(English captions & Hindi subtitles available)
About the Seminar:
Prof. Lowe estimates the effects of collaborative and adversarial intergroup contact and randomly assigned Indian men from different castes to participate in cricket leagues or to serve as a control group. League players faced variation in collaborative contact, through random assignment to homogeneous-caste or mixed-caste teams, and adversarial contact, through random assignment of opponents. Collaborative contact increases cross-caste friendships and efficiency in trade, and reduces own-caste favoritism. In contrast, adversarial contact generally reduces cross-caste interaction and efficiency. League participation reduces intergroup differences, suggesting that the positive aspects of intergroup contact more than offset the negative aspects in this setting.
About the Speaker:
Matt Lowe is an Assistant Professor of Economics at the University of British Columbia. His work is at the intersection of development economics, political economy, and behavioural economics. Most recently, his research has explored the effects of intergroup contact in two contrasting settings: caste in India and politics in Iceland. He received his Ph.D. in Economics from MIT in 2018 and was a postdoctoral fellow at the briq Institute from 2018-19.
Welcome everyone back to CASI's weekly seminars. I'm Tariq Thachil, and I'm the Director of CASI. This is our penultimate seminar of the semester, so next week's seminar with Mable Gergan is going to be our final weekly seminar of the semester. Thank you to all of you who've been joining us throughout this strange semester for our series of webinars, and we're delighted for this session to have Matt Lowe join us.
Matt, as many of you know, is an assistant professor of Economics at the University of British Columbia. He received his PhD in Economics from MIT in 2018, and following which he was a postdoc at the briq Institute, the Institute on Behavior and Inequality in Bonn, Germany, and has subsequently been at UBC. Matt's work has a lot of interesting features, but primarily uses field and natural experiments to study the origins, nature and malleability of relations between groups.
These groups can be those relating to caste as the work that he'll present on today, politics, religion or gender. He really has a wide array of interest, and some of the papers that he's been working on, I think the titles are kind of indicative of the array of topics that interest him. He has a joint project with Nishith Prakash and Raghuraj Rajendran on whether bureaucrats acculturate, Evidence From a Long-Running Natural Experiment in India. He has a paper with Gareth Nellis at UCSD: WhatsApp with India, The Effects of Social Media on Political Preferences and Violent Majoritarianism. And a paper whose title is near and dear to my own heart having worked on street vendors, the market structure and oversupply of deli food vendors that he's working on with Ben Roth.
So, I think that gives you a kind of sense of the array of interests that Matt has. The paper that he's going to present on today, Types of Contact: A Field Experiment on Collaborative and Adversarial Caste Integration is a paper that's received a lot of attention for revisiting one of the most venerable theories within social psychology, social contact theory, but doing so from the lens of a behavioral economist and someone who's got that particular kind of training, and I think it's received the attention that it has for a reason and I'm very delighted that we're going to host a discussion on it today.
Most of you are familiar with the format. We're going to have Matt speak for the first half of our hour together and then if you have any questions, please enter them to me, Tariq Thachil, in the chat box and I will then select and ask you to pose your question directly to Matt, and I will try and get to as many as I can. Apologies in advance if we can't get to all of them, but hopefully we can. So with that, Matt, the screen is yours. Welcome again.
Awesome. Thank you, Tariq, for that very kind introduction, and thank you everyone for tuning in to listen. It's actually cool I can see a couple of names I recognize including... I was just kind of remembering... Just as I was starting this project, and maybe even in 2016, I had cold emailed Simon [Socha 00:03:21] and kind of talked to him about caste and stuff. It's nice to eventually present to these people.
Okay. Before I get started, I actually have just one request which is if you could have your video on... So, some people will have reasons not to have their video on, and that's fine, but if you're someone who's happy to have it on then that's kind of helpful for me because it's nice to feel like I'm talking to people rather than to black screens.
Okay. I mean, Tariq actually already nicely summarized a little bit what this paper is about, so I can start by giving a slightly more intellectual history on what this idea is from social psychology. So, this is broadly a paper about the contact hypothesis. Now, what is the contact hypothesis? Well, what you should have in mind is that social psychologists started thinking about this idea basically in America in the 1930s and 1940s in the context of racial divisions, and so they did a bunch of correlational studies mainly about the integration of black people with white people and whether this can produce the discrimination prejudice you see.
It was on the basis of these mainly correlational studies that Gordon Allport, who's this very well-known social psychologist at Harvard, and he writes this book called The Nature of Prejudice. It's in this book that he states what the contact hypothesis is essentially, right? The statement is that intergroup contact reduces prejudice, but only under certain conditions. That's kind of the inference he makes from this existing work. If you want contact to work, you need the people to have common goals, you need there to be intergroup cooperation, you need these groups to have equal status and you need an authority figure that supports the contact.
The strange thing is in a way that this was a theory that basically was written in the '50s, but there's been essentially very little to no evidence on actually whether these conditions matter. What we've seen more recently over maybe the past 10 years is people starting to do more rigorous studies of the effects of contact in general, so some field experiments, some natural experiments, but where my paper comes in is trying to actually go back to this original idea that the type of contact should matter, and test that rigorously in the field.
Now, why should we care about that? My argument would be that if we want to get to the point where policymakers can actually design integrative policies in an optimal way then we want to know ideally also the conditions that should be there to maximize the effects of these policies. In a way the paper is its conceptual in that it tests this theory, but it's very policy-orientated too in trying to push us in a direction of, "How do we actually make contact policies work effectively as they can do?" Then the related point is that there are basically various pieces of evidence of naturally occurring contact and sometimes also experimental contact that either has null effects or somewhat limited effects, or even negative effects, and so by getting into the mechanisms of these different types of contact, we might be able to learn whether we can restructure existing forms of contact to turn these null and negative effects positives. That at least is the broad idea.
In the paper, what I do is I test primarily for the effects of two different types of contact and it really maps to this first condition that Allport talked about: The condition of common goals. So, I'm going to have a type of contact where people have common goals and I call that collaborative contact, and then I'm going to have a type of contact where they do not have common goals and I'm going to call that adversarial contact.
If you want to read the whole paper, or whatever, I won't get time for it all today. There's also work in the paper on other aspects of the Allport hypothesis, so also the conditions of cooperation and equal status. I might get chance to mention it a bit, but the main focus is going to be on common goals.
This was then basically the experiment. So, I ran this experiment in Northern India, in UP. This audience will know more probably than most audiences that the practice of untouchability is still pretty prevalent in India in general and especially prevalent in UP. If you look at the IHDS, about 37% of upper castes openly admitted practicing untouchability in UP, and then again, people will know... People are very starkly segregated, so we can think of that segregation by caste in terms of marital networks, but also social networks and geographically in the village. People will have been there and know that people tend to live separately in different [inaudible 00:08:08] by caste.
What you want in this situation for a contact paper is a means of integrating people to begin with, right? What I did here was I organized these cricket leagues in eight different rural locations. The nice thing about these leagues is that there's a broad interest in cricket across all castes, and in principle... Although I actually don't have great data on this, but in principle, one thing that's nice about the leagues is that even people who might be quite caste-prejudiced may still want to be involved because they like cricket so much. They're like, "Okay, I don't really want to be on a team with a low-caste guy, but I live for cricket so I'll make the sacrifice just this time." In my sample, 47% of them said they'd play cricket every single day, so that gives you a sense of the popularity.
Now, I had over 1000 people sign up for these leagues, and there's three main randomizations. The first one is that the recruits are randomly assigned to the leagues which were pure control groups. That's nice because that allows me to evaluate this program as a whole, which might be what we're interested in if we want to advise an NGO to expand this program or to do something else. Those that are in the leagues, I randomly assigned them to teams. That's the first key contact variation. Some people are going to have no one on their team from a different caste, so they have no collaborative contact, and other people are going to have a bunch of people from a different caste. That's what I'm calling collaborative contact. Then these teams were randomly assigned to opponents, so the variation there in your cross-caste opponent exposure is what I'm calling adversarial contact.
Okay. One second. Okay, so this is what I'll kind of go through today. First I'll show you that these two types of contact will have opposite effects on the demand for social interaction. People self-report their social networks. The more collaborative contact one of these men has, the more cross-caste friendships they end up with, whereas the more adversarial contact they have, the fewer cross-caste friendships they end up with. These are opposite effects. Each of these effects to some extent extend to the caste group as a whole, so you know it's not just that people become friends with the people on their team, or they lose friendships with their opponents. It also seems to generalize to people they didn't immediately interact with.
The social interaction is a self-reported outcome. As an incentivized and economically relevant outcome, I designed a trading exercise, which I'll explain. In this trading exercise, you can see collaborative contact also increases efficiency, so there's some economic gains from collaborative contact where there are negative though insignificant effects of adversarial contact.
Yeah, one thing that's difficult in a short talk is that I got a bit carried away with outcomes in this paper. So, rather than go through them one-by-one, I'll show you then these effects on an index putting all the outcomes together, so you can think of this index of cross-caste behaviors. That kind of summarizes the main results on contact in the paper, but there's a positive effect on collaborative contact on this index and there's a negative effect of adversarial contact. It's also a good way of summarizing the magnitudes, which I'll go through.
Then the final piece I'll emphasize is this program evaluation part. So, we know that collaborative contact is good, adversarial contact is bad, but what's the net effect? You've got these people in the leagues, they have bets of each and they have other things from being in the leagues. They interact with people that weren't even in the matches, and they get some money. Anyway, you put all of that together and it seems like participation has net positive effects. So, even though there are some negative aspects of contact, which you can isolate in this setting, the leagues as a whole seem like a reasonable policy to reduce intergroup differences.
Okay. Let me get into the heart of the talk. Let me describe first exactly how this experiment ran. Like I said, there were eight leagues, so it's one league per location and in each of these leagues, there are 20 teams, and in each of these teams, there are five players. These are smaller than you would normally have in cricket teams, and five was about as small as I could push it without them being really difficult to play these matches, but the advantage of having smaller teams here is that you have more statistical power to identify effects.
Now, how intensive was the intervention? Well, each team played eight matches over the course of a month, and each match lasted 40 minutes on average, so you have about five to ten hours of contact with your teammates and your opponents. This is what one of the matches would look like. Usually I'm spending time describing to American audiences what cricket is and how it's kind of like baseball but different. There'll probably be more knowledge of cricket than usual in this audience. A lot of people probably know, I mean, there's two men batting here. The collaboration comes through their batting partnership. They have to run between the wickets. They have three teammates who are sat by the sidelines, and then they're playing against this bowling team. There's the bowler who's about to throw the ball, and the fielder who's waiting to receive the ball.
Okay. Before these leagues start, basically had a field team visit villages and recruit people, putting up posters in the village and going door-to-door to some extent to ask people if they want to be involved. The eligibility criteria here is that they should be men aged 14 to 30, and the important thing actually to note is that I have... The science will come from three broad caste categories. These are not castes in Jatis, these are castes in the categories, and that's the way I'll be defining... Well, it's the way that I stratify the randomization is the way that I also define cross-caste exposure. If you're from the general caste, you have cross-caste exposure if you're exposed to either someone from OBC or someone from SCST, and vice versa, so that's going to be how that works.
These 1261 people signed up, and in the first step, I randomized 800 of them to be playing in the league, and I randomized 461 of them to be a control group. This control group are also going to serve as back-up players, which I'll explain through in one second. The league players I randomly assigned to teams. At this point, I rigged the randomization to make sure that about 35% of the teams were completely caste-homogeneous. In those teams, every player in the team belongs to the same broad caste category, and then the remaining mixed teams are across the spectrum, right? You could have one other caste in your team, two, three, or four. So, you have variation from no exposure to 100% exposure.
Now, one issue here is that when you call up these teams, which we would do the day before the match, in the morning of the match to play, not everyone can play every match. You won't always get full attendance, but like I said, with a team of five, it's really difficult to play a cricket match with only four people showing up, or three people showing up because you end up having to keep running all the way to the [inaudible 00:15:19] to get them more. To deal with that we had a system for drafting in substitutes, and this is how it works: I take these control group players, I split them into the three caste groups and I randomly ordered them.
This serves as a priority order for substitutes. If a player can't play, like if a general caste player can't play, one of these black dots, then the field team would call up the first priority general caste backup and ask them to play instead. If they can't, you go down to number two and if they can't, you go down to number three and so on. Just two things: Make sure you have a full attendance and also make sure that the caste composition of the team remains the same as initially randomly assigned, even though you've had to swap someone out.
Okay. These people are randomly assigned to teams, and then to create the adversarial contact, I wanted to have random exposure to your opponents. To do that, I randomized what the match schedule was. It turns out that to do, it's actually equivalent to a network problem. Having a schedule where each team will play one other team, and they'll play eight teams in total, you can think of that as a network where each team is a node and the link is whether you play a match with them.
Anyway, so it turns out there's algorithm, Bollobás' pairing method, that you can use to do that randomization so that it's not totally standard randomization, but this creates randomness in the schedule and that creates randomness in what fraction of your opponents belong to a different caste. Now, this is the empirical specification I'll use for the effects of contact. This is just keeping the 800 people assigned to the leagues, dropping those backup players for now. In fact, I'll have them dropped from most of the analysis I'll talk about. This outcome Yicl, that's outcome for participant I from one of the three caste groups in league L. All of these outcomes were measured one to three weeks after the league ended. For those interested in the contact hypothesis, there are certain things that we still kind of know very little about in terms of the effects of contact. My paper can't speak to the issue of persistence effects nor can many papers on contact actually. One thing that's still very open is understanding both whether effects persist and also what you can do to make effects more persistent.
Okay, so alpha-cl, these caste-by-league fixed effects. These are the strata that I need for identification. Then I have these two key treatment variables that are always both there at the same time because they actually [inaudible 00:17:51] to control the other one if you're including one. The first one is the fraction of my team that belong to a different caste, the second one is the fraction of my opponents that belong to a different caste. So, beta and gamma are giving me these effects of collaborative and adversarial contact. Then I control for the other-caste friends you have at baseline, and the variables I use for re-randomization, it doesn't really matter whether you include these controls or not, but this probably the best way to do it. I clustered the standard errors at the team level.
Okay, let me talk about effects. The first thing I'll talk about is these effects on social interaction. Again for people familiar with these literatures on caste for examples, a lot of the I guess interesting work in econ on caste has been about the importance of these caste networks for economic outcome, so for [foreign language 00:18:39] within your Jati, for migration to the city for occupational ability for one thing. The first thing we can ask in this experiment is how does contact affect these networks, and in particular, does it create new cross-caste links? The way that I did this was actually really simple. People self-report their social networks after the leagues are over. The survey goes to them with a tablet, they see a list of all the people who signed up in their area. The list is about 100 or 150 people. It's actually quite long. It's randomly ordered, and they go through this list and they do two clicks.
Sorry, the first thing I should say is they see the photo of the player and they see the full name. I avoid making caste salient. They're not seeing caste actually written down, they're just seeing a name, but the name usually signals caste. Then they make two selections. First, they select those that are either friends or people that they want to spend time with in future, and then of those, they select the people that they actually have as friends, so kind of a weak form of link and a strong form of link. Then for each person, I calculate based on these selections the number of other caste participants they selected.
Okay, so this is the first result. These two types of contact have opposite effects on the number of other caste you select for the weak form of friendship. Let me explain these figures briefly, they'll come up a bit. The figure on the left, that's showing you the effect of collaborative contact. The X axis is what fraction of my teammates come from a different caste. Each of these bubbles are bins, means that the different caste the distribution, right? Homogenous caste team. One other caste on my team, two, three, four. The bubble is reflecting the sample size in that bin. The bubbles in both sides are showing you these effects non-parametrically and then the dash lines are showing you the linear estimated effects, so the beta and the gamma that are also here. Then the right is the same, but this is the fraction of opponents from a different caste, and here there is less support of this variable, the way the randomization works. So, you actually have less power there.
Anyway, you can see then that there's a positive effect on this side, negative effect on this side. If you want to interpret the magnitudes a bit, if you're in a completely caste-homogenous team, you select about seven people, and then if you have up to four other caste [inaudible 00:20:59] you end up with 2.2 additional people selected.
Okay, and then on the right hand side, you have the negative effect of adversarial contact, which is larger in magnitude. These effects are pretty similar if you instead look at cross-caste friendships. That's the stronger form of the network link. This beta of one is actually nicely interpretable of every... For every four additional other castes on my team, I end up with one more friend from a different caste, or one more person I say is a friend. The effects of adversarial contact, again negative and again larger.
Okay. These types of contact have different effects on networks, but it's not clear from that whether that's because people are making friends with their teammates, or if they're actually expanding their network even beyond the people that they met. The way that I test that is I change the outcomes a little bit. For collaborative contact, I define an outcome that excludes the people that you played with. This outcome is looking at all the other people on the other teams, what percentage of other-caste participants on those other teams do you become friends with? Because they're on other teams, by definition you never played with them. Then for adversarial, I do something similar, but now you want to exclude the people you played against. Now, I go to the backups, and I look at the backups that didn't play a single match, and I say, "What percentage of them did you become friends with?"
So, if you do this, I find something fairly similar. It's not quite as strong here for collaborative, although I think... Yeah, I think it's actually stronger for the friendship outcome, but even when you exclude fully the people that you interacted with, you still see the positive and this negative effects. Some evidence for generalization. This is a part of the paper which actually I think could be a bit stronger. I can kind of give some more evidence on where this generalization comes from by ruling out network effects, some network effects. So, it doesn't seem like you become friends with the friends of your teammates for example, but it's actually not totally clear whether this is really about you changing your beliefs towards other people in the caste group in general, or whether there are other ways that these interactions are leading to other network links. I can talk more about that in the questions because I'm probably already going on too long.
Let me talk about trading. That was social networks and one limitation there is that it's also self-reported, but it'd be nice also to understand the economic effects of these types of contact and in a way it's incentivized. To do that, I turn to the domain of trade and there's a huge literature and economics of course showing that both geographic and social barriers can impede intergroup trade, and what's particular nice is that a colleague of mine, [Sue Ann 00:23:57] Anderson has this paper that demonstrates the importance of social barriers in this specific setting. She has this paper that shows that in UP, I think maybe it's in Bihar as well. If you're a low-caste buyer of water, then you have much higher yields if you're in a village where the sellers of water are also low-caste and that's because it's hard for you to contract and hard for you to trade with other casts for various reasons.
She would say that there's this trade-break breakdown between low and upper-caste. In this paper, I can now ask, "Would this contact effect such social barriers?" Ideally, we might actually want to run a study like mine in a context where you have this irrigation water breakdown, and then look at the causal effect on the yields of the low-caste. In practice, I have a bunch of 19 and 20 year olds running that don't... Not many of them I think trade in irrigation water, so to operationalize trading here, I designed an exercise that involved trade, and so it's more artificial than the irrigation water, but it also has the advantage of giving me a lot more control of how this works. You'll see kind of how in a second.
Yeah, so this is the trading exercise. After the leagues are over, the field team returns to these participants in their homes and gives each person two goods, so they give them a pair of gloves like these beautiful gloves here and they give them a pair of flip-flops. These things all have unique IDs. You can see here, this is the number 93 glove. Somewhat cruelly, these goods were intentionally mismatched. You can see that these gloves are two left hand gloves, so they're pretty useless. They would either get two left-handed gloves or two right-handed gloves. For flip-flops, they either get two left-footed or two right-footed flip-flops. What that does is it creates immediately gains from trade, they have a reason to go out and swap with someone so they can complete the set. We give them that, and then we tell them, "You have five days to trade. You can trade with anyone else that played in the..." Actually, didn't have to play in the league, they could be control group as well. You can play with anyone else that was signed up in this area.
Now, the additional thing which speaks to this point of adding more control of what you can do by doing an exercise like this is that I randomly assign people to monetary incentives to trade with a different color. So, you can see here the color of the sticker is green. There were three colors. So, some people would get money. It would be said, "We'll give 50 rupees if you can come back with a pink right-handed glove rather than a green one," and the reason why I did that was because I wanted to create gains from cross-caste trade, and I assigned the stickers so that they were almost perfectly correlated with caste. If you were a general caste, you'd almost always get green, and if you're SCST you'd almost always get red. This is by the way as well trying to introduce little tricks into the experiment design to obfuscate the caste component, which seems to have been actually relatively successful, but you might think people would work this out quite easily.
To be explicit, the advantage of having these gains from cross-caste trade is that you're creating more of a scenario that was true of the irrigation water situation. That you're creating a situation in which you actually need to trade with someone from a different caste to be able to get those gains, and if you can't, then you're not going to get those gains, so you're creating the efficiency motive.
Okay, 88% of these goods were traded, so most people are managing to complete a trade. So, you can see in column one here that this is a regression of whether you traded the good on the two types of contact and then also these bonuses. In the first column, you can see there's basically no effect, or at least no detectable effect on the extensive margin of whether you trade. All the action comes on the intensive margin of whether you trade with someone from a different caste. Right, so if you're in a homogenous caste team, about 52% of these goods were traded with someone from a different caste. That increases a good amount if you're assigned to these incentives to find a different color, and then in the full sample, so including those without the incentives, there's this positive effect of collaborative, negative, adversarial, but both of these are insignificant, although... Whatever collaborative contact is marginally insignificant.
There if you restrict to just the people with the monetary incentives, then you kind of see the actions. Collaborative contact here if you're going from zero to four is increasing cross-caste trade by 11 percentage points, and then you see that reflected in the payouts as well. They make more money. This is kind of demonstrating that when those efficiency gains exist, collaborative contact is facilitating cross-caste trades.
I guess I'm already at 9:30. Let me just take a couple of minutes on summarizing and then maybe two minutes on policy. Like I said, I got carried away with outcomes. In the paper, I have all these other things as well. The number of other casts you choose for a team for a future match and then also measures of trust. One thing I do which turned out to be very helpful I think... Actually this was based on a referee's suggestion, so I'm thankful to constructive referees. I put all of these measures that I had into an index, so just equally weighting the Z scores. Actually, I really don't like saying... I'm British, so I feel like I should say Z scores, but Z scores also seems kind of wrong, so I always feel a little bit conflicted when to explaining.
Anyway. This is now showing again the figure that I showed you before just for social networks, but now for the full index. This is a really nice way I think of summarizing the paper, or at least the effects of these two types of contact. You can see that if you put everything together [inaudible 00:29:57] the effect of collaborative contact that you can reject the null. There's a negative effect of adversarial contact where there's less power, but you can still reject the null at the 10% level, and you can also confidently reject that these things to the same, so this is actually I think pretty clean and pretty nice. Now, there's also a bit more clarity about the magnitude. If you go from zero contact to four contacts, it's about a .2 standard deviation effect. Then this is the linear effects, so this is if you could go from zero to four adversarial, it would be -.4. You also kind of summarize that adversarial contact is a bit stronger than collaborative contact in the opposite direction.
Okay. Let me then talk very briefly about policy. This uses a variation that I haven't really discussed so far, I haven't used so far. To get the effects of league participation, I want it also to introduce the control group. This shows the three outcomes: Friends, trade and the index. Three groups. Pure control, which is actually the low priority backups, homogenous caste teams, mixed caste teams. For all three of these outcomes, mixed caste teams have positive and statistical significant, the 1% level effect relative to control, but what's true as well is that even the homogenous caste teams either have a positive or kind of somewhat null effect relative to control. Even if your homogenous caste team is kind of a positive effect overall, or a null effect, and so we can say overall the lead participation seems to have a positive effect.
Then going back to the very start. I said Allport has defined the contact hypothesis as saying you need four conditions. I also have variation in the experiment allows me to shed light on two more of those conditions. One condition is this condition of intergroup cooperation, which is the idea of conditional on the goals you have. Do you have incentives to really cooperate with each other, and so here I cross-randomized monetary incentives. Some of the teams got paid based on individual performance, and some got paid based on team performance. It's the team performance teams that have these incentives to really help each other out and collaborate. Anyway, it turns out that those incentives didn't matter at all. Whether you have the team incentives or individual incentives, the effects of collaborative contact are pretty much the same.
Then the other variation I can look at, which is not randomized, but I can look at this condition of equal status. You can basically see in various ways that if there is a mixed-caste team, the different caste do not really have equal status in any reasonable sense, so the other castes are more liked to be picked to be captains, and there's also evidence of discrimination of lower castes, even if you control for ability, but despite this lack of equal status, there's little caste heterogeneity of the treatment effect. It seems to be that even though you have this integration environment in which the caste hierarchy is somewhat reproduced, you still see positive effects with the lower caste for example.
Now, one way to summarize these two results, which I feel like maybe is not actually in the paper, but has struck me that it's probably the most kind of reasonable summary is that if you literally what my study says, and it's just one study, so I don't know if there's any reason for this to be very general, then what my study says in a way is that providing common goals for these groups is a sufficient condition for contact to work. Once you do that, once they're on the same team, you can kind of mess with the contact a bit and it's still going to have positive effects. So, you could mess with the incentives and you could also mess with the status, and it doesn't really matter. That's one way of actually thinking about the punchline of the study including this stuff is to say that the common goals appear to be pretty sufficient for creating these positive effects.
Okay. Let me summarize. I guess I've actually just said, let me not say anymore on that and I can just say that in future, I think an obvious direction for research would be to continue to study what conditions matter for contact, but I think the kind of point I'd make is that it's not obvious that we should really stick to these conditions of Gordon Allport. I think sometimes we're a bit wedded to the past because these people become famous and whatever, but what he came up with was based on a bunch of correlational studies which naturally you wouldn't trust that much. Now, I think there are new contenders for conditions that we think might really matter for the effectiveness of contact, and one which I think's kind of interesting actually is understanding the importance of group size. It's very different having 50% low caste, 50% high caste in a team than say, one low caste person surrounded by nine of the high caste people. That's actually I think one of the condition that people might be doing I imagine would mediate the effects of contact.
Okay, so I'm done and I'm excited to hear any questions and criticisms and clarifications people have.
Okay. Thanks, Matt. Thanks a lot for that presentation, for giving us a lot to think about. The first question we have is from Sumitra. Sumitra, do you want to ask your questions to Matt?
Thank you. This is a great presentation. I actually have a lot of questions, but I'm only going to ask you a few. Feel free to ignore even those. First, I was curious to know whether there's something that's theoretically different about social contact in sports as opposed to in other settings. So, I'm curious to know why you chose cricket relative to say, a classroom setting where you can imagine having a competition even in a classroom setting? If you've read, and I know you have because you cited the [Shana Warren 00:35:43] and [Alex Gackle 00:35:44] paper where they have students from different races compete on group assignments together. Is there something theoretically different about sports and should we expect the context itself to have an effect on outcome if you had to rerun this in a non-sports setting?
Secondly, this is small, I might have missed it, but did you say you had any outcomes that measured prejudice and discrimination, and if yes, what were the effects? Thank you.
Good questions. Yes, I don't have a great answer. What I wonder about sport is that... I mean, I guess the idea I have in mind on your sport question is that essentially the emotions might mediate the effects of contact and that sport much more euphoric. Well, it also has kind of big negative emotions of losing together, whatever it is, but it's much more I think emotionally turbulent than doing homework. That's not really a theory, it's just a kind of guess. Yeah, I wonder if you could have an experiment where... You can imagine an experiment where you also randomize whether the contact is based on sport or education and then you can compare those two things directly, and maybe you'd see that sports would be more effective. You could also imagine trying to just isolate the effects of having an emotional connection, or whatever with the people you interacted with.
I think it's a good question. I don't have a good answer. The one thing I do think, which is not really an answer to your question is that I think an advantage of sport over education, which I mentioned a little bit at the start of the presentation was that there might be selection into contact, which is actually very important. I think this is something that there's no good work on yet that I know of, which is that it might be hard for intergroup contact to work well because the people you really want to get contact might not want it. The really prejudice people might not want to integrate. Sport might overcome that because people love it and it's enjoyable, whereas I'm not sure about vocational training purposes.
Okay, second question. There was only one outcome I think that I have... Well, there was one outcome I have in the paper that I didn't talk about which is that people vote after leagues are over on which players should get to receive cricket training, like in Varanasi. That I think of as kind of a measure of discrimination, but what you might mean more is attitudinal measures, and that I didn't have at all actually. I go back and forth whether I regret that because I guess one reason I didn't have that is because when you're doing the end line, you're surveying 1000 people.
Once you've surveyed the first person, you've asked them what they think about caste marriage, then obviously they can kind of tell everyone, "Oh, we're getting asked about caste marriage." Then the game is out, or whatever, the cat is out of the bag and everyone knows it's about caste. That was what made me a bit reluctant, but now it's a shame because there were other studies that also look at attitudes. For example, [Salma Moosa 00:38:54] has this paper on... I think she mostly doesn't find effects on attitudes. Now I am very curious of were there even effects on attitudes. I just don't know yet.
Okay, thanks. Next we have Ashley Simone. Simone, you have a question, go ahead.
Hi! Hi, Matt. Hi, Tariq. I also had the same question about different variables. I think in this literature, it's a little bit of a mess, but all the types of different variables, so it's not only the attitudinal variables, it's also among the behavior of ones, I was wondering if you could tell us a little bit more about how your measures actually correlate to some of the, let's say, more game-based measures we get in experimental economics, let's say, measures of altruism or collaborative behavior in the context of a game, which it seems to me to reproduce this, but it's also a little bit different. There's a question there.
If I had time to sneak one quick second one in is I keep on hearing about the sort of punchline in terms of policy of your findings, which is if the effect of adversarial contact is stronger than the effect of collaborative contact, what does it say about sport, right? We just not play. Is that the idea? Because it seems to me that overall, I mean, the world is worse off by us playing sports then, is that the idea? Anyway, just stop here.
Okay. Okay, so on your first question. One answer is that this other measure I just mentioned then about voting. I'd actually hoped for that to be kind of like a naturalistic dictator game. In this case, instead of allocating money, you're allocating this reward of who should get this cricket training. In the end, I actually think I made a bit of a mistake here, but I didn't use any dictator games. I just used this measure, and when I use this measure, I do find that collaborative contact reduces caste bias. Well, it reduces caste bias by about a third basically. So, there is caste bias in who you vote for, but it turns out to not be the cleanest measure of social preferences, which I wanted it to be because when you're deciding who to allocate the training to, people allocate people who are better at cricket. Then you have this issue that it might be capturing statistical discrimination, that you learn more about their ability... I tried to deal with that, and to some extent, you can show that it's probably not fully explained by statistical discrimination changing, but it's not the cleanest.
That's kind of the first thing, and then the second thing is that I do have trust games. People play trust games and that's obviously a kind of common measure in experimental econ, and with the trust games, what's interesting is that there's actually some evidence that both types of contact reduce trust, but adversarial contact reduces it more than collaborative contact. You still kind of have the divergent, but the surprising thing there is that collaborative contact is actually causing people to give less in the trust game afterwards. That was a bit of a puzzle to me, but something I find which also helps a bit is that conditional on how much contact, the more pre-existing friends you have in your team, the more you give in the trust game. There seems to be maybe this thing that the more in-group contact you had, or the more contact you had with people you like and that you know then the more trusting you are afterwards.
This is one of those things that's actually is... Was the biggest puzzle in the paper for me, but maybe actually shifts your [inaudible 00:43:04] the most because it's unexpected. That can also kind of show you a downside of contact, that contact with out-groups is not as fun as contacts with in-groups and that might have consequences. Dammit. Forgot your other question now. What was the other one?
Just the fact that-
Oh, wait. I know, I know.
Should we play sports.
Yeah, yeah. I know. On that one, actually the... This was also kind of a puzzle. I think that the answer is that we should play sports because if you compare with a control group, you still have positive effects overall, but then the question becomes: Well, if adversarial contact has a negative effect, why are there positive effects overall? That is not even really something I can unpack, but part of it I think is that when you play these leagues, you're not just treated by the matches, you also hang out after the match and you meet these other people, and that's probably the same regardless of who you had on your opposing team, so this kind of spill-over effects that are also part of program. It ends up being I think that no, we should play sports. You don't need to cancel your weekly football matches, or whatever.
The next question is from Hashar. Hashar, go ahead.
Yeah. Hey, Matt. Nice paper. Two questions. The first one's really straightforward which is just I wasn't clear in the signups what proportion of the eligible population that you were getting, so I'm curious about the selection and how much people knew going into the leagues. Then the second question is about the network access mechanism, and I'm curious to get your thoughts on what role you think community size may play in this, in the generalizability of the results. Where I'm coming from is that... So, when you think about the networks of the teammates, or the opponents, the chances that you kind of already know those folks from prior interactions in a small [foreign language 00:45:01] is perhaps a lot higher and so do you think that's sort of coloring the results to some degree? Could the results potentially be more muted in a larger setting where you don't have as many prior social interaction to go on?
Yeah. Those are great questions. The first question on selection. Yeah, unfortunately I don't have a great answer. What would be nice is to have a census of all the eligible people and even to have behavioral measures from them, so you can characterize selection, but I don't have that. I just have really 1261 people that signed up. I might have worked this out at some point, but essentially, I'm having 150 or so people sign up from a [foreign language 00:45:50], so a cluster of villages. This is a very rough number, but it's probably something like 10% the eligible population. Maybe slightly more. Maybe 10 to 20%. You might have about 1000, or 500 to 1000 people who are eligible. I think it's roughly that, but unfortunately I can't really characterize the selection on prejudice.
Okay, what was your second question again?
It's about the role of village size, community size in a-
Okay. Yeah, so let me tell you one thing that... This doesn't quite answer your question, but one thing that's a little bit suggestive, which is somewhat related is that some of the leagues were better attended than others. When I look at the generalization effects, the effects on people you didn't play with, they're much bigger for the well-attended leagues. So, what it seems like is happening is that people become friends with the other people that came to watch the matches. That's actually why on that slide I said it's not clear whether it's generalization just in terms of beliefs, or whether it is some kind of network spill-over.
Anyway, so if you take that and you kind of connect it to your question, I think it does suggest that getting the generalizations or the spill-overs is going to depend on things that are maybe not too obvious, but maybe stuff you mentioned of are we playing in a pitch that's in the middle of a village. We all live close together, or are we going miles away to a place to play at a pitch with people who we're never going to see again? That I actually find very interesting because I think another type of study, which would make a lot of sense for someone to do is one that instead of asking a question does contact affect prejudice, or whatever, is one that asks the question, "What do you have to do to make the effects of contact generalized?" One thing could be exactly this community size thing. So, I think that's actually a very interesting kind of set of questions that are still pretty much unanswered.
Okay. The next question we have is from Sanjive.
Yeah. My question was how common are cricket leagues... I mean, this is from a very generalized understanding, but if we have a spontaneous [inaudible 00:48:15] and so, these participants would be alerted by how organized this was, and how can you control for artificiality and expectations in that case?
Yeah. Okay, cricket leagues are fairly common at least in this area, but what's different is that they're usually district-level leagues where a village would send a couple of teams. That was my impression. At some point, we also collected some data on that. A village would send a couple of teams, and they would be the best players in the village and they would be pretty caste-homogenous, these teams. What I did was introduce a league which had much broader participation that anyone could play, even if you're not that good at cricket.
Now, how did I control for artificiality? Yeah, there's no way that I kind of explicitly controlled for it, but what I was mostly worried about was... Yeah, I was more worried I guess about demand effects that would relate to people knowing the study is about caste. That's kind of what I spent some time trying to deal with by obfuscating a little bit the purpose of the study and the purpose of some of the measurement, but on the artificiality of the leagues itself, that's just kind of it is what it is, really.
Just as a follow-up, did you ever ask at the very end why they thought the study was done?
No, I didn't, but this was my job market paper and I've learnt... This is one of the lessons I think I've learnt. I definitely would ask people in future. It was just a schoolboy error on my part.
A couple of quick questions that I had, Matt. It is really interesting stuff. One is as someone who hasn't done this kind of work, but has really enjoyed engaging with it, I think one of the things that's really impressive about the paper is as you said, the kind of going outcome crazy, but the question that sometimes raises for me is how do I think about what the treatment is? I understand that you're randomizing this treatment, but you're also doing a study where you're exposed to a certain amount of cross-caste, adversarial, or collaborative contact, but then I'm also lumped in with that, getting a lot of different things on these outcome measures. I'm getting gloves that I'm having to trade. I'm doing a trust exercise. So, do you think that even any findings that you get would even generalize too if we just played the sport?
In other words, I'm getting a treatment, but then it's being reinforced by the number of kind of outcomes that are being collected on me, which are specific to... I don't know. I'm just trying to think about that because so many of these studies are really great about collecting outcomes, but I sometimes wonder what the kind of take away from that is, because that also kind of heightens some of these concerns of artificiality, or even how much of this is just about the sport. So, just kind of thinking about how that relates to this.
The second is were you surprised? In some ways, I was surprised that given that you conceptualize caste in terms of these broad aggregate categories, were you surprised that you got finding them and just thinking about how caste operates in social settings and in Indian village that these aggregations are often not very socially salient. I mean, to be an OBC in UP is a huge cluster of different kinds of Jatis and same with the SC, same with upper-caste. So, just trying to think about whether that surprised you, whether you see that actually in individual villages, there are actually only one or two Jatis within these categories, and so while you're lumping them under these... I'm just trying to think through that. That was actually kind of surprising to me and not really commented on that.
Yeah. Let me answer your first question and then I'll ask you to remind me what the second question was. Okay, so first question on reinforcing. I mean, yeah I think I don't worry too much about that. One issue which I never thought of until you mentioned it is that the trading exercise is maybe the biggest problem for this point because it actually does create contact, right? You have an exercise where they're actually going to go out and find people, and it's true that I have that exercise just before they do the other outcomes. That's kind of a problem because if you have a big treatment effect on cross-caste trade then you're then confounding your own treatment by accident.
I think it doesn't matter that much because the treatment effects are not so big that you'd... That a 10% [inaudible 00:52:43] effective going and having this one interaction with another caste through the trade is very small relative to the treatment, the variation you create from a team assignment. I think you're kind of right that there's something there, but I don't think it worries me. What worries me a lot more, and which is something I'd really like to do better in the future is I really like the idea of having completely obfuscated outcomes, so outcomes where people don't realize they're in an experiment.
I'll give you an example, which is something that I'd hoped that me and Gareth could try for this WhatsApp experiment. I think it's not going to work actually, but it's this... Okay. I don't know. I guess we still have a bit of time. Does everyone know Stanley Milgram, who was famous for these electric shock experiments? He's also well-known for this thing called the Lost Latter Paradigm, which I just think is really cool, where he dropped off envelopes in different communities and the envelopes would either be addressed to some black civil rights organization, or to some white supremacist organization, different things like this. What is interesting is how many of these envelopes will actually get sent, and he's using that as a measure of local support for that organization. He's trying to say, "This is much better than asking someone, 'Do you support black civil rights?' Because they might lie to you."
Anyway, so another version you can do of that is you can do instead of a lost letter experiment, you can do a wrong number paradigm where you text people pretending to be from someone else and then you can randomize that it's from a Hindu or a Muslim name. Anyway, I like this kind of thing because... Then the outcome is just whether the person replies and says, "I'm sorry, you've got the wrong number," because that's kind of the nice thing to do in that situation. Anyway, I like this kind of outcome because the person has no idea they're in an experiment and they're kind of dealing with this in total privacy of their own home. None of their friends know they got that text, and so, that probably actually does capture something. Anyway, all that's to say is...
We piloted this and the reason why I don't think we can do it is because we piloted it Bihar, and I guess what's weird in India, maybe people here already knew this, maybe it's just Bihar is that a lot of people call back to these wrong numbers, which you'd never do that in England, I think. You'd never call the person and then tell them they got the wrong number. That was very strange, and that kind of ruins the paradigm because you don't have a real person to answer the phone.
Okay. That's actually kind of more in the direction where I want to go, which is having outcomes where they don't know they're in an experiment. Okay, now you had a second question? Can you repeat what it was?
I have forgotten my own second question. It was obviously very profound.
Can anyone remember what the second question was?
No, I think the second question was about the Jati versus Varna. Kind of these artificial aggregations. I mean, you're finding effects, but I'm just trying to think through what that means as a [inaudible 00:55:56] setting.
Yeah. I think the reason why it works is that... When you say an OBC has many Jatis or whatever. One reason why it still works is that if you're an OBC, you're still much more likely to be friends with another OBC Jati than a general or an SC Jati. These three groups broadly still capture social segregation. The other thing is that I think the modal Jati in my experiment, in a given broad caste group comprises 60% of that caste group. It almost just approximates cross-Jati contact.
I think one way to think about it is just these right-hand side variables are extremely highly correlated with... If I defined it as cross-Jati. I always thought that a referee was going to ask me to redo everything using cross-Jati contact, and I was glad that they didn't because it's just annoying to have stuff where every table changes and you have to have a full second version paper. They never actually asked me to do that, and so I never did. I mean, maybe I should have just done it out of the goodness of my heart, but my sense is that it's just really, really highly correlated, so it doesn't matter that much.
Yeah. If you were submitting to sociology journals, you'd have definitely got that from a referee, but we will leave it here, but if I can just say one question that I wasn't able to get to, just something for you to think about is whether... There was a question about whether this would generalize to a Hindu-Muslim divide. You don't talk about Muslims. I'm not sure the presence of them in your villages, whether you just coded them according to their... Some of them can be coded within an OBC spectrum, but I'm not sure how you thought about that and whether... Again, this is a question that we can leave you with since we're at time, but actually we have one minute. I don't know, have you thought of that? A minute you have to respond to-
Yeah. I haven't thought that much. So, for my study, I chose villages that had very few Muslims. The idea was just to basically not have to worry about that. Yeah, I don't know whether it would generalize. The only thing I can say is that I'm going to learn I guess a lot more about that because the next similar field experiment I'm working on, this project with Gareth Nellis, we're doing this project on the effects of WhatsApp in West Bengal ahead of their state elections next year. That's going to be trying to understand effects on Hindu-Muslim relations and all that. Yeah, that won't be about intergroup contact, but then I'll actually learn a bit more about that I guess.
Okay. Well, thanks very much for patiently taking all of our I guess double-barrelled questions. Nobody's asked a single question, so thank you for your patience in dealing with us and it was a pleasure to engage with your work, Matt. Yeah, all the best and thanks everyone for joining us, and we'll see you next week for our last seminar of the semester with Mable Gergan. Thanks very much.
Thank you, everyone.