Kartik Srivastava on Referral-Based Hiring, Caste Networks, and Breaking Barriers in India's Labor Markets

SHRUTI RAJAGOPALAN: Welcome to Ideas of India, a podcast where we examine academic ideas that can propel India forward. My name is Shruti Rajagopalan, and we are kicking off the 2025 job market series, where I speak with young scholars entering the academic job market about their latest research on India. 

Our first scholar in the series is Kartik Srivastava, who is a PhD candidate at the Kennedy School at Harvard University. Before this, he received his bachelor's degree from Yale University, where he majored in Economics and Engineering Sciences. 

His research focuses on development economics, labor economics, and political economy. We spoke about his job market paper titled, Familiar strangers: Evidence from referral-based hiring experiments in India. We talked his large-scale experiment at a footwear manufacturing firm in Delhi, on how referral-based hiring improve firm productivity, cohesion, and inclusion, differences in hiring between higher caste versus lower caste networks, feudalism and labor opportunities, and much more.

For a full transcript of this conversation, including helpful links of all the references mentioned, click the link in the show notes or visit mercatus.org/podcasts.

Hi, Kartik. Welcome to the show. It’s a pleasure to have you here.

KARTIK SRIVASTAVA: Hi, Shruti. Great to be here. Thank you.

Referrals and Retentions

RAJAGOPALAN: I’m very excited to read your paper because it touches on so many things that I really care about. You’re doing multiple experiments in the labor market. What you’re really trying to study is if someone just drops into a particular firm or a team from nowhere, is that different from when someone actually comes in through a referral? Now, this matters for every single job market category that we can think of, including academic job markets, because we write such detailed letters of reference for every single candidate.

You look at this in a very practical way for employers in high-turnover, low-wage settings. If a firm deliberately steers employee referrals towards an underrepresented group, which in the context that you study happens to be relatively lower-caste workers in the pecking order, how does this referral system that may bring in an additional lower caste worker actually help firm output or team output, team cohesion, and so on and so forth?

You do the experiment within a firm. You also do the experiment in the field to look at the broader context of the labor market from where people are hiring. What you find - which shouldn’t be surprising - is that when you hire an additional lower caste worker as someone’s referral, as opposed to them just showing up as a lower caste worker in a particular team setting, that changes everything. It changes cooperation within a team; it changes cohesion; it changes how people view them. Incidentally, it also increases output. It allows the new worker who’s been hired to actually stay much longer in the firm, and all these hosts of positive effects.

First up, did I get what you’re trying to do right? 

SRIVASTAVA: Great. Thank you for that summary. I think it gets at a lot of the right points that I’m trying to hit with this work. The story starts with what I think of as quite a striking fact in labor markets in our settings in India and elsewhere: where turnover or the retention problem, firms struggling to keep and retain workers, is about thrice as large in LMICs [low- and middle-income countries] than in richer countries. Workers transition in and out of jobs at a much, much higher rate.

That’s not necessarily bad from a worker perspective because, in many cases, job transitions can lead to upward mobility and climbing the job ladder, as we like to say in labor economics. The access to these transitions tends to be deeply unequal in a lot of these settings. Some groups of workers are more freely able to move across space sectors, firms, and jobs, and others are just stuck in low-value work or underemployment or unemployment.

On the firm side, this is really costly because anytime you talk to a firm in a developing country, I can bet some money that retention or turnover will be probably in the top three problems that they face in these conversations because they have to spend money and resources and time hiring workers, training them, and then lost output, et cetera. My story is about how network-conscious hiring enters this picture.

Networks are a really important part of this picture because on both sides of the market, networks play a huge role. For the workers, they help them find jobs. You ask your friends where to work. For the firms, it’s really important because, again, to reduce the search and matching costs, they’ll ask their incumbent employees. Also, referrals tend to be the default dominant hiring technology in these settings. Referrals have been shown in the economics literature to lower search costs and moral hazard. The reputation of my incumbent employee is on the line, if the new worker does something wrong. Then that allows me as a firm to outsource some of my training costs onto the incumbent worker, et cetera.

Networks as Barriers

In practice, what tends to happen with referrals, though, is that they tend to disadvantage net groups of workers that are what we think of as having bad networks or poorer networks. Referrals tend to be saturated by the privileged groups and well-connected groups. This sets up this very peculiar equilibrium where firms want to hire and retain workers at a high rate, there is this big pool of workers that wants to take these jobs, has a high willingness to accept these jobs, but doesn’t have a way to break into the network and to get these jobs.

We have this high turnover and high unemployment problem in equilibrium, which is an odd thing. While a lot of people have studied this, I think what I’m trying to do in this paper is to pool a lot of these things together. I study this in the context of caste, as you said, in India. I don’t need to introduce caste to your listeners, but if there was ever a social stratification system to study this dynamic of occupational segregation and networks, I think caste is what an economist would design in a lab to study something like this.

The Factory Experiment

I do this through a large-scale experiment at a footwear manufacturing firm in Delhi. What tends to happen in this firm, in status quo as usual, is most workers are hired through referrals. What that means is anytime there’s a vacancy in the firm, a supervisor of the team where there is a vacancy will ask one of his incumbent employees to bring a friend, a previous coworker, a relative, et cetera. These opportunities to bring referrals are allocated discretionarily by supervisors, so they will choose who they ask to bring a friend or a coworker.

RAJAGOPALAN: More generally, one would expect that supervisors tend to be of upper caste. Is that part of the puzzle that we’re trying to solve? Given the identity of the supervisor, this creates barriers to entry for those who are of lower-caste groups who may already be workers but are not exposed to that referral system or the benefits of that system.

SRIVASTAVA: Absolutely, that’s part of the picture. Though I do find that there are some lower-caste supervisors, there are so few of them that they behave as if they’re upper caste. There’s this mimicking thing that we see in the literature all the time. There isn’t enough of them to drive change at that level. That’s a more complicated process because, again, in these kinds of high-turnover, low-wage settings, you don’t find firms promoting people that often. They hire from different pools, and that’s a different complicated dynamic.

I find at baseline that upper-caste incumbent workers are three times more likely to receive an opportunity to refer in their tenure, even though there is no observable difference between upper and lower-caste workers. My experiment then flows naturally from that. I say, “Okay, what if the firm then allocates the opportunity to refer disproportionately towards its lower-caste incumbents?” There are lower-caste people working in the firm.

What if we just, in the treatment group, made them bring their friends or coworkers—chose a random lower-caste worker anytime there’s a vacancy. Relative to a control group where, again, the discretion is taken away from supervisors there as well, but the referral allocation is done randomly, irrespective of caste. So, a worker is chosen regardless of his caste. I’ll say this repeatedly here because in the setting there are exactly zero women in the factory that I work in, which is a context that I want to work on in the future, but for now, yes.

RAJAGOPALAN: “Lower caste” and “upper caste” are handed to us over two millennia. They’re not normative judgments that you are making. I guess that’s another additional disclaimer to add here.

SRIVASTAVA: Absolutely. That’s another disclaimer towards adding for sure. Thank you for that. The third disclaimer is that these are extremely coarse categories, so your Indian listeners will probably quibble with the brush strokes that I’m making here. It’s true that the analysis necessitates this kind of broad categorization. Although I’m discretizing what is inherently a continuous spectrum of hierarchies—infinitely hierarchical—and the main insights still go through it, but you just lose statistical power, if you look at a granular level.

RAJAGOPALAN: When you study something like this, it’s quite clear what your intervention is, which is, now you open up the referral system to groups that either didn’t have access to that referral system, or even if they did, their referrals weren’t taken up as much. Now, what are the outcomes we think about, because the outcomes are different from the point of view of all the different participants?

From the point of view of the person who’s referring, the outcome is, “Did they actually get hired?” From the point of view of the remaining team members, it is, “Does this person actually fit well into our team, and are they a good team member?” From the point of view of the supervisor, it is both team output, like, “Does this person add to the team output, and at least certainly doesn’t diminish it?” but also, “Am I going to have to keep doing this hiring and retraining thing, or is this person really going to stay?” You’re looking at these different kinds of outcomes and what is desirable for different groups. What is a good way to think about this, given your experiment?

SRIVASTAVA: That’s a great point. As you said, each agent in my story has their own point of view. The design had to be careful to account for the differences that might show up. This experiment could be good for some and bad for others. From a logistical design sense, I was careful about the way I did it. 

I survey each agent that you described. That’s useful. I have firm-level data on output and retention, which is at a high frequency, which is helpful to see dynamics. Another important feature of my design is that before I started the experiment, I elicited from each incumbent worker a network of potential referral candidates. I asked them, “If you could bring anyone to the firm, whom would you bring?” This is prior to even myself knowing whether a person I’m talking to is going to be in a treatment or a control team. They give me names. Now, this gives me three or four times as many people to track outside the firm as there are in the firm.

Over time, what I do is, after the experiment ends, I then survey all those people, the network that I built at baseline. Some of these people have friends who happen to be in control teams. Some of these people have friends who happen to be in treatment teams. Some of these people are upper caste. Some of these people are lower caste. There are all sorts of combinations of people, both by caste and by treatment status.

That is the variation that I use to say, “If you happen to get a job offer, being a lower-caste person who’s in the village versus your neighbor who doesn’t happen to get a job offer, remains in the village, how does that improve your labor market prospects? On the flip side, if you’re two upper-caste people who are already in the city, one of you gets a job offer from a different factory and one of you doesn’t, how does that affect your labor market trajectory?” That’s the exercise that I do at the end of the paper to get at the welfare effects.

Retention, Productivity, and Cohesion

RAJAGOPALAN: What do you find?

SRIVASTAVA: I find that on the firm side, the first question is, “Do more lower-caste workers join the firm?” The answer to that is yes. Their representation increases substantially. The first effect you see is on the retention. I already knew before the experiment, having collected this baseline data, that the outside options of lower-caste workers are dramatically worse on average. They’re less likely to already be in the city, and they’re more likely to be unemployed. So once they join the firm, they’re more likely then to stick around for longer. Retention improves by about 40% in treatment teams. Turnover is much less of an issue there.

The risk of this kind of experiment, when you introduce more diversity in teams, is that you’ve done the diversity bit, but that might make cohesion costlier. We have all this other work that suggests that cohesion, especially in tasks that require coordination, gets worse with more diversity, with mixed teams, at least in the short term. So there’s some short-run costs that the firm has to pay.

I was expecting to see similar effects in my setting, where there might be a short-run decline in productivity or output as a result of these cohesion costs in treatment teams, but I did not find that in my setting. If anything, productivity improves over time. The improvement comes almost entirely from this retention. So, more stable teams tend to be more productive in this setting, which is completely intuitive. Those gains are coming from high-coordination tasks as well. So, tasks where it really matters that the four of us working on something together all stay in the same room repeatedly, relative to one of us leaving every week, will make my task costlier and less productive.

I think that productivity positive effect, maybe you could have expected, but you would have guessed that there would be some short-run costs. If I don’t see those short-run costs, then that’s a puzzle. I have some other results on the firm’s side where supervisors have to spend less time training workers, naturally again because retention is not as much of an issue anymore.

RAJAGOPALAN: You have a friend that you can ask questions, right? 

SRIVASTAVA: Exactly.

RAJAGOPALAN: There are so many simple, intuitive things going on here.

SRIVASTAVA: Completely, and that goes back to why referrals work in the first place. The supervisors are able to outsource these things to the incumbent workers. They’ve just done that on steroids. I think there’s other stuff going on at the firm we can come back to, and especially in terms of how the firm then adjusts once the experiment ends. My takeaway at this stage of the study was, “Okay, why am I not seeing the declines? This is odd.”

Part of it could be that I’m maybe only getting self-reported data on cohesion. I ask people how much they like working, how much cohesion is there on the team. From the literature that those self-reported things can be a bit hokey, you want to make sure you’re getting those right. Or it could be that maybe cohesion doesn’t map onto productivity like we thought. The third answer would be that maybe cohesion costs truly don’t get triggered in this experiment, which would be at odds with the literature. Then I invested in doing two lab and field experiments to investigate that question.

RAJAGOPALAN: Before we get into the lab and field experiment, which is also fascinating, could it also be that some of this could just be self-selection, like the kinds of upper-caste workers and lower-caste workers who work in these settings are already selecting into willingness to work with a more diverse group? Because there is other experimental literature where people are willing to take a pay cut to have to not work with someone from another caste, upper or lower, right?

SRIVASTAVA: Completely.

RAJAGOPALAN: There’s a fair bit of literature on how people think about cohesion, and here you’re already in Delhi, which is fairly cosmopolitan. You’re at a footwear factory. You’ve agreed to do this, and presumably, there is some of that selection effect going on, which makes cohesion just that much easier.

SRIVASTAVA: Yes, exactly. That could be driving a lot of this. Now, that’s also why it’s important to make sure that it’s not just coming from the selection piece, and that’s why the lab and field become necessary. Sticking to the story that you just said, I think, interestingly, I find at baseline in the firm, it’s not the upper-caste workers of the firm have what you would think of as discriminatory or biased beliefs. They don’t think lower castes are less productive. This is something that we say in Arkadev’s paper. Arkadev’s paper is more about the correction of perceptions. That does not seem to be the case here for all the reasons that you mentioned, because caste is a different vector of identity, and most people know people from other castes.

RAJAGOPALAN: It would be hilarious if at a footwear factory upper-caste people thought that lower caste [people] are less productive, when historically that is an occupation that was identifying . . .

SRIVASTAVA: That would be an irony of history.

RAJAGOPALAN: That would be really crazy if that were the prior. 

SRIVASTAVA: The funny thing is with this one, on that note, there is no leather footwear in this factory, which I thought that would be—they think of this as mechanized polymer extrusion. They make fancy (or not fancy) croc, not crocs, knockoffs, basically.

RAJAGOPALAN: Oh, that’s awesome. This is not a leather factory, so that takes away one part of the potential bias or the irony.

SRIVASTAVA: The identity vector, yes, exactly.

RAJAGOPALAN: More generally, this is not exactly a factory setting where I would expect that in the first place. I’m not surprised that you don’t find that.

SRIVASTAVA: Yes, exactly. There is no taste-based bias in that sense. But at the same time, you do find that networks are still extremely saturated by caste.

RAJAGOPALAN: Of course.

SRIVASTAVA: Even in the city, the lower-caste workers—when you ask them to bring someone, you don’t tell them to bring their family members—they still want to bring people from the village. They don’t have people in the city that they know. This also goes back to the Ambedkar versus Gandhi, city versus village. The caste networks have all reproduced in full flow in cities in these settings as well. That’s a little bit of a detour.

I think the question on why the lab and field was necessary is going to come back to this selection versus whether it is something about what the referral does. I think I’m pooling all of these model hazard type explanations. Maybe you have to be nice to the new entrant when this entrant is hired as a referral, because you’re working with his brother, so you want to extend that benefit of the doubt. I think that’s the can of explanations that I try to pool together. I won’t talk about one of the lab and fields, because that’s more of a vignette survey, although I think it’s interesting and important.

Familiar Strangers

SRIVASTAVA: Maybe just to talk about the more detailed one, which is the stylized experiment where I myself played the role of the employer. This was important to do, and separately from the firm because at the firm when I was doing surveys—this is something that you would have heard from other people as well—each survey is production lost for the firm, so they really policed how much I could ask, and there was no room to do any behavioral measurement. That was the main logistical hurdle.

I then played the role of the employer where I hired workers to do a task split over two different sessions. The important thing about the design is that in this exercise, I’m able to keep selection, meaning the identity of the new workers, constant. Each team goes from three workers to four workers, and the only difference between the teams is the mode of the recruitment. I’m able to hold fixed the identity of the new workers who are hired, and I’m able to hold fixed the level of diversity. Each team is going from 33% lower caste in session one to 50% lower caste in session two.

The key comparison is whether the lower-caste worker that’s hired is hired through a referral from an incumbent worker, or the lower-caste worker is hired simply labeled as an outsider. This is what sets up the title of my paper, “Familiar Strangers,” because the idea is they’re both strangers. One is familiar because he’s coming through a referral, and the other one is truly a stranger.

The important thing to figure out about the latter “stranger” stranger, or the true outsider, is that the stranger happens to be a connection of a different team’s lower-caste worker. He’s observationally, in terms of ability, in terms of distribution of capacity, or productive capacity, is identical. The only thing that’s different is the labeling of how this person is being hired. That drives the same results that you see in the firm. I replicate basically exactly what I find in the firm, which is that productivity goes down in the cohesion—the productivity does not go down when you hire workers through referrals. There is no negative effect on productivity, but the cohesion cost that the literature warns you about does get triggered in the outsider’s arm.

When this same observationally similar worker is hired—this time labeled as an outsider without an incumbent link—the cohesion goes down. And this time I’m able to measure this not only in output, so output goes down, but also through revealed measures of cohesion, where I get them to play these behavioral games where I can actually measure a real sense of how much they like working with each other. That cost is quite clear there. There are some other smaller results, but I think that’s the key takeaway: that the network links between incumbent workers blunt this cost. I’m able to answer that question through a lab and field.

RAJAGOPALAN: What is the magnitude of the effect that you find? Now we know the direction of it, but how large are these effects really?

SRIVASTAVA: The retention effects are quite large. At baseline, each team in the firm loses about 10% of its workers, which is quite high. That goes down by 4 percentage points in treatment teams. The retention improves by 40%. That translates to a productivity gain over the entire study period—which is about six months—of just under a 10th of a standard deviation, which I mapped to about 5%. 

RAJAGOPALAN: That’s nontrivial. That’s pretty good.

SRIVASTAVA: That’s highly nontrivial. The margins here on optimization are so, so, so fine that if you can squeeze 5% out of productivity, that’s a big gain. I have slightly different magnitudes in the lab and field, but then, obviously, I’m playing the role of a firm, which makes things quite different. I find there that output changes systematically between the referral and outsider groups, but the key result is on cohesion.

In teams that are hired, that increase diversity through referrals versus teams that hire the lower-caste worker as outsiders, the latter has 9% lower cohesion on average. This is just a standardized index. That maps almost close to what I see on the productivity side. I think there’s a dynamic element to the retention where the retention pays off in nonlinear ways. It’s really helpful to have a stable team over time. If a control team keeps losing workers repeatedly, the gains will not be immediately obvious to the firm until they compare it to another team that’s just stable over time. It turns out that there’s very high returns to just having a stable team over six to eight weeks.

RAJAGOPALAN: Here, I have two follow-up questions just on this specific point. One is, how do you think about cohesion in the longer term in both the control and treatment? Presumably, someone is a stranger only once, right?

SRIVASTAVA: Yes.

RAJAGOPALAN: Once you start working with them, they become less of a stranger, and then over a period of time . . . How much of the first introduction matters that it’s coming through a network, versus how much of that effect just disappears after six months of working with a complete stranger, and then they also become part of your group? Is there an easy way to think about that?

SRIVASTAVA: Completely. Yes. I don’t have a great test for that in the firm because both treatment groups hire through referrals—treatment and control hire through referrals—and I don’t have dynamic data in the lab and field. So, I can’t speak from within my sample on this question. But I really like Arkadev’s paper and Jonas’s paper. There we do see that there are these short-run costs that ameliorate over time, so I wouldn’t expect anything different.

I think the position I’m taking is not at all that, in the control group for my lab and field, the cohesion cost I see is permanent. I think it will attenuate over time. Again, Matt Lowe has work on the contact hypothesis, and all of that, I think, is in the same direction. The challenge with this story, though, is that the fear of the immediate short-run cost is so high that it could make this seem like an unattractive proposition for firms. We might be stuck in this equilibrium because firms perceive this cost to be high.

In fact, in my experiment, when I ask supervisors at baseline why their referral allocations are so skewed towards upper-caste workers, they answer as if they’ve read Arkadev and Jonas’s papers. They say that mixed teams are going to be less cohesive, which is an instinctive thing. Or it could be informed by some bad event that happened at the firm like a few years ago.

RAJAGOPALAN: Yes, or just a two-millennia caste system. We don’t mix very much.

SRIVASTAVA: Completely.

RAJAGOPALAN: We are a very segregated group of people.

SRIVASTAVA: Extremely segregated society. I think that instinct can trigger fear so much that you don’t get over the initial cost, even though this is a highly profitable trajectory to take on over time. In my paper, I have a theoretical model to rationalize as equilibrium. This is like a classic—what we call a bandit problem in theoretical economics, which is just: Do you keep doing what you’re doing, or do you take the risk to try something that might be more profitable in the long run? This is a risky enterprise.

In addition to being segregated, we tend to be conservative. In particular, there are all sorts of principal-agent problems inside the firm. Supervisors don’t want to take on a risky bet, especially because they don’t see returns from the upside anyway. I think this is a difficult problem to solve.

RAJAGOPALAN: On cohesion, I think there might be something hidden in your paper itself, even though I understand that you don’t have a clean test for it, which is through your retention mechanism. Presumably, if there’s a lot of attrition, that is bad for cohesion. Given that you have greater retention means that cohesion is more likely in the longer run, in your setting, than you would find otherwise. I know it’s hard for you to measure it, but it’s almost kind of common sense and intuitive in one sense, right?

SRIVASTAVA: Completely, yes. In fact, from a modeling perspective, if you believe that cohesion has an effect on productivity—which I think is a safe assumption to make—it must be the case from that a more stable team is more cohesive and thus more productive. I think that definitely is a factor.

RAJAGOPALAN: If they’re not cohesive, they must leave. [chuckles]

SRIVASTAVA: Exactly. The other thing I didn’t say about the leaving—the turnover problem—the equilibrium that I mentioned in the beginning, it’s not like these workers leave for better opportunities.

RAJAGOPALAN: It’s a similar job at similar pay somewhere else where they know slightly better people, or it’s marginally closer to home, or the hours are slightly better.

SRIVASTAVA: Exactly. Those are the two exact most common reasons. The commuting cost mechanism, and some idiosyncratic thing happened where you maybe had a fight with someone or maybe your friend is in a different firm. I think it is really about the retention-cohesion nexus, [which] is quite important to investigate.

The Tipping Point

RAJAGOPALAN: Cohesion depends very much on the status quo. Let’s say everyone comes from the exact same caste, linguistic, gender group, or village; you’re likely to have a very high degree of cohesion. As you start diversifying more, it may go up and down in the short run and then stabilize in the long run. Since you’ve done both the firm experiment and also the field experiment, how much does the original composition of the group matter? Because you do vary that in the field experiment. You start at 33% and go to 50%. You are a more diverse group to start with, in some sense, and then you are tweaking one group relative to another group. What is a good way to think about that, both in terms of direction and magnitude?

SRIVASTAVA: I think that’s a great question. Again, going into the firm experiment, there is some natural variation in the share of the team that is lower caste to begin with, that I thought would be somewhat predictive of how this goes in different—like the treatment effects would depend on that on some level. I don’t find that to be the case. Partly, it’s [that] there’s variation, but it’s all relatively low on the lower side. The teams vary between 20% to 40% lower caste on average. Maybe that’s not quite the tipping point.

Informed by that lack of variation or explanatory power, I then, in the lab and field, collapse this to just one change, 33 to 50 [percent]. I do ask workers their instinct on how much diversity is too much or how much diversity is enough? Their instinct is—in unincentivized responses and surveys, they always say that homogenous teams are just more productive. Even lower-caste workers say that about upper-caste homogenous teams, and vice versa.

I think there is some sort of likely tipping point, and I try to model this a little bit in the structural exercise I do in the paper, where there may be some sort of tipping point where the returns start to become negative. I don’t think I find it in the empirics in my paper, but I’m happy to consider that intellectual exercise. where there is . . .

RAJAGOPALAN: Because the intuition is simple. Like if you’re looking at a diverse group that started at one particular point, maybe there is another point where people think, “Oh, now there are too many of the other type of person,” if at all there is a tendency to “other” the different group, the different caste group from your own group.

SRIVASTAVA: Completely.

RAJAGOPALAN: What you’re basically finding is from 33% to 50%—which is a pretty decent jump—is not that tipping point. It could be maybe 80%, we don’t know. That doesn’t mean it doesn’t exist, but it’s not relevant yet.

SRIVASTAVA: You’re right. That’s not the question that the lab and field is catering towards. If I had to answer that question, I would ask it—I would set up the experiments slightly differently.

RAJAGOPALAN: Of course.

SRIVASTAVA: I totally agree with you. I think the question is also that this is not a policy prescription that can be permanently attributed to what these kinds of firms should do. This is not the clear, optimal forever. I think there is a transitory element to this. There is an element where firms don’t need to surgically manage the level of diversity in the teams.

It’s more about positively discriminating towards worker outside options, is how I look at it. Like over time, if lower-caste workers come to the city more and stay in the city more, the networks get better, and their outside options will increase—at which point this retention advantage will diminish. Then you don’t need to do this anymore. Maybe even you need to do the opposite.

Not that I expect any of this to happen, but I think there is that intellectual exercise where we think of what we are in—the current equilibrium—as inefficient. This might be a transitory thing to do to get to a better equilibrium. It’s not like—I don’t think it’s a good policy as a blanket policy to attribute referrals or allocate referrals by demographics or by identity to surgically manipulate the level of diversity in teams.

RAJAGOPALAN: Yes. No, I didn’t think of your paper as peddling a particular kind of policy intervention; it’s more almost proof of concept.

SRIVASTAVA: Exactly.

Wider Labor Market Effects

RAJAGOPALAN: Tweaking the referral system to a particular disadvantaged group actually has pretty good effects. No real undesirable effects at the output level, at the firm level, at the cohesion level. I really think of it more along those margins rather than any kind of direct policy intervention. 

Since you track people’s job market prospects, both within the firm and outside the firm, now you’ve created an intervention where the lower-caste referral is more likely to be hired relative to the upper-caste referral. Now, were you worried at all about their job market prospects outside of the firm? Is this something you’re able to track and see because it seems important?

SRIVASTAVA: Yes. Completely. I was definitely worried in the sense that I was intellectually expecting there to be—I mean, there is displacement. That much is clear, that there’s a finite number of job offers, and certain people don’t get as many job offers as they would have gotten in the counterfactual world. I was careful about being able to track those outcomes. I was open to the possibility that there will be some costs. I don’t find any significant adverse effects on displaced upper-caste workers. Upper caste here, again, I’m collapsing generals and OBCs [Other Backward Classes]. You’re correct to caution against that. All of them are not privileged or well-to-do. They’re still working a minimum wage job. By definition, they’re not well-to-do.

RAJAGOPALAN: In the social hierarchy, they’re higher, which is the relevant question that we’re looking at here, right?

SRIVASTAVA: Exactly.

RAJAGOPALAN: No one’s looking at elites, really.

SRIVASTAVA: Completely. The social hierarchy is really what shows up in this. Social capital is what acts as the buffer, like a social insurance. The effects—these people tend to be infra marginal, anyway. So they’re more likely to already be employed. Additional job offers have this effect of improving their bargaining power presumably at their current workplaces. Again, on average. That is the dimension on which I see a decrease. A decrease in this job, getting this job offer likely affects your ability to bargain. But that doesn’t translate to negative effects on earnings or probability to be in salaried work.

On measurable things, you’re not worse off on average. Then again, once you do the math to account for the gains to the lower-caste people, I think all of this is a blowout. The lower-caste people are better off on all these things by large magnitudes.

RAJAGOPALAN: What would be your intuition, and also what do you find, if a low-caste person doesn’t get referred for the job? Are they also likely to have other employment options but just lower bargaining power, or are they less likely to actually find jobs outside?

SRIVASTAVA: On average they’re more likely to be in rural areas, to be unemployed. That’s really where the welfare gains are coming from. The probability of being in salaried work, which is I think a low bar, but a bar that we are failing to clear so much in . . .

RAJAGOPALAN: No, but that’s the relevant bar for them.

SRIVASTAVA: Exactly. We don’t clear that bar often enough in our labor market. I think that share increases dramatically. They’re much more likely to be back in the villages, which is why I think there is a broader macro structural transformation element to this as well. This is more speculative, so I don’t get into this in the paper at all. The idea is that upper-caste workers were able to migrate in earlier generations much more often and much more easily than lower-caste workers. The poorest lower-caste people did—the completely landless—because again, there’s nothing to fall back on, so the opportunity cost is too low. What we have now is this oversupply of workers in agriculture who want to get different jobs, move to the cities, but can’t move to the city without a job; otherwise they will be subjected—

RAJAGOPALAN: Yes, it’s very precarious. Even with a job, it’s very precarious.

SRIVASTAVA: —to the harshness of urban life. I have some other work that studies the harshness of urban life in different ways, but I think the idea is just that getting a job offer is really what makes it stick, and then that changes the market trajectory. I hope to follow these people for a longer time even and see over time how they develop their roots in urban areas.

RAJAGOPALAN: Were you at all concerned at any point about the reputational cost? Because the person who’s making the referral presumably bears some of the reputational cost if the person that they’re referring is not that great. I know that’s very incentive-aligned from the point of view of your experiment. But were you also worried that “Hey, if they have a poor-quality network, whomever they’re referring is going to get hired, and then people within the firm may look at them more poorly because of it.” I ask these questions only because now we’re talking about real people and their real everyday outcomes and their wages and so on, and it does seem to matter.

SRIVASTAVA: Completely. Yes, as you said, it’s incentive-aligned, but there might be a more ethical or logistical lens through which you might worry about this. In practical terms, what I did was I never forced any team to elicit a referral from one person in particular. There was always an ordered rank, ordered list of people they’ve had to go through, which in practice meant that if you don’t want to refer someone in a given month, you don’t have to. You’re not going to be forced to give me a name. We went to the next person, the next person after that.

In reality, most people are desperate to refer people. I think there’s a lot of status and positional utility from doing this. In a somewhat unrelated exercise at the firm, where I did a BDM-style elicitation exercise, where I said, “Would you choose to refer someone or have a slightly longer lunch break?” Seventy percent would choose a referral over a 30-minute longer lunch break, which is a 50% increase in their break time. The ability to choose your coworker or improve the life of your friend or a relative, I think, is huge.

RAJAGOPALAN: But also your status within your community, right?

SRIVASTAVA: Exactly.

RAJAGOPALAN: Now you’re the guy who can help someone get a job.

SRIVASTAVA: You’re the guy in Delhi who’s pulling people from the village, exactly.

RAJAGOPALAN: Yes, and they owe you.

SRIVASTAVA: Yes, exactly. Then again, in my data, I find that there are some transactions that happen between these people. They help smooth liquidity for each other, and there are other inter-referral/referee transfers that happen. I wanted to make sure there is a way out for each person in case they don’t want to refer, they don’t feel like they have a good person to refer, but that did not get triggered very often. There are cases where people gave up that opportunity, but it didn’t happen very often.

RAJAGOPALAN: That, again, intuitively makes sense given the social setting of this experiment. The way I think about your experiment in the broader context—again, I’m not trying to generalize the results to anything else. I’m not trying to make policy based on this. It’s just pretty much in every business that you look at—elite, minimum wage, factory setting, academia—referrals play a very, very important role, especially in getting your foot through the door.

Now, in India, social capital is very disproportionately distributed, and upper-caste groups tend to have longer-standing networks in urban areas, longer-standing networks in pretty much every firm or every category of occupation. Therefore, referrals tend to come through them more easily. What you basically show is it doesn’t matter which group the referral comes from. It’s the referral which is powerful when it comes to hiring more workers, as opposed to the identity of the person that’s being referred. Is that a good way to think about the broader point of the paper?

SRIVASTAVA: Yes, I think that’s exactly right. I think we should think of this network-conscious or network-based hiring as a tool to counter the problems that come with segmentation in our labor markets. We’re turning what is essentially a problem that drives inequality to a tool that could maybe, on the margin, fix inequality a little bit. I think that’s exactly right.

RAJAGOPALAN: Hopefully, with this kind of experimental evidence, there are more firms that are willing to just take referrals from all groups of people as opposed to just one. They’re willing to get referrals from women, for instance, or people who may not have the same educational qualification; it doesn’t just have to be caste. Clearly, this has some broader role of that familiarity place.

SRIVASTAVA: Exactly. Just to return to my firm—after I leave, the experiment ends. I give back the supervisors the ability to allocate referrals the way they choose to. The treatment supervisors continue to do this. That’s optimistic.

RAJAGOPALAN: Oh, yes. I found that fascinating that they continue to make more lower-caste referrals even after your experiment ended.

SRIVASTAVA: Exactly. They show evidence of having learned that cohesion costs don’t get triggered if we do this. Now, they don’t continue to give all their referrals to lower-caste workers. That’s also, I think, good and healthy. Their allocation is still elevated relative to the baseline. I think that’s optimistic.

RAJAGOPALAN: No, this was fascinating. This is a really cool experiment. The fact that there’s an experiment going on within the firm and outside the firm makes this picture super rich.

The Long Shadow of Feudalism

Before I let you go, maybe we have time to talk about another one of your papers. This is called “The Long Shadow of Feudalism.” It’s about concentration of land and then how it matters for labor market outcomes and labor power in India many, many decades and centuries later. Incidentally, it’s with Steven Brownstone, who was on the series last year. We’ve already had your coauthor come here.

Once again, you’re looking at labor market outcomes, but now the intervention is not the one you did in the experiment. This is a weird intervention done centuries ago, and now you’re trying to figure out what happens in the future. Can you walk us through the setup for the paper, and then maybe we can discuss it briefly?

SRIVASTAVA: Yes, absolutely. This can be like a two-episode series on what’s been going on in this paper over the period of one year between Steven’s and my appearance. This paper, as you said, is trying to use variation from the princely state of Hyderabad, where we had certain land grants that were given to noble families by the Nizam in exchange for their support during the 1700s and 1800s.

What that does for us 200 years later is create what we think is a quasi-random variation in the extent of land concentration in the present day. There are these parcels of land where land ownership, at the left tail in particular, tends to be much, much more concentrated relative to other land parcels that are within 20 kilometers of each other and within most conventional bandwidths in this kind of empirical exercise.

What that means in practice is that the smallest land parcel in a feudal area is going to be about 19% to 20% larger than the smallest land parcel in a nonfeudal area. These two villages are separated by, in most cases, 5 to 10 kilometers—highly arbitrageable distance that you could travel from one place to another to work if you wanted to.

What we find in present-day labor markets that are separated by the feudal/nonfeudal barrier is that women, in particular in the feudal areas, have a wage markdown of about 8%. For the same kind of task—growing the same crop as their fellow women, female agriculture laborers in the nonfeudal village, growing the same crop in the same season, using the same inputs, exposed to the same climate, the same soil, and the same yields—they earn about 8% less on average, which is a substantial difference. We don’t find any effect on the men. In the paper, we argue that this is suggestive of women just being closely tied to the frictions in the labor market.

In the monopsony power that the landowners have in the feudal areas has more bite for female agricultural laborers because they lack the ability to escape these labor markets and travel, migrate, or maybe go to a different village and work, or maybe open a shop and work there. Since Steven’s been on last year, we’ve done a little bit more work to establish this link between elite capture and the labor market effect.

In the paper, we show that a big lever through which this works is MNREGA implementation, which is much, much worse during exactly the time when landowners compete with MNREGA for labor. In terms of labor demand, it’s about 70% worse in feudal areas than in nonfeudalfeudalareas. The disadvantage disappears when the landowners are not competing with MNREGA during the agricultural lean season.

We establish this nexus between elite capture and local politics through three new things that we’ve done. The first thing that we did was, through a random subsample of our villages that we have in this sample, we find that 71% of the elected leaders belong to the same subcaste or jāti as the literal largest landowner in the village. Ninety-one percent belong to one of the 20 largest landowners. Now, this is probably not a surprise to those familiar with rural India. But it still feels staggering to see these numbers. This is such a deep nexus between the literal single largest landowner.

The second thing we find is that in previously feudal areas, elected leaders tell us that their preferred NREGA wages are lower on average than in the nonfeudal areas. This indicates this instinctive adversarial relationship that elected leaders take on relative to the public option. Then lastly, we find that all of our results are intensified and stronger in villages that are farther away in space from district and block headquarters.

Leaning on the recent literature on the proximity to centers of power—and a lot of your thinking on federalism as well comes into this—but how close you are to centers of power, state capacity has this receding influence as far as you go. We argue that this distance allows elected leaders to crowd out bureaucratic influence more easily. If you’re in a village that’s farther away from the block headquarter or district headquarter, the wage markdown actually increases. 

RAJAGOPALAN: The long and short of it is, places that were feudal 200 years ago are still acting like they’re totally feudal today, right? In the politics and the way the labor market operates in terms of agriculture and what they grow, and in terms of how they hire people.

SRIVASTAVA: Yes. The government, in terms of the input, the mechanization, the ability to grow the same crops in the same season—all of that is equalized, and the state can only do so much. But the nexus between the landowners and the elected officials or the bureaucrats still allows this kind of labor market friction to persist, and [it] shows up in everyone’s wages.

RAJAGOPALAN: The only way out of feudalism, it seems to me structural transformation is the only way out. Except what you find is, structural transformation is harder in places that were erstwhile feudal than in places that were not feudal, right?

SRIVASTAVA: Exactly.

RAJAGOPALAN: This gives us a bizarre problem now.

SRIVASTAVA: Yes. That’s where Steven’s job market paper, I think, is relevant as well. Coming back to my job market paper, I think a lot of this is also explained by outside options. MNREGA being worse is a big driver in this. If you just get MNREGA to be better as a public option, and in the same way that lower-caste workers have bad outside options—if you strengthen the outside options, the private sector can target people with bad outside options and improve their welfare, and the state can strengthen outside options to improve their welfare. I think we have this equilibrium where we can do more on both sides of the market.

RAJAGOPALAN: Yes. The complicated part of the story is that the feudal places are more likely to capture the state and the institutions, which make the outside options disappear. You may need some external intervention. Here, maybe a centrally sponsored scheme may be better than a locally sponsored scheme or something like that. Something has to cut through this.

SRIVASTAVA: Absolutely, exactly. Karthik Muralidharan and others, and Paul Niehaus, Sandip Sukhtankar, they show that the implementation of biometric—they use that introduction as a technological shock that improves the NREGA implementation, and it has more bite in places that have higher land concentration. That intervention has the promise of cutting through all of these structures a little bit. We should think along those lines, maybe.

RAJAGOPALAN: Yes, and also just stop giving agricultural subsidies. In some way, we got to weaken this nonsense and stop. But the core of it is that resources must go to their highest-valued use—separate from the welfare effects that you’re talking about, or the caste effects and the network effects. So much of what’s going on here is bad also because we don’t let resources go to their highest-valued use, right?

SRIVASTAVA: Right, exactly.

RAJAGOPALAN: Feudalism is bad not just because it’s feudal, but because of this other huge economic elephant in the room that we are unwilling to address when it comes to agricultural subsidies.

SRIVASTAVA: Yes, you’ll find that that’s how a lot of these conversations end, with something about subsidy [chuckles]. You know this more than I do, but yes, I think that’s often the final word.

RAJAGOPALAN: This was a pleasure to read. Thank you so much. I think this might be the first time we’ve had a paper discussed in two parts by two different coauthors. That’s also very fun. Our job market series is progressing.

SRIVASTAVA: Hopefully, given publishing timelines, we don’t discuss this again, but you should never bet against it.

[laughter]

RAJAGOPALAN: No, I’m sure this is all going to land pretty soon somewhere. Thank you so much for doing this. This was such a pleasure.

SRIVASTAVA: Thank you. Really appreciate it. Thank you so much.