, and this difference is statistically significant at the 1% level. In the empirical analysis, we focus on how outcomes differ when SC households face SC shopkeepers, as opposed to considering how outcomes differ by sub-caste (jaati). There are two reasons for this. First, in most villages there is only one dominant sub-caste from any single caste residing in a village. In Uttar Pradesh, there are only five villages where the first and second dominant sub-castes by population share are from the same caste. Given the limited heterogeneity within caste in each village, the social network is effectively proxied by our caste variables. Second, although we have information on household sub-caste, we do not have information on the sub-caste of shopkeepers. Table 2 examines how SCs differ depending on whether they face a SC shopkeeper or a non-sc shopkeeper. In the first four rows, there is some indication that SCs facing SC shopkeepers are poorer than those facing non-sc shopkeepers, but none of these differences are statistically signifi- 10 Quality is a self-reported measure taking three values: 1, better than market quality; 2, same as market quality; 3, worse than market quality. Since kerosene is not legally sold on the open market, it is not feasible to compare to market quality. 11 While the standard deviations for quantities are higher, in the case of grain this is driven by extreme values. 12 Of those making PDS rice purchases, 91.3% of households purchase 3 kg. Of those making wheat purchases, 89.6% of households purchase 7 kg. Of the 173 households purchasing either rice or wheat, 121 (70%) purchase 10 kg of total grain, with an additional 22 households purchasing 7 kg of wheat and not purchasing any rice and 6 households purchasing 3 kg of rice and no wheat. 10

11 cant. Moreover, those in SC shopkeeper villages actually appear to be more educated on average. The large differences emerge when comparing measures of SC economic or political dominance. In villages with SC shopkeepers, SC households are more likely to be the plurality caste, more likely to be the same caste as the village Pradhan, and more likely to have a SC sub-caste be the village s largest sub-caste and land dominant sub-caste. The remaining rows of the table speak to the households relationship to the PDS. The data contain information on the largest caste (by population) in the tola (habitation) where the PDS shop is located. With a SC shopkeeper, SC households are much more likely to have SCs dominate the tola of the PDS shop. We can also examine distance to the PDS shop. In the data, households indicate whether they are closer than 0.5 km from the PDS shop, between 0.5 km and 3 km from the shop, between 3 km and 10 km from the shop, or more than 10 km from the shop. Very few households are more than 3 km from their PDS shop, so we aggregate the latter three categories and create a dummy variable for being less than 0.5 km from the shop. The table shows that SCs are much more likely to be close to the shop when facing a SC shopkeeper. These facts begin to suggest that having a SC shopkeeper may confer monitoring advantages to SCs relative to other castes. The last three rows show that take-up of grain, sugar, and kerosene are all significantly higher with SC shopkeepers. 3.4 Follow-up Survey In order to better understand the role of caste in the take-up of the PDS, we conducted a 2012 follow-up survey of 296 households in the same set of SLC villages that are in our main sample. The 2012 sample was roughly divided in half between SC households and non-sc households. The objective was to illustrate how the PDS operated within a village and what role caste played in facilitating access to PDS goods. We obtained the village rosters and BPL list from the village council offices. 13 From the BPL list, 4 SC and 4 non-sc households were randomly chosen from each village to be interviewed in-person. All respondents who were contacted were found. They were asked to show the ration card to the enumerators to verify their BPL status. We administered a quantitative survey with 15 questions to these respondents. Some of these questions required binary answers, while others required ordering the importance of response options or choosing from multiple response 13 The BPL lists were obtained from the council office but are also posted in Hindi on the website of the Ministry for Rural Development, Government of Uttar Pradesh. 11

12 options. 14 All non-binary close-ended questions had an option to specify other, which respondents could then explain in an open-ended way. We also interviewed 2 APL households using a short qualitative survey. While BPL households were administered close-ended questions, the APL households were asked open-ended questions. We coded the responses of open-ended questions to investigate the correspondence between responses of the BPL and APL households. Using these survey data, we demonstrate in Appendix Figure 1 that the tendency for friends, relatives, and fellow caste members to obtain better treatment from the shopkeeper has been persistent over time. We see that friends and relatives of the shopkeeper and members of the shopkeeper s caste are reported to receive better treatment by more than half the sample in Below, we use this follow-up survey to inform our interpretation of the empirical results from the SLC. 4 Empirical Analysis of Take-Up and Caste 4.1 Estimation Strategy We saw in Table 2 that take-up of all goods is higher for SCs when they face a SC shopkeeper. In this section, we conduct our benchmark analysis and use the emergent results to distinguish between competing models that could explain the influence of shopkeeper caste on SCs take-up of PDS goods. In comparing SCs who face SC shopkeepers versus those who face non-sc shopkeepers, the main empirical concern for interpretation is that the caste of the shopkeeper may be endogenously chosen. For instance, in villages where SCs are politically pivotal, leaders may direct both lucrative bureaucratic posts like PDS shopkeeper and available PDS goods toward SCs. 15 A policy change in India provides a source of variation in the caste of the shop keeper that is plausibly exogenous conditional on certain co-variates. In 1992, the 72nd amendment of India s constitution reserved positions of village council head (Pradhan) for SCs on a rotating basis, based on population shares of SCs in the local administrative area. In Uttar Pradesh, villages with reservations for the head of the village council also had reservations for PDS shopkeeping. Uttar Pradesh had village council elections in 1995, before the SLC data were collected; therefore, the village council heads at the time of the SLC survey were selected under this new policy. Since PDS shopkeepers are not elected, but are instead appointed by district officials after a solicitation process, the existing PDS shopkeepers in 1995 were not removed from SC-reserved seats; instead, 14 The survey instrument is available from the authors on request. 15 We explore these and other mechanisms in the section on alternative channels below. 12

13 any new openings in these areas were filled by SC shopkeepers. If we had data on the tenure of the PDS shopkeepers, we could directly test for this, but the SLC unfortunately does not have such data. However, these rules suggest that a shopkeeper was more likely to be SC in 1997 if village positions happened to be reserved for SCs. We could potentially use the determinants of reservation of council seats for SCs as an instrument to address the endogeneity problem. In Table 2 above, we do observe that the plurality of SC population in the village, which may be strongly correlated with the population shares of SCs in the local administrative area, is positively related with having an SC shopkeeper. Also, across villages, having an SC Pradhan is positively correlated with having a SC shopkeeper. However, we do not use plurality of SC population as an instrument because in our application, the population share of SCs can have an independent effect on take-up. Villages with a larger presence of SCs may face lower monitoring costs since BPL status is higher among SCs, or may face more difficulties in enforcement if higher level government officials are less responsive to SC-dominated areas. Queues may be longer in these villages, since SCs tend to have greater access to BPL cards and may have greater reliance on public services regardless. Thus, we control for village share of SC population in our specifications for take-up rather than using an instrumental variable approach. 16 When providing an explanation for our results that relies on informal monitoring and enforcement in the following section, our identifying assumption is that conditional on population share of SCs, the village s largest and land dominant sub-caste, and whether the household and Pradhan are of the same caste, the variation in the interaction between household caste and PDS shopkeeper caste is plausibly exogenous to unobservable components of households preferences and costs of PDS access. We also explicitly address several endogeneity concerns in the discussion below, and systematically cast doubt on twelve alternative explanations for our results in a later section. We formalize the results in Figure 1 with the following empirical specifications for take-up of grain, sugar, and kerosene. Y iv = α + βs v + γb iv + ηs v B iv + ɛ iv (1) Where Y iv is an index for take-up of the subsidized goods (grain, sugar, or kerosene). S v is an indicator which takes value 1 if the shopkeeper of the village belongs to a Scheduled Caste, B iv is an indicator which equals 1 if the buyer is a Scheduled Caste household, and ɛ iv is the error term. The coefficient η on the interaction term is the parameter of interest. We refer to this as 16 As Anderson (2011) shows, the SC population in these villages has been historically very stable. Reservations did not lead to movement of SC populations. 13

14 the caste pairing effect or caste interaction effect. In a linear probability model, η would be a difference-in-difference (DID) coefficient. Below, we show estimates from OLS estimation of this equation. However, since take-up is a binary variable, we focus on using probit models to estimate Equation (1). We prefer the probit specifications because average take-up of kerosene is quite high, so that predicted values may be more likely to fall outside the unit interval when continuous covariates are included as independent variables or an exhaustive set of dummies is not included. 17 Because of our non-linear specification, η cannot be interpreted as the usual difference-in-difference parameter. Therefore, we explicitly calculate and report the following difference for every specification we run: = [P r(y = 1 S v = 1, B iv = 1) P r(y = 1 S v = 0, B iv = 1)] [P r(y = 1 S v = 1, B iv = 0) P r(y = 1 S v = 0, B iv = 0)] (2) where the first term equals the take-up of SC buyers when they face an SC shopkeeper minus their take-up when they face a non-sc shopkeeper, and the second term equals take-up of non-sc buyers with a SC shopkeeper minus their take-up with a non-sc shopkeeper. That is, indicates SC households differential gain from having a SC shopkeeper; in a linear specification, would be equivalent to η. Table 3 presents the results. Each column represents a specification with additional controls and methodological changes that account for a large class of omitted variables bias concerns. Each panel contains empirical models for take-up of the three goods (grain, sugar, and kerosene). Within each panel, the first two columns are standard probit models, the third column is a random effects probit model with random effects at the village level, and the fourth column is a fixed effects logit model with fixed effects at the SC Shopkeeper-SC Pradhan-Survey Stratum level. For the probit models, we include the implied estimate of at the bottom of each column. 18 We highlight the suitability and limitations of each approach in the discussion of our 17 Of course, in the simple model including only the SC shopkeeper dummy, the SC household dummy, and the interaction, the probit model and the linear probability model yield identical predicted values. 18 Note that, because of the use of non-linear models, the other covariates matter for the size of. In specifications including other covariates besides the household and shopkeeper caste, we evaluate for two sets of values. Set 1 is evaluated for a relatively disadvantaged profile of characteristics: a household with zero land, no pucca housing, primary schooling, more than 0.5 km from the PDS shop, not the same caste as the Pradhan, not the plurality caste, not a land dominant caste, and with the average share of SCs in the village. Set 2 is instead evaluated at means of these variables. We report using set 1 because a significant number of our control variables are binary, and thus 14

15 results. 4.2 Take-up Results and Robustness Tests Columns 1, 5, and 9 of Table 3 show the results for the most basic probit specification with only a SC buyer household dummy, the SC shopkeeper dummy, and the interaction of the two (the caste pairing effect). Standard errors are clustered at the village level. For both grains and sugar, the caste pairing effect is positive and significantly different from zero; for kerosene, it is not. This implies that for grain and sugar, SC buyers benefit more than non-scs from having a SC shopkeeper. The bottom rows show the estimates of, which suggest that the SC buyers differential benefit from having a SC shopkeeper is very large for grain, approximately 54 percentage points. Several endogeneity concerns might arise in interpreting this as the effect of having a SC shopkeeper on take-up. First, omitted household characteristics can be confounding the results. One possibility is that in areas where SCs are the poorest and consequently most dependent on the PDS for food they mobilize to ensure that there is a SC shopkeeper. In other words, wealth determines food take-up and is correlated with the interaction of interest. Second, caste-based residential segregation results in pairings of SC households with SC shopkeepers such that the distance to the PDS shop is reduced. Third, pairings of SC households with SC shopkeepers may be occurring in places where SCs are more politically powerful. In these examples, it is plausible that kerosene take-up is not positively associated with the omitted variable. In the first case, kerosene may be a luxury good relative to food items. In the second case, PDS entitlements of grain and sugar may be more difficult to carry than kerosene. And in the third case, SCs may have a differentially higher demand for food products because of their greater poverty. Columns 2, 6, and 10 of Table 3 address such concerns. In addition to the caste variables, these specifications control for a rich set of characteristics that might be correlated with the interaction term. These include land holdings, type of housing structure, education dummies, a dummy variable indicating whether the household is less than 0.5 km from the PDS shop, a dummy variable for whether the household s caste forms a plurality in the village, a dummy for whether the household is the same caste as the Pradhan, a dummy for whether the household s caste has a sub-caste that has the highest land ownership in the village, and the share of households in the village that are SC. For grain, we continue to find a positive interaction effect that is significant at the 1% level. As indicated in the bottom rows, remains similar. For sugar, falls and it is now significant their mean values are not very informative. 15

16 at 11 and 10 percent, respectively at the two sets of values of other covariates. For kerosene, the caste pairing effect is negative and insignificant at conventional levels of significance. Another concern might be that take-up is strongly correlated among households within villages, perhaps because all households within a village see the same queues at the PDS shop, or they see the same market prices for goods. If villages with many SCs and an SC shopkeeper just happen to have higher market prices for grain in our sample (even though this may not be the case in the population), we could be erroneously attributing the effect to the caste pairing. We address this type of concern in Columns 3, 7, and 11 of Table 3. These columns use random effects probit models of take-up with random effects at the village level that account for village-specific heterogeneity. The results for grain are robust to controlling for village-level random effects. For grain, we see that the coefficient estimate for the interaction term remains positive and significant, though falls slightly. For sugar, the caste interaction effect is not significantly different from zero, and the same is true for kerosene. The limitation of random effects models is that the random effect is assumed to be uncorrelated with the included covariates. Another class of concerns is that SC shopkeepers are chosen for unobservable reasons involving village leaders political desire to ensure households access to PDS grain. Since BPL cards are directed toward SC households as discussed earlier, it may be natural to direct the bureaucratic post serving BPL cardholders to a SC. To address this, we use fixed effects logit models with fixed effects at the SC shopkeeper-sc Pradhan-Survey stratum level. The results in Columns 4, 8, and 12 confirm the earlier results: The caste pairing effect is significantly different from zero at the 1% level for grain, but the effect is not statistically distinguishable from zero at conventional levels for sugar or kerosene. An additional concern might be that an SC shopkeeper treats poor or ill-informed households differently and that SCs tend to be poor and ill-informed. We do control for land owned, pucca housing, and schooling in our regressions to ensure the socio-economic characteristics of SC buyers are not confounding the results. But to allay this concern more directly, we interact SC shopkeeper with land owned and with knowledge of entitlement in addition to the controls specified above. In results not shown here, the caste pairing effect for grain is 1.75 and significant at 1 percent when we control for the interaction of SC shopkeeper and land. The coefficient for the interaction of SC shopkeeper and land is with a standard error of Thus it is not statistically different than 0. The caste pairing effect for grain when we control for knowledge of entitlement and the interaction of SC shopkeeper and knowledge of entitlement is 2.04 and statistically significant at 1 16

17 percent. The coefficient for knowledge of entitlement for grain is positive 1.6 and highly statistically significant at 1 percent. But the coefficient for the interaction of SC shopkeeper and knowledge of grain entitlement is 1.08 and is statistically insignificant with a standard error of In order to show that our non-linear models are not the source of our results, we report the estimates using a linear probability model in Appendix Table A3. The specifications include the same covariates as in Columns 2, 6, and 10 of Table 3 (A linear probability model with a constant term and the three covariates in Columns 1, 5, and 9 yields identical predicted values and hence estimates of as the probit model). Panel A shows the estimates with standard errors clustered at the village level for grains, sugar, and kerosene respectively. Panel B reports estimates from the same specifications, except with standard errors calculated using the wild cluster bootstrap procedure to account for the small number of villages in the sample. We find that SCs are almost 52 percentage points more likely to buy grain when facing an SC shopkeeper, and this effect is statistically significant at the 1 percent level. In contrast, there is no statistically discernible difference for sugar or kerosene. Panel B shows very similar results. These results are in line with the estimates of presented in Table 3. Another concern might be that the purchases of grain, sugar, and kerosene are not being treated jointly above. Some households may purchase these goods simultaneously. 19 Appendix Table A4 addresses this by estimating a model of take-up of all three goods that allows for the unobserved errors for each good to be correlated with one another. The basic patterns continue to hold up. For grain, the caste interaction effect remains positive and significant (still at the 1% level), and the magnitude of is in the same range as before. The bottom line of the table presents the P-values from the test that for each good is different from for grain. The marginal effect for grain is significantly different from that for kerosene, though not from that of sugar. Nevertheless, the most important point is that even when we allow for the errors to be correlated across goods, the caste interaction effects are not significantly different from zero for sugar or kerosene. In Appendix Table A5, we show that our basic estimation approach using random effects probit produces qualitatively similar results regardless of what sample we choose. We first pool both Uttar Pradesh and Bihar and use the full sample including Muslims and non-bpls. We use all covariates specified in Table 3 (Columns 3, 7, and 11), except household and Pradhan being from the same caste, as we do not have data on caste of the Pradhan in Bihar. Row 1 shows the estimates for the 19 This may not always be the case, however. In one village observed by the authors in Uttar Pradesh, the PDS shop sold kerosene on a particular day without selling other goods that day. 17

18 three goods. We find a positive and significant caste pairing effect for grain but not for sugar and kerosene. We then restrict to only BPL households in row 2 and exclude Muslims in row 3. Row 4 shows results from pooling Bihar and Uttar Pradesh, including non-bpl but excluding Muslims. Then we repeat this exercise for only Uttar Pradesh and report the results in rows 5 through As mentioned before, qualitatively our results are the same. We do see a significant caste pairing effect for grain but not for the other goods, regardless of the sample restrictions we impose. To summarize, a robust finding is that the caste interaction effect is positive and statistically significant for grain, but not for sugar and kerosene. These results beg the questions: Why is the take-up of grains larger for SC buyers when they face SC shopkeepers relative to non-sc shopkeepers, whereas having an SC shopkeeper does not affect take-up of grain in the same way for other castes? And why is the same not true for sugar and kerosene? In the subsequent section, we provide detailed evidence to show that caste plays out as a monitoring and enforcement device. 5 Role of Monitoring and Enforcement The patterns observed in our data are consistent with a hypothesis that caste serves as an informal monitoring and enforcement mechanism. In the first sub-section, we formalize this hypothesis with a simple theoretical model that focuses on the role of caste in informal monitoring. In the second sub-section, we derive basic comparative statics that we can test. In the third sub-section, we test these implications of the model with reduced-form specifications that parallel the specifications from the previous section. In the fourth sub-section, we discuss the results of a survey we conducted in the SLC villages recently, in order to gain direct insight into the monitoring and enforcement mechanism. Finally, in the last sub-section, we parameterize and estimate the model and use the structural estimation results to assess the welfare consequences of improving monitoring through this informal mechanism. 5.1 Model Overview A monitoring and enforcement mechanism can explain the fact that SC households benefit differentially from SC shopkeepers for grain, but not for sugar and kerosene. Consider the actionable evidence that a household would need to conclude that the shopkeeper had sold its entitlement 20 In rows 5 through 8, we also control for the household being the same caste as the pradhan. 18