Federal Reserve Bank of New York Staff Reports Stereotypes and Madrassas: Experimental Evidence from Pakistan

paper presents preliminary fi ndings and is being distributed to economists and other interested readers solely to stimulate discussion and elicit comments. The views expressed in this paper are those of the authors and are not necessarily refl ective of views at the Federal Reserve Bank of New York or the Federal Reserve System. Any errors or omissions are the responsibility of the authors. Aslam, their local field teams, and participating institutions, for without their assistance this project would not have reached its conclusion. Funding through a RAND Independent Research and Development grant is gratefully acknowledged. Any errors that remain are the authors'. The views expressed in this paper are those of the authors and do not necessarily reflect the position of the Federal Reserve Bank of New York or the Federal Reserve System. Abstract Little is known about the behavior of Madrassa (Islamic religious seminaries) students, how Madrassas shape their behavior, and how other groups in their communities interact with them. To investigate this, we use experimental data that we collected from students pursuing bachelor's-equivalent degrees in Madrassas and other educational institutions of distinct religious tendencies and socioeconomic background in Pakistan. First, we find that Madrassa students are the most trusting, exhibit the highest level of other-regarding behavior, and expect others to be the most trustworthy. Second, there is a high level of trust among all groups. Third, within each institution group, we fail to find evidence of in-group bias or systematic out-group bias either in trust or tastes. Fourth, we find that students from certain backgrounds underestimate the trustworthiness of Madrassa students .


I. Introduction
Madrassas -Islamic religious seminaries-have received considerable attention recently, especially since 9/11. Despite scant research, claims made by policy makers and in the popular press suggest that they may be responsible for fostering militancy, Islamic extremism, international and national terrorism and violence. Madrassas have sometimes been labeled as "weapons of mass instruction" or "factories for global jihad", and as such have been perceived as a threat for the West and for individual countries hosting them (Rashid, 2000;Stern, 2000;Malik, 2008;Rahman, 2008;Ali, 2009). In fact, the United States has been encouraging Madrassa reform in the Muslim world, in Pakistan especially where Madrassas are thought to be linked to the Taliban (The 9/11 Commission, 2004;Fair, 2008). Madrassas are widespread around the world and educate an estimated 6 million Muslims (Haqqani, 2004). In Pakistan alone, some estimates suggest that nearly 2 million students attend Madrassas (Candland, 2008).
Madrassa students in Pakistan tend to come from modest origins, exhibit high levels of religiosity, base their studies on texts dating back to before the 14 th century, emphasize rote learning, and are thought to be exposed to teachings that reject Western ideas (Rahman, 2008).
Many Madrassa graduates go on to play an important religious and political leadership role in their communities (Malik, 2008), and are therefore important social and economic actors.
Despite their alleged influence both nationally and internationally (as well as substantial interest in understanding their role), we know very little about the behavior of young men attending Madrassas, the causal impact of Madrassas in shaping their students' behavior, and how other groups in their communities interact with these students. This is primarily because Madrassas tend to be closed institutions and data on their students is rather rare. In this paper, we use unique experimental data that we collected from Madrassa students and from students in other educational institutions in Pakistan to investigate (i) how Madrassa students and other members of the Pakistani society interact with each other, and (ii) what role Madrassas play with regards to promoting pro-social behavior vis-à-vis other groups of the society.
Pakistani society can be characterized as fragmented and polarized along social, religious, and ethnic lines. It has also been afflicted by violent conflicts in the last several years. Figure 1 shows that, in 2009, Pakistan ranked third world-wide in terms of terrorism-related deaths. In a survey we conducted in two Pakistani cities in 2010, we find that religious 5 evidence of in-group or systematic (positive or negative) out-group bias. But we find that Madrassa students give the most.
However, we do find important differences in expected trustworthiness across the groups.
Our data on subjective expectations-beliefs of how much students think was sent back to their peers, on average, by other students in the matched institution in the trust game-that measure expected trustworthiness reveal that students from Liberal Universities expect Madrassa students to send back less relative to other groups. Moreover, these beliefs held about Madrassa students are statistically very different and lower than the amount that Madrassa students actually send back. These incorrect expectations (or stereotypes) could negatively influence the social and economic interactions of those two groups outside of the lab. We also find that Madrassa students have the highest levels of expectations of the trustworthiness of other groups, and they over-estimate the trustworthiness of Liberal University students.
In summary, our results show that Madrassa students treat all the groups we consider equally, as do students belonging to other groups. Madrassa students, therefore, do not stand out as discriminating students from institutions with strong Western influence. However, their behavior stands out in another way: Madrassa students are the most trusting, exhibit strongest other-regarding behavior and they expect others to be most trustworthy. An important question that follows from this analysis centers on whether Madrassa students' pro-social behavior can be attributed to Madrassa attendance or to characteristics that lead students to attend a Madrassa in the first place (i.e., selection). Selection into Madrassas would be an issue only if one believes that students enrolling in Madrassas are even more trusting and pro-social prior to entry than levels that we observe. However, given the very high levels of trust and pro-social behavior in the data, this is quite unlikely. Regardless, we investigate the causal role of Madrassas using various identification strategies, including an instrumental variable approach where Madrassa attendance is instrumented with parental years of education, a measure of budget constraint to attend school, and perceived ability in a Madrassa relative to a non-Madrassa institution. Our results suggest that the distinct behavior of Madrassa students cannot be wholly attributed to selection, and that it is causally affected by Madrassa attendance. In particular, we can rule out Madrassas playing a role in promoting distrust and anti-social behavior. This casts doubt on the general perception that Madrassas teach hatred and ideological extremism -at least with regards 6 to groups within the Pakistani society -and is consistent with Madrassas promoting religious teachings and offering an environment that emphasizes selflessness.
Our paper shows how distinct groups within the Pakistani society interact with each other -something that is crucial for the functioning of the society. Given the current divide of the society, the omnipresent violence, and the perception of the Pakistani general population that religious institutions are playing a role in cultivating violence, our result of equal treatment is quite striking. We do not know how our results would extend to groups outside of Pakistan, if students were matched with foreign (particularly, Western) individuals. We also do not know how our results would extend to interactions of less educated groups in Pakistan. We focus on interactions of inter-elite groups, defined as college-level students, primarily for two reasons.
First, because individuals belonging to these groups will eventually become policy makers and dictate future policy, understanding their behavior is of particular relevance. Second, because we seek to isolate the effect of attending different types of institutions (abstracting from the effect of education level) on behavior. 5 Because of the very distinct social identity of the three groups that we study, our paper fits within the literature that focuses on group identity and behavior. 6 5 With regards to the affect of education on behavior, Helliwell and Putnam (2007) find that increases in average education levels improve trust, while Milligan, Moretti, and Oreopoulos (2004) find that education leads to increased civic sense. On the other hand, Algan et al. (2011) find that teaching practices (such as lecture versus work group) influence student beliefs about cooperation. This literature can be divided into two major strands: one that uses induced group membership, and the other that examines the effects of existing groups, such as ethnic groups, clans, and residential groups, on behavior. The first typically finds evidence of a strong impact of group membership on individual behavior (see for, example, Charness, Rigotti, and Rustichini, 2007;Chen and Li, 2009;Heap and Zizzo, 2009;Sutter, 2009;Benjamin, Choi, and Strickland, 2010), while the second shows more mixed results (Fershtman and Gneezy, 2001;Bernhard, Fehr, and 6 The influence of group membership on individual behavior has been widely studied in social psychology (Tajfel, Billig, and Flament, 1971), where group identity is induced exogenously by assigning participants to "minimal" groups, which are arbitrary labels such as blue or red group. These studies have found that even ad-hoc and trivial group categorizations typically lead to in-group bias and discrimination against the out-group (Tajfel and Turner, 1986). Since the introduction of identity into economic analysis by Akerlof and Kranton (2000), several economic studies have analyzed the impact of social/group identity and behavior. 7 Fischbacher, 2006;Falk and Zehnder, 2007). 7 With real groups, the impact of group membership on behavior is more varied. For example, Freshtman and Gneezy (2001) match Israeli students with objectively-recognizable ethnic names and find strong evidence of discrimination against Eastern Jews by both Ashkenazic and Eastern Jews in the trust game. Similar results are found in the social psychology literature when groups are unequal, in which case the disadvantaged group often favors the out-group (e.g., Jost, Banaji and Nosek, 2004;Brown, 2000).
Our paper uses the latter approach and investigates the behavior of existing groups. Although controlling for selection into groups makes the causal inference of group membership harder to identify (an issue avoided when group membership is randomly induced, and mitigated in our context by using several identification strategies including an instrumental variable approach for Madrassa attendance), using existing groups is a valuable approach to understanding the interactions of relevant social and economic actors from a policy and real-world perspective. 8 The paper is organized as follows. We provide background information on Pakistan, Madrassas, and the other groups we consider in Section II. Section III describes the data, sample and experimental procedures, while Section IV presents the empirical results. In Section V, we investigate the causal effect of Madrassa attendance on the behavior of Madrassa students. We While our paper takes a similar approach as Freshman and Gneezy (2001), our context is made unique by having two dimensions of inequality: religiosity and socioeconomic status, both of which are valued by the Pakistani society. Moreover, these two dimensions are negatively correlated: Liberal University students are the highest social status group when judged by socioeconomic characteristics but the lowest status group when judged on the metric of religiosity, and the reverse is true for Madrassa students. This may explain why these two groups trust more and exhibit more other-regarding behavior than the Islamic University students, who fare in the middle on both scales. The interaction between the two dimensions of social status may also explain why we observe no outgroup bias from any of the groups. 7 A third approach is to use real social groups with random assignment, as in Goette, Huffman, and Meier (2006; 8 discuss potential confounding factors for our results in Section VI, and provide a discussion of our results and concluding remarks in Section VII.

II.A Pakistan: A segmented society mired in conflict
With a population of 184 million and a GDP per capita of $2,400 (The World Factbook, 2010), Pakistan is a populous and rapidly growing middle income country. Since its inception from the violent partition of India in 1947, Pakistan has been in search of a national identity. 9 The first divide is economic. While an estimated 24% of the population live under the official poverty line, estimates based on a multidimensional poverty index such as financial poverty, illiteracy or children out of school, poor housing and physical household assets show that 54% of Pakistanis live in a state of multiple deprivations, with vast differences between rural (69%) and urban (21%) poverty rates (Jamal, 2009). About 35% of the population lives in urban areas.
To this day, it continues to be segmented along various lines and mired in violent conflict.
The second divide is religious. Ninety-five percent of the population is Muslim (Sunni 75%, Shia 20%) while the remaining 5% includes Christian and Hindu (The World Factbook,  9 along with large remittances from the Middle East (Talbot, 2009). Fifty-two percent of the Punjab population is classified poor according to Jamal's (2009) index, compared to 74% in the Balochistan province. This disparity has been another cause of increasing resentment among the regions.
In addition to the segmentation highlighted above, another characteristic of today's Pakistan is violence and terrorism. Figure 1 shows that Pakistan had 2,670 terrorism-related deaths in 2009, placing it third in a worldwide rank. Terrorism-related incidents are not confined to certain troubled areas, but are widespread across the country. 10

II.B The Madrassas in Pakistan
In a survey we conducted in 2010 on a random sample of people living in Islamabad/Rawalpindi and Lahore, we find that 14% of the respondents report knowing a victim of a violent attack. These attacks are attributed to a number of causes: sectarian violence, secessionist movements, backlash effect of the Afghan war ("Kalashnikov culture" and jihad mentality), conflict with India over Kashmir, Islamist insurgent groups and forces such as the Taliban, and the society's segmentation (Talbot, 2009).
A unique feature of this paper is to have data from a large pool of Madrassa students in Pakistan. In recent years, and in particular after 9/11, claims made by US policy makers and the popular press suggest that Madrassas-Islamic religious schools-in Pakistan are responsible for nurturing militancy and violence. Despite popular thinking, there is considerable controversy about the link between Madrassas and militancy (Billquist and Colbert, 2006;Fair, 2008).
According to Winthrop and Graff (2010), while some Madrassas are linked to sectarian militancy, most are non-extremist. Asal, Fair, and Shellman (2008)  jurisprudence, the Qur'an and its commentaries) and rational sciences such as Arabic grammar and literature, logic, and rhetoric (Rahman, 2008). 11 An important factor in understanding the extent of Madrassas' influence in Pakistan is how many students study in them. The number of Madrassas has undeniably increased, especially in the 1980s during the Soviet war in Afghanistan, when Madrassas were established in Afghan refugee camps to train fighters for the resistance movement (Winthrop and Graff, 2010). However, there is considerable disagreement over the extent of the penetration of Madrassas: Estimates of Madrassas' enrollment vary from less than 1% (Andrabi et al., 2006) to 33% (International Crisis Group Report, 2002) of all enrolled students.
The materials for these subjects are texts dating to before the 14 th century, and classes are typically taught in Urdu (Fair, 2006;Rahman, 2008). The majority of Madrassas do not impart any secular or vocational training, but they have rigid curricula emphasizing rote memorization, and it has been argued, albeit with scant evidence, that they deliberately educate their students in narrow worldviews and rejection of Western ideas, and do not train them sufficiently for the real world (Ali, 2009).

12
A key feature of Madrassas is that they generally tend to be free. In a country with a dilapidated public educational system (Winthrop and Graff, 2010), Madrassas may offer a viable alternative for families unable to afford more expensive private schools (Singer, 2001).
Therefore, it is believed that it is the poorest families that send their children to Madrassas (Rahman, 2004; New York Times).
One reason why an accurate measure of Madrassa enrollment remains challenging is that few are registeredaccording to Rashid (2000), fewer than a third of Madrassas are registered. Recent studies put the enrollment in registered Madrassas in the 1-7% range (Fair, 2008; Pakistan Ministry of Education). Regardless of the source that one choose to favor with regard to Madrassa enrollment, the overall picture indicates that a non-trivial fraction of Pakistani youth study in Madrassas.

II.C Group identity
We seek to investigate how Madrassa students and other groups of Pakistani youth interact with each other. We focus on groups that vary in socioeconomic characteristics, religiosity, exposure to Western ideas, and type of education they receive.
In our set-up, the "group" is the undergraduate institution where the students study. Our focus is therefore on a highly educated segment of the Pakistani population. Overall, in 2008Overall, in /2009.3% of the males aged 21 and above had attained at least a Bachelor degree in Pakistan. The rate increases to 14.8% among those currently working in an urban area. 14 Madrassas and their curricula have been discussed above: students typically come from modest origins, have limited exposure to Western ideas in school, study in Urdu and base their studies on religious texts. Advanced study within the Madrassas produces an Alim (Islamic scholar and/or teacher). Most students who graduate from a Madrassa go on to work in the religious sector.
The groups we consider are endogenous because families and individuals self-select into schools. We consider three main groups: Madrassas, Islamic Universities, and Liberal Universities.
Islamic Universities provide a Liberal Arts curriculum combined with Islamic teachings and courses. For example, Economics is taught with a focus on Islamic principles of finance.
These universities have segregated campuses for males and females, and classes are taught in Arabic or English. These institutions tend to be public and, therefore, are accessible to low and middle income groups. While they have certain admission requirements, they are relatively easier to get accepted into (relative to liberal universities). A relatively large proportion of students at such universities have typically studied for some time at Madrassas before enrolling.
Liberal Universities-the third kind of institution-are similar to American colleges. They teach a Liberal Arts curriculum in English, and have gender-mixed campuses. Since tuition at such institutions tends to be very expensive, they are generally accessible only to individuals 12 from high socioeconomic backgrounds. In addition, these institutions are quite selective and their entry requirements are such that they primarily accept students who graduate from private high schools (which tend to have higher academic standards and which, in most cases, cater to the rich).
These three groups clearly represent three different identities within the Pakistani society. At one end of the spectrum we have young males from poorer backgrounds who attend religious schools that are thought by many outside of (and to some extent inside of) Pakistan to be linked to militancy and extremism. At the other end of the spectrum we have wealthy students exposed to Western-type education. Our measure of group identity is a measure of both religious identity as well as social class.

III. Data
We conducted experiments in two male Sunni Madrassas from the Barelvi school of thought, two male Sunni Madrassas from the Deobandi school of thought, one Islamic University The institutions in our sample are among the five largest and best-regarded institutions in the relevant category in each city. Among all the institutions we contacted, one Liberal University and one Madrassa declined participation. We sampled the most senior students in the four Madrassas since they are similar in age to university students, and are pursuing the Madrassa equivalent of a Bachelor degree. Though participation was voluntary, almost everyone in the Madrassas participated in the study. At the other institutions, a random sample of students was selected to participate based on a listing of students provided by the registrar's office.
Average response rate at the universities was about 70%. Overall 1,521 male students 13 participated in the experiments. 15 They also answered a questionnaire asking about demographic characteristics, school choice, and attitudes on social issues. Below we describe our sample and the experimental procedure. However, because they differ in their students' characteristics and tuition level (as we show below), we classify the two Liberal Universities (LU) into two separate groups: a Liberal

III.A Sample
Western-style university (LU-W) and a Liberal modern (LU-M) University. LU-W is more selective and liberal than the LU-M, and it caters to a higher socioeconomic segment of the society. Table 1 shows observed differences among respondents from the four groups, ranked by most-to-least liberal institutional affiliation. Students at more liberal schools have parents with higher income, education, and asset ownership. For example, the average number of years of father's (mother's) schooling is 14 (13) for students from the Liberal Western-style University compared to 7.1 (3.5) years in the Madrassas. Similarly, the monthly parental income in the liberal Western-style University is nearly 10 times the income in the Madrassas. Moreover, the characteristics of students at LU-W and LU-M, the two Liberal Universities, are significantly 15 Female students from the Islamic and Liberal Universities also participated in the experiments. We restrict the analysis in this paper to male students who were matched with other male students in order to focus on group identity, defined by socioeconomic class and religiosity (proxied by institution). The full sample (before excluding female students from IU and LU, and male students from all institutions matched with female students) consists of 2,836 students. The Madrassas we surveyed cater to male students only. Female Madrassas tend to be small. Since large sample sizes are needed for randomization in the experiment, we did not include them in our sample. different from each other in most cases. For example, average parental income for LU-W students is almost twice that of LU-M students.
Self-reported religiosity and the number of prayers per day also vary by group. Students were asked to rate how religious they considered themselves to be on a scale from 0 (not religious at all) to 10 (very religious). The average religiosity is 5.3 in the liberal Western-style University compared to 9.2 in the Madrassas. The former also pray much less frequently every day (1.5 as compared with 4.9 times per day).
Finally, students differ in their exposure to information and media, as well as in peer group characteristics. Just 23% of the Madrassa students report watching BBC and CNN, compared with 60% of the students of the other groups. In addition, while fathers of only 4% of students attending LU-W spent more than two years studying in a Madrassa either on a part-time or full-time basis, the corresponding proportion for Madrassa students is 21%. Similarly, about 6% of the LU-W students have at least a sibling or a friend who spent more than two years in a Madrassa either part-or full-time, compared with nearly a quarter of students at  Institutional sorting based on socioeconomic and other characteristics is stark but unsurprising given Pakistan's divided history. As we move from left-most LU-W (column 1) toward Madrassas (column 4) in Table 1, the average socioeconomic characteristics deteriorate (for example, parental income and education decrease). At the same time, extent of religiosity increases. If we compare the students to the City sample (column 5), we see that Madrassa students seem to hail from humbler backgrounds than do those from the general population in the cities, and that all other institutions fare better in terms of most indicators of wealth. This is consistent with the hypothesis mentioned above that poverty may drive families to send their children to Madrassas.
This also suggests that the various groups in our setting do interact with and have exposure to each other at some level.
The last row of the table also shows that a non-trivial proportion of respondents (14-34%) in each setting have an acquaintance who died or was injured as a result of the violence in Pakistan. This shows again, as pointed out in Section II.A, that violence is widespread and has affected a large proportion of the general public.

III.B Perception of Madrassas
While the Madrassa-militancy link has been widely discussed in the popular press and remains controversial (see section II.B above), an important question pertains to how the Pakistani public views this linkage. Because this is relevant to the interpretation of our findings, we report the responses of some survey questions administered to the IU and City samples that were designed to shed light on this.
Columns (1) and (3) of Table 2 report the mean responses for respondents at IU and City.
The two groups have different exposure to Madrassas, as nearly half of the students at IU report having studied in a religious institution either part-time or full-time in the past, compared to only 9% of the City sample. While these numbers may seem high, they reflect a culture in South Asian countries whereby even affluent families send their children (either full-time or part-time) to religious institutions (which could be Madrassas or schools run in mosques) to learn to properly recite or memorize the Qu'ran (Billquist and Colbert, 2006).
In order to evaluate respondents' perceptions of the influence of Madrassas in the society, we first asked them to guess the percentage of 18-year old males currently enrolled in registered Madrassas. Consistent with inflated estimates in the press about Madrassa enrollment (Andrabi et al., 2006), our respondents tend to over-predict the number of students enrolled in Madrassas with estimates ranging from 38% for IU to 26% for the city sample. 17 We also asked respondents to rank a list of seven possible causes for extremism and violence in Pakistan (7 denoting the most important).
This suggests that they see Madrassas as providing education to, and therefore influencing, a large proportion of the population.
18 17 Note that our survey question asks about the enrollment proportion of 18-year old male students, which could be different from the proportion of enrolled students of all ages.
The mean rating assigned to religious institutions is 3.1 for IU students, and 5.7 for the City sample. Religious institutions are perceived to be the least important cause of extremism by IU students, but the most important 16 source by City respondents. Since IU students have stronger exposure to Madrassas (see Table   1), their relatively positive view of Madrassas is not surprising. Even so, it should be noted that both groups perceive religious institutions to be playing some role in fostering extremism.
A third question asked respondents to assign a percent chance that each of the four entities (US and Western countries; Radical Islamic organizations; Afghanistani government; Indian government) were responsible for violence against civilians in Pakistan in recent times.
Both groups assigned to "radical Islamic organizations" the lowest chance of being responsible for such acts. However, the mean likelihood assigned to religious organizations is still very different from zero: 45% for IU students and 34% for the City respondents.
Exposure to religious institutions could affect how respondents perceive Madrassas.
Columns (2) and (4) of the table report the corresponding mean responses for the IU and City respondents who had never attended a Madrassa. We do not find that to be the case-their responses are statistically similar to their counterparts who have been exposed to religious institutions.
Finally, both IU and City respondents overwhelmingly favor the government's plan to reform Madrassas, which, among other things, requires them to register and teach secular subjects such as math and science. This sentiment is similar to that found in Fair, Ramsay, and Kull (2008), where two-thirds of their sample reportedly supports Madrassa reform.
We interpret these statistical findings to imply that the general public views Madrassas as playing a large role in educating the youth, as being somewhat complicit in militancy and extremism, and that reforming them would be positive for society.

III.C Experimental design
We now present the details of the experiments that we run to understand how the groups interact with each other.

Procedure:
The experiments were conducted in sessions of 50-100 students in a classroom of the student's institution. The rooms were large enough to ensure respondent anonymity. The instructions were given to each participant, read out aloud by the experimenters and projected on 17 a retro-projector. 19 Respondents played the games on a paper questionnaire and were matched with an actual partner ex-post, so they did not learn the actual identity or action of their partner while playing the game. The questionnaire was administered in Urdu at all places except the Western-style liberal University where it was conducted in English, since students there are more used to reading and writing in English. 20 Moreover, the questionnaires were identical across all the institutions up to the section leading into the experiments.
Games: Students were asked to play the following games: -Trust game: Player A (the sender) is given a fixed amount of money (Rs. 300) and decides whether to keep it or give it to Player B (the receiver), i.e. to invest it. If given to Player B, the experimenter triples that amount and gives it to Player B who is asked to choose whether to transfer any money back to player A (which can be any amount between zero and Rs. 900). This is a binary version of the "trust game" introduced by Berg, Dickhaut, and McCabe (1995)-it is binary in the sense that player A can choose to send either nothing or the entire amount. The efficient outcome is for A to invest the money by transferring it to player B, while the subgame perfect equilibrium is to keep the money. Lack of trust toward the partner may lead to inefficiencies. In our setting, all respondents first played the role of Player A and then the role of a Player B, who received the money. When put in the role of Player B, we use the strategy method and ask the respondent to report the amount he would like to send back conditional on Player A deciding to invest.
-Dictator game: This is a one-stage game in which Player A (the sender) divides a fixed amount of money (Rs 400) between himself and Player B (the receiver). Player B does not make any decision. Again, respondents play first in the role of Player A and then in the role of Player B (in which case there is no decision to make).
-Expectations: For both the trust and dictator games, respondents were asked to guess (i) the average amount that students from their own institution chose to give to their 19 Full instructions are available from the authors upon request. 20 The translation was supervised by Basit Zafar (co-author) who speaks both English and Urdu fluently, to ensure that nothing was lost in translation.
18 partners, and (ii) vice versa, the average amount that students from the partner's institution chose to give to their match in the respondent's institution. 21 Note that when students are asked to provide their expectations, they are asked about the average payoffs for an identical pair of partners (see the exact instructions in the Appendix). Expectations were elicited after the respondent had played the two games.
Treatment: The treatment in this experiment is the randomization of institution of the pair of players. 22 Each student was randomly matched with one of the following partners: a male Madrassa student, a male student from a Liberal University, or a male student from an Islamic University. Students from the Liberal Modern (Western-style respectively) university who were selected to be matched with a student from a Liberal University were informed that they were matched with a student from their own university. All other students who were selected to be matched with a student from a Liberal University were informed that they were matched with a student from the Liberal Western-style University. The description of the match (with the exact name of the match's educational institution) was already printed on the paper questionnaire received by each participant, so students were not aware that other participants in their session were possibly matched with partners of different educational institutions. Each student was informed that they would play all the games with the same partner. Table 3 presents the sample sizes for each institution, and for the various matches. 23 21 While we want to measure the respondent's expectation of the amount his partner sends back, we ask the respondent to guess the average amount sent back by all students from the partner's institution (who are matched with students in the respondent's institution). This is because, asking the respondent for his expectation of the amount sent back by his partner, may prompt the respondent to report expectations that rationalize his own investment decision in the trust game. We believe our approach mitigates this concern of ex-post rationalization, and is hence superior.
Students were given a short description of the institution they were matched with but since the selected institutions are among the most well-known institutions, most students would have some prior knowledge of them.

IV. Experiment results
We now discuss the results of our experiments. We postpone the discussion on the mechanisms behind the results to later sections. To better understand interactions between the different groups, we address five main questions.

Question 1: Does investment behavior in the trust game vary by group (i.e., institution type)?
We begin by investigating whether investment behavior varies systematically by group. The first column of It is also of interest to evaluate whether group membership solely leads to the distinctive behavior, or whether observable characteristics partly explain this difference. We investigate this in columns (1)-(3) of Appendix Table A1, which reports the marginal effects of a probit regression of the decision to send money in the trust game on observables as well as institution dummies. We find that observable characteristics, such as measures of socioeconomic background (parental wealth, parents' income, parents' education), age, risk preference and selfreported religiosity have little power in explaining the decision to invest in the trust game (column 1), while column 2 shows that attending a Madrassa (and attending LU-M) is associated with a higher probability of sending money in the trust game (coefficient statistically significant at 1%). Moreover, the coefficient on the Madrassa and LU-M dummies stays statistically significant even after controlling for observables (column 3), suggesting that sorting on observables into these institutions cannot explain the differential behavior across groups.
When conducting pairwise t-tests, we find that the mean investment rate of the Madrassa students is statistically significantly different from that of the other groups, except for LU-M.
We summarize these results below: 25 One needs to be cautious in making any comparisons with the few studies that employ a binary trust game, since small modifications in the design (such as stake size, stake increase in the investment game-in our case three times, specifics of the match that the respondent is informed about, sample characteristics) can result in large differences. The continuous trust game has been employed by more studies. Players A (trustors) send about 50% of their endowment in such games in developed as well as developing countries (Camerer, 2003;Cardenas and Carpenter, 2008). 26 Note that this result is not driven by the larger sample size of pairs of Madrassa students matched with other Madrassa students (see footnote 19) as we obtain similar results if we look at the investment behavior for a given partner. For example, among subjects matched with students from the IU, we see that 63% (80%) of the IU (Madrassa) students decided to invest. We discuss these results below.

21
RESULT 1: Investment behavior in the trust game varies by group, with Madrassa students being more likely to invest. Differences in observable characteristics do not explain the differential behavior across groups.
Question 2: Do students exhibit in-group bias and is there differential treatment (discrimination) in terms of investment against a particular group?
As pointed out earlier, a large body of literature suggests that individuals tend to favor members of their own group, though other work finds that there is out-group favoritism from lower status groups when groups are unequal. We now investigate whether there is in-group or out-group bias, or systematic discrimination in favor of or against a particular group in our data. For each group, Columns (2)-(4) of Table 4 show the proportion of respondents who sent the Rs. 300 in the trust game conditional on the group they are matched with. Within each group (i.e., each row institution), we do not reject the hypothesis that the proportion of respondents who invest varies by the matched group (as indicated by the F-test for equality of proportions within group, and Kruskal-Wallis that tests whether the data come from the same distribution in columns 5 and 6, respectively). This suggests that there is no systematic discrimination against a particular group in any group's investment decision. Moreover, none of the two sets of pairwise hypothesis tests between having a match from one's own institution type versus another institution type (Wilcoxon rank-sum, and t-test) are statistically significant at levels of significance of 5% or lower. This implies that students do not invest more extensively when interacting with a partner from their own group. 27 Table 1 shows that at all institutions, a non-negligible proportion of students' parents or siblings attended a Madrassa. These students may have a systematically different perception of Madrassa students. To investigate this, we replicated Table 4 excluding respondents who had no relatives who had attended a Madrassa and found similar patterns (table not shown).
We summarize this result below: 27 These results also hold if we control for respondents' observables in within-institution regressions with match dummies.

22
RESULT 2: There is no evidence of in-group bias or of differential treatment to any particular group.
Results from the trust game do not allow identification of the relative roles of those dimensions (Cox, 2004). While Result 2 emphasizes a homogenous investment behavior toward the various groups, such a result could still be consistent with different levels of trust and of unconditional other-regarding behavior toward certain groups. 28 Our multiple-game experimental design allows us to separately measure unconditional other-regarding behavior and expected trustworthiness. In the dictator game, the only motive for sending money to the partner is unconditional other-regarding behavior. We can thus learn more about other-regarding behavior by analyzing how students played that game. In addition, the elicitation of expected average amount sent back by each group to students from their own institution gives us a measure of expected trustworthiness or stereotype toward each group. This is developed in the two following questions.
For example, Madrassa students may invest similarly in IU students and Liberal University students because they do not trust IU students but are altruistic toward them while they trust Liberal University students but do not exhibit altruism toward them. 28 Since students were randomly assigned a treatment (i.e., match type), differences in risk preferences cannot explain any of the results since there is no reason to believe that risk preferences would change by match type. Therefore, we do not focus on this explanation when decomposing behavior in the trust game. However, we do have qualitative measures of risk preferences from the respondents, and they in fact are similar within each treatment conditional on the student's institution.
Question 3: Is there taste-based discrimination against a group? Table 5 shows the average amount sent in the dictator game for all pairs of partners. As shown in the first column, on average, students sent Rs. 171 (42.7% of the total amount) to their partner and only 6.7% did not send anything at all. This is a very high level of unconditional otherregarding behavior when compared to the standard of dictators typically sending between 20% and 30% of their endowment in both developed as well as developing countries (Camerer, 2003;Cardenas and Carpenter, 2008 We summarize this in the following result: One difference in behavior is, however, noticeable: Madrassa students give on average more than any other group and are less likely to give nothing. We also investigate the predictive power of observable characteristics at explaining the amount sent in the dictator game in columns (4)-(6) of Appendix Table A1. Again, we find that the higher unconditional other-regarding behavior of Madrassa students (and lower other-regarding behavior of IU students) is not explained by differences in observable characteristics.
RESULT 3: Within each group (institution type), there is no taste-based discrimination.
Madrassa students exhibit stronger unconditional other-regarding behavior than any other group.
Question 4: Is there systematic difference in expected trustworthiness (stereotype) for a particular group?
29 Note that we cannot disentangle the exact mechanisms leading to each group's unconditional other-regarding behavior. As mentioned above, unconditional other-regarding behavior may be prompted by altruism, warm glow, inequity-aversion, or maximin preferences. For example, it could be that an average LU student has distaste (low altruism) toward Madrassa students (relative to tastes toward students at other institutions), but is very averse to income inequality. This could lead to similar average behavior in the dictator game toward different groups since Madrassa students tend to be from less affluent backgrounds.  Figure 3 show the expectations reported by individuals regarding the average amount expected back from the matched group, i.e., it reveals expected trustworthiness (or stereotype). Note that respondents choose an interval for the average and do not report a point estimate for the exact average. The mean and median amounts presented in Table 6 are those obtained by allocating the middle of the chosen interval as expected average. To conduct hypothesis testing on various quantiles of the distribution of expectations, we also present in Table 6 the proportion of respondents who expect to receive more than Rs. 200, more than Rs.
300, and more than Rs. 400 from the match. Three points from this table and Figure 3 are of note. First, Madrassa students most expect other groups to be trustworthy. Column (1) of Table 6 shows that Madrassa students expect back Rs 405 on average while all other institutions expect less than Rs 370, with the differences being statistically significant (see test for equality of means, medians, and distributions of the four groups presented at the bottom of Table 6).
Combined with the fact that Madrassa students exhibit the strongest other-regarding behavior, this helps explain why they are more likely to invest in the trust game. Second, students from LU-M and Madrassas expect all groups to be equally trustworthy (none of the P-values of the tests for equality of means, medians and distributions of the three matches presented in the last three columns of Table 6 are less than 5%). Finally, students from LU-W and IU expect different levels of trustworthiness across the various groups, as is also apparent in Figure 3. LU-W students believe IU students to be the most trustworthy and Madrassa students to be the least trustworthy. For example, more than half of the LU-W students expect IU students to send back more than Rs. 400 compared to less than a quarter who expect Madrassa students to send more than Rs. 400. This difference is statistically significant as shown on the test based on the imputed expectations and on the proportion of respondents who expect to receive more than Rs 400 (Pvalue less than 5% in the last three columns of Table 6). In contrast, IU students expect Liberal University students to be the least trustworthy. We summarize the results from Madrassa students are the most trusting, they also exhibit the strongest unconditional other-regarding behavior and they expect to receive back the most from the match in the trust game. Similarly, LU-M students, who are second-most likely to invest, have the second highest levels of other-regarding behavior and expectations of return. This would imply that trusting behavior is related to both unconditional other-regarding behavior as well as expectations of return. We return to this point in section VI.
Question 5: Are the stereotypes correct?
We now compare the expected amount sent to the actual amount sent in Table 7. We show the proportion of students who expected more than Rs. 300 from a given group and the proportion of students from that group who actually sent more than Rs. 300. In addition, we also show the proportion of students who had "accurate" expectations, i.e. chose the interval that contained the actual average, and the proportion of students who under-estimated the amount sent, i.e., chose an interval whose upper-bound was below the actual average.
Several interesting findings stand out. First, Liberal University (LU-W and LU-M) students have inaccurate expectations about Madrassa students. While 81% of the Madrassa students sent more than Rs. 300 to Liberal University students, only 52% of LU-W and 69% of LU-M students expected to receive more than this amount of money. The differences between actual proportions that sent and expected proportions are statistically significant at 5% (as indicated by the P-value in the third row of each panel in Table 7). Moreover, a large proportion of respondents from Liberal Universities under-estimated what Madrassa students would send back (76% of the LU-W and 54% of the LU-M students). Note that this result is not driven solely by the fact that Madrassa students actually send back the most amount relative to other groups (last row in each panel in Table 7), but also by the fact that Liberal Universities' students expect Madrassa students to be the least trustworthy (Table 6). Second, Madrassa students expected more from Liberal University students than what they actually received from them: only 13% of the Madrassa students had accurate expectations while 64% over-estimated the amount they would receive. On the contrary, IU students tended to expect less from Liberal University students than what they actually sent (the t-test for equality of proportion of students who expect more than Rs. 300 and the proportion who actually sent more than Rs. 300 is statistically significant at 10%). 26 These differences, however, do not seem to be large enough to generate differences in investment behavior in the trust game (see Result 2), possibly because of the binary nature of the decision.
RESULT 5: There is incorrect stereotyping. Liberal University students systematically underestimate the trustworthiness of Madrassa students, while Madrassa students systematically overestimate the trustworthiness of Liberal University students.

V. Why do Madrassa Students Behave Differently?
We find that Madrassa students behave systematically differently from other groups: They trust the most, exhibit the strongest other-regarding behavior, and have highest expectations of others' trustworthiness. Unless there is positive selection into Madrassa-i.e., students enrolling in Madrassas are more trusting and pro-social to start out with than the levels that we observe (something that is unlikely given the high levels of trust and other-regarding behavior we find)-this would suggest that Madrassas are not teaching hatred and distrust, but are on the contrary promoting trust and pro-social behavior towards other members of the Pakistani society. We seek to investigate this in more detail in this section. In particular, we investigate whether the difference in behavior of the Madrassa students is due to characteristics that lead students to attend Madrassas (selection) versus exposure to Madrassa teachings and environment.
One potential explanation for the difference in behavior across groups could be systematic difference in preferences. In particular, lower risk-aversion could explain why Madrassa students are more likely to invest in the trust game. This lower risk aversion could be due to either selection or Madrassa attendance. However, Table 1 reveals that Madrassa students are on the contrary more risk-averse on average than any of the other groups.
We next investigate the role of socioeconomic status. We have seen that Madrassa students hail from humbler backgrounds (Table 1), so their different social origin could possibly explain their trusting behavior. However, in the U.S context, Alesina and La Ferrara (2002) find on the contrary that high-income groups tend to trust more. Using data from the 2000 World Value Survey, we also find that in Pakistan, higher income groups exhibit higher levels of trust, 27 as measured by the proportion of people who state that "most people can be trusted these days" when asked "Generally speaking, would you say that most people can be trusted or that you need to be very careful in dealing with people?" (Panel A in Appendix Table A2). Among the educated groups we consider, socioeconomic status does not have predictive power in the decision to invest in the trust game (see Appendix Table A1). So the difference in socioeconomic status is unlikely to be the main driver of differences in behavior.
Another possibility is differences in levels of religiosity. This would be an issue in the interpretation of our results if Madrassa students had higher pre-existing levels of religiosity before enrolling into a Madrassa. Existing evidence suggests that religious rituals promote prosocial behavior (Iannacone, 1998;Ruffle andSossis, 2007, Clingingsmith, Khwaja andKremer, 2009), though the evidence on the relationship between religiosity and trust and trustworthiness is mixed (Welch et al., 2004;Tan and Vogel, 2008;Daniels and von der Rurh, 2010;Putnam and Campbell, 2010). Religiosity alone, however, cannot explain the patterns that we observe across the institutions. Note that IU students rank second in terms of their religiosity and adherence to religious practices, but are less likely to trust (Table 4) and to exhibit other-regarding behavior (  Table A2). So religiosity is also unlikely to be the main driver of differences in behavior.
Overall, it seems unlikely that selection into Madrassas by observable characteristics explains the trusting behavior of the Madrassa students. We now investigate the possibility that selection into Madrassa by unobservable characteristics explain our results. For this, we present in Panel A of Table 8 a regression using a dummy variable for whether the respondent invested in the trust game and as independent variables Madrassa attendance, parental income, level of risk aversion, and a city dummy. The first column shows an OLS specification while the second column presents an instrumental variable (IV) specification where Madrassa attendance is instrumented by the following variables: (i) father's and mother's years of education, (ii) the 28 reported maximum monthly expenses (including tuition, room and board) that a respondent and his family would be able to pay to be enrolled in school without any external financial aid 30 , and (iii) perceived ability rank conditional on graduating from a Madrassa (relative to a non-Madrassa institution). 31 The idea with the first instrument is that we expect more educated parents to be less likely to send their children to a Madrassa. The second instrument is included because we expect students who are less financially constrained to be less likely to enroll in a Madrassa (note that the correlation between parental income and this maximum amount is 0.18, suggesting that parental income may be an imprecise measure for the level of financial constraints a household faces to educate its children). Finally, we include relative ability in a Madrassa as an instrument as we expect students who believe they would perform better at a For both the OLS and IV specifications, we find that the coefficient associated with Madrassa attendance is positive and statistically significant at 1%, suggesting that Madrassa attendance leads student to be more likely to invest in the trust game. In particular, attending a Madrassa leads to an increase in the probability to invest by 14.7 percentage points in the IV specification. Note, moreover, that the instruments are strong predictors of Madrassa attendance in the first stage with the expected sign, and that we do not reject the null hypothesis that the instruments are uncorrelated with the error terms and that the model is correctly specified (as indicated by the p-value of Sargan's chisquared test for overidentifying restrictions). 31 In a different part of the survey, students were asked to report their ability rank (on a 1-100 scale, where 1 means the best student and 100 means the last-ranked student) if they had attended and were to graduate from five possible institutions, which included both Madrassas and non-Madrassas. Relative ability is defined as max (Ability rank in a Madrassa) -max (Ability rank in a non-Madrassa institution). A negative gap implies higher perceived ability in a Madrassa. 32 We use two-stage-Least-Squares (2SLS) for the IV specification, and following the discussion in Angrist and Pischke (2009) use linear IV models with our binary endogenous regressors.

29
To further investigate the extent to which our results are driven by selection, we next conduct the following thought experiment within the pool of Madrassa students. In another part of the survey, respondents were asked which type of institutions they would attend if they were admitted to all institutions types and all expenses would be covered. Eleven percent of the Madrassa students stated that they would choose to attend a non-Madrassa institution under those conditions (i.e., would "switch"). We can think of these students as not selecting into Madrassas on the basis of unobservable characteristics since they would have attended another institution without budget or qualification requirements constraints. That is, the difference in behavior between these students and those who would attend a Madrassa regardless should identify the extent to which selection into Madrassas drives our results. We find that this group of students behaves similarly to students who would have chosen a Madrassa anyway: the proportion of students who invest in the trust game is respectively 80.1% and 80.2% and the difference is not statistically significant. This indicates that the higher trusting behavior of Madrassa students is not a result of selection (on observables or unobservables) into Madrassas. This result holds even if we control for parental income, level of risk aversion and a city dummy (last column of Panel A in Table 8). 33 We next investigate how Madrassa attendance influences the amount sent in the dictator game in Panel B of Table 8. We find that the instrumented Madrassa attendance increases the amount sent in the dictator game by Rs. 31, and that the effect is precisely estimated (significant at 1%). We also find that Madrassa students who would switch to another school give similarly as other Madrassa students.
Finally, using a specification similar to that in Panels A and B of All in all, these results are strong suggestive evidence that Madrassas promote trust and selflessness, and that the differential behavior we identify in our data is not driven solely by selection into Madrassas. In particular, we can rule out Madrassas playing a negative role by promoting distrust and anti-social preferences.
So what exactly is it about attending Madrassas that may lead their students to be more trusting and selfless? Given that Madrassas are closed institutions, one may argue that because their students have limited interactions with other society members, different aspects of their behavior are not fully-formed yet and that, in that regard, their behaviors may be similar to those of children and teenagers. Existing research on behavior of children in experiments similar to ours shows that they are less trusting than adults and that younger children are not more trusting than older ones (Harbaugh et al., 2002), and that altruistic behavior in the dictator game is weakly increasing with age (Benenson, Pascoe, and Radmore, 2007). It is therefore unlikely that something about the Madrassa lifestyle drives our results. However, one possible channel that may lead to our results is the nature of Madrassa teachings. Since religious teachings in general promote selflessness, and studying certain secular subjects is associated with different behavior with regards to cooperation and generosity (Carter and Irons, 1991;Frank, Gilovich, and Regan, 1993) 34 , the emphasis of Madrassas on religious teachings and their neglect of secular subjects may be one reason for our findings of high trust and high levels of pro-social behavior of Madrassa students.

VI. Potential Confounding Factors
In this section, we discuss a list of factors that could influence our findings, and argue that they can be ruled out as possible explanations for our results.

Understanding the game
Given that most students in our sample had never taken a survey, let alone participated in an experimental game, the possibility of students not being able to comprehend experimental instructions has to be considered. When designing the questionnaire, we made a conscious effort 34 These papers find that students in fields such as Economics are less trusting, less cooperative, and more selfish.
However, it seems that this relationship is driven primarily by selection into fields (Frey and Meier, 2005).

31
to keep the instructions as simple as possible. It was precisely for this reason that we chose the dictator and trust game to measure discrimination, and not a public good game. 35 To highlight the point that students did not randomly make decisions in the experiments, we undertake a set of naïve probit regressions, where we regress an indicator of whether the student sent money in the trust game to his match on both 1) the amount he sent to the match in the dictator game (a measure of unconditional other-regarding preferences) and 2) beliefs about the match's trustworthiness.
The instructions for the games were read out line by line, and projected on a screen. Students were allowed to ask any clarifying questions by raising their hands. Finally, the decisions involved in both the dictator and trust game really are very simple. 36 35 For example, Castillo and Petrie (2010) use a public good game to measure discrimination in the lab. Another reason why we could not use a public game setup, aside from being harder to comprehend, was that it would have been challenging to give feedback (for example, about others' contributions) to participants in real time using paper surveys.
The marginal effects of these regressions are reported in column (a) of Table 9 for each institution panel. All the estimates are of the expected sign (positive), and statistically significant (except for the coefficient on trustworthiness for LU-W). This result is similar to that of Ashraf et al. (2006), who also find that trust is related to both expectations of return and unconditional kindness. In columns (b) of this Table, we show the marginal effects of being matched with IU and Madrassa students, with the excluded category being the Liberal University match. These estimates are overall not very precisely estimated, indicating that there is little statistical difference in trust game behavior by match type within an institution (something that is also reflected in Table 4). In columns (c) of Table 9, we include this set of dummies along with the dictator game behavior and trustworthiness expectations as explanatory variables. The results show that, as in columns (a), the dictator game behavior and trustworthiness expectations are significant determinants of the trust game behavior. The institution match dummies become less precisely estimated than in column (b). This suggests that students' actions in the trust game are consistent with their preferences (amount sent in the dictator game) and trustworthiness expectations.
Theory suggests that respondents send weakly larger amounts back in the trust game than they do in the dictator game because of the added concern of reciprocity (Cox, 2004). Column 1 of Table 10 shows the summary statistics of amount sent in the dictator game by institution.
Column 4 shows the corresponding statistics for the amount sent back in the trust game (out of Rs. 900 but scaled down to Rs. 400). Abstracting from effects of stake size, this is consistent with theory. 37 It is also interesting to note that the findings in the trust game match up fairly well with the self-reported response to the question ''Generally speaking, on a scale from 0 to 10 would you say that most people can be trusted?'', where zero means "all people cannot be trusted" and 10 "all people can be trusted." This question is a variant of the question asked in the General Social Survey (see Glaeser et al., 2000, for a discussion on this). As shown in Table 1, the mean response to this question is highest for Madrassa students (5.1) followed by LU-M students (4.8), who are second-most likely to send money in the trust game (Table 4).
The null hypothesis that the distributions of amount sent in the dictator game and the rescaled (from Rs. 900 to Rs. 400) amount sent back in the trust game are the same is rejected for each of the institutions.
All of this provides strong suggestive evidence that respondents understood the setup of the games.

Anonymity and Stakes
Experimental payoffs for each student had to be computed by matching a student with another student belonging to the designated matched institution, and taking into account the choices of both students in the match pair. While students knew the institution of the student they were matched with, they were never told who they were matched with. 38 37 Stake size has been found not to matter for amount sent in dictator games (Forsythe et al., 1994;Carpenter, Verhoogen, and Burks, 2005) or for amount sent back in the trust game (Johansson-Stenman, Mahmud, and Martinsson, 2005).
Also during the administration of the experimental component of the survey, all enumerators were pulled aside 38 Payoffs were only made available to students starting one week after the completion of the survey. Therefore, we needed some way to link students to the questionnaire, while at the same time keeping their identity confidential. In order to achieve this, each student was given a ten rupee bill (that was initially stapled to their questionnaire) and was asked to enter the serial number of the bill on the questionnaire. The computed subject payoffs for each student were then put in sealed envelopes with the corresponding serial number on them. Students simply had to return their ten rupee bill one week later when they came to collect their payoff.

33
and there was ample space between students so that respondents did not feel that their responses were being observed by anyone else. Our survey administration method is therefore equivalent to an anonymous questionnaire.
As mentioned earlier, our experiments involved fairly high stakes. Students earned Rs.
600 (USD 7) on average from the experiments. Rescaling the stakes using per capita GNI numbers at PPP, this corresponds to about USD 120 in the US. Therefore, the stakes involved in the experiments were considerable. This is particularly true for the low-income students in our sample, who predominantly belong to Madrassas and who are found to exhibit the strongest trust and other-regarding behavior. Therefore, our results cannot be attributed to the stakes being low.

Credibility
Nearly all the students in our sample had never been exposed to incentivized experimental tasks. Moreover, as mentioned above, payoffs for the games were made at least one week after the student had taken the survey. Both of these factors may lead subjects to believe that payments may not actually be made, and they might not find the offer of monetary incentives credible. While it is not possible to rule out this factor, we worked with the administrations of the participating institutions to convey the seriousness of the incentives. Our field teams always included students and faculty members of the participating institutions.
Moreover, participants were informed that the experimental payoffs would be available for pickup from the Registrar's Office in sealed envelopes. Finally, participants received an instantaneous show-up fee of Rs. 200 for completing the questionnaire. These measures should have confirmed the legitimacy of the undertaking from the student's perspective.
The show-up rate (to pick up compensation) varied between 69% and 100% at the participating institutions. 39 39 The claim rate does not include those respondents who showed up to receive compensation and had either misplaced their Rs. 10 notes, or had no record of their bill serial numbers.
One could argue that respondents who did not claim their compensation may never have intended to do so, and therefore did not play the games the way they would have if they had intended to collect their compensation. We do not find this to be the case-our results are very similar if we exclude this set of respondents (results available from the authors upon request).

34
A related point is the possibility that students may not have found the matching and randomization procedure credible. Based on how the study was setup, we do not believe that to be the case. Moreover, in Delavande and Zafar (2011), using gender variation in matches, we find significant evidence of (taste-based) discrimination toward certain female students by Madrassa students.
Finally, one may be concerned that the match characteristic (institution of matched student) is not salient enough in our design, and hence the finding that behavior does not vary by match type. The names of institutions printed on the questionnaire are well-known so that is unlikely to be the case. 40 The match characteristics in our setup were at least as salient as in other studies that use real world groups, such as Fershtman and Gneezy (2001) and Falk and Zehnder (2007); both of these studies find differential behavior by match type. 41

Binary nature of the trust game
One possible reason for the lack of an effect of group membership may be that the binary nature of the trust game prevents us from identifying effects that we would find with a continuous trust game. However, this is unlikely to be the case because (i) we do find differences by institutions in the binary decision to invest the money in the trust game, and (ii) we also do not find a group membership effect in the dictator game where the decision is not binary.

VII. Discussion
Using experiments of economic decision-making (that measure trust and unconditional other-regarding behavior), we investigated how Madrassa students in Pakistan interact with other groups of male youth of very different religiosity, socioeconomic background and exposure to 40 In another part of the questionnaire, respondents were asked to report average earnings of students aged 30 who had graduated from each of those institutions. All groups report large difference in wages, with highest wages for LU graduates, followed by IU and then by Madrassa graduates, suggesting that those institutions were indeed salient. 41 We keep the games anonymous and conceal identities of the players from each other because (1) this is the standard approach used in comparable studies, and (2) this approach allows us (the experimenters) control over the environment. Varying social distance (identification) has been shown to change behavior in games (Hoffman, Mccabe, and Smith, 1996;Bohnet and Frey, 1999). Since we are interested in differences in behavior by match type, and we use the same design across matches, this feature of our design cannot explain our results.
Western ideas. Our first finding is that, in this context, Madrassa students are the most trusting, exhibit the highest other-regarding behavior and expect others to be the most trustworthy. We provide evidence that this distinct behavior can be attributed to Madrassa attendance, and is not driven entirely by selection. This suggests that Madrassas promote selflessness and inter-group trust among their students, at least toward other segments of the Pakistani society. It is important to note that the Madrassas we surveyed are mainstream, and in urban centers. Since most Madrassas are similar to the ones that participated in the study and most prominent Madrassas Another important finding from the experiment is the high levels of trust, trustworthiness and other-regarding behavior in our data when compared to existing studies in the literature. It has been argued that hierarchical religions, such as Catholicism and Islam, weaken trust (Putnam 42 Our experimental setup focuses on games of trust and other-regarding behavior, and so we are unable to say how these students would play in games of cooperation (such as public-good provision) and norm enforcement, in both the absence and presence of punishment. We focus on trust because it is a central component of a community's social capital and is crucial for growth and development (Knack and Keefer, 1997;La Porta et al., 1997;Putnam, 2000;Guiso, Sapienza, and Zingales, 2004). These alternative games may shed light on other important aspects of social and economic life. However, implementing them generally requires feedback in real time, which is usually not possible with paper-based experiments. Given the low computer literacy of certain subgroups of our sample, we chose not to include these other games.

1993, La Porta et al. 1997). 43
A third important finding is that there is no evidence of in-group or out-group bias for any of the groups we consider in the investment decision in the trust game and no systematic discrimination in the dictator game. We do not find that Madrassa students behave differently when matched with students attending Universities with strong Western influence. Moreover, despite the fact that opinions collected from a subset of our respondents reveal that Madrassas are not viewed positively and are perceived to play some role in fostering extremism and violence, and a large proportion of the students have an acquaintance who died or was injured as a result of the violence in the country (Tables 1 and 2), we do not find that any of the other groups behave differently when matched with Madrassa students. It is possible that students make a distinction between mainstream Madrassas and radical ones. It is hard to know in practice which Madrassas are radical, and even harder to interview their students, so we cannot say anything about this.
Since our sample consists of Muslims entirely, the high level of trust found in this study is not consistent with this hypothesis. One possibility could be that Pakistan stands out in terms of trust when compared to other countries, but evidence from other surveys suggests that this is not the case. The World Value Surveys of 2000/2001 ask respondents from 70 countries whether "most people can be trusted" or one "needs to be very careful in dealing with people." In Pakistan, 31% of the people surveyed agreed with the statement that most people can be trusted, which is very similar to the UK (30%) and the US (36%), and well below countries that are highly ranked in terms of trust such as Iran (65%) or Denmark (66%).
A possible explanation for the lack of group membership bias could be the fact that groups are unequal in terms of two important social attributes: religiosity and socioeconomic background. Moreover, those attributes are negatively correlated across the groups. When groups are unequal, the lower status group tends to favor the higher status group (e.g., Jost, Banaji and Nosek, 2004). In our context, each student may feel they are of higher status in one of the dimensions when matched with another student. For example, Liberal University students may 43 There is an extensive body of work investigating the relationship between religion and trust or behavior in games arguing, for example, that Catholicism inhibits trust while Protestantism promotes it (Putnam, Leonardi, and Nanetti, 1993;Benjamin, Choi, and Fisher, 2010 in trust-based transactions and to contribute to a collective good within the community (Gilligan, Pasquale, and Samii, 2010), enhances altruistic behavior toward neighbors (Voors et al., 2010), and promotes local collective actions and political participation (Bellows and Miguel, 2009;Blattman, 2009). Our paper does not directly test for the impact of community exposure to violence on trust, and takes place in a country where conflict is still on-going. However, given how widespread violence is in Pakistan, conflict may play a role in explaining our results.
Another important aspect of our results is that Liberal Universities under-estimate the trustworthiness of Madrassa students, suggesting that an important segment of the society has mistaken stereotypes about students in religious seminaries, while Madrassa students overestimate the trustworthiness of Liberal University students. These incorrect stereotypes could negatively influence the social and economic interactions of those two groups outside of the lab.
Moreover, since graduates of Liberal Universities are the most likely to be future policy makers, their incorrect stereotypes could result in inefficiencies in society.
Overall, our findings offer some cautiously optimistic perspectives for Pakistan's future. 44 44 Our results focus exclusively on interaction between males, but it is also of interest to understand how various segments of the society interact with women, particularly in Pakistan, where gender discrimination appears somewhat paradoxical. On the one hand, Pakistan has the most imbalanced sex ratio in the world but on the other hand, women have prominent political leadership. In Delavande and Zafar (2011), we present results of experiments where students were matched with females from Islamic Universities and Liberal Universities. We find that Madrassa students tend to discriminate against women. However, because they tend to give and trust more than any Of course, the high and non-discriminatory levels of trust we find pertain to highly 38 educated groups. However, those groups are likely to be important actors in the economic activity of the country. Several African countries have experienced remarkable post-conflict economic recovery and one of the many channels may be that institutions, including trust, have improved as a result of the conflicts (Cramer, 2006). We can only hope that Pakistan will have a similar fate.
other male group, they actually treat women almost as well or better than other groups of males in the society. They simply treat men better than they treat women. We also find that there is interplay between the gender and social identity of the parties that interact.

Appendix: Instructions for the elicitation of beliefs
For this choice scenario, you will be asked two questions about Choice 1 [dictator game]. If this choice is chosen for compensation, one question will be picked randomly (by the roll of a die) at the end of the session to determine which question you will be paid for. The payoff will be given to you after one week once we have aggregated everyone's actions.
1) We ask you to guess the interval that contains the average amount that students from your institution chose to give to gender (where gender = {male, female}) students from "institution X" in situation 1 of Choice 2) Gender (where gender = {male, female}) students from institution X also participated in Choice 1 as deciders, in which they were randomly matched with students at your institution.
45 Actual instructions had gender of the student, and the name of the institution printed.

Now we ask you to guess the interval that contains the average amount that
"male/female" students from "institution X" chose to give to students in your institution         a T-test for the equality of proportion that expect more than 300 and the proportion of match group that actually sent back more than 300. This table also shows two pairwise hypothesis tests on the proportions between having a match from own institution type versus another institution type: (1) Wilcoxon rank-sum tests signi…cant at *p<0.10,**p<0.05,***p<0.01.
(2) T-tests signi…cant at + p<0.10, ++ p<0.05, +++ p <0.01. respondent currently studies in a Madrassa. b Self-reported risk preference on a scale of zero (totally unwilling to take risk) to 10 (fully prepared to take risks). c What are the maximum MONTHLY expenses (including tuition, room and board) that you and your family would be able to pay for you to be enrolled in school without any external …nancial aid? d Ability di¤erence is the perceived relative ability in Madrassa, with ability elicited on a 1 (highest rank) -100 (lowest rank) scale. De…ned as max (Ability in a Madrassa) -max (Ability in a non-Madrassa institution). e Binary variable that equals 1 if the Madrassa respondent would switch to a non-Madrassa institution if there were no …nancial or ability constraints. f P-value of Sargan's chi-squared test for overidentifying restrictions.  Table reports marginal e¤ects of a probit regression of trust game decision (coded as 1 or zero) on dictator game behavior and expectations of match's trustworthiness.
a Reported expectation of match's trustworthiness (center of the interval) on a 0-Rs. 900 scale, which is then scaled down by 100 to 0-9.
b Amount sent to match by the respondent in the dictator game on a 0-Rs. 400 scale, which is then scaled down by 100 to 0-4.

c(d)
Dummy that equals 1 if respondent is matched with a student at IU (Madrassa). Excluded category is a match with a LU student.
c Self-reported risk preference on a scale of zero (totally unwilling to take risk) to 10 (fully prepared to take risks).