WPS5040 Policy Research Working Paper 5040 Caste and Punishment The Legacy of Caste Culture in Norm Enforcement Karla Hoff Mayuresh Kshetramade Ernst Fehr The World Bank Development Research Group Macroeconomics and Growth Team September 2009 Policy Research Working Paper 5040 Abstract Well-functioning groups enforce social norms that low-caste individuals exhibit a much lower willingness to restrain opportunism, but the social structure of a society punish norm violations that hurt members of their own may encourage or inhibit norm enforcement. This paper caste, suggesting a cultural difference across caste status in studies how the exogenous assignment to different the concern for members of one's own community. The positions in an extreme social hierarchy--the caste system lower willingness to punish may inhibit the low caste's --affects individuals' willingness to punish violations of ability to sustain collective action and so may contribute a cooperation norm. Although the analysis controls for to its economic vulnerability. individual wealth, education, and political participation, This paper--a product of the Growth and the Macroeconomics Team, Development Research Group--is part of a larger effort in the group to investigate the behavioral foundations of economics. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at khoff@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 CASTE AND PUNISHMENT The Legacy of Caste Culture in Norm Enforcement Karla Hoff Mayuresh Kshetramade Ernst Fehr The World Bank Affinnova Inc, Waltham, MA University of Zurich khoff@worldbank.org mayurvk@yahoo.com efehr@iew.uzh.ch JEL Classification : D02, D64 Key Words : Norms, informal sanctions, third-party punishment, altruism, endogenous preferences, social exclusion, caste Acknowledgements. We thank Robert Boyd, Robert Keohane, Vijayendra Rao, and Rob Willer for valuable discussions and comments. We are indebted to Sonal Vats for superb research assistance at every stage of this project, and to Siddharth Aryan, Manoj Gupta, Mukta Joshi, Shiv Mishra, Priyanka Pandey, and Dinesh Tiwari for their help in implementing the experiment. We benefited from participants' comments at seminars at Cornell,Georgetown University Law School, George Washington University, Harvard, the Indian Statistical Institute,the Institute for Advanced Studies, the MacArthur Foundation Research Network on Inequality, Princeton,the University of Texas at Dallas, and the World Bank. We acknowledge research support from the World Bank (for Hoff) and from the Research Priority Program on the "Foundations of Human Social Behavior--Altruism vs. Egoism" at the University of Zurich (for Fehr). Hoff thanks the Princeton University Center for Health and Wellbeing for its hospitality in 2008-9. "When a Dalit [formerly an Untouchable] argued with an upper caste farmer..., the upper caste villagers attacked 80 Dalit families in retaliation. When the same Dalit man then went to the police to report the incident, a social boycott was imposed on all of the Dalits from [his village]; they were thrown out of their village and denied every opportunity to earn their livelihood." Tejeshwi Pratima, "Dalits thrown out of their village for raising their voice against discrimination," June 29, 2006 1 Every society requires restraints on opportunism. The conventional simplifying assumption in economics is that government provides these and that individuals obey them because it is in their self-interest to obey the law. More recently, scholars have questioned the usefulness of this convention, arguing that social norms are a key source of restraints and that without them even formal rules would be unenforceable.2 Underneath the level of behavior that most of economics is concerned with are the social capabilities to enforce restraints on opportunism, underpinning even modern societies with well-developed legal institutions. Economic historians such as Douglass North (1990) and Joel Mokyr (2008) have emphasized the role of appropriate social norms for the enforcement of contracts and property rights and, thus, for economic development and growth.3 As social norms are part of cultural traditions that may inhibit or enhance trade and production by affecting honesty, trust and trustworthiness, and cooperation, these arguments bring cultural factors into the focus of economists' attention (e.g. Gintis 1972, 2008; Bowles 1998; Bowles and Gintis 1998; Guiso, Sapienza and Zingales 2004, 2006; Banerjee and Iyer 2005; Fisman and Miguel 2006; Herrmann, Thöni and Gächter 2008; and Tabellini 2008). Social norms are enforced by informal sanctions that are often imposed by those who obey the norm even though sanctioning is costly and yields no material benefits to the punisher (Fehr and Fischbacher 2004). Such punishment can have very powerful effects because in the absence of sanctioning opportunities, norms tend to erode and eventually may collapse. Thus, the 1 Cited in New York University Center for Human Rights and Global Justice and Human Rights Watch, 2007, p. 60. (hereafter, NYU). The incident took place in June 2006, 2 See, e.g., Hayek (1973, ch. 2), Greif (1993), Platteau (1994), Weingast (1997), Lindbeck, Nyborg and Weibull (1999), Basu (2000), and Ostrom (2000). 3 Mokyr argues that the unusual strength of Britain's social norms against dishonest behavior in business helped elevate Britain to the leading position it took in the Industrial Revolution. In 18th century Britain, "opportunistic behavior was made sufficiently taboo that only in a few cases was it necessary to use the formal institutions to punish deviants" (p. 532). In this secure environment, unprecedented levels of cooperation occurred between individuals with commercial acumen and those with technical skill: "Boulton found his Watt, Clegg his Murdoch, Marshall his Murray, and Cooke his Wheatstone" (p. 535). willingness to sanction plays a key role in the maintenance of norms (Ostrom 2000, Fehr and Gächter 2000, Sobel 2005, Carpenter and Matthews 2006). In laboratory markets and real world markets, the efficiency and volume of trade can critically depend on the altruistic willingness to punish those who do not keep their formal or informal obligations (Fehr, Gächter and Kirchsteiger 1997; Logan and Shah 2009).4 Recent evidence indicates that reciprocating cooperation is a prevalent behavioral pattern and is considered as the normatively right action across many cultures (Henrich et al. 2001; Herrmann, Thöni and Gächter 2008 and Kocher et al. 2008), but still little is known about how the social structure of a society affects people's willingness to sanction norm violations. In our research we were able to compare the willingness to sanction violations of a cooperation norm across groups in which there is an exogenous source of variation in social position, namely, caste. In principle, the existence of, and affiliation to, extreme social positions could affect informal norm enforcement in a variety of ways. An extreme social hierarchy could lead to conflict and hostility between groups that would cause dysfunctional punishment in inter- group interactions. Alternatively, a history of repression could lead members of the repressed group to punish selectively--tolerating norm violations committed by members of the dominant group but punishing norm violations committed by others. A third possibility is that a history of repression undermines--within the historically repressed group--the capability to be prepared psychologically to punish others for norm violations. The caste system can serve as an excellent setting for studying the effect on norm enforcement of affiliation to the top or bottom of an extreme social hierarchy, for three reasons. 1. Stark historical differences in social and economic rights. For thousands of years the caste system was associated with extreme inequalities between the groups at the top of the hierarchy (hereafter, the "high castes") and those at the bottom (hereafter, the "low 4 Logan and Shah measure in an illegal market ­ the male sex market ­ the economic impact of altruistic punishment, demonstrating its power to permit honest providers to signal their type. The enforcement in this market is decentralized and altruistic in the sense that it yields no material benefit to the individual. One enraged client for 10 years policed the web to warn others that he had been robbed by a specific provider (personal communication of Trevon Logan). 2 castes").5 The high castes had high ritual status, basic freedoms, and the power to extract forced labor from low castes. The low castes were denied any basic social rights, could not sell their labor and goods in markets, were barred from schools, temples, and courthouses and were relentlessly stigmatized through the practice of Untouchability (Shah et al. 2006). The low castes provided forced labor to high caste individuals on demand, and the demand for cheap labor was a factor in the persistence of Untouchability (Bayley 1999). 2. Heritability of caste and rigidities of caste ranking at the extreme ends of the caste hierarchy. An individual's specific caste is determined by the accident of birth, and individual mobility across castes is basically not possible in an individual's lifetime. In rural India, castes are endogamous. There are strong norms against cross-caste marriages. Marriages between high and low caste persons are particularly harshly punished and sometimes lead to "public lynching of couples or their relatives, murder (of the bride, groom or their relatives), rape, public beatings and other sanctions" (NYU 2007, p. 11). Although caste boundaries and caste rank can change over long time periods, the status of the specific high-ranked castes (Brahmin and Thakur) and low-ranked castes (Chamar and Pasi) from which we draw our experimental subjects goes back millennia (Gupta 2000). The absence of across-caste mobility for these groups rules out selection bias and enables us to study the impact of caste status on individuals' willingness to punish norm violations. The high castes in our sample thus constitute a meaningful control group for the low castes in our sample. 3. Overlapping wealth distributions. Despite the vast difference in social status between high and low castes, there is today considerable individual variation in wealth, consumption, and education within the high and low castes.6 We have not only poor low-caste subjects, but also many poor high-caste subjects in our sample. By controlling for individual differences in wealth and education, we can rule out that differences in punishment behavior across castes are caused by those individual differences. 5 Once called Untouchables, low caste individuals are today called Dalits, a non-pejorative term that connotes oppressed or ground down. They made up 16 percent of the Indian population in the 2001 Census. They are not ethnically distinct from other castes (Gupta 2000). 6 With the abolition of the Zamindari (landlord) system in Uttar Pradesh in 1952, many Dalit agricultural laborers acquired ownership rights of the lands that they had been cultivating. 3 We implemented a third party punishment experiment with subjects from high and low castes. We use this experiment to study the strength of informal punishment of norm violators. The essence of such an experiment is that one player, whom we call player B, can obey or violate a social norm in an interaction with another player, player A. Then the third party, player C, learns what B did and has the opportunity to sanction him. In order to elicit C's preferences to sanction the norm violation, the punishment is costly for C. In our experiment, the social norm is that B reciprocates a cooperative choice of A in a sequential social dilemma game. Previous research across different cultures suggests that a substantial proportion of third parties are willing to punish the failure to reciprocate cooperation (Fehr and Fischbacher 2004; Carpenter and Matthews 2006; Herrmann, Thöni and Gächter 2008). We report here data from our study of 205 triples composed of adult males in over 150 villages in one of the poorest states of India, Uttar Pradesh. We implemented four treatments, called HHH, HLH, LLL, and LHL. The first letter in each treatment indicates the caste status ­ high (H) or low (L) ­ of player A, the potential injured party; the second letter indicates the caste status of player B, the potential norm violator; and the third letter indicates the caste status of player C, the third party punisher. The subjects in the experiment were informed about the caste status of the matched players in an unobtrusive way (see Section I.C). This design enables us to test for double standards in punishment and for differences between high and low caste individuals in the willingness to punish. Since our experiment takes place against a backdrop of a decades-long effort at political mobilization of the low castes, it might be that hostility and conflict between high and low castes would lead to particularly high punishment levels in triples that include both high and low caste players. For example, in an HLH treatment, in which the victim of a norm violation is high caste, the norm violator is low caste, and the third party is high caste, hostility might induce the third party to punish the norm violator much more harshly than if all three players had come from a high caste. That is, punishment in HLH would be higher than in HHH. We call this the caste conflict hypothesis, which also predicts that punishment in an LHL treatment, in which the norm violator comes from the high caste and the punisher from the low caste, would be greater than in an LLL treatment. Alternatively, a history of fierce retaliation by the high caste when low-caste individuals refused to submit to the prevailing social 4 hierarchy,7 might lead low caste individuals to tolerate norm violations by high caste individuals. The prediction that low caste individuals refrain from punishing high caste individuals in an LHL treatment is what we call the caste submission hypothesis. The third hypothesis that our design enables us to examine is that, controlling for the caste status of the norm violator, high caste compared to low caste individuals punish more severely norm violations that hurt members of their community. Sen (2000) and Rao and Walton (2004) argue that inequality of agency is a consequence of social exclusion. To adapt Sen's language, the aspect of agency that we investigate is the capability to punish those who violate social norms. Repression and social exclusion might induce fatalism, undermining the self- confidence and the motivation required to punish violators of cooperation norms. We call this the caste culture hypothesis. Our results unambiguously refute the caste conflict and the caste submission hypotheses. Instead we find, in line with the caste culture hypothesis, that low caste compared to high caste individuals punish norm violations less often and less severely. This result is robust to controls for wealth, education, and participation in village government. In fact, the effect of individual differences in wealth is very small, always insignificant, and sometimes not in the expected direction. Further, we show that the low castes' lower propensity to punish has nothing to do with differences in the underlying social norm. We measure the underlying social norm in our experiment by player B's expectations about punishment. Regardless of caste status, the vast majority of subjects in the role of player B expected that they would not be punished for cooperation in the social dilemma game but would receive high punishment for defection ­ a clear indication that cooperation was considered the normatively right thing to do in this game. Thus, low and high castes have the same cooperation norm in the social dilemma game. 7 Even in contemporary rural India events like the following are still reported: "When a Dalit ... refused to sell bidis [hand-rolled cigarettes] on credit to the nephew of an upper caste village chief, the upper caste family retaliated by forcibly piercing his nostril, drawing a string through his nose, parading him around the village, and tying him to a cattle post" (cited in NYU, 2007, p. 60, from Indian Express (Bombay), April 28, 1998). 5 Why then do low caste individuals punish norm violations less severely than high caste individuals in our experiment? A factor known to influence altruistic third party punishment is in-group affiliation (Bernard, Fehr and Fischbacher, 2006; Goette, Huffman, and Meier 2006; Chen and Li 2009 show a similar result for second party punishment). We examined in a second experiment the hypothesis that high caste compared to low caste individuals show more concern for the members of their specific caste.8 In the experiment discussed above, we always ensured that the potential victim of the norm violation (player A) and the potential punisher (player C) belonged to the same specific caste, while the potential norm violator (player B) belonged to a different specific caste. In this setting, if a player C who is Brahmin has a strong concern for the victim of the norm violation from his own specific caste, he will be more willing to punish than would a player C who is Chamar, who may not care much for the victims from his own specific caste. To test this hypothesis, we conducted a second experiment in which we ensured that this kind of in-group concern cannot affect the punishment pattern. If the willingness to punish is affected by differential in-group concerns between high and low castes, then the caste gap in punishment should be reduced in the second experiment. In fact, we observe that the caste gap in punishment vanishes in the second experiment: if in-group concern for the victim cannot play a role, the high castes punish at about the same level as the low castes. This result provides support to the hypothesis that high caste compared to low caste individuals exhibit a stronger concern for others from their own specific caste. This result is interesting in light of the evidence that the high castes seem superior in organizing collective action (Kohli 1987, Drze and Gazdar 1997), often for the purpose of sustaining the caste status and power differences with collective force. The quote at the beginning of the introduction nicely illustrates this point. When a single Dalit argued with a high caste farmer over discrimination, the high caste villagers collectively attacked 80 Dalit families and imposed a collective boycott on all Dalits from the village. In contrast, the Dalits were unable to respond to this with collective action. In view of the importance of social preferences for the achievement of group solidarity, it is plausible that the higher degree of in-group concern 8 In the context of this argument, it is very important to recognize that an individual belongs to a specific caste, such as the Brahmins, the Thakurs, the Chamars or the Pasis. The specific castes constitute a large part of an individual's social network and social life. Therefore, the specific castes represent the relevant in-group. For example, for a Brahmin the relevant in-group are the other Brahmins, and for a Pasi the other Pasis constitute the relevant in-group. 6 for the members of their specific caste helps the high castes to sustain collective action to support norms that benefit them collectively.9 The difference between high and low castes in the willingness to sanction violations of cooperation norms, and the relative absence in the low castes of in-group concerns for the members of their caste, may have far-reaching implications. Experimental evidence (Fehr and Gächter 2000; Carpenter 2006; Herrmann, Gächter and Thöni 2008) as well as field evidence (Miguel and Gugerty 2005) indicates the importance of voluntary informal punishment for the provision of public goods. Thus, if high castes are better able to sanction free riders, they are in a better position to organize collective action and produce public goods compared to the low castes. This advantage may be one reason why Untouchability continues to be practiced in almost 80 percent of Indian villages despite the constitutional abolition of Untouchability (Shah et al. 2006). Another consequence may be that the high castes are better able to enforce contracts and ensure their property rights, which advantages them with respect to trading opportunities and production incentives. Taken together, these consequences may contribute to the persistence of poverty and inequalities between high and low castes. The lower willingness in the low caste communities to sanction violations of cooperation norms may also mean that conventional policies for improving governance through decentralization and community-based development projects may need to be reconsidered. These policies assume a broad-based capacity for norm enforcement in society, whereas in fact that capacity may be unevenly distributed across groups.10 Finally, if future 9 A British official in 1947 wrote that attempts by low caste individuals to exercise the right to use public wells, a right granted to them by law under British rule, was commonly met with "social boycotts" ­ collective punishments imposed by high castes who might refuse to trade with an entire low caste community, or might destroy their crops and dwellings. The official concluded that "No legislative or administrative action can restore to the depressed class people the right to use public wells" (cited in Galanter, 1972, p. 234). Social boycotts have also occurred in recent years. In a social boycott in 1998, high caste individuals who gave employment to low caste individuals were fined by the village council (Human Rights Watch, 1999, p. 30). 10 The lower willingness in the low caste communities to sanction violations of cooperation norms appears to reflect patterns of interaction in everyday life in Uttar Pradesh that have contributed to the failure of public sector interventions to help disadvantaged groups. The most detailed study is of the village of Palanpur, but the general picture that emerges from that study seems representative of the state. After reviewing all public services in the village of Palanpur, Drze, Lanjouw and Sharma (1998) conclude that predatory actors had derailed essentially every public initiative to help disadvantaged groups. "With few exceptions, Jatabs [the main low caste in Palanpur] have remained outside the scope of constructive government intervention" (p. 220). In one particularly egregious case, the intended low caste beneficiaries became "victims of merciless extortion" by the high caste managers of a cooperative lending society (p. 220). Because of examples like this one, Drze, Lanjouw and Sharma emphasize that "state 7 research generalizes our findings to other settings with extreme historical inequalities, it would suggest a new mechanism by which elites can sustain their superiority after the legal basis for it is removed. The historical legacy of the denial of basic social and economic rights may be to diminish the repressed groups' capability to organize collective action and enforce contributions to public goods, which perpetuates the vulnerability and exploitability of these groups. In the next section of our paper (Section I), we present the experimental design. In Section II, we discuss our hypotheses regarding the role of caste affiliation in punishment behavior and in Section III, we report our results. Section IV reports the follow-up experiment to investigate the effect of caste status on in-group affiliation. Section V summarizes our findings and relates them to the question: How does an elite resist reforms after it loses control over the political institutions? Whereas earlier work looked for an explanation in a framework of purely selfish agents, our findings suggest that the affiliation to different social positions in an extreme hierarchy affects people's social preferences: members of the repressed group are made more self-interested than members of the elite. One consequence may be a difference between the elite and the historically repressed group in the ability to sustain cooperative behavior among themselves and thereby to coerce others. I. The Experimental Design We are interested in how the affiliation to castes with different social status affects the willingness to punish violations of a conditional cooperation norm. For this purpose, we developed a simple experimental game in which certain behaviors are likely to constitute a normative obligation that, if violated, may be punished by an "impartial" observer. In the following, we describe first the game and then the different treatment conditions. I.A. The Experimental Game Figure 1 depicts the game between three individuals, A, B, and C. Each individual plays the game in his home village with anonymous players from other villages. Players A and B interact in a sequential exchange game. Each player has an endowment of 50 rupees, which is a initiatives do not operate in a social vacuum"; they require social pressures to sustain them. However, the required social pressures from the low castes were absent during the fifty-year period spanned by the five surveys of Palanpur. 8 considerable amount of money compared to the daily wage of an unskilled agricultural worker of about 50-75 rupees. Player A has to choose one of two actions: he can "send" his total endowment to B, in which case the experimenter triples its value so that B has altogether 200 rupees;11 or he can opt out, in which case A keeps his endowment and the game ends. If A sends his endowment to B, then B has to make a binary choice: to keep everything for himself or to "send" 100 rupees back to A. We allowed players A and B only binary choices because we wanted "keeping the money" by player B to be an unambiguous norm violation. We expected a widely shared understanding that if A sends money to B, then B should send money back to A; i.e. that there is a social norm of conditional cooperation. Player C is an uninvolved outside party who can punish B at a cost to himself. His endowment is 100 rupees. For each two-rupee coin that player C spends on punishment, Player B is docked a 10-rupee note. We asked C to make a choice for the case where B keeps all the money, and also for the case where B sends money back to A. Player C makes this choice before he learns B's decision. He indicates his choice by moving coins on a game board in private. Figure 1. Sequential Exchange Game with Third Party Punishment pc = 10 A has 100 C B has 100 - 10pc C has 100 ­ 2pc pc = 0 B pd = 20 A has 0 B has 200 ­ 10pd C has 100 ­ 2pd A C pd = 0 A has 50 B has 50 C has 100 Note: pc and pd , respectively, indicate the number of two-rupee coins that Player C spends to punish Player B conditional on B's cooperation or defection (norm violation). 11 200 rupees is of the order of one week's per capita gross state product in Uttar Pradesh (based on the official estimate of per capita annual gross state product in 2004-05, the latest year for which such data are available). 9 The conventional assumption that individuals are purely self-interested implies a unique subgame perfect Nash equilibrium in which the third party never punishes, the second party keeps all the money if it is offered to him, and so the first party opts out. Thus the equilibrium would be (Opt out, Keep all the money, Don't punish). In this outcome, A avoids being a sucker and does not enter into a relationship with B, and C keeps clear of punishing C. Players A and B each then earn 50 rupees, whereas if they cooperate they would each earn 100 rupees. The sequential exchange problem is akin to A having a good that B values more ­ that is the tripling of the money. By entering into an exchange, A gives B the opportunity to keep all the resources ­ a risk that typically arises in exchanges with separation between the quid and the quo over time or space and imperfect contractibility. In developing countries a large proportion of all exchange is probably characterized by such features. If B sends back nothing and C punishes him, the punishment by C mimics a disinterested third party's sanctioning a norm violator ­ for example, reproaching or bullying him or gossiping about him, which entails some cost or risk to the punisher but a larger cost to the individual punished. The deterrent to defection by a purely self-interested player is 10( pd pc ) , where pd denotes what Player C spends to punish defection and pc denotes his spending to punish cooperation. We measure pd and pc in units of two-rupee coins. The game instructions avoid value-laden terms. For example, we never use the words "cooperate," "defect," and "punish." Instead we use neutral terms such as "send the money," "keep the money," and "impose a loss." I.B. Treatment Conditions Before we describe the treatments in detail, it is important to distinguish between two meanings of the term "caste." First, belonging to a caste means that the individual belongs to a specific endogamous social grouping consisting of thousands of families ­ such as belonging to the caste of Brahmins or the caste of Chamars. Each such social grouping is associated with a traditional set of occupations and culture and, generally, individuals of the same caste are clustered together in neighborhoods. Networks organized around the specific endogamous castes provide mutual insurance, which contributes to the very low rate of migration from villages in 10 India (Munshi and Rosenzweig 2007). The second meaning of belonging to a caste is that the individual is assigned the social status of the caste. For example, both Brahmins and Thakurs are castes at the high end of the caste hierarchy, but they constitute nevertheless clearly distinct social groups. Chamars and Pasis are castes in the lowest stratum of the caste hierarchy, but constitute clearly distinct endogamous groups. In the following, we use the terms "caste divide," "high caste," and "low caste" to indicate the status dimension of caste affiliation. We reserve the term "caste" without a modifier to mean the specific endogamous social group. Because we were interested in the effect of affiliation to an extreme position in a social hierarchy on the willingness to punish norm violations altruistically, we implemented four treatments that varied the composition in a triple of individuals with high (H) and low (L) caste status. In the "single status" treatment, all three players are either high caste or low caste (treatments HHH and LLL). In the "mixed status" treatment, there is a deep status divide among players: in HLH only player B is low caste, and in LHL only player B is high caste. The number of triples by treatment was 62 in HHH, 61 in HLH, 41 in LLL, and 41 in LHL. Because caste affiliation is associated with a certain status and with membership in a specific social group, we developed an experimental design that controls for the in-group-out- group relationships among the players. If we had not done this, we would confound the effect of the caste divide between players with the effect of in-group favoritism or out-group hostility. To see this, notice that in the single status treatments, player B could, in principle, be a member of the same specific caste as A and C, whereas in the mixed status treatments, player B would necessarily be of a different specific caste than A and C. In this case, it would be impossible to know whether any treatment difference between, say, HHH and HLH, was due to the caste divide or to in-group affiliation or to both. To avoid this confound, we formed triples in which player B ­ the potential norm violator ­ was always from a different specific caste than players A and C. Figure 2 illustrates this by giving examples of the triples we used. 11 Figure 2. Examples of Interacting Players A B C Potential Potential norm Third party victim violator punisher HHH, e.g. Brahmin Thakur Brahmin HLH, e.g. Brahmin Chamar Brahmin LLL, e.g. Chamar Pasi Chamar LHL, e.g. Chamar Brahmin Chamar We drew our subjects from Brahmins and Thakurs, who are high caste, and Chamars and Pasis, who are low caste.12 Figure 2 shows, for example, that in both the HHH condition and the HLH condition, the potential victim of a norm violation (player A) and the third party punisher (player C) are from the same specific caste ­ the Brahmins ­ while the potential norm violator (player B) comes from a different specific caste ­ the Thakurs or the Pasi ­ and is, therefore, an out-group member relative to A and C. Figure 2 shows that this feature ­ A and C belong to the same specific caste and B belongs to a different specific caste ­ holds across all four treatment conditions.13 I.C. Procedures We recruited adult male subjects for each role (A, B, and C) from three non-overlapping sets of villages in central Uttar Pradesh: players A came from 61 villages, B from 53 villages, and C from 48 villages, randomly chosen from the hundreds of villages within 2.5 hours' drive from the town of Unnao. Informants in each village told us the neighborhoods in which the different castes in the village lived. In public places in a village, the recruiters asked individuals if they were interested in participating in an experiment about decision making that would last two hours and in which they would earn some money. We generally recruited five subjects for a single treatment (and never more than six) from a given village, none of them brothers. The 12 The number of observations where C is Brahmin is 63, where C is Thakur is 60, where C is Chamar is 39, and where C is Pasi is 43. 13 We varied this feature in a follow-up experiment, described in Section IV, which was designed to examine the psychological mechanisms behind the observed caste differences in the willingness to punish norm violators. 12 average age of players was 35 years (with standard deviation 8.0) for high caste players and 34 years (with standard deviation 7.6) for low caste players. Overall, the age range was between 24 and 50 years. No subject participated in more than one treatment. To ensure that subjects understood the instructions, the rules of the game were explained to them at great length. A few subjects who did not pass a basic test of comprehension did not go on to participate in the game. Individual sessions were held inside a Qualis car. Each subject made his decision in private while the experimenter waited outside the car. After Player C had indicated his decision by moving coins on a game board, the experimenter reentered the car, informed C of A's and B's decisions, and paid C. Implementing this experiment in rural India raised two ethical concerns. First, the players should never learn the identity of those with whom they interact in the game. Second, our concern with caste relationships, a politically fraught issue, should not be salient. To address these concerns, we recruited subjects for each role (A, B, and C, respectively) from three distant sets of villages. We carefully thought about how to communicate the caste affiliation of the participants to the players. One possibility was to use names, since names generally convey both a person's individual identity and his specific caste. The use of names was thus an unobtrusive way of indicating caste. We checked explicitly in a pre-experiment that individuals were generally willing to reveal their last names to our team of recruiters, and we also verified the ability of individuals to recognize caste membership from names. During the individual sessions with each subject, the experimenter conveyed information about his partners by saying, e.g. in the case of Player C, "You are playing the game with two other people. You are person C. [NAME], who is from another village, is person A; and [NAME], who is from another village, is person B." The advantage of using names is that we can convey information about caste but still maintain effective anonymity among the players, since thousands of people with the same last name live in the state and each of the players in a given triple came from a different and distant village. The experimenter emphasized that no player would ever know the villages to which the two players with whom he was matched belonged. The experimenter also never used the word "caste" in his interactions with the subjects before or during the game. Only at the end of the individual session did the experimenter ask the subject his caste (and he would do that indirectly) in order to verify that we had not made a mistake about his caste membership. In post-play 13 interviews, we also asked questions about a subject's beliefs about the other players' actions, about the reasons for his own actions, and about his wealth and other demographic variables. II. Hypotheses Since our experiment takes place against a backdrop of a decades-long effort at political mobilization of the low castes, it might be that hostility and conflict between high and low castes would lead to particularly high punishment levels in triples that include both high and low caste players. The caste conflict hypothesis predicts that mean punishment in an HLH treatment, in which the norm violator comes from a low caste and the punisher from a high caste, would be higher than in an HHH treatment. Similarly, in an LHL treatment, in which the norm violator comes from a high caste and the punisher from a low caste, the punishment would be higher than in an LLL treatment. We express this by punHLH > punHHH and punLHL > punLLL, where punXYZ denotes the mean punishment for defection in treatment XYZ.14 Alternatively, a history of fierce retaliation by the high caste when low caste individuals refused to submit to the prevailing social hierarchy may have led low caste individuals to accept their fate. As caste hierarchy is part of everyday life, it is not implausible that low caste individuals refrain from punishing high caste individuals in an LHL treatment. We call this the caste submission hypothesis. This hypothesis predicts, contra the caste conflict hypothesis, that low caste third parties will punish high caste defectors less harshly than low caste defectors: punLHL < punLLL. Whereas the preceding two hypotheses examine differences in punishment as the caste status of the norm violator varies, our third hypothesis examines differences in punishment as the caste status of the punisher varies. This hypothesis examines whether the social preferences that underlie altruistic punishment are similar across high and low castes, or whether the affiliation to different social position in an extreme social hierarchy has shaped social preferences. The caste culture hypothesis predicts that among the members of the historically repressed group, there is a 14 Throughout, we measure punishment in units of two-rupee coins spent. Recalling that player B loses one 10-rupee note for each coin spent by player C, 2·pun is the rupee spending on punishment and 10·pun is the rupee punishment imposed. 14 lesser willingness to punish those who violate the cooperation norm of the group. Controlling for individual differences in education and wealth, this hypothesis implies punH > punL where punH denotes the mean punishment of a norm violation when the punisher is high caste (i.e. punishment in HHH and HLH), while punL denotes the mean punishment when the punisher is low caste (i.e. punishment in LLL and LHL). III. Results Our data give us two measures of punishment for defection: pd, which measures the absolute punishment of defectors (i.e. norm violators) in terms of the number of two-rupee coins paid for punishment; and pd - pc, which measures the extent to which defectors are more strongly punished than cooperators. In the following, we denote pd - pc as relative punishment. Although our data also provide us with a measure of the sanctions imposed on cooperators, pc, in this paper we focus on the role of caste affiliation in the sanctioning of defectors.15 Before we present the results, we present evidence that cooperation by player B is indeed viewed as the normatively right thing to do. III.1. Testing the existence of a conditional cooperation norm We measure whether there is a social norm of conditional cooperation by player B's beliefs about punishment. After player B had chosen his action, we asked him how much punishment he expected in the case of cooperation and how much punishment he expected in the case of defection.16 If there is a widely shared belief in the existence of a normative obligation to reciprocate cooperation, the beliefs of player B should reflect this obligation. In Figure 3 we show player B's beliefs about punishment. The figure shows little variation across treatments and between castes. On average, in each treatment and for each caste, player B expected that C would spend nearly 10 coins to punish B if he defected, and would spend almost nothing on punishment 15 We find non-negligible levels of punishment for cooperation but no significant differences across treatments (see Fehr, Hoff and Kshetramade 2008). Although the first result may seem surprising, recent evidence, including evidence from cross-cultural studies (Herrmann, Thöni and Gächter 2008; Nikiforakis 2008 and Cinyabuguma, Page and Putterman 2006) indicates that punishment of cooperators is frequent. 16 We asked these questions in a neutral language, i.e. we did not us the terms cooperation, defection, reciprocation or punishment, as we discussed in Section I. 15 if B cooperated. Thus, in each treatment the null hypothesis of equal expected punishment across cooperation and defection can be unambiguously rejected (Mann-Whitney test, P < 0.01 in all cases), indicating a conditional cooperation norm in each treatment. Moreover, the differences between expected punishment of defection and cooperation are almost identical across treatments (Kruskall-Wallis test, P = 0.994), suggesting that the same conditional cooperation norm applies across treatments and castes. Figure 3. Player B's Beliefs about the Absolute Punishment of Cooperation and Defection 12 Punishment of cooperation Punishment of defection 10 8 2-rupee coins 6 4 2 0 HHH HLH LLL LHL III.2. Testing the caste conflict and caste submission hypotheses The caste conflict hypothesis suggests that we should observe a difference between mean punishments in the single status treatments and the mixed status treatments. With regard to punishment by high caste members, this means that we should observe a higher level in HLH than in HHH. Contra this hypothesis, Figure 4 shows that mean relative punishment in HLH is not higher than in HHH; and the difference between conditions is not significant according to a Mann-Whitney test (P = 0.18). This result is reinforced if we examine absolute punishment levels. Mean punishment in HLH (6.80 two-rupee coins) is again not higher than in HHH (8.45 two-rupee coins), and the difference is not significant (Mann-Whitney test, P = 0.09). The caste conflict hypothesis also predicts a higher level of punishment in mixed status treatments by low caste members, that is, punishment should be higher in LHL than in LLL. 16 Contra this hypothesis, mean relative punishment in LHL is not significantly different than in LLL (Mann-Whitney test, P = 0.85). The same picture emerges if we examine absolute punishment: the mean in LHL (6.15) is not significantly different than in LLL (5.37; Mann- Whitney test, P = 0.91). Figure 4. Relative Punishment of Norm Violators 6 Mean relative punishment 4.82 5 3.97 4 3 2.59 2.07 2 1 0 HHH HLH LLL LHL High Low Caste status of third party punisher The caste submission hypothesis predicts that punishment in LLL is higher than in LHL. Figure 4 shows that this hypothesis is not borne out, and the results discussed above show that punishment in LLL does not significantly differ from that in LHL. In order to check the robustness of these results, we conducted OLS and Tobit regressions that included a number of factors that capture social and economic differences among individuals in the role of player C.17 For each individual, we control for land owned, education, and house quality. In rural Uttar Pradesh, where we conducted our experiments, land owned and house quality ­ whether an individual lives in a mud house or a pure brick house ­ are major indicators of wealth; in Appendix Table A1, we show that these variables are also important predictors of 17 Tobit analyses support all results reported here. Tobit deals with censored data better than OLS. The censoring problem may play a role in the case of absolute punishment for defection, which ranges from 0 to 20, but much less so in the case of relative punishment for defection, which ranges from -10 to 20. In supporting materials (not for publication), Tables A1-A3 present the Tobit regressions for absolute punishment. 17 per capita consumption. In addition, the regressions control for whether the individual cultivates his own land, which might affect his attitudes towards the norm that we are investigating, and whether he has participated in village government.18 In regressions (1), (2), (4) and (5) of Table 1, we show the results of OLS regressions that are based on the following model: pun = + (player C is high caste) + (treatment is mixed status) + (player C is high caste treatment is mixed status) + + error where "pun" denotes absolute or relative punishment and Z is a vector of variables measuring individual characteristics of player C. The omitted category in regressions (1), (2), (4) and (5) is a low caste punisher who is in the single status treatment. Thus, the constant measures the punishment level in LLL. The next three coefficients are measures of the caste and treatment effects when we control for individual characteristics: measures the difference between a high and low caste player's punishment in the single status treatments, i.e., the difference between HHH and LLL, and thus, + indicates the punishment level in HHH. The coefficient for the mixed status treatment, , measures the difference between the LLL and LHL treatments, implying that + represents a measure of punishment in LHL. Finally, punishment in the HLH treatment is measured by + + + . With respect to the high caste, the caste conflict hypothesis is that punHLH > punHHH, which implies that + > 0, which we assess with an F-test in Table 1. In all four regressions the F-test indicates that + is not significantly different from zero. For example, in regression models (2) and (5), which control for the above-mentioned socioeconomic characteristics, the P- values for the F-statistics are 0.343 and 0.108, respectively. With respect to the low caste, the caste conflict hypothesis is that punLHL > punLLL, which implies that > 0, while the caste submission hypothesis implies that < 0. In all the regressions, the coefficient for the mixed status treatment () is not significantly different from zero (P > 0.4 in all cases). Thus, after controlling for important socioeconomic characteristics such as land ownership, education, and house quality, we refute the caste conflict or caste submission hypotheses. 18 The regressor "Has political experience" is a dummy variable equal to one if the respondent has ever been a village government chief ("Pradhan"), vice-Pradhan, or member of the village government council. 18 III.3. Testing the caste culture hypothesis The caste culture hypothesis predicts a lower level of punishment in the treatments in which the third party is a low caste member relative to the treatments in which he is a high caste member. Figure 4 provides preliminary support for this hypothesis. From the previous section, we know already that the distribution of punishments in HHH is very similar to that in HLH. Likewise, we know that punishment behaviors in LLL and LHL are similar to each other. Therefore it makes sense to pool HHH and HLH, on the one hand, and LLL and LHL, on the other hand, to examine whether punishment by high caste members is generally higher. We find that mean relative punishment imposed by high caste individuals is roughly 90 percent higher than mean relative punishment by low caste individuals (4.40 compared to 2.33). A similar picture emerges for mean absolute punishment: Members of high castes punish 32 percent more than low caste members (7.63 compared to 5.76). These differences in punishment across castes are significant according to Mann-Whitney tests at P 0.01. To what extent is this systematic difference in the willingness to punish a norm violation a result of differences in wealth and education across high and low castes? We collected information in post-play questions on land ownership, housing wealth, and education, which are all known to be important predictors of per capita consumption.19 In this context it is important to know that although people belonging to high castes are on average wealthier, there is a substantial overlap across castes: many people who belong to a high caste are nevertheless very poor. Likewise, a significant number of low caste people have managed to acquire more wealth than poor high caste people. Therefore, we are in a position to examine the differences in punishment across castes while holding wealth constant. This shows up clearly in our sample of subjects. If we examine the distribution of castes across house quality, we find that among the 81 subjects who live in a mud house, 42% belong to a high caste and 58% belong to a low caste. If we look at those 124 subjects who live in a brick house or a mixed mud and brick house, we 19 We show this in Table A1 in the appendix, where we use data from the 1997-98 Survey of Living Conditions in Uttar Pradesh. The table indicates that land ownership, housing wealth, and education have a large and significant impact on adult per capita consumption. Together, they explain between 30 and 40 percent of the variation in consumption for both high and low caste individuals. 19 find that 71% belong to a high caste and 29% to a low caste. A similar pattern emerges in the case of land ownership. Among the 95 subjects who own land below the median, 36% belong to the high caste and 64% to the low caste. Among the 110 subjects who own land above the median, 83% belong to the high caste and 17% to the low caste. In Figures 5a and 5b we show the mean punishment of defectors conditional on the caste status, house quality, and land ownership of the punisher. Figure 5a indicates that regardless of whether subjects live in a mud house or a house that is built at least partly with bricks, the high caste subjects punish more on average than the low caste subjects. Moreover, the high caste subjects exhibit very similar mean punishment levels regardless of house quality. The same holds true for low caste subjects: they punish on average less than high caste subjects but display similar mean levels regardless of house quality. Figure 5b indicates a very similar pattern with regard to land ownership. Regardless of whether subjects own land below or above the median, the high caste subjects punish on average more than the low caste subjects. Also, within the high caste, the mean punishment level is similar for those below and above median land ownership and, among the low caste individuals, the mean punishment level is also similar for those below and above the median. Taken together, Figures 5a and 5b suggest that wealth does not have a big effect on punishment. This hypothesis is clearly supported by regressions (3) and (6) of Table 1, which examine whether punishment is correlated with wealth, education and other characteristics of the punishers, and whether the caste gap in punishment is robust to the introduction of these controls. Note that in these regressions, we do not include the mixed status dummy or a term for the interaction between the mixed status dummy and caste. This means that the omitted category in these regressions consists of the treatments LLL and LHL, in which only low caste players can punish. Therefore, the high caste dummy measures the extent to which high caste players are generally more willing to punish defectors. Both regressions confirm that high caste players spend roughly two two-rupee coins more on punishment of defection, a difference that is significant in the case of both relative punishment (P = 0.055) and absolute punishment (P = 0.022). Thus the caste gap in punishment is robust to the introduction of these additional controls. In a dprobit regression (not shown), we find that controlling for other individual 20 characteristics, high caste individuals are 16.6 percent more likely to punish norm violations than low caste individuals (P = 0.023) and that no individual characteristic other than caste has a significant effect on the probability of imposing non-zero punishment. Figure 5a. Punishment by Those Who Live in Mud Houses and Those Who Do Not Punisher is high caste 6 Punisher is low caste 5 4.47 4.37 Mean relative punishment 4 3 2.66 2 1.89 1 0 N = 34 N = 47 N = 89 N = 35 Lives in a mud house Lives in a brick house or a mixed mud and brick house Figure 5b. Punishment by Those Who Own Below-median Land and Those Who Own Above-median Land Punisher is high caste 6 Punisher is low caste 5.05 5 Mean relative punishment 4.12 4 3 2.35 2.24 2 1 0 N = 33 N = 62 N = 90 N = 20 Owns below median land Owns at least median land 21 The regressions in Table 1 also show that land ownership has little effect on punishment and is statistically insignificant; likewise, the effect of house quality is small and insignificant. These results confirm the message conveyed by Figures 5a and 5b. Among the control variables, only completion of secondary school (10 years of schooling) has an effect on punishment that is significant at the 5 percent level in at least some of the model specifications. Controlling for the other factors, completion of secondary school raises absolute and relative punishment by roughly 1.8 two-rupee coins, an increase of about 0.3 standard deviation units.20 We also conducted further tests (not shown here) to examine whether the education variable changes the caste gap. For this purpose we interacted "completion of high school" in regressions like those in (3) and (6) with the high caste dummy. The interaction effect is small and insignificant (P > 0.70 for both absolute and relative punishment), indicating that the completion of secondary school leaves the caste gap unaffected. Similarly, the interaction of political participation with the high caste dummy is not significant and does not change the qualitative results. A final robustness check for the validity of the caste culture hypothesis concerns whether our results hold over subpopulations. Do the results reflect the peculiarities of a single specific caste, rather than a general status difference between high and low caste members? If caste status drives the caste gap, we should observe that the punishment behavior of members of the two high castes in our sample ­ the Brahmins and the Thakurs ­ is each significantly greater than that of the two low castes in our sample ­ the Chamars and the Pasis. We examine this question in Table 2. Here columns (2) and (4) duplicate regressions (3) and (6) of Table 1 except that now we distinguish specific castes rather than caste status. In Table 2, Chamar, a low status caste, is the omitted category. As predicted by the caste culture hypothesis, the coefficients on Brahmin and Thakur, which are high status castes, are significant and positive, and the coefficient on Pasi, which is a low status caste, is not significant. Thus, whatever explains the caste gap in the willingness to punish defectors operates at the level of caste status, not at the level of specific 20 We find that the effect of a subject's education on his level of punishment for defection is highly nonlinear. When we measure education in years, the estimated coefficient on education is not significant, whereas in Tables 1-3, which measures education as a dummy variable that equals one if a person has completed 10 years of schooling, the estimated coefficients are nearly always significant. In India, secondary school ends at tenth grade. Among the 125 subjects without secondary school completion, 45 percent were high caste and 55 percent were low caste; among the 80 subjects who had completed high school, 84 percent were high caste and 16 percent were low caste. 22 individual castes (descent groups). Just as plants in spatially separated deserts have indepen- dently developed similar adaptations that enable them to live in the desert, one can conjecture that different low castes have developed similar attitudes towards altruistic punishment that reflect the common constraints on their social life. The attitudes converge between the two low castes in our sample and diverge from those developed by the two high castes in our sample. IV. Interpretation and Discussion of Results Having established in Section III.1 that the perceived obligation to reciprocate is identical between high and low castes, we can be confident that the generally greater willingness to punish among the high caste members does not reflect different normative standards but rather a higher preference to punish norm violations. What could explain this? As noted above, a factor that influences altruistic third party punishment is in-group affiliation. Bernard, Fehr and Fischbacher (2006) and Goette, Huffman, and Meier (2006) experimentally show in-group bias in punishment in two respects. If the victim of a norm violation is a member of the punisher's in-group, or if the norm violator is a member of the punisher's out-group, then punishment tends to be higher. An intuitive explanation of these results is that third parties empathize more with a victim who is an in-group member compared to an out-group member, and "give a break" to a norm violator who is an in-group member. However, in-group affiliation is also affected by what the group means to an individual. Social identity theorists (Tajfel 1982, Turner et al. 1994) argue that group identification exists in part to provide self-esteem as the individual construes himself as associated with some group that he values. This theory predicts lower levels of identification among members of groups that are lower in prestige. A classic field study is Cialdini et al. (1976), which finds that students from universities with major football teams identify themselves more with their team when it wins than when it loses. Applying social identity theory to our setting, the social stigma of the low castes and their limited means to acquire an alternative group identity (e.g. because historically they lacked a literate class) might induce members of low castes to develop little concern for other (unknown) members of their specific caste. An individual who belongs to the Chamars ­ a caste at the lower end of the status hierarchy ­ may feel ashamed of being a Chamar in view of the inferior status of his group and the humiliation he experiences in daily life, which may in turn 23 induce him to have little concern for Chamars whom he does not know. For low caste individuals, the in-group is the category that serves as his discrediting. In contrast, the high social and ritual status of the high castes may induce high caste individuals to show more in-group concern for unknown members of their specific caste. An individual belonging to the Thakurs, for example, a caste at the high end of the status hierarchy, derives status, prestige and material benefits from being a Thakur. It thus seems more likely that this individual would care more about the welfare of other Thakurs.21 In the treatments that we have discussed, players A and C were always members of the same specific caste, while the potential norm violator ­ player B ­ always came from a different specific caste (recall Figure 2). This feature of the experimental design enabled us to examine the role of caste status while controlling for in-group affiliation. By varying this design, we can now learn whether members of the high castes display different degrees of in-group bias relative to members of low castes. We conducted a second experiment in which players B and C were members of the same specific caste, while player A belonged to a different specific caste. We implemented this condition in the single status treatment and thus denote the new triples by HHH BC and LLLBC . In Figure 6 we show an example. In the example, the specific caste composition for HHH BC is Thakur-Brahmin-Brahmin in the roles of players A, B, and C, respectively, and for LLLBC it is Pasi-Chamar-Chamar as players A, B, and C, respectively. Figure 6. Examples of Interacting Players under the BC Condition A B C Potential Potential norm Third party victim violator punisher HHHBC e.g. Thakur Brahmin Brahmin LLLBC e.g. Pasi Chamar Chamar 21 Recent experimental work shows that reinforcement through selective status awards to those individuals who altruistically contribute to the group enhances the individuals' willingness to contribute to the group in the future (Willer 2009a, 2009b). The low castes historically had few means to engage in the kind of rituals in which selective status awards are bestowed, or to enjoy symbolic public goods (the term is due to Rao 2008) that contribute to a positive collective identity. Prohibitions on public celebrations of low caste marriages and other restrictions on low caste rituals still exist in parts of rural India (Thorat 2002 and Shah et al. 2006). 24 A crucial difference between the BC condition and the previous treatments is that in the BC condition, the potential victim of the norm violation (player A) is no longer a member of the punisher's specific caste. Thus, if the punisher exhibits in-group favoritism towards the victim, punishment will be lower in the BC condition than in the previous treatments. Furthermore, in the BC condition, a norm violator is a member of the punisher's in-group. Therefore, a punisher who tends to give the members of his own specific caste a break when they violate a social norm will punish less in the BC condition. Since the BC condition has players B and C of the same specific caste, it necessarily has players B and C of the same caste status. The only treatments in our main experiment that have players B and C of the same caste status are HHH and LLL. Therefore, to study the effect of the BC condition, we compare the new treatments with our earlier single caste treatments (LLL and HHH). We have 39 triples in HHHBC and 36 triples in LLLBC. Figure 7 shows the results. The caste gap in relative punishment for defection nearly vanishes in the BC condition. While there is a large gap between HHH and LLL in the previous experiment as shown by the first two bars, which are repeated from Figure 4, the gap between HHHBC and LLLBC is small and insignificant (Mann-Whitney test, P > 0.8 for both relative and absolute punishment). The elimination of the caste gap reflects the fact that the high caste members punish significantly less in the BC condition (Mann-Whitney test: P 0.05 for both relative and absolute punishment), while the low caste members do not show a difference in punishment across the LLL and LLLBC conditions (Mann-Whitney test: P = 0.99 for relative punishment and P = 0.356 for absolute punishment).22 22 We observe these differences despite the fact that the perceived social norm is the same across these two conditions. Just as we found in our main experiment (see Figure 3), we also observe in the BC condition that players B expect no or little punishment in the case of cooperation and high punishment in the case of defection. Thus, the difference between these two conditions cannot be attributed to differences in perceived social norms. 25 Figure 7. Relative Punishment When the Victim and Third Party Punisher Are Members of the Same Specific Caste, and When They Are Not 6 4.82 5 Mean relative punishment 4 2.87 3 2.07 2.22 2 1 0 HHH LLL HHHBC LLLBC We further examine the role of in-group favoritism in Table 3, which reports the regression results for the pooled data of the HHH, HHH BC , LLL and LLLBC treatments. In these regressions, the LLL treatment is the omitted category. Therefore the BC-dummy measures the difference between LLL and LLLBC . The table shows that the coefficient on the BC-dummy is small and insignificant, suggesting that there is no in-group favoritism among the low caste members. The difference between the HHH condition and the HHH BC condition is measured by the sum of the coefficients for the BC dummy and the interaction term between the BC-dummy and the high caste dummy. Table 3 shows the F-test for the null hypothesis that the sum of these coefficients is zero. Controlling for the socioeconomic characteristics of the punisher, we can reject this hypothesis with high confidence (P < 0.02 for both relative and absolute punishment). This result is consistent with the hypothesis that high caste members' punishment decisions are driven by in-group favoritism towards their specific caste. Finally, Table 3 provides information about whether there is a significant caste gap in the BC condition. The difference between HHH BC and LLLBC is measured by the sum of the coefficients for the high caste dummy and the interaction term. The F-test for the null hypothesis that the sum of these coefficients is equal to zero shows that the null cannot be rejected (P > 0.6 26 in all four regressions of Table 3). Thus, the punishment behavior of low and high caste individuals is indistinguishable in the BC treatment, suggesting that in-group favoritism among the members of the specific castes with high status plays an important role in explaining the caste gap in our main treatments. V. Concluding Remarks In this study we have shown how individuals' life-long position at the top or bottom of an extreme social hierarchy ­ the Indian caste system ­ affects their willingness to punish norm violations. The Indian caste system is an excellent setting for studying this question because people are born into specific castes and mobility across caste status groups is basically absent, whereas the greater freedoms that low caste individuals have enjoyed in the last 50 years, and the greater levels of landlessness among the high castes, have created a substantial overlap between high and low caste groups with respect to wealth, education, and political participation in village government. This means that we can both rule out self-selection into castes and also compare individuals who differ in caste status but do not differ with respect to our measures of wealth, education and political experience. We are thus able to identify the impact of caste status on individuals' willingness to punish norm violations. We have put forward three plausible hypotheses: the caste conflict hypothesis, the caste submission hypothesis and the caste culture hypothesis. Our findings unambiguously refute the first two and support the third: compared to low caste subjects, high caste subjects have a considerably higher willingness to altruistically punish violations of a cooperation norm. This finding is robust to controls for wealth, education and political participation. Our results also suggest that the differences in the willingness to punish can be attributed to the caste status of individuals: individuals from each specific high caste in our sample ­ the Brahmins and the Thakurs ­ exhibit a significantly greater willingness to punish compared to individuals from each specific low caste, while the differences between the two specific high castes and between the two specific low castes in our sample are negligible. We can also rule out that subjects from the high caste have a different view about the prevailing social norm. Both members of high castes and members of low castes believe that the third party punisher will dock their payoffs by a large 27 amount if they defect, but not otherwise. Our findings therefore suggest that the punishment differences between high and low castes truly reflect a difference in preferences for the punishment of defection. In a further experiment, we isolated the proximate mechanism behind the high castes' stronger willingness to punish. In our main experiment, the injured party always belonged to the punisher's specific caste while the norm violator did not. In the follow-up experiment, the norm violator always belonged to the punisher's specific caste, while the injured party did not. If the punisher is motivated by in-group favoritism ­ taking revenge if the injured party is a member of his own specific caste, or giving a norm violator from his own specific caste a break ­ we should observe less punishment in the second experiment. We observe a substantial reduction in the severity of punishment imposed by high caste individuals but no significant change in the level of punishment imposed by low caste individuals. As a result, the caste gap in punishment vanishes in the second experiment. This suggests that in-group favoritism ­ being more socially minded, but only towards those whom they consider part of their community ­ is an important driving force behind the higher castes' stronger willingness to punish norm violations altruistically. These findings are of particular interest in view of the evidence that the high castes are better able to organize collective action, often for the purpose of maintaining their economic privileges with collective force. Our results relate to a longstanding question in political economy: How does an elite resist reforms after it loses control over the political institutions? Acemoglu and Robinson (2006) propose a model of purely self-interested agents in which each agent within the elite is large enough to internalize the benefits from resistance to reform through lobbying, bribery, intimidation, or violence, whereas each actor in the repressed group is not large enough to do so. In their model, changes in de jure political power (i.e. political power allocated by political institutions) induce offsetting changes in the distribution of de facto political power as each agent in the elite increases the intensity of his collective action, whereas agents in the masses do not. Our results suggest another, complementary mechanism to explain the puzzle of persistence of repressive institutions despite major de jure reform. In line with a large--and rapidly increasing--literature on altruistic norm enforcement, we have provided evidence that the assumption that individuals are purely selfish is inadequate to explain the 28 observed enforcement of norms. The novelty of our findings is to provide evidence that an individual affiliated to the top stratum of an extreme social hierarchy has a substantially greater willingness to altruistically enforce a cooperation norm that helps his group than does an individual affiliated to the bottom stratum of the hierarchy.23 In our experiment, individuals at the bottom of the caste hierarchy adopted an attitude toward norm enforcement that was closer to pure self-interest than did individuals at the top of the caste hierarchy. This result is reminiscent of an older perspective (e.g. Gellner 1994) that stressed that in order to dominate a group thoroughly, the group had to be pulverized and atomized. In this view, many of the restrictions historically imposed on the low castes ­ such as exclusions from public celebrations and bans on marriage ceremonies and other shared rituals ­ make sense because they prevent the low castes from developing positive group identities that promote collective action. In a world in which everybody is completely selfish such restrictions would make little sense, whereas if one takes into account the possibility of altruism towards one's own group, these restrictions may help the high castes to maintain their superior position. If the evidence reported here is replicated across cultures, it would suggest a new mechanism through which a repressive regime has persistent effects even after the legal basis for it is removed and the economic forces that gave rise to it have changed. The regime may have shaped the repressed group's culture in ways that moderate members' willingness to altruistically punish violations of cooperation norms, with effects on the group's ability to enforce informal agreements and to sustain collective action. 23 The example that Acemoglu and Robinson use to illustrate their argument is the persistence of a repressive economy in the southern United States after the Civil War that looked remarkably like that of the antebellum South. However, another factor that may have played a key role were the actions of Southerners who viewed dismantling white supremacy as a deep violation of a norm and who were willing to enforce the norm at a cost to themselves. As Foner (1988, pp. 431-32) reports, members of the Ku Klux Klan crossed class lines: "ordinary farmers and laborers constituted the bulk of Klan membership," which included men so poor that it was said they were "not worth the bread they eat." In one South Carolina county, Foner reports that nearly the entire white male population joined the Ku Klux Klan. This lends support to the view that material self-interest was not the only factor in the intimidation. Observers at the time noted that blacks did not have this power to act in common for self-protection (Foner, p. 436). 29 References Acemoglu, Daron and Robinson, James A. "De Facto Political Power and Institutional Persistence." American Economic Review, Papers and Proceedings, 2006, 96(2), pp. 325- 330. Banerjee, Abhijit, and Iyer, Lakshmi. "History, Institutions and Economic Performance: The Legacy of Colonial Land Tenure Systems in India." American Economic Review, 2005, 95 (4), pp. 1190-1213. Basu, Kaushik. Prelude to political economy: A study of the social and political foundations of economics. Oxford, UK: Oxford University Press, 2000. Bayley, Susan. Caste, society and politics in India from the eighteenth century to the modern age. Cambridge, Massachusetts: Cambridge University Press, 1999. Bernhard, Helen; Fehr, Ernst and Fischbacher, Urs. "Group Affiliation and Altruistic Norm Enforcement." American Economic Review, 2006, 96 (2), pp. 217-221. Bowles, Samuel. "Endogenous Preferences: The Cultural Consequences of Markets and Other Economic Institutions." Journal of Economic Literature, 1998, 36(1), pp. 75-111. Bowles, Samuel and Gintis, Herbert. "The Moral Economy of Communities: Structured Populations and the Evolution of Pro-Social Norms." Evolution and Human Behavior, 1998, 19(1), pp. 3-25. Carpenter, Jeffrey P. and Matthews, Peter Hans. "Norm Enforcement: Anger, Indignation or Reciprocity?" Manuscript, Middlebury College, 2006. Chen, Yan and Li, Sherry Xin. "Group Identity and Social Preferences," American Economic Review, 2009, 99, pp. 431-457. 30 Cialdini, Robert; Borden, Richard; Throne, Avril; Walker, Marker; Freeman, Stephen and Sloan, Lloyd. "Basking in Reflected Glory: Three (Football) Field Studies," Journal of Personality and Social Psychology, 1976, 34, pp. 366-75. Cinyabuguma, Michael; Page, Talbot and Putterman, Louis. "Can Second-Order Punishment Deter Perverse Punishment?" Experimental Economics, 2006, 9(3), pp. 265-79. Drze, Jean and Gazdar, Haris. "Uttar Pradesh: The Burden of Inertia," in Drze and Amartya Sen, eds. Indian development. Oxford, UK: Oxford University Press, 1997. Drze, Jean; Lanjouw, Peter and Sharma, Naresh. "Economic development in Palanpur, 1957- 93," in Lanjouw and Nicholas Stern, Economic development in Palanpur over five decades, Oxford, UK: Clarendon Press Oxford, 1998. Fehr, Ernst and Gächter, Simon. "Cooperation and Punishment in Public Goods Experiments." American Economic Review, 2000, 90(4), pp. 980-94. Fehr, Ernst; Gächter, Simon and Kirchsteiger, Georg. "Reciprocity as a Contract Enforcement Device: Experimental Evidence." Econometrica, 1997, 65(4), pp. 833-60. Fehr, Ernst and Fischbacher, Urs. 2004. "Third Party Punishment and Social Norms." Evolution and Human Behavior, 2004, 25, pp. 63-87. Fehr, Ernst; Hoff, Karla and Kshetramade, Mayuresh. 2008. "Spite and Development," American Economic Review, Papers and Proceedings 98(2): 494-99. Fisman, Raymond and Miguel, Edward. "Corruption, Norms, and Legal Enforcement: Evidence from Diplomatic Parking Tickets." Journal of Political Economy, 2007, 115 (6), pp. 1020-1048. Foner, Eric.[1988] Reconstruction: America's unfinished revolution. Reprinted: New York: Perennial Classics,1989. 31 Galanter, Marc. "The Abolition of Disabilities, Untouchability and the Law," in J. Michael Mahar, ed., The Untouchables in Contemporary India. Tucson: University of Arizona Press, 1972, pp. 227-314. Gellner, Ernest. Conditions of liberty: Civil Society and its rivals. New York: The Penguin Press, 1994. Gintis, Herbert. "A Radical Analysis of Welfare Economics and Individual Development." Quarterly Journal of Economics, 1972, 86(4), pp. 572-99. Gintis, Herbert. "Punishment and Cooperation." Science, 2008, 319(5868), pp. 1345-46. Goette, Lorenz; Huffman, David and Meier, Stephan. "The Impact of Group Membership on Cooperation and Norm Enforcement: Evidence Using Random Assignment to Real Social Groups." American Economic Review, 2006, 96(2), pp. 212-16. Greif, Avner. "Contract Enforceability and Economic Institutions in Early Trade: The Maghribi Traders' Coalition." American Economic Review. 1993, 83(3), pp. 525-548. Guiso, Luigi; Sapienza, Paola and Zingales, Luigi. "The Role of Social Capital in Financial Development," American Economic Review, 2004, 94(3), pp. 526-556. Guiso, Luigi; Sapienza, Paola and Zingales, Luigi. "Does Culture Affect Economic Outcomes?" Journal of Economic Perspectives, 2006, 20(2), pp. 23-48. Gupta, Dipankar. Interrogating caste: Understanding hierarchy and difference in Indian society. New York: Penguin Books, 2000. Hayek, Friedrich A. Law, legislation, and liberty: Volume 1 - Rules and order. Chicago, IL: University of Chicago Press, 1973. 32 Henrich, Joseph; Boyd, Robert; Bowles, Samuel; Camerer, Colin; Fehr, Ernst; Gintis, Herbert and McElreath, Richard. "Cooperation, Reciprocity and Punishment in Fifteen Small- scale Societies." American Economic Review, 2001, 91, pp. 73­78. Herrmann, Benedikt; Thöni, Christian and Gächter, Simon. "Antisocial Punishment across Societies." Science, 2008, 319(5868), pp. 1362-7. Human Rights Watch. Broken people: Caste violence against India's "Untouchables." New York: Human Rights Watch, 1999. Kocher, Martin G.; Cherry, Todd; Kroll, Stephan; Netzer, Robert J. and Sutter, Matthias. "Conditional Cooperation on Three Continents." Economics Letters, 2008, 101(3), pp. 175-78. Kohli, Atul. The state and poverty in India. Cambridge: Cambridge University Press, 1987. Lindbeck, Assar; Nyberg, Sten and Weibull, Jorgen W. "Social Norms and Economic Incentives in the Welfare State." Quarterly Journal of Economics, 1999, 114, pp. 1­35. Logan, Trevon D. and Shah, Manisha. "Face Value: Information and Signaling in an Illegal Market." NBER Working Paper 14841, 2009. Miguel, Edward and Gugerty, Mary Kay. "Ethnic Diversity, Social Sanctions, and Public Goods in Kenya." Journal of Public Economics, 2005, 89 (11-12), pp. 2325-68. Mokyr, Joel. The enlightened economy: An economic history of Britain, 1700-1850. Yale University Press and Penguin Press, 2008 (in press). Munshi, Kaivan and Rosenzweig, Mark. "Why is Mobility in India so Low? Social Insurance, Inequality, and Growth." Unpublished manuscript, 2007. 33 Nikoforakis, Nikos. "Punishment and Counter-Punishment in Public Good Games: Can We Really Govern Ourselves?" Journal of Public Economics, 2008, 92, pp. 91-112. New York University School of Law Center for Human Rights and Global Justice and Human Rights Watch. Hidden Apartheid: Caste Discrimination against India's "Untouchables." Shadow Report to the UN Committee on the Elimination of Racial Discrimination, 19 (3) 2007. North, Douglass. Institutions, institutional change, and economic performance. Cambridge, UK and New York: Cambridge University Press, 1990. Oakes, Penelope, Haslam, S. Alexander, and Turner, John C. Stereotyping and social reality. Oxford, England: Blackwell, 1994. Ostrom, Elinor. "Collective Action and the Evolution of Social Norms." Journal of Economic Perspectives, 2000, 14(3), pp. 137-58. Platteau, Jean-Philippe. "Behind the Market Stage Where Real Societies Exist--Part II: The Role of Moral Norms." Journal of Development Studies, 1994, 30(3), pp. 753-817. Rao, Vijayendra. "Symbolic Public Goods and the Coordination of Collective Action: A Comparison of Local Development in India and Indonesia," in Bardhan, Pranab and Isha Ray, eds., The Contested Commons: Conversations between Economists and Anthropologists. ch. 18, 2008. Rao, Vijayendra and Walton, Michael. "Culture and Public Action: Relationality, Equality of Agency, and Development," in Rao and Walton, eds., Culture and public action. Stanford, CA: Stanford University Press, 2004, pp. 3-36. Sen, Amartya. "Social Exclusion: Concept, Application, and Scrutiny." Social Development Papers No. 1, Office of Environment and Social Development, Asian Development Bank, 2000. 34 Shah, Ghanshyam, Mander, Harsh, Thorat, Sukhdeo, Deshpande, Satish and Baviskar, Amita. Untouchability in rural India. New Delhi: Sage Publications, 2006. Sobel, Joel. "Interdependent Preferences and Reciprocity." Journal of Economic Literature, 2005, 43, pp. 392-436. Tabellini, Guido. "Presidential Address - Institutions and Culture." Journal of the European Economic Association, 2008, 6(2-3), pp. 255-94. Tajfel, Henri."Social Psychology of Intergroup Relations."Annual Review of Psychology, 1982,33, pp.1-39. Thorat, Sukhadeo. "Oppression and Denial: Dalit Discrimination in the 1990s." Economic and Political Weekly, 2002, pp. 572-578. Weingast, Barry. "Political Foundations of Democracy and the Rule of Law." American Political Science Review, 1997, 91(2), pp. 245-63. Willer, Robb. "Groups Reward Individual Sacrifice: The Status Solution to the Collective Action Problem." American Sociological Review. 2009a, 74, pp. 23-43. Willer, Robb. "A Status Theory of Collective Action," in Shane R. Thye and Edward J. Lawler, eds., Advances in group processes. 2009b, 26. 35 Table 1 Determinants of Third Party Punishment (1) (2) (3) (4) (5) (6) Relative Punishment Absolute Punishment Constanta 2.073*** 4.548 4.783* 5.366*** 7.258** 7.575** (0.789) (2.868) (2.831) (0.765) (3.061) (3.040) Player C is high caste, 2.749** 2.552** 1.737* 3.086*** 3.444*** 2.112** (1.070) (1.114) (0.898) (1.062) (1.095) (0.918) Treatment is mixed status, 0.512 0.678 0.78 1.021 (1.284) (1.296) (1.341) (1.340) Player C is high caste and treatment is mixed status, -1.368 -1.652 -2.429 -2.703 (1.642) (1.643) (1.701) (1.694) Individual characteristics of player C Land owned in acres (centered) 0.012 0.009 -0.184 -0.191 (0.143) (0.142) (0.141) (0.140) Lives in a mud houseb 0.849 0.707 0.665 0.425 (0.913) (0.920) (0.894) (0.896) Has completed at least 10 years of schooling 1.955** 1.896** 1.819** 1.725* (0.892) (0.882) (0.906) (0.899) Is an owner-cultivator -1.505* -1.568* -1.572* -1.680* (0.852) (0.838) (0.872) (0.865) Has political experience -1.36 -1.248 -1.159 -0.96 (1.375) (1.388) (1.491) (1.517) R2 0.035 0.081 0.076 0.037 0.088 0.074 HHH F stat: pun = punHLH 0.70 0.90 2.48 2.60 Prob F 0.404 0.343 0.117 0.108 Notes: The table reports on player C`s punishment of norm violations in treatments LLL, HHH, LHL, and HLH. Each column gives the results of an OLS regression based on 205 observations. Robust standard errors are in parentheses. Significance at the 10% level is represented by a *, at the 5% level by a **, and at the 1% level by ***. a The constant is the baseline group mean with respect to which we measure changes. It is LLL in regressions (1),(2),(4), and (5) and LLL and LHL in regressions (3) and (6). b The omitted category is a house constructed at least partly of brick. 36 Table 2 Determinants of Third Party Punishment, Including Specific Caste Affiliation (1) (2) (3) (4) Relative Absolute Punishment Punishment Constant (baseline group mean: Player C is Chamar) 1.933* 4.124 5.263*** 6.808** (1.068) (2.908) (1.090) (3.143) Player C is Brahmin 3.294** 3.217** 3.482** 3.945*** (1.364) (1.378) (1.417) (1.404) Player C is Thakur 2.458* 2.988** 2.875** 3.999*** (1.421) (1.403) (1.421) (1.415) Player C is Pasi 0.273 1.095 0.2 1.04 (1.290) (1.282) (1.343) (1.347) Treatment is mixed status 0.506 0.653 0.776 1.002 (1.288) (1.297) (1.343) (1.338) Player C is high caste and treatment is mixed status -1.354 -1.621 -2.419 -2.679 (1.647) (1.650) (1.705) (1.699) Individual characteristics of Player C Land owned in acres (centered) 0.015 -0.182 (0.143) (0.142) Lives in a mud housea 0.858 0.691 (0.923) (0.903) Has completed at least 10 years of schooling 2.001** 1.900** (0.896) (0.935) Is an owner-cultivator -1.619* -1.710** (0.830) (0.860) Has political experience -1.408 -1.183 (1.350) (1.472) R2 0.038 0.085 0.039 0.091 Notes: The table reports on player C`s punishment of norm violations by the specific caste of player C in treatments LLL, HHH, LHL, and HLH. Each column gives the results of an OLS regression based on 205 observations. Robust standard errors are in parentheses. Significance at the 10% level is represented by a *, at the 5% level by a **, and at the 1% level by ***. a The omitted category is a house constructed at least partly of brick. 37 Table 3 Determinants of Third Party Punishment, Including the Effect of In-group Affiliation (1) (2) (3) (4) Relative Absolute Punishment Punishment Constant (baseline group mean: LLL) 2.073*** 5.401* 5.366*** 7.882** (0.790) (2.804) (0.767) (3.265) Player C is high caste 2.749** 2.766** 3.086*** 3.745*** (1.072) (1.167) (1.064) (1.104) Players B and C have the same specific caste but different from player A's 0.149 -0.164 -0.505 -0.817 (1.179) (1.138) (1.262) (1.203) Interaction of BC-condition and Player C is high caste -2.1 -3.050* -2.742 -3.857** (1.752) (1.734) (1.776) (1.717) Individual characteristics of Player C Land owned in acres - (centered) -0.251 0.466*** (0.193) (0.172) Lives in a mud housea 0.221 -0.242 (0.892) (0.887) Has completed at least 10 years of schooling 2.823*** 2.339*** (0.922) (0.889) Is an owner-cultivator -1.618* -1.821** (0.844) (0.843) Has political experience -1.909 -1.455 (1.295) (1.573) R2 0.043 0.125 0.071 0.161 F-stat: HHH = HHHBC 2.270a 6.530a 6.750a 15.320a Prob F 0.134 0.011 0.010 0.000 F-stat: LLLBC = HHHBC 0.220b 0.040b 0.060b 0.010b Prob F 0.640 0.845 0.809 0.937 Notes: The table reports on player C`s punishment of norm violations in treatments LLL, HHH, LLLBC , and HHHBC . Each column gives the results of an OLS regression based on 178 observations. Robust standard errors are in parentheses. Significance at the 10% level is represented by a *, at the 5% level by a **, and at the 1% level by ***. a The omitted category is a house constructed at least partly of brick. 38 Table A1 Determinants of Per Capita Consumption Per capita adult-equivalent consumption (ln) High caste Low caste Ln (land owned)a 0.282*** 0.084*** (0.039) (0.024) Lives in a mixed mud and brick houseb -0.052 0.100 (0.111) (0.081) Lives in a brick houseb 0.179* 0.238** (0.102) (0.079) Has completed at least 10 years of schooling 0.304*** 0.118* (0.093) (0.067) Constant 8.001*** 7.875*** (0.082) (0.033) Number of observations 131 204 R2 0.330 0.385 Notes: The table shows OLS regressions of per capita adult-equivalent consumption for the high caste and low caste households, respectively, covered in the survey of Uttar Pradesh in the World Bank 1997-98 Survey of Living Conditions: Uttar Pradesh and Bihar. Robust standard errors are in parentheses. Significance at the 10% level is represented by a *, at the 5% level by a **, and at the 1% level by ***. a Land ownership of each household is shifted up by 1/6 acre so that landless households can be included in the regression. b Omitted category is a house of walls constructed only of mud, with either thatch or tiled roof. 39