Download - Making up People- The Effect of Identity on Preferences and Performance in a Modernizing Society

Journal of Development Economics 106 (2014) 118–131

Contents lists available at ScienceDirect

Journal of Development Economics

j ourna l homepage: www.e lsev ie r .com/ locate /devec

Making up people—The effect of identity on performance in amodernizing society

Karla Hoff ⁎, Priyanka PandeyThe World Bank, USA

⁎ Corresponding author at: 1818 H St., Washington, DC4077; fax: +1 202 522 3518.

E-mail addresses: [email protected] (K. Hoff), ppa(P. Pandey).

0304-3878/$ – see front matter © 2013 Elsevier B.V. All rhttp://dx.doi.org/10.1016/j.jdeveco.2013.08.009

a b s t r a c t
a r t i c l e i n f o
Article history:Received 22 September 2012Received in revised form 7 August 2013Accepted 16 August 2013

JEL classification:C93D02D03I2O15

Keywords:Randomized experimentLabelsSocial identityInstitutional changeStereotype threatCaste

It is typically assumed that being hard-working or clever is a trait of the person, in the sense that it is always there, in afixedmanner. However, in an experimentwith 288 high-caste and 294 low-caste students in India, cues to one's placein the caste system turned out to starkly influence the expression of these traits. The experiment allows us to discrim-inate between two classes of models that give different answers to the question of how someone's identity affects hisbehavior.Models of the fixed self assume that identity is a set of preferences.Models of the frame-dependent self assumethat identity entails a set of mental models that are situationally evoked and that mediate information processing. Ourfindings suggest that theeffectof identityon intellectualperformancedepends sensitivelyon the social setting. Thisper-spective opens up new policy options for enhancing human capital formation and development.

© 2013 Elsevier B.V. All rights reserved.

1. Introduction

A number of models in economics give different answers to thequestion of how someone's identity—an individual's sense of the socialcategories to which he belongs—might affect his behavior. We presentan experiment that allows us to discriminate between some of thesemodels. We show that situational cues to identity can alter intellectualperformance. Our findings suggest that identity can have a first-ordereffect on human capital formation and development.

A central goal in many disciplines is to understand how identity af-fects behavior. Historians have documented that societies all over theworld have invented social identities and used symbols, etiquette, rit-uals, dress codes, and segregation to impress on people the notion thatcertain individuals constituted significantly different categories andwere subject to different constraints. For example, in Growing up JimCrow: How Black and White Southern Children Learned Race, JenniferRitterhouse (2006, p. 4) describes how the unwritten rules that

20433, USA. Tel.: +1 240 473

[email protected]

ights reserved.

governed interactions across race lines were used “not only as a formof social control but also as a script for the performative creation of …‘race’ itself.” The Victorian clothing system with its frock coats for menand tight-laced corsets for women “produced the existence of [certain]categories of behaviour and kept them habitually in being by mouldingbodily configuration and movement”: The role of men was to be activeand strong (their clothes allowed them movement and emphasizedbroad chests and shoulders); the role of women was to be inactive andsubmissive (their clothes inhibited movement and were constricting)(Connerton, 1989, p. 34). Our title alludes to and amplifies the title ofan essay, “Making Up People,” in which the philosopher Ian Hacking(1986) argues that the creation of new slots in which to fit and catego-rize people, e.g. the perverted, the suicidal, and the heterosexual or homo-sexual person, molds individuals' sense of themselves and producesbehavior that would not occur in the absence of these labels.

Our concern in this paper is to measure how labels affect learningand intellectual performance. A large body of work supports the viewthat identity matters for human capital achievement. Social normsand, sometimes unknowingly, schools and teachers can encourageor discourage students from given social backgrounds. Akerlof andKranton (2002) present evidence that whether students invest inschooling depends in part on their cultural identity, wherein payoffsdiffer among “jocks,” “nerds,” and “burnouts.” Anthropological stud-ies document that school routines and curricula can convey to black

http://crossmark.crossref.org/dialog/?doi=10.1016/j.jdeveco.2013.08.009&domain=pdf

http://dx.doi.org/10.1016/j.jdeveco.2013.08.009

mailto:[email protected]

mailto:[email protected]


http://www.sciencedirect.com/science/journal/03043878

1 Through a political process, social mobilitywas possible on the part of the community,but not by individuals (see Rao and Ban, 2007 for a measure of the process of change incaste groups). For an exception, in which individuals were able to change their status forpurposes of British law in India, see Cassan (2011).

2 Two examples are illustrative. (i) In July 1998 in the state of Uttar Pradesh, a HighCourt judge had his chambers “purifiedwith ‘Ganga jal’ (water from the River Ganges) be-cause it had earlier been occupied by a Dalit judge” Times of India (Bombay), July 23, 1998.(ii) Increases inwealth among the low-caste community of Pallars from remittances fromthe Persian Gulf created tensions that led the entire Dalit population of the village to bedriven out by the marginally higher caste Thevas, who set fire to their homes and fields(Human Rights Watch, 1999, pp. 85); a similar episode occurred elsewhere in 2012(“When development triggers caste violence,” The Hindu, May 8, 2013). A recent IndianTV show provided one of the first public settings where viewers heard low-caste individ-uals speak first-hand about their harsh and horrifying experiences of growing up as un-touchables (www.satyamevjayate.in, episode 10).

3 High-caste boys were drawn from Thakurs (62%), Brahmins (32%), Vaishayas (4%),and others (2%). Low-caste boys were drawn only from Chamars. In Hardoi district,Chamar is considered the lowest caste. The set of high castes and the Chamar caste eachmakes up somewhat less than 20% of the population of the district.

4 See Déliege (1999) and Gupta (2000). In contrast, Hindu surnames domark caste. Forthat reason some low-caste groups have sought a constitutional amendment to abolishHindu surnames (The Telegraph, October 15, 2005, “Slash Surname to Kill Caste”). Thanksto Pranab Bardhan for this reference.

119K. Hoff, P. Pandey / Journal of Development Economics 106 (2014) 118–131

students that there is something “wrong” with them and their back-ground (Ogbu, 1999). In schools in which students are from onebackground and teachers are from a different background, the ten-sion between students and teachers, Ogbu suggests, may give riseto an oppositional culture among students. Studies of adolescentsfind some evidence of racial differences in the relationship betweensocial status among peers and academic achievement; blacks, for ex-ample, may face a tradeoff between acceptance and academicachievement that whites do not face. Not wanting to be rejected bytheir peer group for “acting white,” they may put less effort intotheir schoolwork (see Fryer, 2011 for a review). Ferguson (2003)presents evidence that teachers' perceptions, expectations, and be-havior differ across students of different social groups and that theinteraction between the expectations of teachers and those of stu-dents contributes to the black–white test score gap. On the brighterside, all these findings suggest that schools have room to reducethe disparity in educational outcomes by ethnicity and race by ad-dressing at least two sources for this difference—teachers and stu-dents. Minority groups, such as African-American students, aremore likely to report discrimination by teachers. In a sample of eth-nically diverse US schools, a large percentage of students reportedbeing bullied by peers based on their ethnicity (Bellmore andTomonaga, 2009). Students who reported ethnicity-based discrimi-nation were more likely to experience depressive symptoms.

The sociologists AnnSwidler (1986, 2001, especially p. 161) andPaulDiMaggio (1997) argue that culture, as a matter of self-conscious orien-tation or identity, is not a set of values or preferences, but instead is afragmented set of mental models, understandings, worldviews, andguides to action, and these elements may not be consistent. Culture af-fects action when it is given force by contexts that organize culturalmeanings and bring them to bear on action. Culture shapes behaviorthrough cognitive frames (mental models) that are situationally evokedand that determine how an individual construes a situation and whichactions seem possible and desirable in that situation, given a person'svalues. “If one asked a slum youth why he did not take steps to pursuea middle-class path to success … the answer might well be not ‘I don'twant that life,’ but instead, ‘Who, me?’” (Swidler, 1986, pp. 275).

This perspective suggests that the effect of identity on performancemay depend on which of a set of possible mental models is evoked in agiven situation. To test the hypothesis, we draw on our experiment inrural India that manipulates the classroom setting in which high-casteand low-caste students are asked to learn and perform a new skillunder incentives (Hoff and Pandey, 2006). A feature of the caste systemthat makes it well-suited to identifying the effect of identity is that anindividual's caste is determined by the accident of birth. Althoughcaste boundaries and caste rankings change over long periods, the sta-tus in north India of the specific high castes (primarily Thakur and Brah-min) and the specific low caste (Chamar) from which we draw ourparticipants goes back millennia (Gupta, 2000).

A second reason that caste is well-suited for the study of the effect ofidentity on performance is that themeaning of the caste categories is notin doubt. Two central characteristics of the caste system are hierarchyand repulsion (or difference) between castes (Gupta, 2000). Hierarchyreflects beliefs in essential differences between the castes:

For centuries it was believed that a man's social capacities wereknown from the caste or the lineage into which he was born, andthat no further test was necessary to determine what these capaci-ties were (Béteille, p. 99).

Repulsion is expressed in endogamy and restrictions on contact be-tween castes, in particular, between those distant in rank. A body ofwork in anthropology supports the view that vertical mobility of the in-dividual within the caste hierarchy was nearly impossible for someonewho remained in or near his village, where his caste identity would beknown (e.g. Gupta, 2000; Srinivas, 2009). A low-caste boy could not

move up, and a high-caste boy could not move down.1 At the bottomof the traditional caste hierarchy are the castes whose members weremarked as polluted. They were called “untouchables” and are todaycalled Dalits. Untouchability is the imposition of social disabilities onpersons by reason of their birth in certain low castes. Dalits were notallowed to own land or use public wells, were made to use separatecrockery in food stalls, and were not allowed to sit inside a schoolhousebut instead forced to remain outside.

Untouchability is illegal under the Constitution of India, and atti-tudes in Uttar Pradesh towards Dalits are radically different todayfrom what they were in the recent past (Kapur et al., 2010). Evidenceof a new social order is visible to every schoolchild in the stipends pub-licly distributed to Dalits to encourage school enrollment and in thebroad participation of Dalits in the political process. Yet children arealso likely to encounter the traditional order of caste, segregation, anduntouchability in their own experiences, through the fables they learn,and in the continued insults and atrocities against upwardly mobileDalits.2 Two surveys give some indication of how untouchability playsout in schools in the form of humiliation and segregation:

One common example of social prejudice in the classroom is the dis-paraging attitude of upper caste teachers towards Dalit children. Thiscan take various forms, such as telling Dalit children that they are ‘stu-pid,’making them feel inferior, using them formenial chores, and giv-ing them liberal physical punishment (PROBE, 1999, pp. 51).In one outof four primary schools in rural India, Dalit children are forced by theirteachers or by convention to sit apart fromnon-Dalits. Asmany as 40%of schools practice untouchability while serving mid-day meals, mak-ing Dalit children sit in a separate rowwhile eating (Shah et al., 2006,pp. 168, based on a 2001–02 national survey).

The two social orders coexist in uneasy tension—one in which “we areall equal now,” as some villagers remarked to us, and another inwhich up-wardmobility by low-caste individuals is not liked and could be dangerous.

Our experiment assesses the effect on children's intellectual perfor-mance of making caste identity public and of segregating children bycaste. Participants in our experiment were junior high school boysfrom the top three castes and from the lowest caste.3 In groups of six,the participants were asked to solve mazes under incentives and wererandomly assigned to one of three conditions.

(1) In the control condition, caste identity, which is not discerniblefrom natural physical markers,4 was not made public in a sessionof three high-caste and three low-caste boys. Since there is

http://www.satyamevjayate.in

5 A secondexample is that the categories for colors in the language one speaks influenceone's ability to discriminate between colors. See Alter (2013, ch. 2) for a popular, lively ac-count of the influence of categories on cognition.

120 K. Hoff, P. Pandey / Journal of Development Economics 106 (2014) 118–131

a large overlap in poverty between high- and low-caste house-holds in the local population, quality of dress is not an automaticsign of caste status. We call the control condition Caste Not Re-vealed.

(2) The second condition (RevealedMixed)was the same as the con-trol except that caste identity was made public.

(3) The third condition (Revealed Segregated) was the same asthe second except that a session was composed of only high-caste boys or only low-caste boys. Participants would likelyhave been aware that the composition of their sessionreflected deliberate segregation by caste status. This is so be-cause participants were driven to the experimental site withother boys from their village, whose caste they would know,in groups equally divided between high and low castes. Afterarriving at the school that was the site of the experiment, indi-viduals were assigned to new groups of six boys to participatein an experimental session. The probability that a boy wouldfind himself in a groupwith five other boys of his same caste sta-tus as a result of a random draw of the entire local school popu-lation was very small—less than (0.2)5 = 0.00032. If childrenperceived the segregation as deliberate, it is likely that theywould also perceive it as meaningful in the context of the tradi-tional caste system in which high castes are not supposed tomingle with Dalits. The practice of segregation remains inmany villages an issue over which high and low castes struggle(Human Rights Watch, 1999).

The control condition shows that low-caste boys solve mazes just aswell as high-caste boys. This is true under piece-rate incentives. It isalso true under the tournament incentive. We have three main resultsfor the effects of revealing caste. (All effects reported here are significantand control for individual characteristics.)

Result 1. Under piece-rate incentives, publicly revealing caste in mixed-caste groups creates a 23% caste gap in totalmazes solved in favor of thehigh castes. Since therewas no caste gap in the control, we infer that inother possible worlds the low castes could have been an equal or adominant group. Here, a social identity has affected behavior.Underlying the caste gap is an increase, when caste is revealed, inthe proportion of low-caste boys who fail to learn to solve mazes.Our first result extends to a new social group—low-caste individualsof India—a large body of work in psychology that finds that cues toa person's identity, if it is stereotyped as intellectually inferior,may undermine the person's ability to perform cognitive tasks(stereotype threat).Result 2. Under piece-rate incentives, both high- and low-caste boysunderperform by over 20%when caste is publicly revealed in segregatedgroups. Stereotype susceptibilitywould predict this result for the lowcaste, but it would not predict this result for the high caste. Further-more, all available evidence—from failure rates to learn how to solveamaze in this experiment and from a direct test of self-confidence inanother study (Hoff and Pandey, 2005)—show that the RevealedSegregated condition does not lower the self-confidence of high-caste boys. To us the most plausible interpretation of the decline inhigh-caste performance is that segregation is a marker of high-caste dominance and evokes a sense of entitlement, in which thehigh-caste boys feel less need to achieve. As the Indian sociologistAndré Béteille notes, in the caste system social preeminence isassigned by birth rather than by competition (Béteille, 2011, II[2003], p. 11).Result 3. Under tournament incentives, both high-caste and low-caste boys underperform when caste is revealed, but in the segregat-ed condition, the effect is much larger for low-caste boys. In a tourna-ment, strategic rather than psychological factors could explain

why revealing identity impairs performance. Becausewe cannot dis-entangle these factors, we relegate the analysis of the tournamentsto Appendix A.

2. The fixed self and the frame-dependent self

Our experiment will allow us to discriminate between two broadsets of theories about how social identity might affect performance.We will call the two sets of theories the fixed self and the frame-dependent self.

2.1. The fixed self

A broad class of theories takes the view that an individual at a mo-ment in time has fixed,well-defined preferences and abilities. They pro-vide all the information that is relevant for describing the individual'schoices and outcomes for a given set of opportunities. In the textbookmodel in economics, an individual has preferences in which a sense ofidentity with others has no influence. This theory is one of the funda-mental differences between the standard model of economics and theconcept of the individual that other social sciences find useful, inwhich socially defined variables, such as conformity, affect preferences.

The theory that an individual has, at any moment in time, a well-defined set of preferences and that they are always salient, ismaintained in recent work that substantially broadens the notion ofpreferences by incorporating one's sense of group membership. InAkerlof and Kranton (2000), a social category constitutes part of anindividual's identity. Associated with the category are a set of normsor ideals for how someone in the category should behave. The individuallikes conforming to the ideals of that category, and dislikes actions byothers that deviate from the ideals.

An individualmaybe able to choose his social identities, i.e.he cande-fine himself and his relationships to others at a categorical level (Akerlofand Kranton, 2002; Fang and Loury, 2005; Hoff and Sen, 2006; Munshiand Rosenzweig, 2006;Munshi andWilson, 2008; Sen, 2006). For exam-ple, a descendant of Irish immigrants to the US can define himself asIrish-American or not. The individual's choice problem makes senseonly under the assumption that an individual has ameta-utility function.However, just as in the two theories above, an individual has well-defined preferences that provide all the information that is relevant fordescribing his choices and outcomes for a given set of opportunities.

2.2. The frame-dependent self

The other broad class of theories draws on a certain view of how themind works, namely, that information is processed in relation to mentalmodels (references are in Tverksy and Kahneman, 1983 and an elabora-tion is in DiMaggio, 1997). Individuals have multiple mental models,they may not be consistent, and they are situationally evoked.

An experiment from Tversky and Kahneman illustrates the powerfuleffect of cues on cognition. A group of subjects is asked, How manyseven-letter words of which the sixth letter is “N” (_ _ _ _ _ N _)would you expect to find in four pages of a novel (about 2000words)? The same group of subjects is also asked, How many seven-letter words of which the last three letters are “ING” (_ _ _ _ ING)would you expect to find in four pages of a novel (about 2000words)? The median estimate is several times greater for ING wordsthan for _N_ words. The responses violate logic but can be explainedthis way: _ING words are a standard category (a very simple mentalstructure), _N_ words are not, and categories shape how we think.5

There is substantial evidence that the setting in which alternativesare offered affects choices, and that one mechanism underlying this

6 In the 1998–99 IndianNational FamilyHealth Survey, householdswere asked to nametheir caste. Most low-caste respondents gave their actual caste name (e.g. Chamar), but afew used themore generic and politically correct names, Dalit, harijan, or Scheduled Caste(Marriott, 2003).


influence is that the setting evokes a particular mental model or aspectof the self. For example, in the studies of Benjamin et al. (2010), Cohnet al. (2013), and LeBoeuf et al. (2010), individuals were randomlyassigned to fill out background questionnaires that either includedquestions related to an aspect of their identities or that did not. A ques-tionnaire that primedAsian identitymade Asian-Americansmore coop-erative and patient, a questionnaire that primed a family-orientedidentity strengthened values related to family obligations, and a ques-tionnaire that primed the prisoner identity of prison inmates led themto cheat more. Even the language of thought affects behavior; studieswith bilingual students find that randomly directing them to use oneor the other of their two languages shifts their implicit attitudes and be-havior (Danziger andWard, 2010; Ogunnaike et al., 2010). In the studyof Lambarraa et al. (2012), which also uses bilingual subjects, the lan-guage of thought in the period before the subject has to make a choiceinfluences his behavior. Since the decision-making situation is exactlythe same but only the prior language has changed, it is clear that contexthas changed notwhat it is objectively appropriate to do but instead thewaythat the subject evaluates his choices. The effect of context on behavior isanalogous to the effect of the environment on the expression of anindividual's genes. DNA are the instructions for making an individual,but as yet poorly understood features of the environment determinethe on-and-off states of genes. (See Salant and Rubinstein, 2008 for amodel of frame-dependent utility.)

The idea of an extended utility function is interesting when it leadsto the observation of inconsistent choices. Of course, if we knew allthe stimuli to the individual, then the theory of rationality (i.e. consis-tency) would be trivial. Since we do not observe all stimuli, and our un-derstanding of how individuals process information is limited, itbecomes a useful construct to posit multiple preferences, one for eachself-construal or worldview.

Useful for what purpose? It may be useful for understanding long-run social change, which entails changes in the set of possible identities,the salience of particular identities, and the possible ways of under-standing a situation. In the process of economic development, the stim-uli to an individual can change in a way that leads to the expression ofone set of preferences rather than another; that is, preferences dependon context and frames.

Finally, an influential body of evidence in psychology on stereotypethreat (and to a lesser extent, stereotype boost) relates to the influenceof context on human productivity. In experiments over many domains(e.g. SATs and sports), priming a negatively or a positively stereotypedaspect of an individual's identity shifts performance in the direction ofthe stereotype: for instance, African-Americans do worse on scholasticaptitude tests if before the test they are asked to check a box for theirrace (Steele and Aronson, 1995). Asian-American women, if the Asianaspect of identity is made salient, do better on math tests than womenin the no-prime condition, but if their gender is made salient, do worsethan women in the no-prime condition (Shih et al., 1999). Children inboth lower elementary grades and middle school grades (but notthose in upper elementary grades) show shifts in performance consis-tent with the patterns of stereotype threat and stereotype boost(Afridi et al., 2010; Ambady et al., 2001). However, other studies donot find evidence of stereotype threat (e.g. Fryer et al., 2008).When ste-reotype threat occurs, results in neuroscience support the view that thestereotype is a powerful distractor in the task that subjects are trying toaccomplish. In Krendl et al. (2008), women taking a math test in condi-tions of stereotype threat did not recruit the neural regions associatedwith mathematical learning but instead showed heightened activationin a neural region associated with social and emotional processing.Neuroscientists also find that early environmental exposure can af-fect the chemistry of the DNA, with a long-term effect on one's re-sponses to stress (Begley, 2007; Pollak, 2008). Children who have ahistory of exposure to stressful events react more strongly to stress.We conjecture that for a low-caste child who has encounteredstressful events related to his caste identity, increasing the salience

of caste is a source of stress. In contrast, a high-caste child, becauseof his early experiences, may find comfort and power in contexts inwhich his caste identity is salient.

3. Participants and design

288 high-caste (hereafter H) and 294 low-caste (hereafter L) in 6thor 7th grade in the district of Hardoi in the Indian state of Uttar Pradeshparticipated in the study. Hardoi was under feudal rule in the 19th cen-tury, and the feudal elitewasmade up of the high castes. A legacy of feu-dal rule is greater high-caste dominance compared to regions of UttarPradesh not under such rule (Pandey, 2010). Thus the site of our exper-iment is a relatively high-caste dominant district of a region of India (theNorth) in which caste divisions run deep.

In the experiment, participants, in groups of six, solved mazes in aclassroom. The six boys in a session were generally recruited from sixdifferent villages, but since this was not always the case, wewill controlfor the number of other participants that a participant knew. Partici-pants were brought to the experiment site by car. Just before enteringthe car, each participant was asked in private his name, village name,father's name, grandfather's name, and caste. On arriving at the site,we verified in private with each participant his name and caste beforerandomly assigning him to a treatment.

Three conditions varied the salience of caste in the session, whichwas always led by a high-caste young female experimenter:

Caste Not Revealed (the control condition). A session was composedof 3 H and 3 L. No personal information about the participants wasrevealed.Revealed Mixed (i.e. caste revealed in a mixed-caste session). Thecomposition of a sessionwas the same as in the preceding condition,but now the experimenter began a session by saying that she wouldlike to confirm some information with each participant, who shouldnod if it is correct. Then the experimenter turned to each participantand stated his name, village name, father's name, grandfather'sname, and caste.Revealed Segregated (i.e. caste revealed in a segregated session). Thiswas the same as the preceding condition except that a session wascomposed of either 6 H or 6 L.

The priming mechanism reflects a way in which caste identity isactually made salient in classroom settings. This increases the exter-nal validity of our results. Although an individual's caste is widelyknown in a village, publicly referring to a child's caste is not uncom-mon in rural schools. There is anecdotal evidence of teachers tellinglow-caste children to not drink from the tap at the school lest it pol-lute the water for others. While implementing this study, we cameacross some such instances. Caste is commonly recorded in schoolenrollment books, often using different colors for high and lowcastes, to identify caste-targeted entitlements such as stipends anduniforms provided by state governments. In villages, people are fre-quently called by their caste names. Following the common usage inthis area and the way that caste is recorded in school enrollmentbooks, we used the traditional name for each caste (Thakur, Chamar,etc.) in revealing caste identity.6

Fig. 1. Experiment design. Note. P/P means that the piece rate incentive applies in both rounds of maze-solving. P/T means that the piece rate incentive applies in round 1 and the tour-nament incentive applies in round 2.


We next describe the incentive schemes. Participants were givena packet of 15 mazes to solve in each of two 15-minute rounds.7

Some participants had piece-rate incentives in both rounds (the “P/Ptreatments”); others had piece-rate incentives in round 1 and tourna-ment incentives in round 2 (the “P/T treatments”). Under the piece-rate scheme, a participant earned one rupee per maze solved. Underthe tournament scheme, he earned six rupees per maze solved if hesolved the most mazes in his session; otherwise he earned nothing. Incase of a tie, both winners received the prize. The tournament providedhigh-powered incentives: a winner could (and some did) earn 15 × 6rupees, equivalent to almost two days' unskilled adult wages.

Fig. 1 gives the organization of the experiment. Experimental condi-tions were identical in the first round of treatments (1) and (4), (2) and(5), and (3) and (6), and so we will pool them when reporting first-round results.

3.1. Recruitment

We conducted the experiment in January and March 2003 and inMarch 2005. In January 2003, on days that schools were open, wewent to public schools near the site of the experiment and chosehigh- and low-caste children for each day after pooling the enrollmentdata for all nearby public schools. A letter from the District Magistrateinstructed the teachers to cooperate with our team. On days thatschools were closed, we visited homes in nearby villages each eveningto ask parents' permission to pick up their children the next day todrive them to the junior high school that served as the site of the exper-iment. In only rare instances did parents refuse to let their children par-ticipate. In March 2003 and March 2005, to choose the subjects, everyday our team went to six randomly selected villages within a 20-kilo-meter radius of the experiment site. From each village, we drew anequal number of high-caste and low-caste children. At most ten partic-ipants came from a single village, nearly always an equal number of Hand L. On each day, we recruited participants from a new set of villages.

3.2. Implementation

On arrival at the experiment site, participants waited in silence in alarge common room. A research assistant provided each child a snack.When the cars bringing the participants for the morning or afternoonsessions had all arrived, the participants were directed in groups of six

7 The mazes are Xerox copies from http://games.yahoo.com/games/maze.html, level 3.Gneezy et al. (2003) showed that individuals do not solvemazes just for fun, they respondto incentives.

to a new set of classrooms, where they remained for the rest of the ex-periment. Theywere not told anything about how orwhy the particulargroups were formed.

We next describe what took place during an experimental session,which lasted about 70 min. Under the Revealed Mixed and Revealed Seg-regated conditions but not in theControl, the experimenter began a sessionby making public the identity of the participants, as described above. Afterthat, all sessions proceeded in the same way. The experimenter told theparticipants that they would “take part in two games of solving puzzles.”She gave participants the show-up fee of 10 rupees and described how tosolve a maze in this way:

… there is one child. The child has to go to the ball. The solution is apath that takes the child to the ball. The black lines are walls. Thechild cannot cross a wall.

She gave participants 5 min to practicewith an additional maze. Sheexplained that for eachmaze they solved, participants would receive anadditional one rupee. She checked to make sure each child understoodthe incentive scheme. She said that the incentive payments “will begiven to every child in an envelope after the games are over.” Thenshe told the participants that they would have 15 min to solve a packetof mazes, and the first round of maze-solving began.

After the first round, and without giving feedback on performance,she explained that there would be one more round of solving mazes,explained the incentive scheme (piece rate or tournament), andchecked that each child understood it. In a post-play survey, the partic-ipants gave information about their background in private. Mazes weregraded blind. The participants received their earnings in sealed enve-lopes and were then taken home.

3.3. Predictions

Under piece rates, the payoff to a participant depends on only his ownactions. Therefore the theories of thefixed selfwould predict that increas-ing the salience of castewould have no effect on behavior. In contrast, thetheories of the frame-dependent selfwould predict that increasing the sa-lience of caste could evoke for a low-caste individual themental model inwhich Dalits are accepted only so long as they stay “in their place,”whichwould reduce the utility from high achievement and thus reduce perfor-mance. For a high-caste individual, in contrast, the prediction of the theo-ries of the frame-dependent self is ambiguous. On the one hand, makinghim more aware of his caste should, if anything, enhance his desire toconform to the ideal of a high-caste person, which is to be superior. Onthe other hand, making caste more salient could evoke a mental model

http://games.yahoo.com/games/maze.html

Predicted effect of increasing caste salience on the performance of:TheoryLow casteHigh caste

The Fixed SelfIndividuals have well-defined preferencesthat are always salient.

None None

The Frame-Dependent SelfIncreasing an individual’s awareness of anaspect of his identity may cue a mental model.Individuals have multiple sets of preferences,one for each mental model. Cues to anegatively stereotyped identity can alsoimpair the ability to perform.

Ambiguous—Cueing an identity whose normis to be superior increases theutility from achievement, whichimproves performance, butevoking a worldview in whichlife chances depend less oneffort than on caste impairsperformance.

Declines—Making a low-caste personmore aware of his caste (i)evokes a worldview in which itis a norm violation for him toexcel, and (ii) may triggerstereotype threat.

Fig. 2. Predicted effects of increasing the salience of caste under piece rate incentives.

Table 1Descriptive statistics for participants.

High caste Low caste

Caste NotRevealed

Identityconditions

Caste NotRevealed

Identityconditions

Mother's educationNone 32% 25% 75% 68%Years (0, 6) 26% 29% 17% 17%At least 6 years 42% 46% 8% 15%⁎

Father's educationNone 6% 6% 26% 31%


inwhich he has less need to achieve because he has an entitlement to sta-tus. In addition, under the theory of stereotype threat or boost, makingcastemore salientmay entail a negative productivity shock to L and apos-itive productivity shock to H—see Fig. 2.

4. Descriptive statistics

In this section, we describe the participants' characteristics andbroadly summarize the results.8 Table 1 shows that parents of H havemuch greater education than parents of L. For simplicity, the tablegroups together Revealed Mixed and Revealed Segregated as the “iden-tity conditions.” The table shows that 45% of all H compared to 12% of allL have amotherwith at least six years of schooling. (These areweightedaverages across conditions, calculated using Fig. 1.) Both parents are il-literate among only 5% of H, compared to 28% of L. Only 8% of H have fa-thers who are day laborers, compared to 18% in the case of L. Thesedifferences highlight the need to examine whether the correlates ofcaste can explain the differences between H and L in our results. Wecan do that because the distribution of parents' characteristics for Hshares a common support with that for L. For example, there are notonly L who have mothers with no schooling; there are also H whosemothers have no schooling. We collected data on two other variablesin the post-play survey: exposure to mazes and the number of otherparticipants in a session that a subject knows.

Table 1 shows that the randomization between the control and iden-tity conditions was largely successful. However, in the identity condi-tions, participants have parents with a significantly higher level ofeducation and are significantly more likely to have had some exposureto mazes. These differences should, if anything, improve performancein the identity conditions compared to the control. An effect that goesthe other way is that the low caste is on average slightly more likelyto be in 6th than in 7th grade in the identity conditions.9 We control

8 In each time period in whichwe conducted the experiment (January andMarch 2003and March 2005), we held at least six sessions under P/P incentives in the control condi-tion. As shown inWeb Appendix Table S1, there were no significant differences in outputby time period. Thereforewe pool thedata across the three time periods.We also found noexperimenter effects on the number of mazes solved per round.

9 Unlike for L, the randomization in terms of grade in school is perfect for H across con-trol and identity treatments. For L, the mean grade in school is 6.53 for control (Caste NotRevealed) and 6.34 for the identity treatments and this difference is significant, as Table 1shows. Further disaggregating the data by treatment reveals that the problem of imbal-ance in grade for L lies with the control versus all other treatments, and that this is notso in the case of H. For H, for piece rate conditions, the means for grade in school (withstandard deviations in parentheses) are 6.53 (0.50) for Caste Not Revealed, 6.53 (0.50)for Revealed Mixed, and 6.69 (0.47) for Revealed Segregated. For the tournament condi-tions, the means are 6.47 (0.50) for Caste Not Revealed, 6.44 (0.50) for Revealed Mixed,and 6.43 (0.50) for Revealed Segregated. On the other hand, for L, while there is little dif-ference in the mean grade in school among the four identity treatments, the mean gradefor Caste Not Revealed is higher. For L, for piece rate conditions, the means are 6.57(0.50) for CasteNot Revealed, 6.37 (0.48) for RevealedMixed, and 6.30 (0.46) for RevealedSegregated. For the tournament conditions, the means are 6.43 (0.50) for Caste Not Re-vealed, 6.30 (0.46) for Revealed Mixed, and 6.39 (0.49) for Revealed Segregated.

for these factors in the analysis, and all results described in Section 1are robust to these controls. In particular, the results that the lowcaste underperforms when caste is revealed are robust to controllingfor grade in school.

Fig. 3 shows the average number of mazes solved by H. The figure isdivided into three blocks. Block 1 is round 1, block 2 is round 2-piecerate, and block 3 is round 2-tournament. In each block, H output is low-est in Revealed Segregated. Under theMann–WhitneyU-test, the differ-ences between Revealed Segregated and the control are significant atp b .05 in all blocks. In block 2, average output is higher in RevealedMixed than in the control, but the difference is not significant.

Fig. 4 superimposes on Fig. 3 the average outputs of L. All threeblocks show that when caste is not revealed, the average output of His almost the same as that of L. However, when caste is made public,the performance declines for L are generally steeper than those forH. A significant caste gap emerges in Revealed Mixed in both rounds.

Fig. 5 shows how the identity conditions impair L relative to H perfor-mance at the top of the performance distribution. The figure reports theratio of L participants to all participants with output at or above each

Years (0, 6) 7% 13%⁎ 22% 19%At least 6 years 86% 81% 52% 50%

Both parents illiterate 7% 4% 26% 29%Mother works outside thehome

4% 7% 7% 5%

Father is a day laborer 8% 9% 16% 19%Grade in school 6.51 6.51 6.53 6.34⁎

Previous exposure tomazes 7% 15%⁎ 4% 16%⁎

Mean number of otherparticipants known

0.55 1.14⁎ 0.56 1.03⁎

Note. This table looks at the balance between treatments inwhich caste identity is revealed(“Identity conditions”) and those in which it is not. Except for the last row, thecharacteristics reported in this table are binary. For example, “both parents illiterate” =1 if both parents have no formal education and otherwise it is zero; “previous exposureto mazes before” = 1 if the subject had seen mazes before and otherwise it is zero; andgrade in school is equal to either 6th or 7th. For binary variables, the tests of equality ofmeans across conditions for the high caste are based on logit regressions, one for eachcharacteristic; and similarly for the low caste. For “mean number of other participantsknown,” the test of equality of means is based on a t-test.⁎ Denotes rejection of the equality of means for the control and identity conditions at

p b 0.05.

image of Fig.�2

0

1

2

3

4

5

6

7

8

Ave

rag

e o

utp

ut

ROUND 1 ROUND 2TournamentPiece Rate Piece Rate

Fig. 3. Average output of high-caste participants. Note. Brackets indicate differences be-tween treatments with 95% confidence based on the Mann–Whitney U-test.

0

1

2

3

4

5

6

7

8

Ave

rag

e o

utp

ut

High Caste

Low Caste

ROUND 1 ROUND 2Piece Rate Piece Rate Tournament

Fig. 4.Average output of high-caste and low-caste participants.Note. Vertical lines indicatecaste gaps that are statistically significant based on the Mann–Whitney test with 95%confidence.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

10 20 30 40 50 60 70 80 90 100

Pro

po

rtio

n

Decile (10 is top decile)

Tournament Caste Not Revealed

Piece rate Caste Not Revealed

Piece rate Revealed Segregated

Tournament Revealed Mixed

Piece rate Revealed Mixed

Tournament Revealed Segregated

Fig. 5. Proportion of the low caste above each performance decile in round 2 (Cumulative).Note. There is, in general, more than one participant whose performance ranks him at theborder between two deciles. To explain the adjustment we make for this case, we use anexample. Let n(20) = the number of L in decile 20, n(30) = the number of L in decile 30,and n(b) = the number of L at the border between deciles 20 and 30. Finally, denote an Lwhose performance ranks him at the border between these twodeciles as a border person,Lb. Consider a case where n(20) = 2, n(30) = 1, and n(b) = 3. How do we allocate thethree border persons?We allocate them so that the proportion of border persons in decile


decile in round 2. (If H and L were equally represented in each decile andif varying caste salience had the same effect on H and L, then all points inthe figure would lie along the horizontal line at 0.5; i.e. in any cut of thedistribution, the proportion of L participants would be one-half.) The fig-ure shows that if the top 10% of participants were selected based on theirperformance in the control (CasteNot Revealed), Lwould be in themajor-ity. But if selectionwas based onperformance in Piece rate or TournamentRevealedMixed, Lwould be in theminority. And if selectionwas based onperformance in Piece rate Revealed Segregated, it would result in an equalrepresentation of H and L. Thus, whether H or L is overrepresented in thetop decile depends on the context in which the boys perform.

5. Measuring treatment effects

5.1. Number of mazes solved—full sample

We find patterns similar to those in Fig. 4 in regressions that controlfor individual and family characteristics. We pool all the observationsand allow for interactions among caste, context, and incentives.Table 2, columns (1)–(4), report OLS estimates,with robust standard er-rors clustered at the individual level, for the following specification:

Mazes solved in a round¼ α þω � round is 2ð Þ þ β � subject is Hð Þ þ γ � session cues identityð Þþδ � subject is H � session cues identityð Þ þ τ � Tournamentð Þþλ � Tournament � subject is Hð Þþξ � Tournament � session cues identityð Þþθ � Tournament � subject is H � session cues identityð Þ þ μ � Ζ þ error

ð1Þ

where Z is a vector of individual and family characteristics. The term αmeasures the predicted output in the omitted case: an L in Piece ratecontrol in round 1. The next eight coefficients (from ω to θ) measurethe effects of round, caste, and treatments and the two-way andthree-way interactions.10

10 For example, γ is a vector that measures the difference for L between an iden-tity condition (Revealed Mixed or Revealed Segregated) and the control, underpiece rate incentives. Using a subscript s for Revealed Segregated, α + ω + γs isthe predicted output of L in round 2 of Revealed Segregated under piece rate incen-tives. The predicted output of H in Revealed Segregated under tournament incen-tives is α + ω + β + γs + δs + τ + λ + ξs + θs.

One result is immediate. Low-caste boys solve mazes just as well ashigh-caste boys when caste is not revealed. The estimated coefficientson H and T ∗ H are always very small and insignificant. We get sensibleresults for round 2 and grade in school: both factors significantly im-prove round performance in every specification.

Specification (1) uses only caste and treatment indicators. Speci-fication (2) adds controls for individual characteristics: grade inschool, previous exposure to mazes, and number of other partici-pants known in a session. Specification (3) adds controls for family

20 is equal to the ratio n(b)/[n(20) + n(30) + n(b)] = 0.5, and the proportion of borderpersons in decile 30 is also equal to this ratio. Thus we allocate 2 border persons to decile20 and the remaining border person to decile 30. After the allocation, decile 20 has2L + 2Lb; the proportion of boundary Ls in the decile is .5, as required. And decile 30has 1L + 1Lb; the proportion of boundary Ls in the decile is also .5, as required. Intuitively,since there are twice asmany Ls in decile 20 as in decile 30,we allocate twice asmany bor-der persons to decile 20 as to decile 30. For H, the procedure is analogous.

Table 2OLS estimates of the determinants of output per round and output change between rounds.

Dependent variable Output per round Output changebetween rounds

Without individual andfamily characteristics

With individualcharacteristics

With individual andfamily characteristics

Excluding participantswho solved zero mazes

With individualcharacteristics

(1) (2) (3) (4) (5)

High caste (H) 0.29 0.16 0.35 0.56 0.25(0.35) (0.36) (0.39) (0.34) (0.42)

Round 2 2.14⁎⁎⁎ 2.17⁎⁎⁎ 2.27⁎⁎⁎ 2.33⁎⁎⁎

(0.15) (0.16) (0.16) (0.16)Revealed Mixed −0.70⁎⁎ −0.58 −0.51 −0.07 −0.54

(0.34) (0.37) (0.38) (0.35) (0.39)Revealed Segregated −0.97⁎⁎⁎ −0.93⁎⁎ −0.74 −0.70⁎ −0.86⁎⁎

(0.37) (0.40) (0.46) (0.40) (0.43)Tournament (T) 1.40⁎⁎ 1.45⁎⁎ 1.44⁎⁎ 1.28⁎ 1.06⁎

(0.65) (0.66) (0.66) (0.66) (0.55)Revealed Mixed ∗ H 0.75 0.73 0.65 −0.12 0.64

(0.48) (0.50) (0.53) (0.47) (0.60)Revealed Segregated ∗ H 0.02 −0.01 −0.16 −0.52 −0.64

(0.54) (0.58) (0.65) (0.56) (0.64)T ∗ H −0.26 −0.12 −0.14 −0.04 −0.44

(0.89) (0.90) (0.96) (0.86) (0.77)Revealed Mixed ∗ T −1.35⁎ −1.59⁎⁎ −2.02⁎⁎⁎ −1.48⁎ −1.02

(0.76) (0.78) (0.78) (0.77) (0.69)Revealed Segregated ∗ T −2.77⁎⁎⁎ −3.05⁎⁎⁎ −3.02⁎⁎⁎ −2.82⁎⁎⁎ −1.38⁎

(0.76) (0.77) (0.82) (0.77) (0.76)Revealed Mixed ∗ T ∗ H −0.07 0.02 0.67 −0.16 −0.26

(1.08) (1.11) (1.20) (1.08) (1.00)Revealed Segregated ∗ T ∗ H 1.73 1.73 1.91 1.92⁎ 2.56⁎⁎

(1.14) (1.21) (1.33) (1.16) (1.05)Grade in school 0.43⁎⁎ 0.51⁎⁎ 0.45⁎⁎ 0.34

(0.21) (0.23) (0.21) (0.21)Previous exposure to mazes 0.37 0.51 0.35 −0.19

(0.30) (0.33) (0.29) (0.36)Number of participants 0.06 0.10 0.01 0.02known (0.09) (0.09) (0.08) (0.09)Mother's education (0, 6) 0.28

(0.30)Mother's education ≥ 6 0.44

(0.33)Father's education (0, 6) −0.64⁎

(0.39)Father's education ≥ 6 −0.91⁎⁎⁎

(0.34)Mother employed outside home 0.05

(0.53)Father not a day laborer 0.55

(0.35)Constant 3.26⁎⁎⁎ 2.97⁎⁎⁎ 2.76⁎⁎⁎ 2.98⁎⁎⁎ 2.16⁎⁎⁎

(0.24) (0.28) (0.50) (0.28) (0.32)R2 0.189 0.197 0.221 0.223 0.080n 1164 1076 928 1008 538

Notes. Standard errors in parentheses are robust to heteroskedasticity, and observations are clustered at the level of the individual. The omitted case is L in Caste Not Revealed under piecerate incentives. Column (4) excludes participants whohave zero output in both rounds. Round 2 is a dummy variable that equals 1 for round 2 and zero for round 1. Grade in school = 1 ifthe participant is in grade 7, 0 if he is in grade 6. Previous exposure to mazes = 1 if some time before the experiment, the participant had seen mazes; 0 otherwise. Number of other par-ticipants known is the number of others in the experimental session known to a given participant.⁎⁎⁎ p b 0.01.⁎⁎ p b 0.05.⁎ p b 0.10.


characteristics. In this section we will analyze the results under thepiece rate incentive; we will analyze the results under the tourna-ment incentive in Appendix A.

Between specifications (1) and (2), there is only one change in theset of significant treatment effects: the output decline by L in RevealedMixed is no longer significant. Table 3 reports the treatment effects inthe first two columns. The treatment effects of Revealed Mixed foreach caste (0.16 for H and −0.58 for L) are individually insignificantbut jointly produce a significant caste gap in favor of H. The treatmenteffect of moving from the control to Revealed Segregated is a significant

decline of −0.93 mazes for both H and L. This effect is roughly doublethe effect on output of being in 6th instead of 7th grade (=−0.43, asshown in Table 2).

In order to checkwhether the channel throughwhich social identityinfluences behavior is class—a poor-versus-rich effect on performance,as in Croizet and Claire (1998)—rather than caste, we next considerthe effect of controls for family characteristics. In Table 2, column (3),we control for parents' education, mother's employment outside thehome, and father's employment as a day laborer. Because stigma is asso-ciated with daily wage-labor, our post-play survey did not ask, “Is your

Table 3Effects of making identity public in mixed and segregated sessions under the piece-rate incentive.

Output per round, full sample Output per round, excludingparticipants who solved zero mazes

Output change between rounds, full sample

H L Caste gapsignificant



(1) (2) (3) (4) (5) (6) (7) (8) (9)

Estimated treatment effect ofRevealed Mixed 0.16 −0.58 ⁎⁎ −0.19 −0.07 0.10 −0.54 ⁎⁎

(0.36) (0.37) (0.34) (0.35) −0.46 −0.39Revealed Segregated −0.93⁎⁎ −0.93⁎⁎ −1.23⁎⁎⁎ −0.70⁎ −1.50⁎⁎⁎ −0.86⁎⁎

(0.42) (0.40) (0.41) (0.40) (0.49) −0.43Mean outcome in Caste Not Revealed 4.22 4.06 4.71 4.15 2.41 2.16Effect of Revealed Segregated as % of

mean in Caste Not Revealed−22% −23% −26% −17% −62% −40%

Notes. Cluster-robust standard errors are in parentheses. The mean output per round pools both rounds. The treatment effects can be derived from the regressions in Table 2. Effects incolumns (1)–(3) are from regression (2); those in columns (4)–(6) are from regression (4); those in columns (7)–(9) are from regression (5). However, it is easier to obtain the effectsand standard errors for H by running regressions in which H is the benchmark caste.⁎⁎⁎ p b 0.01.⁎⁎ p b 0.05.⁎ p b 0.1.


father a day laborer?” Instead, the survey asked about the father's occu-pation. We formed a binary variable for daily wage labor based on theresponses. Adding controls for class reduces thedeclines in performancein Revealed Segregated. The treatment effect ofmoving from the controlto Revealed Segregated ismarginally significant for H (−0.90, p b 0.10),not for L (−0.74, p = 0.11). However, two key results are robust toadding controls for class. First, merely making caste public without seg-regating H from L does not impair H performance (the effect is0.14 = − .51 + .65 and insignificant). Second, H and L are equallygood at solving mazes when caste is not revealed, whereas in RevealedMixed a significant caste gap favoring H emerges (=0.35 + 0.65 = 1.0;p = 0.01).

We have emphasized specification (2) in Table 2 more than spec-ification (3) because we cannot reject the hypothesis that parentalvariables have no effect on performance: F(6,486) = 1.58, p-value =0.11. The only proxy for class that is individually significant is father'seducation, and its effect is not in the direction that is predicted by thehypothesis that class stigma impedes performance.

Itmight be, however, that parental variablesmatter for L, but not H, be-cause having educated parents alleviates low-caste stigma. Therefore, inunreported regressions, we reran specification (3) separately for H and Lparticipants. We still find that parental variables have little explanatorypower and are insignificant by an F-test. We also checked for the effect ofhaving two illiterate parents. We find that this is not significant (resultnot shown). In these and all other regressions that we have run, we findno evidence that class is the channel through which caste influences be-havior. However, since we do not have measures of income and wealth,the concern that unobserved class variables may matter remains.

5.2. Between-round change in the number of mazes solved

As an additional check on our results, we consider the treatment ef-fects on the change in output between rounds. This is shown in the lastthree columns in Table 3. In Piece-rate control, H and L are significantlyimproving their skills across rounds (2.17 mazes, p b .01). RevealedSegregated reduces output change between rounds for H by 62%(p b .01) and for L by 40% (p b .05).

11 Otherwise the estimates are unbounded, since some cells in Table 4 are empty.

5.3. Success or failure in learning how to solve a maze

In the remainder of this section, we decompose performance intotwo stages:

Stage 1: Learning a new skill. The participant learns what it means tosolve amaze. The outcome is binary—success or failure.Wemeasurefailure by zero output by a participant during the 30 min of maze-solving.Stage 2. Applying the skill. In this stage, the outcome is the number ofmazes solved conditional on success in learning how to solve amaze.

Table 4 reports the failure rate by identity condition, treatment,and caste. For H in P/P, failure is greater in the control (9%) than inthe two identity conditions (2% and 3%). However, the effect of re-vealing caste is not consistent across the P/P and P/T treatments. InP/T 7% of H (2/30) fails in the control, 0% (0/60) fails in RevealedMixed, and 10% (3/30) fails in Revealed Segregated. Since the exper-imental conditions in P/P and P/T were identical until round 2, andsince if caste was revealed, it happened at the beginning of a session,it is reasonable to pool P/P and P/T within the Revealed Mixed condi-tion andwithin the Revealed Segregated condition.Whenwe do this,a pattern emerges in which revealing caste decreases the failure rateamong H: failure is greater in the control (8% or 9/108) than in eitherRevealed Mixed (2% or 2/120) or Revealed Segregated (7% or 4/60).For L, the pattern is the reverse: failure is smaller in the control (2%or 2/108) than in either RevealedMixed (13% or 15/120) or RevealedSegregated (9% or 6/66).

To fit a logit model, it is necessary to collapse the two identity condi-tions and also the two incentive conditions.11 We estimate:

Failure ¼ αþ β � subject is Hð Þ þ γ � session cues identityð Þþδ � subject is H � session cues identityð Þ þ μ � Ζ þ error;

ð2Þ

where Z are individual characteristics. The benchmark case is L, CasteNot Revealed. We use the logit results, reported in SupportingTable S2, to predict the probability of failure. Fig. 6 presents the re-sults. The figure shows that revealing caste reduces failure amongH from 8% to 2%, and increases failure among L from 1% to 11%, con-trolling for individual characteristics. The treatment effects are sig-nificant and robust to the addition of controls for householdcharacteristics. The effects are consistent with the predictions of ste-reotype susceptibility: when participants are made more aware ofcaste, H are less likely and L are more likely to fail to learn how tosolve a maze.

Fig. 6. Predicted probability of failure. Note. Based on the logit regression in SupportingTable S2, column (1). The control variables are grade in school, previous exposure tomazes, and number of other participants known in the session. The predicted probabilitiesare estimated at the means of the control variables.

Table 4Proportion of participants with zero output.

Treatment Number of participantswith zero output/Totalnumber of participants ofthe respective caste inthe treatment

Proportion

High caste Low caste High caste Low caste

P/P-Caste Not Revealed 7/78 2/78 0.09 0.03P/P-Revealed Mixed 1/60 9/60 0.02 0.15P/P-Revealed Segregated 1/30 2/30 0.03 0.07P/T-Caste Not Revealed 2/30 0/30 0.07 0P/T-Revealed Mixed 0/60 6/60 0 0.10P/T-Revealed Segregated 3/30 4/36 0.10 0.11


5.4. Number of mazes solved by the subsample excluding non-learners

An advantage of decomposing performance into stages is that wecan consider treatment effects on performance conditional on know-ing how to solve a maze. We report the effects in Table 3, columns(4)–(5). Are the qualitative results for the full sample robust in thesubsample?

5.4.1. The high casteThe treatment effects of making caste public are stronger in the sub-

sample than in the full sample. The reason is that in the subsample, weare not capturing the stage 1 effect, in which making caste public re-duces the probability that H will fail to learn how to solve mazes. Inthe full sample the treatment effect of Revealed Segregated is −0.93(p b 0.05), compared to −1.23 (p b 0.01) in the subsample. We viewthis latter figure (−1.23), which is a 26% decline in performance, asour best estimate of the entitlement effect. By this we mean the effecton a high-caste boy's output, conditional on his knowing how to solvea maze, of moving from the control condition (where the authority fig-ure is silent about caste) to Revealed Segregated (where segregationcues high-caste boys' place in the social order). We call it the entitle-ment effect because we conjecture that the treatment effect reflects areduced need to achieve in a situation that cues one's place in the tradi-tional, ascriptive social order. The entitlement effect on H is thus about

the same size as the performancedecline by L that could be attributed tostereotype threat (=−23%; see column 2). These treatment effects arelarge. They are the same order of magnitude—but of opposite sign—asthe effect of switching from the piece rate to the winner-take-all tour-nament in the control condition (25% for H and 28% for L).

Besides the entitlement effect, other hypotheses might explain thedecline in H performance in Revealed Segregated. The first is a perversekind of stereotype susceptibility. Such an effect occurred in one of thetwo experiments in Shih et al. (2002). They compared amild versus bla-tant activation of a positive stereotype. They found in one experimentthat priming Asian identity had no effect on Asian-Americans' perfor-mance on amath test, and in the second experiment that it significantlyimpaired performance, perhaps by creating anxiety in participants'minds that their performance would not confirm the high expectationsfor Asians in the US. But our finding that the identity conditions in ag-gregate reduced the proportion of individuals who failed to learn howto solve a maze does not support the view that the identity conditionsincreased anxiety.

Second, because the caste order is always contested, it could be thatin Revealed Mixed, H feel a need to demonstrate (even if only to them-selves) their superiority, and that they do not feel this need in RevealedSegregated. Ultimately, the second hypothesis comes down to largelythe same thing as our preferred one, namely, that contexts that reinforcethe complacency of the high caste in their superior status induce them tovalue less the rewards from individual achievement.

5.4.2. The low casteNext consider the treatment effects for L in the subsample com-

pared to the full sample (Table 3, columns (2) and (5)). We findthat making caste public reduces output less in the subsample be-cause in the subsample, we are not capturing the stage 1 effect, inwhich making caste identity public increase the probability that Lwill fail to learn how to solve amaze. In the subsample, the treatmenteffect of Revealed Segregated is no longer significant under piecerates. This suggests that under piece rate incentives, the identity condi-tions impair L performance primarily by reducing their success at learn-ing the new task (maze-solving).

6. Further evidence

In this section, we discuss evidence that bears on the mental framesthat Revealed Segregated evokes in high-caste and low-caste boys.

6.1. Revealed Segregated reduces the self-confidence of the low-caste boys

In an earlier experiment (Hoff and Pandey, 2005), we used ran-dom assignment of participants to the three conditions of castesalience (Caste Not Revealed, Revealed Mixed, and Revealed Segre-gated) to assess the relationship between caste salience and self-confidence. In a six-person session, H and L were taught how tosolve a wooden puzzle that we had constructed along the lines ofthe game Rush-Hour Traffic Jam. At the end of the session, partici-pants had to make a choice between a sure payoff and a lotterywith a high payoff if the individual solved a new puzzle successfullyand zero otherwise. In choosing the lottery, a participant was bettingon his own success. The outcome is thus a test of self-confidence. Theresults showed no significant caste gap in the acceptance rate of thelottery in both the Caste Not Revealed and Revealed Mixed condi-tions. In contrast, in Revealed Segregated, there was a large and a sig-nificant caste gap in the proportion that accepted the lottery whenthe puzzle was difficult and the judge had some discretion in evalu-ating a player's success: 70% of H compared to only 33% of L acceptedthe lottery (p = .004). This evidence supports the hypothesis thatRevealed Segregated evokes a mental model in which a low-casteboy may feel, “I can't or don't dare to excel.”

Fig. A1. Predicted output in round 2: piece rate versus tournament incentives. Note. Predicted output is based on Table 4, column (3), which controls for the participant's grade in school,previous exposure to mazes, and number of other participants known in the session. Bars show standard errors.


6.2. Observational studies of school enrollment point to a causal effect ofcaste salience on low school enrollment by both high- and low-castestudents

Consider again our conjecture that when the social division betweenhigh and low castes is salient, a high-caste boy may think, “I don't needto excel.” If this is correct, it would predict that when caste boundariesare eroded, a high-caste boy (or his family on his behalf) may think, “Ineed to work harder to better myself.” This prediction is consistentwith the findings in Kochar (2004). She analyzes an Indian governmentpolicy to construct schools in low-caste hamlets. The policy led to an in-crease in low-caste enrollment. The increased school enrollment of low-caste children had, however, an unintended effect: it increased the en-rollment rate of the upper castes! Thus, aid targeted to Dalits did not,as policy-makers had expected, narrow the schooling gap between theDalits and the rest of society. The increase in high-caste enrollmentmaintained the relative superiority of the high caste in years of educa-tion.12

Finally, an observational study in Pakistan identifies a causal impactof high-caste dominance on low enrollment of low-caste children inschool. This finding is consistent with our conjecture that cues to thecaste order evoke a mental frame that impairs low-caste individuals'ability or desire (or both) to achieve in the classroom. Jacoby andMansuri (2011) analyze data from a survey of over 3000 householdsand 1000 elementary schools in rural Pakistan. They define a settlementas high-caste-dominant if a high caste owns the majority of land in thesettlement, giving them the power to enforce the traditional casteorder. They show that low-caste children are deterred from enrollingin schools in high-caste dominant settlements. They find that the verylow enrollment of low-caste children13 for whom the closest availableschool is in a high-caste-dominant hamlet can account for the entire en-rollment gap favoring high-caste over low-caste children. The followingresponses from low-caste women to the question, “Do children receivethe same treatment from teachers?” illustrate the kinds of exclusionarynorms imposed on low-caste individuals:

“They let the daughters of [high castes] use the latrines, but tell ourdaughters to use the fields because you stink.” “The teachers makethe daughters of Zamindar Zaats [high castes] sit inside the rooms,

12 We thank Anjini Kochar for bringing her work to our attention.13 Especially girls, for whom honor considerations restrict the freedom to travel outsidetheir own hamlet.

under the fans. Our poor children are outside, under the sun anddust” (Jacoby and Mansuri, p. 7).

These two observational studies—the only studies of which we areaware that examine the effect on achievement of increasing or decreas-ing the salience of the caste order—indicate that when one's traditionalstatus in the social order seemsmore fixed, both high- and low-caste in-dividuals are less likely to enroll in school.

7. Conclusion

The experimental data show that being hard-working or clever isnot a trait of the person, in the sense that it is always there, in a fixedmanner. Instead, situational cues to one's place in the social order influ-ence the expression of these traits. Cues to caste identity inmixed-castegroups depressed the ability of the lowcaste to learn a new skill. Segre-gation by caste status did not impair the ability of high-caste boys tolearn a new skill, but sharply reduced their performance in using thenew skill. The influence of identity on performance suggests the useful-ness of the theories that posit a frame-dependent self. In this view, con-text may affect performance by changing not what it is objectivelyappropriate to do, but instead by changing the set of associations thatare elicited and the mental model through which the individual inter-prets the situation. Borrowing from the title of a recent book, Framedby Gender (Ridgeway, 2011), one might say that our subjects wereframed by caste. Our findings provide evidence of one possible sourceof the caste gap in school performance in north India. Pandey et al.(2010) find that after controlling for observed family characteristicsand school fixed effects, high-caste students have significantly highertest scores than low-caste students in rural public primary schools inthe Indian states of Uttar Pradesh and Madhya Pradesh, where casteand class hierarchies run deep.

Our findings contribute to a richer view of institutional change thanone that follows from the standard definition of institutions-as-rules.Besides rules, institutions also entail social identities and worldviews(see Greif, 2006, Section 2.16). Social identities made salient by situa-tional cues—for example, by frockcoats and tight-laced corsets, racenames, or caste names—have an independent influence on “making uppeople.” After the rules of an institution have been abolished and thestructure of power that underpinned them has changed, the identitiesfounded on the superseded rules will continue to affect chronic waysthat people think about themselves and interpret the world unlessprevailing situations make those identities less salient. For example,when towns emerged in medieval Europe, “[i]t was not only that the


serf, having escaped from the countryside, found legal freedom in thetown, but that the whole social atmosphere there was open to ambitionand talent” (Cipolla, 1994, pp. 119). The new possibilities for upwardmobility depended on change in the social as well as the legal environ-ment. No one would argue with this, although these ideas rarely enterinto economic models.

Rural India is in transition between a feudal and a capitalist econom-ic system. Our findings suggest that boys in India hold, in uneasy ten-sion, a traditional worldview in which caste is destiny and a modernworldview in which each person shapes his own destiny. Wemeasuredthe effects on children's performance of manipulating the publicnessand salience of the traditional caste identities. Our findings are consis-tent with the hypothesis that changing the publicness and salience ofcaste affectswhich of the twoworldviews is uppermost in the children'sminds.

We have argued that it is a useful construct to posit that an individ-ual has multiple preferences, one for each of his mental models orworldviews, because this perspective opens up a new set of policy op-tions for enhancing human capital formation and development. It isalso useful for understanding long-run social change, which entailschanges in the salience of particular identities, the set of possible iden-tities, and the possible ways of understanding a situation. This perspec-tive highlights the relevance to economics of understanding how the setof possible identities (and not merely the possible technologies and thestock of resources) evolves in a society. Recentwork in economics takesup this exciting question (e.g.Greif and Laitin, 2004; Greif and Tabellini,2012; Hoff and Sen, 2006; Hoff and Stiglitz, 2010; Munshi andRosenzweig, 2006).

Supplementary data to this article can be found online at http://dx.doi.org/10.1016/j.jdeveco.2013.08.009.

Acknowledgments

Over the course of many years, we have benefitted from veryhelpful comments from many people. We would particularly like tothank Rachel Croson, Allison Demeritt, Anjini Kochar, Leigh Linden,Tauhidur Rahman, Vijayendra Rao, Joe Stiglitz, and Ann Swidler.We also thank the referees. We owe a special debt to Anaka Narayanan,RamPratap, andMayuresh Kshetramade for assistancewith data collec-tion and to Shweta Arya, Sonal Vats, and Sam Zhongxia Zhang for re-search assistance. This work was made possible by grants from theWorld Bank-Netherlands Partnership Program, the World Bank Re-search Support Budget, and the MacArthur Foundation Research Net-work on Inequality and Economic Performance.

Table A1Effects of making identity public in mixed and segregated sessions under the tournament ince

Output per rou

H L

(1) (

Estimated treatment effect ofRevealed Mixed −1.42⁎

(0.79)Revealed Segregated −2.25⁎⁎

(0.92)Mean outcome in Caste Not Revealed 6.63Effect of Revealed Mixed as % of mean outcome in Caste Not Revealed −21% −Effect of Revealed Segregated as % of mean outcome in Caste Not Revealed −34% −

Note. Cluster-robust standard errors are in parentheses. The mean output per round poolseffects in columns (1)–(3) from regression (2), and effects in columns (4)–(6) from regresby running regressions in which the tournament is the incentive in the baseline case. ⁎⁎⁎

Appendix A. Analysis of the tournaments

Recall from Fig. 1 that the experiment used two different financialincentives. Some treatments used piece rate incentives in both rounds(“P/P treatments”), and others used piece rate incentives in round 1and tournament incentives in round 2 (“P/T treatments”). Under thetournament incentive, a participant earned six rupees per maze solvedif he solved themostmazes in his session; otherwise he earned nothing.A winner could (and some did) earn 15 × 6 rupees, equivalent to al-most two days' unskilled adult wages. In case of a tie, both winners re-ceived the prize.

In piece rates, payoffs depend only on own performance. In contrast,revealing the caste identity of the other participants in a tournamentmight affect beliefs about the individual's own chances of winning thetournament. In addition, if a participant knows that he is competingagainst members of his own community, he might not wish to bestthem in the game. Thus in analyzing the results in this section, thereare many more possible influences on behavior, and we cannot disen-tangle them. We first report all treatment effects, and then discuss thepossible explanations.

Results

When caste is not revealed, the shift from the piece-rate to the tour-nament incentive significantly increases performance. In Table 2, col-umns (1)–(4), all coefficients on tournament are significantly positive.The table also shows that in all specifications the coefficient on T ∗ His insignificant. This means that the response of H to tournament incen-tives is statistically indistinguishable from that of L.

Revealing caste, whether in mixed or segregated groups, eliminatesthe performance boost from tournament incentives. Fig. A1 graphs thepredicted output in round 2 under all conditions (two incentiveconditions × three identity manipulations). The left panel shows re-sults for H, and the right panel shows results for L. As indicated by thedotted lines, when caste is not revealed, the tournament boosts outputby 25% for H and 28% for L above performance under piece rates(p b .05). As shown by both the dark and light solid lines, when casteis made public, there is no performance boost for H or L. In fact, in Re-vealed Segregated the tournament incentive reduces L output. The de-cline is 1.6 mazes (p-value b 0.01), which corresponds to a 38%decline from the predicted level under Piece-rate control. These resultsare robust to controls for class—see Table 2, column 3.

Table A1 shows the effect of the identitymanipulations, holding con-stant the tournament incentive. The first three columns show the re-sults for the full sample, and columns 4–6 show the results for the

ntive.

nd, full sample Output per round, excluding participants whosolved zero mazes

Caste gap significant H L Caste gap significant

2) (3) (4) (5) (6)

−2.17⁎⁎⁎ −1.82⁎⁎ −1.54⁎⁎

(0.77) (0.75) (0.77)−3.97⁎⁎⁎ ⁎ −2.13⁎⁎ −3.52⁎⁎⁎ ⁎⁎

(0.75) (0.86) (0.75)6.59 7.11 6.59

33% −26% −23%60% −30% −53%

both rounds. All values reported here can be derived from the regressions in Table 2:sion (4). However, it is easier to obtain the treatment effects and their standard errorsp b 0.01, ⁎⁎p b 0.05, and ⁎p b 0.1.




subsample that excludes participants who solved zero mazes. It is easyto see that in the full sample, the treatment effect of revealing caste isalways negative. The declines are much steeper for L than for H. For ex-ample, compared to the control, Revealed Segregated decreases outputby 34% for H and by 60% for L. The treatment effects are only slightly re-duced in the subsample.

Discussion

In the tournaments, we cannot disentangle context effects on perfor-mance from strategic responses to equilibrium play. The finding thatmaking caste public eliminates the positive response to tournament in-centives is consistent with an interpretation based on context effects:i.e. when caste is salient, H may feel, “Why try?” and L may feel, “Ican't, or don't dare to, excel.” But there are three additional possible ex-planations, which we discuss next.

Strategic responses to equilibrium playAs an example, suppose wemake the plausible assumption that H

believe that low-caste boys will choose not to compete because foran L to win against an H would violate the traditional caste order.Or suppose we assume that H believe that low-caste boys cannotcompete effectively since they are less able than high-caste boys.Under either assumption, the H-segregated tournaments wouldhave low expected returns to effort compared to the control tourna-ments because the participants know they are competing against fiveH boys.

Social preferencesThe 30 high-caste boys in Revealed Segregated were 20 Thakurs, 8

Brahmins, and 2 Vaishyas. The overall distribution in the treatmentwas mirrored in the individual sessions. Thus, a high-caste boy in Re-vealed Segregated did not find himself among boys of only his own spe-cific caste. Earlier experiments in the state of Uttar Pradesh revealsolidarity only among men of the same specific high-status caste (Hoffet al., 2011), not among men of different specific high-status castes.The experiments also reveal no solidarity among low-caste men, evenwhen they are from the same caste. Evidence fromFehr et al. (2008) sup-ports the view that many high-caste men are averse to disadvantageousinequality. In a tournament, someone must win, and we have no evi-dence from any study of social preferences that if someone must win, ahigh-caste man would not wish to be the person that won.

Work normsBandiera et al. (2005) find evidence from personnel data that

workers' productivity is much lower under a relative incentive schemethan under piece rates. In the relative incentive scheme in their study,workers' daily pay depended on the ratio of individual productivity toaverage productivity among all coworkers on the same field and day;whereas under piece rates, individual pay depended only on individualproductivity. Thus the harder a person worked, the lower the pay of hiscoworkers. The authors find that workers internalize the externality ofhigh productivity by working less only when they can monitor othersand be monitored by them. Thus the finding reflects work normsenforced by sanctions, not social preferences.

It seems to us unlikely that work norms explain the declines inperformance in tournaments when caste is revealed. First, the tour-nament incentive is a one-shot, winner-take-all incentive. The per-son with the best performance wins, and all others lose. Byworking less, a participant does not increase the earnings of his co-workers, as would occur under the scheme analyzed in Bandieraet al., unless he completely forgoes a payoff. Further, Bandiera et al.find that the decline occurs only when workers can monitor eachother. In our experiment, payoffs were not public: each participantwas given his payment such that others did not know the amountat the time of giving, and instructions prior to the game (“This [the

earnings from the piece rate incentive in round 1 and from tourna-ment in round 2] will be given to every child in an envelope afterthe games are over”) tried to make this clear. However, it is possiblewe did not make it clear enough. We cannot rule out that partici-pants feared social sanctions from others in their group if they wonthe tournament.

We comment finally on the question of work norms in Hardoi. Thehigh castes in the villages shun positions in which theywork for others;this is considered a violation of their honor. High-caste individualsworkfor others only if they leave the village to work in the towns in otherdistricts; there is essentially no industry in Hardoi district. The issue ofwork norms in firms thus does not arise for either high or low castesin Hardoi.

References

Afridi, Farzana, Xin Li, Sherry, Ren, Yufei, 2010. Social Identity and Inequality: The Impactof China's Hukou System. Manuscript University of Texas-Dallas.

Akerlof, George, Kranton, Rachel, 2000. The economics of identity. Q. J. Econ. 115 (3),715–753.

Akerlof, George, Kranton, Rachel, 2002. Identity and schooling: some lessons for the eco-nomics of education. J. Econ. Lit. 40 (4), 1167–1201.

Alter, Adam, 2013. Drunk Tank Pink and Other Unexpected Forces that Shape how WeThink, Feel, and Behave. Penguin Press, New York.

Ambady, Nalini, Shih, Margaret, Kim, Amy, Pittinsky, Todd, 2001. Stereotype susceptibilityin children: effects of identity activation on quantitative performance. Psychol. Sci. 12(5), 385–390.

Bandiera, Oriana, Barankay, Iwan, Rasul, Imran, 2005. Social preferences and the responseto incentives: evidence from personnel data. Q. J. Econ. 120 (3), 917–962.

Begley, Sharon, 2007. Train Your Mind, Change Your Brain. Ballantine Books, New York.Bellmore, Amy, Tomonaga, Ayako, 2009. When Kids Use Ethnicity and Gender to Bully.

Web http://www.education.com/reference/article/ethnicity-gender-bullying/.Benjamin, Daniel J., Choi, James J., Joshua Strickland, A., 2010. Social identity and preferences.

Am. Econ. Rev. 100 (4), 1913–1928.Béteille, André, 2011. The André Béteille Omnibus. Oxford University Press, Delhi.Cassan, Guilhem, 2011. Law and Identity Manipulation: Evidence from Colonial Punjab.

Paris School of Economics Manuscript.Cohn, Alain, Michel, André Maréchal, Noll, Thomas, 2013. Saliency of Criminal Identity

Causes Dishonest Behaviour: an Experiment Behind Bars. University of ZurichManuscript.

Cipolla, Carlo M., 1994. Before the Industrial Revolution: European Society and Economy1000–1700, 3rd edition. W.W. Norton & Company, New York.

Connerton, Paul, 1989. How Societies Remember. Cambridge University Press, Cambridge.Croizet, Jean-Claude, Claire, Theresa, 1998. Extending the concept of stereotype threat to

social class: the intellectual underperformance of students from low socioeconomicbackgrounds. Pers. Soc. Psychol. Bull. 24 (6), 588–594.

Danziger, Shai, Ward, Robert, 2010. Language changes implicit associations between eth-nic groups and evaluation in bilinguals. Psychol. Sci. 21 (6), 799–800.

Déliege, Robert, 1999. The Untouchables of India. Berg Publishers, Oxford, UK.DiMaggio, Paul, 1997. Culture and cognition. Annu. Rev. Sociol. 23 (1), 263–287.Fang, Hanming, Loury, Glenn C., 2005. Dysfunctional identities can be rational. Am. Econ.

Rev. 95 (2), 104–111.Fehr, Ernst, Hoff, Karla, Kshetramade, Mayuresh, 2008. Spite and development. Am. Econ.

Rev. Pap. Proc. 98 (2), 494–499.Ferguson, Ronald F., 2003. Teachers' perceptions and expectations and the black–white

test score gap. Urban Educ. 38 (4), 460–507.Fryer Jr., Roland G., 2011. The importance of segregation, discrimination, peer dy-

namics and identity in explaining trends in the racial achievement gap. In:Benhabib, Jess, Bisin, Alberto, Jackson, Matthew O. (Eds.), Handbook of SocialEconomics. North-Holland, Amsterdam, pp. 165–1192.

Fryer Jr., Roland G., Levitt, Steven D., List, John A., 2008. Exploring the impact of financialincentives on stereotype threat: evidence from a pilot study. Am. Econ. Rev. Pap. Proc.98 (2), 370–375.

Gneezy, Uri, Niederle, Muriel, Rustichini, Aldo, 2003. Performance in competitive environ-ments: gender differences. Q. J. Econ. 118 (3), 1049–1074.

Greif, Avner, 2006. Institutions and the Path to the Modern Economy: Lessons from Me-dieval Trade. Cambridge University Press, Cambridge.

Greif, Avner, Laitin, David, 2004. A theory of endogenous institutional change. Am. Polit.Sci. Rev. 98 (4), 14–48.

Greif, Avner, Tabellini, Guido, 2012. The Clan and the City: Sustaining Cooperation inChina and Europe. Working Paper 445 IGIER.Bocconi University.

Gupta, Dipankar, 2000. Interrogating Caste: Understanding Hierarchy and Difference inIndian Society. Penguin Books, New Delhi.

Hacking, Ian, 1986. Making up people. In: Heller, T.C., Sosna, M., Wellbery, D.E. (Eds.),Reconstructing Individualism. Stanford University Press, Stanford.

Hoff, Karla, Pandey, Priyanka, 2005. Opportunity is not everything: how belief systemsand mistrust shape responses to economic incentives. Econ. Transit. 13 (2), 445–472.

Hoff, Karla, Pandey, Priyanka, 2006. Discrimination, social identity, and durable inequal-ities. Am. Econ. Rev. Pap. Proc. 96 (2), 206–211.

Hoff, Karla, Sen, Arijit, 2006. The kin system as a poverty trap? In: Bowles, Samuel, Durlauf,Steven, Hoff, Karla (Eds.), Poverty Traps. Princeton University Press, pp. 95–115.

http://refhub.elsevier.com/S0304-3878(13)00124-7/rf0275














http://www.education.com/reference/article/ethnicity-gender-bullying/



















































Hoff, Karla, Stiglitz, Joseph E., 2010. Equilibrium fictions: a cognitive approach to societalrigidity. Am. Econ. Rev. Pap. Proc. 100 (2), 141–146.

Hoff, Karla, Kshetremade, Mayuresh, Fehr, Ernst, 2011. Caste and punishment: the legacyof caste culture in norm enforcement. Econ. J. 121 (556), F449–F475.

Human Rights Watch, 1999. Broken People: Caste Violence against India's “Untouchables”.Human Rights Watch, New York.

Jacoby, Hanan,Mansuri, Ghazala, 2011. Crossing Boundaries: Caste, Stigma, and Schoolingin Rural Pakistan. WPS 5710.The World Bank, Washington, DC.

Kapur, Devesh, Prasad, Chandra Bhan, Pritchett, Lant, Shyam Babu, D., 2010. Rethinking in-equality: Dalits in Uttar Pradesh in the market reform era. Econ. Polit. Wkly. 45 (35),39–49.

Kochar, Anjini, 2004. Reducing Social Gaps in Schooling: Caste and the Differential Effectof School Construction Programs in Rural India. Manuscript Stanford University.

Krendl, Anne C., Richeson, Jennifer A., Kelley, William M., Heatherton, Todd F., 2008. Thenegative consequence of threat: a functional magnetic resonance imaging investiga-tion of the neural mechanism underlying women's underperformance in math.Psychol. Sci. 19 (2), 168–175.

Lambarraa, Fatima, Riener, Gerhard, 2012. On the Norms of Charitable Giving in Islam: AField Experiment. DICE Discussion Paper 59.Heinrich-Heine-Universitat, Dusseldorf.

LeBoeuf, Robyn A., Shafir, Eldar, Bayuk, Julia B., 2010. The conflicting choices of alternatingselves. Organ. Behav. Hum. Decis. Process. 111 (1), 48–61.

Marriott, Alan, 2003. Dalit or Harijan? Self-naming by scheduled caste interviewees. Econ.Polit. Wkly. 38 (36), 3751–3752.

Munshi, Kaivan, Rosenzweig, Mark, 2006. Traditional institutions meet the modernworld: caste, gender, and schooling choice in a globalizing economy. Am. Econ. Rev.96 (4), 1225–1252.

Munshi, Kaivan, Wilson, Nicholas, 2008. Identity, Parochial Institutions, and CareerDecisions: Linking the Past to the Present in the American Midwest. ManuscriptBrown University.

Ogbu, John U., 1999. Beyond language: Ebonics, proper English and identity in a Black-American speech community. Am. Educ. Res. J. 36 (2), 147–184.

Ogunnaike, Oludamini, Dunham, Yarrow, Banaji, Mahzarin R., 2010. The language of im-plicit preferences. J. Exp. Soc. Psychol. 46, 999–1003.

Pandey, Priyanka, 2010. Service delivery and corruption in public services: does historymatter? Am. Econ. J Appl. Econ. 2, 190–204.

Pandey, Priyanka, Goyal, Sangeeta, Sundararaman, Venkatesh, 2010. Public participation,teacher accountability and school outcomes in three states. Econ. Polit. Wkly. 45 (24),75–83.

Pollak, Seth D., 2008. Mechanisms linking early experience and the emergence of emo-tions. Curr. Dir. Psychol. Sci. 17 (6), 370–375.

Public Report on Basic Education in India (PROBE Team), 1999. Public Report on Basic Ed-ucation in India. Oxford University Press, New Delhi.

Rao, Vijayendra, Ban, Radu, 2007. Political Construction of Caste in South India, Manuscript.Ridgeway, Cecilia L., 2011. Framed by Gender. Oxford University Press, Oxford.Ritterhouse, Jennifer, 2006. Growing up Jim Crow: How Black and White Southern

Children Learned Race. University of North Carolina Press, Chapel Hill.Salant, Yuval, Rubinstein, Ariel, 2008. (A, f): choice with frames. Rev. Econ. Stud. 75 (4),

1287–1296.Sen, Amartya, 2006. Identity and Violence: The Illusion of Destiny.W.W. Norton & Co, NY.Shah, Ghanshyam, Mander, Harsh, Thorat, Sukhadeo, Deshpande, Satish, Baviskar, Amita,

2006. Untouchability in Rural India. Sage Publications, New Delhi.Shih, Margaret, Pittinsky, Todd, Ambady, Nalini, 1999. Stereotype susceptibility:

identity salience and shifts in quantitative performance. Psychol. Sci. 10 (1),80–83.

Shih, Margaret, Ambady, Nalini, Richeson, Jennifer, Fujita, Kentaro, Gray, Heather, 2002.Stereo-type performance boosts: the impact of self-relevance and the manner of ste-reotype activation. J. Personal. Soc. Psychol. 83 (3), 638–647.

Srinivas, M.N., 2009. The Oxford India Srinivas. Oxford University Press, Oxford.Steele, Claude, Aronson, Joshua, 1995. Stereotype threat and the intellectual test perfor-

mance of African-Americans. J. Personal. Soc. Psychol. 69 (5), 797–811.Swidler, Ann, 1986. Culture in action: symbols and strategies. Am. Sociol. Rev. 51 (2),

273–286.Swidler, Ann, 2001. Talk of Love: How Culture Matters. University of Chicago Press,

Chicago.Tverksy, Amos, Kahneman, Daniel, 1983. Extensional versus intuitive reasoning: the con-

junction fallacy in probabilistic reasoning. Psychol. Rev. 90 (293–315).































































