In 2004, David Atkin—then a graduate student at Princeton—enrolled in Gene Grossman’s course on international trade. Five years later, Atkin completed his Ph.D. at Princeton and, in the decade that followed, went on to become a leader in international economics himself, working at the intersection of trade and development economics.
Today, Atkin is a Professor of Economics at MIT and the first former Princeton student to be interviewed for our new “Alumni Conversations” series. In this episode, Atkin talks to Grossman, the Jacob Viner Professor of International Economics at Princeton, about:
We’re grateful Professor Atkin was able to take time during a trip to California to dial in to our Zoom call. You can watch the entire interview on YouTube or read the transcript below.
Gene Grossman: Hi, David. How’s life in California treating you?
David Atkin: Very good, thanks.
Gene Grossman: So, David, you came to Princeton after completing your master’s degree at LSE where, if memory serves me correctly, you were a classmate of Dave Donaldson, your current colleague.
And then you came to Princeton and took my trade course in the fall of 2004, if my arithmetic is correct, and went on to write your dissertation under Angus Deaton and me.
And then you made stops at Yale and UCLA before joining the MIT faculty in 2015, were promoted to tenure in 2017, and these days you’re one of the leaders in the intersection between trade and development economics, and you’re just about the most creative empirical economist I know.
So let me start by having you reflect on your Princeton experience. In particular, you were of only two students ever to work with Angus and me as advisors, and I’m wondering whether that was a bit disorienting, or how you managed to manage our different approaches and styles.
David Atkin: Thank you very much for the kind introduction. So that’s a great question. As you’re aware, having a great supervisor’s a blessing, so having two is even better.
I guess I didn’t find it too disorientating. If you’re interested in doing empirical work and trade, it’s naturally quite a theoretical field, and so you need to definitely think through all your empirics in terms of a theoretical base. So having an empirically-minded supervisor in Angus and a more theoretically-minded one in you, I think, was an excellent combination, and I struggle to think how a trade empiricist would get by with only one of those two bits of guidance. And of course, you are a notable empiricist yourself, though maybe slightly longer ago.
Gene Grossman: So, are you claiming that we never gave you any conflicting advice?
David Atkin: I think, on the specific papers, not so much conflicting advice, maybe on what topics are more interesting or less interesting. You know, some differences in opinion, maybe. But on specific projects…on one end you are trying to think about how to make a point more convincingly and then, on the other end, you’re asking why you’re asking that point, and whether or not that’s the right point to be asking, and both of those are complements, I think. Maybe I’m misremembering. I don’t remember any discord.
Gene Grossman: Yes, graduate school always looks better when you look back on it, then when you’re in the midst of it.
Let’s turn to the research you’ve been doing since you left Princeton, and why don’t we begin with a high- level overview. Do you have a research agenda? Do you think of it in those terms, that you could describe? Is there a unifying theme, or do you just go for whatever problem seems interesting at the moment? What’s your research approach?
David Atkin: Certainly at Princeton and coming out of Princeton, I definitely saw a quite focused agenda in thinking about the impacts of trade on the lives of the poor and in developing countries. That’s continued to be a focus of the work I do. You know these were the mid-2000s when I started the Ph.D., and I’d look around me and most of the discussion you’d see about trade would involve poor countries, impacts on poverty, etc.. And it seemed like there was a bit of a disconnect between how much focus was on developing country issues when it came to trade versus the media’s attention.
That’s obviously changed a bit. Thanks to Donald Trump, there’s now been a renewed interest in a variety of a developed country trade issues. But, at the time, I thought that was a particularly important set of questions and I still do. I guess recently I’ve also become quite interested in maybe an adjacent subject, which is industrial policy and firms in developing countries. Maybe because of our tool sets and the type of data we use, that’s typically been an area where there’s been a lot of interest from people who are also interested in trade and that’s also an area which is of considerable interest to me now.
And again, a lot of important policy debates are going on, and we as economists should be providing more inputs into a lot of those debates, and the hope is to have an agenda that is a kind of complement with current policy questions.
Gene Grossman: One of the unique aspects of your work, especially your work in some of the developing countries, is that you’ve resorted to field experiments. This is quite popular in development economics but quite novel in trade economics, and coming from a student of Angus Deaton it’s particularly interesting since, as you know, he’s one of the most vocal critics of this kind of methodology.
Maybe I could get you to talk a little bit about what you see as the role of experiments in trade research and whether you think this is going to be a once in a while thing, or it’s going to become a more standard part of our toolkit. How do you see experiments?
David Atkin: Excellent question. I guess my views and Angus’ aren’t so far apart in terms of the fundamentals. Maybe some of the details we differ on. But the way I see it is: What is an experiment doing? It’s providing us with exogenous variation, and it’s particularly valuable in settings where we struggled to get exogenous variation to identify causal impacts. Now, of course, that doesn’t mean it’s a substitute for a well-posed question, or it’s a substitute for careful data work. I’d say that it’s a very, very strong complement with something which is very important in a lot of our trade questions, which is good data collection. Often, we don’t have in our standard data sets exactly what we need, specifically when we’re thinking about things such as global value chains or even what happens within firms when they start exporting. And if you’re going to have to collect that data yourself, then that’s often done alongside randomized control trials. So there’s a compliment, in that sense.
Where do I see this being most valuable? When you have a well-posed causal question, which is important to the field, then obviously being able to generate this exogenous variation is very important. I guess I’ll give one example from my own work, which is we were very interested in understanding learning by exporting. It’s kind of a classic question in the trade literature, but as has been highlighted in the last 20 or 30 years, it’s not random which firms start exporting, and in fact we have a large body of theory suggesting that firms are selecting into exporting based on their characteristics or their productivity. And so being able to disentangle whether a firm is more productive because they’re exporting or because they were more productive and that’s why they decided to export has always been a challenge.
So that’s kind of a natural causal question where, if you can generate variation in participation in exporting, you can answer that question, at least for some sample where you are able to do that. So we picked a number of small firms in the textiles industry in Egypt and were able, with the help of an NGO to generate exogenous variation in access to foreign export markets and, through very detailed measurement, to actually see what happened within those firms as they upgraded quality, increased their productivity, increased their profits, etc.
So there’s an example where, am I able to conclude that in every industry in the world exporting to every country in the world has big learning by exporting impacts? No, and you want to do these types of exercises in other contexts in order to be more certain of those things. But it’s the type of question where we struggled with observational data to make a lot of progress.
So is there a space for this type of approach in the trade literature? I think absolutely. A lot of the critiques that come from Angus Deaton and others are that essentially it biases us towards answering certain questions that may not be the most important questions. Well given there’s maybe three, four, or five RCTs in the whole literature, I think we’re far away from reaching the point where everyone’s answering questions that aren’t so interesting because they’re forced to use this methodology. But as part of our toolkit, alongside natural experiments, alongside some more quantitative approaches, alongside just descriptive statistics and theory-based narrowing of hypotheses, I think this is an important part of our toolkit for certain types of questions, and I expect there to be more such experiments in the future.
Gene Grossman: If I can just push a little bit and channel my Angus Deaton, how do we get from a few small firms in the textile industry in Egypt to knowing whether learning by exporting is something we should really think is first order in our thinking about trade policy and export behavior?
David Atkin: Well, a couple of ways you can do that. One is you could propose experiments in other contexts or settings and then see if they line up. And obviously that’s going to be a lengthy process, but you can build up an evidence base that way. The other is you can use existing techniques and see if they give us somewhat similar answers in somewhat similar settings. And then maybe the biases aren’t so large in other settings, and then, you know, utilize those non-experimental technique answers in those other settings to shed light, having validated them where you have experimental variation.
Famously, maybe the birth of us being very fond of using experimental analysis in economics was a famous a long study out of Princeton, which showed that the existing techniques did very, very badly compared to the kind of answers you got from a randomized experiment when thinking about job programs in the U.S. So these are important questions for us to try and get to the bottom of. If learning by exporting is an important feature of our policy, we need to know how big it is to make the correct choices, for example.
Gene Grossman: Let’s stick with empirical methodology for little longer, but let’s go to the other end of the spectrum. As you know, our field has become enamored with quantitative modeling and intricate models that have been calibrated to match “moments in the data” and then used to conduct “counterfactual analysis.” I don’t see you doing a lot of that. You don’t see me doing a lot of that. Do you have reservations? Is it just you haven’t gotten there yet? How do you view that whole trend in our field?
David Atkin: An excellent question and one I’ve been thinking about a bunch recently. I have two recent papers where I’ve dabbled somewhat.
Gene Grossman: Oh no, don’t tell me!
David Atkin: I have one paper with Treb Allen and one with Arnaud Costinot. So what’s my view on this type of methodology? You know, again, my background is definitely coming more from reduced-form applied micro approaches, thinking very much about exogenous variation and natural experiments and other techniques and having particular moments in the data which are, you know, very revealing about a particular conjecture or hypothesis or comparative static. And we lose some of that certainly with some of these quantitative approaches where you’re matching to a whole bunch of moments that may not be particularly closely related to the particular mechanism you’re studying.
And so, certainly I have some reservations there, but you know there’s obviously a lot of very good quantitative work where you are matching to moments that are very close to some of these kind of moments the reduced-form empiricist would be looking at.
I guess my bigger reservation is I see these as very powerful tools for understanding theory. You know you have a model in mind, and you start playing with it and you get very unexpected things from your quantitative model. And then you change your quantitative model, and you find out what works and what doesn’t.
I guess my sadness is that most of that information is only held by the authors of the paper, because they’re encouraged to write up the paper as if this was the first model they wrote down when, in fact, this is a model that accorded to our priors, which is why they’ve left it in the paper and why they stopped there. So we’re often not learning things beyond our priors, because you’re not going to be able to publish a paper that gives you quantitative predictions that are way outside what we think is realistic. And the authors aren’t writing down clearly what they’ve learned and instead kind of claiming it’s all about just answering this particular question about a 3.7% rise in GDP when I did X, Y and Z. I guess I would really push for us to share that knowledge more widely by explaining what models did work, what models didn’t work, and helping us learn a little bit about some of these kinds of theoretical relationships rather than the current norms.
But I think this is a new methodology in trade. The applied micro methodology is now much more mature, and I’ve seen that change dramatically over my career, what people do and what they don’t do. And I suspect the same will be true for quantitative work. The type of quantitative work will see in 15-20 years will be much more sophisticated and provide much, much quicker knowledge accumulation than the current forays.
Gene Grossman: I could speak endlessly about my reservations, but let me try one and see whether it gets you to agree or defend, which is that models are abstractions. The better the model, the more stylized it is because it allows us to focus on one particular mechanism. When you force a simple model to fit a complicated world and you don’t get standard errors, you don’t know what damage you’re doing. And then you immediately run to counterfactuals, and what am I to make of those numbers that it generates? How am I to know that the model is not missing the key forces because you intentionally simplified it?
David Atkin: I think that’s all valid. I get the sense, which may be overly optimistic, that in the macro literature, where they’ve been doing this longer and people, at least the better economists, have a good sense of what forces you’ll hit with this type of model and which ones you won’t, and for certain types of questions there’s some sort of agreement that it’s okay to be missing on this but not okay to be missing on that, and we don’t really have that in trade yet—that kind of discipline—which maybe prevents some of what you’re describing there. Although, like you, I have a fondness for comparative statics which, I think often, as much as we might learn from some of these exercises at present, given our methodologies, focusing on those types of predictions seems particularly valuable. Those are also easy to verify or validate in the data.
Gene Grossman: I’ll give you one more complaint before we move on, which is: So the models often fit the data by introducing things we can’t measure like amenities and productivity. In my generation, we used to call that the error term, but now it’s part of the model. Then when you do counterfactuals you don’t know what the hell they are, but you have to hold them constant. So why would we think these things would remain constant once we start changing the environment?
David Atkin: (laughing) I guess that’s the model they wrote down where those things were made constant. I do think there’s exciting work now in starting to think seriously about amenities and trying to relate them to observables. We have a very promising student on the job market who is thinking about racial inequality in cities and thinking about these amenity terms and how they relate to Black-white interactions and the like, and I think there’s exciting work there.
I share many of these reservations but at the same time we’re now being bombarded with more data than we can possibly deal with, and these types of frameworks allow us to use a rich, very disaggregated and detailed data—which I think will eventually be a huge plus. But again, we need to know how to use these tools to maximize what we’re learning from these data.
Gene Grossman: Okay, I won’t to try to force you to offend your friends. Next question I like to ask people about is which of their papers they’re most proud of. I know, in my own case, it’s not necessarily the ones that have received the most external notice or the most citations. Is there a paper that stands out in your mind? Maybe it is one that has a lot of citations? Is there one that stands ahead of the others that you’re proud of?
David Atkin: I don’t know it stands ahead, but certainly one which I think has been read less or had less influence than I think it deserves is this paper on the impacts of retail globalization on households in Mexico. And this goes back to something you said right at the beginning of this interview…you’ve only had two students who were jointly supervised by you and Angus, and I guess the other was Guido Porto. Starting with some work of Deaton then followed up by Guido Porto, he wrote what I think was a fabulous job market paper that was maybe held back by the data at the time and the methodologies of causal inference at the time. But basically he shows that you can get, and this is based on some early work by Angus, quite a long way in characterizing the welfare impacts of trade reforms or general reforms that change prices through a whole bunch of first order terms, many of which you can read off the data, and others you need more causal evidence for.
And the world since then has advanced dramatically in terms of data access and also in terms of the availability of natural experiments, and the ability now for us to calculate extremely rich distributional effects to the first order using data that’s now available is an incredibly powerful tool, and I think trade economists should be using it much more. And then we can go into some more complicated second order effects and the Gini effects as well, and that’s icing to the cake. But we can learn an enormous amount, with a reasonably limited number of assumptions from just looking at things such as, you know, what share of your consumption is exported or imported from abroad or subject to particular tariff lines and the like.
And now I’m working with Dave Donaldson and co-authors on a project in Chile where we’re actually able to see the individual products the consumers purchase from which firms thanks to the incredibly elaborate VAT collection data. You can start to see how you can really make enormous progress working out the impacts of trade reforms or other price-changing reforms. And that paper tries to do that for the impact of the arrival of foreign supermarkets in Mexico, and I think we make a lot of progress in really widening that toolkit in terms of allowing for quite a wide range of impacts to be credibly identified and then put into this framework of these first order impacts of price changes.
Gene Grossman: I’m very much a fan of that paper, as well, but there was one piece of a sentence that just prompted a different question: Can we ever have too much data? Can it lead us to focus on trivia and lose the big picture? Do we sometimes want to aggregate or not focus on, you know, whether the shipments were in red boxes or blue boxes just because the data set tells us that?
David Atkin: I think one of the challenges I found hardest in my career, at least, when you do have an extremely rich data set, is working out how to not get tangled up in those minutiae, and step back to what is actually valuable information. And you’re exactly right that it’s not just that we are spending too much time on the red and the blue, it’s that you might not make any progress and never write that paper because you haven’t worked out a way to aggregate the data in some way that you’re able to pull out meaningful economic relationships.
I think that’s a big problem, and is also related to this quantification discussion we had a little before, which is when you are confronted with these huge data sets, and now we have these machines that are obviously well-suited to utilize these huge data sets, it is low hanging fruit. I have a huge data set that people will like, plus a machine that can use that data set. And you know I think one thing that’s missing then is the most important one: What’s the question? Is there an interesting economic mechanism for us to uncover? Those three together can be very powerful, but starting with just the first two you may never develop that interesting question and the economics behind it.
Gene Grossman: In my day, that version of that was being told that when people had to calculate regressions by hand, they were very careful about which regression they calculate, and didn’t try hundreds of them just because they could get them back in a fraction of a second.
David Atkin: (laughing) Yeah, that’s definitely changed.
Gene Grossman: Let’s wrap up by talking just a little bit about teaching and advising. It’s been a while since you’ve been a student. You have classes and advisees of your own, of course. Do you enjoy these roles? Would you say there’s something distinctive about your approach? Did we teach you well how to do it at Princeton? And then, which of your students are you most proud of?
David Atkin: An excellent question. I guess one thing we do continue to do at MIT, which very much comes from you—and you know Arnaud Costinot is there, as well, who was also a Princeton student—is we find it very important to intersperse theory with empirics in our trade course. At many schools there would be a more theoretical side and then maybe they’d teach the empirics separately, or as some subset. Here we jump back between it week-by-week or class-by-class in order to teach the appropriate, relevant empirical work straight after the theory, in order to keep those connections.
That’s something we certainly learned from you, and it’s been very valuable to the students. In terms of a distinct style…I don’t know how distinct it is, but I do feel very strongly that you want to get students excited by the research questions and by the topics. I still remember complaining to you during class when there was something that didn’t seem of first-order importance. And then we got to, I think it was a Venables-Krugman paper or a Krugman-Venables paper, and you said “Is this important enough for you, David?…This explanation of the whole world through history?” or something. So I do think it’s important to get the students excited, and one area to do that is highlight gaps between where policy discussions are, where big concerns in the world are, and where our research is.
For example, recently I’ve added to the course a little bit on multinationals and taxation and tax evasion. It’s not an area where we really have much research, if at all, maybe one paper. But spending some time on that highlights that this is an area we should have much more. Maybe you have to put blog posts in the slides and such, but hopefully you push students into these understudied areas. I think that has high returns.
This comes back to the discussion we had a few minutes ago. As you know a really great job market paper often needs a hook, a question, something interesting that catches people’s attention. And if you start out with that, then the path is a lot easier than having most of the paper done, but no hook, and then having to try and work out how to make the thing that you’ve already done three quarters of interesting.
Gene Grossman: Let me ask you, before we turn to advising, one question on teaching. My approach, if you remember, is very much an integrative approach. I try to teach the lessons of the theory rather than particular papers. Esteban [Rossi-Hansberg], for example, my colleague of many years, very much likes to go carefully through leading papers. I think both of those methods can work. Does your style lean one way or the other?
David Atkin: I mainly teach the empirical components, and for that we do a bit of both. Certainly, the syllabus as-is was more your summary style. As we’ve adjusted the course, I’ve added a few more deep dives. I think with that empirical work, you do learn a lot about the decisions the authors made, why they decided to do what they did. The discussions in class about is this credible? How could it be better? That’s part of, I think, how you learn that craft.
My thought is that with more theoretical papers, a lot of that craft is also learned by doing it. And particularly with more quantitative exercises. You need to be talked through one to give you confidence, and then you get a nightmarish problem set that takes you weeks of your life. But then you have learned some serious skills. It’s a little hard to get people to actually do empirical work, and so much of that is thinking through endogeneity concerns and coming up with you know, a valid methodology, which is hard to do for a problem set. Although, certainly, implementation is something we can do. So I think a bit of both. You want to know where the literature is, what we’ve learned. But at the same time you want to have the chance to go through some of the thought processes in writing an empirical paper, which you know takes maybe half a class going through one paper.
Gene Grossman: Do you give your students data sets and have them to work on those data? Or is it more left for a third-year paper and the like?
David Atkin: We do a bit of that, maybe not so much for the trade course, just because the thought is that often the empirical students are also taking courses where they’re doing more of that, and we naturally spend a lot of time, for good or for bad, on some of these quantitative exercises which do take a lot of practice, as well as you know, some of the canonical theoretical models which knowing them back-to-front has high returns.
Maybe we have one assignment that involves much more data work of the sort you’ve described.
Gene Grossman: Let’s finish up with advising. A quote I like was from the physicist [Richard] Feynman, who said that advising Ph.D. students is a lot like doing research, but with your hands tied behind your back. Are you a very active advisor? Do enjoy advising?
David Atkin: I do enjoy advising. It’s like doing research, but without having to do all the boring bits in between those meetings, where you actually have to implement the things you discussed. And obviously it’s a real pleasure to see your students succeed and thrive and go off and do great work at their respective institutions. It’s certainly one of the perks of the job. Really good students are a perk of the job. And then otherwise I think it’s an important part of you know, working in universities, to make sure that we provide a good learning environment to all students, as well. That’s someone I value even if office hours with undergraduates just before the exam is not the highlight of my week.
Gene Grossman: One of your excellent students, recent students, just arrived in Princeton yesterday. Mayara Felix is going to be our postdoc this year, and I know you think very highly of her and we’re really looking forward to spending the year with her.
David Atkin: Yes, we were thrilled she decided to go back to Princeton to learn from the wonderful faculty there, just as we benefited from them earlier in our careers. As you say she’s wonderful, brings a huge amount of energy and ideas and she’s also a wonderful person, so you will enjoy the year with her.
Gene Grossman: Great, well let’s wrap this up. Let me thank you so much for doing this, and it was great to see you in person recently in Cambridge. As you know, I have a grandson in Cambridge now, so I guess we’ll be seeing more of each other than we did during the pandemic.
David Atkin: I very much look forward to it, and thank you so much for chatting. That was that a lot of fun.