Andrey: Welcome to Economic Frontiers. Today our guest is Alex Teytelboym, who is a research fellow at the Institute for New and Economic Thinking at Oxford University, and will be starting there as an associate professor next fall. So the topic for our conversation today is going to be market design, and specifically the design of matching systems. And in a bit of a departure of what we usually talk about, the topic will be not primarily digital nature, it will be refugees, and how one might use economic thinking to help alleviate the refugee problem. So, welcome to the show Alex.
Alex: Thanks Andrey.
Seth: Alex, thanks so much for joining us.
Andrey: Yeah, and of course Seth Benzell is my new co-host, and he will be joining the podcast from now on.
Seth: The pleasure's all mine.
Andrey: With that in mind, I think maybe a good place to start before we delve into the refugee problem, is to explain what economics has to say about matching and maybe to give us an example from a more familiar setting, which would be the school choice setting.
Alex: Thanks. Well, yeah you know matching's been around as a discipline for about 50 years. It started with a beautiful work of David Gale and Lloyd Shapley. And since then we've been looking for ways of extend the theory and looking for applications in which we can apply this theory. And one of the ways in which we routinely apply it all over the world, especially here in Boston, but all over the UK, is trying to match kids to public schools. So how would this work? Well, kids and parents they have preferences over the sorts of schools that will be good for their kids, and the schools they have priorities of the children. So, the kids who have a sibling and live in the neighborhood, they would get the highest priority, and the kids who live in the neighborhood but don't have a sibling would have the second highest priority, and so on. And so we look for ways to match the kids to the schools. Now, what is a good match? So, a good match we might think is one where no kid's priority is being violated. So if I really want to go to another school, it must be the case that I couldn't get into the school cuz my priority's too low. And there is no space to take me because if you bump somebody out, they would've had a higher priority, and that seems unfair, right? So we don't want to have a matching that would be unfair in that way. We also don't want to match that would be inefficient. So if two kids can swap a school and both prefer the schools that they swap into, that also sounds a little silly. We want to prevent something like that. The other thing we want to do we want to avoid giving parents incentive to strategize and manipulate the preferences over the schools that they submit to the system. So, you know fifty years ago the system that was used in Boston allowed parents to do that. They'd allowed them to strategize and misreport, which schools they actually prefer. Because then the outcome you get you don't really know what preferences it's based on. It's based on some preferences that maybe were not real. So you might say okay well can we design a system like this that would be fair and efficient and wouldn't give any incentives to manipulate, and the answer is no. As ever in economics there are trade-offs, and so what would have to pick is decide whether fairness is more important than efficiency and so on.
Andrey: So in this specific example of school choice let's think about the differed acceptance algorithm, which is a commonly used algorithm for this. Which of the three criterias violated and why?
Alex: So the only criteria that will be violated in general is efficiency. So if the Gale-Shapley algorithm will give you an outcome which will be fair, but it and it will give you no incentives to manipulate it if your a parent it will give you the best possible outcome out of all the possible fair outcomes for the students, which is good. But there could be opportunities for two students to swap schools and make themselves better off. You could run another algorithm, which it gets used in places like San Francisco and New Orleans, which is the [inaudible 00:04:20] algorithm. That allows you to find an efficient outcome and there is no way to manipulate that either for the parents, but in general it won't be fair.
Seth: Can you give us an intuition for why a Pareto efficient outcome is sometimes incompatible with a fair outcome?
Alex: They are just two completely different motions in a way, right? What we're really doing is we're allocating school seats, which are objects we are worried about welfare as something that happens from the size of the students. And that's the efficiency criteria, it's really a one-sided criteria in this case. Whereas the fairness criteria is really about a two-sided notion, it takes into account the priorities of the schools as well, whereas efficiency it doesn't matter. Of course, if you cared about welfare from both sides and the way that we say allocate doctors to hospitals, we care about welfare from both sides. Then every stable outcome going to be very efficient so the trade-off disappears. But then you might think, well, the hospitals in that particular case also strategic thinking. So, really depends on how you set up the system. And the way the schools are set out these priorities are really given, schools do not strategize. And so the strategic agents are the ones who welfare we care about, and these are the students and the parents.
Andrey: I see. So other than the school choice example, are there any private sector cases where people might use these sorts of algorithms?
Alex: You can allocate, you could use this to allocate any sort of objects you might want, you know. You might want to think about how you allocate you know, slots for meetings and you might have to take in the preferences over different slots and people have some seniority in the organization that gives them priority about what sort of slots get chosen, and then you want to make sure that there's certain people present in a meeting and you might want to allocate these meeting spots according to the system like this. So as long as you have your two sides, some side that has preferences and one side is objects or maybe it also has preferences or priorities, you can use a mechanism like this.
Andrey: I see. But in the real world we often times don't observe such mechanisms. So for example, meetings are unfortunately not allocated according to any, like, centrally designed algorithm. Is there an intuition for why that's the case?
Alex: These things are hard to build, right? So in order to create a market that allows you to apply the Gale-Shapley algorithm, you need to have all the inputs into the Gale-Shapley algorithm. So in the case of school choice, you need to make sure that all the schools are on board, that parents know what they're doing, they need to know which schools they want to choose. You need to be able to aggregate all this information. And really the main thing is that this has to be, it's a contract between the parents and schools. So if you get allocated to a school, this has to be credible, this has to really mean something. This is not something that you just run for fun to check what's going on and then allow the system to work in a decentralized way. So it's hard to set up institutions that will keep these things centralized, and it's hard to ensure that people really can participate in a way that is convenient for them and safe. So if I have to, you know, rank 400 different time slots in order for you to be able to run your algorithm, well I'm not going to want to do that. Actually somebody asked me to do this recently and I just, I don't want to have to do say yes or no to this huge matrix of time slots. Well I can be bothered to do this, right? And so if the participation is too difficult and too costly, people don't want to do it.
Andrey: That makes sense. So, I guess just generalizing here, we might expect to see such mechanisms in cases where there is really a lot at stake for the participants, and where every kind of extra bit of efficiency or fairness is really important exactly as in the school choice case, but also in I guess the kidney exchange cases or the national residency match.
Alex: That's right.
Alex: So, you know, the more that's at stake, the higher the incentives are to build systems that actually make sense, or remember these the systems are built exactly because there is something that's left on the table. These, the outcomes that we might observe in a market that will be decentralized will be very poor and we will think well it doesn't have to be like this. Some degree of decentralization allows us to improve things in a dramatic way.
Seth: So, one setting in which the quality of the match might be very important is the one where you were recent work has focused, which is on the topic of refugees. Can you tell us how you got interested in that setting for the problem?
Alex: Sure. So my original idea with my co-author Will Jones was to think about why is it that certain refugees end up in particular European countries? There's a convention in the EU that the first country you land in is the country in which you're supposed to apply for asylum. But what we discovered is that actually countries have very different ways of deciding on whether an asylum-seeker should be a refugee. Even though we might think in Europe these conventions would be applied in very similar ways, actually they're not, and for particular kinds of refugees you might be much more likely to gain asylum in France verses in the UK. So I thought that was a bit disconcerting, and it looked a lot like a school choice problem. People have I feel like a slightly different priorities in a way that's what they would be. But of course people would also have very different preferences over the kind of country they actually want to move to if they're already fleeing war. And so, we started thinking about what might happen in international contacts, and then the more we thought about this, the more we realized that this intuition can also be applied to sub-national entities. So we thought about the case of Germany, which at the time a year and a half ago, was experiencing a huge influx of refugees. And we were thinking about well we're in Germany should they go should they go, which [inaudible 00:10:07] should they end up in. And the more we thought about this, the more we realize really what really matters is what is the local area where the refugees end up. And then we discovered that there's actually a great deal of evidence that the match between the local area and the refugee dramatically affects their outcomes over the course of their life, and in fact of the life of their kids. And the evidence for that comes from randomly assigning refugees in Sweden and Denmark in the early '90s, which allows us to precisely estimate what is the neighborhood effect on a particular refugee. Well if we know that, then we should think there must be room for improvement because how do we know, how did the refugees know where they're going to end up? How to local areas decide about what kind of refugees come. Is it possible to centralized a market like this and use this knowledge in order to improve in this market?
Andrey: So just following up on that point. There seems to be an important question of whether these refugees they have vertical preferences as in there's uniformly like Stockholm is better than any other place in Sweden or do some people prefer certain places to other places. because if everyone had the same exact preferences, then someone would get it and it's not clearly prioritized, that we care about which person specifically.
Alex: So that turns out to make no difference to the way that the logical, the matching system works. So consider, it let's take the school choice case again. Suppose there's a really, really popular school that every kid in the city wants to go to. Well in any outcome that's going to be fair. What's going to happen is you're going to send, the school is just going to pick its highest priority kids, right? And it will get them because everybody wants to go to school as a first choice. And not everybody will be able to get in, so it will really matter is what are the kids second, third, and fourth and fifth choices. So, you might be completely right, and in the case of refugees it's possible that they often might want to move to large cities. In the case of the UK, for example, we know a lot of the refugees have said that they want to move to London. But London is not a place that even takes refugees, because the housing is so expensive. So the question is, what is then the refugees second choice, if you know they're not going to be able to move to London. What is their second and third and fourth choice? And that's where I think you will get a lot more differences across different families.
Seth: You partially ask my follow-up question, which is why is this even a matching problem at all? When we think about normal migration, we just think people go where they want to go and they take into account the relative benefits of different locations and the relative prices. Why is this the matching problem, this putting immigrants, the refugees places?
Alex: Yeah so that's a very good point. The context in which we think we're matching would be most useful is in the case of organized resettlement programs. So let me just give you a bit of background. There are a lot of people who migrated and they move around and there's hundreds of millions of people who move around the world. These are not the people we're concerned about, we're concerned about refugees and in fact a very particular group of refugees, which refugees score eligible for resettlement. Which basically says you're already a refugee but even if the war ends in your country, you would not be able to go back there safely. Perhaps because you're being persecuted on the grounds of your political beliefs or your sexuality whatever it might be. And that's about a million people in a given year are deemed to have been in such a dreadful situation that the only way in which we can possibly let them have a life which will be fulfilling is by taking them to a third country, and typically this would be the U.S., the U.K., Australia, and Norway and so on. Out of those one million people that need resettlement, only about a hundred or so thousand actually get it. So the demand for resettlement vastly exceeds the supply of those places. Now these resettlement programs are already very well-organized. They've been going on for since the end of the second World War, and they sometimes they get scaled up depending on what conflicts there are. But when the refugees are resettled to a country, they're usually put to a particular community that really helps them get going. these are not typically young economic migrants that they can sort of traverse an entire continent and then find their job and do it. These are often large families. They're very mobile, they did not come with a lot of assets. They often come with a lot of human assets, actually, but not a lot of physical, financial assets. And they find it very hard to move. And what the community typically does is that it helps with housing. And the cost of moving, because the housing is provided free or at a very low rate for them is actually very high because then they lose that ability outside of their community to have that housing. So in the UK, for example, you would go to the end of the housing queue, whereas if you resettled at a particular local area, you go right to the top you get your housing when you move. So actually, moving is very costly and that's why getting the match right initially actually makes a big difference. So in the case of economic migrants, were not too concerned, because we know that they will gravitate toward where there are jobs. But in case of resettlement, because people stick and because it affects their outcomes so much getting that match right at the very start is what matters.
Seth: So one difference that you point out between the refugee matching problem and for example the school matching problem is that when refugees enter a location, they don't just need a house they also need medical services. they presumably might need a couple of places at schools for their children. How does this change how you attack the matching problem?
Alex: So you know refugees come in families, right? We do everything we can to not separate families, and the families can be very large. So in the UK for example, the modal size of a resettled family is six. So, when we think about simple matching problems where we are assigning kids to schools, there's one kid and there's one school seat, and that makes the math a lot easier. It gets immediately much more complicated when you have to assign for example couples in the case of when we assign doctors to hospitals and the Residency match. We often have to take into account the fact that there are couples and they might have preferences together or individually. And that creates a lot of difficulties, and actually we've had to work our way around that. In the case of refugees is even more complicated exactly as you said. People come in and they have very different if you like sizes and different demands for a lot of different services and that actually creates a lot of mathematical difficulties for finding fair outcomes. Because what often ends up happening is a fair outcome was what we might think of as a reasonable definition of their outcome might be quite wasteful. So we might end up with a lot of space that would be unassigned. So if we really insist on ensuring that the local area priorities are satisfied, and the outcome is fair, there might be a lot of services that are not being used up. And maybe that's not a good thing.
Seth: Would it be fair to say that there's a trade-off between welfare maximization and maximizing the amount of refugees we can handle?
Alex: Absolutely, yeah. So if you think about this as a problem trying to just simply maximize the number of refugees that's resettled, that's interesting enough on its own. Forget preferences and priorities, what you're really thinking of is the following kind of problem, so here's a problem you can compare it to. Let's say you have a trunk of your car, and you're trying to pack boxes into your car, right? How many boxes can you fit in? So, a refugee is a sort of multidimensional sort of object, it's like a box, right? It's a family, it's got many dimensions, got kids, and it's got sick relatives and it's got people who might want to get a job. And the local area has dimensions, too. It provides the school seats and it provides the jobs and so on. So these problems are known to be computationally difficult, and so what we sort of have to think of is of ways of trying to find solutions to problems like that. So we borrow a lot from you know computer science and combinatorial optimization. People have thought about these problems in the case of helping FedEx. We use the same kind of math to try and figure out how we can allocate the refugees to a local areas.
Andrey: So is the system that you proposed, is it also incentive compatible or strategy proof?
Alex: So, you know, if you insist on strategy proofing in this sort of setting, if you really want to make sure that the refugees do not misreport their preferences, and we have to really figure out how they even report their preferences. But let's say they can, you lose a lot in efficiency. So, this will typically really affect the number of families you resettle, a lot of people would be left on a side.
Seth: When you say efficiency, you mean in terms of empty spaces? It seems like you don't mean Pareto efficiency in this sense.
Alex: Oh they would also be Pareto improvements, yeah.
Alex: So you would not get Pareto efficiency it will be a lot of wasted services. So imposing, really wanting a mechanism to be strategy-proof is very costly in efficiency terms in general. And so here, it manifests itself in a very drastic way because of this sort of bulkiness problem, this combinatorial nature of families.
Andrey: I guess one of the things that you might think though is that even if there isn't like a proof or the fact that it's optimal for refugees to report their preferences truthfully they might do so anyway. How do you think about that problem?
Alex: Well at the moment the refugees don't really get asked at all about where they get resettled. So you might need, you know, as I did recently, might need somebody who has been resettled from Iraq to Boston and his job in Iraq was driving trucks. And what he really wants to do is drive trucks but it's very expensive to get a truck driving license, and it's very hard to get truck driving jobs in Massachusetts. So if he, for example, lived in North Dakota, there's a lot of truck driving jobs there and it's much easier to get truck driving license. We might have thought that if we had given them the information and given them some sort of ability to express preferences, he would have said really I would love to go to North Dakota. so we don't really even have a way of understanding how we can get refugees to reveal these things. At the moment, they don't get asked at all, they just get sent somewhere where there is space and where very well meaning resettlement agency workers think they will do well. And often, you know, we might expect that the resettlement workers know enough that they know what might be good for the family. But very often they won't, because families have a lot of private information that they don't really share with resettlement agencies.
Andrey: So this is actually a really interesting aspect of this problem, which also relates to the digital focus of the initiative on the digital economy. How do you set up a system that allows refugees to express their preferences, and furthermore is it possible that you want to combine information from the refugees with the information from the agencies?
Alex: So this is a real difficulty. It's very easy to ask parents "Hey what's the best school for your kids?" because they live in the local area and they just can go around they can ask other people what's the best school that can find the best schools guide and there is just loads of information that allows them to actually express their preferences. What about refugees, you know, in the UK we have 353 local authorities. These are sort of the administrative units that decide on education and social services. How on earth can anybody rank 353 local areas having never lived in England. I mean, if you ask somebody who is from England, they will probably not heard of most of these places. So how do we do this? So the way we think is the best way to do it is to understand what features of local areas are important for refugees. So think about how you might decide which area of the city you want to live in. What matters really is what is the area like? Does it have a lot of crime, is it well connected, is it close to university, is it green, does it have good schools, does it have a hospital if you're very sick, does it have jobs in the area that you want to be employed at. This is the same thing for refugees with things. So you can create profiles of areas, which will collect information that will be relevant to how well they think they can do. And then what we can really get them to do is to tell us what's important, which feature of the area is really important for them. So is it that they're really, really want to be in an area that's extremely safe and very green, but can be potentially quite rural rather than urban. Or is it that they're so keen on being in a city that they're prepared to maybe have a little crime and you know a bit more pollution. So if we can get these preferences over the features of the local area, we can then construct what would have been their preferences had they known every single one of the UK's 353 local areas. It's even harder in the U.S. because there's probably 353 local areas in Massachusetts and then you've got to scale this to the whole country. So the question is how can we present this information in a way that is intuitive and simple and figure out what the preferences would've been. So a lot of this is digital work and we're building a information platform to try and deploy and test in the field to get those preferences.
Andrey: So what has that process been like?
Alex: Tricky you know. A lot of it is we need to make sure that it's this information is easily presented ,so you don't always want to present it in terms of words you maybe want to show people pictures. And, also it matters the way that people understand something like crime is really going to vary from some refugees to others from different countries. People have very different perception of what something like crime is or what does it mean to have little pollution? And so we have to be quite sensitive to these issues as well. So this is a kind of process of learning and at the moment we're gathering a lot of local area information both in the UK and the U.S. to at least make sure we have a comprehensive database that will allow us to construct these preferences. And in that process you know even if we never use them in the actual matching, we will learn a lot about actually what refugees want, and what is it they're looking for in the places where they're going to be resettled.
Seth: You've talked about attributes of these settlement localities, both as things like hard caps, like number of beds, and in terms of things that refugees might have preferences over. To what extent do you decide what you want to think of as something that is an input into welfare versus some sort of hard constraint you have to fit. Is that something you're talking to governments, about what they feel about it?
Alex: That's a great question. So we think about the hard constraints of the ones that being ones that need to meet certain needs. So if you think about a local area, they're really sort of three types of constraints in a way. The first constraint is a true physical constraint. There might not be a hospital that can cater for a particular kind of illness that the family might have, so you really don't want to send a family to that area. The school might be too small to accommodate a kid with disabilities. Something like this. There might not be a house that is big enough for the family or at least in might take too long to find a house that's big enough for the family.
So these are hard constraints. The other type of constraints is sort of self constraints, and they can manifest themselves in sort of pretty unpredictable way depending on the local area. They might say you know we're not very good accommodating single parent families. It's not exactly what was clear why that might be the case. They might say we are not able to provide for families that speak a particular language, because there are not any people in this local area that speak this language. Now again you might think these are slightly softer constraints these are a bit easier to relax nevertheless we also take them seriously. So we think that these are the kinds of things that are important to respect.
Some look more like physical constraints and the other ones look like willingness if you like constraints almost. But nevertheless, we take them in. So we have information on refugees that would allow us to determine whether this constraint is going to be violated or not. Now, the last time is really the priority, so like what sort of area what sort of refugees does the local area really would prefer to resettle. Is it that they really keen on having doctors because they're a rural area that finds it difficult to attract doctors, or is it they would like to have teachers, or maybe they want to have orphan kids because it's a lot of families in that local area that want to adopt orphan children. So these three things will interact together, and we tried to take them as seriously as we can, because local areas deserve and should have the sort of sum degree of control and if they can see that their constraints and their priorities are being, met they may be more likely in the future to actually relax then a little bit. So that's why we take them seriously.
Andrey: I guess one useful piece of context would be how is this currently done? So let's say you're a government and you've agreed to take in some amount of refugees. How do you split it up amongst all your locality? Do localities get to veto every single profile like some sort of job applicant system or is it like here you go here's your five refugees Town "X or whatever?"
Alex: So it works in different countries in different ways, and in the US there are nine resettlement agencies that handle load of the resettlement that their state department takes on. And they first allocate the refugees among themselves for a sort of nice, fair system. And then they work with local areas directly, and they collect profiles of the local areas that they work with. They work with the communities, they make sure that they're happy with the refugees that they get or they talk to them all the time. And typically they agree on the kind of profile of refugees that would fit in the local area exactly as I described, and then they try to make sure that the refugees that arrived match that profile. There isn't really a job applicant system because there isn't a case-by-case review, but you need to make sure that the constraints at least are met, right? So the next step would be would it be, can we incorporate some more information from local areas that would allow them to express further preferences than just the constraints.
Andrey: I see, yeah that's exactly what I was wondering because if it was a system where it was a case-by-case review then they be potentially giving up a lot of control by opting into the system.
Alex: That's exactly right.
Andrey: But in fact, you're giving them more control than what they previously had. Which should encourage them to enter the system.
Alex: That's exactly right. So we hope that the communities that take in refugees will be willing to take even more, and there might be new communities that might also want to do it. Now mind you, the system is a little, right? Because the government decides on what its target is. It decides on the budget that it wants to spend, and this is true in every country. And then it has to sort of convinced local areas to take people on. So there is this sort of public good aspect of it. Well how do you incentivize local areas to take any refugees, cuss what if every local area said we don't want to take any refugees. The government would find it very hard especially kind of in more federal systems to really untwist their local areas into it. And something like a matching system is a very soft way to be able to mediate you know concerns and control from both sides and give local areas I think what they need just to make sure that this process is, you know, takes their needs into account and gives them some control.
Andrey: The last thing that I want to talk about is actually what about the outcomes? So, the hope is that this system is going to improve people's lives. So how does one go about tracking that and what are the outcomes that you'd be looking for?
Alex: So you can think about it in two ways, right? One is we might just think hey we don't care about outcomes what we care about is refugees preferences. Refugees should determine what the outcomes that matter to them are. And so, in a case of schools, sure we tried kids outcomes, but when we decide which kid should go to which school what we really look at is the preferences of the parents. So we don't want to decide for them what's important for their kids, they should have decide this themselves and therefore choose the schools. That is not of you that most governments take at the moment. They are really concerned about the outcomes of refugees, partly I think because they are using taxpayer money to help refugees when they move. And so they accountable for the dollars that day spent or might be charity money and so on. In any case, the outcomes really matter and the sort of outcomes you might worry about is how quickly do refugees get into employment? How well do their kids do at school? How quickly do they learn English and so on. And these things unfortunately we are not actually very good at tracking because it's very expensive and resettlement agencies don't have a lot of money to be able to do this. So can we use better technology to be able to trace the outcomes of refugees, because even if we didn't have their preferences, we could still do better by being able to do combinatorial optimization and deciding what's a good outcome. We might say you know if a particular type of refugee family does really well in a particular kind of area, well that's great, that should have a lot of weight in our combinatorial optimization problem. We should really make sure that this family goes in there. And this might be something like getting into employment within some kind of time frame, so maybe within a year or so. Again if you're opposed something like this, you're sort of deciding that this is good for the refugees. Well it might not be, right, because this employment might not be very good, it might be poorly paid, maybe it would be better if the adults had delayed and gone into further training to then get a better job. So these things always have these difficult trade-offs, but in the, you know you work with institutions and you work in the real world where these things are the way they are often and you try to do the best under these kind of difficult constraints. And so we think that at least by measuring how well refugees do will be able to shine a light on the actual match really matters. Because if you know that, people do really, really differently. You should think about whether they where they go first.
Andrey: Alright so this has been really fascinating. If the listeners want to learn more or help out, where would they go?
Alex: We've got a website setup. It's www. Refugees-say. Com and that's got a bunch of information, it's got our contact details. You can also follow us on Twitter, which is, the handle is @refuge essay, and you can also just Google my name and there's a bunch of papers and things on my website, which are more academic-y. And, you know, we are looking for anybody with skills especially if you're a tech-y and you are interested in, you know, digital markets. Reach out, because you know we would love some people with some really good software and programming skills to build some of these tools.
Andrey: Awesome. Thanks so much for joining us.
Seth: Thanks so much Alex.
Alex: Thanks guys.