WEBVTT 00:00:00.000 --> 00:00:06.331 align:middle line:84% [SQUEAKING] [RUSTLING] [CLICKING] 00:00:06.331 --> 00:00:18.510 align:middle line:90% 00:00:18.510 --> 00:00:22.890 align:middle line:84% PROFESSOR: All right, so I have a fair amount 00:00:22.890 --> 00:00:27.120 align:middle line:84% to cover today, especially if I want to go all the way to what 00:00:27.120 --> 00:00:33.120 align:middle line:84% a lot of you were interested in in reading the Indonesia paper. 00:00:33.120 --> 00:00:36.900 align:middle line:84% I may not get to it, but we are going to do our best. 00:00:36.900 --> 00:00:39.570 align:middle line:84% Today, I want to talk about private and social returns 00:00:39.570 --> 00:00:40.530 align:middle line:90% to education. 00:00:40.530 --> 00:00:43.770 align:middle line:90% 00:00:43.770 --> 00:00:45.690 align:middle line:84% What are the private returns to education? 00:00:45.690 --> 00:00:47.410 align:middle line:84% Well, we discussed them a little bit. 00:00:47.410 --> 00:00:51.270 align:middle line:84% They are, on the one hand, the impact on the labor market, 00:00:51.270 --> 00:00:59.280 align:middle line:84% and on the other hand, the impact on making it easier 00:00:59.280 --> 00:01:01.290 align:middle line:84% to raise a future generation that itself 00:01:01.290 --> 00:01:03.240 align:middle line:84% has a lot of education, which we can 00:01:03.240 --> 00:01:06.060 align:middle line:84% take as a shortcut for human capital. 00:01:06.060 --> 00:01:10.710 align:middle line:84% And the private return also includes the happiness 00:01:10.710 --> 00:01:16.860 align:middle line:84% that comes out of educating your children. 00:01:16.860 --> 00:01:19.710 align:middle line:84% There is plenty of evidence that there 00:01:19.710 --> 00:01:22.800 align:middle line:84% is a strong correlation between human capital and earnings. 00:01:22.800 --> 00:01:26.400 align:middle line:84% One of these come, for example, from one 00:01:26.400 --> 00:01:30.780 align:middle line:84% of the descriptive [? first ?] stats in the Indonesia study, 00:01:30.780 --> 00:01:33.810 align:middle line:84% where we see that people who have more education 00:01:33.810 --> 00:01:39.450 align:middle line:84% make more money, potentially with some evidence of a jump 00:01:39.450 --> 00:01:40.980 align:middle line:90% towards middle school. 00:01:40.980 --> 00:01:43.710 align:middle line:84% Might even look like a s-curve if you squint. 00:01:43.710 --> 00:01:46.770 align:middle line:90% 00:01:46.770 --> 00:01:48.330 align:middle line:90% So that's there. 00:01:48.330 --> 00:01:51.060 align:middle line:84% There is similar evidence, by the way, on health. 00:01:51.060 --> 00:01:55.720 align:middle line:84% This is a correlation between people's body mass index 00:01:55.720 --> 00:01:58.620 align:middle line:84% or their height and their earnings. 00:01:58.620 --> 00:02:00.870 align:middle line:84% And you can also potentially-- you 00:02:00.870 --> 00:02:03.600 align:middle line:84% might be tempted to interpret this as the return 00:02:03.600 --> 00:02:05.230 align:middle line:90% to physical health. 00:02:05.230 --> 00:02:07.600 align:middle line:84% So we're going to be talking today about education, 00:02:07.600 --> 00:02:10.440 align:middle line:84% but a lot of what we are doing, both 00:02:10.440 --> 00:02:12.720 align:middle line:84% in terms of the intellectual content 00:02:12.720 --> 00:02:14.880 align:middle line:84% and the methodological content, can 00:02:14.880 --> 00:02:17.730 align:middle line:84% be found again thinking about the return to physical health 00:02:17.730 --> 00:02:19.950 align:middle line:90% instead. 00:02:19.950 --> 00:02:23.880 align:middle line:84% And then now, also evidence of all sorts of other health 00:02:23.880 --> 00:02:25.650 align:middle line:84% benefits of health and education, which 00:02:25.650 --> 00:02:29.310 align:middle line:84% could be represented as forms of consumption, 00:02:29.310 --> 00:02:32.910 align:middle line:84% like how healthy your own kids are, 00:02:32.910 --> 00:02:37.000 align:middle line:90% political attitude, et cetera. 00:02:37.000 --> 00:02:38.740 align:middle line:84% There is also evidence of interaction 00:02:38.740 --> 00:02:41.480 align:middle line:84% between different forms of human capital. 00:02:41.480 --> 00:02:45.880 align:middle line:84% So educated people are healthier. 00:02:45.880 --> 00:02:48.910 align:middle line:84% The children of educated people are healthier. 00:02:48.910 --> 00:02:50.920 align:middle line:84% Healthier children means fewer days of school. 00:02:50.920 --> 00:02:53.170 align:middle line:84% And then they earn more as adults. 00:02:53.170 --> 00:02:55.060 align:middle line:84% For example, when you deworm children, 00:02:55.060 --> 00:02:57.220 align:middle line:90% they miss fewer days of schools. 00:02:57.220 --> 00:02:58.300 align:middle line:90% They earn more as adults. 00:02:58.300 --> 00:03:00.008 align:middle line:84% I think we saw that in the first lecture. 00:03:00.008 --> 00:03:02.950 align:middle line:90% 00:03:02.950 --> 00:03:05.050 align:middle line:84% So if we go back to the basic correlation, 00:03:05.050 --> 00:03:06.610 align:middle line:90% what's the basic problem? 00:03:06.610 --> 00:03:08.890 align:middle line:84% Why was I tentative when I was looking 00:03:08.890 --> 00:03:11.890 align:middle line:84% at showing you these graphs, in interpreting these 00:03:11.890 --> 00:03:13.510 align:middle line:90% as returns, necessarily? 00:03:13.510 --> 00:03:16.180 align:middle line:90% 00:03:16.180 --> 00:03:18.760 align:middle line:84% And that is the causal effect of human capital, 00:03:18.760 --> 00:03:23.395 align:middle line:84% say, on earnings, education, or height, or BMI. 00:03:23.395 --> 00:03:26.518 align:middle line:90% 00:03:26.518 --> 00:03:27.940 align:middle line:90% Yeah, [? Karesh? ?] 00:03:27.940 --> 00:03:29.410 align:middle line:90% AUDIENCE: Selection issues. 00:03:29.410 --> 00:03:31.535 align:middle line:84% ESTHER DUFLO: Yeah, there could be selection issues 00:03:31.535 --> 00:03:35.890 align:middle line:84% of various kinds, ability bias, omitted valuable of the kind 00:03:35.890 --> 00:03:37.300 align:middle line:84% that people who have enough money 00:03:37.300 --> 00:03:39.850 align:middle line:84% to feed their children also have enough money 00:03:39.850 --> 00:03:44.200 align:middle line:84% to make sure that they get the right apprenticeship. 00:03:44.200 --> 00:03:46.690 align:middle line:84% People who live in towns will have better schools. 00:03:46.690 --> 00:03:49.377 align:middle line:84% They will also have better job opportunities, 00:03:49.377 --> 00:03:50.335 align:middle line:90% and so on and so forth. 00:03:50.335 --> 00:03:53.380 align:middle line:90% 00:03:53.380 --> 00:03:56.070 align:middle line:84% Another issue that we want-- so that's basically 00:03:56.070 --> 00:03:58.500 align:middle line:84% what we are going to try and tackle 00:03:58.500 --> 00:04:03.810 align:middle line:84% today, what a fair amount of the literature is concerned by. 00:04:03.810 --> 00:04:07.740 align:middle line:84% Another issue that comes up very much in the Ghana paper 00:04:07.740 --> 00:04:11.740 align:middle line:84% is the distinction between private and social returns. 00:04:11.740 --> 00:04:14.460 align:middle line:84% So the private returns are for an individual, 00:04:14.460 --> 00:04:17.375 align:middle line:84% how much money they make with or without education. 00:04:17.375 --> 00:04:19.500 align:middle line:84% And in fact, from the point of view of the parents, 00:04:19.500 --> 00:04:23.190 align:middle line:84% you could add a layer of private estimated return, which 00:04:23.190 --> 00:04:26.400 align:middle line:84% is how much they think the child is going to make 00:04:26.400 --> 00:04:28.530 align:middle line:90% with or without education. 00:04:28.530 --> 00:04:30.370 align:middle line:84% And then there are the social returns, 00:04:30.370 --> 00:04:33.480 align:middle line:84% which is the value for society of a particular individual 00:04:33.480 --> 00:04:35.220 align:middle line:90% getting more education. 00:04:35.220 --> 00:04:37.680 align:middle line:84% And so what are the sources of the discrepancy 00:04:37.680 --> 00:04:40.500 align:middle line:84% between private and social returns 00:04:40.500 --> 00:04:42.660 align:middle line:90% that we might worry about? 00:04:42.660 --> 00:04:46.148 align:middle line:90% 00:04:46.148 --> 00:04:47.690 align:middle line:84% AUDIENCE: In the Ghana paper, there's 00:04:47.690 --> 00:04:51.077 align:middle line:84% the displacement effect, if some people get the government jobs, 00:04:51.077 --> 00:04:52.490 align:middle line:90% and other people get displaced. 00:04:52.490 --> 00:04:53.407 align:middle line:90% ESTHER DUFLO: Exactly. 00:04:53.407 --> 00:04:55.190 align:middle line:84% So there can be displacement effect. 00:04:55.190 --> 00:04:58.220 align:middle line:84% In the Ghana paper, it's pretty obvious 00:04:58.220 --> 00:05:02.240 align:middle line:84% because education is used as a ticket for rent-seeking jobs. 00:05:02.240 --> 00:05:03.990 align:middle line:84% And then there are only so many of those. 00:05:03.990 --> 00:05:05.930 align:middle line:84% It's not that you can create more. 00:05:05.930 --> 00:05:09.260 align:middle line:84% You could, to some extent, if the educated people who 00:05:09.260 --> 00:05:13.718 align:middle line:84% are so much better at the job of running the government-- you 00:05:13.718 --> 00:05:16.010 align:middle line:84% could increase the size of the government a little bit. 00:05:16.010 --> 00:05:18.680 align:middle line:84% But the truth is that the government is probably 00:05:18.680 --> 00:05:22.070 align:middle line:90% mostly rent-seeking jobs. 00:05:22.070 --> 00:05:23.820 align:middle line:84% There are only so many that can happen. 00:05:23.820 --> 00:05:26.910 align:middle line:84% So it's basically a one-for-one displacement. 00:05:26.910 --> 00:05:31.140 align:middle line:84% And I don't think that's an exception. 00:05:31.140 --> 00:05:33.530 align:middle line:84% I don't think that's a special case in Ghana. 00:05:33.530 --> 00:05:37.940 align:middle line:84% I think this idea that we see in the Ghana paper-- 00:05:37.940 --> 00:05:40.070 align:middle line:84% I'll go back to that, but I'll show you the graph-- 00:05:40.070 --> 00:05:45.380 align:middle line:84% that we see in the Ghana paper, where people think that someone 00:05:45.380 --> 00:05:48.620 align:middle line:84% with education, what they might be doing 00:05:48.620 --> 00:05:54.410 align:middle line:84% is to get a government job very likely. 00:05:54.410 --> 00:05:57.020 align:middle line:84% And that is one key reason why people are 00:05:57.020 --> 00:05:59.300 align:middle line:90% looking to educate their kids. 00:05:59.300 --> 00:06:02.510 align:middle line:84% Because there are usually restrictions on whether 00:06:02.510 --> 00:06:05.720 align:middle line:84% you can even apply as a function of your education degree. 00:06:05.720 --> 00:06:07.520 align:middle line:84% You need maybe a tertiary degree for which 00:06:07.520 --> 00:06:09.560 align:middle line:90% you need a secondary degree. 00:06:09.560 --> 00:06:11.420 align:middle line:84% Or you need just a secondary degree. 00:06:11.420 --> 00:06:13.610 align:middle line:84% This is something that we are seeing. 00:06:13.610 --> 00:06:15.950 align:middle line:84% This queue for the government jobs 00:06:15.950 --> 00:06:19.880 align:middle line:84% is something that we are seeing in country after country 00:06:19.880 --> 00:06:22.730 align:middle line:84% after country in the developing world. 00:06:22.730 --> 00:06:25.550 align:middle line:84% In India, there is a very interesting fact, which 00:06:25.550 --> 00:06:30.110 align:middle line:84% is that you can only apply until you're 30, 00:06:30.110 --> 00:06:32.660 align:middle line:90% I believe-- maybe 35. 00:06:32.660 --> 00:06:34.380 align:middle line:84% The specific date doesn't matter. 00:06:34.380 --> 00:06:35.990 align:middle line:90% Let's say 30. 00:06:35.990 --> 00:06:39.680 align:middle line:84% And Abhijit and [INAUDIBLE] have a paper 00:06:39.680 --> 00:06:42.650 align:middle line:84% where they show that the unemployment rate falls 00:06:42.650 --> 00:06:45.560 align:middle line:90% dramatically exactly at 30. 00:06:45.560 --> 00:06:47.570 align:middle line:84% So people basically until 30 still 00:06:47.570 --> 00:06:49.490 align:middle line:90% hope that they might get a job. 00:06:49.490 --> 00:06:51.950 align:middle line:84% They're holding out for that job. 00:06:51.950 --> 00:06:53.610 align:middle line:84% Live with a parent, figure it out. 00:06:53.610 --> 00:06:55.943 align:middle line:84% And then when it's clear that they can't get it anymore, 00:06:55.943 --> 00:06:57.270 align:middle line:90% then they go and get a job. 00:06:57.270 --> 00:07:01.130 align:middle line:84% So these queues for those government jobs, 00:07:01.130 --> 00:07:05.570 align:middle line:84% I've come to believe, has a very destructive impact 00:07:05.570 --> 00:07:06.770 align:middle line:90% on the whole labor market. 00:07:06.770 --> 00:07:10.680 align:middle line:84% And in the Ghana paper, we are seeing them all over the place. 00:07:10.680 --> 00:07:14.160 align:middle line:84% So that's an extreme example of displacement. 00:07:14.160 --> 00:07:19.400 align:middle line:84% But in the absence of government jobs, 00:07:19.400 --> 00:07:22.730 align:middle line:84% could you nonetheless have displacement effects 00:07:22.730 --> 00:07:24.620 align:middle line:84% in the absence of government, or in fact 00:07:24.620 --> 00:07:29.150 align:middle line:84% any rent-seeking opportunity with a perfectly competitive 00:07:29.150 --> 00:07:30.050 align:middle line:90% labor market? 00:07:30.050 --> 00:07:32.340 align:middle line:84% Would you nonetheless potentially see 00:07:32.340 --> 00:07:33.215 align:middle line:90% displacement effects? 00:07:33.215 --> 00:07:42.260 align:middle line:90% 00:07:42.260 --> 00:07:43.070 align:middle line:90% Yeah. 00:07:43.070 --> 00:07:44.487 align:middle line:84% AUDIENCE: [INAUDIBLE] displacement 00:07:44.487 --> 00:07:47.175 align:middle line:90% with tertiary [INAUDIBLE]? 00:07:47.175 --> 00:07:49.050 align:middle line:84% ESTHER DUFLO: Yes, so there is that for sure. 00:07:49.050 --> 00:07:52.350 align:middle line:84% There might be a ripple effect on tertiary education. 00:07:52.350 --> 00:07:57.810 align:middle line:84% But even if we consider the labor market and just the fact 00:07:57.810 --> 00:08:03.030 align:middle line:84% that there are more educated workers in the labor market, 00:08:03.030 --> 00:08:05.250 align:middle line:84% might that create displacement of other people? 00:08:05.250 --> 00:08:09.700 align:middle line:90% 00:08:09.700 --> 00:08:11.551 align:middle line:90% Yeah? 00:08:11.551 --> 00:08:14.626 align:middle line:84% AUDIENCE: In standard search mode in labor, 00:08:14.626 --> 00:08:18.000 align:middle line:84% these [INAUDIBLE] people are searching for the same job. 00:08:18.000 --> 00:08:20.773 align:middle line:84% It should be pretty easy job [INAUDIBLE] for everyone else. 00:08:20.773 --> 00:08:21.690 align:middle line:90% ESTHER DUFLO: Exactly. 00:08:21.690 --> 00:08:31.650 align:middle line:84% And would we find that it's very unusual labor markets, or type 00:08:31.650 --> 00:08:33.000 align:middle line:90% of modeling the labor market? 00:08:33.000 --> 00:08:39.828 align:middle line:84% Or would we expect this to happen fairly easily? 00:08:39.828 --> 00:08:40.622 align:middle line:90% Yeah, [? Aisha. ?] 00:08:40.622 --> 00:08:42.289 align:middle line:84% AUDIENCE: It's pretty standard, I think. 00:08:42.289 --> 00:08:43.747 align:middle line:84% ESTHER DUFLO: It's pretty standard. 00:08:43.747 --> 00:08:48.400 align:middle line:84% It's going to depend, of course, on the production function. 00:08:48.400 --> 00:08:50.800 align:middle line:84% So one case where you're obviously going to find it 00:08:50.800 --> 00:08:55.760 align:middle line:84% is if people are working with a limited resource-- for example, 00:08:55.760 --> 00:08:56.770 align:middle line:90% land. 00:08:56.770 --> 00:08:58.460 align:middle line:90% And there's only so much of it. 00:08:58.460 --> 00:09:04.120 align:middle line:84% And if educated workers are more productive, 00:09:04.120 --> 00:09:07.720 align:middle line:84% there is going to be a substitution between educated 00:09:07.720 --> 00:09:10.730 align:middle line:84% workers and workers in general, educated or not. 00:09:10.730 --> 00:09:15.460 align:middle line:84% So we have-- marginally, take any production function. 00:09:15.460 --> 00:09:18.190 align:middle line:84% There is going to be, depending on the form, 00:09:18.190 --> 00:09:20.410 align:middle line:84% the parameters of the CES function, 00:09:20.410 --> 00:09:22.900 align:middle line:84% or the Cobb-Douglas glass function of whatever it is that 00:09:22.900 --> 00:09:25.690 align:middle line:84% you choose that is appropriate to model the labor market-- 00:09:25.690 --> 00:09:28.290 align:middle line:84% there is always degrees of substitutability 00:09:28.290 --> 00:09:30.950 align:middle line:90% between workers. 00:09:30.950 --> 00:09:34.180 align:middle line:84% And then that subsitutability is worse 00:09:34.180 --> 00:09:36.190 align:middle line:90% when there is a limiting factor. 00:09:36.190 --> 00:09:40.000 align:middle line:84% Land is an extreme setup because there is only so much of it. 00:09:40.000 --> 00:09:45.100 align:middle line:84% But in capital also, it depends how well the capital market 00:09:45.100 --> 00:09:49.300 align:middle line:84% functions and adjusts to the arrival of these new workers. 00:09:49.300 --> 00:09:55.070 align:middle line:84% If, for example, as [INAUDIBLE] was working on before, 00:09:55.070 --> 00:09:57.100 align:middle line:90% eventually came to MIT. 00:09:57.100 --> 00:10:01.310 align:middle line:84% The labor market is slow to move from one place to another. 00:10:01.310 --> 00:10:04.660 align:middle line:84% Sorry, capital is slow to move from one place to another. 00:10:04.660 --> 00:10:07.690 align:middle line:84% Then there is not an immediate adjustment 00:10:07.690 --> 00:10:09.280 align:middle line:84% of the demand of the labor demand 00:10:09.280 --> 00:10:12.010 align:middle line:84% to the new effective labor supply for your more productive 00:10:12.010 --> 00:10:13.240 align:middle line:90% worker. 00:10:13.240 --> 00:10:15.100 align:middle line:84% Then whether or not the externality-- 00:10:15.100 --> 00:10:18.550 align:middle line:84% who the incidence of the externality is going to be on, 00:10:18.550 --> 00:10:21.020 align:middle line:84% again, depends on the production function. 00:10:21.020 --> 00:10:25.420 align:middle line:84% So you could have models where educated and uneducated workers 00:10:25.420 --> 00:10:28.120 align:middle line:84% are perfectly substitute, except the educated worker do 00:10:28.120 --> 00:10:31.040 align:middle line:84% everything a little faster, in which case, 00:10:31.040 --> 00:10:35.300 align:middle line:84% the educated workers are going to be displacing everyone, 00:10:35.300 --> 00:10:37.540 align:middle line:84% including the uneducated workers. 00:10:37.540 --> 00:10:40.510 align:middle line:84% If, on the other hand, the educated and uneducated workers 00:10:40.510 --> 00:10:43.720 align:middle line:84% are not perfectly substitute, or are very imperfectly 00:10:43.720 --> 00:10:45.740 align:middle line:84% substitute, then the educated workers 00:10:45.740 --> 00:10:49.450 align:middle line:84% are going to push out the other educated workers. 00:10:49.450 --> 00:10:51.070 align:middle line:84% And what you're going to get is not 00:10:51.070 --> 00:10:53.980 align:middle line:84% the decline of the wage of the base wage, 00:10:53.980 --> 00:10:56.920 align:middle line:84% but the decline of the returns to education, 00:10:56.920 --> 00:11:00.110 align:middle line:84% the differential between educated and uneducated worker. 00:11:00.110 --> 00:11:02.720 align:middle line:84% So this is what Richard Friedman now many, 00:11:02.720 --> 00:11:04.630 align:middle line:84% many decades ago was worried about when 00:11:04.630 --> 00:11:07.570 align:middle line:84% he wrote this book, The Overeducated American. 00:11:07.570 --> 00:11:10.000 align:middle line:84% It's that a lot of people would go to school. 00:11:10.000 --> 00:11:12.830 align:middle line:84% And then they were, at this point, going to college. 00:11:12.830 --> 00:11:14.080 align:middle line:90% And they would acquire skills. 00:11:14.080 --> 00:11:18.280 align:middle line:84% And then the return to education would collapse. 00:11:18.280 --> 00:11:21.160 align:middle line:84% Didn't happen in the US, but that is a concern. 00:11:21.160 --> 00:11:23.560 align:middle line:84% That's in the back of people's minds, something 00:11:23.560 --> 00:11:27.100 align:middle line:84% that Heckman also, for example, is talking about. 00:11:27.100 --> 00:11:30.850 align:middle line:84% And those are equilibrium effect, or general equilibrium 00:11:30.850 --> 00:11:31.690 align:middle line:90% effect of education. 00:11:31.690 --> 00:11:36.130 align:middle line:84% So there is a basic, very standard reason 00:11:36.130 --> 00:11:38.200 align:middle line:84% for the social and the private return 00:11:38.200 --> 00:11:44.440 align:middle line:84% not to align, even in the absence of very uncompetitive 00:11:44.440 --> 00:11:46.970 align:middle line:90% sectors, like the labor market. 00:11:46.970 --> 00:11:48.670 align:middle line:90% So that's that. 00:11:48.670 --> 00:11:52.030 align:middle line:84% On the other hand, of course, what 00:11:52.030 --> 00:11:54.490 align:middle line:84% an institution like the World Bank 00:11:54.490 --> 00:11:59.860 align:middle line:84% was mainly motivated by when they were thinking of investing 00:11:59.860 --> 00:12:02.230 align:middle line:84% in education-- there are the positive spillovers 00:12:02.230 --> 00:12:06.670 align:middle line:84% of educated workers coming from easier adoption 00:12:06.670 --> 00:12:08.380 align:middle line:90% of technologies, for example. 00:12:08.380 --> 00:12:10.270 align:middle line:84% That might make everyone more productive 00:12:10.270 --> 00:12:14.030 align:middle line:84% and hence increase everyone's productivity wages 00:12:14.030 --> 00:12:16.070 align:middle line:90% and so on and so forth. 00:12:16.070 --> 00:12:22.160 align:middle line:84% So that's from the A in the production function. 00:12:22.160 --> 00:12:24.595 align:middle line:84% And then there could be impacts in other spheres. 00:12:24.595 --> 00:12:27.580 align:middle line:90% 00:12:27.580 --> 00:12:32.380 align:middle line:84% An educated population might support institutions 00:12:32.380 --> 00:12:35.620 align:middle line:84% that themselves lead to better policies, which again, would 00:12:35.620 --> 00:12:37.240 align:middle line:90% increase one layer above-- 00:12:37.240 --> 00:12:41.150 align:middle line:84% improve the institution, therefore improve education. 00:12:41.150 --> 00:12:45.550 align:middle line:84% So those would be the reasons for the social return 00:12:45.550 --> 00:12:47.770 align:middle line:90% to go in the other direction. 00:12:47.770 --> 00:12:52.720 align:middle line:84% That might be a little hard to pick up 00:12:52.720 --> 00:12:55.570 align:middle line:84% because if they are the very macro level, 00:12:55.570 --> 00:12:57.830 align:middle line:84% we don't have so many countries to even think of it. 00:12:57.830 --> 00:13:00.160 align:middle line:84% But people have looked, at least, 00:13:00.160 --> 00:13:02.120 align:middle line:84% to see whether it is true, for example, 00:13:02.120 --> 00:13:05.410 align:middle line:84% that more educated workers support more democracy, 00:13:05.410 --> 00:13:06.530 align:middle line:90% or so on and so forth. 00:13:06.530 --> 00:13:08.200 align:middle line:84% I'm not going to do much of that today. 00:13:08.200 --> 00:13:10.680 align:middle line:84% But just keep it in the back of your mind 00:13:10.680 --> 00:13:16.250 align:middle line:84% that that's an important, that's an interesting topic. 00:13:16.250 --> 00:13:18.700 align:middle line:84% So other point that's interesting in Ghana 00:13:18.700 --> 00:13:22.750 align:middle line:84% is, of course, people have a completely wild overestimate 00:13:22.750 --> 00:13:29.050 align:middle line:84% of how much of a ticket the secondary education is 00:13:29.050 --> 00:13:33.430 align:middle line:84% to tertiary degrees and hence the government. 00:13:33.430 --> 00:13:36.520 align:middle line:84% They vastly overestimate, as a result, the probability 00:13:36.520 --> 00:13:39.580 align:middle line:84% that sending their kids to secondary school 00:13:39.580 --> 00:13:43.390 align:middle line:84% would give them access to this job. 00:13:43.390 --> 00:13:45.490 align:middle line:84% And therefore, they vastly estimate the returns 00:13:45.490 --> 00:13:47.170 align:middle line:84% to education, which in their opinion, 00:13:47.170 --> 00:13:51.130 align:middle line:84% comes mostly from one of these jobs. 00:13:51.130 --> 00:13:54.100 align:middle line:84% So where they are correct is that these jobs are hugely 00:13:54.100 --> 00:13:56.710 align:middle line:90% valuable once you get them. 00:13:56.710 --> 00:13:58.810 align:middle line:84% They are fantastic, very well paid. 00:13:58.810 --> 00:14:02.740 align:middle line:90% We saw during COVID-- 00:14:02.740 --> 00:14:05.140 align:middle line:84% very protective against shocks because you're 00:14:05.140 --> 00:14:06.730 align:middle line:90% paid no matter what. 00:14:06.730 --> 00:14:08.620 align:middle line:84% But where they are wrong is that they 00:14:08.620 --> 00:14:11.590 align:middle line:84% are pretty hard to get even after education. 00:14:11.590 --> 00:14:14.460 align:middle line:90% 00:14:14.460 --> 00:14:20.940 align:middle line:90% So one way to get across-- 00:14:20.940 --> 00:14:22.620 align:middle line:84% we are going to look at two ways today 00:14:22.620 --> 00:14:27.210 align:middle line:84% to get around the problem of selection into education. 00:14:27.210 --> 00:14:30.120 align:middle line:90% One is random assignment. 00:14:30.120 --> 00:14:35.670 align:middle line:84% And the other is the school construction building 00:14:35.670 --> 00:14:38.430 align:middle line:84% that provides a natural experiment for increasing 00:14:38.430 --> 00:14:39.510 align:middle line:90% education. 00:14:39.510 --> 00:14:43.270 align:middle line:84% And both have advantages and disadvantages. 00:14:43.270 --> 00:14:46.080 align:middle line:84% So of course, it's hard to directly 00:14:46.080 --> 00:14:48.480 align:middle line:90% randomly assign education. 00:14:48.480 --> 00:14:51.990 align:middle line:84% We were able to randomly assign an [INAUDIBLE] program 00:14:51.990 --> 00:14:55.410 align:middle line:84% and compare people who get it to people who don't get it. 00:14:55.410 --> 00:14:58.863 align:middle line:84% Even then we had some people who managed to not comply 00:14:58.863 --> 00:15:00.030 align:middle line:90% to our treatment assignment. 00:15:00.030 --> 00:15:03.300 align:middle line:84% But at least, this is something we could manipulate directly. 00:15:03.300 --> 00:15:06.870 align:middle line:84% But education-- you can't really force people 00:15:06.870 --> 00:15:08.070 align:middle line:90% not to get an education. 00:15:08.070 --> 00:15:12.390 align:middle line:84% You can't force them to get an education, either. 00:15:12.390 --> 00:15:17.474 align:middle line:84% So what's randomly assigned in Ghana? 00:15:17.474 --> 00:15:20.000 align:middle line:84% Tends to be-- education is hard to randomly assign. 00:15:20.000 --> 00:15:23.700 align:middle line:90% 00:15:23.700 --> 00:15:27.380 align:middle line:90% Right side [INAUDIBLE]. 00:15:27.380 --> 00:15:28.440 align:middle line:90% AUDIENCE: Scholarships? 00:15:28.440 --> 00:15:30.380 align:middle line:84% ESTHER DUFLO: It's a scholarship to attend 00:15:30.380 --> 00:15:32.960 align:middle line:90% school, secondary school. 00:15:32.960 --> 00:15:35.390 align:middle line:90% In what sample? 00:15:35.390 --> 00:15:38.070 align:middle line:84% AUDIENCE: People who tested high enough 00:15:38.070 --> 00:15:40.990 align:middle line:84% to go to secondary school, but for some reason, 00:15:40.990 --> 00:15:42.520 align:middle line:90% saying they could not afford it. 00:15:42.520 --> 00:15:43.312 align:middle line:90% ESTHER DUFLO: Yeah. 00:15:43.312 --> 00:15:44.830 align:middle line:84% In fact, more than claimed, right? 00:15:44.830 --> 00:15:46.663 align:middle line:84% Not only they claimed they could not afford, 00:15:46.663 --> 00:15:47.706 align:middle line:90% but they didn't show up. 00:15:47.706 --> 00:15:48.520 align:middle line:90% AUDIENCE: Yeah. 00:15:48.520 --> 00:15:49.937 align:middle line:84% ESTHER DUFLO: The first trimester. 00:15:49.937 --> 00:15:53.020 align:middle line:84% So they took some costly steps to demonstrate that they 00:15:53.020 --> 00:15:54.160 align:middle line:90% couldn't. 00:15:54.160 --> 00:15:56.290 align:middle line:84% And they could discuss this step when there 00:15:56.290 --> 00:15:57.590 align:middle line:90% was no scholarship in mind. 00:15:57.590 --> 00:16:00.460 align:middle line:84% So presumably, they were close to could not afford. 00:16:00.460 --> 00:16:00.970 align:middle line:90% Exactly. 00:16:00.970 --> 00:16:05.560 align:middle line:84% So therefore, if we say in word, what can we confidently 00:16:05.560 --> 00:16:16.220 align:middle line:84% estimate by our standard methods of analyzing 00:16:16.220 --> 00:16:19.835 align:middle line:84% randomized controlled trial in table 2 of our paper? 00:16:19.835 --> 00:16:26.200 align:middle line:90% 00:16:26.200 --> 00:16:27.450 align:middle line:90% [? Shaina? ?] 00:16:27.450 --> 00:16:31.134 align:middle line:84% AUDIENCE: [INAUDIBLE] the scholarship on education? 00:16:31.134 --> 00:16:32.050 align:middle line:90% ESTHER DUFLO: For? 00:16:32.050 --> 00:16:35.460 align:middle line:84% AUDIENCE: For the people who tested but don't go. 00:16:35.460 --> 00:16:36.210 align:middle line:90% ESTHER DUFLO: Yes. 00:16:36.210 --> 00:16:40.290 align:middle line:84% So we can estimate the effect of getting secondary school 00:16:40.290 --> 00:16:45.510 align:middle line:84% for people who qualify on pretty hard exams, 00:16:45.510 --> 00:16:52.110 align:middle line:84% but could not afford to go in the absence of the program. 00:16:52.110 --> 00:16:56.590 align:middle line:84% So we had some discussion about external validity later on. 00:16:56.590 --> 00:16:59.670 align:middle line:90% That's a pretty specific sample. 00:16:59.670 --> 00:17:03.630 align:middle line:84% It's a sample of interest because, for example, it's 00:17:03.630 --> 00:17:07.530 align:middle line:84% a sample that would be relevant if you were considering 00:17:07.530 --> 00:17:10.319 align:middle line:84% making secondary education free, but not changing 00:17:10.319 --> 00:17:17.069 align:middle line:84% the requirement to get in, which is precisely what the Ghana 00:17:17.069 --> 00:17:20.550 align:middle line:84% government did shortly after the study started 00:17:20.550 --> 00:17:22.008 align:middle line:90% without asking us. 00:17:22.008 --> 00:17:23.550 align:middle line:84% So it's not something we could claim. 00:17:23.550 --> 00:17:27.990 align:middle line:84% But that was, at least at the time, 00:17:27.990 --> 00:17:30.810 align:middle line:90% very much the debate in India. 00:17:30.810 --> 00:17:33.000 align:middle line:84% We are not getting very necessarily very informative 00:17:33.000 --> 00:17:34.950 align:middle line:90% estimate to what will happen. 00:17:34.950 --> 00:17:37.350 align:middle line:84% Instead, once they have done it now, 00:17:37.350 --> 00:17:40.720 align:middle line:84% well, there are two debates in Ghana currently. 00:17:40.720 --> 00:17:42.540 align:middle line:84% There's one party that says we want 00:17:42.540 --> 00:17:46.890 align:middle line:84% to revert to what we had before, remove free secondary school. 00:17:46.890 --> 00:17:50.580 align:middle line:84% And then there is the party who did free secondary school who 00:17:50.580 --> 00:17:56.950 align:middle line:84% wants to go further and eliminate the exam requirement. 00:17:56.950 --> 00:18:01.630 align:middle line:84% So if we consider eliminating the exam requirement, 00:18:01.630 --> 00:18:05.100 align:middle line:84% these estimates might not be very informative on them 00:18:05.100 --> 00:18:07.110 align:middle line:84% because that's a different group of people. 00:18:07.110 --> 00:18:10.500 align:middle line:84% This might be a group of people for whom returns to education 00:18:10.500 --> 00:18:11.490 align:middle line:90% are higher or lower. 00:18:11.490 --> 00:18:13.080 align:middle line:90% We don't know. 00:18:13.080 --> 00:18:15.340 align:middle line:90% So we have to keep that in mind. 00:18:15.340 --> 00:18:18.400 align:middle line:84% So we can confidently estimate that. 00:18:18.400 --> 00:18:24.150 align:middle line:84% And so these are your standard, basic way 00:18:24.150 --> 00:18:26.370 align:middle line:84% of looking at an experiment table, which 00:18:26.370 --> 00:18:30.390 align:middle line:84% you will come to know and love, or at least learn 00:18:30.390 --> 00:18:34.860 align:middle line:90% to read and maybe write. 00:18:34.860 --> 00:18:39.300 align:middle line:84% Usually, you write the mean in the comparison group, 00:18:39.300 --> 00:18:45.090 align:middle line:84% the treatment effect, the P-value of the treatment 00:18:45.090 --> 00:18:46.800 align:middle line:90% effect, the AR. 00:18:46.800 --> 00:18:49.980 align:middle line:84% And we eventually could have done away with stars 00:18:49.980 --> 00:18:50.940 align:middle line:90% to my great delight. 00:18:50.940 --> 00:18:52.410 align:middle line:90% And not just delight-- 00:18:52.410 --> 00:18:55.720 align:middle line:84% I made that happen because I hate stars. 00:18:55.720 --> 00:18:56.220 align:middle line:90% [LAUGHTER] 00:18:56.220 --> 00:19:00.060 align:middle line:84% This doesn't make any sense that 10% or 5% 00:19:00.060 --> 00:19:01.560 align:middle line:90% become magical numbers. 00:19:01.560 --> 00:19:03.090 align:middle line:90% It's annoying. 00:19:03.090 --> 00:19:05.640 align:middle line:84% So instead, you can report P-value 00:19:05.640 --> 00:19:07.230 align:middle line:84% if you think people shouldn't have 00:19:07.230 --> 00:19:09.400 align:middle line:90% to make division in their head. 00:19:09.400 --> 00:19:13.200 align:middle line:84% And so that's the comparison mean. 00:19:13.200 --> 00:19:14.430 align:middle line:90% That's the treatment effect. 00:19:14.430 --> 00:19:17.340 align:middle line:84% If you added them both, if there are no control variables, 00:19:17.340 --> 00:19:20.740 align:middle line:84% that's the education in the treatment group. 00:19:20.740 --> 00:19:30.030 align:middle line:84% And so the program was effective at increasing completion 00:19:30.030 --> 00:19:31.080 align:middle line:90% of secondary school. 00:19:31.080 --> 00:19:32.880 align:middle line:90% Some people went anyways. 00:19:32.880 --> 00:19:35.700 align:middle line:84% So we're going to keep that in mind. 00:19:35.700 --> 00:19:40.140 align:middle line:84% Almost half of the comparison group goes to secondary school 00:19:40.140 --> 00:19:44.250 align:middle line:84% and completes it anyways, not because they were lying 00:19:44.250 --> 00:19:47.340 align:middle line:84% about their inability to afford secondary school [? tuition, ?] 00:19:47.340 --> 00:19:50.400 align:middle line:90% but because someone came along. 00:19:50.400 --> 00:19:53.430 align:middle line:84% And then basically, it's pretty common for people 00:19:53.430 --> 00:19:58.260 align:middle line:84% to try and figure it out during their first trimester or even 00:19:58.260 --> 00:20:00.910 align:middle line:84% their first year and enroll one year later. 00:20:00.910 --> 00:20:04.110 align:middle line:84% So eventually, some of them could have found someone 00:20:04.110 --> 00:20:07.320 align:middle line:84% to give them a scholarship-- people like us, 00:20:07.320 --> 00:20:10.050 align:middle line:84% or the sugar board, or the government. 00:20:10.050 --> 00:20:13.230 align:middle line:84% But still, the chance to enroll and complete 00:20:13.230 --> 00:20:17.010 align:middle line:84% was larger by about 27%, corresponding 00:20:17.010 --> 00:20:19.690 align:middle line:90% to about 1.2 years of education. 00:20:19.690 --> 00:20:27.600 align:middle line:84% So that's basically completing secondary school divided by 4. 00:20:27.600 --> 00:20:30.040 align:middle line:84% Conditional on starting, most people complete. 00:20:30.040 --> 00:20:33.450 align:middle line:84% So that's why we have 1.2 extra years. 00:20:33.450 --> 00:20:34.380 align:middle line:90% Some people get 0. 00:20:34.380 --> 00:20:37.560 align:middle line:90% And some people get 4. 00:20:37.560 --> 00:20:39.400 align:middle line:84% If they don't go to secondary school, 00:20:39.400 --> 00:20:43.070 align:middle line:84% some kids get to go to a technical institute. 00:20:43.070 --> 00:20:44.520 align:middle line:84% There's a small number, but one we 00:20:44.520 --> 00:20:46.680 align:middle line:84% need to keep in mind because that means we 00:20:46.680 --> 00:20:49.410 align:middle line:84% are also changing the form of education people 00:20:49.410 --> 00:20:52.260 align:middle line:90% get, potentially. 00:20:52.260 --> 00:20:55.380 align:middle line:84% Once you go to secondary school-- 00:20:55.380 --> 00:20:57.810 align:middle line:84% so 27% of them completed secondary school. 00:20:57.810 --> 00:21:01.182 align:middle line:84% Of those 27%, over the year, they keep trying. 00:21:01.182 --> 00:21:01.890 align:middle line:90% They keep trying. 00:21:01.890 --> 00:21:03.510 align:middle line:90% They keep trying. 00:21:03.510 --> 00:21:07.290 align:middle line:84% By 2019, some of those 27% had managed 00:21:07.290 --> 00:21:11.323 align:middle line:84% to enroll in a tertiary program and completed. 00:21:11.323 --> 00:21:12.990 align:middle line:84% So this is the people who have enrolled. 00:21:12.990 --> 00:21:14.668 align:middle line:84% This is the people who have completed. 00:21:14.668 --> 00:21:16.710 align:middle line:84% Some of the people who have enrolled but have not 00:21:16.710 --> 00:21:20.010 align:middle line:84% completed-- they are still ongoing because this operation 00:21:20.010 --> 00:21:20.940 align:middle line:90% takes so long. 00:21:20.940 --> 00:21:25.380 align:middle line:84% That's part of an interesting finding in this paper. 00:21:25.380 --> 00:21:27.240 align:middle line:84% Again, I don't think unique to Ghana-- 00:21:27.240 --> 00:21:29.700 align:middle line:84% but it was news to me is how long people 00:21:29.700 --> 00:21:34.320 align:middle line:84% wait to get into a degree and to cobble the money 00:21:34.320 --> 00:21:36.538 align:middle line:90% together to do it and all that. 00:21:36.538 --> 00:21:38.580 align:middle line:84% And in the meantime, they live with their parents 00:21:38.580 --> 00:21:41.810 align:middle line:90% or grandparents. 00:21:41.810 --> 00:21:46.070 align:middle line:84% And so this is the number of enrolled people, et cetera. 00:21:46.070 --> 00:21:47.090 align:middle line:90% So that's for all. 00:21:47.090 --> 00:21:48.830 align:middle line:90% And then we have women and men. 00:21:48.830 --> 00:21:50.630 align:middle line:90% Similar effect on education. 00:21:50.630 --> 00:21:53.100 align:middle line:84% Basically, larger effect on tertiary school. 00:21:53.100 --> 00:21:55.880 align:middle line:84% In fact, all of the effect of enrolling in tertiary education 00:21:55.880 --> 00:22:01.010 align:middle line:84% comes from women, something that's pertinent 00:22:01.010 --> 00:22:03.380 align:middle line:84% because although they don't catch up 00:22:03.380 --> 00:22:07.760 align:middle line:84% on-- even the treatment girls, if you add to that, 00:22:07.760 --> 00:22:10.700 align:middle line:84% the treatment girls are still less likely to have completed, 00:22:10.700 --> 00:22:13.310 align:middle line:84% or about as likely to have completed 00:22:13.310 --> 00:22:16.700 align:middle line:84% secondary school as the control boys 00:22:16.700 --> 00:22:19.100 align:middle line:84% and stay behind the control girls. 00:22:19.100 --> 00:22:21.465 align:middle line:84% But it is not true for secondary school, 00:22:21.465 --> 00:22:23.840 align:middle line:84% where are the treatment girls catch up with the treatment 00:22:23.840 --> 00:22:26.215 align:middle line:84% boys because the treatment boys have much lower treatment 00:22:26.215 --> 00:22:26.760 align:middle line:90% effect. 00:22:26.760 --> 00:22:30.320 align:middle line:84% So for secondary school, actually, 00:22:30.320 --> 00:22:35.330 align:middle line:84% just giving them the entry, the ticket into secondary school 00:22:35.330 --> 00:22:37.670 align:middle line:90% makes them catch up with-- 00:22:37.670 --> 00:22:41.195 align:middle line:84% so that says something about the dynamic in the family. 00:22:41.195 --> 00:22:43.070 align:middle line:84% And we were discussing that a little bit when 00:22:43.070 --> 00:22:46.220 align:middle line:84% we were discussing the model, that parents either 00:22:46.220 --> 00:22:50.017 align:middle line:84% dislike their girls, or have less of an expectation 00:22:50.017 --> 00:22:52.100 align:middle line:84% that they might do something with their education. 00:22:52.100 --> 00:22:55.640 align:middle line:84% They probably underestimate the value 00:22:55.640 --> 00:22:57.770 align:middle line:84% of investing in their girls, from their own point 00:22:57.770 --> 00:23:00.710 align:middle line:84% of view, which is sending them to secondary school, which 00:23:00.710 --> 00:23:03.755 align:middle line:84% is why we have such a larger treatment effect for the girls. 00:23:03.755 --> 00:23:08.780 align:middle line:90% 00:23:08.780 --> 00:23:10.610 align:middle line:90% And then they learn stuff. 00:23:10.610 --> 00:23:11.270 align:middle line:90% So that's nice. 00:23:11.270 --> 00:23:14.540 align:middle line:90% 00:23:14.540 --> 00:23:17.270 align:middle line:90% So we did test score. 00:23:17.270 --> 00:23:20.600 align:middle line:90% Is Where is my test score? 00:23:20.600 --> 00:23:21.140 align:middle line:90% Yeah. 00:23:21.140 --> 00:23:26.420 align:middle line:84% Total cognitive score-- so as we were following them up in 2013, 00:23:26.420 --> 00:23:27.890 align:middle line:84% we administered to them a version 00:23:27.890 --> 00:23:31.010 align:middle line:90% of the PISA international tests. 00:23:31.010 --> 00:23:34.070 align:middle line:84% And reassuringly, people have learned 00:23:34.070 --> 00:23:35.300 align:middle line:90% something in the schools. 00:23:35.300 --> 00:23:37.430 align:middle line:84% This, by the way, was a potential criticism 00:23:37.430 --> 00:23:39.190 align:middle line:84% of sending people to secondary schools. 00:23:39.190 --> 00:23:40.430 align:middle line:84% The schools are so bad people are not 00:23:40.430 --> 00:23:41.570 align:middle line:90% going to learn anything. 00:23:41.570 --> 00:23:43.010 align:middle line:84% That doesn't seem to be the case. 00:23:43.010 --> 00:23:46.460 align:middle line:84% They've learned something across domains. 00:23:46.460 --> 00:23:48.590 align:middle line:84% They also have greater political knowledge. 00:23:48.590 --> 00:23:52.820 align:middle line:84% They have greater engagement with the media. 00:23:52.820 --> 00:23:57.050 align:middle line:84% They are more likely to be conversant with technology, 00:23:57.050 --> 00:23:59.585 align:middle line:84% like having a bank account, know how to use the internet, 00:23:59.585 --> 00:24:02.840 align:middle line:84% will have a Facebook account, such marvelous things. 00:24:02.840 --> 00:24:05.900 align:middle line:90% 00:24:05.900 --> 00:24:09.080 align:middle line:84% We don't find any effect on fertilizer use. 00:24:09.080 --> 00:24:11.510 align:middle line:84% But we find a effect on health that I'm not 00:24:11.510 --> 00:24:15.254 align:middle line:90% going to talk much about here. 00:24:15.254 --> 00:24:17.720 align:middle line:84% Continuing and moving as this tends 00:24:17.720 --> 00:24:22.700 align:middle line:84% to get to secondary school, one of the pretty striking results 00:24:22.700 --> 00:24:26.850 align:middle line:84% in the context of this improvement in human capital, 00:24:26.850 --> 00:24:30.110 align:middle line:84% which are real, is the fact that there is really not 00:24:30.110 --> 00:24:33.720 align:middle line:84% much happening on the labor market. 00:24:33.720 --> 00:24:37.310 align:middle line:84% In fact, there is really nothing happening on 00:24:37.310 --> 00:24:38.630 align:middle line:90% whether you're working or not. 00:24:38.630 --> 00:24:41.720 align:middle line:84% And few people work more and more over time. 00:24:41.720 --> 00:24:45.770 align:middle line:84% But even by 2019, we still have 73% 00:24:45.770 --> 00:24:48.200 align:middle line:84% of people who have done any work. 00:24:48.200 --> 00:24:49.580 align:middle line:90% And this kept increasing. 00:24:49.580 --> 00:24:53.340 align:middle line:84% This was much lower at the beginning of this period. 00:24:53.340 --> 00:24:57.650 align:middle line:84% So people start really, really slowly to start working. 00:24:57.650 --> 00:25:01.850 align:middle line:90% They are not making much more. 00:25:01.850 --> 00:25:05.690 align:middle line:84% They are not more likely to be able to cope with a crisis. 00:25:05.690 --> 00:25:12.050 align:middle line:84% But they are more likely to have a job with benefits or a wage 00:25:12.050 --> 00:25:13.100 align:middle line:90% contract. 00:25:13.100 --> 00:25:17.850 align:middle line:84% So one of these jobs was potential rent-seeking job. 00:25:17.850 --> 00:25:20.090 align:middle line:84% And if we focus on women, who are the ones who 00:25:20.090 --> 00:25:25.435 align:middle line:84% were likely to be enrolled in tertiary, moving down the line, 00:25:25.435 --> 00:25:26.810 align:middle line:84% they are also more likely to have 00:25:26.810 --> 00:25:30.440 align:middle line:84% obtained one of these precious public sector jobs. 00:25:30.440 --> 00:25:33.950 align:middle line:84% However, note how few of them have one 00:25:33.950 --> 00:25:35.450 align:middle line:90% of those private sector jobs. 00:25:35.450 --> 00:25:39.560 align:middle line:84% Instead of the 80% or something that parents were expecting 00:25:39.560 --> 00:25:43.550 align:middle line:84% for their secondary school graduate, 00:25:43.550 --> 00:25:47.990 align:middle line:84% after almost 10 years of this, we 00:25:47.990 --> 00:25:52.340 align:middle line:84% are at about 10 percentage points on average, 00:25:52.340 --> 00:25:58.160 align:middle line:84% and 10 percentage point boost for girls, and for boys, who 00:25:58.160 --> 00:25:59.760 align:middle line:90% are now young men and women. 00:25:59.760 --> 00:26:03.830 align:middle line:84% So in that sense there is an overeducated Ghanaian. 00:26:03.830 --> 00:26:07.790 align:middle line:84% The hope that they had of secondary education 00:26:07.790 --> 00:26:12.890 align:middle line:84% being a ticket was dashed, squashed by what 00:26:12.890 --> 00:26:13.850 align:middle line:90% happened over time. 00:26:13.850 --> 00:26:18.490 align:middle line:90% 00:26:18.490 --> 00:26:21.790 align:middle line:84% One advantage of having these people go on, and on, and on, 00:26:21.790 --> 00:26:23.650 align:middle line:84% and on, and never finishing is eventually, 00:26:23.650 --> 00:26:25.580 align:middle line:90% we hit the COVID crisis. 00:26:25.580 --> 00:26:30.160 align:middle line:84% So we could look at the impact on labor market outcomes 00:26:30.160 --> 00:26:31.960 align:middle line:90% during the COVID crisis. 00:26:31.960 --> 00:26:35.050 align:middle line:84% And for women, who are more likely to have 00:26:35.050 --> 00:26:39.730 align:middle line:84% this public sector job, we see the protection at work. 00:26:39.730 --> 00:26:43.360 align:middle line:84% Because they are doing much better during the COVID crisis. 00:26:43.360 --> 00:26:45.970 align:middle line:84% And that's driven-- they're doing much better on average. 00:26:45.970 --> 00:26:50.950 align:middle line:84% But that appears to be driven mostly by these women who 00:26:50.950 --> 00:26:54.100 align:middle line:84% have this much more secure job-- in particular, teachers 00:26:54.100 --> 00:26:56.560 align:middle line:90% or nurses, employment job. 00:26:56.560 --> 00:26:58.922 align:middle line:90% 00:26:58.922 --> 00:26:59.630 align:middle line:90% AUDIENCE: Esther? 00:26:59.630 --> 00:27:00.050 align:middle line:90% ESTHER DUFLO: Yeah. 00:27:00.050 --> 00:27:01.280 align:middle line:84% AUDIENCE: I have a [INAUDIBLE] question 00:27:01.280 --> 00:27:02.363 align:middle line:90% about something like that. 00:27:02.363 --> 00:27:05.840 align:middle line:84% So whether or not women got a public sector job 00:27:05.840 --> 00:27:07.280 align:middle line:90% is itself an outcome. 00:27:07.280 --> 00:27:08.322 align:middle line:90% ESTHER DUFLO: Yeah, yeah. 00:27:08.322 --> 00:27:13.520 align:middle line:84% So here all I know is that on average, these women are-- 00:27:13.520 --> 00:27:16.070 align:middle line:84% that's why there is nothing here that tells you that it's 00:27:16.070 --> 00:27:17.300 align:middle line:90% them who are doing better. 00:27:17.300 --> 00:27:19.342 align:middle line:84% And it would be hard to say because that's itself 00:27:19.342 --> 00:27:19.910 align:middle line:90% an outcome. 00:27:19.910 --> 00:27:20.577 align:middle line:90% AUDIENCE: Right. 00:27:20.577 --> 00:27:25.070 align:middle line:84% ESTHER DUFLO: So all you can do is eyeball the data 00:27:25.070 --> 00:27:27.950 align:middle line:90% and say, well, let me look at-- 00:27:27.950 --> 00:27:32.210 align:middle line:84% there are very few of these women who got these jobs. 00:27:32.210 --> 00:27:34.037 align:middle line:84% If I look at all of the other women, 00:27:34.037 --> 00:27:35.870 align:middle line:84% they are not doing any better than the woman 00:27:35.870 --> 00:27:38.570 align:middle line:90% who didn't get the private job. 00:27:38.570 --> 00:27:40.610 align:middle line:84% But I can't really do that because that's 00:27:40.610 --> 00:27:42.570 align:middle line:90% a selected sample. 00:27:42.570 --> 00:27:47.720 align:middle line:84% So that's why I was tentative in saying it. 00:27:47.720 --> 00:27:49.520 align:middle line:90% And it's not in our table. 00:27:49.520 --> 00:27:51.440 align:middle line:84% If we wanted to put it in table, we 00:27:51.440 --> 00:27:53.833 align:middle line:84% wanted to take this message home, 00:27:53.833 --> 00:27:55.250 align:middle line:84% then we would have to do something 00:27:55.250 --> 00:27:58.250 align:middle line:84% about selecting into the selection 00:27:58.250 --> 00:28:02.390 align:middle line:84% into the tertiary job, and do some bounds about what would 00:28:02.390 --> 00:28:05.450 align:middle line:84% the effect be for people who didn't get 00:28:05.450 --> 00:28:09.140 align:middle line:84% into the secondary job, which we could 00:28:09.140 --> 00:28:13.100 align:middle line:84% do with an exercise similar to the one we did for working 00:28:13.100 --> 00:28:14.510 align:middle line:90% at all in the paper. 00:28:14.510 --> 00:28:17.000 align:middle line:84% We have bounds for selection into having 00:28:17.000 --> 00:28:19.230 align:middle line:90% any kind of earnings. 00:28:19.230 --> 00:28:23.840 align:middle line:84% And we create bounds by saying, well, 00:28:23.840 --> 00:28:26.030 align:middle line:84% suppose the people who have any kinds of earnings 00:28:26.030 --> 00:28:29.550 align:middle line:90% are the worst or the best. 00:28:29.550 --> 00:28:32.790 align:middle line:84% So you could do something like that. 00:28:32.790 --> 00:28:34.580 align:middle line:84% That would be the least model driven. 00:28:34.580 --> 00:28:39.050 align:middle line:84% Or you could have a more parametric version 00:28:39.050 --> 00:28:41.730 align:middle line:84% of doing that, a Heckman selection equation of some kind 00:28:41.730 --> 00:28:46.120 align:middle line:84% for trying to select not getting the government job. 00:28:46.120 --> 00:28:47.870 align:middle line:84% And then you could control for that, a bit 00:28:47.870 --> 00:28:51.300 align:middle line:84% like I did in the other paper, where I said, well, I'm only 00:28:51.300 --> 00:28:52.550 align:middle line:90% going to look at wage workers. 00:28:52.550 --> 00:28:55.430 align:middle line:84% Because that's the data I have, but let 00:28:55.430 --> 00:28:58.940 align:middle line:84% me at least control for the probability of having 00:28:58.940 --> 00:29:01.550 align:middle line:84% one of those jobs, finding another instrument, 00:29:01.550 --> 00:29:03.170 align:middle line:84% using a Heckman correction equation. 00:29:03.170 --> 00:29:04.670 align:middle line:84% So you could do something like that, 00:29:04.670 --> 00:29:09.060 align:middle line:84% which we haven't done because we left it very, very mellow, 00:29:09.060 --> 00:29:10.530 align:middle line:90% so to speak. 00:29:10.530 --> 00:29:11.850 align:middle line:90% Yeah? 00:29:11.850 --> 00:29:14.552 align:middle line:90% You go ahead, and then her. 00:29:14.552 --> 00:29:17.490 align:middle line:84% AUDIENCE: [INAUDIBLE] so in 2019, I 00:29:17.490 --> 00:29:21.780 align:middle line:84% look at the sequence number 3 and 4, 00:29:21.780 --> 00:29:24.360 align:middle line:84% I think on jobs with benefits and depending 00:29:24.360 --> 00:29:27.660 align:middle line:90% upon this [INAUDIBLE] mode. 00:29:27.660 --> 00:29:30.750 align:middle line:84% In males, there is a [INAUDIBLE] anything 00:29:30.750 --> 00:29:33.570 align:middle line:84% on getting a better job with benefits for males. 00:29:33.570 --> 00:29:34.420 align:middle line:90% ESTHER DUFLO: Yeah. 00:29:34.420 --> 00:29:34.920 align:middle line:90% Yeah. 00:29:34.920 --> 00:29:37.020 align:middle line:84% AUDIENCE: That disappears in 2020? 00:29:37.020 --> 00:29:39.120 align:middle line:84% ESTHER DUFLO: Well, in a job with benefits, 00:29:39.120 --> 00:29:41.610 align:middle line:90% it's already a bit-- 00:29:41.610 --> 00:29:49.170 align:middle line:84% it wasn't the most robust effect anyways. 00:29:49.170 --> 00:29:50.760 align:middle line:90% You have a P-value of 0.13. 00:29:50.760 --> 00:29:52.926 align:middle line:90% So-- 00:29:52.926 --> 00:29:55.500 align:middle line:84% AUDIENCE: Also, for females, the effect 00:29:55.500 --> 00:30:01.480 align:middle line:84% doesn't exist first for during a job with benefits in 2019. 00:30:01.480 --> 00:30:07.240 align:middle line:84% But in 2020, there is a stronger effect. 00:30:07.240 --> 00:30:10.023 align:middle line:90% Not strong, but close. 00:30:10.023 --> 00:30:11.440 align:middle line:84% At least, the coefficient estimate 00:30:11.440 --> 00:30:14.320 align:middle line:84% becomes very large, or almost as large 00:30:14.320 --> 00:30:17.248 align:middle line:84% as the effect on being a public centered employee. 00:30:17.248 --> 00:30:18.040 align:middle line:90% ESTHER DUFLO: Yeah. 00:30:18.040 --> 00:30:21.640 align:middle line:84% So in general, job with benefits should 00:30:21.640 --> 00:30:27.130 align:middle line:84% be almost a subset of government employees 00:30:27.130 --> 00:30:28.630 align:middle line:84% because all the government employees 00:30:28.630 --> 00:30:30.530 align:middle line:90% have some kind of benefits. 00:30:30.530 --> 00:30:34.900 align:middle line:84% So the two should really be relatively similar, 00:30:34.900 --> 00:30:36.250 align:middle line:90% especially for women. 00:30:36.250 --> 00:30:39.650 align:middle line:84% Some men get some of these wage jobs. 00:30:39.650 --> 00:30:41.680 align:middle line:84% So in general, they should look quite similar. 00:30:41.680 --> 00:30:44.500 align:middle line:90% 00:30:44.500 --> 00:30:48.950 align:middle line:84% But then that also comes from another place in the survey. 00:30:48.950 --> 00:30:51.220 align:middle line:84% So it depends a bit on what people tell us. 00:30:51.220 --> 00:30:54.160 align:middle line:84% But if you look at the point estimate, 00:30:54.160 --> 00:30:59.920 align:middle line:84% they are almost the same between a job with benefits 00:30:59.920 --> 00:31:01.430 align:middle line:90% and government job. 00:31:01.430 --> 00:31:05.270 align:middle line:84% And if you look at the means, they also are almost the same. 00:31:05.270 --> 00:31:09.180 align:middle line:84% So basically, that's the same variable for women 00:31:09.180 --> 00:31:10.710 align:middle line:84% because there are very few women who 00:31:10.710 --> 00:31:14.460 align:middle line:84% get a job with benefits, especially probably in 2020, 00:31:14.460 --> 00:31:17.040 align:middle line:84% where everybody lost their job if they weren't a government 00:31:17.040 --> 00:31:17.550 align:middle line:90% employee. 00:31:17.550 --> 00:31:18.240 align:middle line:90% AUDIENCE: I see. 00:31:18.240 --> 00:31:21.210 align:middle line:84% ESTHER DUFLO: Now if you look at the men, they didn't have-- 00:31:21.210 --> 00:31:23.250 align:middle line:84% that's a bit the same thing, which is you 00:31:23.250 --> 00:31:26.580 align:middle line:84% have four percentage points more men still have 00:31:26.580 --> 00:31:30.867 align:middle line:90% a job with benefits in 2020. 00:31:30.867 --> 00:31:32.700 align:middle line:84% Imagine that all of the government employees 00:31:32.700 --> 00:31:34.960 align:middle line:90% are in that number as well. 00:31:34.960 --> 00:31:36.690 align:middle line:84% And then the treatment effect are gone. 00:31:36.690 --> 00:31:39.960 align:middle line:84% And presumably, a lot of people have lost these jobs, 00:31:39.960 --> 00:31:41.740 align:middle line:90% both in treatment and control. 00:31:41.740 --> 00:31:44.100 align:middle line:84% So the extra job with benefits, the few, 00:31:44.100 --> 00:31:47.070 align:middle line:84% like working with an NGO, or working for the hospital, 00:31:47.070 --> 00:31:49.140 align:middle line:84% or being a driver in a big company-- 00:31:49.140 --> 00:31:51.390 align:middle line:84% they've gone during the beginning of COVID. 00:31:51.390 --> 00:31:53.287 align:middle line:84% So that's why you get this difference. 00:31:53.287 --> 00:31:53.954 align:middle line:90% AUDIENCE: I see. 00:31:53.954 --> 00:31:56.700 align:middle line:90% 00:31:56.700 --> 00:31:59.810 align:middle line:84% ESTHER DUFLO: So just so that I don't say good things 00:31:59.810 --> 00:32:02.300 align:middle line:90% about my own paper-- 00:32:02.300 --> 00:32:03.470 align:middle line:90% on camera, in particular-- 00:32:03.470 --> 00:32:05.102 align:middle line:90% [LAUGHTER] 00:32:05.102 --> 00:32:08.810 align:middle line:84% What's pretty cool about this experiment? 00:32:08.810 --> 00:32:12.137 align:middle line:90% What's fairly remarkable? 00:32:12.137 --> 00:32:15.720 align:middle line:90% 00:32:15.720 --> 00:32:17.340 align:middle line:84% AUDIENCE: The extremely long timeline 00:32:17.340 --> 00:32:19.470 align:middle line:90% coupled with very low attrition. 00:32:19.470 --> 00:32:20.220 align:middle line:90% ESTHER DUFLO: Yes. 00:32:20.220 --> 00:32:21.690 align:middle line:90% So the extremely long timeline-- 00:32:21.690 --> 00:32:24.630 align:middle line:84% I don't know if it's very cool or if it means we cannot get 00:32:24.630 --> 00:32:25.830 align:middle line:90% around to write a paper. 00:32:25.830 --> 00:32:28.860 align:middle line:90% But we continue finding them. 00:32:28.860 --> 00:32:32.250 align:middle line:84% But that's useful because we are certainly 00:32:32.250 --> 00:32:35.790 align:middle line:90% learning things over the years. 00:32:35.790 --> 00:32:38.700 align:middle line:90% It's necessary because people-- 00:32:38.700 --> 00:32:41.580 align:middle line:84% some amount of time was definitely necessary 00:32:41.580 --> 00:32:45.180 align:middle line:84% because you need to let people finish secondary school, 00:32:45.180 --> 00:32:46.537 align:middle line:90% start in the labor market. 00:32:46.537 --> 00:32:48.120 align:middle line:84% The people who are in secondary school 00:32:48.120 --> 00:32:50.560 align:middle line:84% start with a gap four years later. 00:32:50.560 --> 00:32:54.470 align:middle line:84% So you need to give them time to catch up in terms of experience 00:32:54.470 --> 00:32:55.470 align:middle line:90% and something like that. 00:32:55.470 --> 00:32:58.320 align:middle line:84% So you really need to continue a little bit. 00:32:58.320 --> 00:33:03.070 align:middle line:84% But that creates the issue that you might lose them. 00:33:03.070 --> 00:33:08.370 align:middle line:84% So we were very concerned about not losing them. 00:33:08.370 --> 00:33:11.340 align:middle line:84% We gave them these cell phones, which 00:33:11.340 --> 00:33:14.730 align:middle line:84% we kept changing from time to time, and sim cards, 00:33:14.730 --> 00:33:16.630 align:middle line:90% and called them regularly. 00:33:16.630 --> 00:33:20.640 align:middle line:84% And if we couldn't find them, we went 00:33:20.640 --> 00:33:23.610 align:middle line:84% to look for them in their homes and convinced 00:33:23.610 --> 00:33:24.870 align:middle line:90% them to do a survey. 00:33:24.870 --> 00:33:26.670 align:middle line:84% I should say, it becomes harder and harder. 00:33:26.670 --> 00:33:27.810 align:middle line:90% They are a bit tired now. 00:33:27.810 --> 00:33:30.490 align:middle line:84% So I don't know how long we can keep them. 00:33:30.490 --> 00:33:32.070 align:middle line:90% We are now doing their kids. 00:33:32.070 --> 00:33:34.830 align:middle line:84% We are doing a lot of work with their children 00:33:34.830 --> 00:33:37.960 align:middle line:84% along with a psychologist at Harvard, Liz Belcher. 00:33:37.960 --> 00:33:41.290 align:middle line:84% I'm going to talk more about her work on Wednesday as well. 00:33:41.290 --> 00:33:45.780 align:middle line:84% We are trying to see whether their children have 00:33:45.780 --> 00:33:48.030 align:middle line:90% higher human capital. 00:33:48.030 --> 00:33:49.740 align:middle line:84% But they are starting to get tired of us. 00:33:49.740 --> 00:33:51.323 align:middle line:84% And they don't want to see us anymore. 00:33:51.323 --> 00:33:54.180 align:middle line:84% But otherwise, during all this time, we keep finding them. 00:33:54.180 --> 00:33:56.430 align:middle line:84% We pay them small amounts of money for the interview. 00:33:56.430 --> 00:33:57.840 align:middle line:90% They're used to us, et cetera. 00:33:57.840 --> 00:34:00.180 align:middle line:90% But it takes a lot of effort. 00:34:00.180 --> 00:34:05.370 align:middle line:84% And the low attrition rate is rare for this. 00:34:05.370 --> 00:34:08.232 align:middle line:84% So that's what's remarkable about this experiment. 00:34:08.232 --> 00:34:09.690 align:middle line:84% But of course, we don't really have 00:34:09.690 --> 00:34:12.750 align:middle line:84% what we want because we are comparing the scholarship 00:34:12.750 --> 00:34:17.310 align:middle line:90% to not getting the scholarship. 00:34:17.310 --> 00:34:19.420 align:middle line:84% And this is a lecture on return to education, 00:34:19.420 --> 00:34:22.590 align:middle line:90% not return on free education. 00:34:22.590 --> 00:34:26.190 align:middle line:84% So the question is, can we go one step further? 00:34:26.190 --> 00:34:28.650 align:middle line:84% Now there are good reasons why we are not doing it 00:34:28.650 --> 00:34:33.389 align:middle line:84% in the paper, which we will go to towards 00:34:33.389 --> 00:34:35.190 align:middle line:90% the end of the lecture. 00:34:35.190 --> 00:34:37.230 align:middle line:84% But it's still a good example to go to. 00:34:37.230 --> 00:34:41.460 align:middle line:84% It's easy to think about what you might want to do, and to do 00:34:41.460 --> 00:34:43.260 align:middle line:90% it for during the class. 00:34:43.260 --> 00:34:46.739 align:middle line:84% And then we can discuss the limitation 00:34:46.739 --> 00:34:48.750 align:middle line:84% of the exercise in this context and why 00:34:48.750 --> 00:34:51.850 align:middle line:90% we decided not to perform it. 00:34:51.850 --> 00:34:56.520 align:middle line:84% So if the question we had was how education improves 00:34:56.520 --> 00:35:01.590 align:middle line:84% earnings, or test scores, or health, or whatever and we, 00:35:01.590 --> 00:35:04.260 align:middle line:84% again, said that we cannot manipulate the education 00:35:04.260 --> 00:35:09.840 align:middle line:84% directly, but we have this thing that has an effect on it, 00:35:09.840 --> 00:35:11.290 align:middle line:90% then what would we do? 00:35:11.290 --> 00:35:14.040 align:middle line:84% Well, to start with simplest case, 00:35:14.040 --> 00:35:18.300 align:middle line:84% assume that we can write this, where the treatment effect is 00:35:18.300 --> 00:35:19.150 align:middle line:90% constant. 00:35:19.150 --> 00:35:21.450 align:middle line:84% So we don't have this individual treatment 00:35:21.450 --> 00:35:22.710 align:middle line:90% effect we started with. 00:35:22.710 --> 00:35:24.430 align:middle line:84% I'm going to change that very soon. 00:35:24.430 --> 00:35:27.300 align:middle line:90% But we'll start with that. 00:35:27.300 --> 00:35:33.040 align:middle line:90% Then that's easy. 00:35:33.040 --> 00:35:38.460 align:middle line:84% We have what's called an instrument in Ghana, which 00:35:38.460 --> 00:35:40.560 align:middle line:90% is an emphasis on potential. 00:35:40.560 --> 00:35:43.080 align:middle line:84% Because we'll discuss what characteristics 00:35:43.080 --> 00:35:44.910 align:middle line:84% the instrument has and whether we think 00:35:44.910 --> 00:35:47.905 align:middle line:84% they are satisfied here and so on and so forth. 00:35:47.905 --> 00:35:49.530 align:middle line:84% The scholarships were randomly assigned 00:35:49.530 --> 00:35:52.020 align:middle line:90% to students who qualified. 00:35:52.020 --> 00:35:57.690 align:middle line:84% So let's call Z a dummy equal to 1 00:35:57.690 --> 00:36:04.350 align:middle line:84% if you get the scholarship in this sample, and 0 otherwise. 00:36:04.350 --> 00:36:06.030 align:middle line:84% Getting the scholarship, as we saw, 00:36:06.030 --> 00:36:08.790 align:middle line:84% increased the probability to ever enroll in a high school 00:36:08.790 --> 00:36:10.740 align:middle line:90% by 25 percentage points. 00:36:10.740 --> 00:36:12.810 align:middle line:90% Some kids enrolled anyway. 00:36:12.810 --> 00:36:14.490 align:middle line:90% About half of them complete. 00:36:14.490 --> 00:36:17.400 align:middle line:84% Some kids don't enroll even with the scholarship. 00:36:17.400 --> 00:36:20.310 align:middle line:84% About a quarter of people say, no, thank you very much. 00:36:20.310 --> 00:36:22.090 align:middle line:84% So there is noncompliance both ways. 00:36:22.090 --> 00:36:24.800 align:middle line:90% 00:36:24.800 --> 00:36:31.640 align:middle line:84% So still, despite that, we can construct an instrumental 00:36:31.640 --> 00:36:34.130 align:middle line:90% variable estimates. 00:36:34.130 --> 00:36:35.960 align:middle line:84% In this setup, we can say, well, let's 00:36:35.960 --> 00:36:38.390 align:middle line:84% write the treatment effect on school participation-- 00:36:38.390 --> 00:36:41.960 align:middle line:84% enrollment, or years of schooling. 00:36:41.960 --> 00:36:44.630 align:middle line:90% Whatever we like. 00:36:44.630 --> 00:36:47.570 align:middle line:84% And then down the road, we can do the same-- 00:36:47.570 --> 00:36:49.640 align:middle line:84% as we have already shown you, the table 00:36:49.640 --> 00:36:53.480 align:middle line:84% where we're comparing earnings in treatment and control group, 00:36:53.480 --> 00:36:55.880 align:middle line:84% where treatment is defined by getting the scholarship 00:36:55.880 --> 00:36:57.660 align:middle line:90% versus not. 00:36:57.660 --> 00:37:03.950 align:middle line:84% And then since I had written Yi as alpha plus beta 00:37:03.950 --> 00:37:09.800 align:middle line:84% Si plus epsilon, I can just write the expectation here. 00:37:09.800 --> 00:37:11.270 align:middle line:90% And this one as well-- 00:37:11.270 --> 00:37:13.760 align:middle line:84% take the difference between the two. 00:37:13.760 --> 00:37:16.952 align:middle line:90% And I'm getting this-- 00:37:16.952 --> 00:37:19.160 align:middle line:84% and once I have taken the difference between the two, 00:37:19.160 --> 00:37:23.870 align:middle line:84% I'm getting that E of this difference 00:37:23.870 --> 00:37:30.470 align:middle line:84% is alpha plus beta times this difference. 00:37:30.470 --> 00:37:34.180 align:middle line:84% So if I'm interested in beta, I just 00:37:34.180 --> 00:37:36.580 align:middle line:84% have to divide this one by this one. 00:37:36.580 --> 00:37:38.950 align:middle line:90% And there's some assumption. 00:37:38.950 --> 00:37:40.350 align:middle line:84% And the assumption, of course, is 00:37:40.350 --> 00:37:47.850 align:middle line:84% I want to make that E of epsilon given Zi equal 1 minus E 00:37:47.850 --> 00:37:50.640 align:middle line:90% of epsilon given Zi equals 0. 00:37:50.640 --> 00:37:52.680 align:middle line:84% I would like to be able to assume that it's 0. 00:37:52.680 --> 00:37:56.500 align:middle line:90% 00:37:56.500 --> 00:37:58.960 align:middle line:84% So when can I make this assumption? 00:37:58.960 --> 00:38:03.250 align:middle line:84% What underlies this assumption in this case, in words? 00:38:03.250 --> 00:38:04.090 align:middle line:90% What helps? 00:38:04.090 --> 00:38:05.620 align:middle line:90% And is it a good assumption? 00:38:05.620 --> 00:38:09.120 align:middle line:90% 00:38:09.120 --> 00:38:10.103 align:middle line:90% Yep. 00:38:10.103 --> 00:38:12.270 align:middle line:84% AUDIENCE: In this case, the instrument Z is randomly 00:38:12.270 --> 00:38:17.120 align:middle line:84% assigned, so you wouldn't have confidence that whether you're 00:38:17.120 --> 00:38:20.460 align:middle line:84% assigned or not is independent of any unobservables that are 00:38:20.460 --> 00:38:21.788 align:middle line:90% captured [? and is ?] epsilon. 00:38:21.788 --> 00:38:22.580 align:middle line:90% ESTHER DUFLO: Yeah. 00:38:22.580 --> 00:38:23.450 align:middle line:90% So that's good. 00:38:23.450 --> 00:38:24.500 align:middle line:90% That helps. 00:38:24.500 --> 00:38:31.500 align:middle line:84% But remember that the epsilon is in the education equation. 00:38:31.500 --> 00:38:36.920 align:middle line:84% So from the discussions we had last time, 00:38:36.920 --> 00:38:39.470 align:middle line:84% we know we have a reliable causal estimate 00:38:39.470 --> 00:38:42.470 align:middle line:84% of the impact of the scholarship on Y 00:38:42.470 --> 00:38:45.110 align:middle line:84% and the impact of the scholarship on Z 00:38:45.110 --> 00:38:48.080 align:middle line:90% because the scholarship is-- 00:38:48.080 --> 00:38:49.947 align:middle line:84% and so the impact of the scholarship on Y 00:38:49.947 --> 00:38:51.530 align:middle line:84% and the impact of the scholarship on S 00:38:51.530 --> 00:38:56.460 align:middle line:84% because the scholarship is randomly assigned. 00:38:56.460 --> 00:39:00.000 align:middle line:90% So that's a good start. 00:39:00.000 --> 00:39:02.270 align:middle line:90% But what is hiding? 00:39:02.270 --> 00:39:05.328 align:middle line:84% There's one more thing that's hiding in this assumption. 00:39:05.328 --> 00:39:06.703 align:middle line:84% AUDIENCE: It [? seems ?] like you 00:39:06.703 --> 00:39:08.411 align:middle line:84% might be worried about unobservables that 00:39:08.411 --> 00:39:11.462 align:middle line:90% would determine the treatment-- 00:39:11.462 --> 00:39:13.170 align:middle line:84% ESTHER DUFLO: Would you mind speaking up? 00:39:13.170 --> 00:39:15.462 align:middle line:84% AUDIENCE: You might be worried about unobservables that 00:39:15.462 --> 00:39:17.370 align:middle line:90% are affected by the treatment. 00:39:17.370 --> 00:39:21.450 align:middle line:84% So it could be the cash that you get to go to school. 00:39:21.450 --> 00:39:23.663 align:middle line:90% It could also be having a-- 00:39:23.663 --> 00:39:24.580 align:middle line:90% ESTHER DUFLO: Exactly. 00:39:24.580 --> 00:39:26.590 align:middle line:84% So what we might be worried about-- 00:39:26.590 --> 00:39:29.530 align:middle line:84% we know that the people who got the scholarship are no 00:39:29.530 --> 00:39:32.690 align:middle line:84% different than the people who didn't. 00:39:32.690 --> 00:39:33.850 align:middle line:90% So that's your point. 00:39:33.850 --> 00:39:35.890 align:middle line:90% I'm continuing that. 00:39:35.890 --> 00:39:37.570 align:middle line:90% I randomly assigned them. 00:39:37.570 --> 00:39:39.250 align:middle line:90% I've done this work. 00:39:39.250 --> 00:39:42.140 align:middle line:90% But then they got a scholarship. 00:39:42.140 --> 00:39:45.790 align:middle line:84% So the further assumption that is necessary to make that claim 00:39:45.790 --> 00:39:48.340 align:middle line:84% is that the scholarships themselves 00:39:48.340 --> 00:39:53.950 align:middle line:84% don't have a direct impact on earnings, which they might, 00:39:53.950 --> 00:39:56.860 align:middle line:84% in [? Kailesh's ?] example, for example, 00:39:56.860 --> 00:39:59.800 align:middle line:84% because I gave cash to the family. 00:39:59.800 --> 00:40:03.550 align:middle line:84% Maybe for the half of the kids who 00:40:03.550 --> 00:40:06.970 align:middle line:84% would have gone to school anyways, 00:40:06.970 --> 00:40:08.920 align:middle line:90% it's now much easier for them. 00:40:08.920 --> 00:40:12.020 align:middle line:84% There is not this financial strain, or it's not there. 00:40:12.020 --> 00:40:17.270 align:middle line:84% And so it's easier for the family. 00:40:17.270 --> 00:40:18.130 align:middle line:90% They're better fed. 00:40:18.130 --> 00:40:20.470 align:middle line:84% And because of that, they earn more money later. 00:40:20.470 --> 00:40:23.890 align:middle line:90% 00:40:23.890 --> 00:40:25.180 align:middle line:90% So that would be one. 00:40:25.180 --> 00:40:27.760 align:middle line:84% Because I give the scholarship to everyone, even those 00:40:27.760 --> 00:40:30.430 align:middle line:84% who would have gone to school anyways, right? 00:40:30.430 --> 00:40:35.200 align:middle line:84% So in randomized experiments, we have 00:40:35.200 --> 00:40:39.170 align:middle line:84% to be careful because the first one is for free from the trial. 00:40:39.170 --> 00:40:43.030 align:middle line:84% But when we use the experiment as an instrument for something 00:40:43.030 --> 00:40:45.700 align:middle line:84% else, the second one doesn't come for free. 00:40:45.700 --> 00:40:47.740 align:middle line:90% And it's not always satisfied. 00:40:47.740 --> 00:40:52.420 align:middle line:84% And you get a lot of dicey people 00:40:52.420 --> 00:40:55.000 align:middle line:84% from going one step too far and saying, well, 00:40:55.000 --> 00:40:56.110 align:middle line:90% I'm going to use-- 00:40:56.110 --> 00:40:57.938 align:middle line:90% I'm estimating my first stage. 00:40:57.938 --> 00:40:59.230 align:middle line:90% That's the effect on education. 00:40:59.230 --> 00:41:02.540 align:middle line:84% My reduced form-- that's the effect on wages, for example. 00:41:02.540 --> 00:41:03.790 align:middle line:90% And then I can divide it, too. 00:41:03.790 --> 00:41:05.510 align:middle line:90% I have the returns to education. 00:41:05.510 --> 00:41:17.190 align:middle line:84% It may or may not apply, depending on the setup. 00:41:17.190 --> 00:41:20.730 align:middle line:84% And here this is something that we might want to discuss. 00:41:20.730 --> 00:41:22.410 align:middle line:84% Well, we already started the discussion. 00:41:22.410 --> 00:41:24.090 align:middle line:90% There might be other reasons. 00:41:24.090 --> 00:41:26.860 align:middle line:90% Oops. 00:41:26.860 --> 00:41:28.930 align:middle line:84% Nonetheless, if we believe them, then we 00:41:28.930 --> 00:41:31.750 align:middle line:84% obtain the effect of education-- so on knowledge, earnings, 00:41:31.750 --> 00:41:35.710 align:middle line:84% and anything else by dividing the effect of the program 00:41:35.710 --> 00:41:38.200 align:middle line:84% on, say, cognitive test scores, if that's knowledge 00:41:38.200 --> 00:41:40.810 align:middle line:84% by the effect of the program on education, or earnings, 00:41:40.810 --> 00:41:43.060 align:middle line:90% or whatever you want. 00:41:43.060 --> 00:41:46.690 align:middle line:84% We call that guide the reduced form. 00:41:46.690 --> 00:41:48.550 align:middle line:90% This one is the first stage. 00:41:48.550 --> 00:41:53.950 align:middle line:84% The simple ratio of the two is the Wald estimate. 00:41:53.950 --> 00:41:56.590 align:middle line:84% If you estimate it-- and that's a simple form 00:41:56.590 --> 00:42:00.280 align:middle line:84% of the instrumental variable estimator. 00:42:00.280 --> 00:42:04.360 align:middle line:84% Instead, you could run two SLS, which basically involves 00:42:04.360 --> 00:42:10.660 align:middle line:84% projecting an education on the scholarship, 00:42:10.660 --> 00:42:15.580 align:middle line:84% and then regressing the outcome on the projection. 00:42:15.580 --> 00:42:17.170 align:middle line:84% And then you'll get standard error. 00:42:17.170 --> 00:42:20.770 align:middle line:84% That's your two [INAUDIBLE] square. 00:42:20.770 --> 00:42:23.470 align:middle line:84% But let's calculate the Wald estimate, just to make sure. 00:42:23.470 --> 00:42:26.260 align:middle line:84% So for example, if we wanted to do it 00:42:26.260 --> 00:42:28.330 align:middle line:84% for completed-- so first of all, one choice 00:42:28.330 --> 00:42:35.140 align:middle line:84% we have is what independent variable you want to use. 00:42:35.140 --> 00:42:37.750 align:middle line:84% You have one instrument for all of these things. 00:42:37.750 --> 00:42:39.460 align:middle line:84% It's a little bit cosmetic, in a way. 00:42:39.460 --> 00:42:41.950 align:middle line:84% It's whether you want to-- because the information that's 00:42:41.950 --> 00:42:43.990 align:middle line:84% underlying it is always the same. 00:42:43.990 --> 00:42:46.450 align:middle line:84% But you might want to, say, let me 00:42:46.450 --> 00:42:48.910 align:middle line:84% do it for years of education, or let me do it 00:42:48.910 --> 00:42:50.860 align:middle line:90% for completed secondary school. 00:42:50.860 --> 00:42:54.670 align:middle line:84% Maybe it makes more sense to use completed secondary school 00:42:54.670 --> 00:42:57.130 align:middle line:84% as a mental frame because again, the years-- it's 00:42:57.130 --> 00:42:59.210 align:middle line:84% not that you're giving people an extra year. 00:42:59.210 --> 00:43:02.380 align:middle line:84% It's you're giving a quarter of people four extra years. 00:43:02.380 --> 00:43:04.010 align:middle line:90% So I prefer to use that. 00:43:04.010 --> 00:43:06.130 align:middle line:90% So let's say it's 27%. 00:43:06.130 --> 00:43:14.750 align:middle line:84% And then I want to take the cognitive score. 00:43:14.750 --> 00:43:16.400 align:middle line:90% It's 0.15. 00:43:16.400 --> 00:43:23.270 align:middle line:84% So if you recall, the 0.27%, 0.25 that 00:43:23.270 --> 00:43:25.380 align:middle line:84% makes it easier to do the calculation. 00:43:25.380 --> 00:43:27.860 align:middle line:84% What's the Wald estimate of education on test scores? 00:43:27.860 --> 00:43:33.710 align:middle line:90% 00:43:33.710 --> 00:43:36.430 align:middle line:84% I made it easier by saying the effect on completed 00:43:36.430 --> 00:43:41.320 align:middle line:90% primary school is 0.25. 00:43:41.320 --> 00:43:41.820 align:middle line:90% Yeah. 00:43:41.820 --> 00:43:44.610 align:middle line:90% 00:43:44.610 --> 00:43:45.930 align:middle line:90% AUDIENCE: 0.16. 00:43:45.930 --> 00:43:48.091 align:middle line:90% ESTHER DUFLO: A point? 00:43:48.091 --> 00:43:50.466 align:middle line:84% AUDIENCE: Or sorry, so that one's 0.1 times [INAUDIBLE].. 00:43:50.466 --> 00:43:51.508 align:middle line:90% ESTHER DUFLO: Yeah, 0.16. 00:43:51.508 --> 00:43:54.380 align:middle line:90% And then first set test scores. 00:43:54.380 --> 00:43:56.660 align:middle line:84% AUDIENCE: Yep, and then you divide it by the other one 00:43:56.660 --> 00:43:57.160 align:middle line:90% that we had. 00:43:57.160 --> 00:43:58.088 align:middle line:90% So you get 64. 00:43:58.088 --> 00:43:58.880 align:middle line:90% ESTHER DUFLO: Yeah. 00:43:58.880 --> 00:44:03.020 align:middle line:84% So you divide it by 0.25, which is like multiplying it by 4. 00:44:03.020 --> 00:44:08.730 align:middle line:90% We get about 0.64. 00:44:08.730 --> 00:44:10.770 align:middle line:84% So we can do the same thing for-- 00:44:10.770 --> 00:44:14.160 align:middle line:84% so that gives us now the effect of completing secondary school 00:44:14.160 --> 00:44:17.950 align:middle line:84% on standard deviation of cognitive test scores. 00:44:17.950 --> 00:44:21.120 align:middle line:84% That's very helpful to think, even for cognitive skills. 00:44:21.120 --> 00:44:23.310 align:middle line:84% Now we can compare that with other things 00:44:23.310 --> 00:44:25.170 align:middle line:84% we could do that change test scores. 00:44:25.170 --> 00:44:27.360 align:middle line:84% And we say, well, that's a pretty large impact. 00:44:27.360 --> 00:44:29.010 align:middle line:84% 0.6 standard deviation on test scores-- 00:44:29.010 --> 00:44:30.468 align:middle line:84% there are not very many things that 00:44:30.468 --> 00:44:33.780 align:middle line:84% move standard deviation on test scores by 0.6. 00:44:33.780 --> 00:44:35.262 align:middle line:90% We can also do it for wages. 00:44:35.262 --> 00:44:37.470 align:middle line:84% So that's going to be less impressive because nothing 00:44:37.470 --> 00:44:38.310 align:middle line:90% is significant. 00:44:38.310 --> 00:44:43.110 align:middle line:84% It's not going to give you a bigger ratio, which once you 00:44:43.110 --> 00:44:45.330 align:middle line:84% do it as standard error with huge standard error-- 00:44:45.330 --> 00:44:47.730 align:middle line:84% but this is the exercise you could do. 00:44:47.730 --> 00:44:52.510 align:middle line:90% 00:44:52.510 --> 00:44:55.630 align:middle line:84% Now once we have done that, so suppose our Wald estimate-- 00:44:55.630 --> 00:44:57.490 align:middle line:90% so from the Wald estimate-- 00:44:57.490 --> 00:45:00.752 align:middle line:90% 00:45:00.752 --> 00:45:03.480 align:middle line:84% oh, we don't have cognitive score. 00:45:03.480 --> 00:45:04.710 align:middle line:90% So that's OK. 00:45:04.710 --> 00:45:07.710 align:middle line:84% We can do it with something else-- for example, 00:45:07.710 --> 00:45:10.230 align:middle line:90% with technology adoption. 00:45:10.230 --> 00:45:15.600 align:middle line:84% If we compute the Wald estimate or the IV of education 00:45:15.600 --> 00:45:19.140 align:middle line:84% on adopting technology, if you remember, 00:45:19.140 --> 00:45:22.750 align:middle line:84% the reduced form was very small and insignificant. 00:45:22.750 --> 00:45:26.970 align:middle line:84% So the IV, similarly, is small and insignificant. 00:45:26.970 --> 00:45:31.620 align:middle line:84% But then if you note, the OLS is simply 00:45:31.620 --> 00:45:37.710 align:middle line:84% regressing whether someone adopt technology, 00:45:37.710 --> 00:45:40.140 align:middle line:84% and whether they completed secondary school. 00:45:40.140 --> 00:45:43.140 align:middle line:84% If we had run the naive regression that we started 00:45:43.140 --> 00:45:46.410 align:middle line:84% with, what we have concluded education 00:45:46.410 --> 00:45:51.400 align:middle line:84% does to the probability of adopting a technology-- 00:45:51.400 --> 00:45:53.110 align:middle line:90% technology in general. 00:45:53.110 --> 00:45:56.890 align:middle line:84% Fertilizer, internet, health technologies, et cetera. 00:45:56.890 --> 00:46:03.140 align:middle line:90% 00:46:03.140 --> 00:46:06.748 align:middle line:90% You understand what this OLS is? 00:46:06.748 --> 00:46:10.920 align:middle line:84% It's regressing the outcome on education. 00:46:10.920 --> 00:46:13.710 align:middle line:84% So what would we conclude from that with 0.15 00:46:13.710 --> 00:46:16.274 align:middle line:90% with a standard of 0.01? 00:46:16.274 --> 00:46:17.040 align:middle line:90% Yeah. 00:46:17.040 --> 00:46:19.710 align:middle line:84% AUDIENCE: That education increases technology adoption. 00:46:19.710 --> 00:46:20.610 align:middle line:84% ESTHER DUFLO: Yeah, that education 00:46:20.610 --> 00:46:21.990 align:middle line:90% increases technology adoption. 00:46:21.990 --> 00:46:23.055 align:middle line:90% And it's very precise. 00:46:23.055 --> 00:46:25.590 align:middle line:90% 00:46:25.590 --> 00:46:26.880 align:middle line:90% And it's pretty large. 00:46:26.880 --> 00:46:31.060 align:middle line:84% But then when we do the IV, it's a much smaller coefficient. 00:46:31.060 --> 00:46:32.880 align:middle line:90% So that's relatively intuitive. 00:46:32.880 --> 00:46:35.730 align:middle line:84% It's saying, well, people who are prone to adopt 00:46:35.730 --> 00:46:37.770 align:middle line:84% go to school, even without scholarship, of all 00:46:37.770 --> 00:46:39.000 align:middle line:90% of the people who have money. 00:46:39.000 --> 00:46:41.590 align:middle line:84% And then with the money, they can do other things. 00:46:41.590 --> 00:46:43.365 align:middle line:84% So we have a natural selection bias. 00:46:43.365 --> 00:46:46.130 align:middle line:90% 00:46:46.130 --> 00:46:50.950 align:middle line:84% So that shows what the IV can do sometimes. 00:46:50.950 --> 00:46:54.730 align:middle line:84% I'm not going to talk now about the GML, but I'll go to that. 00:46:54.730 --> 00:47:00.970 align:middle line:84% Now you can see from this equation the danger with IV 00:47:00.970 --> 00:47:06.650 align:middle line:84% is that that thing is in the denominator. 00:47:06.650 --> 00:47:08.830 align:middle line:84% And if you have anything in IV, for example, 00:47:08.830 --> 00:47:16.560 align:middle line:84% that doesn't have a huge impact on your variable-- 00:47:16.560 --> 00:47:18.260 align:middle line:90% 0.25 is already pretty good. 00:47:18.260 --> 00:47:21.400 align:middle line:84% But sometimes in IV, the effect is much smaller. 00:47:21.400 --> 00:47:23.950 align:middle line:84% You have an instrument that really doesn't move education 00:47:23.950 --> 00:47:26.080 align:middle line:90% that much. 00:47:26.080 --> 00:47:30.550 align:middle line:84% If it has a direct impact on your outcome of interest, 00:47:30.550 --> 00:47:33.340 align:middle line:84% that's going to be divided by a small number. 00:47:33.340 --> 00:47:35.560 align:middle line:90% That's going to blow up. 00:47:35.560 --> 00:47:39.310 align:middle line:84% So any bias that exists in the reduced 00:47:39.310 --> 00:47:43.060 align:middle line:84% form due, for example, to a direct impact 00:47:43.060 --> 00:47:45.460 align:middle line:84% of your instrument on your outcomes, 00:47:45.460 --> 00:47:49.660 align:middle line:84% or in nonexperimental cases due to selection biases that 00:47:49.660 --> 00:47:52.420 align:middle line:84% are not properly taken into account, 00:47:52.420 --> 00:47:57.200 align:middle line:84% any bias like that is going to explode, 00:47:57.200 --> 00:48:00.640 align:middle line:84% which is one reason why that, combined 00:48:00.640 --> 00:48:05.830 align:middle line:84% with a certain amount of selection bias of what gets 00:48:05.830 --> 00:48:08.440 align:middle line:84% published, that an editor will not publish 00:48:08.440 --> 00:48:12.220 align:middle line:90% a huge negative result-- 00:48:12.220 --> 00:48:16.030 align:middle line:84% so the huge negative results get canceled. 00:48:16.030 --> 00:48:18.400 align:middle line:90% They stay in drawers. 00:48:18.400 --> 00:48:20.675 align:middle line:90% In general, there is more bias. 00:48:20.675 --> 00:48:22.300 align:middle line:84% So the estimate goes all over the place 00:48:22.300 --> 00:48:25.870 align:middle line:84% because the bias adds to the noise to create-- 00:48:25.870 --> 00:48:29.290 align:middle line:84% estimates can be very large or very low, 00:48:29.290 --> 00:48:31.720 align:middle line:84% depending on which side the bias goes. 00:48:31.720 --> 00:48:33.750 align:middle line:84% If you only publish the positive one, 00:48:33.750 --> 00:48:35.950 align:middle line:84% then you're finding yourself in a situation where 00:48:35.950 --> 00:48:37.600 align:middle line:84% all the IVs that are published tend 00:48:37.600 --> 00:48:40.540 align:middle line:90% to be much larger than the OLS. 00:48:40.540 --> 00:48:46.150 align:middle line:84% So that is an observation that people have done. 00:48:46.150 --> 00:48:47.950 align:middle line:84% Colm Harmon and David [INAUDIBLE] 00:48:47.950 --> 00:48:51.220 align:middle line:84% have a paper on that, showing how the IV is always 00:48:51.220 --> 00:48:55.180 align:middle line:84% bigger than the OLS for returns to education in particular. 00:48:55.180 --> 00:48:56.800 align:middle line:84% There could be substantive reason. 00:48:56.800 --> 00:48:59.050 align:middle line:84% You can definitely think of that. 00:48:59.050 --> 00:49:02.560 align:middle line:84% And we'll come in a minute for why IV might be better than 00:49:02.560 --> 00:49:04.240 align:middle line:90% or OLS for legitimate reasons. 00:49:04.240 --> 00:49:10.090 align:middle line:84% But it might also just be because the IV are biased. 00:49:10.090 --> 00:49:12.340 align:middle line:84% All the negative ones-- the fact that they are biased 00:49:12.340 --> 00:49:14.507 align:middle line:84% doesn't mean they are necessarily positively biased, 00:49:14.507 --> 00:49:17.200 align:middle line:84% but just have bias, which is enlarged. 00:49:17.200 --> 00:49:19.210 align:middle line:84% All of the negative ones are dead. 00:49:19.210 --> 00:49:23.600 align:middle line:84% And then you only are left with the positive large ones. 00:49:23.600 --> 00:49:25.930 align:middle line:90% So that's the problem with IV. 00:49:25.930 --> 00:49:29.200 align:middle line:84% And that's why you want to be very careful handling IV. 00:49:29.200 --> 00:49:32.920 align:middle line:84% And in general, I like it because it gives you 00:49:32.920 --> 00:49:35.930 align:middle line:90% a good sense of magnitude. 00:49:35.930 --> 00:49:40.652 align:middle line:90% But in general, I'm wary of it. 00:49:40.652 --> 00:49:42.610 align:middle line:84% So that's why, for example, in the Ghana paper, 00:49:42.610 --> 00:49:45.640 align:middle line:84% we are accepting that one table, which is 00:49:45.640 --> 00:49:47.140 align:middle line:90% meant to compare with the OLS. 00:49:47.140 --> 00:49:51.070 align:middle line:84% And therefore, we need comparable magnitude. 00:49:51.070 --> 00:49:55.390 align:middle line:84% I often don't tend to print them. 00:49:55.390 --> 00:49:57.880 align:middle line:84% A lot of the literature is more going 00:49:57.880 --> 00:49:59.710 align:middle line:84% in this direction, at least in development, 00:49:59.710 --> 00:50:03.310 align:middle line:84% in saying, even when you have something that affects what you 00:50:03.310 --> 00:50:06.160 align:middle line:84% might be interested in and the other, 00:50:06.160 --> 00:50:09.070 align:middle line:84% and then that affects the outcome of interest, 00:50:09.070 --> 00:50:14.640 align:middle line:84% it's a little dangerous to do an IV because of the bias. 00:50:14.640 --> 00:50:21.290 align:middle line:84% So in this particular case, we were already worried about-- 00:50:21.290 --> 00:50:24.740 align:middle line:84% one potential source of direct impact and therefore bias 00:50:24.740 --> 00:50:29.030 align:middle line:84% is giving money to the families to attend school. 00:50:29.030 --> 00:50:30.740 align:middle line:84% Now this is counterbalanced by the fact 00:50:30.740 --> 00:50:33.560 align:middle line:84% that you're taking away money from the families that 00:50:33.560 --> 00:50:35.390 align:middle line:84% do go to school that would otherwise not 00:50:35.390 --> 00:50:38.300 align:middle line:90% have gone to school, right? 00:50:38.300 --> 00:50:41.840 align:middle line:84% Some families now need to pay for the school expenditure, 00:50:41.840 --> 00:50:45.630 align:middle line:84% and the scholarship doesn't cover all of it. 00:50:45.630 --> 00:50:49.040 align:middle line:84% So on the one hand, you're giving extra money 00:50:49.040 --> 00:50:52.280 align:middle line:84% they wouldn't have had for people who would 00:50:52.280 --> 00:50:54.050 align:middle line:90% have gone to school anywhere. 00:50:54.050 --> 00:50:56.990 align:middle line:84% On the other hand, you are taking money away, 00:50:56.990 --> 00:50:59.870 align:middle line:84% or you're making people spend more on education 00:50:59.870 --> 00:51:00.770 align:middle line:90% than they would have. 00:51:00.770 --> 00:51:01.680 align:middle line:90% Those are the compliant. 00:51:01.680 --> 00:51:03.638 align:middle line:84% They end up spending more than they would have. 00:51:03.638 --> 00:51:05.930 align:middle line:84% In terms of budget, it balances out, 00:51:05.930 --> 00:51:12.410 align:middle line:84% given how the point estimate shook out. 00:51:12.410 --> 00:51:14.600 align:middle line:84% But that's one effect going one way, 00:51:14.600 --> 00:51:16.290 align:middle line:90% one effect going the other way. 00:51:16.290 --> 00:51:18.200 align:middle line:84% You have no guarantee that they even out. 00:51:18.200 --> 00:51:19.280 align:middle line:90% So that's one problem. 00:51:19.280 --> 00:51:23.120 align:middle line:84% What would be another problem, another potential violation 00:51:23.120 --> 00:51:25.820 align:middle line:84% of exclusion restriction coming from the scholarship? 00:51:25.820 --> 00:51:41.070 align:middle line:90% 00:51:41.070 --> 00:51:42.042 align:middle line:90% Yep. 00:51:42.042 --> 00:51:43.750 align:middle line:84% AUDIENCE: You followed them for 10 years. 00:51:43.750 --> 00:51:46.033 align:middle line:84% So maybe they feel very motivated to do the things 00:51:46.033 --> 00:51:47.825 align:middle line:84% because a team of international researchers 00:51:47.825 --> 00:51:49.320 align:middle line:90% followed them around? 00:51:49.320 --> 00:51:51.445 align:middle line:84% ESTHER DUFLO: Well, I'm following the control group 00:51:51.445 --> 00:51:52.070 align:middle line:90% as well. 00:51:52.070 --> 00:51:54.930 align:middle line:84% AUDIENCE: Oh, but they know they were given-- 00:51:54.930 --> 00:51:56.940 align:middle line:90% ESTHER DUFLO: Yeah, yeah. 00:51:56.940 --> 00:51:58.200 align:middle line:90% They might be grateful. 00:51:58.200 --> 00:52:00.060 align:middle line:84% I wish they were even more grateful so we 00:52:00.060 --> 00:52:01.380 align:middle line:90% can follow them even easier. 00:52:01.380 --> 00:52:03.180 align:middle line:90% But they might be grateful, yes. 00:52:03.180 --> 00:52:05.610 align:middle line:84% And in general, and this gratefulness 00:52:05.610 --> 00:52:09.185 align:middle line:84% might be even bigger because we are following them. 00:52:09.185 --> 00:52:10.560 align:middle line:84% We are calling them all the time. 00:52:10.560 --> 00:52:11.610 align:middle line:90% We are monitoring them. 00:52:11.610 --> 00:52:14.050 align:middle line:90% They feel they have to perform. 00:52:14.050 --> 00:52:16.440 align:middle line:84% And you mentioned got given, I think. 00:52:16.440 --> 00:52:18.780 align:middle line:84% So in general, even in the absence of the experiment, 00:52:18.780 --> 00:52:22.170 align:middle line:84% they might feel psychologically, oh, I got a scholarship. 00:52:22.170 --> 00:52:25.140 align:middle line:84% I'd better do something with it, and therefore 00:52:25.140 --> 00:52:29.970 align:middle line:84% be motivated to do more, which might bias their earnings 00:52:29.970 --> 00:52:30.510 align:middle line:90% either way. 00:52:30.510 --> 00:52:31.885 align:middle line:84% Because they might say, well, I'm 00:52:31.885 --> 00:52:33.210 align:middle line:90% not going to take any odd job. 00:52:33.210 --> 00:52:35.100 align:middle line:84% And therefore, I'm going to wait for something better. 00:52:35.100 --> 00:52:36.180 align:middle line:90% I already had this luck. 00:52:36.180 --> 00:52:37.347 align:middle line:90% I'm going to have some more. 00:52:37.347 --> 00:52:39.970 align:middle line:84% Alternatively, I'd better perform. 00:52:39.970 --> 00:52:42.060 align:middle line:84% So there might be direct psychological effect 00:52:42.060 --> 00:52:44.218 align:middle line:90% of getting the scholarship. 00:52:44.218 --> 00:52:45.760 align:middle line:84% There might be direct-- in this case, 00:52:45.760 --> 00:52:50.040 align:middle line:84% the scholarship was not tied to performance. 00:52:50.040 --> 00:52:52.810 align:middle line:90% But it often is, for example. 00:52:52.810 --> 00:52:56.250 align:middle line:84% There is a famous paper by Josh Angrist and Michael Kremer 00:52:56.250 --> 00:53:00.240 align:middle line:84% on estimating the return to private school. 00:53:00.240 --> 00:53:02.220 align:middle line:84% But they gave people a scholarship 00:53:02.220 --> 00:53:04.980 align:middle line:84% to go to private school that were 00:53:04.980 --> 00:53:08.100 align:middle line:90% removed if people repeated. 00:53:08.100 --> 00:53:10.770 align:middle line:84% And I always found it very funny to use that as an instrument 00:53:10.770 --> 00:53:14.190 align:middle line:84% because you might expect a direct effect of motivation 00:53:14.190 --> 00:53:18.690 align:middle line:84% effect just to get the scholarship irrespective of-- 00:53:18.690 --> 00:53:19.935 align:middle line:90% so that's one issue. 00:53:19.935 --> 00:53:22.750 align:middle line:90% 00:53:22.750 --> 00:53:25.350 align:middle line:84% The incentives, the feeling very confident, 00:53:25.350 --> 00:53:28.710 align:middle line:90% you want something, et cetera-- 00:53:28.710 --> 00:53:30.520 align:middle line:90% any other that you can think of? 00:53:30.520 --> 00:53:31.260 align:middle line:90% Yeah. 00:53:31.260 --> 00:53:34.620 align:middle line:84% AUDIENCE: Maybe since they feel grateful that you give them 00:53:34.620 --> 00:53:38.340 align:middle line:84% a scholarship, they feel tempted to say they're doing 00:53:38.340 --> 00:53:40.618 align:middle line:90% better than they actually are. 00:53:40.618 --> 00:53:41.410 align:middle line:90% ESTHER DUFLO: Yeah. 00:53:41.410 --> 00:53:45.680 align:middle line:84% So then there might be a social experiment effect, 00:53:45.680 --> 00:53:51.280 align:middle line:84% or social bias in reporting, which I don't think 00:53:51.280 --> 00:53:52.447 align:middle line:90% is a huge deal in this case. 00:53:52.447 --> 00:53:54.488 align:middle line:84% But it's a very good thing you're pointing it out 00:53:54.488 --> 00:53:56.440 align:middle line:84% because in a lot of experiments, it might be. 00:53:56.440 --> 00:53:58.000 align:middle line:84% And by the way, this is something 00:53:58.000 --> 00:54:00.700 align:middle line:84% that it's not even a problem of instruments. 00:54:00.700 --> 00:54:02.950 align:middle line:84% Or it could be even more because the people who 00:54:02.950 --> 00:54:05.650 align:middle line:84% are on scholarship could feel even more that they 00:54:05.650 --> 00:54:08.900 align:middle line:84% feel they need to report higher wages because of that. 00:54:08.900 --> 00:54:09.400 align:middle line:90% Yeah. 00:54:09.400 --> 00:54:13.030 align:middle line:84% So you could have a social desirability in the reporting. 00:54:13.030 --> 00:54:17.100 align:middle line:90% 00:54:17.100 --> 00:54:19.500 align:middle line:84% For example, it's there in the reporting of their wages. 00:54:19.500 --> 00:54:21.690 align:middle line:84% And then for example, on their test scores, 00:54:21.690 --> 00:54:23.738 align:middle line:84% they might put a bit more effort. 00:54:23.738 --> 00:54:25.530 align:middle line:84% So in fact, for effort, on the test scores, 00:54:25.530 --> 00:54:28.980 align:middle line:84% we try to track that-- how much effort they 00:54:28.980 --> 00:54:33.130 align:middle line:90% are putting on the test itself. 00:54:33.130 --> 00:54:35.820 align:middle line:84% So all of this would create issues. 00:54:35.820 --> 00:54:39.010 align:middle line:84% And I'm sure you could cook up some more, given a bit more 00:54:39.010 --> 00:54:39.510 align:middle line:90% time. 00:54:39.510 --> 00:54:42.440 align:middle line:90% 00:54:42.440 --> 00:54:45.140 align:middle line:84% Forgetting that for a moment, let's assume 00:54:45.140 --> 00:54:46.910 align:middle line:90% this is not an issue. 00:54:46.910 --> 00:54:49.010 align:middle line:84% Some of this can be looked at in the data. 00:54:49.010 --> 00:54:51.920 align:middle line:90% Many of them cannot. 00:54:51.920 --> 00:54:53.780 align:middle line:90% Any money can be looked at. 00:54:53.780 --> 00:54:58.010 align:middle line:84% It can be looked at only in circumstantial ways 00:54:58.010 --> 00:54:58.952 align:middle line:90% on the back side. 00:54:58.952 --> 00:55:00.410 align:middle line:84% For example, with the effort thing, 00:55:00.410 --> 00:55:02.930 align:middle line:84% you can try and measure how much effort they put in. 00:55:02.930 --> 00:55:05.510 align:middle line:90% 00:55:05.510 --> 00:55:09.170 align:middle line:84% Now forgetting that issue for a moment, at least 00:55:09.170 --> 00:55:12.670 align:middle line:84% for a pedagogical purpose, we can ask ourselves 00:55:12.670 --> 00:55:15.770 align:middle line:84% for whom the estimates-- we are calculating them from whom. 00:55:15.770 --> 00:55:19.340 align:middle line:84% So already, we know it's for a sample of kids who qualified 00:55:19.340 --> 00:55:20.630 align:middle line:90% but didn't have money. 00:55:20.630 --> 00:55:23.370 align:middle line:84% So that's already a first restriction. 00:55:23.370 --> 00:55:28.130 align:middle line:84% Now with an IV, there is one more. 00:55:28.130 --> 00:55:32.630 align:middle line:84% As soon as you stop assuming the constant return to education, 00:55:32.630 --> 00:55:36.800 align:middle line:84% some kids would have gone to school anyways. 00:55:36.800 --> 00:55:40.790 align:middle line:84% And we call these guys the compliers. 00:55:40.790 --> 00:55:42.500 align:middle line:90% They always take your-- 00:55:42.500 --> 00:55:46.010 align:middle line:84% they always go to school, so they always 00:55:46.010 --> 00:55:47.190 align:middle line:90% take your treatment. 00:55:47.190 --> 00:55:51.560 align:middle line:84% So some kids did not go to school even 00:55:51.560 --> 00:55:52.610 align:middle line:90% without the scholarship. 00:55:52.610 --> 00:55:56.210 align:middle line:84% We call these guys the never-takers. 00:55:56.210 --> 00:55:59.550 align:middle line:84% And some kids were moved by the scholarship to go to school. 00:55:59.550 --> 00:56:01.190 align:middle line:84% We know how many of those exist-- 00:56:01.190 --> 00:56:03.320 align:middle line:90% 25%. 00:56:03.320 --> 00:56:07.490 align:middle line:84% We can also know a lot of things about them 00:56:07.490 --> 00:56:10.430 align:middle line:84% by looking at the average characteristics of those 00:56:10.430 --> 00:56:12.770 align:middle line:84% who go to school with or without the scholarship. 00:56:12.770 --> 00:56:15.020 align:middle line:84% This is from work by Alberto Abadie. 00:56:15.020 --> 00:56:18.200 align:middle line:84% But we never know who they are, exactly. 00:56:18.200 --> 00:56:19.970 align:middle line:84% We call these guys the compliers. 00:56:19.970 --> 00:56:22.770 align:middle line:84% Those are the people who go to school with the scholarship 00:56:22.770 --> 00:56:25.790 align:middle line:90% but do not go without. 00:56:25.790 --> 00:56:30.562 align:middle line:84% So one question of interest is how the returns line up 00:56:30.562 --> 00:56:33.020 align:middle line:84% if they are not homogeneous, if these different groups have 00:56:33.020 --> 00:56:33.728 align:middle line:90% different groups. 00:56:33.728 --> 00:56:36.770 align:middle line:84% Who are we estimating the return to education? 00:56:36.770 --> 00:56:40.650 align:middle line:84% So that takes a little bit of algebra. 00:56:40.650 --> 00:56:42.140 align:middle line:90% So we have now our instrument. 00:56:42.140 --> 00:56:46.070 align:middle line:84% We go back to the-- and a little bit of notation 00:56:46.070 --> 00:56:50.360 align:middle line:84% because we have to keep track of both the instrument 00:56:50.360 --> 00:56:52.040 align:middle line:90% and the treatment. 00:56:52.040 --> 00:56:55.430 align:middle line:84% I'm going back to the [INAUDIBLE] notation of Wi 00:56:55.430 --> 00:56:57.570 align:middle line:90% for the treatment. 00:56:57.570 --> 00:57:02.420 align:middle line:84% So we have now a treatment of 1 if you got the scholarship, 00:57:02.420 --> 00:57:05.910 align:middle line:84% treatment 0 if you didn't get the scholarship. 00:57:05.910 --> 00:57:08.420 align:middle line:84% So we have potential outcome with respect to scholarship 00:57:08.420 --> 00:57:12.080 align:middle line:84% and potential outcome with respect to the treatment. 00:57:12.080 --> 00:57:16.910 align:middle line:84% The observed treatment is like we had before-- 00:57:16.910 --> 00:57:22.670 align:middle line:84% Zi, Wi of 1 plus 1 minus Ci, Wi of 0, OK? 00:57:22.670 --> 00:57:27.710 align:middle line:84% We observe people-- either they won or they lost. 00:57:27.710 --> 00:57:29.600 align:middle line:84% And so that's where the notation becomes 00:57:29.600 --> 00:57:32.390 align:middle line:84% a little bit tricky because we have 00:57:32.390 --> 00:57:38.130 align:middle line:84% Yi of 1 is the potential outcome if treated by Wi, 00:57:38.130 --> 00:57:40.760 align:middle line:90% not treated by Zi, OK? 00:57:40.760 --> 00:57:44.390 align:middle line:90% And similarly for Y0. 00:57:44.390 --> 00:57:48.050 align:middle line:84% The identification assumption-- so this 00:57:48.050 --> 00:57:54.560 align:middle line:84% is from a very classic paper by Imbens and Angrist, 00:57:54.560 --> 00:57:56.810 align:middle line:84% which one day will win them the Nobel Prize-- 00:57:56.810 --> 00:58:00.710 align:middle line:90% sometime soon, hopefully. 00:58:00.710 --> 00:58:03.290 align:middle line:84% So the assumption is that all the potential outcomes 00:58:03.290 --> 00:58:05.390 align:middle line:84% have to be independent of the instrument. 00:58:05.390 --> 00:58:12.380 align:middle line:84% That's the parallel of the assumption 00:58:12.380 --> 00:58:15.830 align:middle line:84% that we were discussing before about our epsilon, right? 00:58:15.830 --> 00:58:19.910 align:middle line:84% And again, this implies not only random assignment 00:58:19.910 --> 00:58:28.670 align:middle line:84% of the instrument, which ensures directly that this is 00:58:28.670 --> 00:58:33.170 align:middle line:84% independent, but also no-direct effect on the Y's, which gives 00:58:33.170 --> 00:58:34.950 align:middle line:90% you independence of the Y's. 00:58:34.950 --> 00:58:36.200 align:middle line:90% Otherwise, you don't get that. 00:58:36.200 --> 00:58:39.750 align:middle line:90% 00:58:39.750 --> 00:58:43.170 align:middle line:84% In addition, we are going to make one extra assumption, 00:58:43.170 --> 00:58:45.230 align:middle line:90% which is monotonicity. 00:58:45.230 --> 00:58:51.920 align:middle line:84% What monotonicity says is that you're either 00:58:51.920 --> 00:58:55.190 align:middle line:84% as likely to be treated with the instrument down to 1, 00:58:55.190 --> 00:59:00.290 align:middle line:84% or more likely to be treated, but not less. 00:59:00.290 --> 00:59:05.180 align:middle line:84% So people either never go to school regardless, or always 00:59:05.180 --> 00:59:07.070 align:middle line:84% go to school regardless, or are more 00:59:07.070 --> 00:59:09.290 align:middle line:84% likely to go to school because of the instrument. 00:59:09.290 --> 00:59:11.690 align:middle line:84% But there are nobody who is a defier, which 00:59:11.690 --> 00:59:14.847 align:middle line:84% is they would go to school if they don't have 00:59:14.847 --> 00:59:17.180 align:middle line:84% the scholarship, but they refuse to go to school if they 00:59:17.180 --> 00:59:19.250 align:middle line:90% have a scholarship. 00:59:19.250 --> 00:59:21.680 align:middle line:90% So that's an extra assumption. 00:59:21.680 --> 00:59:25.790 align:middle line:84% In further work by [INAUDIBLE] in particular, 00:59:25.790 --> 00:59:29.915 align:middle line:84% you can work on some relaxing on this monotonicity. 00:59:29.915 --> 00:59:33.470 align:middle line:90% 00:59:33.470 --> 00:59:35.420 align:middle line:84% It's also not particularly testable, 00:59:35.420 --> 00:59:40.623 align:middle line:90% so it's good to relax it. 00:59:40.623 --> 00:59:43.040 align:middle line:84% But for now, we are going to assume that's the assumption. 00:59:43.040 --> 00:59:46.310 align:middle line:84% So you are either compliant and always 00:59:46.310 --> 00:59:49.730 align:middle line:84% take care, or never take care, but you are not a defier. 00:59:49.730 --> 00:59:51.980 align:middle line:84% In most experiments, honestly, that's 00:59:51.980 --> 00:59:55.760 align:middle line:84% a much more plausible assumption than independence anyways. 00:59:55.760 --> 00:59:57.380 align:middle line:84% So there is not much to worry about, 00:59:57.380 --> 01:00:00.680 align:middle line:84% but it's good to know that it's happening. 01:00:00.680 --> 01:00:02.488 align:middle line:90% So we now have our groups. 01:00:02.488 --> 01:00:03.530 align:middle line:90% We don't have our defier. 01:00:03.530 --> 01:00:06.080 align:middle line:90% 01:00:06.080 --> 01:00:11.970 align:middle line:84% And now we can start doing some algebra. 01:00:11.970 --> 01:00:19.100 align:middle line:84% So like before, we are going to add and subtract some terms. 01:00:19.100 --> 01:00:24.810 align:middle line:84% And we are going to express Yi as a full function of the W. 01:00:24.810 --> 01:00:27.310 align:middle line:90% So that's what I'm doing here. 01:00:27.310 --> 01:00:29.720 align:middle line:90% This Yi is this one. 01:00:29.720 --> 01:00:31.490 align:middle line:90% And then this is this one. 01:00:31.490 --> 01:00:34.950 align:middle line:90% 01:00:34.950 --> 01:00:38.250 align:middle line:90% And then I can replace some-- 01:00:38.250 --> 01:00:41.820 align:middle line:84% I can move things because the expectation 01:00:41.820 --> 01:00:45.720 align:middle line:90% can be combined nicely. 01:00:45.720 --> 01:00:50.190 align:middle line:84% To go from here to here, you can just write things down 01:00:50.190 --> 01:00:51.720 align:middle line:90% and combine them. 01:00:51.720 --> 01:00:53.820 align:middle line:84% You will be convinced that it works. 01:00:53.820 --> 01:00:57.690 align:middle line:84% So I get here this product of Yi of 1 01:00:57.690 --> 01:01:01.800 align:middle line:84% minus Yi of 0 multiplied by Yi of 1 minus Yi of 0. 01:01:01.800 --> 01:01:04.920 align:middle line:84% And these guys-- so what can I do with these guys? 01:01:04.920 --> 01:01:08.670 align:middle line:90% 01:01:08.670 --> 01:01:11.810 align:middle line:84% I can get rid of them by independence. 01:01:11.810 --> 01:01:12.770 align:middle line:90% That's nice. 01:01:12.770 --> 01:01:19.430 align:middle line:84% So now that we have that, we have to assess how that is. 01:01:19.430 --> 01:01:23.960 align:middle line:84% And it's easy to deal with because this is really 01:01:23.960 --> 01:01:29.300 align:middle line:90% only 1 or 0 or minus 1, right? 01:01:29.300 --> 01:01:33.680 align:middle line:90% In fact, can it be minus 1? 01:01:33.680 --> 01:01:37.220 align:middle line:84% It cannot be minus 1 by monotonicity. 01:01:37.220 --> 01:01:39.410 align:middle line:90% So I'm left with-- 01:01:39.410 --> 01:01:42.610 align:middle line:90% so this one goes away. 01:01:42.610 --> 01:01:43.950 align:middle line:90% Then there is this one. 01:01:43.950 --> 01:01:48.830 align:middle line:84% So basically, I'm taking all of the cases one by one. 01:01:48.830 --> 01:01:51.230 align:middle line:90% Either this is minus 1-- 01:01:51.230 --> 01:01:54.110 align:middle line:84% the property of that is 0, so that's going to go away-- 01:01:54.110 --> 01:01:55.940 align:middle line:90% or this is this one. 01:01:55.940 --> 01:01:58.850 align:middle line:84% That's also going to go away simply because it's 0. 01:01:58.850 --> 01:02:01.970 align:middle line:84% And then I'm left with only this one. 01:02:01.970 --> 01:02:05.710 align:middle line:90% 01:02:05.710 --> 01:02:08.770 align:middle line:90% And so that's the product of-- 01:02:08.770 --> 01:02:12.025 align:middle line:84% so then we want to modify that a little bit. 01:02:12.025 --> 01:02:15.580 align:middle line:90% 01:02:15.580 --> 01:02:18.220 align:middle line:90% And that turns out to be-- 01:02:18.220 --> 01:02:22.270 align:middle line:90% 01:02:22.270 --> 01:02:30.780 align:middle line:84% this probability is simply the reduced form on education. 01:02:30.780 --> 01:02:34.620 align:middle line:84% So if I divide this by the reduced form on education, 01:02:34.620 --> 01:02:39.510 align:middle line:84% which is my Wald estimate, it is equal to the treatment 01:02:39.510 --> 01:02:47.040 align:middle line:84% effect for people who have a Wi of 1 minus Wi of 0 equal to 1. 01:02:47.040 --> 01:02:48.585 align:middle line:90% So it's a treatment effect for-- 01:02:48.585 --> 01:02:52.470 align:middle line:90% 01:02:52.470 --> 01:02:57.960 align:middle line:84% so the ratio of this difference, the reduced form 01:02:57.960 --> 01:03:00.510 align:middle line:84% by the first stage, which is the Wald estimate, 01:03:00.510 --> 01:03:05.947 align:middle line:84% is the treatment effect for this group of people, who are who? 01:03:05.947 --> 01:03:06.780 align:middle line:90% AUDIENCE: Compliers. 01:03:06.780 --> 01:03:08.440 align:middle line:84% ESTHER DUFLO: Who are the compliers. 01:03:08.440 --> 01:03:10.050 align:middle line:84% So that's something you probably know. 01:03:10.050 --> 01:03:13.500 align:middle line:84% But I think it's good to go to why it's true once. 01:03:13.500 --> 01:03:15.870 align:middle line:90% And I even spared you that. 01:03:15.870 --> 01:03:18.540 align:middle line:84% Normally, I do it on the board, but I want to have a little-- 01:03:18.540 --> 01:03:21.150 align:middle line:90% 01:03:21.150 --> 01:03:25.290 align:middle line:84% just very simple, beautiful pieces of algebra 01:03:25.290 --> 01:03:29.985 align:middle line:84% that summarize why this gives you the effect on the complier. 01:03:29.985 --> 01:03:33.280 align:middle line:90% 01:03:33.280 --> 01:03:35.180 align:middle line:90% So that's very nice. 01:03:35.180 --> 01:03:38.500 align:middle line:90% There are some special cases. 01:03:38.500 --> 01:03:44.140 align:middle line:84% If the controls take control, for example, in our case, 01:03:44.140 --> 01:03:46.180 align:middle line:84% in the case of the [INAUDIBLE] program, 01:03:46.180 --> 01:03:52.330 align:middle line:84% where some of the treatment refused to take the program, 01:03:52.330 --> 01:03:56.080 align:middle line:84% then the compliers are the treated. 01:03:56.080 --> 01:03:59.350 align:middle line:84% Then it's the effect of treatment on the treated. 01:03:59.350 --> 01:04:01.360 align:middle line:84% In the general case, these are those 01:04:01.360 --> 01:04:02.860 align:middle line:84% who are compelled by the instruments 01:04:02.860 --> 01:04:05.060 align:middle line:90% to get the treatment. 01:04:05.060 --> 01:04:08.772 align:middle line:84% And so there is further issue of external validity-- 01:04:08.772 --> 01:04:11.290 align:middle line:84% all above the fact that we were in a small sample. 01:04:11.290 --> 01:04:13.120 align:middle line:84% In addition, within that sample, we 01:04:13.120 --> 01:04:16.480 align:middle line:84% selected the people who would have gone to school 01:04:16.480 --> 01:04:20.620 align:middle line:90% just because of the instrument. 01:04:20.620 --> 01:04:23.440 align:middle line:90% And then that tells us that-- 01:04:23.440 --> 01:04:26.770 align:middle line:84% it gives us some guidance on how to pick instruments. 01:04:26.770 --> 01:04:28.750 align:middle line:84% Because if you pick instruments that 01:04:28.750 --> 01:04:34.480 align:middle line:84% are going to select some exotic group of people, 01:04:34.480 --> 01:04:39.160 align:middle line:84% then your estimate is also not going to be of great relevance. 01:04:39.160 --> 01:04:42.310 align:middle line:84% Here we are selecting the people who, 01:04:42.310 --> 01:04:45.710 align:middle line:84% because of the scholarship, are going to school. 01:04:45.710 --> 01:04:48.050 align:middle line:84% So that's a group that is of interest. 01:04:48.050 --> 01:04:49.490 align:middle line:90% So that's nice. 01:04:49.490 --> 01:04:51.640 align:middle line:90% We like that. 01:04:51.640 --> 01:04:55.690 align:middle line:84% So that means that you want to go 01:04:55.690 --> 01:04:58.750 align:middle line:84% beyond finding an instrument that is going to-- not only 01:04:58.750 --> 01:05:02.380 align:middle line:84% your instrument has to satisfy independence, 01:05:02.380 --> 01:05:05.650 align:middle line:84% satisfy independence of the potential outcomes 01:05:05.650 --> 01:05:08.620 align:middle line:84% with respect to the instrument to satisfy monotonicity-- 01:05:08.620 --> 01:05:11.950 align:middle line:84% in addition, it has to be interesting in the sense 01:05:11.950 --> 01:05:15.040 align:middle line:84% that it moves an interesting group of compliers. 01:05:15.040 --> 01:05:17.680 align:middle line:84% So this one satisfies because in the real world, 01:05:17.680 --> 01:05:20.410 align:middle line:84% if a government were thinking of removing fees of education, 01:05:20.410 --> 01:05:23.072 align:middle line:84% presumably, they would do that, keeping in mind 01:05:23.072 --> 01:05:25.030 align:middle line:84% the fact that it's going to help the people who 01:05:25.030 --> 01:05:26.320 align:middle line:90% are credit-constrained today. 01:05:26.320 --> 01:05:29.185 align:middle line:84% And those are those people because they 01:05:29.185 --> 01:05:30.310 align:middle line:90% would like to go to school. 01:05:30.310 --> 01:05:32.768 align:middle line:84% And they go to school as soon as they have the opportunity. 01:05:32.768 --> 01:05:36.663 align:middle line:90% 01:05:36.663 --> 01:05:38.330 align:middle line:84% So I think we went already through this. 01:05:38.330 --> 01:05:40.520 align:middle line:90% So that's good. 01:05:40.520 --> 01:05:42.920 align:middle line:84% So an advantage of doing the IV is 01:05:42.920 --> 01:05:46.290 align:middle line:84% that we can compare with other strategies. 01:05:46.290 --> 01:05:48.140 align:middle line:84% So we already compared with the OLS. 01:05:48.140 --> 01:05:52.190 align:middle line:84% Already did that before showing you what the IV was. 01:05:52.190 --> 01:05:55.830 align:middle line:84% We can do a more fancy version of the OLS, 01:05:55.830 --> 01:05:58.340 align:middle line:84% which is the oils after controlling for everything 01:05:58.340 --> 01:06:03.110 align:middle line:84% we have in the database that could lead to selection. 01:06:03.110 --> 01:06:05.930 align:middle line:84% We have a very, very, very, very rich database. 01:06:05.930 --> 01:06:08.840 align:middle line:84% So you might think that maybe this 01:06:08.840 --> 01:06:10.673 align:middle line:84% is enough to capture all the selection. 01:06:10.673 --> 01:06:12.590 align:middle line:84% And in that case, we don't need to go over all 01:06:12.590 --> 01:06:16.370 align:middle line:90% of this complicated machinery. 01:06:16.370 --> 01:06:19.040 align:middle line:84% But there is one issue in doing that. 01:06:19.040 --> 01:06:21.800 align:middle line:84% So what we did in that paper is we 01:06:21.800 --> 01:06:24.110 align:middle line:84% applied the double machine learning algorithm 01:06:24.110 --> 01:06:28.640 align:middle line:84% that we developed in the paper by Victor, 01:06:28.640 --> 01:06:31.670 align:middle line:84% and Whitney, and many, many other coauthors, 01:06:31.670 --> 01:06:38.060 align:middle line:84% which is basically a fancy method to do Frisch-Waugh. 01:06:38.060 --> 01:06:40.285 align:middle line:84% If you have taken first-year econometrics, that's 01:06:40.285 --> 01:06:41.660 align:middle line:84% one of the first things you learn 01:06:41.660 --> 01:06:44.480 align:middle line:84% is Frisch-Waugh decomposition, which is basically 01:06:44.480 --> 01:06:47.660 align:middle line:84% you can project what you're interested in on what 01:06:47.660 --> 01:06:49.340 align:middle line:90% you are not interested in. 01:06:49.340 --> 01:06:55.120 align:middle line:84% And then what you're interested in is the Y and the S. 01:06:55.120 --> 01:06:57.050 align:middle line:84% And then when you regress Y on S, 01:06:57.050 --> 01:07:01.970 align:middle line:84% you have-- when you regress the residual 01:07:01.970 --> 01:07:03.920 align:middle line:84% from the projection on each other, 01:07:03.920 --> 01:07:08.340 align:middle line:84% you get the same beta coefficient of interest. 01:07:08.340 --> 01:07:11.493 align:middle line:84% So this is basically an application of Frisch-Waugh 01:07:11.493 --> 01:07:13.660 align:middle line:84% and saying, well, we are going to do this projection 01:07:13.660 --> 01:07:18.470 align:middle line:84% in as flexible a way as possible using machine learning, 01:07:18.470 --> 01:07:21.000 align:middle line:90% and then run the regression. 01:07:21.000 --> 01:07:22.520 align:middle line:84% So that's double machine learning. 01:07:22.520 --> 01:07:27.260 align:middle line:84% Then you have to do that in two samples, hence the double. 01:07:27.260 --> 01:07:31.040 align:middle line:90% But the idea is that. 01:07:31.040 --> 01:07:31.790 align:middle line:90% So that was nice. 01:07:31.790 --> 01:07:33.500 align:middle line:84% We were very happy with our paper. 01:07:33.500 --> 01:07:36.650 align:middle line:84% And this is, in a sense, the first test of this paper 01:07:36.650 --> 01:07:40.670 align:middle line:84% because it works very nicely in Monte Carlo simulation, where 01:07:40.670 --> 01:07:42.230 align:middle line:84% all the assumptions are satisfied. 01:07:42.230 --> 01:07:44.480 align:middle line:84% But we never know whether all the assumptions 01:07:44.480 --> 01:07:46.500 align:middle line:90% are satisfied for DML. 01:07:46.500 --> 01:07:49.190 align:middle line:84% So what we have here is comparing the double machine 01:07:49.190 --> 01:07:53.090 align:middle line:84% learning estimates to the OLS and then to the IV. 01:07:53.090 --> 01:07:57.050 align:middle line:84% And it's really not that great because most of the DML 01:07:57.050 --> 01:07:59.300 align:middle line:90% look much closer to the OLS. 01:07:59.300 --> 01:08:02.000 align:middle line:84% And despite throwing all this machinery at the problem, 01:08:02.000 --> 01:08:06.390 align:middle line:90% you don't make so much headway. 01:08:06.390 --> 01:08:08.420 align:middle line:84% So it seems that whatever is the selection, 01:08:08.420 --> 01:08:11.360 align:middle line:84% it's not easy to capture it from even very, very 01:08:11.360 --> 01:08:12.770 align:middle line:90% rich [? observers. ?] 01:08:12.770 --> 01:08:14.450 align:middle line:84% Now you might think, well, all of that 01:08:14.450 --> 01:08:20.689 align:middle line:84% is not correct because in fact, the IV is 01:08:20.689 --> 01:08:23.830 align:middle line:90% the effect for the compliers. 01:08:23.830 --> 01:08:26.819 align:middle line:84% So they can't be compared to the double machine learning 01:08:26.819 --> 01:08:29.490 align:middle line:84% estimate or to the OLS, which is the effect 01:08:29.490 --> 01:08:31.529 align:middle line:90% for the population at large-- 01:08:31.529 --> 01:08:35.880 align:middle line:84% or at least, some other weighting scheme. 01:08:35.880 --> 01:08:39.862 align:middle line:84% The weighting scheme in OLS gives more weight 01:08:39.862 --> 01:08:41.279 align:middle line:84% to the observation that contribute 01:08:41.279 --> 01:08:42.569 align:middle line:90% the more to the variation. 01:08:42.569 --> 01:08:45.149 align:middle line:84% The weighting scheme for DML is who knows? 01:08:45.149 --> 01:08:45.850 align:middle line:90% I don't know. 01:08:45.850 --> 01:08:47.830 align:middle line:90% It's a mess. 01:08:47.830 --> 01:08:49.899 align:middle line:84% So what we did here is that we said, 01:08:49.899 --> 01:08:52.200 align:middle line:84% well, let's do a double machine learning estimate. 01:08:52.200 --> 01:08:57.000 align:middle line:84% That is going to mimic as much as we 01:08:57.000 --> 01:08:59.010 align:middle line:84% can from the observable variable-- 01:08:59.010 --> 01:09:02.760 align:middle line:84% the LATE estimate, which is we are going to estimate 01:09:02.760 --> 01:09:05.460 align:middle line:84% the treatment effect for a population that 01:09:05.460 --> 01:09:10.229 align:middle line:84% looks on the observable variable, of which we are many. 01:09:10.229 --> 01:09:13.200 align:middle line:84% Remember, this is the hope-- the HD row here 01:09:13.200 --> 01:09:15.510 align:middle line:84% is that the observer will do a good job in capturing 01:09:15.510 --> 01:09:17.093 align:middle line:84% the [? endogeneities ?] in the sample. 01:09:17.093 --> 01:09:19.500 align:middle line:90% 01:09:19.500 --> 01:09:23.460 align:middle line:84% Trying to estimate the machine learning estimate, 01:09:23.460 --> 01:09:27.300 align:middle line:84% the machine learning corrected estimate of education 01:09:27.300 --> 01:09:32.279 align:middle line:84% on a population that is weighted appropriately 01:09:32.279 --> 01:09:34.979 align:middle line:84% to look observationally, like people who 01:09:34.979 --> 01:09:38.439 align:middle line:90% are affected by the instrument. 01:09:38.439 --> 01:09:41.960 align:middle line:84% So what we need to do now is to apply our machinery 01:09:41.960 --> 01:09:45.660 align:middle line:84% in our other people to look at heterogeneous treatment effect 01:09:45.660 --> 01:09:48.120 align:middle line:90% by axes. 01:09:48.120 --> 01:09:51.450 align:middle line:84% We calculate the heterogenous treatment effect by axes. 01:09:51.450 --> 01:09:53.710 align:middle line:84% We don't actually want to give them a causal estimate. 01:09:53.710 --> 01:09:57.600 align:middle line:84% But we just use machine learning to estimate, 01:09:57.600 --> 01:10:01.540 align:middle line:84% to get a sense of the heterogeneity in the treatment 01:10:01.540 --> 01:10:02.040 align:middle line:90% effect. 01:10:02.040 --> 01:10:05.880 align:middle line:84% And then we use those estimates to reweight so that we have 01:10:05.880 --> 01:10:09.330 align:middle line:84% the DML LATE is a machine learning estimate 01:10:09.330 --> 01:10:12.960 align:middle line:84% for a population that looks like the complier in terms 01:10:12.960 --> 01:10:15.390 align:middle line:84% of all of the observables that we have. 01:10:15.390 --> 01:10:18.360 align:middle line:84% That is, it puts more weight on people-- 01:10:18.360 --> 01:10:20.430 align:middle line:84% on the type, on the kind of people 01:10:20.430 --> 01:10:24.680 align:middle line:84% who seems to be more affected by the instruments. 01:10:24.680 --> 01:10:27.680 align:middle line:84% After all of this effort, this is almost the same as the DML 01:10:27.680 --> 01:10:29.120 align:middle line:90% and almost the same as the OLS. 01:10:29.120 --> 01:10:31.880 align:middle line:84% So this is not-- this didn't make much of a difference. 01:10:31.880 --> 01:10:32.880 align:middle line:90% Yeah? 01:10:32.880 --> 01:10:35.650 align:middle line:84% AUDIENCE: I just don't understand-- 01:10:35.650 --> 01:10:39.540 align:middle line:84% if you don't believe the exclusion restriction, 01:10:39.540 --> 01:10:41.522 align:middle line:90% why did you use the IV as the-- 01:10:41.522 --> 01:10:42.480 align:middle line:90% ESTHER DUFLO: Oh, yeah. 01:10:42.480 --> 01:10:44.790 align:middle line:90% Until now, I believe it. 01:10:44.790 --> 01:10:49.110 align:middle line:84% Yeah, so you could say, well, this is all-- 01:10:49.110 --> 01:10:51.390 align:middle line:84% all I can tell you now is that the DML is 01:10:51.390 --> 01:10:55.200 align:middle line:84% very similar to OLS despite the addition of the controls. 01:10:55.200 --> 01:10:57.570 align:middle line:84% The ML LATE is still very similar to the OLS 01:10:57.570 --> 01:11:00.420 align:middle line:84% despite weighting appropriately for the observable. 01:11:00.420 --> 01:11:02.610 align:middle line:84% And they're all quite different from the IV. 01:11:02.610 --> 01:11:06.390 align:middle line:84% But you could tell me I don't believe the IV anyways. 01:11:06.390 --> 01:11:10.740 align:middle line:84% So there is nothing here that tells you 01:11:10.740 --> 01:11:13.920 align:middle line:84% you have to believe in one versus the other. 01:11:13.920 --> 01:11:18.190 align:middle line:84% AUDIENCE: Could you compare it to the [INAUDIBLE]?? 01:11:18.190 --> 01:11:19.210 align:middle line:90% ESTHER DUFLO: No. 01:11:19.210 --> 01:11:21.340 align:middle line:84% You could, but you would get nothing from it. 01:11:21.340 --> 01:11:24.030 align:middle line:90% 01:11:24.030 --> 01:11:25.440 align:middle line:90% Well, what you could do-- 01:11:25.440 --> 01:11:30.480 align:middle line:84% and would be interesting, but we couldn't do in this data set-- 01:11:30.480 --> 01:11:33.060 align:middle line:90% is if you knew who-- 01:11:33.060 --> 01:11:38.010 align:middle line:84% you could compare a DML estimate of getting a scholarship 01:11:38.010 --> 01:11:42.660 align:middle line:84% to an OLS estimate of getting a scholarship in the control. 01:11:42.660 --> 01:11:44.820 align:middle line:84% Not going to school, but you don't 01:11:44.820 --> 01:11:51.030 align:middle line:84% want to compare the going to school to getting a scholarship 01:11:51.030 --> 01:11:52.950 align:middle line:84% because that's not the same thing. 01:11:52.950 --> 01:11:55.440 align:middle line:84% You definitely have to scale them somehow. 01:11:55.440 --> 01:11:57.150 align:middle line:84% And then whether or not it's appropriate 01:11:57.150 --> 01:12:01.410 align:middle line:84% depends on whether or not your exclusion restrictions 01:12:01.410 --> 01:12:02.250 align:middle line:90% are valid. 01:12:02.250 --> 01:12:05.460 align:middle line:84% But to not have this problem, you 01:12:05.460 --> 01:12:11.640 align:middle line:84% could find a setup where you are going to compare-- 01:12:11.640 --> 01:12:13.770 align:middle line:84% you would like to compare an OLS to an OLS. 01:12:13.770 --> 01:12:15.930 align:middle line:84% So in my case, it would be, can I 01:12:15.930 --> 01:12:21.180 align:middle line:84% compare the effect of getting a scholarship not randomly-- 01:12:21.180 --> 01:12:22.650 align:middle line:84% in the control group, for example, 01:12:22.650 --> 01:12:24.900 align:middle line:84% some kids got from the sugar board. 01:12:24.900 --> 01:12:27.000 align:middle line:84% Some kids got from a Christian church. 01:12:27.000 --> 01:12:29.520 align:middle line:90% Some kids got from whatever. 01:12:29.520 --> 01:12:33.090 align:middle line:84% Can I compare the effect of getting a scholarship 01:12:33.090 --> 01:12:37.080 align:middle line:84% to the effect of getting a randomly assigned scholarship? 01:12:37.080 --> 01:12:40.410 align:middle line:84% And that would be a much easier comparison. 01:12:40.410 --> 01:12:43.890 align:middle line:84% But we don't really have that in our data set. 01:12:43.890 --> 01:12:46.470 align:middle line:84% In fact, a lot of people scramble the money somehow 01:12:46.470 --> 01:12:48.240 align:middle line:90% without, really, a scholarship. 01:12:48.240 --> 01:12:50.833 align:middle line:84% So we don't have a way of doing that. 01:12:50.833 --> 01:12:52.500 align:middle line:84% But that's what you would like, exactly. 01:12:52.500 --> 01:12:54.610 align:middle line:84% In the original LaLonde exercise-- 01:12:54.610 --> 01:12:56.730 align:middle line:84% Bob LaLonde is a researcher at Chicago 01:12:56.730 --> 01:13:01.980 align:middle line:84% who compared a lot of non-experimental estimates 01:13:01.980 --> 01:13:06.060 align:middle line:84% of training programs with the experimental estimates. 01:13:06.060 --> 01:13:09.330 align:middle line:90% And the way he did it is by-- 01:13:09.330 --> 01:13:11.880 align:middle line:84% this was comparing like to like because he was comparing 01:13:11.880 --> 01:13:16.960 align:middle line:84% assignment to training to going into training of your own will. 01:13:16.960 --> 01:13:20.425 align:middle line:84% He didn't have DML, but he did matching and whatever. 01:13:20.425 --> 01:13:22.800 align:middle line:84% Actually, he didn't do matching because he didn't have it 01:13:22.800 --> 01:13:23.430 align:middle line:90% at the time. 01:13:23.430 --> 01:13:27.270 align:middle line:84% He did OLS, OLS with control, diff and diff, et cetera. 01:13:27.270 --> 01:13:28.950 align:middle line:84% And then some people added matching. 01:13:28.950 --> 01:13:31.770 align:middle line:90% And one could now add this. 01:13:31.770 --> 01:13:33.580 align:middle line:84% But I haven't seen people doing that. 01:13:33.580 --> 01:13:35.640 align:middle line:84% So you're absolutely right that ours 01:13:35.640 --> 01:13:38.370 align:middle line:84% is a bit more fragile because you have to believe in the IV 01:13:38.370 --> 01:13:40.320 align:middle line:84% to believe in the comparison being valid. 01:13:40.320 --> 01:13:44.190 align:middle line:90% 01:13:44.190 --> 01:13:44.690 align:middle line:90% OK. 01:13:44.690 --> 01:13:46.940 align:middle line:84% So a minute on the substantive findings. 01:13:46.940 --> 01:13:48.950 align:middle line:84% So we find positive effect on all sorts 01:13:48.950 --> 01:13:51.350 align:middle line:90% of sides of education. 01:13:51.350 --> 01:13:54.110 align:middle line:90% Financial return to education-- 01:13:54.110 --> 01:13:57.800 align:middle line:84% they are in the sense, maybe not so much on earnings, 01:13:57.800 --> 01:13:59.190 align:middle line:90% but on this-- 01:13:59.190 --> 01:14:01.880 align:middle line:90% some people get these perks. 01:14:01.880 --> 01:14:06.840 align:middle line:84% But they seem to be mostly private returns. 01:14:06.840 --> 01:14:09.840 align:middle line:84% I'm almost running out of time, so I'm 01:14:09.840 --> 01:14:13.500 align:middle line:84% going to skip most of describing how we build 01:14:13.500 --> 01:14:19.050 align:middle line:84% an IV from the diff and diff estimator in the INPRES paper 01:14:19.050 --> 01:14:20.050 align:middle line:90% because you all read it. 01:14:20.050 --> 01:14:21.425 align:middle line:84% And once you've done that, that's 01:14:21.425 --> 01:14:26.340 align:middle line:84% just the same except that's a valid diff and diff instead of 01:14:26.340 --> 01:14:28.320 align:middle line:90% a valid from simple difference. 01:14:28.320 --> 01:14:30.330 align:middle line:84% And then you construct your instrument 01:14:30.330 --> 01:14:34.995 align:middle line:84% as the cohort interacted with the treatment effect-- 01:14:34.995 --> 01:14:37.980 align:middle line:84% or cohort interacted with the number of schools 01:14:37.980 --> 01:14:41.670 align:middle line:90% built in the region of birth. 01:14:41.670 --> 01:14:44.790 align:middle line:90% So I'll skip that part. 01:14:44.790 --> 01:14:47.680 align:middle line:84% Then we find that differently from this setup, 01:14:47.680 --> 01:14:53.010 align:middle line:84% we find that the IV is very, very similar to the OLS. 01:14:53.010 --> 01:14:57.910 align:middle line:84% The IV and the OLS look really almost the same. 01:14:57.910 --> 01:15:02.460 align:middle line:84% So in that case, we didn't find much evidence of-- 01:15:02.460 --> 01:15:06.660 align:middle line:84% but what I want to talk about for two minutes 01:15:06.660 --> 01:15:11.260 align:middle line:84% is once we've done that, so we estimate the treatment effect. 01:15:11.260 --> 01:15:16.890 align:middle line:84% One thing that-- the follow-on on the [INAUDIBLE] paper, 01:15:16.890 --> 01:15:20.130 align:middle line:84% follow-on in the compliers-- if your instrument takes 01:15:20.130 --> 01:15:23.250 align:middle line:84% a lot of discrete value, not only 01:15:23.250 --> 01:15:25.440 align:middle line:84% the effect for the complier, but you 01:15:25.440 --> 01:15:27.330 align:middle line:84% can show that it's a weighted average 01:15:27.330 --> 01:15:31.380 align:middle line:84% of the effect of each year of education, where the weight is 01:15:31.380 --> 01:15:35.550 align:middle line:84% given by the fraction of people who are moved to complete 01:15:35.550 --> 01:15:40.480 align:middle line:84% each extra year, which is why, if you'll remember, when I 01:15:40.480 --> 01:15:42.670 align:middle line:84% started this project, I wanted to estimate 01:15:42.670 --> 01:15:46.030 align:middle line:84% the nonparametric return to education. 01:15:46.030 --> 01:15:49.420 align:middle line:84% But that didn't work out because all of my variation 01:15:49.420 --> 01:15:50.770 align:middle line:90% comes from primary school. 01:15:50.770 --> 01:15:53.445 align:middle line:90% 01:15:53.445 --> 01:15:55.070 align:middle line:84% So there was no way I was going to look 01:15:55.070 --> 01:15:58.100 align:middle line:84% at the effect of secondary school. 01:15:58.100 --> 01:16:00.830 align:middle line:84% I wouldn't worry about this negative estimate. 01:16:00.830 --> 01:16:06.650 align:middle line:84% These confidence bends are hugely too small because they 01:16:06.650 --> 01:16:07.790 align:middle line:90% are point-by-point. 01:16:07.790 --> 01:16:09.652 align:middle line:84% And they don't take into account clustering. 01:16:09.652 --> 01:16:12.110 align:middle line:84% And when you do all that, the standards are actually large. 01:16:12.110 --> 01:16:16.130 align:middle line:84% So you can treat that as a whole big 0 01:16:16.130 --> 01:16:19.400 align:middle line:84% on both this one and this one-- the small spillover 01:16:19.400 --> 01:16:22.912 align:middle line:84% on junior high and the negative on senior high. 01:16:22.912 --> 01:16:24.620 align:middle line:84% Really, it's driven by the fact that they 01:16:24.620 --> 01:16:29.480 align:middle line:84% built more senior high schools where they had more 01:16:29.480 --> 01:16:30.820 align:middle line:90% primary schools to start with. 01:16:30.820 --> 01:16:32.990 align:middle line:84% So the school building of senior high school 01:16:32.990 --> 01:16:35.220 align:middle line:84% was negatively correlated with school building 01:16:35.220 --> 01:16:38.120 align:middle line:84% of private school, which itself is an issue for the exclusion 01:16:38.120 --> 01:16:39.320 align:middle line:90% restriction. 01:16:39.320 --> 01:16:40.940 align:middle line:84% But anyway, all the variation was 01:16:40.940 --> 01:16:44.090 align:middle line:84% coming from my complier, or my poor kids, 01:16:44.090 --> 01:16:48.050 align:middle line:84% who are driven from doing one more year of education, 01:16:48.050 --> 01:16:49.430 align:middle line:90% two more years of education. 01:16:49.430 --> 01:16:51.320 align:middle line:84% So there was no way I was going to estimate 01:16:51.320 --> 01:16:53.913 align:middle line:84% a nice, nonparametric curve like the one I showed you 01:16:53.913 --> 01:16:55.580 align:middle line:84% at the beginning of the graph, which had 01:16:55.580 --> 01:16:57.260 align:middle line:90% been my hope at the beginning. 01:16:57.260 --> 01:17:00.410 align:middle line:90% So that never happened. 01:17:00.410 --> 01:17:03.590 align:middle line:84% But what did happen is something many of you asked. 01:17:03.590 --> 01:17:07.430 align:middle line:84% So I want to at least advertise thinking 01:17:07.430 --> 01:17:09.860 align:middle line:84% about the social returns to education. 01:17:09.860 --> 01:17:13.890 align:middle line:90% 01:17:13.890 --> 01:17:16.205 align:middle line:84% And something which we talk about in the Ghana paper, 01:17:16.205 --> 01:17:18.830 align:middle line:84% but couldn't do because we have an instrument that just affects 01:17:18.830 --> 01:17:23.570 align:middle line:84% individuals-- but here are the effects, the education policy-- 01:17:23.570 --> 01:17:25.580 align:middle line:90% it [INAUDIBLE]. 01:17:25.580 --> 01:17:28.100 align:middle line:84% And unless people migrate like crazy 01:17:28.100 --> 01:17:31.460 align:middle line:84% to undo all of the variation in education, which 01:17:31.460 --> 01:17:34.760 align:middle line:84% we know they don't, then that really 01:17:34.760 --> 01:17:37.580 align:middle line:84% provides, in a sense, a shock in the education level 01:17:37.580 --> 01:17:39.410 align:middle line:90% at the market level. 01:17:39.410 --> 01:17:43.070 align:middle line:84% So we at least have one hope to estimate not only 01:17:43.070 --> 01:17:46.190 align:middle line:84% that the effect of education on myself, 01:17:46.190 --> 01:17:47.870 align:middle line:90% but it should not be a beta. 01:17:47.870 --> 01:17:49.730 align:middle line:84% It should be a beta 1 or a gamma-- 01:17:49.730 --> 01:17:53.360 align:middle line:84% the effect of the average education of other people 01:17:53.360 --> 01:17:58.350 align:middle line:84% in my district on my own earnings. 01:17:58.350 --> 01:17:59.870 align:middle line:90% So the estimation problem-- 01:17:59.870 --> 01:18:00.890 align:middle line:90% now we have two things. 01:18:00.890 --> 01:18:03.920 align:middle line:84% We need an instrument for my own education 01:18:03.920 --> 01:18:08.390 align:middle line:84% and an instrument for the average education 01:18:08.390 --> 01:18:11.360 align:middle line:90% in the neighborhood. 01:18:11.360 --> 01:18:14.240 align:middle line:84% And here we have a nice setup for that 01:18:14.240 --> 01:18:19.910 align:middle line:84% because if you consider a cohort that was 12 or older in 1973, 01:18:19.910 --> 01:18:23.270 align:middle line:84% and that's not exposed by the program, 01:18:23.270 --> 01:18:25.040 align:middle line:84% then what the program does to them 01:18:25.040 --> 01:18:27.980 align:middle line:84% is that when they enter the labor market-- 01:18:27.980 --> 01:18:33.710 align:middle line:84% let's say they start entering the labor market in 1980-- 01:18:33.710 --> 01:18:36.590 align:middle line:84% every year they age, they get a new cohort 01:18:36.590 --> 01:18:42.650 align:middle line:84% of educated kids who are coming to compete with them. 01:18:42.650 --> 01:18:44.880 align:middle line:84% Themselves-- never benefited from the school. 01:18:44.880 --> 01:18:48.920 align:middle line:84% So for this group of people, there is no worry 01:18:48.920 --> 01:18:52.970 align:middle line:84% that the education level, the same thing that 01:18:52.970 --> 01:18:54.860 align:middle line:84% moved the average education level also 01:18:54.860 --> 01:18:57.440 align:middle line:90% moved their own education. 01:18:57.440 --> 01:19:02.120 align:middle line:84% We don't have that bias that might occur in other setups 01:19:02.120 --> 01:19:06.860 align:middle line:84% because we know they weren't exposed to the schools. 01:19:06.860 --> 01:19:12.110 align:middle line:84% So we can-- but starting in '79, there 01:19:12.110 --> 01:19:17.270 align:middle line:84% is a slow increase in the average education 01:19:17.270 --> 01:19:18.420 align:middle line:90% of the other guys. 01:19:18.420 --> 01:19:21.560 align:middle line:84% So this is the effect of my own-- 01:19:21.560 --> 01:19:26.960 align:middle line:84% of the program on the education of the older cohort. 01:19:26.960 --> 01:19:30.350 align:middle line:84% So this is the same pattern of coefficient 01:19:30.350 --> 01:19:34.340 align:middle line:84% we had before except that now I had to take one labor 01:19:34.340 --> 01:19:35.960 align:middle line:90% survey for every year. 01:19:35.960 --> 01:19:41.670 align:middle line:84% Because I'm now looking at the education of-- all cohorts are 01:19:41.670 --> 01:19:44.690 align:middle line:90% observing-- '86, '87, '88-- 01:19:44.690 --> 01:19:49.320 align:middle line:84% as a function of DD relative to the first 0 here. 01:19:49.320 --> 01:19:51.680 align:middle line:84% So the old cohort doesn't get more educated 01:19:51.680 --> 01:19:54.300 align:middle line:90% as they get older. 01:19:54.300 --> 01:19:55.850 align:middle line:90% They have missed out. 01:19:55.850 --> 01:19:59.570 align:middle line:84% But this is now the average of the young around them. 01:19:59.570 --> 01:20:02.270 align:middle line:84% And you can see that the average of the young around them 01:20:02.270 --> 01:20:06.230 align:middle line:84% keeps increasing because these new guys finish-- 01:20:06.230 --> 01:20:10.740 align:middle line:84% or girls-- guys, in this instance, finish school. 01:20:10.740 --> 01:20:13.160 align:middle line:90% And therefore, they show up. 01:20:13.160 --> 01:20:18.060 align:middle line:84% So now we can do the same thing for the wage of the older. 01:20:18.060 --> 01:20:21.180 align:middle line:84% And if we find any effect, positive or negative, that's 01:20:21.180 --> 01:20:25.530 align:middle line:84% going to tell us that if there is a rising 01:20:25.530 --> 01:20:30.240 align:middle line:84% effect of the program with every year 01:20:30.240 --> 01:20:33.570 align:middle line:84% that passes on their wages, it's the fact 01:20:33.570 --> 01:20:36.880 align:middle line:84% that they compete with those guys. 01:20:36.880 --> 01:20:42.160 align:middle line:84% And what we see is that there is actually a decline-- 01:20:42.160 --> 01:20:45.670 align:middle line:84% the coefficient-- the first one is higher. 01:20:45.670 --> 01:20:47.860 align:middle line:84% But then they tend to go down, and down, 01:20:47.860 --> 01:20:50.030 align:middle line:90% and down, and down, and down. 01:20:50.030 --> 01:20:51.410 align:middle line:90% This is super imprecise. 01:20:51.410 --> 01:20:53.920 align:middle line:90% So the actual tables are mush. 01:20:53.920 --> 01:20:57.400 align:middle line:84% But if anything, what you're looking at, what you're finding 01:20:57.400 --> 01:21:03.400 align:middle line:84% is, in fact, a decrease in wages for the old educated, 01:21:03.400 --> 01:21:07.060 align:middle line:84% for the old cohort as these other people 01:21:07.060 --> 01:21:08.650 align:middle line:90% enter the labor market. 01:21:08.650 --> 01:21:12.280 align:middle line:84% And this decrease in wages is as large for the educated 01:21:12.280 --> 01:21:13.760 align:middle line:90% and the uneducated worker. 01:21:13.760 --> 01:21:16.990 align:middle line:84% So it doesn't go to a decline return to education. 01:21:16.990 --> 01:21:20.620 align:middle line:84% So they appear to be substitute and substituted away 01:21:20.620 --> 01:21:22.640 align:middle line:90% by those guys. 01:21:22.640 --> 01:21:26.800 align:middle line:90% And even though the table-- 01:21:26.800 --> 01:21:30.790 align:middle line:84% the results are super imprecise, you can hold your nose 01:21:30.790 --> 01:21:33.220 align:middle line:84% and do a calibration to say, what does this 01:21:33.220 --> 01:21:36.670 align:middle line:84% imply about the movement of capital 01:21:36.670 --> 01:21:40.750 align:middle line:84% to this region when more people got educated? 01:21:40.750 --> 01:21:42.160 align:middle line:84% And basically, you can't rule out 01:21:42.160 --> 01:21:45.080 align:middle line:84% that the capital is a fixed factor. 01:21:45.080 --> 01:21:48.760 align:middle line:84% So no one realized that there were suddenly 01:21:48.760 --> 01:21:50.290 align:middle line:84% all these educated people around. 01:21:50.290 --> 01:21:52.900 align:middle line:84% And they should have built more factories, 01:21:52.900 --> 01:21:53.900 align:middle line:90% and so on and so forth. 01:21:53.900 --> 01:21:56.483 align:middle line:84% It's either they didn't realize or they didn't have the credit 01:21:56.483 --> 01:21:57.140 align:middle line:90% to do it. 01:21:57.140 --> 01:22:00.760 align:middle line:84% And therefore, the capital end up 01:22:00.760 --> 01:22:06.010 align:middle line:84% being a limiting factor in this case. 01:22:06.010 --> 01:22:08.620 align:middle line:84% So I write down a model, which has land. 01:22:08.620 --> 01:22:12.460 align:middle line:84% And people can work in factories. 01:22:12.460 --> 01:22:14.830 align:middle line:84% And then I calibrate how much capital-- 01:22:14.830 --> 01:22:17.560 align:middle line:84% land is fixed, and then how much capital could have come in. 01:22:17.560 --> 01:22:22.420 align:middle line:84% And I show that this is compatible with none 01:22:22.420 --> 01:22:23.840 align:middle line:90% in some simulation. 01:22:23.840 --> 01:22:28.300 align:middle line:84% This was more of an intellectual exercise than a contribution 01:22:28.300 --> 01:22:32.050 align:middle line:84% that I want you to take as-- don't take, necessarily, 01:22:32.050 --> 01:22:33.880 align:middle line:90% the substantive lessons home. 01:22:33.880 --> 01:22:37.810 align:middle line:90% Because it's very imprecise. 01:22:37.810 --> 01:22:42.170 align:middle line:84% But technically, it's a useful exercise to do. 01:22:42.170 --> 01:22:43.420 align:middle line:90% OK, let's stop here. 01:22:43.420 --> 01:22:45.400 align:middle line:84% Sorry for running over to some extent. 01:22:45.400 --> 01:22:49.330 align:middle line:84% And we'll talk about the supply side of education next week. 01:22:49.330 --> 01:22:52.380 align:middle line:90% [SIDE CONVERSATIONS] 01:22:52.380 --> 01:23:04.000 align:middle line:90%