1 00:00:00,090 --> 00:00:02,490 The following content is provided under a Creative 2 00:00:02,490 --> 00:00:04,059 Commons license. 3 00:00:04,059 --> 00:00:06,330 Your support will help MIT OpenCourseWare 4 00:00:06,330 --> 00:00:10,720 continue to offer high-quality educational resources for free. 5 00:00:10,720 --> 00:00:13,350 To make a donation or view additional materials 6 00:00:13,350 --> 00:00:17,280 from hundreds of MIT courses, visit MIT OpenCourseWare 7 00:00:17,280 --> 00:00:18,480 at ocw.mit.edu. 8 00:00:30,444 --> 00:00:31,986 AUDIENCE: Was that going to be-- were 9 00:00:31,986 --> 00:00:33,790 we going to do that session after you do this? 10 00:00:33,790 --> 00:00:34,090 ROBERT TOWNSEND: No. 11 00:00:34,090 --> 00:00:34,750 AUDIENCE: [? Or ?] [? will ?] [? we ?] do it before? 12 00:00:34,750 --> 00:00:37,500 ROBERT TOWNSEND: We'll get your questions first. 13 00:00:37,500 --> 00:00:40,950 That way, I can't time how quickly 14 00:00:40,950 --> 00:00:43,870 or how slowly to go through this lecture. 15 00:00:43,870 --> 00:00:47,543 So I'd just as soon start off. 16 00:00:47,543 --> 00:00:48,710 AUDIENCE: I have a question. 17 00:00:48,710 --> 00:00:49,390 ROBERT TOWNSEND: Yes. 18 00:00:49,390 --> 00:00:50,765 AUDIENCE: I wondered if you could 19 00:00:50,765 --> 00:00:53,880 explain in more depth how-- 20 00:00:53,880 --> 00:00:55,740 so, when you were-- 21 00:00:55,740 --> 00:00:57,473 when you're simulating structural models, 22 00:00:57,473 --> 00:00:58,890 and there's an obvious distinction 23 00:00:58,890 --> 00:01:00,598 between structural models in which people 24 00:01:00,598 --> 00:01:02,860 are forward-looking and which people are not. 25 00:01:02,860 --> 00:01:06,120 And it makes it easier if people aren't, but it's not 26 00:01:06,120 --> 00:01:07,240 as realistic. 27 00:01:07,240 --> 00:01:10,355 So I wondered if you can kind of explain like what are 28 00:01:10,355 --> 00:01:14,680 the trade-offs in making people forward-looking and then how 29 00:01:14,680 --> 00:01:16,100 you-- 30 00:01:16,100 --> 00:01:17,950 I think there was like a [? Vieira ?] paper 31 00:01:17,950 --> 00:01:20,860 about how he was able, for the first time, 32 00:01:20,860 --> 00:01:24,568 to like get at the transition and not just steady state. 33 00:01:24,568 --> 00:01:26,550 So I didn't understand [INAUDIBLE].. 34 00:01:26,550 --> 00:01:28,092 ROBERT TOWNSEND: Yeah, OK, so there's 35 00:01:28,092 --> 00:01:32,472 like three or four different things in there. 36 00:01:32,472 --> 00:01:34,800 It's a good question, questions. 37 00:01:37,510 --> 00:01:38,970 First of all, this forward-looking 38 00:01:38,970 --> 00:01:42,090 or not forward-looking, part of it is just tractability. 39 00:01:42,090 --> 00:01:46,530 The first generation of these models, half of them 40 00:01:46,530 --> 00:01:50,850 basically assumed what's called warm-glow bequest motive. 41 00:01:50,850 --> 00:01:55,290 So you just sort of feel warmly about passing your money 42 00:01:55,290 --> 00:01:58,000 on to your kids, but, actually, it's you, 43 00:01:58,000 --> 00:02:01,680 and you are reborn the next period. 44 00:02:01,680 --> 00:02:03,900 We went through LEB, which is this Lloyd-Ellis 45 00:02:03,900 --> 00:02:09,340 and Bernhardt, which is a simple sort of bequest motive. 46 00:02:09,340 --> 00:02:11,920 So you don't take into account the utility that's 47 00:02:11,920 --> 00:02:13,750 generated by the bequest. 48 00:02:13,750 --> 00:02:17,530 You just model it as a reduced form. 49 00:02:17,530 --> 00:02:19,750 It's a bit like overlapping generations, 50 00:02:19,750 --> 00:02:24,670 and that's why overlapping generations models are still 51 00:02:24,670 --> 00:02:25,960 widely used. 52 00:02:25,960 --> 00:02:26,680 We had another. 53 00:02:26,680 --> 00:02:31,060 There was a QJE paper on financial volatility, which 54 00:02:31,060 --> 00:02:34,720 was overlapping generations, another one on China 55 00:02:34,720 --> 00:02:39,390 and savings rates, which was overlapping generations. 56 00:02:39,390 --> 00:02:43,510 It's just far easier to solve a two- or a three-period problem 57 00:02:43,510 --> 00:02:46,600 and splice together these overlapping generations 58 00:02:46,600 --> 00:02:50,620 than it is to have a forward-looking behavior. 59 00:02:50,620 --> 00:02:52,150 So it's a judgement call. 60 00:02:52,150 --> 00:02:58,510 Now Paco, way back to his dissertation, but also then 61 00:02:58,510 --> 00:03:02,200 subsequent papers with Kaboski and so on, 62 00:03:02,200 --> 00:03:04,610 have this forward-looking aspect to it. 63 00:03:04,610 --> 00:03:11,230 So not only do you have to kind of work 64 00:03:11,230 --> 00:03:16,330 backwards on your own decision problem, but, to the extent 65 00:03:16,330 --> 00:03:19,090 that prices and wages are endogenous, 66 00:03:19,090 --> 00:03:22,930 you kind of have to forecast those too. 67 00:03:22,930 --> 00:03:24,820 And, in rational expectations, you 68 00:03:24,820 --> 00:03:29,380 would require that the forecast be rational, that households 69 00:03:29,380 --> 00:03:32,200 are kind of getting it right on average. 70 00:03:32,200 --> 00:03:34,750 Actually, that reminds me to say that one 71 00:03:34,750 --> 00:03:39,010 of those flow of funds papers we reviewed briefly 72 00:03:39,010 --> 00:03:42,540 was Monika Piazzesi and Martin Schneider. 73 00:03:42,540 --> 00:03:46,450 And, although they were using US data, 74 00:03:46,450 --> 00:03:53,520 they were looking at the effect of various things, 75 00:03:53,520 --> 00:03:59,370 including a demographic shift, sort of baby boom generation 76 00:03:59,370 --> 00:04:03,030 having more young people in the population, relatively 77 00:04:03,030 --> 00:04:04,080 speaking, than before. 78 00:04:07,490 --> 00:04:09,950 And they did not have the forward-looking behavior 79 00:04:09,950 --> 00:04:14,150 because they basically assumed that the expectations you 80 00:04:14,150 --> 00:04:18,720 see in the Michigan surveys are accurate. 81 00:04:18,720 --> 00:04:20,433 So-- 82 00:04:20,433 --> 00:04:22,100 AUDIENCE: What were the expectations of? 83 00:04:22,100 --> 00:04:23,100 ROBERT TOWNSEND: Prices. 84 00:04:23,100 --> 00:04:24,225 AUDIENCE: Would they go up? 85 00:04:24,225 --> 00:04:25,820 ROBERT TOWNSEND: Inflation rates. 86 00:04:25,820 --> 00:04:29,030 So it wasn't like solving for an endogenous price in the model 87 00:04:29,030 --> 00:04:30,830 and then having people forecast that. 88 00:04:30,830 --> 00:04:34,250 They just short circuited the whole algorithm 89 00:04:34,250 --> 00:04:37,820 by assuming that people's expectations in the models 90 00:04:37,820 --> 00:04:42,890 were as measured in the data, even if those decisions that 91 00:04:42,890 --> 00:04:45,410 got generated might be inconsistent 92 00:04:45,410 --> 00:04:48,230 with those inflation forecasts. 93 00:04:48,230 --> 00:04:51,620 And it was a reminder to many of us 94 00:04:51,620 --> 00:04:55,400 how straight-jacketed we've become, 95 00:04:55,400 --> 00:04:58,640 in a good and a bad way, with the rational expectations 96 00:04:58,640 --> 00:05:01,070 because, once you have a way of pinning down 97 00:05:01,070 --> 00:05:03,470 the future via these measured expectations, 98 00:05:03,470 --> 00:05:09,200 you can expand the heterogeneity in the model and deal 99 00:05:09,200 --> 00:05:11,810 with a lot more realism. 100 00:05:14,438 --> 00:05:16,105 AUDIENCE: So, in that inflation, how far 101 00:05:16,105 --> 00:05:18,767 did they forecast inflation in the survey? 102 00:05:18,767 --> 00:05:21,100 ROBERT TOWNSEND: I think it was just the following year. 103 00:05:24,362 --> 00:05:28,398 So the last thing you asked about were the transitions. 104 00:05:28,398 --> 00:05:30,830 AUDIENCE: [INAUDIBLE] 105 00:05:30,830 --> 00:05:32,580 ROBERT TOWNSEND: Sorry, I didn't hear you. 106 00:05:32,580 --> 00:05:34,500 AUDIENCE: Oh, I was just telling him they 107 00:05:34,500 --> 00:05:37,692 measured for different ranges of times. 108 00:05:37,692 --> 00:05:40,077 I was like one year ahead and five years ahead. 109 00:05:43,430 --> 00:05:45,140 ROBERT TOWNSEND: They were concerned 110 00:05:45,140 --> 00:05:50,070 with consumer durables and housing purchases and so on. 111 00:05:50,070 --> 00:05:53,970 So you'd want to have expectations 112 00:05:53,970 --> 00:05:55,425 of some of those asset prices. 113 00:05:58,080 --> 00:06:01,120 The transitions are the hardest thing to do. 114 00:06:01,120 --> 00:06:05,710 You can have a model, which is forward-looking, 115 00:06:05,710 --> 00:06:08,590 and solve for the steady state. 116 00:06:08,590 --> 00:06:13,840 It's like taking a little tease off of all the variables. 117 00:06:13,840 --> 00:06:15,580 So, you know, things will settle down 118 00:06:15,580 --> 00:06:20,110 to one interest rate and one wage rate. 119 00:06:20,110 --> 00:06:23,800 That doesn't mean the decision problems are static. 120 00:06:23,800 --> 00:06:25,360 People are still forward-looking, 121 00:06:25,360 --> 00:06:29,350 but they're looking forward at those prices. 122 00:06:29,350 --> 00:06:32,470 And, in a steady state, there's always these offsetting 123 00:06:32,470 --> 00:06:34,630 fractions of the population. 124 00:06:34,630 --> 00:06:36,790 So some fraction become urgent. 125 00:06:36,790 --> 00:06:38,080 The rest become patient. 126 00:06:38,080 --> 00:06:40,270 There's always a fixed fraction. 127 00:06:40,270 --> 00:06:43,660 And there's always a fixed demand and supply of funds. 128 00:06:43,660 --> 00:06:45,910 And so you get these. 129 00:06:45,910 --> 00:06:48,790 And you can almost analytically, sometimes, 130 00:06:48,790 --> 00:06:51,280 get the steady state of some of these models. 131 00:06:51,280 --> 00:06:54,520 Or, even if not, you can compute them 132 00:06:54,520 --> 00:07:00,740 much more readily, whereas the transitions, by definition, 133 00:07:00,740 --> 00:07:04,210 you have a whole path, and you have 134 00:07:04,210 --> 00:07:09,760 a path of guessed interest rates and wages and so on. 135 00:07:09,760 --> 00:07:14,230 And people take that path as given and optimized. 136 00:07:14,230 --> 00:07:16,930 And then you have to look at the excess demand 137 00:07:16,930 --> 00:07:19,540 and supply for labor and for capital 138 00:07:19,540 --> 00:07:22,510 not only in the current period, but all these future periods 139 00:07:22,510 --> 00:07:24,880 for which they've already made a guess, 140 00:07:24,880 --> 00:07:31,510 and then have this algorithm to somehow adjust the guessed path 141 00:07:31,510 --> 00:07:34,940 and hope, somehow, it will converge. 142 00:07:34,940 --> 00:07:37,630 So Buera and Shin did that originally 143 00:07:37,630 --> 00:07:41,450 and have modified the code. 144 00:07:41,450 --> 00:07:45,430 Yan, you've played around with that code. 145 00:07:45,430 --> 00:07:47,170 Other people have as well. 146 00:07:47,170 --> 00:07:51,160 It seems to work, but it-- 147 00:07:51,160 --> 00:07:52,760 if you take it to a new application, 148 00:07:52,760 --> 00:07:55,810 you've got to kick it and push it around a bit 149 00:07:55,810 --> 00:07:58,970 and get scared that it's not going to work at all. 150 00:07:58,970 --> 00:08:03,400 And so it's certainly not a silver bullet, 151 00:08:03,400 --> 00:08:06,460 but it does seem to work for particular applications. 152 00:08:09,020 --> 00:08:10,920 Am I answering all-- 153 00:08:10,920 --> 00:08:11,550 AUDIENCE: Yeah. 154 00:08:11,550 --> 00:08:12,383 ROBERT TOWNSEND: OK. 155 00:08:12,383 --> 00:08:14,600 AUDIENCE: So I actually have a related question. 156 00:08:14,600 --> 00:08:16,610 So I remember when we were talking 157 00:08:16,610 --> 00:08:20,300 about the computation issues in one of the extra sessions. 158 00:08:20,300 --> 00:08:22,620 We were talking about having myopic agents 159 00:08:22,620 --> 00:08:27,110 and saying, well, in the profession, 160 00:08:27,110 --> 00:08:30,170 we're not that interested in myopic agents. 161 00:08:30,170 --> 00:08:35,720 But I feel like there is potentially 162 00:08:35,720 --> 00:08:37,260 some middle ground between saying, 163 00:08:37,260 --> 00:08:38,677 well, I have to have an agent that 164 00:08:38,677 --> 00:08:41,570 looks to the infinite future, or I 165 00:08:41,570 --> 00:08:44,150 have to have an agent that is totally myopic. 166 00:08:44,150 --> 00:08:47,210 And I feel like it's actually not 167 00:08:47,210 --> 00:08:51,560 more sensible to say this agent looks to period infinity 168 00:08:51,560 --> 00:08:53,960 than to say this agent doesn't look forward at all. 169 00:08:53,960 --> 00:08:56,797 So I'm wondering like what way you see like-- 170 00:08:56,797 --> 00:08:58,880 is there a reason I'm not understanding why one is 171 00:08:58,880 --> 00:09:00,150 more ridiculous than the other? 172 00:09:00,150 --> 00:09:00,650 Or like I mean-- 173 00:09:00,650 --> 00:09:02,275 ROBERT TOWNSEND: People take positions. 174 00:09:02,275 --> 00:09:05,000 There's no question you're going to get opinions. 175 00:09:05,000 --> 00:09:11,240 Some people would dismiss the myopia out of hand as outdated. 176 00:09:11,240 --> 00:09:12,740 Other people are convinced there's 177 00:09:12,740 --> 00:09:15,290 a lot of irrationality and behavioral things. 178 00:09:15,290 --> 00:09:19,190 And I do think that taking the expectations as measured 179 00:09:19,190 --> 00:09:26,090 in the data is quite reasonable thing to do. 180 00:09:26,090 --> 00:09:28,520 And some overlapping generations models 181 00:09:28,520 --> 00:09:32,810 are, with realistic lifetimes and mortality and all of that, 182 00:09:32,810 --> 00:09:37,775 are really, in some sense, the most realistic of all. 183 00:09:37,775 --> 00:09:40,150 Granted, you still have to deal with this bequest motive, 184 00:09:40,150 --> 00:09:44,030 but, you know, I'm practical. 185 00:09:44,030 --> 00:09:47,220 I think it depends on the question at hand. 186 00:09:47,220 --> 00:09:51,860 And it's nice to be fully rational and forward-looking 187 00:09:51,860 --> 00:09:55,430 and all of that, but computational, 188 00:09:55,430 --> 00:10:00,900 dimensional issues will limit the model you want to analyze. 189 00:10:00,900 --> 00:10:03,330 And, sometimes, it seems unreasonable. 190 00:10:03,330 --> 00:10:05,070 So I mean, for example, in IO-- 191 00:10:05,070 --> 00:10:08,010 I'm not sure if we'll get to this today-- 192 00:10:08,010 --> 00:10:13,050 the state spaces just expand dramatically when 193 00:10:13,050 --> 00:10:17,560 you start thinking about strategic behavior. 194 00:10:17,560 --> 00:10:19,380 And then you can actually show, if you 195 00:10:19,380 --> 00:10:23,460 wanted to keep track of where all financial providers are 196 00:10:23,460 --> 00:10:26,010 going to locate with a finite number of nodes, 197 00:10:26,010 --> 00:10:29,400 kind of working backwards and having-- 198 00:10:29,400 --> 00:10:35,350 that becomes a very computationally intensive 199 00:10:35,350 --> 00:10:35,890 problem. 200 00:10:35,890 --> 00:10:38,950 And it seems reasonable to say, well, let's just have them 201 00:10:38,950 --> 00:10:43,390 forward-looking for two, three periods or something, 202 00:10:43,390 --> 00:10:46,558 a finite number of periods, as an approximation. 203 00:10:46,558 --> 00:10:48,100 AUDIENCE: I think it's so interesting 204 00:10:48,100 --> 00:10:49,930 because, from outside the profession, 205 00:10:49,930 --> 00:10:51,740 a lot of the criticism we get is, oh, you 206 00:10:51,740 --> 00:10:54,222 assume that these agents are totally forward-looking 207 00:10:54,222 --> 00:10:55,180 and perfectly rational. 208 00:10:55,180 --> 00:10:56,770 You guys should see what happens in your models 209 00:10:56,770 --> 00:10:58,110 if you don't assume that. 210 00:10:58,110 --> 00:10:59,350 But then, within the profession, if you 211 00:10:59,350 --> 00:11:00,388 were trying to assume that, people 212 00:11:00,388 --> 00:11:02,350 would be like this is a waste of my time. 213 00:11:02,350 --> 00:11:03,243 Leave me alone. 214 00:11:03,243 --> 00:11:05,410 ROBERT TOWNSEND: Well, the profession comes and goes 215 00:11:05,410 --> 00:11:09,520 in terms of its rational expectations, 216 00:11:09,520 --> 00:11:15,860 which was kind of invented by Muth in 1960-something I think, 217 00:11:15,860 --> 00:11:17,900 late '60s. 218 00:11:17,900 --> 00:11:21,500 And then it took over macro with a vengeance, 219 00:11:21,500 --> 00:11:26,360 well, in the sense that there were a lot of unhappy campers. 220 00:11:26,360 --> 00:11:31,645 So people had stakes in doing certain-- 221 00:11:31,645 --> 00:11:34,310 one way or the other. 222 00:11:34,310 --> 00:11:36,770 And then you need to start putting back 223 00:11:36,770 --> 00:11:40,400 ingredients in the models to explain rigidities 224 00:11:40,400 --> 00:11:41,990 and slow-moving behavior. 225 00:11:41,990 --> 00:11:44,900 I mean, personally, I'm not so sure sticky prices is 226 00:11:44,900 --> 00:11:52,690 any more compelling to me than sort of some fixed expectation. 227 00:11:52,690 --> 00:11:55,460 So those are-- you'll just have to develop your own judgment 228 00:11:55,460 --> 00:12:00,830 somehow, but there isn't a single right way to look at it. 229 00:12:00,830 --> 00:12:03,150 I think it really does depend on the applications. 230 00:12:03,150 --> 00:12:08,020 And, especially, as I said, if you want to break new ground 231 00:12:08,020 --> 00:12:11,590 and explore a different kind of model, 232 00:12:11,590 --> 00:12:13,720 you may have to make some approximations. 233 00:12:19,010 --> 00:12:19,510 Yes? 234 00:12:19,510 --> 00:12:22,690 AUDIENCE: So I have a more general question, 235 00:12:22,690 --> 00:12:27,040 sort of how to interpret one of the classic models that we saw. 236 00:12:27,040 --> 00:12:31,270 So it made a lot of sense to me like the last part 237 00:12:31,270 --> 00:12:34,410 of the other course where we looked at insurance 238 00:12:34,410 --> 00:12:39,160 and how to fit data with different models. 239 00:12:39,160 --> 00:12:46,320 But there was that one lecture on TFP that I am kind of not 240 00:12:46,320 --> 00:12:48,210 clear exactly how to think about it 241 00:12:48,210 --> 00:12:49,740 in the context of this class. 242 00:12:49,740 --> 00:12:57,660 So, the paper by Zilibotti and other people on China, 243 00:12:57,660 --> 00:12:59,640 how do we-- 244 00:12:59,640 --> 00:13:04,080 how should we think about it in the context of this class? 245 00:13:04,080 --> 00:13:09,000 ROBERT TOWNSEND: Well, the way I think about it is we're 246 00:13:09,000 --> 00:13:13,440 interested in development and development issues. 247 00:13:13,440 --> 00:13:17,710 We're interested in the impact of financial programs. 248 00:13:17,710 --> 00:13:20,670 And, for both those reasons, it seems 249 00:13:20,670 --> 00:13:23,250 like we have to take into account how 250 00:13:23,250 --> 00:13:25,485 the entire economy is evolving. 251 00:13:28,430 --> 00:13:30,560 People don't narrowly stay in their village. 252 00:13:30,560 --> 00:13:32,600 They can migrate. 253 00:13:32,600 --> 00:13:34,190 They're forward-looking. 254 00:13:34,190 --> 00:13:35,900 Potentially, there's urbanization 255 00:13:35,900 --> 00:13:37,220 going on and so on. 256 00:13:41,170 --> 00:13:42,930 The starting point-- and I put it 257 00:13:42,930 --> 00:13:46,110 at the beginning, rather than the end of the course, 258 00:13:46,110 --> 00:13:48,250 was what models do we have. 259 00:13:48,250 --> 00:13:52,200 Let's take an inventory of the different kinds of models 260 00:13:52,200 --> 00:13:56,370 we have to try to deal with some of the puzzles 261 00:13:56,370 --> 00:13:59,790 like why is money leaving China while the rates of return 262 00:13:59,790 --> 00:14:03,480 are so high in that paper that you mentioned, 263 00:14:03,480 --> 00:14:04,470 "Growing like China." 264 00:14:08,050 --> 00:14:11,320 It is true, for example, that a lot of macro guys 265 00:14:11,320 --> 00:14:17,530 are happy sort of modeling TFP as a process. 266 00:14:17,530 --> 00:14:21,070 The business cycle literature, following 267 00:14:21,070 --> 00:14:23,780 the lead of Prescott and Kydland, 268 00:14:23,780 --> 00:14:25,390 is very much like that. 269 00:14:27,940 --> 00:14:30,790 The development literature still has those ingredients 270 00:14:30,790 --> 00:14:33,940 in the sense of allowing individual talent 271 00:14:33,940 --> 00:14:36,070 and productivity to vary, but part 272 00:14:36,070 --> 00:14:38,890 of what we see in the aggregate data 273 00:14:38,890 --> 00:14:40,720 is an equilibrium phenomenon. 274 00:14:40,720 --> 00:14:45,640 TFP has to do with potentially improving allocation 275 00:14:45,640 --> 00:14:49,340 of resources across individuals and households. 276 00:14:49,340 --> 00:14:52,540 So that brings us back to the micro data. 277 00:14:52,540 --> 00:14:57,940 And the way I see the inter-relationship is I 278 00:14:57,940 --> 00:15:01,990 could have equally well started with the micro underpinnings 279 00:15:01,990 --> 00:15:03,670 and built up to the macro, rather 280 00:15:03,670 --> 00:15:08,320 than starting with the macro and then getting into the micro. 281 00:15:08,320 --> 00:15:14,170 It is true, I think, that you need to iterate back and forth. 282 00:15:14,170 --> 00:15:19,150 You tend to lose the perspective when you're just reading 283 00:15:19,150 --> 00:15:22,100 macro development papers. 284 00:15:22,100 --> 00:15:24,450 So what does this have to do with the poor people 285 00:15:24,450 --> 00:15:25,510 in an Indian village? 286 00:15:25,510 --> 00:15:29,800 Or, vise versa, when you're trapped in a village, 287 00:15:29,800 --> 00:15:33,220 you can't see what's going on in the rest of the economy. 288 00:15:33,220 --> 00:15:36,460 I mean, it's just my firm belief that we absolutely 289 00:15:36,460 --> 00:15:38,980 have to do both. 290 00:15:38,980 --> 00:15:41,440 Although, how much of one versus the other 291 00:15:41,440 --> 00:15:44,180 depends on the application. 292 00:15:44,180 --> 00:15:46,480 But this is a macro development class. 293 00:15:46,480 --> 00:15:50,020 So it really has to get us all up 294 00:15:50,020 --> 00:15:54,430 to speed and comfortable with sort of macro. 295 00:15:54,430 --> 00:15:56,380 And the cool thing about macro development 296 00:15:56,380 --> 00:15:58,270 is this whole generation of models 297 00:15:58,270 --> 00:16:01,900 that's coming along that are explicit about micro 298 00:16:01,900 --> 00:16:07,090 underpinnings and potentially testable, whereas so much 299 00:16:07,090 --> 00:16:10,900 of the macro literature, real business cycle 300 00:16:10,900 --> 00:16:13,660 literature, new Keynesian, and so on, 301 00:16:13,660 --> 00:16:15,340 just make assumptions about what's 302 00:16:15,340 --> 00:16:20,980 going on at the individual micro level and never test. 303 00:16:20,980 --> 00:16:24,310 Markets, institutions, the nature of financial contracts 304 00:16:24,310 --> 00:16:27,970 are all primarily just assumed and taken 305 00:16:27,970 --> 00:16:30,970 as given in the more standard macro 306 00:16:30,970 --> 00:16:34,000 literature, whereas development is kind 307 00:16:34,000 --> 00:16:38,250 of a wonderful playground, if you want to put it that way, 308 00:16:38,250 --> 00:16:40,630 for seeing just, whether you're going from urban 309 00:16:40,630 --> 00:16:43,630 to rural areas or one country to the next, 310 00:16:43,630 --> 00:16:49,420 you can see these constraints being estimated and varying. 311 00:16:49,420 --> 00:16:52,570 And the work we did in the class shows that that 312 00:16:52,570 --> 00:16:54,762 matters for the macro. 313 00:16:54,762 --> 00:17:00,220 AUDIENCE: Yeah, so I guess that my point of view, maybe 314 00:17:00,220 --> 00:17:04,150 also because of my background, but it seemed to me 315 00:17:04,150 --> 00:17:10,210 that, the last part, we really were in villages like seeing 316 00:17:10,210 --> 00:17:15,310 how people do risk sharing and stuff like that and your paper 317 00:17:15,310 --> 00:17:20,500 with Kaboski about the Million Baht program. 318 00:17:20,500 --> 00:17:23,170 But them it seemed maybe we didn't 319 00:17:23,170 --> 00:17:26,410 do that much on urban, or most of the emphasis 320 00:17:26,410 --> 00:17:27,800 was on the rural side. 321 00:17:27,800 --> 00:17:29,467 I mean, we said some of the implications 322 00:17:29,467 --> 00:17:32,360 for the urban model and how the data do in the urban setting. 323 00:17:32,360 --> 00:17:38,620 So it's kind of a little hard to me to go back to that macro 324 00:17:38,620 --> 00:17:43,132 where we kind of don't have firms. 325 00:17:43,132 --> 00:17:45,280 Like it's unclear to me exactly what 326 00:17:45,280 --> 00:17:48,840 we're going when we're talking about individuals 327 00:17:48,840 --> 00:17:49,870 having projects. 328 00:17:49,870 --> 00:17:54,400 And, oh, they might employ some other people in the village. 329 00:17:54,400 --> 00:17:57,970 But that still sounds to me as like little things. 330 00:17:57,970 --> 00:17:59,440 And then how do we translate that 331 00:17:59,440 --> 00:18:02,412 into if we think more about it in a macro perspective 332 00:18:02,412 --> 00:18:04,370 where there's really no firms and [INAUDIBLE]?? 333 00:18:04,370 --> 00:18:06,453 ROBERT TOWNSEND: Well, the Kaboski paper it's true 334 00:18:06,453 --> 00:18:10,130 used the annual rural resurvey. 335 00:18:10,130 --> 00:18:12,900 So it's about village and village size. 336 00:18:12,900 --> 00:18:17,790 But, before we did that, we did the paper with Alex. 337 00:18:17,790 --> 00:18:19,480 And that's about obstacles. 338 00:18:19,480 --> 00:18:22,480 And there we used the annual data 339 00:18:22,480 --> 00:18:27,610 for both the rural and the urban or the annualized 340 00:18:27,610 --> 00:18:29,290 version of the rural. 341 00:18:29,290 --> 00:18:31,810 But, anyway, half of the estimation 342 00:18:31,810 --> 00:18:33,747 was about urban households. 343 00:18:33,747 --> 00:18:35,080 AUDIENCE: Yeah, I remember that. 344 00:18:35,080 --> 00:18:35,540 ROBERT TOWNSEND: And so-- 345 00:18:35,540 --> 00:18:36,670 AUDIENCE: And we said that the rural [? other ?] 346 00:18:36,670 --> 00:18:37,540 was very important. 347 00:18:37,540 --> 00:18:42,386 But it still kind of don't resonate with me like how-- 348 00:18:42,386 --> 00:18:44,410 like what are we thinking about when we think-- 349 00:18:44,410 --> 00:18:45,830 when we think about rural? 350 00:18:45,830 --> 00:18:47,730 Are we thinking about firms? 351 00:18:47,730 --> 00:18:52,662 Are we thinking about little resalers? 352 00:18:52,662 --> 00:18:54,410 [INAUDIBLE] 353 00:18:54,410 --> 00:18:58,010 ROBERT TOWNSEND: There's a whole spectrum of size and scale 354 00:18:58,010 --> 00:19:00,800 of firms in these countries. 355 00:19:03,380 --> 00:19:06,560 And it's true that, in the rural areas, 356 00:19:06,560 --> 00:19:11,510 they tend to be small, small or medium enterprise, household 357 00:19:11,510 --> 00:19:12,890 enterprise. 358 00:19:12,890 --> 00:19:16,040 That's actually true as well in the urban areas. 359 00:19:16,040 --> 00:19:19,790 Most of the firms in these countries are small. 360 00:19:19,790 --> 00:19:23,960 Most of them don't even hire employees. 361 00:19:23,960 --> 00:19:26,990 Now, when we're doing macro proper, 362 00:19:26,990 --> 00:19:29,600 we want to make sure that we also 363 00:19:29,600 --> 00:19:34,670 include the medium-large firms if we 364 00:19:34,670 --> 00:19:40,470 have the relevant data on them, but I'm not 365 00:19:40,470 --> 00:19:43,050 sure I see so much of a difference 366 00:19:43,050 --> 00:19:45,300 between urban and rural. 367 00:19:45,300 --> 00:19:47,910 For example, there are very active money markets 368 00:19:47,910 --> 00:19:49,620 in the urban areas. 369 00:19:49,620 --> 00:19:52,110 Credit constraints, in some respects, 370 00:19:52,110 --> 00:19:57,090 seem as severe or at least constrained, 371 00:19:57,090 --> 00:20:00,990 very high interest rates, money lending, 372 00:20:00,990 --> 00:20:07,900 despite the proximity of financial infrastructure. 373 00:20:07,900 --> 00:20:09,880 So, yeah, you can study urban neighborhoods. 374 00:20:09,880 --> 00:20:11,710 I once did this in Chicago. 375 00:20:11,710 --> 00:20:15,820 I started looking at urban neighborhoods, 376 00:20:15,820 --> 00:20:18,070 a Mexican neighborhood, a Korean neighborhood, 377 00:20:18,070 --> 00:20:19,690 an African-American neighborhood, 378 00:20:19,690 --> 00:20:25,690 and just seeing the differences and whether the institutions 379 00:20:25,690 --> 00:20:29,410 and the data seemed consistent with certain outcomes. 380 00:20:29,410 --> 00:20:34,580 So, as far as the method goes, I don't see a big difference. 381 00:20:34,580 --> 00:20:39,400 It's true that you're often persuaded by the data you have, 382 00:20:39,400 --> 00:20:43,690 and this monthly data is rather unique to the rural areas. 383 00:20:43,690 --> 00:20:49,390 So we've done more with that, but that's just a constraint. 384 00:20:49,390 --> 00:20:53,230 It's not some-- we're thinking about, god willing, 385 00:20:53,230 --> 00:20:56,080 we will gather these monthly financial accounts 386 00:20:56,080 --> 00:21:00,945 data in urban areas soon. 387 00:21:00,945 --> 00:21:02,610 AUDIENCE: So, from my perspective, 388 00:21:02,610 --> 00:21:06,040 I think that, when we talk about rural and urban, 389 00:21:06,040 --> 00:21:10,130 if it's not Bangkok, urban isn't really like urban here. 390 00:21:10,130 --> 00:21:11,540 It's just like small towns. 391 00:21:11,540 --> 00:21:13,230 So I wouldn't think that like-- 392 00:21:13,230 --> 00:21:16,550 I would think that the results from the rural data [INAUDIBLE] 393 00:21:16,550 --> 00:21:18,300 generalize to urban areas. 394 00:21:18,300 --> 00:21:19,670 But when you think about-- 395 00:21:19,670 --> 00:21:21,830 I think your question is more like big, big firms 396 00:21:21,830 --> 00:21:23,890 or big firms in developing countries. 397 00:21:23,890 --> 00:21:26,650 AUDIENCE: Yeah, so my point is that, sometimes, 398 00:21:26,650 --> 00:21:31,150 most of the time, like, when I think about TFP and studies 399 00:21:31,150 --> 00:21:36,870 about TFP, they have this concept about TFP growth 400 00:21:36,870 --> 00:21:39,770 like in a country, which is possibly unclear 401 00:21:39,770 --> 00:21:40,790 exactly what it needs. 402 00:21:40,790 --> 00:21:44,210 But it's like how much of what we're 403 00:21:44,210 --> 00:21:46,880 doing at the micro level then translates into this 404 00:21:46,880 --> 00:21:48,162 like super aggregate measure? 405 00:21:48,162 --> 00:21:49,870 AUDIENCE: And you think that most of this 406 00:21:49,870 --> 00:21:51,412 may be coming from the large firm? 407 00:21:51,412 --> 00:21:52,370 AUDIENCE: I don't know. 408 00:21:52,370 --> 00:21:54,370 I mean, I just-- again, like China is different, 409 00:21:54,370 --> 00:21:57,590 but like that one paper was about state-owned enterprises, 410 00:21:57,590 --> 00:21:59,270 which I'm assuming are big. 411 00:21:59,270 --> 00:22:01,850 Like that one made sense and was sort of-- 412 00:22:01,850 --> 00:22:05,450 which is why I kind of can't reconcile the two. 413 00:22:05,450 --> 00:22:08,810 ROBERT TOWNSEND: The Jeong, Townsend paper is about TFP 414 00:22:08,810 --> 00:22:11,370 at the macro level in Thailand. 415 00:22:11,370 --> 00:22:14,930 And there we use socioeconomic survey data, 416 00:22:14,930 --> 00:22:17,120 which includes some data on larger firms. 417 00:22:17,120 --> 00:22:23,452 I suspect it's dramatically under sampling larger firms. 418 00:22:23,452 --> 00:22:25,160 And, nevertheless, we explained something 419 00:22:25,160 --> 00:22:29,780 like 72% of TFP movements at the aggregate level. 420 00:22:29,780 --> 00:22:33,800 So I mean that was a real eye-opener that development 421 00:22:33,800 --> 00:22:35,970 in micro is all about-- 422 00:22:35,970 --> 00:22:39,220 largely about macro TFP. 423 00:22:39,220 --> 00:22:40,560 It's just efficiency. 424 00:22:40,560 --> 00:22:44,950 You know, TFP literally means nothing more than how much-- 425 00:22:44,950 --> 00:22:48,220 or, changing TFP, how much more output 426 00:22:48,220 --> 00:22:51,550 can you get period after period, even 427 00:22:51,550 --> 00:22:54,850 holding the level of capital and labor and so on fixed? 428 00:22:54,850 --> 00:22:58,420 So, to the extent that intermediation allows a better 429 00:22:58,420 --> 00:23:01,390 allocation of resources based on individual talent, 430 00:23:01,390 --> 00:23:06,610 you should expect it to show up in TFP. 431 00:23:06,610 --> 00:23:08,110 AUDIENCE: You do have a new survey 432 00:23:08,110 --> 00:23:09,970 that focuses on big firms. 433 00:23:09,970 --> 00:23:14,300 Do you have any work in progress that maybe the result is 434 00:23:14,300 --> 00:23:16,272 different from what-- 435 00:23:16,272 --> 00:23:17,730 ROBERT TOWNSEND: Well, it turns out 436 00:23:17,730 --> 00:23:20,610 that big firms are harder to find in some sense. 437 00:23:20,610 --> 00:23:22,830 They're harder to get into. 438 00:23:22,830 --> 00:23:26,040 This TEPS survey in Thailand ends up 439 00:23:26,040 --> 00:23:27,990 looking a lot like the urban survey, 440 00:23:27,990 --> 00:23:31,920 even though it was designed to try to pick up the larger ones. 441 00:23:31,920 --> 00:23:34,170 I mean, there is something called the missing middle, 442 00:23:34,170 --> 00:23:38,460 and I don't know whether it's missing because of the survey 443 00:23:38,460 --> 00:23:41,610 design techniques that are used or just missing 444 00:23:41,610 --> 00:23:43,860 because it doesn't exist. 445 00:23:43,860 --> 00:23:46,890 But the missing middle refers to the sort 446 00:23:46,890 --> 00:23:50,890 of life cycle of firms and the fact that we tend to see either 447 00:23:50,890 --> 00:23:54,690 the very big ones, as in surveys, census 448 00:23:54,690 --> 00:23:58,620 of manufacturers, or the very small ones, as in household 449 00:23:58,620 --> 00:24:01,230 survey data. 450 00:24:01,230 --> 00:24:03,720 And the interesting thing is, where do the big ones 451 00:24:03,720 --> 00:24:05,160 come from. 452 00:24:05,160 --> 00:24:06,240 What is their history? 453 00:24:06,240 --> 00:24:07,410 Go back in time. 454 00:24:07,410 --> 00:24:10,290 Did people move out of villages into the town 455 00:24:10,290 --> 00:24:12,600 and set up their businesses there? 456 00:24:12,600 --> 00:24:16,470 And that's something we don't know as much about 457 00:24:16,470 --> 00:24:18,570 as we would like to know. 458 00:24:18,570 --> 00:24:21,720 It's even hard, actually, to imagine 459 00:24:21,720 --> 00:24:26,060 how to sample to figure it out because everything 460 00:24:26,060 --> 00:24:28,190 is endogenous. 461 00:24:28,190 --> 00:24:29,840 You can comfort yourself when you 462 00:24:29,840 --> 00:24:33,470 have a fixed enumeration of households, as in a population 463 00:24:33,470 --> 00:24:34,040 census. 464 00:24:34,040 --> 00:24:36,360 Then you just sample at random. 465 00:24:36,360 --> 00:24:41,300 Or you have an official registry of large businesses, 466 00:24:41,300 --> 00:24:43,710 and then you sample them at random. 467 00:24:43,710 --> 00:24:46,640 So we can take comfort that we have 468 00:24:46,640 --> 00:24:49,250 an unbiased, stratified, random sample, 469 00:24:49,250 --> 00:24:53,480 but that doesn't get at this missing middle if-- 470 00:24:55,995 --> 00:24:56,495 yeah. 471 00:24:59,880 --> 00:25:00,583 Yes? 472 00:25:00,583 --> 00:25:02,625 AUDIENCE: So I have another very general question 473 00:25:02,625 --> 00:25:03,730 about structural model. 474 00:25:03,730 --> 00:25:07,180 So we've seen a lot of models with this heterogeneous agent. 475 00:25:07,180 --> 00:25:11,470 So, obviously, there's all kinds of sources of heterogeneity. 476 00:25:11,470 --> 00:25:14,140 And, in your model, you only choose to model-- 477 00:25:14,140 --> 00:25:15,460 you cannot model all of them. 478 00:25:15,460 --> 00:25:16,835 And what you're going to model is 479 00:25:16,835 --> 00:25:19,270 what you think is playing a key ingredient in driving 480 00:25:19,270 --> 00:25:21,180 the mechanism of the model. 481 00:25:21,180 --> 00:25:23,650 So, for the other heterogeneity, it's what? 482 00:25:23,650 --> 00:25:26,795 Like, in the paper with Kaboski, you just filter-- 483 00:25:26,795 --> 00:25:31,150 adjust your data by filtering the other heterogeneities. 484 00:25:31,150 --> 00:25:34,460 I'm just wondering whether like, for example, people 485 00:25:34,460 --> 00:25:38,368 think that there are three or four key heterogeneities that 486 00:25:38,368 --> 00:25:40,660 are driving the decision-making at the household level, 487 00:25:40,660 --> 00:25:42,660 but you're only modeling one of them, 488 00:25:42,660 --> 00:25:44,672 and you're filtering out the others will be 489 00:25:44,672 --> 00:25:45,880 getting a lot of [INAUDIBLE]. 490 00:25:45,880 --> 00:25:47,713 ROBERT TOWNSEND: Integrating up I would say, 491 00:25:47,713 --> 00:25:49,690 rather than filtering out. 492 00:25:49,690 --> 00:25:51,460 There is a question of the dimension 493 00:25:51,460 --> 00:25:55,480 of the unobserved heterogeneity that's in that model 494 00:25:55,480 --> 00:25:59,320 and in what we're going to talk about momentarily. 495 00:25:59,320 --> 00:26:02,520 You have-- like an econometrician, 496 00:26:02,520 --> 00:26:04,560 it's very similar in spirit. 497 00:26:04,560 --> 00:26:07,890 You have the observed covariates x on the right-hand side, 498 00:26:07,890 --> 00:26:11,400 and then you have some unobserved household 499 00:26:11,400 --> 00:26:13,470 or other fixed effects on the right-hand side. 500 00:26:13,470 --> 00:26:20,100 And you can assume all kinds of things about those unobserved 501 00:26:20,100 --> 00:26:23,400 heterogeneities, but you have to assume something basically. 502 00:26:23,400 --> 00:26:27,300 It's hard to do something completely non-parametric. 503 00:26:27,300 --> 00:26:28,710 So you're making assumptions. 504 00:26:28,710 --> 00:26:31,740 I'm not denying that at all. 505 00:26:31,740 --> 00:26:34,290 And it matters what you assume, but it's not 506 00:26:34,290 --> 00:26:38,335 as though the heterogeneity isn't there. 507 00:26:38,335 --> 00:26:39,960 AUDIENCE: So it's not about unobserved. 508 00:26:39,960 --> 00:26:41,840 So, for example, household [? that respond ?] 509 00:26:41,840 --> 00:26:44,052 in that paper, the heterogeneity is 510 00:26:44,052 --> 00:26:46,805 in permanent income and wealth, liquidity, wealth, 511 00:26:46,805 --> 00:26:47,680 and permanent income. 512 00:26:47,680 --> 00:26:49,060 ROBERT TOWNSEND: Yeah, wealth might be observed. 513 00:26:49,060 --> 00:26:50,770 Permanent income is not observed. 514 00:26:50,770 --> 00:26:53,320 AUDIENCE: And household composition is observed, 515 00:26:53,320 --> 00:26:55,540 but it's not modeled. 516 00:26:55,540 --> 00:26:58,620 ROBERT TOWNSEND: Oh, now that's something we did for the data. 517 00:26:58,620 --> 00:27:00,625 And that's more a trade-off in methods. 518 00:27:03,430 --> 00:27:04,790 Let me draw an analogy. 519 00:27:04,790 --> 00:27:08,530 If you did real business cycle modeling, 520 00:27:08,530 --> 00:27:12,820 you would basically take a trend out of the data. 521 00:27:12,820 --> 00:27:15,190 I mean, in the US, there's just this huge trend. 522 00:27:15,190 --> 00:27:17,110 I mean, if you looked at the business cycles 523 00:27:17,110 --> 00:27:20,740 around the trend, you'd like what is this subject called 524 00:27:20,740 --> 00:27:21,917 business cycle. 525 00:27:21,917 --> 00:27:23,500 Of course, when you're in a recession, 526 00:27:23,500 --> 00:27:25,600 it seems real enough. 527 00:27:25,600 --> 00:27:26,830 So people have a filter. 528 00:27:26,830 --> 00:27:28,690 You could take a linear trend out. 529 00:27:28,690 --> 00:27:31,450 You could use this Hodrick-Prescott filter 530 00:27:31,450 --> 00:27:35,650 and take out some kind of smooth moving average. 531 00:27:35,650 --> 00:27:37,480 And then, having taken it out, you'd 532 00:27:37,480 --> 00:27:39,750 just focus on the deviations around the trend. 533 00:27:39,750 --> 00:27:44,380 So what we did was take out all the demographics, which 534 00:27:44,380 --> 00:27:49,030 is not pleasant either, but, I mean, it at least-- 535 00:27:49,030 --> 00:27:52,510 the other thing to do is to leave it in the data 536 00:27:52,510 --> 00:27:54,700 and use a model that doesn't have it. 537 00:27:54,700 --> 00:27:56,710 And that seems weird too. 538 00:27:56,710 --> 00:27:58,540 That's like doing business cycle modeling 539 00:27:58,540 --> 00:28:00,742 and leaving the trend in. 540 00:28:00,742 --> 00:28:02,450 AUDIENCE: So, id you don't-- for example, 541 00:28:02,450 --> 00:28:04,220 but, if the question that you're-- 542 00:28:04,220 --> 00:28:07,220 like the model, like, if people think that household 543 00:28:07,220 --> 00:28:11,050 composition or demographics is playing some sort of key role 544 00:28:11,050 --> 00:28:12,700 in the household decision-making, 545 00:28:12,700 --> 00:28:14,712 but you're filtering it out-- 546 00:28:14,712 --> 00:28:16,420 ROBERT TOWNSEND: That's a fair criticism. 547 00:28:16,420 --> 00:28:18,750 AUDIENCE: So there are criticisms like that? 548 00:28:18,750 --> 00:28:20,710 ROBERT TOWNSEND: Yeah, yeah, absolutely. 549 00:28:20,710 --> 00:28:25,390 Yeah, well, it sounded like all the research got done. 550 00:28:25,390 --> 00:28:28,860 Oh wow, we'll see you on Thursday. 551 00:28:28,860 --> 00:28:32,110 No, no, those are good, productive thoughts, 552 00:28:32,110 --> 00:28:34,300 which is it's not like-- 553 00:28:34,300 --> 00:28:37,150 you can be sort of uncomfortable with research in general 554 00:28:37,150 --> 00:28:39,760 and the apparently arbitrarily way 555 00:28:39,760 --> 00:28:42,590 in which different researchers make assumptions. 556 00:28:42,590 --> 00:28:45,730 But where research kicks in in a very positive way is to say, 557 00:28:45,730 --> 00:28:49,450 no, I don't believe that is a reasonable assumption. 558 00:28:49,450 --> 00:28:51,100 I think it would matter a lot. 559 00:28:51,100 --> 00:28:54,610 Let me try to incorporate it and maybe leave out 560 00:28:54,610 --> 00:28:58,120 some other things and see what difference it makes. 561 00:28:58,120 --> 00:29:00,956 That's where good research is born, absolutely. 562 00:29:04,220 --> 00:29:05,490 Yes? 563 00:29:05,490 --> 00:29:10,320 AUDIENCE: So a lot of the models that we've covered 564 00:29:10,320 --> 00:29:14,032 has a wage as a key parameter. 565 00:29:14,032 --> 00:29:20,616 But, in like some research in, if I remember correctly, 566 00:29:20,616 --> 00:29:25,500 your paper in Econometrica doesn't find a effect 567 00:29:25,500 --> 00:29:28,310 on wage, at least short term. 568 00:29:28,310 --> 00:29:31,700 And, also, like Esther and Abhijit's paper 569 00:29:31,700 --> 00:29:35,230 didn't also find wage, like effect on wage, 570 00:29:35,230 --> 00:29:37,625 of microfinance [INAUDIBLE]. 571 00:29:37,625 --> 00:29:39,500 ROBERT TOWNSEND: Joe and I did find an effect 572 00:29:39,500 --> 00:29:42,020 on the wage in the astructural model. 573 00:29:42,020 --> 00:29:44,533 AUDIENCE: Oh, OK so that's totally [INAUDIBLE].. 574 00:29:44,533 --> 00:29:45,700 ROBERT TOWNSEND: Yeah, yeah. 575 00:29:45,700 --> 00:29:46,700 AUDIENCE: I missed that. 576 00:29:46,700 --> 00:29:47,810 ROBERT TOWNSEND: But, again, that's 577 00:29:47,810 --> 00:29:50,000 a good example again of these trade-offs, which 578 00:29:50,000 --> 00:29:52,580 is you have a sort of a search algorithm 579 00:29:52,580 --> 00:29:56,600 through the lens of a somewhat reduced form, 580 00:29:56,600 --> 00:30:00,170 maybe instrumented variables to see what's big 581 00:30:00,170 --> 00:30:00,950 and what's small. 582 00:30:00,950 --> 00:30:03,020 And we found an effect on the wage. 583 00:30:03,020 --> 00:30:06,020 Now it is true that in that model, 584 00:30:06,020 --> 00:30:08,240 the structural Econometrica paper with Joe, 585 00:30:08,240 --> 00:30:11,510 we did not try to make wages endogenous. 586 00:30:11,510 --> 00:30:15,050 In fact, we allowed investment, but we moved away 587 00:30:15,050 --> 00:30:18,600 from occupation choice and these key ingredients 588 00:30:18,600 --> 00:30:22,040 of a lot of these macro models. 589 00:30:22,040 --> 00:30:24,140 It's not like it couldn't be done. 590 00:30:24,140 --> 00:30:28,550 It was just something we chose not to focus on. 591 00:30:32,030 --> 00:30:35,440 AUDIENCE: I guess my point is-- 592 00:30:35,440 --> 00:30:39,010 so I don't think there are a lot of evidence, especially 593 00:30:39,010 --> 00:30:43,290 [? treating ?] endogeneity of financial development. 594 00:30:43,290 --> 00:30:47,140 But it seems to me that, at least 595 00:30:47,140 --> 00:30:51,900 like RCT of like MFI and [INAUDIBLE],, 596 00:30:51,900 --> 00:30:55,780 they usually didn't find, but I only know two papers, 597 00:30:55,780 --> 00:30:59,810 but they didn't find like effect on wage. 598 00:30:59,810 --> 00:31:04,980 So I'm kind of wondering like what's your [INAUDIBLE].. 599 00:31:04,980 --> 00:31:08,970 ROBERT TOWNSEND: I think, if you think about-- 600 00:31:08,970 --> 00:31:11,860 you have to think about the scale of the intervention, 601 00:31:11,860 --> 00:31:12,360 all right? 602 00:31:12,360 --> 00:31:17,400 So, if you have an RCT, even with a fairly big sample, 603 00:31:17,400 --> 00:31:22,290 in a small, open economy, it's arguably true 604 00:31:22,290 --> 00:31:24,120 that you wouldn't necessarily expect 605 00:31:24,120 --> 00:31:26,130 to see anything with the wage. 606 00:31:26,130 --> 00:31:31,320 But, if you were scaling that same experiment up 607 00:31:31,320 --> 00:31:35,460 to the national level, then you probably would see it. 608 00:31:35,460 --> 00:31:37,890 Actually, that's a very good argument 609 00:31:37,890 --> 00:31:42,780 for doing both basically and, in particular, 610 00:31:42,780 --> 00:31:46,140 not getting misled by program impact. 611 00:31:46,140 --> 00:31:49,830 Ester's thesis in Indonesia was about education reform. 612 00:31:49,830 --> 00:31:54,360 And she was worried about whether the wages would 613 00:31:54,360 --> 00:31:58,330 go down, as you had more educated cohorts. 614 00:31:58,330 --> 00:32:00,030 So people have worried about this kind 615 00:32:00,030 --> 00:32:02,490 of thing for quite a long time. 616 00:32:02,490 --> 00:32:03,720 There is some work in Canada. 617 00:32:03,720 --> 00:32:09,600 RCTs are a bit sort of like they look new, like people maybe 618 00:32:09,600 --> 00:32:12,420 haven't been doing them much, but it's not true actually. 619 00:32:12,420 --> 00:32:14,790 If you can go back into the '60s and '70s, 620 00:32:14,790 --> 00:32:20,250 there was work in Canada on I think 621 00:32:20,250 --> 00:32:22,560 it was either a pension or social security 622 00:32:22,560 --> 00:32:25,920 reform where they did exactly what I was saying. 623 00:32:25,920 --> 00:32:28,110 They kind of backed that impact locally, 624 00:32:28,110 --> 00:32:29,970 and then they extrapolated to what 625 00:32:29,970 --> 00:32:32,790 would happen to the rest of the country 626 00:32:32,790 --> 00:32:34,620 if it were implemented nationally 627 00:32:34,620 --> 00:32:35,931 with the structural model. 628 00:32:40,150 --> 00:32:40,650 Yes? 629 00:32:40,650 --> 00:32:43,290 AUDIENCE: So this is a very broad question, 630 00:32:43,290 --> 00:32:45,563 but [INAUDIBLE] asked something about the financial 631 00:32:45,563 --> 00:32:50,400 [INAUDIBLE],, especially corporate financial [INAUDIBLE] 632 00:32:50,400 --> 00:32:53,140 equity markets and money markets. 633 00:32:53,140 --> 00:32:56,070 So these markets of course are [INAUDIBLE] 634 00:32:56,070 --> 00:33:01,270 to the way economic surplus are divided into different people 635 00:33:01,270 --> 00:33:03,300 or different organizations. 636 00:33:03,300 --> 00:33:06,910 So, as a result of that [INAUDIBLE] impasse, 637 00:33:06,910 --> 00:33:11,560 how, for example, new projects or new policies 638 00:33:11,560 --> 00:33:15,480 have benefited different types of groups or organizations-- 639 00:33:15,480 --> 00:33:20,220 so there are some political economy issues related 640 00:33:20,220 --> 00:33:25,406 to the imperfection of the corporate financial 641 00:33:25,406 --> 00:33:27,300 [INAUDIBLE]. 642 00:33:27,300 --> 00:33:29,820 ROBERT TOWNSEND: Impact of the corporate finance on-- 643 00:33:29,820 --> 00:33:33,690 AUDIENCE: Structure on-- so political economy concerns 644 00:33:33,690 --> 00:33:39,000 about who benefits from new policies and new projects. 645 00:33:39,000 --> 00:33:41,320 ROBERT TOWNSEND: Well, there's two things. 646 00:33:41,320 --> 00:33:42,870 On almost anything we've done, there 647 00:33:42,870 --> 00:33:46,190 is sort of the winners and losers. 648 00:33:46,190 --> 00:33:48,060 And that, in principle, that has nothing 649 00:33:48,060 --> 00:33:49,380 to do with political economy. 650 00:33:49,380 --> 00:33:53,600 That's just mapping out who's going to get the benefits. 651 00:33:53,600 --> 00:33:58,040 Now the next step is, if you think there are barriers, 652 00:33:58,040 --> 00:34:01,940 it may be that the losers are blocking reforms. 653 00:34:01,940 --> 00:34:04,025 And that's definitely a political economy. 654 00:34:04,025 --> 00:34:04,241 AUDIENCE: Oh yes, then my question 655 00:34:04,241 --> 00:34:06,810 is does the literature seriously investigate 656 00:34:06,810 --> 00:34:12,130 the second step you just described, [? fair ?] wages. 657 00:34:12,130 --> 00:34:14,710 ROBERT TOWNSEND: I mean, Daron has occupation choice models 658 00:34:14,710 --> 00:34:17,380 where, basically, the oligopolists are the ones 659 00:34:17,380 --> 00:34:21,230 sort of dictating the outcome. 660 00:34:21,230 --> 00:34:26,380 And they are a force for financial repression 661 00:34:26,380 --> 00:34:30,460 because they want to keep the wages low. 662 00:34:30,460 --> 00:34:31,840 I mean, that's at least a step. 663 00:34:34,750 --> 00:34:38,199 We, in this class-- 664 00:34:38,199 --> 00:34:40,659 let me try to think if we actually covered it. 665 00:34:44,030 --> 00:34:47,120 A bit, yeah, you saw one version of that. 666 00:34:47,120 --> 00:34:50,420 We can model how financial service providers 667 00:34:50,420 --> 00:34:52,460 ought to be behaving if we believe 668 00:34:52,460 --> 00:34:54,199 the structure of the model. 669 00:34:54,199 --> 00:34:56,150 Then we look at the data, and we discover 670 00:34:56,150 --> 00:34:59,390 that there are gaps in some places 671 00:34:59,390 --> 00:35:03,030 and too many banks in other places. 672 00:35:03,030 --> 00:35:04,670 And so, through the lens of the model, 673 00:35:04,670 --> 00:35:08,450 you sort of get suspicious that-- 674 00:35:08,450 --> 00:35:11,090 develop a null hypothesis that there 675 00:35:11,090 --> 00:35:14,410 may be political considerations at work. 676 00:35:14,410 --> 00:35:22,355 Now how explicit or evil or well intended it is, that's-- 677 00:35:22,355 --> 00:35:26,060 you have to go in there with your eyes 678 00:35:26,060 --> 00:35:28,730 open, but not necessarily assume. 679 00:35:32,968 --> 00:35:35,218 AUDIENCE: I have another question if no one else does. 680 00:35:35,218 --> 00:35:36,170 ROBERT TOWNSEND: OK. 681 00:35:36,170 --> 00:35:36,983 AUDIENCE: So-- 682 00:35:36,983 --> 00:35:38,941 ROBERT TOWNSEND: Anything but the lecture, huh? 683 00:35:38,941 --> 00:35:41,300 AUDIENCE: No, I just-- this is-- 684 00:35:41,300 --> 00:35:43,910 so I'm kind of embarrassed to ask this question because it 685 00:35:43,910 --> 00:35:45,280 seems like it's something that I should have been 686 00:35:45,280 --> 00:35:46,488 able to figure out on my own. 687 00:35:46,488 --> 00:35:48,800 But you see a lot of people throwing 688 00:35:48,800 --> 00:35:52,370 around the word structural in ways that are confusing 689 00:35:52,370 --> 00:35:54,530 and in ways that sometimes-- 690 00:35:54,530 --> 00:35:56,300 sort of I guess the natural and then 691 00:35:56,300 --> 00:36:00,830 from what we've looked at and all your papers, 692 00:36:00,830 --> 00:36:03,140 I guess I'd sort of be inclined to say your models 693 00:36:03,140 --> 00:36:06,890 are only structural if you are identifying 694 00:36:06,890 --> 00:36:10,850 like underlying parameters that are in a model of interest 695 00:36:10,850 --> 00:36:13,020 using econometrics or using calibrated simulation. 696 00:36:13,020 --> 00:36:15,500 But I see a lot of people calling stuff structural that 697 00:36:15,500 --> 00:36:17,900 doesn't necessarily actually identify those parameters. 698 00:36:17,900 --> 00:36:19,483 Have I got the wrong end of the stick? 699 00:36:19,483 --> 00:36:23,057 Or is there like a nuance that I don't see? 700 00:36:23,057 --> 00:36:24,140 ROBERT TOWNSEND: I don't-- 701 00:36:24,140 --> 00:36:26,390 I don't actually know for sure, depending 702 00:36:26,390 --> 00:36:28,727 on who's using the words, what they have in mind. 703 00:36:28,727 --> 00:36:29,810 AUDIENCE: Yeah, me either. 704 00:36:29,810 --> 00:36:34,610 ROBERT TOWNSEND: I think, legitimately, 705 00:36:34,610 --> 00:36:39,380 many people are worried that assumptions are being made 706 00:36:39,380 --> 00:36:42,650 in structural models, which are never tested, 707 00:36:42,650 --> 00:36:45,770 and that you're seeing the observables 708 00:36:45,770 --> 00:36:48,270 through these auxiliary assumptions. 709 00:36:48,270 --> 00:36:51,950 And, if they were sort of totally off and totally wrong, 710 00:36:51,950 --> 00:36:54,260 then you're not learning anything. 711 00:36:54,260 --> 00:36:57,620 So that tends to push you toward data summaries. 712 00:36:57,620 --> 00:37:01,350 It tends to push you toward not assuming things. 713 00:37:01,350 --> 00:37:04,460 But then there's a flip side, which is actually 714 00:37:04,460 --> 00:37:09,380 the lecture, which is this notion that there's 715 00:37:09,380 --> 00:37:11,810 theory-free tests. 716 00:37:11,810 --> 00:37:14,540 It's just not true. 717 00:37:14,540 --> 00:37:16,550 Even though RCTs and-- 718 00:37:16,550 --> 00:37:19,460 you know, they're all assuming some things 719 00:37:19,460 --> 00:37:21,950 that you might think are rather innocuous, 720 00:37:21,950 --> 00:37:25,220 but, when you start to think about the structure 721 00:37:25,220 --> 00:37:29,240 of decision-making, there's content there. 722 00:37:29,240 --> 00:37:32,090 And those assumptions could be wrong. 723 00:37:32,090 --> 00:37:35,150 So I view the tension that way, which 724 00:37:35,150 --> 00:37:40,400 is trying to assume as little as possible, 725 00:37:40,400 --> 00:37:42,860 if you can get away with it, in terms of the underlying 726 00:37:42,860 --> 00:37:45,900 distributions of unobservables and so on, 727 00:37:45,900 --> 00:37:53,720 but, at the same time, mindful that data work without theory 728 00:37:53,720 --> 00:37:58,890 has big limitations. 729 00:37:58,890 --> 00:38:02,640 And that's a debate, philosophical and otherwise, 730 00:38:02,640 --> 00:38:06,150 that goes way back, I mean, back to business cycles. 731 00:38:06,150 --> 00:38:07,800 You know, Burns and Mitchells-- 732 00:38:07,800 --> 00:38:13,500 Burns and Mitchell have this business cycle measurement, 733 00:38:13,500 --> 00:38:15,225 the NBER dates actually. 734 00:38:18,330 --> 00:38:22,920 And Koopmans came along, and he said it's just crap. 735 00:38:22,920 --> 00:38:26,310 And there was a lot of fighting. 736 00:38:26,310 --> 00:38:28,500 So how can you measure something if you don't have 737 00:38:28,500 --> 00:38:30,200 a model in mind, et cetera? 738 00:38:30,200 --> 00:38:32,880 So I'm not saying there aren't good answers to that, 739 00:38:32,880 --> 00:38:37,290 but this is not just a debate in development economics. 740 00:38:37,290 --> 00:38:40,140 This is, again, kind of a judgment. 741 00:38:43,920 --> 00:38:46,040 AUDIENCE: So it's a kind of vague question, 742 00:38:46,040 --> 00:38:51,840 but we've seen in the lecture a lot of models and evidence that 743 00:38:51,840 --> 00:38:56,760 financial [INAUDIBLE],, [? let's say, ?] sort of like 744 00:38:56,760 --> 00:39:00,650 what financial development does to the real economy. 745 00:39:00,650 --> 00:39:05,310 And how would you compare that to the situation 746 00:39:05,310 --> 00:39:06,853 in developed countries? 747 00:39:10,640 --> 00:39:14,610 What is like-- what is the difference? 748 00:39:14,610 --> 00:39:17,980 Because like a lot of models-- 749 00:39:17,980 --> 00:39:19,970 at least model-wise, you can apply it 750 00:39:19,970 --> 00:39:22,480 to developed countries. 751 00:39:22,480 --> 00:39:25,510 And, at first glance, like, for example, 752 00:39:25,510 --> 00:39:33,830 like for consumption, people aren't risk sharing quite well. 753 00:39:33,830 --> 00:39:43,130 So like it's quite close to this situation 754 00:39:43,130 --> 00:39:46,520 where the market is complete. 755 00:39:46,520 --> 00:39:52,170 On the other hand, if you look at macro level, 756 00:39:52,170 --> 00:39:55,920 as we saw in the first lecture, like there's-- 757 00:39:55,920 --> 00:39:59,840 or like people tend to think that financial development 758 00:39:59,840 --> 00:40:03,620 [INAUDIBLE] varies over countries 759 00:40:03,620 --> 00:40:07,410 and affect [INAUDIBLE]. 760 00:40:07,410 --> 00:40:09,900 ROBERT TOWNSEND: Well, we've used the US as-- 761 00:40:09,900 --> 00:40:13,080 some of those papers used the US as like the benchmark standard, 762 00:40:13,080 --> 00:40:16,590 as if it had perfect and complete financial markets. 763 00:40:16,590 --> 00:40:22,440 But I think we all know now just how 764 00:40:22,440 --> 00:40:24,750 badly the US and other advanced countries 765 00:40:24,750 --> 00:40:26,010 have done over the last-- 766 00:40:26,010 --> 00:40:28,050 since '08, '09, and so on. 767 00:40:28,050 --> 00:40:30,570 So, actually, what's going on in macro 768 00:40:30,570 --> 00:40:34,410 is this huge rebirth of interest in modeling 769 00:40:34,410 --> 00:40:36,170 the financial sector. 770 00:40:36,170 --> 00:40:40,600 There's no consensus on how to do it. 771 00:40:40,600 --> 00:40:42,370 And I think that literature would 772 00:40:42,370 --> 00:40:44,260 benefit from the advances that have 773 00:40:44,260 --> 00:40:46,690 been made in macro development. 774 00:40:46,690 --> 00:40:48,670 But there's no reason why-- 775 00:40:48,670 --> 00:40:50,200 I mean, the same issues will come up 776 00:40:50,200 --> 00:40:53,800 when thinking about the US or Spain. 777 00:40:53,800 --> 00:40:59,587 I mean, Spain is an example, for a good example, 778 00:40:59,587 --> 00:41:02,170 depending on whether you think it's an advanced industrialized 779 00:41:02,170 --> 00:41:03,190 country or not. 780 00:41:07,780 --> 00:41:12,820 Spain liberalized its financial system in the late '80s 781 00:41:12,820 --> 00:41:17,680 with a series of reforms on interest rates 782 00:41:17,680 --> 00:41:21,420 and the nature of business, of savings and loans, 783 00:41:21,420 --> 00:41:22,420 and all these providers. 784 00:41:22,420 --> 00:41:26,380 So these little regional banks, like Santander up there 785 00:41:26,380 --> 00:41:34,920 in the north and, basically, along with many of the others, 786 00:41:34,920 --> 00:41:38,010 expanded their financial infrastructure, 787 00:41:38,010 --> 00:41:42,930 establishing branches, spreading throughout the rest of Spain 788 00:41:42,930 --> 00:41:47,790 and, eventually, of course, on to Latin America. 789 00:41:47,790 --> 00:41:51,180 And then now we have this mess, right? 790 00:41:51,180 --> 00:41:54,120 So what went wrong? 791 00:41:54,120 --> 00:41:57,885 Or what went wrong in the US mortgage crisis and so on? 792 00:41:57,885 --> 00:42:01,180 So I think those are-- 793 00:42:01,180 --> 00:42:03,240 there's hardly a consensus, as I said, 794 00:42:03,240 --> 00:42:06,000 but those are really important topics. 795 00:42:06,000 --> 00:42:11,940 This is haunting countries, as well as the IMF and the World 796 00:42:11,940 --> 00:42:18,460 Bank and so on, now because countries want to grow. 797 00:42:18,460 --> 00:42:22,920 There's a lot of poverty, a lot of inequality. 798 00:42:22,920 --> 00:42:27,630 And there is this notion that expanding 799 00:42:27,630 --> 00:42:31,170 financial systems, finance, can cause growth, 800 00:42:31,170 --> 00:42:37,600 and growth can trickle down and help reduce poverty and so on. 801 00:42:37,600 --> 00:42:40,650 So there is this sort of development view 802 00:42:40,650 --> 00:42:43,980 that finance ought to expand in some way and cause growth. 803 00:42:43,980 --> 00:42:47,970 And, at the same time, there's this worry that expanded, 804 00:42:47,970 --> 00:42:50,690 deeper financial systems are going 805 00:42:50,690 --> 00:42:53,820 to be prone to instability. 806 00:42:53,820 --> 00:42:57,480 And that's a force for repression. 807 00:42:57,480 --> 00:42:59,760 And that was the first lecture of the class, 808 00:42:59,760 --> 00:43:04,290 actually, was how do we deal with both these, 809 00:43:04,290 --> 00:43:07,210 get both these things on the same page. 810 00:43:07,210 --> 00:43:09,060 But, for sure, today in the world, 811 00:43:09,060 --> 00:43:12,070 throughout Africa and Asia and so on, 812 00:43:12,070 --> 00:43:14,480 people are sort of wanting to expand 813 00:43:14,480 --> 00:43:18,090 and, at the same time, quite scared, as regulators, 814 00:43:18,090 --> 00:43:20,880 that they're going to take the blame in the future 815 00:43:20,880 --> 00:43:22,100 when something goes wrong. 816 00:43:25,540 --> 00:43:29,095 All right, so I'm going to give you a 30-minute summary. 817 00:43:31,910 --> 00:43:36,990 Fortunately, some of these themes have come up. 818 00:43:36,990 --> 00:43:39,300 And, as always, the lecture notes 819 00:43:39,300 --> 00:43:42,990 are on the Stellar website. 820 00:43:42,990 --> 00:43:47,220 So it's OK if we don't cover everything in enormous detail. 821 00:43:47,220 --> 00:43:49,830 Oh, by the way, one last thing, which is I 822 00:43:49,830 --> 00:43:52,260 did write out, for those of you who 823 00:43:52,260 --> 00:43:55,950 are worried about the generals, sort of a reader's guide 824 00:43:55,950 --> 00:43:56,575 to the-- 825 00:43:56,575 --> 00:43:59,790 not a reader's guide, a summary of the questions 826 00:43:59,790 --> 00:44:02,010 that we've been addressing in the class. 827 00:44:02,010 --> 00:44:04,230 It's just kind of a way to organize the material, 828 00:44:04,230 --> 00:44:06,570 and I'll post that on Stellar and-- 829 00:44:06,570 --> 00:44:09,883 who's also going to have some sessions. 830 00:44:09,883 --> 00:44:12,300 I think there's only four or five of you that are actually 831 00:44:12,300 --> 00:44:15,370 going to take the generals, but, if you are, 832 00:44:15,370 --> 00:44:19,820 it's easy to get lost in all this material. 833 00:44:19,820 --> 00:44:21,278 So I wrote up this sort of summary. 834 00:44:21,278 --> 00:44:22,778 AUDIENCE: [INAUDIBLE] [? going to ?] 835 00:44:22,778 --> 00:44:23,770 need it for next year. 836 00:44:28,008 --> 00:44:29,550 ROBERT TOWNSEND: Yeah, well, by then, 837 00:44:29,550 --> 00:44:32,810 research will have advanced, and we'll-- 838 00:44:32,810 --> 00:44:37,730 OK, so this is about this structural 839 00:44:37,730 --> 00:44:39,410 versus reduced form stuff. 840 00:44:43,120 --> 00:44:49,100 And the topic is the impact of expanding financial systems. 841 00:44:49,100 --> 00:44:54,980 So it's both a methods and a topic that basically 842 00:44:54,980 --> 00:44:57,040 is the hallmark of the class. 843 00:44:57,040 --> 00:44:58,580 There's three things. 844 00:44:58,580 --> 00:45:01,340 I'll probably only barely get through the first. 845 00:45:01,340 --> 00:45:07,160 One is sort of IV versus structural analysis, 846 00:45:07,160 --> 00:45:09,800 all in the same model, basically. 847 00:45:09,800 --> 00:45:12,600 The second is this expansion of banks in Spain, 848 00:45:12,600 --> 00:45:17,300 which I was already mentioning, using what the IO guys do, 849 00:45:17,300 --> 00:45:21,410 the way they avoid having to spell out 850 00:45:21,410 --> 00:45:24,350 the entire structure of the equilibrium 851 00:45:24,350 --> 00:45:29,400 by using reduced form Markov processes. 852 00:45:29,400 --> 00:45:31,460 And then, finally, if I were to have time, 853 00:45:31,460 --> 00:45:33,620 I could tell you about the limitations of that. 854 00:45:33,620 --> 00:45:37,490 The point is there's no one thing, 855 00:45:37,490 --> 00:45:42,880 but you need to be aware of the strengths and limitations. 856 00:45:42,880 --> 00:45:48,480 So this is a paper I wrote with Urzua, 857 00:45:48,480 --> 00:45:55,480 and it's about occupation choice and about 858 00:45:55,480 --> 00:45:58,350 unobserved heterogeneity. 859 00:46:01,780 --> 00:46:05,720 So a lot of this is already familiar. 860 00:46:05,720 --> 00:46:08,110 You've got sort of your wealth if you're 861 00:46:08,110 --> 00:46:11,710 a wage earner, your wealth if you're an entrepreneur. 862 00:46:11,710 --> 00:46:14,210 Instead of having just a flat wage for everyone, 863 00:46:14,210 --> 00:46:16,450 there's sort of this productivity shifter, 864 00:46:16,450 --> 00:46:18,730 which varies with households. 865 00:46:18,730 --> 00:46:19,720 Or you can run a firm. 866 00:46:19,720 --> 00:46:25,300 There's some setup costs unobserved, again, 867 00:46:25,300 --> 00:46:28,980 varying with households, [? i. ?] 868 00:46:28,980 --> 00:46:32,500 The profits come from maximizing revenue 869 00:46:32,500 --> 00:46:34,120 less the cost of inputs. 870 00:46:34,120 --> 00:46:36,310 In this case, there's no financial sector. 871 00:46:36,310 --> 00:46:38,700 Ironically, I'll get there. 872 00:46:38,700 --> 00:46:42,388 And you just have to finance the setup costs and your capital 873 00:46:42,388 --> 00:46:43,430 with your initial wealth. 874 00:46:43,430 --> 00:46:44,520 So you saw this. 875 00:46:44,520 --> 00:46:46,050 This was day two. 876 00:46:46,050 --> 00:46:49,200 This was the Lloyd-Ellis and Bernhardt model, 877 00:46:49,200 --> 00:46:53,190 versions of which we took into or modified 878 00:46:53,190 --> 00:46:58,710 in some of the other papers in the literature. 879 00:46:58,710 --> 00:47:03,030 So you get to choose one or the other. 880 00:47:03,030 --> 00:47:07,080 So we can create this decision rule, a dummy D 881 00:47:07,080 --> 00:47:09,960 for do profits dominate wages. 882 00:47:09,960 --> 00:47:13,370 If so, do it, and, otherwise, don't. 883 00:47:13,370 --> 00:47:15,030 OK, so it's an indicator. 884 00:47:15,030 --> 00:47:19,080 Note that it depends on observables, 885 00:47:19,080 --> 00:47:22,410 the economy-wide wage and the wealth, 886 00:47:22,410 --> 00:47:25,200 which varies across households, and these unobservables, which 887 00:47:25,200 --> 00:47:29,760 are these setup costs or talent for both being a wage earner 888 00:47:29,760 --> 00:47:31,980 and being an entrepreneur. 889 00:47:31,980 --> 00:47:33,270 We're not going to see those. 890 00:47:36,690 --> 00:47:40,590 So, as you know, it's sort of naive-- 891 00:47:40,590 --> 00:47:43,560 you know, incredibly simple, but, in some respects, 892 00:47:43,560 --> 00:47:47,070 powerful way to write down the observable 893 00:47:47,070 --> 00:47:50,430 depends on which thing you do. 894 00:47:50,430 --> 00:47:54,510 But, if D equals 1, then 1 minus D is 0, 895 00:47:54,510 --> 00:47:57,540 and you have the profits and earnings of being a firm 896 00:47:57,540 --> 00:48:03,450 versus, if D is 0, you're going to have the wage earnings 897 00:48:03,450 --> 00:48:05,430 from being a wage labor. 898 00:48:05,430 --> 00:48:08,610 So you can write down, if you were running 899 00:48:08,610 --> 00:48:15,240 a regression of earnings, you could let that regression 900 00:48:15,240 --> 00:48:21,270 have wealth in there and wages. 901 00:48:21,270 --> 00:48:25,500 And then you'd have this sort of dummy, 902 00:48:25,500 --> 00:48:31,140 which is a 0, 1 binary variable, plus some epsilon term, 903 00:48:31,140 --> 00:48:33,820 as regressions tend to have. 904 00:48:33,820 --> 00:48:36,510 However, following the structure of this model, 905 00:48:36,510 --> 00:48:40,530 we know exactly what the epsilon is. 906 00:48:40,530 --> 00:48:46,380 It's this D-weighted average of the talent. 907 00:48:46,380 --> 00:48:49,500 Now you don't change your talent when you make an-- 908 00:48:49,500 --> 00:48:52,530 you have your talent in both of those occupations. 909 00:48:52,530 --> 00:48:55,470 And that's what's driving the choice, OK? 910 00:48:55,470 --> 00:49:02,430 So, basically, the worry, correctly, 911 00:49:02,430 --> 00:49:04,590 is that D on the right-hand side is 912 00:49:04,590 --> 00:49:10,410 correlated with epsilon because that's self-selection, OK? 913 00:49:10,410 --> 00:49:15,000 You choose the occupation with which you're most productive, 914 00:49:15,000 --> 00:49:17,670 and part of that choice is based on unobservables. 915 00:49:17,670 --> 00:49:20,610 So we see it both in the choice and in the error. 916 00:49:24,720 --> 00:49:28,758 OK, so we need an instrument. 917 00:49:31,510 --> 00:49:37,840 Let's imagine in this economy there's some subsidy or tax. 918 00:49:37,840 --> 00:49:41,720 And it's lump sum and administered, say, randomly. 919 00:49:45,660 --> 00:49:49,830 Well, then, clearly, the decision variable, 920 00:49:49,830 --> 00:49:55,620 which has a phi in it now is influenced by that. 921 00:49:55,620 --> 00:49:58,020 And, the higher is the subsidy, the more likely 922 00:49:58,020 --> 00:50:02,640 you are to be running a firm, other things equal, right? 923 00:50:02,640 --> 00:50:09,270 But, because it's lump sum, this instrument, this subsidy, 924 00:50:09,270 --> 00:50:13,560 affects the choice of what occupation you're in, 925 00:50:13,560 --> 00:50:16,440 but it does not influence the outcome conditioned 926 00:50:16,440 --> 00:50:18,903 on choosing. 927 00:50:18,903 --> 00:50:20,320 If you choose to be a firm, you're 928 00:50:20,320 --> 00:50:21,890 going get that subsidy or not. 929 00:50:21,890 --> 00:50:26,060 And that leaves all the other decisions about labor to hire, 930 00:50:26,060 --> 00:50:26,970 et cetera, et cetera. 931 00:50:26,970 --> 00:50:29,360 And none of those decisions, none of those outcomes, 932 00:50:29,360 --> 00:50:33,370 depend on the subsidy because that's 933 00:50:33,370 --> 00:50:36,670 sort of bygones are bygones. 934 00:50:36,670 --> 00:50:40,630 You can't do much about it. 935 00:50:40,630 --> 00:50:48,280 All right, so we could imagine running that regression, 936 00:50:48,280 --> 00:50:54,270 but we regress the decision to be a firm in the data 937 00:50:54,270 --> 00:50:56,950 onto observables, like the wage and wealth, 938 00:50:56,950 --> 00:51:00,780 but also onto this subsidy as an instrument. 939 00:51:00,780 --> 00:51:03,470 That's called the IV, the instrumented version 940 00:51:03,470 --> 00:51:05,270 of the decision. 941 00:51:05,270 --> 00:51:08,510 And then we run the outcome function 942 00:51:08,510 --> 00:51:13,560 with the instrumented version, OK? 943 00:51:13,560 --> 00:51:14,995 So, hopefully, this is a review. 944 00:51:19,390 --> 00:51:26,380 And what is the policy outcome? 945 00:51:26,380 --> 00:51:30,690 The coefficient on d after instrumenting it 946 00:51:30,690 --> 00:51:34,900 is supposed to be the gain to being a firm, 947 00:51:34,900 --> 00:51:38,110 as opposed to being "treated," quote, and being a firm rather 948 00:51:38,110 --> 00:51:40,300 than being a wage earner. 949 00:51:40,300 --> 00:51:42,400 But it isn't exactly that. 950 00:51:42,400 --> 00:51:45,520 Or, at least, put on your thinking cap. 951 00:51:45,520 --> 00:51:50,450 It's the local average treatment effect. 952 00:51:50,450 --> 00:51:58,120 It is the average gain in income for those people induced 953 00:51:58,120 --> 00:52:03,900 by the subsidy to switch from being wage earners to being 954 00:52:03,900 --> 00:52:05,205 entrepreneurs. 955 00:52:09,632 --> 00:52:10,590 So you're not getting-- 956 00:52:14,167 --> 00:52:15,750 you're not getting other things, which 957 00:52:15,750 --> 00:52:20,670 are easily, similarly sounding, like treatment on the treated. 958 00:52:20,670 --> 00:52:23,160 Treatment on the treated sounds like, oh, well, people 959 00:52:23,160 --> 00:52:25,420 became firms. 960 00:52:25,420 --> 00:52:28,990 What was their profits relative to the earnings they would have 961 00:52:28,990 --> 00:52:32,130 had if they were wage earners? 962 00:52:32,130 --> 00:52:33,920 That's what this says, right? 963 00:52:33,920 --> 00:52:38,900 What is the average in the population of profits 964 00:52:38,900 --> 00:52:42,230 with the thetas in there less wages, 965 00:52:42,230 --> 00:52:46,130 given that they made the decision to be a firm, 966 00:52:46,130 --> 00:52:48,680 and their wealth is basically b? 967 00:52:48,680 --> 00:52:53,450 So I'm doing this wealth category by wealth category. 968 00:52:53,450 --> 00:52:58,210 And this is the average treatment effect. 969 00:52:58,210 --> 00:53:00,430 This is what the average would be 970 00:53:00,430 --> 00:53:04,810 if the whole population became firms, taking wages 971 00:53:04,810 --> 00:53:07,690 as given and so on. 972 00:53:07,690 --> 00:53:12,390 So, basically, be careful because LATE is not 973 00:53:12,390 --> 00:53:15,870 necessarily equal to-- what you get with the IV is not 974 00:53:15,870 --> 00:53:18,930 necessarily equal to treatment on the treated 975 00:53:18,930 --> 00:53:22,890 and not necessarily equal to the average treatment effect. 976 00:53:22,890 --> 00:53:25,950 Is this a review for you guys? 977 00:53:25,950 --> 00:53:27,080 OK, good. 978 00:53:32,020 --> 00:53:35,700 So you can put more structure as a way 979 00:53:35,700 --> 00:53:38,910 to back out the average treatment effect and so on. 980 00:53:38,910 --> 00:53:41,550 Now this is like assuming that those error 981 00:53:41,550 --> 00:53:46,020 terms are multivariate normal. 982 00:53:46,020 --> 00:53:51,030 And then you run this logistics regression, basically, 983 00:53:51,030 --> 00:53:54,870 binary yes, no, be firm or not. 984 00:53:54,870 --> 00:54:03,680 And you can back out these variance terms, 985 00:54:03,680 --> 00:54:05,510 assuming that there's no-- 986 00:54:08,060 --> 00:54:10,010 that the talent things have say zero 987 00:54:10,010 --> 00:54:13,800 mean, or you could adjust for that. 988 00:54:13,800 --> 00:54:15,545 And then you can get expressions for-- 989 00:54:19,600 --> 00:54:24,060 so then you want to basically run your regression of profits 990 00:54:24,060 --> 00:54:25,710 onto those people who have chosen 991 00:54:25,710 --> 00:54:30,270 to be firms, controlling for their wealth and their subsidy. 992 00:54:30,270 --> 00:54:33,690 And, if you didn't, if you weren't careful and did OLS, 993 00:54:33,690 --> 00:54:35,520 you'd have a bias, but you basically 994 00:54:35,520 --> 00:54:42,082 use that first stage probit to get a propensity. 995 00:54:42,082 --> 00:54:43,540 The structural model tells you what 996 00:54:43,540 --> 00:54:48,460 the exact formula is for the selection based 997 00:54:48,460 --> 00:54:49,600 on the unobservables. 998 00:54:49,600 --> 00:54:51,250 And you put that into the regression. 999 00:54:51,250 --> 00:54:55,250 You literally put this into the regression and run it. 1000 00:54:55,250 --> 00:54:58,360 And it, quote, "solves for the selection bias." 1001 00:54:58,360 --> 00:54:59,530 Yes? 1002 00:54:59,530 --> 00:55:01,780 AUDIENCE: How much more computationally difficult 1003 00:55:01,780 --> 00:55:05,110 is it to use some kind of data distribution 1004 00:55:05,110 --> 00:55:08,370 or like some other kind of-- because normal 1005 00:55:08,370 --> 00:55:10,586 is a bit of a stretch as an assumption for a talent 1006 00:55:10,586 --> 00:55:11,500 distribution. 1007 00:55:11,500 --> 00:55:13,550 ROBERT TOWNSEND: Oh, I think that's quite doable. 1008 00:55:13,550 --> 00:55:17,930 Look, this is just meant to be an example, a classic example, 1009 00:55:17,930 --> 00:55:18,430 but-- 1010 00:55:18,430 --> 00:55:19,302 AUDIENCE: Cool, so it's not too-- 1011 00:55:19,302 --> 00:55:20,223 OK, sounds good. 1012 00:55:20,223 --> 00:55:21,640 ROBERT TOWNSEND: But, I mean, this 1013 00:55:21,640 --> 00:55:27,070 question that you guys were asking, so here you've made-- 1014 00:55:27,070 --> 00:55:29,650 you're able to get average treatment effect 1015 00:55:29,650 --> 00:55:31,420 and treatment on the treated, but only 1016 00:55:31,420 --> 00:55:33,640 if you make some assumptions. 1017 00:55:33,640 --> 00:55:35,020 And some people don't like that. 1018 00:55:35,020 --> 00:55:38,350 They'll say, oh, you know, it's too structural. 1019 00:55:38,350 --> 00:55:40,915 But, in order to get answers, sometimes, you need-- 1020 00:55:40,915 --> 00:55:42,540 AUDIENCE: You have to make assumptions. 1021 00:55:42,540 --> 00:55:47,740 ROBERT TOWNSEND: OK, now, I mean, the most powerful way 1022 00:55:47,740 --> 00:55:51,320 to say this is we actually see the counterfactual 1023 00:55:51,320 --> 00:55:52,570 through the lens of the model. 1024 00:55:52,570 --> 00:55:55,450 We can see what the wages would have been for people 1025 00:55:55,450 --> 00:55:57,430 who chose to be firms. 1026 00:55:57,430 --> 00:55:59,440 By definition, they're not in the data. 1027 00:56:09,520 --> 00:56:12,685 And this is a generalized version. 1028 00:56:12,685 --> 00:56:17,590 And I'll just say that this is like local average treatment 1029 00:56:17,590 --> 00:56:22,900 effect, but really local, like just taking derivatives 1030 00:56:22,900 --> 00:56:27,640 of how that difference between profits and wages 1031 00:56:27,640 --> 00:56:30,550 is going to move, holding wealth constant, 1032 00:56:30,550 --> 00:56:35,800 as you vary something like a propensity score. 1033 00:56:35,800 --> 00:56:40,720 And the propensity score has, not too surprisingly, 1034 00:56:40,720 --> 00:56:44,620 everything to do with those unobservable talents. 1035 00:56:44,620 --> 00:56:46,120 So this is hard to implement. 1036 00:56:46,120 --> 00:56:48,190 This one is really not easy. 1037 00:56:48,190 --> 00:56:53,240 Josh doesn't like it, among other things, 1038 00:56:53,240 --> 00:56:55,720 but he's not disagreeing that it's wrong. 1039 00:56:55,720 --> 00:57:00,970 He's just thinking it's just not as operational. 1040 00:57:00,970 --> 00:57:04,360 Now the nice thing, if you could back this out, 1041 00:57:04,360 --> 00:57:07,690 then you can ignore this page, but, basically, 1042 00:57:07,690 --> 00:57:10,360 treatment on the treatment and the average treatment effect 1043 00:57:10,360 --> 00:57:15,310 are just sort of integrals of these marginal treatment 1044 00:57:15,310 --> 00:57:16,670 effects. 1045 00:57:16,670 --> 00:57:17,170 Yes? 1046 00:57:17,170 --> 00:57:19,820 AUDIENCE: So Josh sometimes recommends 1047 00:57:19,820 --> 00:57:23,210 to apply LATE to non-compliance based 1048 00:57:23,210 --> 00:57:25,888 on observable characteristics. 1049 00:57:25,888 --> 00:57:29,280 So, in the typical data set in the context 1050 00:57:29,280 --> 00:57:33,520 of occupational choice, what can we 1051 00:57:33,520 --> 00:57:39,063 observe among characteristics of [INAUDIBLE]?? 1052 00:57:39,063 --> 00:57:40,480 ROBERT TOWNSEND: So here the model 1053 00:57:40,480 --> 00:57:42,500 is taking a stand that we're going 1054 00:57:42,500 --> 00:57:45,170 to see their initial wealth. 1055 00:57:45,170 --> 00:57:48,350 We're going to see the factor prices that they faced. 1056 00:57:48,350 --> 00:57:51,640 We're going to see the profits that they make. 1057 00:57:51,640 --> 00:57:54,400 And we're not going to see these unobserved things that 1058 00:57:54,400 --> 00:57:57,970 are driving fixed costs around. 1059 00:57:57,970 --> 00:57:59,860 This is not the only model around. 1060 00:57:59,860 --> 00:58:03,190 We're going to use the model to make these examples. 1061 00:58:03,190 --> 00:58:05,470 I haven't gotten sort of to the punchline. 1062 00:58:05,470 --> 00:58:08,380 I'm happy if it's a review, but-- 1063 00:58:11,870 --> 00:58:21,020 OK, so we have the model, right? 1064 00:58:21,020 --> 00:58:24,830 So, if you don't believe this miracle, 1065 00:58:24,830 --> 00:58:27,470 let's simulate the data from the model. 1066 00:58:27,470 --> 00:58:31,670 Instead of looking at some of the cross-section that 1067 00:58:31,670 --> 00:58:35,990 got treated with the subsidy and others that didn't, let's 1068 00:58:35,990 --> 00:58:39,140 take a particular individual in the model at given parameters 1069 00:58:39,140 --> 00:58:43,460 and see what the behavior is, in terms of switching occupation 1070 00:58:43,460 --> 00:58:48,930 and so on, as you turn the subsidy on and then average up 1071 00:58:48,930 --> 00:58:50,970 in that wealth category. 1072 00:58:50,970 --> 00:58:54,600 So believing in the miracle is basically the math that you're 1073 00:58:54,600 --> 00:58:59,950 going to see the average of this model-generated simulation 1074 00:58:59,950 --> 00:59:03,250 of impact will be the IV. 1075 00:59:03,250 --> 00:59:07,215 It's going to be LATE under these assumptions. 1076 00:59:18,860 --> 00:59:22,000 So here is the IV estimate. 1077 00:59:22,000 --> 00:59:28,270 And, roughly averaging over wealth, et cetera, it's 0.45. 1078 00:59:28,270 --> 00:59:30,370 OLS is-- by the way, OLS is terrible, 1079 00:59:30,370 --> 00:59:35,080 but I guess you don't need to know why. 1080 00:59:35,080 --> 00:59:37,150 There's nothing productive about the subsidy. 1081 00:59:37,150 --> 00:59:39,130 The subsidy is just inducing people 1082 00:59:39,130 --> 00:59:41,920 who should be wage earners to be firms. 1083 00:59:41,920 --> 00:59:44,660 Of course, the impact is negative. 1084 00:59:44,660 --> 00:59:47,260 And OLS estimates it to be positive. 1085 00:59:47,260 --> 00:59:48,105 Why? 1086 00:59:48,105 --> 00:59:49,480 Because, in the observables, it's 1087 00:59:49,480 --> 00:59:51,320 the talented people who are firms, 1088 00:59:51,320 --> 00:59:53,650 but they're not being induced by the subsidy. 1089 00:59:53,650 --> 00:59:56,480 They're already firms, right? 1090 00:59:59,032 --> 00:59:59,990 I couldn't help myself. 1091 01:00:02,760 --> 01:00:08,750 OK, and this is the model-generated treatment where 1092 01:00:08,750 --> 01:00:09,920 we actually have the panel. 1093 01:00:09,920 --> 01:00:15,390 And you can see, numerically, it's quite close. 1094 01:00:15,390 --> 01:00:17,720 All right, so that's reassuring. 1095 01:00:17,720 --> 01:00:21,680 Now let's go back to the topic, which 1096 01:00:21,680 --> 01:00:23,840 is what is the impact of expanding 1097 01:00:23,840 --> 01:00:27,600 financial infrastructure. 1098 01:00:27,600 --> 01:00:30,660 Let's invent a village fund. 1099 01:00:30,660 --> 01:00:34,290 Let's let banks establish branches-- 1100 01:00:34,290 --> 01:00:37,140 forgive me for the moment-- at random. 1101 01:00:37,140 --> 01:00:39,540 Like maybe the government told them something, 1102 01:00:39,540 --> 01:00:45,740 or I'll get to that, but, for now, random. 1103 01:00:45,740 --> 01:00:50,770 So, if we observe these travel times or the effectiveness 1104 01:00:50,770 --> 01:00:52,720 of local financial infrastructure, 1105 01:00:52,720 --> 01:00:57,940 we can call this cost here to be an instrument, right? 1106 01:00:57,940 --> 01:01:01,390 So, yeah, so the idea is just like the subsidy. 1107 01:01:01,390 --> 01:01:03,430 The lower is this number, the more inclined 1108 01:01:03,430 --> 01:01:07,045 you are to incur the cost to be able to do the banking. 1109 01:01:10,710 --> 01:01:12,810 OK, we rewrite the problem. 1110 01:01:12,810 --> 01:01:15,450 This is the part of the model with financial infrastructure. 1111 01:01:15,450 --> 01:01:20,970 They get to borrow money to cover the capitalization 1112 01:01:20,970 --> 01:01:24,420 and the fixed cost. 1113 01:01:24,420 --> 01:01:26,130 Wage earners, by the way, also get 1114 01:01:26,130 --> 01:01:30,510 to put their wealth in the bank and earn money, interest 1115 01:01:30,510 --> 01:01:32,070 rate on their savings. 1116 01:01:32,070 --> 01:01:34,150 You've seen this model. 1117 01:01:34,150 --> 01:01:36,180 This was, again, the second lecture. 1118 01:01:41,700 --> 01:01:47,280 And now we have the observed outcome. 1119 01:01:47,280 --> 01:01:49,560 If you are in-- 1120 01:01:49,560 --> 01:01:52,290 if you chose to be in the intermediated sector, 1121 01:01:52,290 --> 01:01:55,080 you have another occupation choice problem 1122 01:01:55,080 --> 01:01:57,130 within that sector. 1123 01:01:57,130 --> 01:01:59,710 You can be a firm or a wage earner. 1124 01:01:59,710 --> 01:02:04,790 These D's denote that choice. 1125 01:02:04,790 --> 01:02:07,140 Notationally, it's got a w and r in there. 1126 01:02:07,140 --> 01:02:10,375 That r is a clue that we're in the intermediated sector 1127 01:02:10,375 --> 01:02:11,750 because there's an interest rate. 1128 01:02:11,750 --> 01:02:13,590 Because, if you were in the autarky sector, 1129 01:02:13,590 --> 01:02:15,340 there's no borrowing and lending of funds. 1130 01:02:19,420 --> 01:02:23,340 And, of course, this is the occupation choice 1131 01:02:23,340 --> 01:02:26,970 we had before without intermediation. 1132 01:02:29,820 --> 01:02:34,800 So we're going to get sort of the observed 1133 01:02:34,800 --> 01:02:39,120 earnings in the autarky sector and observed earnings 1134 01:02:39,120 --> 01:02:40,710 in the intermediated sector. 1135 01:02:40,710 --> 01:02:45,190 And, within both sector, there's this other margin going on, 1136 01:02:45,190 --> 01:02:47,610 which is the occupation choice. 1137 01:02:47,610 --> 01:02:49,800 All right, now we're reaching the point here. 1138 01:02:49,800 --> 01:02:50,880 We have a double margin. 1139 01:02:54,450 --> 01:02:57,420 There are two things people are choosing from, 1140 01:02:57,420 --> 01:03:02,460 whether or not to go to the bank and what occupation to have. 1141 01:03:02,460 --> 01:03:05,055 Sounds like it shouldn't cause a problem. 1142 01:03:11,140 --> 01:03:12,460 This is basically-- 1143 01:03:15,520 --> 01:03:17,600 I'm sorry this is so hard to see, 1144 01:03:17,600 --> 01:03:20,060 and it must be even harder for you than me. 1145 01:03:20,060 --> 01:03:21,890 I'm right next to it. 1146 01:03:21,890 --> 01:03:25,730 This sort of symmetric Y is like an indicator variable 1147 01:03:25,730 --> 01:03:28,220 for choosing to be intermediated. 1148 01:03:28,220 --> 01:03:31,130 These Y's with a little bit of tail off to the right 1149 01:03:31,130 --> 01:03:35,570 are incomes under intermediation and autarky. 1150 01:03:35,570 --> 01:03:38,240 It's this way in the paper, and somehow the slides 1151 01:03:38,240 --> 01:03:39,510 got created this way. 1152 01:03:39,510 --> 01:03:43,640 But, anyway, so the gain is-- 1153 01:03:49,460 --> 01:03:50,510 let's skip that. 1154 01:03:59,300 --> 01:04:00,140 So what can I say? 1155 01:04:00,140 --> 01:04:02,990 It's a double margin. 1156 01:04:02,990 --> 01:04:08,650 So you get to choose whether or not to go to the bank. 1157 01:04:08,650 --> 01:04:10,790 And, within that, each category, you 1158 01:04:10,790 --> 01:04:14,300 get to choose your occupation. 1159 01:04:14,300 --> 01:04:18,320 We can linearize the profits functions to be functions 1160 01:04:18,320 --> 01:04:21,260 of the observables and the unobservables, 1161 01:04:21,260 --> 01:04:26,930 substitute in those decisions, and write out the expression 1162 01:04:26,930 --> 01:04:34,160 for what you see in the data, namely, this income, 1163 01:04:34,160 --> 01:04:39,830 but it's just a weighted average of all the choices, 1164 01:04:39,830 --> 01:04:44,240 depending on whether these D's are 0 and 1's or these 1165 01:04:44,240 --> 01:04:48,632 symmetric Y's are 0 or 1's. 1166 01:04:48,632 --> 01:04:50,590 That was a bit of a rush, but, I mean, I think, 1167 01:04:50,590 --> 01:04:54,130 conceptually, you understand where this is coming from. 1168 01:04:54,130 --> 01:05:00,220 But, not only that, look at this error term. 1169 01:05:00,220 --> 01:05:05,890 Oh my god, the error term is complicated. 1170 01:05:05,890 --> 01:05:08,680 And it also has all these decisions, 1171 01:05:08,680 --> 01:05:13,510 but note it's got the occupation decisions in both sectors. 1172 01:05:13,510 --> 01:05:16,773 Depending on what you do, certain terms 1173 01:05:16,773 --> 01:05:18,190 are going to kick in and kick out. 1174 01:05:18,190 --> 01:05:22,780 There's massive selection going on about whether or not 1175 01:05:22,780 --> 01:05:26,588 to go to the bank, based on what occupation you're 1176 01:05:26,588 --> 01:05:28,630 going to choose, conditioned on going to the bank 1177 01:05:28,630 --> 01:05:30,820 or not going to the bank. 1178 01:05:30,820 --> 01:05:33,160 And we don't see any of this. 1179 01:05:33,160 --> 01:05:34,210 We just see the choices. 1180 01:05:34,210 --> 01:05:35,680 We don't see these error terms. 1181 01:05:39,470 --> 01:05:45,880 So what can an instrumented version of intermediation 1182 01:05:45,880 --> 01:05:47,770 give you? 1183 01:05:47,770 --> 01:05:59,490 It can give you the local average treatment, the income 1184 01:05:59,490 --> 01:06:06,720 gains that are a consequence of a lowered cost that 1185 01:06:06,720 --> 01:06:09,690 allows you-- makes you want to go to the bank, 1186 01:06:09,690 --> 01:06:13,110 as opposed to what you would do if those costs had been higher, 1187 01:06:13,110 --> 01:06:14,780 and you're not going to the bank. 1188 01:06:14,780 --> 01:06:20,280 So the local average treatment is the impact 1189 01:06:20,280 --> 01:06:24,390 on earnings that have to do with those people who are induced 1190 01:06:24,390 --> 01:06:27,750 to join the banking system as a consequence 1191 01:06:27,750 --> 01:06:33,450 of the placement of branches or varied costs of access. 1192 01:06:40,620 --> 01:06:44,040 Now you have to be careful in terms of what this isn't. 1193 01:06:44,040 --> 01:06:48,340 For example, do we want to look-- 1194 01:06:48,340 --> 01:06:50,770 what do we mean when we say we want 1195 01:06:50,770 --> 01:06:52,630 to measure the impact of improved 1196 01:06:52,630 --> 01:06:57,220 financial access on the profits of entrepreneurs? 1197 01:07:01,260 --> 01:07:02,800 You want to do an IV for that? 1198 01:07:02,800 --> 01:07:03,300 Good luck. 1199 01:07:06,280 --> 01:07:09,100 Now what can go wrong? 1200 01:07:09,100 --> 01:07:14,620 Remember monotonicity and independence, OK? 1201 01:07:14,620 --> 01:07:18,880 So monotonicity means that more people 1202 01:07:18,880 --> 01:07:22,360 are inclined to do something as a consequence of the policy 1203 01:07:22,360 --> 01:07:24,220 variation. 1204 01:07:24,220 --> 01:07:26,650 Are more people inclined to be firms 1205 01:07:26,650 --> 01:07:30,520 as a function of having reduced cost of financial access? 1206 01:07:30,520 --> 01:07:34,540 Maybe if they were credit constrained and then 1207 01:07:34,540 --> 01:07:37,060 make a lot of money in business, but there's 1208 01:07:37,060 --> 01:07:39,550 another subset of the population, 1209 01:07:39,550 --> 01:07:43,840 those not so talented people who are in financial autarky, who 1210 01:07:43,840 --> 01:07:47,590 would be actually induced to leave business and put 1211 01:07:47,590 --> 01:07:48,700 their money in a bank. 1212 01:07:51,460 --> 01:07:55,090 So it's not monotonic. 1213 01:07:55,090 --> 01:07:57,130 The effect of improved financial access 1214 01:07:57,130 --> 01:08:05,370 is not monotonic on things like the income of entrepreneurs. 1215 01:08:05,370 --> 01:08:07,400 You can, if you're careful, think 1216 01:08:07,400 --> 01:08:12,470 about what is the impact on profits of entrepreneurs, 1217 01:08:12,470 --> 01:08:15,950 of those people whose decision to be 1218 01:08:15,950 --> 01:08:20,120 an entrepreneur is not affected by financial intermediation. 1219 01:08:20,120 --> 01:08:21,859 Now that's an interesting question, 1220 01:08:21,859 --> 01:08:24,830 but it is not necessarily the one 1221 01:08:24,830 --> 01:08:26,660 that we had in mind to begin with. 1222 01:08:29,210 --> 01:08:32,609 And who are those people? 1223 01:08:32,609 --> 01:08:35,609 Well, it kind of depends on their talent, right? 1224 01:08:35,609 --> 01:08:38,040 So you're making statements about where these talent 1225 01:08:38,040 --> 01:08:40,770 cutoffs are when you're conditioning 1226 01:08:40,770 --> 01:08:44,340 on, quote, "things like people who were not making their"-- 1227 01:08:44,340 --> 01:08:48,830 if you just run an IV in this cross-section, 1228 01:08:48,830 --> 01:08:55,689 just be mindful of what you want to do with the occupation 1229 01:08:55,689 --> 01:08:57,340 categories. 1230 01:08:57,340 --> 01:09:05,180 Now this is where the rubber hits the road. 1231 01:09:05,180 --> 01:09:10,090 Am I just talking fancy theoretical models here? 1232 01:09:10,090 --> 01:09:11,500 I don't think so. 1233 01:09:11,500 --> 01:09:13,810 This is just the most innocuous kind 1234 01:09:13,810 --> 01:09:17,899 of unobserved heterogeneity. 1235 01:09:17,899 --> 01:09:21,149 And yet it tells you that IV is not a panacea. 1236 01:09:21,149 --> 01:09:23,540 You have to be very careful when you 1237 01:09:23,540 --> 01:09:25,397 use it to think about the questions 1238 01:09:25,397 --> 01:09:26,314 you're able to answer. 1239 01:09:37,319 --> 01:09:39,160 So let me just jump. 1240 01:09:41,870 --> 01:09:44,779 So, again, we have the model fortunately. 1241 01:09:44,779 --> 01:09:49,470 So we can generate any data that we want from the model 1242 01:09:49,470 --> 01:09:53,819 and talk about occupation choice or talk 1243 01:09:53,819 --> 01:09:56,280 about financial intermediation. 1244 01:09:56,280 --> 01:10:02,690 So, at the parameters we put in, which 1245 01:10:02,690 --> 01:10:07,880 have to do with that LEB thing you saw in the second lecture, 1246 01:10:07,880 --> 01:10:15,720 roughly, we can talk about moving people 1247 01:10:15,720 --> 01:10:19,590 in the model from autarky to financial intermediation, 1248 01:10:19,590 --> 01:10:22,190 in other words, changing their choice problem, 1249 01:10:22,190 --> 01:10:27,840 damning them to autarky, or allowing them to be 1250 01:10:27,840 --> 01:10:29,340 in the intermediated sector-- 1251 01:10:29,340 --> 01:10:32,280 that's the conceptual experiment-- 1252 01:10:32,280 --> 01:10:37,600 and looking at what happens to people by occupation. 1253 01:10:37,600 --> 01:10:41,100 So we've got guys who are wage earners before and after. 1254 01:10:41,100 --> 01:10:42,690 They're not moving. 1255 01:10:42,690 --> 01:10:46,610 And we kind of get their gains. 1256 01:10:46,610 --> 01:10:50,660 We've got people who move from the wage-earning category 1257 01:10:50,660 --> 01:10:56,492 under autarky to entrepreneurs under financial intermediation, 1258 01:10:56,492 --> 01:10:57,950 and you'd like to think about those 1259 01:10:57,950 --> 01:11:02,307 as the talented poor people who now are able to borrow. 1260 01:11:02,307 --> 01:11:03,890 And then there's this negative number. 1261 01:11:03,890 --> 01:11:05,307 And I looked at this this morning, 1262 01:11:05,307 --> 01:11:07,340 and I'm like, [? oh, ?] what's wrong, negative. 1263 01:11:07,340 --> 01:11:11,090 And then I remembered, oh yeah, the subsidy. 1264 01:11:11,090 --> 01:11:14,310 So the subsidy is in place here. 1265 01:11:14,310 --> 01:11:17,310 So you can actually get adverse selection, 1266 01:11:17,310 --> 01:11:21,290 if you want to call it that, of people who really shouldn't 1267 01:11:21,290 --> 01:11:24,110 be moving from wage earner to the entrepreneurial category. 1268 01:11:24,110 --> 01:11:26,720 And it's precisely those people that 1269 01:11:26,720 --> 01:11:29,330 get impacted on the margin. 1270 01:11:29,330 --> 01:11:31,860 These are local average treatment effects. 1271 01:11:31,860 --> 01:11:33,770 These are people who are induced to change 1272 01:11:33,770 --> 01:11:38,510 their behavior as a consequence of the program. 1273 01:11:38,510 --> 01:11:43,310 Then we've got people who are firms no matter what, 1274 01:11:43,310 --> 01:11:45,140 and they gain. 1275 01:11:45,140 --> 01:11:56,400 And, people who are from entrepreneur under autarky 1276 01:11:56,400 --> 01:11:59,340 to wage earner under intermediation, 1277 01:11:59,340 --> 01:12:03,480 so these are the not so talented guys who basically do 1278 01:12:03,480 --> 01:12:07,860 switch and do gain as a result of the program. 1279 01:12:07,860 --> 01:12:11,520 So, I mean, so you might think that, at plausible parameter 1280 01:12:11,520 --> 01:12:14,010 values, you wouldn't see any of these occupation switchers, 1281 01:12:14,010 --> 01:12:15,690 but you do. 1282 01:12:15,690 --> 01:12:17,220 And, again, the models that we've 1283 01:12:17,220 --> 01:12:20,520 been talking about sort of assume something like that's 1284 01:12:20,520 --> 01:12:21,070 going on. 1285 01:12:21,070 --> 01:12:24,250 That's where the TFP is coming from, except we 1286 01:12:24,250 --> 01:12:26,080 didn't have a subsidy. 1287 01:12:26,080 --> 01:12:29,020 Well, maybe China did. 1288 01:12:29,020 --> 01:12:30,720 Maybe we were removing the subsidy. 1289 01:12:35,120 --> 01:12:39,560 OK, this will be even shorter, but it's 1290 01:12:39,560 --> 01:12:41,420 a nice review in a way. 1291 01:12:41,420 --> 01:12:44,330 You remember that Greenwood and Jovanovic, 1292 01:12:44,330 --> 01:12:48,280 the forward-looking model? 1293 01:12:48,280 --> 01:12:53,080 People solve stochastic, dynamic decision problems 1294 01:12:53,080 --> 01:12:57,100 to maximize expected utility, choosing how much to save 1295 01:12:57,100 --> 01:12:59,050 and where to put their money in terms 1296 01:12:59,050 --> 01:13:01,045 of safe or risky activities or assets. 1297 01:13:03,713 --> 01:13:06,130 We're going to put a little heterogeneity in this discount 1298 01:13:06,130 --> 01:13:06,630 rate. 1299 01:13:09,140 --> 01:13:10,318 It could be other places. 1300 01:13:10,318 --> 01:13:12,110 Now we're just going to experiment and see. 1301 01:13:12,110 --> 01:13:17,930 So here let beta i be equal to some common average, beta bar, 1302 01:13:17,930 --> 01:13:24,810 and then some heterogeneity, unobserved heterogeneity, 1303 01:13:24,810 --> 01:13:26,920 that makes beta i different from beta bar. 1304 01:13:30,570 --> 01:13:33,550 And then we saw this optimization problem. 1305 01:13:33,550 --> 01:13:38,090 We have savings, fractions of wealth 1306 01:13:38,090 --> 01:13:41,710 to put in sort of safe or risky. 1307 01:13:41,710 --> 01:13:43,390 Or, in this case, I generalized it 1308 01:13:43,390 --> 01:13:48,310 to entrepreneurial and wage-earning activities. 1309 01:13:48,310 --> 01:13:50,680 You have aggregate shocks in both. 1310 01:13:50,680 --> 01:13:53,350 You have idiosyncratic shocks in both. 1311 01:13:53,350 --> 01:13:56,680 You have a law of motion for wealth. 1312 01:13:56,680 --> 01:14:01,300 Wealth today, fraction saved, fraction 1313 01:14:01,300 --> 01:14:06,820 put into risky and safe gets you wealth tomorrow. 1314 01:14:06,820 --> 01:14:10,110 We have two sectors here, intermediated 1315 01:14:10,110 --> 01:14:12,320 and non-intermediated. 1316 01:14:12,320 --> 01:14:15,610 The non-intermediated sector is like this financial autarky 1317 01:14:15,610 --> 01:14:17,090 sector. 1318 01:14:17,090 --> 01:14:20,610 So we're going to look for this value function, W0. 1319 01:14:23,920 --> 01:14:27,220 If you're in the intermediated sector-- 1320 01:14:27,220 --> 01:14:30,010 oh, sorry, well, I'll do the next one. 1321 01:14:30,010 --> 01:14:32,440 If you're in the intermediated sector, 1322 01:14:32,440 --> 01:14:35,630 you're looking for another value function that 1323 01:14:35,630 --> 01:14:39,120 solves dynamic optimization. 1324 01:14:39,120 --> 01:14:45,810 Now, for each one of these, if you were there and destined 1325 01:14:45,810 --> 01:14:49,920 to be there forever with these constant relative risk-averse 1326 01:14:49,920 --> 01:14:53,490 utility functions, you have some simple sort 1327 01:14:53,490 --> 01:14:58,470 of facts like the savings rates are constant. 1328 01:14:58,470 --> 01:15:03,702 Savings rates depend on beta, but give me the beta, 1329 01:15:03,702 --> 01:15:05,160 and I'll give you the savings rate. 1330 01:15:05,160 --> 01:15:06,720 And that's true in autarky, and it's 1331 01:15:06,720 --> 01:15:12,830 true as well in the financial intermediated sector, OK? 1332 01:15:17,300 --> 01:15:20,370 But, because it affects the savings rate, 1333 01:15:20,370 --> 01:15:22,980 it's going to affect income. 1334 01:15:22,980 --> 01:15:25,320 And so part of the unobserved error 1335 01:15:25,320 --> 01:15:28,850 is going to be the product of income with this theta i, 1336 01:15:28,850 --> 01:15:31,620 but theta i was that preference parameter we don't see. 1337 01:15:35,410 --> 01:15:37,580 You with me? 1338 01:15:37,580 --> 01:15:39,520 Hopefully, you can at least conceptually see 1339 01:15:39,520 --> 01:15:41,620 where we're going. 1340 01:15:41,620 --> 01:15:42,880 OK, and so there's-- 1341 01:15:42,880 --> 01:15:46,990 similarly, there's a savings rate and consumption 1342 01:15:46,990 --> 01:15:49,990 for the intermediated guys. 1343 01:15:49,990 --> 01:15:55,370 And then you have the big choice here. 1344 01:15:55,370 --> 01:15:56,900 This is Greenwood and Jovanovic. 1345 01:15:56,900 --> 01:16:00,890 So there's this cost Z of joining the financial system. 1346 01:16:00,890 --> 01:16:05,360 And, if this cost Z is sort of a random variable 1347 01:16:05,360 --> 01:16:08,240 in the population, it's going to affect the choice. 1348 01:16:08,240 --> 01:16:16,360 The issue is whether V is greater than W. 1349 01:16:16,360 --> 01:16:22,570 So, if all of this choosing were happening only 1350 01:16:22,570 --> 01:16:25,660 in the initial period at t equals 0, 1351 01:16:25,660 --> 01:16:28,570 then I've got a valid instrument. 1352 01:16:28,570 --> 01:16:32,590 I have something that affects the choice of intermediation, 1353 01:16:32,590 --> 01:16:35,710 even in this dynamic model. 1354 01:16:35,710 --> 01:16:38,560 And then you're in one sector or the other. 1355 01:16:38,560 --> 01:16:42,580 The savings rates depend on the error 1356 01:16:42,580 --> 01:16:45,520 term, which depends on theta, but this occupation choice 1357 01:16:45,520 --> 01:16:46,990 here-- 1358 01:16:46,990 --> 01:16:52,180 sorry, this this participation choice is instrumented with Z. 1359 01:16:52,180 --> 01:16:55,030 And Z has nothing to do with the unobserved preference 1360 01:16:55,030 --> 01:16:56,510 heterogeneity. 1361 01:16:56,510 --> 01:16:57,790 So I have a valid instrument. 1362 01:16:57,790 --> 01:17:01,900 I have monotonicity, OK? 1363 01:17:01,900 --> 01:17:05,710 However, that was not the way we dealt with that model 1364 01:17:05,710 --> 01:17:07,210 in the lecture. 1365 01:17:07,210 --> 01:17:11,030 We talked about endogenous financial deepening. 1366 01:17:11,030 --> 01:17:13,700 That makes households forward-looking, 1367 01:17:13,700 --> 01:17:16,610 planning, potentially, to join the financial system 1368 01:17:16,610 --> 01:17:18,380 at some future date. 1369 01:17:18,380 --> 01:17:23,530 Well, now we're in trouble, even with the same instrument, 1370 01:17:23,530 --> 01:17:27,760 because what we see prior to the selection 1371 01:17:27,760 --> 01:17:30,910 is the endogenous evolution of wealth. 1372 01:17:30,910 --> 01:17:35,200 And these guys will be saving up to cover those fixed costs. 1373 01:17:35,200 --> 01:17:37,660 It's no longer true that the savings rates are just 1374 01:17:37,660 --> 01:17:40,930 constant fractions of income. 1375 01:17:43,840 --> 01:17:46,290 What does it mean for RCTs? 1376 01:17:46,290 --> 01:17:48,810 It means you should try to keep the control 1377 01:17:48,810 --> 01:17:53,020 population in the dark about the program 1378 01:17:53,020 --> 01:17:54,490 because, if they know it's coming, 1379 01:17:54,490 --> 01:17:56,212 and they'll eventually get treated, 1380 01:17:56,212 --> 01:17:58,420 they're going to change their behavior if they behave 1381 01:17:58,420 --> 01:18:01,000 the way they do in the model. 1382 01:18:01,000 --> 01:18:05,980 So Ben surprises people with the road construction 1383 01:18:05,980 --> 01:18:09,100 when he dug up the dirt to see what the pavement was made of. 1384 01:18:11,720 --> 01:18:13,860 That happened after the fact. 1385 01:18:13,860 --> 01:18:16,350 He didn't tell them he was going to do beforehand 1386 01:18:16,350 --> 01:18:18,100 for obvious reasons. 1387 01:18:18,100 --> 01:18:21,565 But it's very hard in the field to maintain this 1388 01:18:21,565 --> 01:18:23,190 because people want to know why they're 1389 01:18:23,190 --> 01:18:25,190 answering all the questions in the questionnaire 1390 01:18:25,190 --> 01:18:26,760 and not getting the program. 1391 01:18:26,760 --> 01:18:29,960 They're not even supposed to know about the program. 1392 01:18:29,960 --> 01:18:35,610 You know, there are ways to try to fix it. 1393 01:18:35,610 --> 01:18:40,168 Now it's true that, if you had like a surprise, 1394 01:18:40,168 --> 01:18:42,210 even though they're following this dynamic model, 1395 01:18:42,210 --> 01:18:47,130 you say, oh, today is we got a special deal. 1396 01:18:47,130 --> 01:18:49,920 It's not going to be repeated tomorrow. 1397 01:18:49,920 --> 01:18:52,140 Then you have this one-time-off experiment. 1398 01:18:52,140 --> 01:18:56,040 You have this Z that's lowered that they got surprised, 1399 01:18:56,040 --> 01:18:59,010 that prior behavior isn't influenced by it, that will 1400 01:18:59,010 --> 01:19:01,613 induce selection and so on. 1401 01:19:01,613 --> 01:19:02,530 Well, what am I doing? 1402 01:19:02,530 --> 01:19:05,640 I'm just reviewing, in a way, what we've already done 1403 01:19:05,640 --> 01:19:06,810 with these kinds of models. 1404 01:19:06,810 --> 01:19:09,810 This one is dynamic, but putting it 1405 01:19:09,810 --> 01:19:14,460 in the language of when RCTs and instruments will work 1406 01:19:14,460 --> 01:19:17,890 and what to be kind of leery about. 1407 01:19:17,890 --> 01:19:23,090 I honestly don't see how you can talk about IV or RCTs, 1408 01:19:23,090 --> 01:19:25,660 random as they may be, without having 1409 01:19:25,660 --> 01:19:29,350 some framework in your head to think about the model. 1410 01:19:33,810 --> 01:19:37,320 And the rest I'll skip. 1411 01:19:37,320 --> 01:19:41,520 It's about that financial expansion in Spain. 1412 01:19:41,520 --> 01:19:45,390 I'll tell you the key insight the IO guys had, 1413 01:19:45,390 --> 01:19:48,420 which is, basically, you worry about an equilibrium 1414 01:19:48,420 --> 01:19:50,690 where everyone is anticipating everybody else. 1415 01:19:50,690 --> 01:19:53,760 You say, well, hey, if you believe it's an equilibrium, 1416 01:19:53,760 --> 01:19:56,430 then it ought to be in the data. 1417 01:19:56,430 --> 01:20:00,150 So you start running reduced form regressions 1418 01:20:00,150 --> 01:20:06,150 of how your competitors behave as a function of what you did. 1419 01:20:06,150 --> 01:20:08,010 And that's kind of the forecasting 1420 01:20:08,010 --> 01:20:13,550 aspect, which you can bring into the individual problems. 1421 01:20:13,550 --> 01:20:15,390 I know it sounds a bit counterintuitive, 1422 01:20:15,390 --> 01:20:17,790 but, basically, the reason it works 1423 01:20:17,790 --> 01:20:20,430 is because everyone is getting these shocks. 1424 01:20:20,430 --> 01:20:23,370 So no node-- you know, you can be at any node 1425 01:20:23,370 --> 01:20:26,310 and go to any other node because there's always 1426 01:20:26,310 --> 01:20:28,545 a shock that's going to rationalize that choice. 1427 01:20:31,050 --> 01:20:36,310 It's like Hutz and Sedlacek way back when. 1428 01:20:36,310 --> 01:20:40,410 And it's the foundation of these discrete choice models. 1429 01:20:40,410 --> 01:20:42,390 Now that doesn't always work. 1430 01:20:42,390 --> 01:20:44,730 And the last thing I won't even tell you 1431 01:20:44,730 --> 01:20:46,770 is a situation where it might fail. 1432 01:20:46,770 --> 01:20:49,860 So the point was not to drive nails 1433 01:20:49,860 --> 01:20:52,140 in the coffin of one way of doing business. 1434 01:20:52,140 --> 01:20:55,090 The point was to have an open conversation 1435 01:20:55,090 --> 01:20:58,060 and have everything on the table and so 1436 01:20:58,060 --> 01:21:00,910 to increase our sensitivity. 1437 01:21:00,910 --> 01:21:02,850 If you don't think heterogeneity is-- 1438 01:21:02,850 --> 01:21:05,610 unobserved heterogeneity is a big deal, 1439 01:21:05,610 --> 01:21:08,610 that would help you a lot, but, when 1440 01:21:08,610 --> 01:21:10,440 we looked at the impact of microcredit, 1441 01:21:10,440 --> 01:21:14,620 we saw a lot of heterogeneity in outcomes, 1442 01:21:14,620 --> 01:21:18,535 whether it was Hyderabad or Thailand or Morocco. 1443 01:21:18,535 --> 01:21:22,220 That seems to suggest this selection on unobservables 1444 01:21:22,220 --> 01:21:25,580 is quite important. 1445 01:21:25,580 --> 01:21:27,460 OK, good.