YT YouTube2Text

YouTube2Text → Transcripts → BREAD-IGC virtual PhD-level course 2026 | Lecture 3 Core Political Economy Models

BREAD-IGC virtual PhD-level course 2026 | Lecture 3 Core Political Economy Models — Transcript

by International Growth Centre · 16,242 words · 2,369 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:34All right, welcome everyone uh back to
0:37the Brad IGC
0:40Bred uh PhD level course on political
0:43economy. We are here for our third
0:45lecture. We're very grateful to have Pat
0:47Patrick Francois with us. He's a
0:50professor at the University of British
0:52Columbia and teaches in the Vancouver
0:54School of Economics.
0:56And before I pass over to him, I just
0:58want to go over our kind of standard
1:01housekeeping items. Um, I will put this
1:03in the chat as well, but all the reading
1:07lists, slides, and past recordings are
1:11available on our website. I'll put that
1:13link in. And so, if you have questions
1:16around you missed a week or or the
1:19reading list for next week, just check
1:22um that location. The second thing is
1:26for attendance
1:28that is captured automatically by Zoom.
1:30So you don't need to send in any
1:32personal note or email saying that that
1:35you were here. And there is a a bar of
1:4010 uh 10 kind of lectures attended for
1:44the cert certification
1:47um at the end of the course. this being
1:50the third, the this will if you if
1:53you're just joining this week and you
1:55attend the next uh this one and the next
1:58nine, then you'll still be able to get
1:59that that certificate.
2:01Um now, if anyone would like to kind of
2:04be a live participant here and have
2:06their video on uh to be in kind of the
2:10virtual classroom, please raise your
2:11hand now and I'll randomly select 15 or
2:1320 of you. uh please still keep your mic
2:17muted but your video on and then during
2:19the Q&A session you'll be able to
2:21interact and ask um questions. For
2:23others that have questions just put them
2:25into the Q&A box and we'll try to answer
2:28those throughout.
2:30Okay, I think that's everything from me.
2:32So I'll pass it over to Patrick to kind
2:35of go over um everything that we need to
2:38know about core political economy
2:39models.
2:42Okay, thanks [clears throat] Caleb and
2:45thanks to IGC and Brad and CPR for
2:48having me for this uh this lecture.
2:50Okay, so uh yeah, I'm not going to teach
2:53you everything you need to know about
2:54political economy models. There's a lot
2:56more that I'm that I'm going to miss
2:57that I'm going to cover, I'd say. Um
3:02uh okay, so I can't see the top of that,
3:05but I guess you can. So yeah, there we
3:07go. So, um,
3:10what's this, uh, lecture going to be
3:12about? It's going to be back to thinking
3:15about, uh,
3:19what we do with models. Okay? So, I'm
3:21not going to survey political economy
3:23models. I I want you to take away from
3:26this lecture answers to the following
3:29questions. what models are for, how
3:33modeling actually begins, applied
3:34modeling, how a canonical model becomes
3:36useful for you, and I'm going to talk
3:38about a few of the canonical models at a
3:40very high level, how model adds value to
3:43data, and where you must exercise
3:45judgment.
3:47Okay? So my aim is for you to see econom
3:50data in a model first way. Okay? So in
3:54order to do that, I want you to think
3:56back to undergraduate economics. Okay.
3:59So, when you first started taking
4:02economics and I think you probably all
4:04did what I did, you start with thinking
4:07about markets and
4:10the actual observables you have in
4:12markets are essentially quantities and
4:14prices. That's what we see in markets.
4:17Yet, that's not really what we deal
4:20with. What we first do is we make up
4:22this abstraction
4:24which is the idea of supply and demand
4:26which are fictions. right? We we make up
4:29a supply curve, a demand curve. We think
4:31about them when we go a little bit
4:33further. We think about like the
4:34underpinnings of that things like
4:36consumer choice and the behavior of the
4:38firm. Um and then we kind of go a little
4:40bit further and we think about
4:41monopolies and things like
4:42externalities.
4:44So
4:46when we're thinking about the actual
4:48observables, quantities and prices,
4:50we're not basically doing this a
4:52theoretical thing, which is to say,
4:54well, you know, a natural question would
4:55be to ask how are these things
4:56correlated? How do quantities and prices
4:59move together? And that's not a sensible
5:02question. That's actually a question
5:05that if you did that, you'd find all
5:06sorts of weird correlations between
5:08quantities and prices depending on which
5:10curve was moving in the background,
5:11supply or demand. So our our way of
5:14thinking about it is really structured
5:16by the fiction that really drives our
5:19empirical approach to them.
5:21And I want to argue to you that this is
5:24how we as economists think implicitly
5:26for the rest of our careers. We're
5:28really conditioned by that way of
5:30thinking and it's a good thing and
5:34we should do that in the political
5:37economy realm as well. Okay. So what do
5:39these simple models actually teach us?
5:42They teach us the same things that we
5:44learn essentially in the supply and
5:46demand world. So in the in the supply
5:49and demand mode just remind you what are
5:50those things in a market we ask what
5:51shifted demand what shifted supply and
5:53remember these are fictions right
5:54they're not the real things we observe
5:56what shifted those things which margin
5:58moved what's the equilibrium response
6:01what happens if we were to do something
6:02like intervene so that the model becomes
6:06so ingrained in our thinking about this
6:10context that it actually becomes the
6:11language of our conceptions in our mind
6:16deeply
6:18So I'm going to argue the political
6:20economy thinking about political economy
6:21as grad students or as advanced
6:23economists needs the same kind of
6:25language. So um a small number of
6:28strategic problems happen again and
6:30again in political economy contexts.
6:33Okay. So just to list them I mean this
6:35is my kind of uh thinking about it and
6:38I'm sure there are others that uh I
6:40don't think about too much but here are
6:41the ones that are important for me.
6:42Politicians competing for voters. Voters
6:45disciplining politicians. Citizens
6:47deciding whether to become politicians
6:49themselves. Politicians bargaining over
6:52how to divide things up. Usually the big
6:54thing is power. Groups expending
6:57resources to actually control political
6:59office. The credibility of promises made
7:03by politicians. Citizens trying to
7:06coordinate punishments and to discipline
7:08politicians and information asymmetries
7:12that arise between the informed actors.
7:14usually the political agents and and and
7:17the principles who are the voters.
7:19So I'm going to argue that these are the
7:21political economy equivalents of supply
7:22and demand.
7:25Okay. So we can I I I put these readings
7:30on the reading list. They're by no means
7:32essential, but these are the ones that
7:34for me are the canonical models that
7:38provide a a sort of a conceptual way of
7:41thinking about these problems. So the
7:43Dian models for politicians competing
7:46for voters. The fair John and Fieran
7:48paper um in in the issue of voters
7:52disciplining politicians citizens entry
7:55into polit political office is the
7:57Osborne Slovinsky and Basian co-papers
8:00politics and bargaining overpowered
8:01Baron Fjon and on and on. So here
8:04they're listed. I'm going to list them
8:05again later. I'm not going to go through
8:07any of these models in any detail.
8:10I think it's not that useful to do it in
8:12a short period of time, but just so you
8:14have them, these are the ones that I go
8:16to first. They come into my mind when I
8:18think about these particular situations.
8:21Okay. So, what's the point of them? The
8:24point of reading these, I think, is
8:25great. They're great papers. They're
8:26famous papers for the most part, but you
8:29don't need to remember anything about
8:30them. They're useful. Why? because they
8:35have a sort of a strategic structure and
8:37architecture
8:39which when you learn to recognize them
8:42you can
8:44see and think about how to apply them in
8:46empirical settings. So once you
8:47recognize them you see this that this
8:50structure sort of pops up in a lot of
8:51empirical contexts and it can motivate
8:53your your way of thinking about the
8:56data.
8:58So applied modeling usually does start
9:00with data and this is a sense in which
9:02as graduate students you will usually
9:05come at a problem with data first and
9:08that's a good thing. So I think this is
9:10very different than and what I want to
9:12talk about today is very different than
9:13being a theorist. You know I'm not a
9:15theorist. I'm I'm an applied theorist. I
9:17I look at data. I usually work with
9:21empirical people or I do some rud
9:23rudimentary empirical work myself and
9:25then I think about how to structure it
9:27using models.
9:28Okay. So, so it starts with data and I
9:30think this is a great this is where we
9:32should start. We should be disciplined
9:33by data and suppose that you have built
9:35a data set from something, right? So, it
9:37could be a a an RCT could be you have
9:40access to administrative data. You've
9:42come to the archives and you've created
9:44one surveys, census, web scraping, which
9:47is incredibly common and cheap and
9:49relatively easy to do now or just obser
9:52ordinary observational data.
9:54Okay? So you do that work, you do it
9:57carefully and you might even find
9:59something interesting. So you find some
10:00interesting correlations or even some
10:02interesting causal relations.
10:04So so what often happens is the data is
10:07telling you something, right? So
10:08something like this. X goes up and Y
10:10then goes up. Okay? And maybe you've
10:13established it well. You've even got a a
10:14clear causal relationship. It's X going
10:16up that makes Y go up. Okay? So you can
10:20write a paper just showing that. And for
10:22some data, for some variables, that's
10:26actually super important just to
10:28establish a causal relation like that.
10:30And usually that's going to be important
10:33as a standalone fact on its own. Once
10:36theory and once the profession is
10:38decided, we need to know that
10:39relationship. So once that's already
10:41decided, then just establishing it
10:44empirically can itself be an important
10:46contribution. But that's not usually the
10:49case for students doing PhDs. They don't
10:52usually happen upon the key empirical
10:56finding that's just going to resolve a
10:57bunch of theoretical or or open
11:00conjectures in in economics. That's not
11:02usually what happens. What usually
11:03happens is you just found a correlation
11:04in some data. Now, when it becomes when
11:08you can elevate that to something that's
11:10a more significant contribution is when
11:12you can answer a different question,
11:14which is the why question. And that's
11:16the modeling question. Why is this
11:18happening?
11:20So the implied modeling sequence is
11:22inserting the modeling thinking into
11:26that data collection process. So this is
11:29where it fits. You have data, you
11:32establish an empirical fact like X goes
11:35up then Y follows. Then you think about
11:37a mechanism and a mechanism is just a
11:40story, a narrative, a parable, whatever
11:43you want to put it, but it's just what's
11:45going on here. Then you formalize it,
11:47okay, with a model. And that's because
11:50means a few reasons why we want to
11:51formalize it. We'll get into that in in
11:53a little bit, but in a nutshell, it's
11:55because telling the story often doesn't
11:57do the full idea justice. It's hard to
12:00tell whether it all hangs together. And
12:02a model helps you know whether it hangs
12:04together. But it does much more. One of
12:05the things that actually can and should
12:08do, I'd argue, is provide new
12:10predictions. Then it tells you where to
12:12go and get new evidence. Often there's a
12:14back and forth between data, theory, and
12:17data. So the model sits inside, it's
12:19embedded in this what I call empirical
12:21research process.
12:23So step one is often very simple and
12:25modest. And what is that step one?
12:28First, it has to explain the thing you
12:29found. That's the point of building the
12:31model, right? If the data says A goes
12:34up, then B goes up and C equals 1
12:37implies B goes down, then your mechanism
12:40should be able to generate both of those
12:42things. So you have to think about a
12:43story. If it can't do that, if the story
12:45can't explain the basic observations you
12:48have, it's failed at step one. So that's
12:50where you exist in the first instance.
12:53You have to try to use the model, your
12:55thinking, your story to explain what
12:57you're observing. Okay? But the problem
13:00is it's not a problem, but the fact is
13:03uh that's pretty easy to do in a lot of
13:05context, right? Rationalization is
13:07cheap. And that's just why because we're
13:10good at telling stories. Like this is
13:11just part of what humans do. We're great
13:14at it, right? So we can tell any sort of
13:16BS story. BS meaning not true and with
13:19enough assumptions you can almost
13:21rationalize any fact. So some of these
13:23stories are not very interesting but
13:25they may be true. So reducing so just
13:28making the story fit the correlation
13:30that's the point of it but that's
13:32obviously not sufficient. Okay. So then
13:34the question is what what would make it
13:36sufficient? And this what would make it
13:38sufficient is when the model's starting
13:40to provide some discipline beyond the
13:42initial story that you came up with. So
13:44what I mean a m a useful model should do
13:46this. It should tell you some of these
13:48things. So one of them is it should tell
13:50you when the result holds. So x going up
13:53implies y going up. That's the data. But
13:56maybe the model tells you and this is
13:59embedded into your structure is that
14:03well that's only if some parameter alpha
14:05is greater than alpha bar. Why is that
14:07useful? Well, firstly, it's it's may
14:10just be a a consequence of something
14:13that's totally arbitrary like the
14:14functional form assumption. If it's
14:16that, then it's not useful. But if it's
14:18profoundly part of the story, that in a
14:20sense that you can't write down a
14:22functional form that wouldn't make it
14:23true. It actually depends on this
14:25parameter alpha being high enough, then
14:28that's giving you a new empirical
14:31question, right? It's saying where in
14:33the data should this hold? Is there a is
14:35there something that corresponds to
14:36alpha in the data? If it's doing that
14:38then it's already giving us more than
14:40just a rationalization.
14:42It should generate some sort of
14:43heterogeneity. Right? So suppose like I
14:46said y goes x goes up y goes up
14:51then we should also have something else
14:54like this that well that depends on zed
14:57as well. So if you see an interaction
14:59term like that second derivative is also
15:02positive. The model says the effect
15:03should be stronger where zed is high.
15:05Maybe nobody looked at that. Maybe you
15:07didn't think about that before the model
15:09existed. If that's the case, again, it's
15:10a useful model.
15:13It should generate some auxiliary
15:15predictions. These are auxiliary. But
15:16here's some more. The motivating factor
15:18is X goes up, Y goes up. Also, the
15:19mechanism should also imply if X goes
15:21up, maybe W goes down. Often W wasn't
15:24even thought of when you were getting
15:26the data when you were initing the facts
15:28or when you were constructing the model.
15:30and and in that case it's it's likely
15:32that the prediction is super
15:33informative. It's again telling you
15:35somewhere else to look empirically. So
15:37the model here is a tool and it's a tool
15:40because we're not smart enough to do all
15:41this abstractly in our head or in or in
15:43narrative or in English. Putting it in
15:45math allows us to get insights that we
15:48couldn't get otherwise.
15:51So um let's just take an example of
15:54that. So this is all that's all abstract
15:56in terms of x's and y's and ws. suppose
15:57the motive fact is this that you know
15:59coastal constituencies so if you get you
16:02know little electorates or little um
16:04political contests that are near the
16:05coast tend to have more corruption I
16:08don't know if this is a fact I'm just
16:09making it up so let's just think about a
16:11story so one one possible mechanism is
16:13there's reduced spatial competition
16:15right so when constituents or businesses
16:18uh have fewer substitutes nearby because
16:20one side is bounded by the sea there's
16:22less discipline on political bad actors
16:25there's there's less latitude but less
16:26places fewer places to go. So then the
16:29model should also predict what that
16:30lakefront constituencies should have
16:33more corruption because they have the
16:34same issue. So lakes were never part of
16:37the facts used to motivate the model
16:38that the model was about data which was
16:41about the coast versus the inland.
16:45This story would tell you look if that's
16:47the right thing then it should be true
16:48in lakes too and in big lakes.
16:52So that comes from the mechanism and
16:54it's not just mechanically extending the
16:56correlation. You could just mechanically
16:57extend it by saying, "Well, it's just
16:59water and there's water there." That's
17:00not it. It's about this spatial
17:02component.
17:05Okay? So, sometimes the models are going
17:07to actually tell you what data to
17:08collect. So, I can point to another
17:10variable, another question, something
17:12about beliefs. You should ask people
17:14about their beliefs, information,
17:15another treatment. So, this arrow from
17:18data to thinking about the model, then
17:20to new data, that's great if that
17:23happens. And sometimes as I said earlier
17:28actually usually lots of mechanisms are
17:30going to fit the same fact. So you know
17:32you might have model A B and C that can
17:34all generate when X goes up Y goes up.
17:37That's normal.
17:40When these models become valuable is
17:42when they start generating differences.
17:44So if they're exactly equivalent then
17:47we've got no way of telling between
17:49them. But that's rarely the case. Right?
17:52So when they diverge then they start
17:56providing possibilities for
17:58distinguishing between them.
18:01So
18:02the model then becomes a machine that
18:06you can use to design your data
18:07collection moving forward. So what
18:10comparative statics differ across the
18:11models? Which subgroups differ? What
18:14treatments would distinguish the
18:16mechanisms? What observables would move
18:19them in different directions? and what
18:21new data should you collect? So theory
18:24can tell the empiricists what to do
18:25next. So we should all as empiricists,
18:28we should all also be theorists in this
18:30sense. So it's particularly useful if
18:32done prior to going to the field or
18:35after piloting after you've been to the
18:36field and got a sense of what's going
18:38on. You might have some prior about the
18:40variables you should think about. When
18:42you see the piloting, when you go to the
18:45field, when you speak to people, you're
18:46like, you know, these are not the big
18:47moving parts. I should think about
18:49something else. I have a mechanism in
18:50mind. It tells me other things that I
18:53should ask.
18:55So I'm going to provide an illustration.
18:57I think this is a great model. I I put
19:00it as an a reading to do. Uh it's a it's
19:03a paper from Gerard Padre Miguel in
19:06Review of Economic Studies and it
19:08basically asks the question and it's a
19:10pure theoretical paper. Okay. Why do
19:12citizens continue to support leaders who
19:15are basically stealing from them or
19:18governing terribly? So
19:21usually and he starts from this position
19:23that this is got something to do with
19:25identity. People tend to support people
19:26who are from the same identity group. So
19:28it happens in in societies that are
19:30polarized or clustered around identity
19:33markers like for instance tribal
19:35markers. So he notes that a tempting
19:38explanation I don't think he actually
19:40notes this but I think a tempting
19:41explanation is just say look we've got a
19:43taste for supporting somebody from our
19:45own group. We just have an inherent
19:46tendency to want to support people vote
19:49for people who are like us
19:50linguistically culturally physically you
19:54could just say this is like a
19:56psychological tendency. Okay. So he
20:00takes I think a much more deep economic
20:03approach. She says look this can emerge
20:04from an interaction that we don't have
20:06to posit that as a feature of our
20:09preferences. This can be an equilibrium
20:11outcome.
20:13Okay. So here's the core mechanism. So
20:15if you've read the paper you'll be
20:16familiar with the notation here. This is
20:20sort of a a slightly bastardized version
20:21of his notation. But I think the
20:24interesting thing is that um that his
20:26his uh his his his
20:29theoretical approach is you know a a
20:32beautiful elegant way of generating the
20:35story as an emergence rather than a
20:37direct assumption. Okay. So suppose the
20:39following is true. He he makes his
20:41assumption. Suppose if you keep the
20:43incumbent the probability that you're
20:44going to hold on to power is greater. So
20:47he calls this gamma overbar and I'm
20:50going to call it for a for group A.
20:52That's the probability you'll stay in
20:54power if you keep your incumbent in
20:55power. If you remove your incumbent,
20:57that probability is lower. Some gamma
20:59underbar a and suppose I value holding
21:02power. So that means the value to me as
21:05an a person if I have an a leader is
21:07bigger than the value if I have a b
21:09leader. Okay.
21:12So he notes and he generates this as an
21:15equilibrium object that supporters may
21:19accept a loss as large as this amount.
21:22Okay. And this is a kind of a political
21:24slack. And the story is very simple. If
21:28I were [snorts] to remove my current
21:30politician then I will have suffer this
21:33decline in the probability of holding
21:36power. How much do I value holding
21:38power? This much. So that's the actual
21:42amount that I fear losing
21:45and delta is a discount rate.
21:49This is a type of a rent that the
21:52politician
21:53can exploit. They can do things up to
21:56this. They can steal up to this. They
21:58can be slack up to this. It's a
22:00political rent. That e kleptocratic rent
22:03that he calls it. So the more I fear the
22:05other group taking power, the more badly
22:07my own leader can treat me. Why is this?
22:09Well, the worse it would be for me if
22:11the other group takes power, the bigger
22:13this rent is. So, this is his mechanism
22:17for explaining why people will be in an
22:21equilibrium where they continually
22:22support leaders who are venal corrupt
22:27mal in whatever way. Okay.
22:31So, it's a nice model and if it were
22:33just a model and it is just a model in
22:35his version, it's nice enough that it
22:37makes this complicated dynamic problem
22:39very simple. You can see the moving
22:41parts and that's why it was published,
22:43you know, 20 years ago in the review of
22:44economic studies. It's a great paper.
22:49Okay, so this is just kind of the
22:51technique of it, which I think is nice
22:53too. It makes the problem simple, but I
22:54won't go into that too much. Now I want
22:56to tell you what my co-author and I did
23:00when we saw a an empirical fact in in
23:04Maharashtra. So we were collecting data
23:05there and we found this very unusual
23:08empirical pattern. We found that when a
23:11political office was reserved for a
23:13member of a particular cast and here it
23:16was OBC but just think about a cast
23:18that's not traditionally a powerful cast
23:21then mil village governance actually
23:24improved and the effect was concentrated
23:27uh in these OBC groups that were large
23:29but not super dominant and we found that
23:32the observed characteristics of the
23:34leader did little to account for the
23:36pattern. So it didn't seem like it was a
23:38improving the quality of people's
23:40standing per se. Okay. So normally we
23:43might have expected the following. So if
23:45you restrict who can stand for office,
23:47that means you're implying there's less
23:49competition and that means you're going
23:51to get worse governance.
23:53Okay, so that's what we would usually
23:55think, but we found that governance was
23:57improving. So what we realized was that
23:59the politics of fear model, that model
24:01that I just gave you, was able to affect
24:05a margin that could be explaining this
24:07data. Okay? So we cannibalized that
24:09model.
24:11Okay. The original model contained a lot
24:12of structure that just wasn't important
24:14in the Maharashtran village setting.
24:16Okay? There was activity specific t
24:18taxation. There was a particular form of
24:20specialization. These things weren't
24:22that relevant. And the reason is because
24:24the capacity to tax in that context was
24:26very low. And this specialization story
24:28just wasn't relevant for these different
24:30OBC groups.
24:33So what we did keep was what was
24:34essential and that is that basic
24:36equation that I showed you. The value of
24:38group control plus leader specific
24:40advantage and retaining it was
24:41generating the rent and what
24:44reservations political reservations were
24:46doing was essentially affecting the
24:49entry margin. They they affected who
24:51could compete for leadership.
24:54And when a reservation was given to a
24:56group, it meant that the competition had
24:58to come entirely from within because
25:00that group was guaranteed to win. Okay?
25:04So that original politics of fear model
25:07didn't have a process for replacing
25:09incumbents.
25:11But political reservation works
25:13precisely by changing that replacement
25:15possibility. Okay. So we built an
25:18extension which modeled the challenger
25:20entry and the governance offer that a
25:22challenge makes. So what often happens
25:24and this is what happened in this
25:25context that modeling in a particular
25:28context requires taking that off the
25:30path off off off the shelf model and
25:34making something which was relatively
25:38minor in that model here it was an off
25:40path behavior into the core of the
25:42model. So that's the case in a lot of
25:45those core models. They're focused on a
25:47generic economic context, but when you
25:49get a particular empirical setting, it
25:51tells you what you need to focus your
25:53model on. So we built down this
25:55extension and here we had two
25:58politically coherent groups or juties
26:00and the leader would obtain office
26:02rents. Citizens would value holding
26:04office and the leader would choose how
26:06much effort to put in. Okay. So I won't
26:09go into the details of that but
26:11essentially we took things that were not
26:13in that model and added them in but kept
26:15that core structure. So this was simpler
26:17than the original tax and patronage
26:19model of his but richer in the dimension
26:21where where the data was rich. Okay. So
26:24so what I want to tell you is this led
26:28to a set of very different predictions
26:31than his original model had. Okay. But
26:33it was still using the same idea. So
26:36here's what a reservation does in using
26:38the thinking of his model. So without a
26:41reservation, remember the probability
26:43that the group will hold office if they
26:46keep their leader is some gamma a
26:50uppercase. If they replace him with a
26:53challenger, it's gamma a lower case and
26:56this is lower. That's the cost of moving
26:58against an incumbent who might be
27:00terrible. That's the cost. So you have
27:02to risk losing control. With a
27:04reservation, both of those just go to
27:07one. So that difference collapses. Why
27:09does it go to one? Because essentially
27:13the electoral office says to you, you
27:15know, you've won. Your group has one
27:17office for sure. It's been reserved for
27:18you. That's going to change the law of
27:21motion of political power. And that
27:23changes everything in that model. So now
27:28the disciplining has come from entry
27:30from within. Before the disciplining on
27:32power was losing to another group, but
27:35that created a lot of bad problems for
27:38the constituents supporting the leader.
27:39Now you could throw out your leader.
27:42You're guaranteed to hold power. So the
27:45discipline now comes from who's going to
27:46want to challenge. What could how how
27:49good could they do? What can they
27:50promise you? So this changes the
27:52equilibrium by changing the off path
27:54behavior.
27:56So the incumbent's electoral advantage
27:58creates an equilibrium rent he consumes
27:59as lower governance in the politics of
28:02fear model with a reservation that's
28:04gone and so has the rent.
28:07So the model delivers
28:09essentially
28:11an empirical prediction which is one
28:13that we wanted to generate right that
28:17this was the level of governance in an
28:20unreserved village. It was whatever the
28:22first best was minus some amount. That's
28:25the politics of fear in the reservation.
28:29That level of governance, the difference
28:30between a reserved and an unreserved
28:34was positive. A reservation would
28:36improve governance. So this was
28:38generating the raw empirical fact that
28:41we found which of of course is step one.
28:44That's what you want a model to do.
28:46Okay. But
28:50and this should be the intuition. I
28:51won't go through it again, but it's on
28:52that slide if you want it again. It's
28:54essentially that it removes the
28:56political advantage of being the
28:57incumbent because now you can be
28:59challenged by anyone and the group
29:01doesn't lose power. Okay,
29:04but here's a non-obvious prediction. The
29:07model says the effects of reservation
29:10are governed by this initial difference
29:12between what would happen if you threw
29:14someone out and if you keep them. So if
29:17you had a very very take a strong group,
29:19a very strong group, they were
29:20guaranteed to win in any case.
29:23So in that case, these numbers, these
29:26gamas are close to one. That is, if they
29:28were strong, even if they threw a guy
29:29out, wouldn't matter. They were probably
29:31going to win anyway. In that case, a
29:33reservation is not going to have much
29:35effect because a reservation collapses
29:36this difference. But if the difference
29:38was small anyway, it's not going to do
29:39much. If you took a very weak group,
29:42they were very unlikely to win again.
29:44And then a reservation would also have
29:46little effect. So what this is telling
29:48you is there's a nonmonoticity. The
29:50model which was generated to explain the
29:53core correlation which is putting in a
29:56reservation improves governance tells
29:57you no no it's only going to happen if a
30:00group is like in the middle size where
30:02it could compete for power but it's not
30:04guaranteed to win it. That's where you
30:06should find it.
30:08So that tells you well if I can observe
30:10these things then I can test this model.
30:14I can test whether this story is the
30:15right story. And we can, you know, we
30:20could just basically look in unreserved
30:21villages and say, what is a what is a a
30:23group that is going to win for sure?
30:25What is a group that is almost never
30:26going to win? And here it is. If you had
30:29less than 25% of the village village
30:31population in your group, you won about
30:332% of the time. So then you're one of
30:35these groups that's almost never going
30:36to win. If you're above 50%, then you
30:40kind of have a plurality. You're almost
30:41always going to win. It's these groups
30:43in the middle where you should see it.
30:44So we could classify
30:46village structures, population
30:48structures and say here's the
30:49reservations that should really be
30:51driving our correlation in the data.
30:54And that's what happened. We we measured
30:56a bunch of observations uh different
30:59types of government governance uh
31:01quality effects and we found that the
31:03effects were only positive in these
31:07villages
31:08in the ones where that population of the
31:10reserve group was a particular size. if
31:12it was bigger or smaller in those
31:14extremes, it had no effect. So the the
31:17model hetrogenity was exactly what we
31:20saw in the data. We had no idea that
31:22this was going to be the case. We were
31:24just fishing around looking for models
31:26that could explain it. And when we found
31:28this one and we found the extra one,
31:29then the model generated a huge amount
31:32more value. And I started to believe it
31:34at that point. Before it was just a
31:36story, then I start believing it. There
31:38was another set of predictions which we
31:40also tested and I'm not going to go
31:43through them because I want to move on
31:44to talking about modeling more in
31:45general. Okay.
31:48So what did the modeling actually add
31:50here? What the modeling added was
31:57an explanation
31:59which came from totally left field. Like
32:02it wasn't the explanation that we
32:03thought would have been true.
32:06But when we started thinking about it
32:07and had to do stuff with it to make it
32:09fit our context, we found this other set
32:12of predictions and they fitted. They
32:14actually told us they they made us run
32:17regressions we would never have run
32:18before. We had no idea xante that this
32:20is what we would end up with
32:21empirically. But when we did them, they
32:24made sense. They made sense in the data
32:26and they started to give us a richer
32:28understanding of the original
32:29correlation that we that we built. Okay.
32:33So the model formalized the story and
32:36made predictions and then it took us way
32:37further than we could ever have gotten.
32:40Okay. So this is the model exercise in
32:43full a little bit more. I'm filling in a
32:45few more arrows here. We have an
32:47empirical puzzle. Essentially we don't
32:50we're not theorists in the sense of
32:52we're coming up with wholly different
32:53mechanisms. I mean some of you will do
32:55that at some point but mostly that's not
32:57the case. You just borrow a mechanism
32:59from one of these canonical models or
33:01from something else and then you throw
33:02out the bits of it that are built for
33:06the generality of the problem. A lot of
33:07them are not going to be relevant in
33:08your context.
33:10You're going to indogenize and and
33:12enrich the bit whether your data is rich
33:14and that will often depend on the
33:15institution. You'll solve the model.
33:17You'll derive conditions and some
33:19predictions from it. You construct then
33:21new data. You'll be driven by the model
33:23to look for new things and you test
33:25again. And that's what I mean by applied
33:28political economy modeling. And it may
33:31not seem obvious to you. It's very
33:32obvious to me that this is exactly what
33:34we do when we move from first learning
33:38about economics, which is essentially we
33:40build models, these very abstract
33:42things, and then we start thinking about
33:43the empirical. We don't do it the other
33:45way. I mean, you'll notice when you
33:47first do econ, you don't start with
33:49econometrics. You start with models.
33:53Okay. So, what's a toolkit here? Okay.
33:56So, so once you've seen one example then
34:01and once you've worked with this a few
34:02times the models become very easy to
34:04apply and again they're not a list of
34:06papers to memorize but when you're
34:09familiar with these models these are the
34:10ones that I'm telling you I think have
34:12useful structures that come up a lot in
34:15political economy contexts and the great
34:17thing about them is they're old so
34:19they're easy to read right they're
34:20written in a way that's you know often
34:23the first paper in in an area they have
34:24to do a lot of the handwaving and a lot
34:27of the contextual structure for you and
34:28they build it in a simple way so you can
34:30see the mechanisms clearly. If you
34:31understand these mechanisms, they're
34:32going to give you a set of tools that
34:34will be useful in lots of empirical
34:36context. These are the I'm just putting
34:38them
34:40these are the ones that were on the
34:41second slide. At at the end of the
34:43slides, I'm going to give you the
34:43precise uh readings for these, but I
34:46think these were shared with you anyway.
34:47So again, you don't need them for every
34:49model. You need to recognize when some
34:52of the structure in these models is
34:54coming up in an empirical problem. And
34:56sometimes multiple models will have
34:59a role in a in a particular empirical
35:01problem. All right? So they're not
35:05things you need to memorize, but they're
35:06like you could think of them like spare
35:08parts. Okay? You don't want to copy the
35:10structure of any particular one. So you
35:11don't want to rebuild that whole car,
35:13but you know, you want might want to
35:14take a part of that car. So it's a piece
35:17of machinery that helps you explain some
35:19fact that seems to be popping up for you
35:21in the data.
35:24Okay. So you you also and and you are
35:27you should and you will cannibalize
35:28models, right? So you'll take bits of
35:30them. You take bit of say this is just
35:31an example the par baron fedon
35:33bargaining model with a contest
35:35technology. That's a paper that that
35:36that you know that I built and and
35:39worked with others on to to to explain
35:41power sharing in in in African politics.
35:44Um and there's a bunch of other ways to
35:46put these things together. So you pull
35:47from different parts of these models to
35:49build your own model. Okay. So all also
35:52it's some of these models are really in
35:54a context got nothing to do the one that
35:56you're going to be studying empirically.
35:57So the corid soil model is a sort of a
35:59sender receiver model about
36:03asym asymmetries in information and how
36:04that affects signaling and and
36:06interpretation of signaling. This is not
36:08really couched as a political model. I I
36:11don't think in that paper that was even
36:12talked about in a political context. But
36:14this is an architecture that's really
36:16useful for thinking about a bunch of
36:17political economy interactions between
36:19parties and voters, bureaucrats and
36:21politicians, experts and citizens and
36:23lobbyists and policy makers because you
36:24can see in all those context there's
36:25that core essential information
36:28asymmetry and knowledge asymmetry and
36:30and that affects how signals are sent
36:32and how they're interpreted.
36:34Okay. So, so, so let me just kind of
36:36talk a little bit at overview level
36:38about um
36:41elements of the modeling, right? So,
36:42these are that was the kind of the
36:44political economy context. Now, there's
36:46kind of some some more general
36:47questions, right? So, so one first
36:49question is um how much model do you
36:51need? Right? So,
36:54if your model explains the mechanisms,
36:57generates predictions, tell you what
36:59tells you what to collect, then that's
37:01maybe enough. You should stop. And
37:04there's a question of parimony in in
37:06applied modeling that's very different
37:09in its in how you judge it relative to
37:11thinking about pure theory. Um so so I
37:16mean the way I think of it is in
37:19applied theory I'm always thinking what
37:21do I need minimally to get to the data
37:23and do interesting things with the data.
37:25In a true theory paper and a lot of the
37:29papers that are canonical models are
37:30written more like that. They're really
37:32looking for the generality of the
37:34result. How general is it? You know,
37:37when does it hold? When does it not
37:38hold? You know, really exploring
37:43the limits of the story. And sometimes
37:47that is going to be something useful for
37:48you because you can get predictions off
37:49that. But often all of that generality
37:52which is how we would judge a good
37:55theory paper like you know has it really
37:58fully illuminated the mechanism what's
38:00necessary what's sufficient that's not
38:02what you need in an applied paper. You
38:04just need
38:06a believable story that rationalizes the
38:10correlation. You need to make it clear
38:12that it's not your it's not being bought
38:14by some sort of cheap assumptions in the
38:16structure. But you can use very simple
38:18assumptions like even quadratic form or
38:20something. If it's clear that that's not
38:22what's buying it, that there's a deeper
38:23intuition, then you can generate the
38:25results from that. So that wouldn't fly
38:26as a theory paper, but as an applied
38:28theory paper, that's exactly the level
38:30you want to be at. You don't want to do
38:31all that machinery for nothing. Okay. So
38:34when do you need more structure? So this
38:36is
38:38well, we'll think about what structure,
38:39what I mean by structure in a second. So
38:41So if you want to ask questions like
38:42this, what happens if you were to
38:44fundamentally change the system? change
38:46the voting rule, change a term limit. So
38:49you might put in a term limit when there
38:51aren't term limits for political office.
38:53Um you want to change who's eligible to
38:55stand. What happens essentially under an
38:58institution that you never observe in
39:00the data.
39:02So now the observed reduced form
39:06relationships aren't going to be enough
39:07because you're not observing these
39:08things. But if you can believe that some
39:11of them are invariant to these changes
39:13and this is where you have to make the
39:14assumption and and find it believable
39:17then you can run counterfactuals
39:19along these lines under the assumption
39:21that the parameters that you estimated
39:24initially in the data will hold across
39:27the counterfactuals. So that's when you
39:29can do structural modeling and that's
39:30when it's worthwhile to do it. So here
39:33you're going to make assumptions that
39:35will essentially be stronger than in the
39:38simple reduced form world. But with
39:40those stronger assumptions, you're going
39:41to be able to get more results. Results
39:45about things that would be more like
39:46systemic change. Now that comes at a
39:50cost. The cost is those assumptions.
39:53So those are stronger claims, right? So
39:56you're you're starting from primitives
39:58that you are taking much more seriously
40:00and actually much more literally, right?
40:02So preferences, technology, information,
40:04institutions, stocks, all those things
40:05are things that are now going to
40:06generate your equilibrium.
40:08So you're actually going to generate the
40:10data with your model. You're generating
40:14a rationalization of the data from your
40:16model. And you're going to be when it's
40:17useful is when you can hold those
40:19primitives, at least some of them fixed.
40:22So primitives on preferences,
40:23technology, and claim that those things
40:24are going to be invariant to the the try
40:26to change that I'm imagining. then
40:28you're going to get predictions about
40:29the world that would happen in that very
40:32different counterfactual world and that
40:34can be useful when when people will buy
40:36when you buy and when people will buy
40:38that those things are invariant
40:40and if they won't then it's a useless
40:42exercise right so so I've certainly met
40:45people who argue for doing this sort of
40:47structural work for or not doing it for
40:49the wrong reasons right so I I mean
40:52maybe people don't do this anymore but
40:54when it was when I was in grad school I
40:56heard people saying this structural
40:57models are just better because they're
40:59more rigorous. They, you know, they they
41:01have more math. There's more theory.
41:02They take theory seriously. That's not a
41:04good reason to do it. To me, the good
41:06reason to do it would be, oh, if I can
41:08actually understand something that I
41:10couldn't understand by doing the simple
41:11reduced form model, that's a good reason
41:14under a set of assumptions that people
41:15find believable, then okay, then it's
41:18worthwhile doing. And some people just
41:21start with an exit position that oh
41:23taking this like assuming preferences
41:25like nobody knows what the preference
41:26you can't just assume that structural
41:28models are never believable. I find that
41:30also an un an unconvincing reason to not
41:33do it. So I think they're they're both
41:35the wrong way to proceed. You should ask
41:37what question can I answer with the
41:39additional structure that I cannot
41:40answer without it? If none not at all
41:43then the question is is that structure
41:45believable? believable enough in
41:48believable enough to be invariant. Okay,
41:50then then it's worth doing.
41:54Okay, so that's my two cents on
41:57structural modeling when to and when
41:58not. And for me as an applied theorist,
42:01I've done both and in different contexts
42:03I think they're useful. Okay, so now
42:07this has changed things a lot. So LLMs
42:10have changed the technology of model
42:13building and they've changed it in
42:16mostly a good way and you should utilize
42:20them. I utilize them. They are useful
42:22for building models. The latest
42:24generation are two years ago. They
42:26weren't. A year ago they weren't that
42:27much. Now they're pretty useful. They're
42:30not perfect, but they do these things
42:32pretty well. You need to check them. But
42:34this is what they can do well. They can
42:35do algebra. And for someone like me that
42:37is just my life was littered with
42:41algebra errors. So I would for me
42:44algebra is a statistical process. What I
42:46mean is I do things five times and if
42:47two of them are the same then I would
42:49assume that that was the right one. So
42:51I'm terrible at algebra. So this made my
42:53life so much better. Right? So now I can
42:56just check it. I check one it looks
42:57right. I'm so happy. Derivations
42:59comparative statics trying alternative
43:01functional forms. So exploring elements
43:03of the model space that would have taken
43:04a long time. You can do it easy checking
43:07candidate models, generating several
43:09modeling approaches quickly. So you have
43:11some vague ideas, you put it in, it
43:13gives you some structure and you can
43:15think is this capturing it well or not?
43:17If it is, then you can say, well, you
43:18know, what are the things this is giving
43:19me? You can do this fast now. This is
43:21fantastically valuable. It's an enormous
43:24improvement in our research technology.
43:27So what's become scarce? Something has
43:29become scarce now. So if models can
43:31generate
43:33many at one time like 10 at once
43:37then the scarce input becomes what it
43:40becomes actually thinking and and you
43:42can't see thinking where do you see it
43:44where do you see the effects of it it's
43:45good judgment
43:47so building intuitive parimonious real
43:52illuminating defensible to you and by
43:54you models it's not always good at doing
43:58that so that you are still very much
44:01involved in the loop. So, it's giving
44:02you a sort of a a big leg up over a lot
44:06of elements of the model building, but
44:08it's not taken us um primates out of the
44:11model building process yet.
44:15Okay? And and and you can see like we
44:17used to be at least when I was in grad
44:19school, we were we I was impressed by
44:20people's mathematical capacities. Not
44:22just because, you know, there's this
44:24kind of thing that a lot of economists
44:25have, which is like if you if you're
44:26math, you're just smart. It was because
44:28they could explore the model space very
44:30fast. They could explore this kind of
44:32weird world, right? This abstract space
44:34of modeling. If you're good at math,
44:36that's the language of models. They
44:37could explore it fast and that's
44:38valuable. So, that was impressive.
44:40That's not impressive anymore. Okay. So,
44:42now we're increasingly asking, you know,
44:44why why are you building this
44:45complicated machinery? What's the
44:47algebra doing? Is it worth doing? Why
44:49couldn't things be simpler? And and more
44:51importantly, what do you actually
44:53believe? What can you defend?
44:57Okay. So, you have to own your model.
45:01That means what? You should use the LMS
45:03to build them. Absolutely, you should.
45:04And I put should in capital letters cuz
45:06actually you're wasting time if you
45:09don't. I think you should. And I don't
45:12know where that should be inserted. It's
45:14different for everyone, but I think they
45:15they are a valuable tool. We should be
45:17using them. But if it helps you build
45:18the model, by the time you present it,
45:20it still has to be your model. So you
45:22should know why every assumption is
45:24there, what results depend on them, what
45:26happens if you pull something out,
45:29why you rejected alternatives. You have
45:31to thought through all the alternatives
45:32and why you rejected them. Okay? It
45:34should become something like, you know,
45:35when you when you have a very good
45:36friend, you know them so well. You know
45:38how they'll react even in situations
45:40that that you haven't seen them in. you
45:42just know, oh boy, if you know, I know
45:43what they're going to do here even
45:44though I have, you know, that's that's
45:46the way you have to know your model. It
45:48has to be completely familiar to you.
45:52So that's the test, right? And this is
45:54the test that's increasingly becoming a
45:56big part of our profession. We using it,
45:58I'm using it all in my grad courses now.
46:00It's actually getting up and presenting
46:02your model and defending it against
46:04questions like this and worse questions
46:06and more difficult questions that you
46:08get. Why do you need that assumption?
46:09Why is information private? Why was this
46:12the sequencing? What happens if I change
46:13the sequencing that seems arbitrary?
46:15Couldn't another mechanism generate the
46:17same fact? Have you thought about that
46:18other mechanism? What do you think about
46:20it now? What happens if I remove
46:22something in the model?
46:25That intuition has to be very clearly in
46:28your mind. And you can't just say, well,
46:29that's in the appendix, right? That now
46:32it's it's even worse than it used to be
46:34if you said it because now it just makes
46:35it look like, well, you actually didn't
46:36generate your model. You don't own it.
46:39And here's the last bit. very important.
46:42LLMs are not good at attributing. Okay?
46:46So, they will take they will do what I
46:48said to you, right, which is cannibalize
46:51a model and give you something that's
46:53useful, but they don't tell you what
46:54model they cannibalize. They're very bad
46:56at that. You have to really push them to
46:59do it. And so, that may just be a a bug
47:04and that might be changed through time
47:06as the LM become better, but right now
47:08that's the case. So, so you can say,
47:11"Oh, I've got this vague idea from a
47:13model." And it goes, "Hey, yeah, like I
47:14I think this is what you should do." And
47:16it gives you a formalization that's just
47:18completely lifted from someone else,
47:20right? Who also had the vague idea maybe
47:22in a different context. You have to site
47:23that paper. You have to push the LM to
47:26tell you what is this like, where did
47:28you get this? Is does it resemble
47:30anything? You have to push it hard right
47:31now for it to give you that information.
47:33And if you don't and if you take a model
47:36that someone else's that the LLM gave
47:38you, you are responsible for that. And
47:40that is plagiarism, right? You're
47:42responsible ultimately for that. So you
47:44have to do a deep dive and make sure
47:46that whatever it's giving you, you go
47:48and find the paper and you read that
47:50paper and you site it properly. Okay? If
47:52not, then I put that in bold at the end.
47:55I have seen careers ended by that sort
47:57of not not in the LLM context, it's too
47:59early, but by plagiarism. Okay?
48:02you that's the one thing you don't want
48:03to be. You have to be correctly
48:05attributing. Okay.
48:08Okay. So let's me go to sort of the over
48:10overview level a bit.
48:12These are five questions to ask when you
48:13have a new empirical result. What is the
48:15simplest economic environment that could
48:17generate that fact?
48:20You have to think about what should be
48:22primitive? What should emerge? And those
48:23canonical models in particular contexts
48:25are going to help you that they're
48:27exactly doing that. Then what else has
48:30to be true? If the mechanism that I'm
48:33I'm suggesting is the one that works,
48:35what has to be true if that's correct,
48:38then think about more broadly what
48:40competing mechanisms could explain the
48:41same fact and what data would help you
48:43to distinguish between them. And then
48:46when do I stop? Do I have enough model?
48:49Is there something else I can get for it
48:51from it? Maybe then I could maybe if I
48:53want to look at a different sort of
48:55world that might be an important policy
48:56world, maybe I want to put more
48:58structure in. These are the questions
49:00that come to me when I do model
49:04building. And like I said, I mostly work
49:07with empirical people who come to me
49:09with I mean there's two sorts of
49:11situations that I kind of do this model
49:13building for. One is with students. So
49:15it's it's remarkable. We used to learn
49:17this kind of skill because a lot of the
49:19fields were much more theoretical. So we
49:21would see these canonical models and
49:23work with them. Now so much of the work
49:25and this is a good thing but we have
49:27evolved towards doing much more
49:29databased work. So many students will
49:32just have they're swimming in data and
49:34have a bunch of correlations and then
49:36they come to people like me and say my
49:39supervisor said I have to come up with a
49:41story to rationalize this and you know
49:44that's good but the story can be much
49:46more than just a rationalization and it
49:48should be otherwise and I see this a lot
49:50people are just putting in math what
49:52they could put in English and calling it
49:53an empirical model it's totally useless
49:56you might as well just say it in English
49:57the model has to do the work of
49:59generating something new for you
50:01empirically, not just rationalizing the
50:04fact.
50:06So, so what I want you to leave this uh
50:09talk with and I'm happy to take Q&A, but
50:11here's the kind of um
50:15questions that I and the thinking that I
50:17want you to take away. So, so when you
50:19see an empirical pattern, you shouldn't
50:22think the first thing shouldn't be you
50:24thinking what regression should I run
50:26next. um should also ask what's the
50:30environment that generated this
50:31empirical pattern
50:34and that's going to tell you what else
50:36you should run next. That's going to
50:38tell you what that particular
50:40environment those moving parts
50:44are generating empirically and then you
50:45go and run that. So, so like with supply
50:48and demand, right? If you just saw a
50:50bunch of correlations between quantities
50:51and prices, you could think about, okay,
50:54maybe I should run a sort of a different
50:56specification. Maybe I should put in a
50:58squared term or something. It's totally
51:00useless to do it that way. But when you
51:02start thinking about supply and demand,
51:03when that's deeply embedded into how you
51:05think about that context, then it gives
51:07you a different structure. Now I'm
51:09thinking, oh, well, what are things that
51:11could move the supply curve maybe
51:13exogenously that would allow me to
51:15identify the demand curve? It totally
51:17structures how you think. So that way of
51:19thinking which I think is deeply
51:21embedded how we think about
51:23the most simple context should be
51:26embedded in how we think in political
51:27economy context as well. And what I'm
51:29suggesting is those canonical models
51:32give you that set of tools for thinking
51:34that way. They're just like the supply
51:36and demand of the standard market.
51:40So the objective is essentially is to
51:42move from an empirical fact.
51:46This is the correlation that you found
51:48set of facts and
51:51think about the story which is the
51:53economic mechanism. Then build a model
51:57use whatever tools you need to do that.
51:59But you have to own it.
52:02That model should give you new things to
52:04look for in the world. And when you
52:07found those things, if they fit, then
52:09you actually have a paper. If they don't
52:11fit, tells you that's not a good model.
52:12Go back and think about it again. So, it
52:15all comes from this this position that,
52:18you know, I have um which is a kind of
52:20an an act of faith in a sense, right?
52:23It's like I think that the economic way
52:25of thinking actually does generate
52:27insights into the world. So if I find a
52:31set of correlations that don't make
52:32sense, I try to think of them with a an
52:35economic model and if that model
52:37generates predictions and they fail,
52:39then I don't say, well, this is just a
52:41mystery. We'll never be able to
52:42understand it. I think no, there
52:43actually must be a different model. And
52:45I will keep exploring the model space
52:48till I find the one that makes sense.
52:50And in fact, the best thing is if the
52:52first five models don't work, they
52:54generate predictions that don't work,
52:55then you've got a super interesting
52:57motivation for a paper, right? It's
52:59like, look, you may think it's X. Well,
53:01model X generates this and that just
53:03doesn't fit. You may think it's Y. Then
53:05you can say, here's my model, which is
53:07this super weird funky thing, but that
53:09actually fits. Then you've got a great
53:12paper. So
53:16I think political economy is for that
53:19process. Political economy is a set of
53:22theories firstly explanations for the
53:24world that we use to generate deeper
53:28ways of going into the world into the
53:32data.
53:34Okay. So those are the papers I think
53:36the reading list has been circulated
53:38with you but I that I think are great
53:39papers to read that have helped me a lot
53:41in thinking about different political
53:43economy contexts. Um I think uh I that's
53:48really all I wanted to say. So that's
53:50I'm now happy to take questions and I
53:52think Caleb's going to moderate the
53:53questions and um just far away. I can
53:57stop screen sharing Caleb.
54:00>> Sure. Yeah, that that works great.
54:02>> Okay.
54:06So, uh if anyone has questions, they
54:09there's many that have been coming in
54:11through the Q&A box, but now's a good
54:13time to send through any more for for
54:15Patrick if you have them. I'll start
54:17with what we have and we'll
54:20begin with uh the live participants here
54:23and we'll just kind of go back and forth
54:24like we have in in the other lectures.
54:27So um I'm going to murder this. Kar Sara
54:31Van would you like to go first?
54:33>> Yes.
54:33>> Yeah. Uh hi Patrick thank you so much
54:35for the presentation. It's quite
54:36interesting. My question is related to
54:38the political reservations in um the
54:41village of Maharashtra that you spoke
54:43about like whether the the socioeconomic
54:46uh status of Pradhan influence the
54:49governance mechanism because since we
54:50work with village punays what happens is
54:53that these pradans they don't get
54:55salaries or like a proper hondorium. So
54:58my question is like on that context
55:01where does the socioeconomic status of
55:04Pradan makes sense or does it uh you
55:08know bring uh changes in the governance?
55:11>> So it's a good question absolutely and
55:13and especially in that context. So um
55:16but let me just kind of because I think
55:18it's a it's sort of a little bit uh a
55:21very specific question that's not
55:22related to the overall theme of the
55:23lecture. I wanted to give that as an
55:24example but I think there is something
55:26related to the theme of lecture in your
55:27question. So I want to kind of touch on
55:29that. So in that context one of the
55:32first things we thought was look the
55:34type of person that's getting the office
55:36now is going to be different than the
55:37type before and that's going to be
55:39related to social economic status
55:41education maybe landholding maybe uh
55:45wealth in other ways you know
55:47we found and that was the first place we
55:49looked. So you know when we saw this
55:51pattern of reservations acting in a
55:53certain way we found that to be
55:57unimportant. So we we we had pretty good
55:59measures of it and it just wasn't moving
56:01the variables at all. So so that was one
56:04of the things we conjectured. So in a
56:05sense that's a theory that we explored
56:07in a very preliminary way with the data
56:09and ruled out pretty quickly. And also
56:12when we once we had developed our theory
56:15it generated this non-monotonicity
56:16prediction. And then we went back and
56:18thought well look maybe we just didn't
56:21measure those observables about predan
56:23type very well if we had measured them
56:25maybe there's mismeasurement there maybe
56:28that is still important could that
56:30generate the nomotenicity so we thought
56:32through that theory we resurrected it
56:33again thought could it explain our data
56:35and it didn't generate it so it actually
56:38failed in terms of the empirical
56:40observables and then once we had the the
56:42kind of interesting extra results from
56:43the other theory which were confirmed in
56:45the data we went back to that theory and
56:47it also didn't make of it. So, so it
56:49comes up a lot when you know and this is
56:50one of those things when we gave the
56:51talk people like oh no they had your
56:54intuition that this might be an
56:55important thing and we could rule it out
56:57for those two reasons. One is it didn't
56:59fit in in our raw data it just wasn't
57:02doing anything but two even it would be
57:04very hard to rationalize our empirical
57:06finding with that story. So that that
57:09seemed to me that was kind of very much
57:11part of the modeling process is to
57:13consider theories like that one and
57:14others in in context. Yes.
57:17Good question.
57:18>> Thank you.
57:20>> Perfect. Um, there have been a few
57:23questions in in the chat kind of I think
57:27going back to what you were saying more
57:28about the beginning around kind of where
57:31you getting your data from and what the
57:32role of theory is. And I I think many
57:35researchers um are told that you can
57:38either you know start with a big theor
57:39theoretical question and then seek
57:41evidence and others are told to begin
57:43with that empirical pattern or work
57:45backwards towards a mechanism and so in
57:48your own work maybe how do you move
57:50between theory and data and what
57:52distinguishes a productive interaction
57:55>> yeah so uh okay so I think it's very
57:58risky to start with I had this theory
58:02about the universe
58:04and I'm going to formalize it and uh
58:08then it's going to generate some
58:09predictions and it's going to take me
58:11like a year or two to formalize it and
58:13then uh I'm going to look for data.
58:16Maybe there'll be some data that could
58:17speak to it and uh and then and
58:21hopefully the data fits. I mean that's
58:23just like a fantastically
58:25scary undertaking I think for a PhD
58:28student, right? Like, so once you're me
58:30and I got tenure, then I actually do do
58:32things like that now, right? I mean, I
58:34it wouldn't be my whole job because if I
58:36if that was the only thing I did, I'd
58:37feel embarrassed cuz for 10 years I
58:39wouldn't publish, right? And then maybe
58:40I publish something. So, but I do I have
58:43these things going on in the background.
58:44So, I wrote a paper recently on
58:45consciousness, right? Which is like
58:47totally wildly not economic and it was a
58:50form of that. It's a pure theory piece.
58:52But I find that as if an applied
58:54theorist, I mean, there are people who
58:56are doing theory and they're making the
58:57tools that we think that's fine. And I'm
58:58not talking about that for the applied
59:00theorist who's trying to explain
59:00something in the world, I think it's a
59:02very risky way to start. So, so that
59:04kind of stylized conception that we come
59:06up with a a theory and it's kind of
59:08sealed and then we we we we look at we
59:12fully, you know, instantiate that theory
59:14in a model and then we look for data to
59:16test it. That's that's not really how
59:19this process can work. I just don't
59:21think it's feasible. What I think is
59:22feasible is
59:24the world now throws up a lot of data in
59:28a way that it didn't when I was in grad
59:29school. So now you can just create data
59:31sets so quickly and you can access data
59:33sets so quickly and most of that data is
59:36in the social realm. So it's really
59:38begging for theoretical economic
59:40explanation. So that's the way it
59:43usually works. You come up with data,
59:44you find these interesting things.
59:47Sometimes it can really establish great
59:49causal relations. So it fits all the
59:51kind of tests of um identification. Then
59:54you think about what what does it mean?
59:57How do I explain it? And if if that
1:00:00process is rich, what I mean is that
1:00:03story which will rationalize it will
1:00:05give you more things to look at in the
1:00:06data and then you have a great project.
1:00:08Because I'm always thinking when I'm
1:00:11doing a project, how am I going to stand
1:00:13up in front of a room of skeptical
1:00:14economists who each have their own ideas
1:00:16which are not the same as mine and
1:00:18convince them that this is something
1:00:20that's worth an hour and a half of their
1:00:22time. And it's like that here's this new
1:00:24set of facts that you didn't know about
1:00:26before. That's not enough. Cuz most
1:00:28people say, "Yeah, okay, big deal. What
1:00:29does it mean?" I tell them what it means
1:00:31and I defend it. So I say, "Look, you
1:00:34know, here's a way of making sense of
1:00:35it." And then they're like most
1:00:37economists there's 30 people in the
1:00:39room. There'll be 40 different opinions,
1:00:41right? No, no, this is not it. It's
1:00:42something else. It's this thing. God,
1:00:44how can you claim that? So people just
1:00:46come up and then you say, "Okay, here's
1:00:48my story." And it generates these seven
1:00:51other predictions. And your stories
1:00:53don't do that. I know that. I thought
1:00:54about some of them. They'd be very
1:00:56surprised if they did. And then they
1:00:58fit. Then you then you've taken somebody
1:01:00from skepticism to like, "Oh, yeah, so
1:01:04what?" to like, oh, that's kind of cool.
1:01:06I think it's just a human thing. We just
1:01:08like explanation. We like a story or a
1:01:10parable or myth that makes makes sense
1:01:14of the observables. And if I can say,
1:01:16look, this one seems to fit, then that
1:01:18that's just interesting to people. So, I
1:01:21think that that process is a better
1:01:22process for a PhD student is to take the
1:01:24the data first, enrich it with a story
1:01:27and a theory, and then do more data
1:01:30based on that enriching. That's that's a
1:01:32more feasible one. So I know it's not
1:01:33pure science cuz I think we have this
1:01:35pure science view. We we we come up with
1:01:37theories in the abstract like you know
1:01:40and then we just go and test them in the
1:01:41empirical world. I don't think that's
1:01:42how it is. It's a back and forth.
1:01:47>> Thank you. I think I think that really
1:01:49hits home to some of the the root of
1:01:51some of these questions.
1:01:53Assuming then that we do start with uh
1:01:56the data first and we're we're going
1:01:59from there. When several models appear
1:02:02that are consistent with the data, what
1:02:04do you use to decide amongst them? Do
1:02:07you prioritize predictive performance,
1:02:10parsimony, externability like h how do
1:02:12you how do you decide which out of all
1:02:13the models that come from your data to
1:02:16go with?
1:02:17>> Yeah. So sometimes, you know, I think
1:02:19it's a combination of things. So the
1:02:21reality is, you know, we have this view
1:02:22of like these Okams rays are the most
1:02:24parsimonious one, but sometimes people
1:02:26just find the parimonious explanation
1:02:28not to be real. So you know if they're
1:02:30really equivalent on observables at
1:02:32least you know there may be all models
1:02:35are going to have some differential
1:02:36predictions but in might be that those
1:02:38are in in realms that are not
1:02:39observable. So if they're equivalent on
1:02:42the observables
1:02:43then I think the right thing to say is
1:02:45look I have a class of models that vary
1:02:48on some dimensions that's why they're a
1:02:50class and they're going to differ on
1:02:51some of these but those differences
1:02:53are in dimensions that we just don't
1:02:56have the data to collect. I can I could
1:02:58tell you the experiments that you would
1:02:59need to run but you know it's not
1:03:01feasible for me as a student to run them
1:03:02but so I don't I don't wed myself to one
1:03:06of those particular ones because I don't
1:03:08have a good reason to do it at that
1:03:09point and I don't need to. If I say look
1:03:11it's a class of models and there may be
1:03:13variation on these at the fringes of
1:03:15those models that would be generating
1:03:17other predictions but I can't do them
1:03:18then I just say I commit to that class
1:03:20of models that class of models is
1:03:21generating a set of predictions that are
1:03:24in the data and that's what I'm going to
1:03:26tell you about. Now when I'm giving the
1:03:28narrative you know when you're giving a
1:03:30narrative on a model it's very hard to
1:03:31do a class of models because they're all
1:03:33a little bit different. Then I pick the
1:03:34one that I think isn't necessarily the
1:03:36parimmonious one. I usually pick the one
1:03:38that I think is the most compelling, the
1:03:40most compelling story. And and again,
1:03:42that's important, you know, like most
1:03:44people, I mean, we learn to not do this,
1:03:46but most people, at least when they
1:03:48approach a newspaper, they approach it
1:03:49like they do reading a newspaper
1:03:51article. Newspaper articles doesn't give
1:03:52you, it gives you a little bit of stats,
1:03:54but it's often with a narrative. So
1:03:55those compelling narratives, I think,
1:03:57are very important. But the but I think
1:03:59that's part of the pitch of a paper and
1:04:02a talk. But you know, in the end, if you
1:04:04have a class of models that are
1:04:06indistinguishable, you have to say, "No,
1:04:08no, all of those are on the table at
1:04:10this point. This is all I can do with
1:04:11the data. I can just tell you it's one
1:04:13of these or maybe all of these or some
1:04:15combination of these." But I So that's
1:04:17the right thing to do. But in terms of
1:04:18the pitch, you want to pitch the one
1:04:20that's the clear, cleanest story.
1:04:23Does that help?
1:04:25>> I can answer for them and say yes. Uh I
1:04:29hope so.
1:04:31If if not then they can say no. Um
1:04:35>> yeah uh whoever submitted that one if
1:04:37you have any follow-ups please send
1:04:39through and in the Q&A. In the meantime
1:04:41>> let me ask there's a there's Aman Ray
1:04:44has a has his hand up and I want to ask
1:04:46him a question because he's one of those
1:04:47very good audience members that smiles
1:04:49and laughs a lot. So um I I want to
1:04:51reward that with a [laughter] with
1:04:53asking is a lot of value added as a as a
1:04:56lecturer to seeing a sort of an animated
1:04:58uh so
1:05:00>> Andre you've been called upon.
1:05:02>> Yeah. Okay. Uh my question was actually
1:05:06so sometimes with these models what
1:05:08happens is some of the assumptions they
1:05:11may not seem that realistic. For
1:05:14example, the hoteling downs model for
1:05:17like the one with the median water
1:05:19theorem, it assumes full voter turnout,
1:05:22right? But empirically that's not well
1:05:24observed. I mean empirically we don't
1:05:25observe that at all. And yeah, so when
1:05:29you relax these assumptions, maybe the
1:05:32original model predictions sometimes
1:05:34they don't hold up anymore. So what do
1:05:37you think of the canonical model in that
1:05:39cases like that? like when it's not
1:05:41robust to its assumptions being relaxed.
1:05:45>> So sometimes so I think it's important
1:05:47so I think you should you have to think
1:05:48about it. So if that critical assumption
1:05:51is changing something that's important
1:05:53for explaining the moments in your data
1:05:55then then then it's important that you
1:05:57actually get that assumption right but
1:05:58often it's not right. So often you're
1:06:01going to get this convergence result you
1:06:03know independently of whether there's
1:06:05full turnout or not. So you have to kind
1:06:07of decide whether that's important in
1:06:08your context but it could be you know
1:06:11when I think about it um so for instance
1:06:14you know Australia where I'm from you
1:06:17know I used to be from so I'm still
1:06:18anyway I'm still from there but I don't
1:06:20live there. um it has compulsory voting.
1:06:25So that element I think so my I guess
1:06:29someone else has said it but it's always
1:06:30occurred to me in Australia you have
1:06:33essentially parties that converge in the
1:06:35middle. You don't have any polarization
1:06:38because everyone votes. So they're all
1:06:40competing for the the median voter and
1:06:43it's because that assumption holds by
1:06:46institutional construction. If you don't
1:06:48vote in Australia you get fined a lot.
1:06:49it's compulsory.
1:06:51So now what we're seeing, you know, say
1:06:54in the US is you're seeing this
1:06:55polarization. Well, they're competing
1:06:57for those fringe voters. So that's
1:06:59clearly a violation of the assumption
1:07:01that everyone must vote. But that that
1:07:03that is exactly what's generating, I
1:07:05think, that polarization because now
1:07:07they're not competing over the middle.
1:07:08There's more to gain from activating the
1:07:10base on either side. So that's that's
1:07:12exactly the context. So if I was trying
1:07:14to explain Australia, it's not I mean
1:07:16some people don't vote. That assumption
1:07:18holds. and that generates a result. If
1:07:20I'm trying to explain, I can still use a
1:07:22lot of the insights of the hoteling
1:07:23model in the US context by adding the
1:07:28restriction by releasing the restriction
1:07:30that people are forced to vote. So, I
1:07:31think it just depends on the context and
1:07:34and you just have to get some things
1:07:36right because they're going to matter
1:07:37for the context. Other things you'll get
1:07:39wrong, but they don't matter for the
1:07:41context. So, you just go with the
1:07:43assumption because it's simple and
1:07:44parimonious. So you know the bottom line
1:07:46is if it's a big moving part in your
1:07:48model and it's being generated by an
1:07:49unrealistic assumption that's a bad
1:07:51model.
1:07:55>> Yeah. Thank you.
1:07:58>> Perfect. Um we have a couple more hands
1:08:00here. So maybe I'll do one more. Sanjo,
1:08:02do you want to ask your question
1:08:03briefly?
1:08:07>> Hey Patrick. Um as a PhD student I I
1:08:11just want to say like your lecture was
1:08:12absolutely brilliant. uh firstly and I'm
1:08:16like sort of in that applied modeling
1:08:18process at step two. So I've like
1:08:20collected data. I have some predictions,
1:08:24correlation, causation, whatever you may
1:08:25call it. And now I'm trying to build
1:08:27this model. And I think uh this is
1:08:30following up on Caleb's question about
1:08:33competing canonical models. But the
1:08:36struggle that I'm facing is that there's
1:08:39no very obvious canonical model in my in
1:08:44my case. So for instance, your model is
1:08:46a politician identity to citizen or like
1:08:48governance outcomes relationship model
1:08:52but I'm looking at a bureaucrat identity
1:08:54and um governance
1:08:58um so I see a bureaucrat as like a
1:09:00mechanism from going from the politician
1:09:04to the bureaucrat to to governance. So
1:09:08in that case, if your if your models to
1:09:11choose from are too different from each
1:09:12other, like they're not a class. They're
1:09:16quite different from each other. Like
1:09:18how would you navigate that?
1:09:20>> So, so just to be clear, so firstly,
1:09:22thanks for saying that. But just to be
1:09:25clear, um you have two different models
1:09:29that will generate the same
1:09:32essentially predictions in your data. Is
1:09:34is that that problem? and you're trying
1:09:35to distinguish between which one to use
1:09:38or is it that you're trying to take
1:09:40different components
1:09:42from two canonical models and decide how
1:09:45to put them together to generate your
1:09:49your explanation. I think it's it's the
1:09:51latter uh of of having one is an
1:09:54asymmetric information model which gives
1:09:56you certain kinds of predictions and
1:09:58another are these like identity based
1:10:00models which give you a certain kind of
1:10:03prediction and you're trying to you feel
1:10:05from your empirical predictions that
1:10:06these moving parts matter
1:10:09>> u but how do you how do you bring that
1:10:12together and like sort of like construct
1:10:14it? Yeah.
1:10:15>> Yeah. So those two, if I understand
1:10:17them, those two um components are not at
1:10:20odds. They could be built into one model
1:10:22and they could have both moving parts.
1:10:25And that's actually better, right?
1:10:27Because then you're contributing by, you
1:10:30know, amalgamating two freestanding
1:10:32models into one framework and generating
1:10:35predictions that are coming from
1:10:37different components of that framework.
1:10:39So you need them both to explain your
1:10:41data. I think that's a a good thing.
1:10:43That's a kind of a a richness. So it's
1:10:45saying look I can't and this is a good
1:10:47thing to do like especially you know
1:10:49we've all been in talks where somebody
1:10:51says now I'm going to give you the model
1:10:52and you just see you know pages of math
1:10:54and you're like oh my god you know so
1:10:57here's a motivation for people to do the
1:10:59work because you go look you you have
1:11:00this one model in mind that doesn't do
1:11:02the work and this other model which you
1:11:04also have doesn't but when I put them
1:11:05together and I have to do a bit of work
1:11:06to do that but when I put them together
1:11:08it's going to give us so much more
1:11:10richness
1:11:11so it motivates the modeling the model
1:11:14building process because neither one can
1:11:16is sufficient to to explain the data
1:11:18that you have. So to me, that's a good
1:11:20thing. It's not a I don't see an
1:11:21inherent tension in that. Where there's
1:11:23a tension is when they're they're
1:11:25actually at odds and you can't build the
1:11:26model because they actually have two
1:11:27very different moving parts. Then you
1:11:29kind of have to go with one or the other
1:11:30and then the data should tell you which
1:11:31one's doing better. Um so yeah, so I
1:11:35think that's the I think it's okay. I
1:11:37think that's not a huge problem. I think
1:11:38it's good. I think you should explore
1:11:40that richer model because you need it
1:11:41for the data. Um,
1:11:45>> thank you.
1:11:48>> Thanks, Sanjanna. I I think that uh I've
1:11:50certainly also seen papers where
1:11:52complementing models built on top of
1:11:55each other do provide more of a
1:11:57richness. Um, but there is another side
1:12:01and someone else has asked about this
1:12:03whether we we were talking about
1:12:05layering models. Now
1:12:08Patrick, do you think that having an
1:12:09explicit model always improves a
1:12:12political economy paper or are there
1:12:14cases where a plain English explanation
1:12:16is actually superior? Maybe you can draw
1:12:19on experiences both, you know, as a
1:12:21model and a consumer of research.
1:12:23>> Yeah. Yeah. No, great great um question.
1:12:26So uh if you can do it all in English,
1:12:30do it in English. So, I just don't think
1:12:34there's any point to putting in math
1:12:36something that just repeats and
1:12:38regurgitates the basic intuition that
1:12:41you already have in English. So, so I
1:12:43think a lot of papers just are fine to
1:12:47be just done as essentially it's it's an
1:12:50empirical result and here's the you
1:12:53already have implicitly in your mind
1:12:55probably what's going on and I'm going
1:12:57to tell you in English and that's fine.
1:12:58you don't even need the um the the
1:13:01theoretical model. So I think that's
1:13:03fine. Yeah. So often it is just the
1:13:05case. So I think when you're going to
1:13:07drag somebody through your formalization
1:13:10and a bunch of results,
1:13:13it has to be doing some work. So So by
1:13:16work I mean the distance between the
1:13:19assumptions and the results. I mean I
1:13:21used to have a uh a uh he was actually
1:13:26one of my supervisors Mckesh Ezarn and
1:13:29he used to say there has to be a
1:13:30gentlemanly distance between assumption
1:13:32and results. It means it just can't be
1:13:35that you know you're essentially just
1:13:36regurgitating your assumptions in your
1:13:38results. If if the model isn't
1:13:42generating
1:13:44insight. So what it means is when you
1:13:46throw these assumptions together, how
1:13:48come this result how come you get this
1:13:50result if that's if that's completely
1:13:52obvious you don't need to build the
1:13:53model probably you can just tell the
1:13:54story. If it's not obvious then you know
1:13:57they need to go through well here's
1:13:58here's the theorem here's here's how it
1:14:00holds together here are the conditions
1:14:03under which the theorem holds and that
1:14:04wasn't obvious and here's the extra
1:14:06prediction that wasn't obvious. All that
1:14:07stuff has to be non-obvious.
1:14:09I mean that's why you're dragging people
1:14:11through reading all your math. I mean we
1:14:13don't very few of us think in math right
1:14:15we mostly think in just narrative so
1:14:18whatever language it is English is the
1:14:20one we tend to write in so we just tell
1:14:21it in English I don't think you need a
1:14:23model for a lot of a lot of applications
1:14:28>> thank you um
1:14:32we'll go now to back to our kind of live
1:14:34panelists here Gagas would you like to
1:14:44You're muted mute there.
1:14:48>> All right. Thank you. Can you hear me
1:14:49now?
1:14:50>> Yes.
1:14:51>> Okay. Thank you so much for the
1:14:53beautiful presentation. I have a couple
1:14:55of questions anyway, but let me just
1:14:56push one for now. Um so you you talked
1:15:00about um the possibility of having um an
1:15:04amazing model so to say in a class of
1:15:07competing models right that could also
1:15:10pro provide you with the best result.
1:15:12But my question is what then happens in
1:15:14an instance whereby you pick a model
1:15:16which as far as you are concerned is the
1:15:18best model in that class of rival
1:15:20alternatives and yet you are having a
1:15:22challenge of fitting it into reality or
1:15:26reality into it. Okay, especially in the
1:15:30context of political economy where the
1:15:32data itself could be political. It is
1:15:34only as good as the political class
1:15:36tells you it is. I'm saying this um with
1:15:40a particular emphasis on where I'm
1:15:41coming from Africa for example that data
1:15:44may have integrity issues. So I'm asking
1:15:47what happens where your model is as far
1:15:50as you are concerned good the best
1:15:52perhaps but then there are issues with
1:15:54fitting it into reality or fitting
1:15:56reality into it and then how do I avoid
1:15:58the tragedy of forcing your model to
1:16:01confess the reality it was not built to
1:16:04to you know mirror and all of that. I'll
1:16:07be glad if you can assist. Thank you.
1:16:10>> Okay, good question. So, so that that um
1:16:13world in which I have my model that I
1:16:17think is a beautiful model and that I
1:16:19think fits and explains the world,
1:16:22but reality is getting in the way and
1:16:26saying, "Yeah, that model's not right."
1:16:27That's that's a world that I've lived in
1:16:29a lot. Um where I built complicated
1:16:32models and thought, "Yeah, these are
1:16:33this is such a cool model. So subtle.
1:16:37But that's useless. So you know if if so
1:16:41one part of it one way to think of it is
1:16:43if really reality just keeps saying no
1:16:46that just fails that just model isn't
1:16:48working then you just have to throw that
1:16:50model out even though it might be the
1:16:51most beautiful thing because I mean
1:16:52ultimately we're not artists right we're
1:16:55we're trying to explain the world
1:16:57but so so if at that level of your
1:17:00question I think yeah it's awful the you
1:17:04know our models don't always work and I
1:17:07wish the world was better but it's not
1:17:08so you We're not mathematicians. When we
1:17:10if we were, we'd be terrible
1:17:11mathematicians. So, we're we're trying
1:17:13to explain the world and if the world
1:17:15keeps saying this isn't fitting, you
1:17:16just got to throw it out. You got to
1:17:18live with it. But I think you said
1:17:20something else too, which is kind of a
1:17:21like a I think another issue, which is
1:17:23that if you really doubt the the
1:17:26veracity of the data, and in some
1:17:28context you are going to doubt it, then
1:17:30maybe that's not the right metric you
1:17:32should be using for your model. So, you
1:17:34might say, look, you know, my model's
1:17:36not fitting. Um, I think it's what's
1:17:39going on and I have these other maybe
1:17:41stylized facts that make me think this
1:17:43is actually really true and important.
1:17:44The data is is rejecting it, but I have
1:17:47reasons to believe [snorts] that that
1:17:49data is incorrect. So, so then you're
1:17:53questioning the data, right? You're
1:17:55questioning then what are you
1:17:56questioning then? You're you're
1:17:57questioning our our empirical picture,
1:17:59our picture of, you know, the the
1:18:01reality. Now, that's a different issue.
1:18:04Like if if you have deep reasons and
1:18:06good reasons to to doubt the the
1:18:09veracity of the picture you're getting
1:18:10from the data then you shouldn't be
1:18:12trying to match that data. So so I think
1:18:15I think but but what would what I would
1:18:17resist you know exant is just building a
1:18:21model it doesn't fit the data and then
1:18:24saying oh that's because you know now I
1:18:25want to say the data is bad. So you have
1:18:27to be careful right? So because you can
1:18:28always just say look I'm not going to
1:18:30you know that thing that I'm seeing
1:18:32outside the window that didn't happen
1:18:34right and you know so you have to resist
1:18:36that. So but yeah um some data is very
1:18:40reliable right so that's the ones and
1:18:42some some isn't and then you know you're
1:18:44right you got a problem and you know
1:18:45this is we don't usually go there in
1:18:47economics too much right so we usually
1:18:49but you know there's a there's a kind of
1:18:50a more profound critique which is like
1:18:52the post structural postmodern critiques
1:18:55which is that all data is generated all
1:18:57data is you know they put a lot of they
1:18:59think we're very naive in the way that
1:19:01we view the world and and partly that
1:19:03there's a truth to that like a lot of
1:19:04this is a postcolonial literature
1:19:06saying, you know, the data was created
1:19:08by the victors and tells a narrative and
1:19:11that narrative is one that's in the
1:19:13favor of a particular class of people or
1:19:15something like that. So, we don't
1:19:17usually do those post post type uh
1:19:20critiques in economics, but I think
1:19:22there's a there's something to that and
1:19:24in some context there really is
1:19:25something to that. So, you know,
1:19:26sometimes the data can be dodgy. Yeah.
1:19:30>> Thank you, sir.
1:19:32>> You're welcome.
1:19:33>> Thank you, Carass. Um, we'll go back now
1:19:36to the chat and there there are a couple
1:19:38questions about counterfactuals.
1:19:41Um, I'll read one. It says, "In
1:19:43structural political economy, how do you
1:19:45prevent overspecifying candidate
1:19:47mechanisms when structural
1:19:48counterfactuals heavily depend on
1:19:50institutional primitives that remain
1:19:52unobservable or underidentified in
1:19:54baseline empirical facts?"
1:19:56>> Yeah, that's a good question. Yeah. So,
1:19:58um, so, so I think there's two things
1:20:00here. So one is um
1:20:04yeah when they're unobservable you know
1:20:07we're really in the realm of you know
1:20:08you have to try to convince people that
1:20:10you know when when I'm doing this
1:20:11counterfactual that those primitives are
1:20:13not going to change and you know how do
1:20:15you do that um
1:20:18various ways I mean you could maybe
1:20:20point to other contexts where they were
1:20:22relatively invariant across such a
1:20:24change that would be one way to do it
1:20:25but you know it's not always going to be
1:20:26available so there may be ways of
1:20:29convincing people by pointing to
1:20:30analogist situations
1:20:33often there isn't and then it's just
1:20:34going to come down to people you know I
1:20:37you see it and people are often too
1:20:38polite to say in the seminar but they're
1:20:40like uh you know come on like you know
1:20:42under that counterfactual everything
1:20:44everything would change it's like you
1:20:46can't hold x fixed you know whatever it
1:20:49is whatever the parameters fixed theta
1:20:52so you know that it's just it's just
1:20:54it's there's no kind of clear formula
1:20:56for that you know it's in a sense we're
1:20:57trying to it's we're running narratives
1:20:59and we're trying to convince convince
1:21:00others that this is a sensible thing to
1:21:01do. [sighs]
1:21:03Um the question of overfitting I think
1:21:06is a slightly different one in the sense
1:21:08that you know uh when you're I mean I
1:21:12think this is one of the advantages of
1:21:13of doing structural work is when you
1:21:16have a a a structural framework you
1:21:20actually have your model you have your
1:21:23set of parameters you have your
1:21:24estimates and you fit the data however
1:21:28you fit it. Now you know I I know I can
1:21:32always write down a model with more
1:21:33parameters, more moving parts that will
1:21:35fit better. I know I can do that. Now
1:21:40how do you compete? How do you test
1:21:42between those models? So there are tests
1:21:43that actually do penalize overfitting.
1:21:46So we can formally do that. So that's
1:21:49that's uh
1:21:51one way one way that we do it. Now it
1:21:54depends on your context, right?
1:21:55sometimes just fitting is what you want
1:21:57to do and then you don't even care about
1:21:59the economic narrative right so if
1:22:01you're you know LLM are essentially
1:22:03doing that they have you know a
1:22:04gazillion which is where gazillion is a
1:22:06large number of parameters that they
1:22:08estimate on a on a on a on a very big
1:22:10data set and it allows them to predict
1:22:13now do we understand no we don't
1:22:15understand how it's doing it I it's
1:22:17totally a mystery how they can provide
1:22:19such coherent answers to things but it
1:22:21is just a big structural estimation
1:22:24problem on a lot of data and that in
1:22:27that context that's the right thing to
1:22:29do right in economics context we don't
1:22:31usually want that what we want is to
1:22:34provide a narrative an explanation for
1:22:36the observables and generally we want to
1:22:40not have a blackbox explanation so so
1:22:43but but in the set of explanations that
1:22:45are not blackbox we can have two that
1:22:47would rationalize you know not
1:22:49rationalize in in a structural setting
1:22:50we're matching we're fitting the data
1:22:52the one with more parameters will often
1:22:53fit better, but you know, it's coming at
1:22:55the cost of more parameters and we have
1:22:56tests to to trade these off. So, I think
1:22:58it's a combination of this story is a
1:23:00good story. It fits the data pretty
1:23:03well. It explains things pretty well. It
1:23:04doesn't seem like it's overfitted. You
1:23:06know, it's kind of more of a handwavy
1:23:08type thing, you know. Um, yeah,
1:23:12elements of science, elements of art
1:23:14there,
1:23:15>> art and science together.
1:23:17>> Um,
1:23:18>> Zoo, uh, over to you. Do you want to ask
1:23:21your question?
1:23:22>> Yeah. Uh can you hear me? Um thank you
1:23:25uh Professor Francois. Uh kind of a big
1:23:27picture question. Uh where do you see
1:23:29the role of applied political economy
1:23:32modeling in development and political
1:23:34economy in the future given that there
1:23:36was a trend in development economics
1:23:38credibility revolution that there's a
1:23:40lot of kind of replete with empirical
1:23:42facts and a lot of improvement in
1:23:45identification of those empirical facts.
1:23:46Do you see the same trend that um we're
1:23:50moving towards you know identifying
1:23:52mechanisms to applied modeling and uh
1:23:55second part of that is what advice
1:23:57particular advice would you give to kind
1:24:00of young researchers who try to navigate
1:24:02that landscape?
1:24:04>> Yeah, it's tough. So, so I think I I you
1:24:07know I think so I I primarily see myself
1:24:10as a development economist. That's what
1:24:12I'm interested in. That's what I've
1:24:14always liked. And I moved into political
1:24:16economy um because I felt like the
1:24:19questions that I wanted to understand
1:24:21weren't really being tackled in in
1:24:23development. And partly it's because um
1:24:26it became so a theoretical. I was always
1:24:29interested in mechanisms like why is
1:24:30there so much variation in development?
1:24:32Why do we see these big differences? And
1:24:34the questions that they're very hard to
1:24:37to answer because it's very hard to
1:24:39identify a particular mechanism. So you
1:24:42never fully identify things very well
1:24:44unless you can actually control the
1:24:45variation. So I think development got
1:24:48very obsessed with and in a good way,
1:24:50right? It got obsessed with trying to
1:24:53say things that were true, trying to
1:24:55identify causal relationships. That was
1:24:58a good thing because again, when I was
1:25:00in grad school, people were just
1:25:02regressing indogenous things on each
1:25:03other all the time and it was very
1:25:05sloppy. So it needed it, right? Um,
1:25:09but I feel like to me that's a kind of
1:25:11a,
1:25:13you know, I mean, I think it's it's not
1:25:15unreasonable for people to say, why
1:25:17isn't anyone doing development anymore?
1:25:19Like, why isn't anyone asking the big
1:25:20questions? Why is everyone just finding
1:25:22out very specific bits of information
1:25:24about things that will work in
1:25:26particular contexts? And and I think
1:25:29that, yeah, that's a reasonable
1:25:30question. So, I I I don't I'm not very
1:25:32good at predicting the future. I I don't
1:25:34think political economy is going to go
1:25:36there. I think the best to me the best
1:25:39case scenario is people are asking
1:25:41questions that are big thinking about
1:25:44with theory usually because it's just
1:25:46too hard to do without it. Thinking
1:25:48about like the the things I should be
1:25:49able to run in my experiments that will
1:25:51help me identify those bigger questions.
1:25:54So so it's a combination of both. I
1:25:56don't I see and so to answer your
1:25:58question I think for a student I think
1:26:01we teach people how to solve complicated
1:26:04models very well in grad school we do
1:26:06complicate and we actually select for
1:26:08math mathiness in that sense but I don't
1:26:11think we teach people very well how to
1:26:13build models and I feel like there's a
1:26:17lot of value added to doing a little bit
1:26:19of that model building in the empirical
1:26:22project. So as a grad student, I would
1:26:24invest a bit of time in that which is in
1:26:26the political economy realm. To me, it's
1:26:27understanding those canonical models,
1:26:29having them in the back of my mind.
1:26:30That's the advice I would that's the
1:26:32advice I do give and I I would give and
1:26:34I teach this course. I teach like a
1:26:3612-week course on how to build models in
1:26:38political economy. That's what I do
1:26:39because I think it's a useful thing.
1:26:40It's a complement to your core and it's
1:26:43still your core business which is the
1:26:44empirical one.
1:26:48>> Thank you very much, professor.
1:26:50>> Yeah, no worries. Thank you. It is
1:26:52pretty incredible um just everyone
1:26:54that's here how condensed
1:26:58uh the professors are making what is
1:27:02normally maybe a semester long class
1:27:04into an hour hour and a half uh every
1:27:08week. And so um we we recognize there is
1:27:11a lot of of information but it's
1:27:15something that you'll be able to
1:27:16reference back to and and build on. Um,
1:27:20I think that this is really the that's
1:27:22kind of the zoomed out perfect place to
1:27:24to end the lecture. And I just really
1:27:27want to thank Professor Vansis for your
1:27:30time and um your your insights.
1:27:34Everyone, I've I've put in just a couple
1:27:37of things into the chat um for links
1:27:40that that you should reference, but but
1:27:42really thank you for coming. Uh we we
1:27:45look forward to to you joining us next
1:27:48week with Professor Tim Bezley u to talk
1:27:52about state capacity and we uh look
1:27:55forward to seeing you then.
1:27:56>> Thank thank you too for um letting me do
1:27:59this and I I enjoyed it a lot. I think
1:28:01the questions were great and I think
1:28:03this is a really good initiative too
1:28:04just to get a a view of how all of us
1:28:07approach this a little bit differently
1:28:09and you should definitely go to Tim's
1:28:10talk. Tim is one of the great applied
1:28:13modelers, right? So, he is does exactly
1:28:16this, I think. Um, and he also, you
1:28:18know, he's just great. So, you should
1:28:20come to it and enjoy it. You'll learn a
1:28:22lot.
1:28:23>> Well, we have a personal endorsement
1:28:25here, guys. So, so we'll very much uh
1:28:28hope to see you next week. Thank you
1:28:29again, Patrick, and hope everyone has a
1:28:31great rest of your days.
1:28:32>> Thank and thank you, Caleb. You did a
1:28:34great job.
1:28:35>> Thank you.
1:28:36>> Was excellent lecture. Thank you.
1:28:38>> Thank you.
1:28:38>> Thank you. Thank you.

About this transcript

This page contains the full transcript of BREAD-IGC virtual PhD-level course 2026 | Lecture 3 Core Political Economy Models by International Growth Centre, generated from the public captions YouTube serves with the video. The transcript has 16,242 words across 2,369 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.