WEBVTT 00:00:00.000 --> 00:00:03.479 align:middle line:84% [SQUEAKING] [RUSTLING] [CLICKING] 00:00:03.479 --> 00:00:15.828 align:middle line:90% 00:00:15.828 --> 00:00:18.120 align:middle line:84% GLENN ELLISON: OK, I guess I will start the tradition I 00:00:18.120 --> 00:00:20.250 align:middle line:84% started last time of starting at 9:04. 00:00:20.250 --> 00:00:23.013 align:middle line:84% So if you remember last time, I finished a lecture right 00:00:23.013 --> 00:00:24.180 align:middle line:90% in the middle of an example. 00:00:24.180 --> 00:00:27.150 align:middle line:84% I was talking about monopoly information 00:00:27.150 --> 00:00:32.377 align:middle line:84% design, which is about how monopolists can affect-- 00:00:32.377 --> 00:00:34.710 align:middle line:84% by changing the information that consumers receive about 00:00:34.710 --> 00:00:36.400 align:middle line:84% their product can affect the prices. 00:00:36.400 --> 00:00:39.813 align:middle line:84% And this is motivation, both for firms 00:00:39.813 --> 00:00:41.730 align:middle line:84% who are giving information about their product 00:00:41.730 --> 00:00:44.130 align:middle line:84% to consumers to try to maximize profits, 00:00:44.130 --> 00:00:47.610 align:middle line:84% or potentially for platforms like Amazon giving information 00:00:47.610 --> 00:00:50.905 align:middle line:84% to consumers to try to increase the consumer surplus 00:00:50.905 --> 00:00:53.280 align:middle line:84% and limit the monopoly power that the monopolist is going 00:00:53.280 --> 00:00:55.200 align:middle line:90% to exert. 00:00:55.200 --> 00:00:57.960 align:middle line:84% I had been multiplying everything by 18 00:00:57.960 --> 00:01:00.880 align:middle line:84% when I talked about the examples last class in my head. 00:01:00.880 --> 00:01:05.880 align:middle line:84% So I made this about the Sharan sandwich at Clover 00:01:05.880 --> 00:01:08.550 align:middle line:84% and said that if consumers had full information 00:01:08.550 --> 00:01:11.820 align:middle line:84% about the sandwich, their values would be uniform. 00:01:11.820 --> 00:01:17.090 align:middle line:84% On 0, 18, there's a cost of $6 of producing the sandwich. 00:01:17.090 --> 00:01:18.970 align:middle line:84% And I discussed a few different-- 00:01:18.970 --> 00:01:21.820 align:middle line:84% how consumer surplus and profit would 00:01:21.820 --> 00:01:24.520 align:middle line:84% change a few different information structures. 00:01:24.520 --> 00:01:26.410 align:middle line:84% If you give the consumers full information, 00:01:26.410 --> 00:01:28.930 align:middle line:84% then their values are uniform 0, 18. 00:01:28.930 --> 00:01:32.650 align:middle line:84% When your cost is 6, the monopoly profit 00:01:32.650 --> 00:01:36.640 align:middle line:84% is halfway in between 6 and 18, which is 12. 00:01:36.640 --> 00:01:38.980 align:middle line:84% Another example I discussed is suppose Clover 00:01:38.980 --> 00:01:41.050 align:middle line:84% puts the sandwich in a bag and tells you nothing 00:01:41.050 --> 00:01:42.350 align:middle line:90% other than it's a new sandwich. 00:01:42.350 --> 00:01:44.180 align:middle line:84% We're not telling you anything about it. 00:01:44.180 --> 00:01:47.890 align:middle line:84% Then all consumers would have values of 9. 00:01:47.890 --> 00:01:51.200 align:middle line:84% And so the monopoly price would be to set a price of 9. 00:01:51.200 --> 00:01:53.230 align:middle line:84% And in this case, they actually make more money, 00:01:53.230 --> 00:01:56.050 align:middle line:84% but consumer surplus is completely eliminated. 00:01:56.050 --> 00:01:59.202 align:middle line:84% And then I gave this partial information example, where 00:01:59.202 --> 00:02:01.660 align:middle line:84% I gave the consumers partial information about the product, 00:02:01.660 --> 00:02:03.890 align:middle line:84% then ended up with a monopoly price of 10. 00:02:03.890 --> 00:02:06.340 align:middle line:84% And consumers and the firm were both better off 00:02:06.340 --> 00:02:07.610 align:middle line:90% than with full information. 00:02:07.610 --> 00:02:10.780 align:middle line:84% And this is what that example looked like. 00:02:10.780 --> 00:02:13.300 align:middle line:84% I know people often ask me, How on earth could 00:02:13.300 --> 00:02:14.600 align:middle line:90% a firm do this in practice? 00:02:14.600 --> 00:02:16.725 align:middle line:84% And let's imagine-- you can think of this two ways. 00:02:16.725 --> 00:02:21.610 align:middle line:84% One is, is this a model of an omniscient firm like Amazon 00:02:21.610 --> 00:02:24.590 align:middle line:84% or Google or Apple that knows so much about you, 00:02:24.590 --> 00:02:26.140 align:middle line:84% they know your preferences better 00:02:26.140 --> 00:02:28.690 align:middle line:84% than you do because they know what people like 00:02:28.690 --> 00:02:29.830 align:middle line:90% you thought of the product? 00:02:29.830 --> 00:02:32.330 align:middle line:84% So they can actually tell you about what your value is going 00:02:32.330 --> 00:02:34.420 align:middle line:90% to be if you eat that sandwich. 00:02:34.420 --> 00:02:36.400 align:middle line:84% The other example could be, let's imagine 00:02:36.400 --> 00:02:39.280 align:middle line:84% that the value for having a sandwich depends on how much 00:02:39.280 --> 00:02:41.650 align:middle line:84% you like cauliflower and how much you 00:02:41.650 --> 00:02:44.630 align:middle line:84% like the Indo-Chinese sauce that it comes in. 00:02:44.630 --> 00:02:48.740 align:middle line:84% And initially, you know nothing about the sandwich. 00:02:48.740 --> 00:02:50.830 align:middle line:84% But then one thing they can tell you is, 00:02:50.830 --> 00:02:54.340 align:middle line:84% I'm just telling you that the sandwich is made of cauliflower 00:02:54.340 --> 00:02:55.490 align:middle line:90% as the base. 00:02:55.490 --> 00:02:57.400 align:middle line:84% And then people who don't like cauliflower 00:02:57.400 --> 00:02:59.350 align:middle line:84% are going to have values between 0 and 6, 00:02:59.350 --> 00:03:01.180 align:middle line:84% depending on how much they like the sauce. 00:03:01.180 --> 00:03:03.340 align:middle line:84% Those who have a medium preference for cauliflower 00:03:03.340 --> 00:03:04.368 align:middle line:90% would have these values. 00:03:04.368 --> 00:03:06.160 align:middle line:84% Those who have a high value for cauliflower 00:03:06.160 --> 00:03:07.070 align:middle line:90% would have those values. 00:03:07.070 --> 00:03:09.010 align:middle line:84% So it's like, we're not going to tell you what the sauce is like 00:03:09.010 --> 00:03:11.410 align:middle line:84% or the vegetables, but we'll tell you that cauliflower 00:03:11.410 --> 00:03:13.270 align:middle line:90% is the base of the sandwich. 00:03:13.270 --> 00:03:16.090 align:middle line:84% Anyway, but what happened with this example was 00:03:16.090 --> 00:03:18.190 align:middle line:84% all the consumers, after getting the signal, 00:03:18.190 --> 00:03:20.350 align:middle line:90% had values of 3, 10, and 16. 00:03:20.350 --> 00:03:23.080 align:middle line:84% And that ended up making the monopoly price 10. 00:03:23.080 --> 00:03:26.800 align:middle line:84% So anyway, the monopoly information design literature 00:03:26.800 --> 00:03:29.510 align:middle line:84% is about how can we think more generally 00:03:29.510 --> 00:03:31.520 align:middle line:84% about what types of information we 00:03:31.520 --> 00:03:34.610 align:middle line:84% want to give consumers to either maximize profits or maximize 00:03:34.610 --> 00:03:38.250 align:middle line:84% consumer surplus or do something in between? 00:03:38.250 --> 00:03:43.260 align:middle line:84% So the Roesler and Szentes paper that I put on the reading list 00:03:43.260 --> 00:03:46.560 align:middle line:84% assumes that the values are now distributed according 00:03:46.560 --> 00:03:49.360 align:middle line:90% to some distribution F on 0, 1. 00:03:49.360 --> 00:03:52.600 align:middle line:84% It specializes that the cost is 0. 00:03:52.600 --> 00:03:56.160 align:middle line:84% I'm going to write mu for the mean value of the consumers. 00:03:56.160 --> 00:03:58.860 align:middle line:84% And what I'm going to do is give this firm 00:03:58.860 --> 00:04:02.070 align:middle line:84% tremendous ability to manipulate the signal distribution. 00:04:02.070 --> 00:04:04.020 align:middle line:84% What the firm can do is choose signals 00:04:04.020 --> 00:04:07.980 align:middle line:84% to choose any joint distribution whatsoever on v and s. 00:04:07.980 --> 00:04:10.750 align:middle line:84% So the most general set of signal structures 00:04:10.750 --> 00:04:15.300 align:middle line:84% is they're just some joint distribution on v and s. 00:04:15.300 --> 00:04:17.935 align:middle line:84% Notice that when consumers get signal s, 00:04:17.935 --> 00:04:19.560 align:middle line:84% what they're going to be willing to pay 00:04:19.560 --> 00:04:21.779 align:middle line:90% is expectation of v given s. 00:04:21.779 --> 00:04:25.410 align:middle line:84% So rather than giving them a signal with a label like low, 00:04:25.410 --> 00:04:28.980 align:middle line:84% medium, high, I might as well just say, 00:04:28.980 --> 00:04:31.950 align:middle line:84% your signal is either 3 or 10 or 16. 00:04:31.950 --> 00:04:35.470 align:middle line:90% 00:04:35.470 --> 00:04:37.720 align:middle line:84% Given the information about the product I'm revealing, 00:04:37.720 --> 00:04:40.660 align:middle line:84% your expected value is 3, your expected value is 10, 00:04:40.660 --> 00:04:42.520 align:middle line:90% your expected value is 16. 00:04:42.520 --> 00:04:44.172 align:middle line:90% And so you might as well-- 00:04:44.172 --> 00:04:46.630 align:middle line:84% consumers are going to do the Bayesian updating of whatever 00:04:46.630 --> 00:04:47.950 align:middle line:90% letters or words you give them. 00:04:47.950 --> 00:04:50.328 align:middle line:84% So you might as well just do the direct signal structure, 00:04:50.328 --> 00:04:52.120 align:middle line:84% just tell them what their expected value is 00:04:52.120 --> 00:04:54.520 align:middle line:90% given what you've known. 00:04:54.520 --> 00:04:59.600 align:middle line:84% Now, obviously, I can't pick any distribution over values of s 00:04:59.600 --> 00:05:02.500 align:middle line:90% I'm going to give them. 00:05:02.500 --> 00:05:05.470 align:middle line:84% v is going to be s plus epsilon, where expectation 00:05:05.470 --> 00:05:07.730 align:middle line:90% of epsilon given s is 0. 00:05:07.730 --> 00:05:10.000 align:middle line:90% And so that means two things. 00:05:10.000 --> 00:05:13.720 align:middle line:84% One thing that it means is that the expectation of the signals 00:05:13.720 --> 00:05:18.100 align:middle line:84% must be mu because the expectation of v 00:05:18.100 --> 00:05:23.140 align:middle line:84% is expectation of s plus epsilon given s-- 00:05:23.140 --> 00:05:24.190 align:middle line:90% iterated expectations. 00:05:24.190 --> 00:05:26.210 align:middle line:84% I guess you must've done [14.]380 already. 00:05:26.210 --> 00:05:29.038 align:middle line:84% And so the expectation of the signal distribution 00:05:29.038 --> 00:05:30.580 align:middle line:84% has to be the same as the expectation 00:05:30.580 --> 00:05:32.800 align:middle line:90% of the prior distribution. 00:05:32.800 --> 00:05:34.690 align:middle line:84% And the other thing is that F is going 00:05:34.690 --> 00:05:39.640 align:middle line:84% to be the distribution of v, the true values, 00:05:39.640 --> 00:05:43.950 align:middle line:84% is a mean-preserving spread of the distribution of g. 00:05:43.950 --> 00:05:46.740 align:middle line:84% And so what you want to think about the monopolist is doing 00:05:46.740 --> 00:05:50.850 align:middle line:84% is just choosing a distribution g over possible signals, 00:05:50.850 --> 00:05:52.510 align:middle line:84% subject to those two constraints-- 00:05:52.510 --> 00:05:57.300 align:middle line:84% I've got to choose a distribution that has mean mu, 00:05:57.300 --> 00:05:59.730 align:middle line:84% and I have to choose a distribution such that F is 00:05:59.730 --> 00:06:04.410 align:middle line:84% a mean-preserving spread of the distribution that I pick. 00:06:04.410 --> 00:06:07.440 align:middle line:84% So, for instance, if this is the true distribution, 00:06:07.440 --> 00:06:11.580 align:middle line:84% I can't give signals like this because this distribution 00:06:11.580 --> 00:06:14.640 align:middle line:84% of finite support is not a mean-preserving spread 00:06:14.640 --> 00:06:17.340 align:middle line:90% of that bigger distribution. 00:06:17.340 --> 00:06:19.620 align:middle line:84% I could choose something like this 00:06:19.620 --> 00:06:22.910 align:middle line:84% because this thing is a mean-preserving spread of that. 00:06:22.910 --> 00:06:29.620 align:middle line:90% 00:06:29.620 --> 00:06:33.860 align:middle line:84% So let me say, the two problems here, 00:06:33.860 --> 00:06:35.920 align:middle line:84% one is the profit maximization problem. 00:06:35.920 --> 00:06:38.760 align:middle line:84% The other is the consumer surplus maximization problem. 00:06:38.760 --> 00:06:41.470 align:middle line:90% Profit maximization is trivial. 00:06:41.470 --> 00:06:44.020 align:middle line:84% It's economically probably more important. 00:06:44.020 --> 00:06:45.810 align:middle line:84% No one ever writes about it because what 00:06:45.810 --> 00:06:47.470 align:middle line:90% do you do to maximize profits? 00:06:47.470 --> 00:06:50.460 align:middle line:84% You just tell consumers your value is greater 00:06:50.460 --> 00:06:54.142 align:middle line:84% than c or your value is less than c. 00:06:54.142 --> 00:06:56.100 align:middle line:84% And then if you tell them your value is greater 00:06:56.100 --> 00:06:59.190 align:middle line:84% than c, you sell it to them at the price of p star, 00:06:59.190 --> 00:07:01.680 align:middle line:84% which is just the expectation of v, 00:07:01.680 --> 00:07:05.400 align:middle line:90% given that v is greater than c. 00:07:05.400 --> 00:07:08.520 align:middle line:84% So this simple pricing scheme, just tell them your value 00:07:08.520 --> 00:07:11.730 align:middle line:84% is above c or below c, when you hear 00:07:11.730 --> 00:07:14.850 align:middle line:84% your value is above c, every consumer who 00:07:14.850 --> 00:07:17.560 align:middle line:84% hears their value is above c is willing to pay at this price. 00:07:17.560 --> 00:07:19.350 align:middle line:84% So we get the fully efficient outcome, 00:07:19.350 --> 00:07:22.980 align:middle line:84% and we fully extract all the consumer surplus. 00:07:22.980 --> 00:07:24.510 align:middle line:84% So that's clearly the best thing you 00:07:24.510 --> 00:07:26.593 align:middle line:84% could do because you're maximizing social surplus, 00:07:26.593 --> 00:07:29.490 align:middle line:84% and the monopolist gets all of it. 00:07:29.490 --> 00:07:31.815 align:middle line:84% What do we do if, instead, we're Amazon, 00:07:31.815 --> 00:07:35.268 align:middle line:84% and we're trying to maximize consumer surplus? 00:07:35.268 --> 00:07:36.810 align:middle line:84% Intuitively, what you're trying to do 00:07:36.810 --> 00:07:40.860 align:middle line:84% is get the monopolist to charge the lowest price it can 00:07:40.860 --> 00:07:43.230 align:middle line:84% and sell to the largest number of consumers that it can, 00:07:43.230 --> 00:07:46.320 align:middle line:84% because that's the way you generate consumer surplus is get 00:07:46.320 --> 00:07:47.985 align:middle line:84% everyone with a value greater than c 00:07:47.985 --> 00:07:50.855 align:middle line:84% to buy and do it at the lowest possible price. 00:07:50.855 --> 00:07:53.570 align:middle line:90% 00:07:53.570 --> 00:07:55.040 align:middle line:84% Could potentially be some trade-off 00:07:55.040 --> 00:07:56.130 align:middle line:90% between those two objectives. 00:07:56.130 --> 00:07:57.110 align:middle line:90% It turns out it's not. 00:07:57.110 --> 00:08:01.070 align:middle line:84% In this problem, you want to get everybody to buy. 00:08:01.070 --> 00:08:02.040 align:middle line:90% So what do you do? 00:08:02.040 --> 00:08:03.385 align:middle line:90% The answer is this. 00:08:03.385 --> 00:08:06.465 align:middle line:90% 00:08:06.465 --> 00:08:08.090 align:middle line:84% I'm going to use some distribution that 00:08:08.090 --> 00:08:09.270 align:middle line:90% looks like that. 00:08:09.270 --> 00:08:13.130 align:middle line:84% And so the two-step argument to show that this is optimal 00:08:13.130 --> 00:08:16.700 align:middle line:84% is I'm going to first look at distributions of this form. 00:08:16.700 --> 00:08:19.157 align:middle line:84% I'm going to show that they work very well. 00:08:19.157 --> 00:08:20.990 align:middle line:84% I'm going to show what the maximum profit is 00:08:20.990 --> 00:08:22.980 align:middle line:84% I can get with the distribution of this form. 00:08:22.980 --> 00:08:25.355 align:middle line:84% And then I'm going to show that no other distribution can 00:08:25.355 --> 00:08:26.430 align:middle line:90% do better than this one. 00:08:26.430 --> 00:08:30.690 align:middle line:84% So it's like a guess and check answer, I guess. 00:08:30.690 --> 00:08:33.419 align:middle line:84% So what is that distribution formally? 00:08:33.419 --> 00:08:35.600 align:middle line:90% So I let the cdf-- 00:08:35.600 --> 00:08:37.470 align:middle line:90% I guess I graphed the pdf here. 00:08:37.470 --> 00:08:40.250 align:middle line:90% So the cdf looks like this. 00:08:40.250 --> 00:08:46.760 align:middle line:84% It starts at some value, s lower bar, goes flat like that, 00:08:46.760 --> 00:08:55.950 align:middle line:84% goes like that, and then has a jump and then goes across at 1. 00:08:55.950 --> 00:08:59.770 align:middle line:90% So this is s upper bar. 00:08:59.770 --> 00:09:04.750 align:middle line:90% So this is my graph of G of s. 00:09:04.750 --> 00:09:07.077 align:middle line:84% So this is the cumulative distribution function 00:09:07.077 --> 00:09:08.535 align:middle line:84% that I'm choosing the signals from. 00:09:08.535 --> 00:09:14.270 align:middle line:90% 00:09:14.270 --> 00:09:19.760 align:middle line:84% Notice that if the platform chooses this distribution, 00:09:19.760 --> 00:09:27.410 align:middle line:84% and the monopolist then picks any price p in this interval, 00:09:27.410 --> 00:09:33.140 align:middle line:84% the profit that the monopolist gets is p times 1 minus G of p. 00:09:33.140 --> 00:09:39.110 align:middle line:84% And p times 1 minus G of p is just p times s lower bar over p 00:09:39.110 --> 00:09:41.420 align:middle line:90% equals s lower bar. 00:09:41.420 --> 00:09:46.040 align:middle line:84% So what this distribution does is I choose that distribution, 00:09:46.040 --> 00:09:53.082 align:middle line:84% and I get this graph of s lower bar, s upper bar. 00:09:53.082 --> 00:09:54.540 align:middle line:84% I get this distribution of profits. 00:09:54.540 --> 00:09:59.900 align:middle line:84% So profits are if I charge a price of anything less than s 00:09:59.900 --> 00:10:01.880 align:middle line:90% lower bar, everybody buys. 00:10:01.880 --> 00:10:06.330 align:middle line:84% And so my profit is linear in the price that I charge. 00:10:06.330 --> 00:10:08.630 align:middle line:84% So I'm putting price on this axis. 00:10:08.630 --> 00:10:12.950 align:middle line:84% Then here, everybody buys, and I get a price of-- 00:10:12.950 --> 00:10:15.210 align:middle line:84% I get a profit of exactly s lower bar. 00:10:15.210 --> 00:10:16.520 align:middle line:90% And then I've made people-- 00:10:16.520 --> 00:10:21.150 align:middle line:84% I've made you indifferent over all prices in this range. 00:10:21.150 --> 00:10:23.220 align:middle line:90% And then my profit drops to 0. 00:10:23.220 --> 00:10:26.250 align:middle line:84% So this, I'm graphing the profit that you get charging p. 00:10:26.250 --> 00:10:31.430 align:middle line:90% 00:10:31.430 --> 00:10:36.650 align:middle line:84% So given that the monopolist is indifferent over all prices 00:10:36.650 --> 00:10:42.060 align:middle line:84% in this interval, one of the optimal prices is s lower bar. 00:10:42.060 --> 00:10:44.700 align:middle line:84% And so by setting this distribution, and then 00:10:44.700 --> 00:10:46.650 align:middle line:84% the monopolist is indifferent over all prices, 00:10:46.650 --> 00:10:48.890 align:middle line:90% it can charge price s lower bar. 00:10:48.890 --> 00:10:52.920 align:middle line:84% We observe we get profit s lower bar. 00:10:52.920 --> 00:10:56.940 align:middle line:84% And then the consumer surplus is the difference 00:10:56.940 --> 00:11:00.210 align:middle line:84% between each of these values and s lower bar. 00:11:00.210 --> 00:11:03.720 align:middle line:84% So if you integrate under this distribution, 00:11:03.720 --> 00:11:05.940 align:middle line:84% you're going to get what the consumer surplus is. 00:11:05.940 --> 00:11:08.220 align:middle line:84% And so by putting s lower bar down here, 00:11:08.220 --> 00:11:11.350 align:middle line:84% the lower I can put it, the more consumer surplus I'm generating. 00:11:11.350 --> 00:11:14.862 align:middle line:90% 00:11:14.862 --> 00:11:18.190 align:middle line:84% So within this class of G's, the consumer surplus optimal 00:11:18.190 --> 00:11:18.950 align:middle line:90% solution is clear. 00:11:18.950 --> 00:11:20.617 align:middle line:84% What I want to do is I just want to pick 00:11:20.617 --> 00:11:24.610 align:middle line:84% the smallest s lower bar that I can subject to the constraint 00:11:24.610 --> 00:11:30.190 align:middle line:84% that G has to have mean mu and F has 00:11:30.190 --> 00:11:35.610 align:middle line:84% to be a mean-preserving spread of G. How low 00:11:35.610 --> 00:11:37.380 align:middle line:90% can I go and still do that? 00:11:37.380 --> 00:11:41.190 align:middle line:84% Well, this distribution has the property that because it's only 00:11:41.190 --> 00:11:46.020 align:middle line:84% going down like 1/s, if I were to set s upper bar going towards 00:11:46.020 --> 00:11:49.470 align:middle line:84% infinity, the mean of this distribution would go 00:11:49.470 --> 00:11:50.950 align:middle line:90% to infinity. 00:11:50.950 --> 00:11:54.760 align:middle line:84% So for any s lower bar, I can just pick a high enough s lower 00:11:54.760 --> 00:11:57.240 align:middle line:90% bar and make the mean mu. 00:11:57.240 --> 00:12:03.240 align:middle line:84% But the problem I run into is I need this distribution to be 00:12:03.240 --> 00:12:06.998 align:middle line:84% a mean-preserving spread of F. You guys-- 00:12:06.998 --> 00:12:07.540 align:middle line:90% I don't know. 00:12:07.540 --> 00:12:09.248 align:middle line:84% Do they still even do this in statistics? 00:12:09.248 --> 00:12:11.840 align:middle line:84% What does a mean-preserving spread look like? 00:12:11.840 --> 00:12:16.160 align:middle line:84% So If this is G, and this is F, G is a mean-preserving spread 00:12:16.160 --> 00:12:19.070 align:middle line:84% of F. And it's a mean-preserving spread of F 00:12:19.070 --> 00:12:24.200 align:middle line:84% because it crosses the cdf of F once from below. 00:12:24.200 --> 00:12:26.570 align:middle line:84% If F looked like this, this thing 00:12:26.570 --> 00:12:30.570 align:middle line:90% would not be valid because-- 00:12:30.570 --> 00:12:33.910 align:middle line:90% oh, no, that would still be OK. 00:12:33.910 --> 00:12:36.580 align:middle line:84% The problem is going to come in if let's suppose 00:12:36.580 --> 00:12:45.390 align:middle line:84% that F were to go like this, then it never crosses. 00:12:45.390 --> 00:12:48.020 align:middle line:84% And so if I'm having to use an s upper bar that's 00:12:48.020 --> 00:12:50.840 align:middle line:84% bigger than the upper bound of the F distribution, 00:12:50.840 --> 00:12:53.090 align:middle line:84% then this thing isn't a mean-preserving spread 00:12:53.090 --> 00:12:55.040 align:middle line:90% of F anymore. 00:12:55.040 --> 00:13:00.230 align:middle line:84% And so what I'm going to want to do is just pick an s lower bar 00:13:00.230 --> 00:13:02.270 align:middle line:84% and pick the smallest s lower bar such 00:13:02.270 --> 00:13:07.820 align:middle line:84% that when I put the upper bound below the upper bound 1, 00:13:07.820 --> 00:13:09.670 align:middle line:90% I'm getting a mean of mu. 00:13:09.670 --> 00:13:13.440 align:middle line:90% 00:13:13.440 --> 00:13:15.430 align:middle line:84% So anyway, so that's the best we can 00:13:15.430 --> 00:13:18.700 align:middle line:84% do within that class of distributions. 00:13:18.700 --> 00:13:21.407 align:middle line:84% And then step 2, the harder step of the proof-- 00:13:21.407 --> 00:13:23.740 align:middle line:84% actually, I think I've done it pretty well on the slides 00:13:23.740 --> 00:13:24.860 align:middle line:84% if you want to look at the slides at home. 00:13:24.860 --> 00:13:26.470 align:middle line:84% But the basic idea in the slides is 00:13:26.470 --> 00:13:29.050 align:middle line:84% that no other distribution could possibly 00:13:29.050 --> 00:13:31.330 align:middle line:90% do better than this one. 00:13:31.330 --> 00:13:33.710 align:middle line:90% And how does that argument go? 00:13:33.710 --> 00:13:36.550 align:middle line:84% Well, what I'm going to do is I'm going to say, 00:13:36.550 --> 00:13:40.380 align:middle line:90% let G be any other valid choice. 00:13:40.380 --> 00:13:43.740 align:middle line:84% If you give me G is a valid choice, 00:13:43.740 --> 00:13:46.320 align:middle line:84% I'm going to pick a distribution in this class 00:13:46.320 --> 00:13:50.460 align:middle line:84% and show that it does at least as well as G and is valid. 00:13:50.460 --> 00:13:56.700 align:middle line:84% So you give me any other valid distribution, 00:13:56.700 --> 00:14:01.620 align:middle line:84% let pi be the profit that you earn when you use G, 00:14:01.620 --> 00:14:04.230 align:middle line:84% I'm going to consider the distribution G of pi 00:14:04.230 --> 00:14:07.920 align:middle line:84% s upper bar where s upper bar is chosen such 00:14:07.920 --> 00:14:10.415 align:middle line:84% that s upper bar gives this thing a mean of mu. 00:14:10.415 --> 00:14:14.310 align:middle line:90% 00:14:14.310 --> 00:14:17.760 align:middle line:84% This distribution, G of pi s upper bar, 00:14:17.760 --> 00:14:21.090 align:middle line:84% it also gives-- it gives the same profit as in G, 00:14:21.090 --> 00:14:24.120 align:middle line:84% because these distributions G of s lower bar s upper bar 00:14:24.120 --> 00:14:25.840 align:middle line:90% give profit s lower bar. 00:14:25.840 --> 00:14:28.800 align:middle line:84% So if I choose the profit I get from using this, 00:14:28.800 --> 00:14:32.110 align:middle line:84% this thing also gives the same profit pi. 00:14:32.110 --> 00:14:34.570 align:middle line:90% It gives the same profit pi. 00:14:34.570 --> 00:14:36.410 align:middle line:90% it's fully socially efficient. 00:14:36.410 --> 00:14:39.010 align:middle line:84% Everybody purchases in that distribution, 00:14:39.010 --> 00:14:42.100 align:middle line:84% so it has to give at least as much social surplus. 00:14:42.100 --> 00:14:45.950 align:middle line:84% Because it's maximizing social welfare, it has the same profit, 00:14:45.950 --> 00:14:48.530 align:middle line:84% consumer surplus has to be at least as high. 00:14:48.530 --> 00:14:53.390 align:middle line:84% So if I switch from G to G of pi s upper bar, 00:14:53.390 --> 00:14:56.960 align:middle line:84% then I've gotten at least as much consumer surplus 00:14:56.960 --> 00:14:59.580 align:middle line:90% to complete the proof. 00:14:59.580 --> 00:15:01.730 align:middle line:84% I just then need to show that G of pi 00:15:01.730 --> 00:15:05.780 align:middle line:84% s upper bar is a valid distribution to choose. 00:15:05.780 --> 00:15:07.670 align:middle line:84% And to do that, I just need to show 00:15:07.670 --> 00:15:15.300 align:middle line:84% that F is a mean-preserving spread of G of pi s. 00:15:15.300 --> 00:15:16.860 align:middle line:90% To do that-- 00:15:16.860 --> 00:15:25.770 align:middle line:84% I'm trying to show that F is a mean-preserving spread of G pi s 00:15:25.770 --> 00:15:26.690 align:middle line:90% upper bar. 00:15:26.690 --> 00:15:28.270 align:middle line:90% I do that in two steps. 00:15:28.270 --> 00:15:36.180 align:middle line:84% I show that-- we know that F is a mean-preserving spread of G 00:15:36.180 --> 00:15:37.750 align:middle line:90% because G was valid. 00:15:37.750 --> 00:15:40.620 align:middle line:84% And so I just need to show that any G is 00:15:40.620 --> 00:15:45.450 align:middle line:84% a mean-preserving spread of G of pi s upper bar. 00:15:45.450 --> 00:15:47.010 align:middle line:84% And the way I'm going to do that is 00:15:47.010 --> 00:15:51.230 align:middle line:90% show that any G looks like this. 00:15:51.230 --> 00:16:00.510 align:middle line:84% Any G that's valid has to start above G of pi s upper bar here-- 00:16:00.510 --> 00:16:07.800 align:middle line:84% has to start above it, and then it has to end below it. 00:16:07.800 --> 00:16:10.030 align:middle line:84% Anyway, I've got the argument here. 00:16:10.030 --> 00:16:12.660 align:middle line:84% But basically, the argument is it has to be above it here 00:16:12.660 --> 00:16:14.370 align:middle line:90% because this one is 0. 00:16:14.370 --> 00:16:17.640 align:middle line:84% In this interval, it has to be above it here 00:16:17.640 --> 00:16:21.390 align:middle line:84% because this one, the maximized profit, is at most pi. 00:16:21.390 --> 00:16:24.140 align:middle line:84% This one, the maximized profit is exactly equal to pi. 00:16:24.140 --> 00:16:26.260 align:middle line:84% So you have to sell to fewer consumers. 00:16:26.260 --> 00:16:28.080 align:middle line:84% So it's above it all the way to here, 00:16:28.080 --> 00:16:30.760 align:middle line:84% and then it has to be below it here because that one's 1. 00:16:30.760 --> 00:16:32.320 align:middle line:90% That's the basic argument. 00:16:32.320 --> 00:16:38.080 align:middle line:84% So it's a-- what are my takeaways-- 00:16:38.080 --> 00:16:41.250 align:middle line:90% let me get my takeaways first. 00:16:41.250 --> 00:16:44.410 align:middle line:84% Takeaways, profit-maximizing structure is simple. 00:16:44.410 --> 00:16:47.557 align:middle line:84% You just tell people that their value is at least c. 00:16:47.557 --> 00:16:49.390 align:middle line:84% A good way to think about information design 00:16:49.390 --> 00:16:51.250 align:middle line:84% is think about choosing the distribution of the consumer's 00:16:51.250 --> 00:16:52.210 align:middle line:90% posterior. 00:16:52.210 --> 00:16:56.480 align:middle line:90% And if you want to do that-- 00:16:56.480 --> 00:16:59.098 align:middle line:90% if you want to minimize-- 00:16:59.098 --> 00:17:01.640 align:middle line:84% if you want to maximize consumer surplus, what you want to do 00:17:01.640 --> 00:17:03.350 align:middle line:84% is get everyone to buy, and get everyone 00:17:03.350 --> 00:17:05.609 align:middle line:84% to buy at the lowest price possible. 00:17:05.609 --> 00:17:08.930 align:middle line:84% The way you get people to buy at the lowest price possible 00:17:08.930 --> 00:17:13.099 align:middle line:84% is to use a distribution like this where there's 00:17:13.099 --> 00:17:15.089 align:middle line:90% a lot of low-value consumers. 00:17:15.089 --> 00:17:18.890 align:middle line:84% And if you raise your price at all, you lose a lot of demand. 00:17:18.890 --> 00:17:22.685 align:middle line:84% And demand is very steep here, so that gets the monopolist 00:17:22.685 --> 00:17:24.560 align:middle line:84% to be indifferent between charging this price 00:17:24.560 --> 00:17:26.371 align:middle line:84% and charging those higher prices. 00:17:26.371 --> 00:17:27.829 align:middle line:84% And so that's the kind of thing you 00:17:27.829 --> 00:17:29.840 align:middle line:84% want to do with the information structure is 00:17:29.840 --> 00:17:32.230 align:middle line:84% create a lot of low-value consumers 00:17:32.230 --> 00:17:36.080 align:middle line:84% and create a distribution where every time the monopolist 00:17:36.080 --> 00:17:37.940 align:middle line:84% raises their price a little bit, they 00:17:37.940 --> 00:17:39.800 align:middle line:84% would be losing lots and lots of consumers. 00:17:39.800 --> 00:17:41.450 align:middle line:84% And distributions like this are the way 00:17:41.450 --> 00:17:43.480 align:middle line:90% you maximize consumer surplus. 00:17:43.480 --> 00:17:47.072 align:middle line:90% 00:17:47.072 --> 00:17:51.610 align:middle line:84% Also, Paolo had actually asked after the last class, What's 00:17:51.610 --> 00:17:54.650 align:middle line:84% the full set of things we can do, and how do I show that? 00:17:54.650 --> 00:17:58.218 align:middle line:84% I added a slide that I won't talk about. 00:17:58.218 --> 00:18:00.760 align:middle line:84% But Roesler and Szentes just have this fairly simple argument 00:18:00.760 --> 00:18:05.110 align:middle line:84% that shows that this is the profit-maximizing point. 00:18:05.110 --> 00:18:08.090 align:middle line:84% This is the consumer surplus maximizing point. 00:18:08.090 --> 00:18:10.850 align:middle line:84% Everything in this triangle is possible. 00:18:10.850 --> 00:18:12.970 align:middle line:84% So if you want to achieve less efficiency, 00:18:12.970 --> 00:18:14.665 align:middle line:90% you can achieve less efficiency. 00:18:14.665 --> 00:18:16.540 align:middle line:84% But basically, what you can do is give them-- 00:18:16.540 --> 00:18:19.010 align:middle line:84% you have to give a monopolist at least this much profit. 00:18:19.010 --> 00:18:20.170 align:middle line:84% If you were trying to hurt the monopolist, 00:18:20.170 --> 00:18:21.310 align:middle line:90% you can't do any better. 00:18:21.310 --> 00:18:23.118 align:middle line:84% And consumer surplus-- anyway, it's 00:18:23.118 --> 00:18:24.660 align:middle line:84% going to have to be in that triangle. 00:18:24.660 --> 00:18:27.580 align:middle line:90% 00:18:27.580 --> 00:18:34.290 align:middle line:84% I really think it's probably mostly this diagonal 00:18:34.290 --> 00:18:36.660 align:middle line:84% that we're typically interested in, how do we maximize 00:18:36.660 --> 00:18:38.680 align:middle line:90% consumer surplus versus profit. 00:18:38.680 --> 00:18:41.310 align:middle line:84% And the way we do this is we choose these distributions 00:18:41.310 --> 00:18:43.230 align:middle line:90% G of s lower bar, s upper bar. 00:18:43.230 --> 00:18:46.680 align:middle line:90% And we just vary s lower bar. 00:18:46.680 --> 00:18:50.130 align:middle line:84% And the higher we make s lower bar, 00:18:50.130 --> 00:18:52.260 align:middle line:84% the higher is the monopolist's profit, 00:18:52.260 --> 00:18:54.240 align:middle line:90% the lower is consumer surplus. 00:18:54.240 --> 00:18:56.550 align:middle line:84% In some sense, this is a special case 00:18:56.550 --> 00:18:59.250 align:middle line:84% of this argument where you just set s 00:18:59.250 --> 00:19:01.980 align:middle line:84% lower bar equals expectation of v, 00:19:01.980 --> 00:19:04.320 align:middle line:90% given that v is greater than c. 00:19:04.320 --> 00:19:06.430 align:middle line:90% That's the monopoly solution. 00:19:06.430 --> 00:19:08.170 align:middle line:84% That's the monopoly pricing solution. 00:19:08.170 --> 00:19:15.060 align:middle line:84% And so as we raise s upper bar to that point, 00:19:15.060 --> 00:19:16.990 align:middle line:90% that's what we're going to do. 00:19:16.990 --> 00:19:20.960 align:middle line:90% 00:19:20.960 --> 00:19:22.216 align:middle line:90% Any questions? 00:19:22.216 --> 00:19:26.750 align:middle line:90% 00:19:26.750 --> 00:19:29.030 align:middle line:90% Let me say, I feel this is-- 00:19:29.030 --> 00:19:31.990 align:middle line:84% just in terms of thinking about topics, I don't really know. 00:19:31.990 --> 00:19:33.140 align:middle line:90% I looked. 00:19:33.140 --> 00:19:35.540 align:middle line:84% I couldn't find any good empirical papers 00:19:35.540 --> 00:19:36.960 align:middle line:90% on monopoly information design. 00:19:36.960 --> 00:19:39.293 align:middle line:84% It seems like that's something that I think there should 00:19:39.293 --> 00:19:40.740 align:middle line:90% be some empirical work on. 00:19:40.740 --> 00:19:44.810 align:middle line:84% And also, this question of translating this from theory 00:19:44.810 --> 00:19:48.950 align:middle line:84% to IO is this "I can choose any joint distribution over v and s 00:19:48.950 --> 00:19:52.920 align:middle line:84% that's possible" is a heroic assumption. 00:19:52.920 --> 00:19:54.590 align:middle line:90% And I guess the question-- 00:19:54.590 --> 00:19:56.600 align:middle line:84% intriguing empirical question is, 00:19:56.600 --> 00:19:59.210 align:middle line:84% how can one do that in practice, and how close can one 00:19:59.210 --> 00:20:04.010 align:middle line:84% get to designing expectation of the distributions 00:20:04.010 --> 00:20:07.190 align:middle line:84% over expectation of v, given the signals, 00:20:07.190 --> 00:20:10.520 align:middle line:84% with a finite set of signals or signals about product 00:20:10.520 --> 00:20:13.490 align:middle line:84% characteristics or a more limited set of ways 00:20:13.490 --> 00:20:15.680 align:middle line:84% I can give signals than just this telling people 00:20:15.680 --> 00:20:17.340 align:middle line:90% their posteriors immediately? 00:20:17.340 --> 00:20:20.290 align:middle line:90% 00:20:20.290 --> 00:20:22.360 align:middle line:84% Anyway, so rest of today's class, 00:20:22.360 --> 00:20:24.580 align:middle line:84% I'm going to start doing empirical papers 00:20:24.580 --> 00:20:27.370 align:middle line:90% for the first time. 00:20:27.370 --> 00:20:31.010 align:middle line:84% And I'm going to mostly talk about two papers. 00:20:31.010 --> 00:20:33.410 align:middle line:84% What I thought I would do first, though, 00:20:33.410 --> 00:20:37.370 align:middle line:84% is go back and give you the 8-minute history of empirical 00:20:37.370 --> 00:20:38.540 align:middle line:90% IO. 00:20:38.540 --> 00:20:41.540 align:middle line:84% I mean, I think the background is useful to understand. 00:20:41.540 --> 00:20:43.250 align:middle line:84% As you start reading IO papers, there 00:20:43.250 --> 00:20:45.625 align:middle line:84% are a lot of IO papers written about ketchup or cardboard 00:20:45.625 --> 00:20:48.350 align:middle line:84% boxes, and people often ask me why on earth 00:20:48.350 --> 00:20:49.770 align:middle line:90% do IO economists do that. 00:20:49.770 --> 00:20:53.360 align:middle line:84% And so I think giving some perspective on what IO people do 00:20:53.360 --> 00:20:57.440 align:middle line:90% and why could be useful. 00:20:57.440 --> 00:20:59.510 align:middle line:84% Let me start my brief empirical history. 00:20:59.510 --> 00:21:02.510 align:middle line:90% So early empirical IO-- 00:21:02.510 --> 00:21:07.670 align:middle line:84% so before the 1960s, IO economists wrote books. 00:21:07.670 --> 00:21:09.330 align:middle line:90% And they wrote big, thick books. 00:21:09.330 --> 00:21:10.940 align:middle line:84% They wrote books-- an IO economist was 00:21:10.940 --> 00:21:15.380 align:middle line:84% an expert in some industry like the airline industry or the car 00:21:15.380 --> 00:21:19.080 align:middle line:84% industry or the coal industry or the steel industry or whatever. 00:21:19.080 --> 00:21:21.630 align:middle line:84% But IO economists were experts in industries. 00:21:21.630 --> 00:21:26.060 align:middle line:84% They wrote these big, thick books discussing lots 00:21:26.060 --> 00:21:28.690 align:middle line:90% of stuff about the industry. 00:21:28.690 --> 00:21:31.190 align:middle line:84% We actually used to have an assignment in this course, which 00:21:31.190 --> 00:21:33.950 align:middle line:84% was try to get a Harvard library card, 00:21:33.950 --> 00:21:36.290 align:middle line:84% go into the stacks of Widener, pick out 00:21:36.290 --> 00:21:39.807 align:middle line:84% one of these many books, read it, and write something 00:21:39.807 --> 00:21:41.390 align:middle line:84% about it, because it was often they're 00:21:41.390 --> 00:21:43.820 align:middle line:84% full of facts and full of interesting things going on 00:21:43.820 --> 00:21:44.840 align:middle line:90% in the industry. 00:21:44.840 --> 00:21:47.960 align:middle line:84% And they often have lots of puzzles about what's going on 00:21:47.960 --> 00:21:50.810 align:middle line:84% and why is this happening when, with modern theory, you're like, 00:21:50.810 --> 00:21:52.580 align:middle line:84% OK, I could tell you why that's happening, 00:21:52.580 --> 00:21:56.640 align:middle line:84% or this is a good example of search affecting pricing 00:21:56.640 --> 00:22:02.100 align:middle line:84% or whatever, some kind of incomplete information story. 00:22:02.100 --> 00:22:04.200 align:middle line:84% It was a very interesting field, in the sense 00:22:04.200 --> 00:22:05.790 align:middle line:84% that you would read books about these industries, 00:22:05.790 --> 00:22:07.990 align:middle line:84% and you would learn a tremendous amount about them. 00:22:07.990 --> 00:22:10.390 align:middle line:84% They were typically very light on theory. 00:22:10.390 --> 00:22:12.538 align:middle line:84% It was really a lot of descriptive work about how 00:22:12.538 --> 00:22:14.580 align:middle line:84% does this industry work, what are the firms, what 00:22:14.580 --> 00:22:17.062 align:middle line:84% are the prices like, what are the markups, what 00:22:17.062 --> 00:22:18.770 align:middle line:84% are all the practices that you see there. 00:22:18.770 --> 00:22:22.590 align:middle line:90% 00:22:22.590 --> 00:22:29.880 align:middle line:84% So in the '60s and '70s, there was this increasing use 00:22:29.880 --> 00:22:31.890 align:middle line:90% of statistics in economics. 00:22:31.890 --> 00:22:36.810 align:middle line:84% And instead of just writing discussions of industries, 00:22:36.810 --> 00:22:39.690 align:middle line:84% people wanted to have regressions in their papers. 00:22:39.690 --> 00:22:42.120 align:middle line:84% And so regression started to come into the profession 00:22:42.120 --> 00:22:43.530 align:middle line:90% in the 1950s. 00:22:43.530 --> 00:22:47.068 align:middle line:84% I have this one anecdote here that my wife got from Zvi 00:22:47.068 --> 00:22:49.110 align:middle line:84% Griliches, who was a famous economist at Harvard, 00:22:49.110 --> 00:22:52.620 align:middle line:84% that his summer job at some point in the 1950s was to run 00:22:52.620 --> 00:22:53.230 align:middle line:90% a regression. 00:22:53.230 --> 00:22:55.230 align:middle line:84% And you might ask, well, OK, was your summer job 00:22:55.230 --> 00:22:55.980 align:middle line:90% to run regression? 00:22:55.980 --> 00:22:58.050 align:middle line:84% It's like, no, he was given the data set, 00:22:58.050 --> 00:23:02.770 align:middle line:84% and his job was to compute x prime x and then invert it. 00:23:02.770 --> 00:23:06.420 align:middle line:84% And it was, like, a 7-variable regression or something like 00:23:06.420 --> 00:23:06.920 align:middle line:90% that. 00:23:06.920 --> 00:23:09.550 align:middle line:84% So his summer job was with a calculator, like, here's 00:23:09.550 --> 00:23:11.200 align:middle line:90% a 7 by 7 matrix. 00:23:11.200 --> 00:23:15.220 align:middle line:84% I want you to invert x prime x so that we can then report 00:23:15.220 --> 00:23:17.860 align:middle line:90% the results of that regression. 00:23:17.860 --> 00:23:22.480 align:middle line:84% But anyway, obviously, this got easier in the '50s and '60s when 00:23:22.480 --> 00:23:25.870 align:middle line:90% computers got invented. 00:23:25.870 --> 00:23:29.770 align:middle line:84% And Joe Bain, who was actually a first one 00:23:29.770 --> 00:23:33.190 align:middle line:84% to do regressions in a within-industry study, 00:23:33.190 --> 00:23:36.040 align:middle line:84% had a view that IO shouldn't be developing 00:23:36.040 --> 00:23:38.750 align:middle line:84% these facts about the coal industry or the steel industry. 00:23:38.750 --> 00:23:41.720 align:middle line:84% We want broad truths that apply to all industries. 00:23:41.720 --> 00:23:43.000 align:middle line:90% And so it became popular. 00:23:43.000 --> 00:23:46.130 align:middle line:84% And IO after that was cross-section regressions. 00:23:46.130 --> 00:23:48.460 align:middle line:84% And what happened in these cross-section regressions 00:23:48.460 --> 00:23:50.590 align:middle line:84% was people typically have a dependent variable 00:23:50.590 --> 00:23:54.610 align:middle line:90% like markups in industry i. 00:23:54.610 --> 00:23:57.460 align:middle line:84% And they would regress markup on industry i. 00:23:57.460 --> 00:24:02.230 align:middle line:84% So it would be like markup in industry i 00:24:02.230 --> 00:24:08.950 align:middle line:84% is alpha 0 plus alpha 1 times the four-firm concentration 00:24:08.950 --> 00:24:22.490 align:middle line:84% ratio plus alpha 2 times the advertising to sales ratio, 00:24:22.490 --> 00:24:24.890 align:middle line:84% and there'd be a bunch of other variables. 00:24:24.890 --> 00:24:26.900 align:middle line:90% But you would just do-- 00:24:26.900 --> 00:24:30.140 align:middle line:84% I'm running a regression like this with six or seven or eight 00:24:30.140 --> 00:24:32.810 align:middle line:84% variables, and I'm going to try to talk about what is it that 00:24:32.810 --> 00:24:35.660 align:middle line:90% makes industries competitive. 00:24:35.660 --> 00:24:38.935 align:middle line:84% And they were looking for these broader truths that 00:24:38.935 --> 00:24:41.060 align:middle line:84% would give us-- instead of having to study industry 00:24:41.060 --> 00:24:46.130 align:middle line:84% by industry, what determines profit levels in general? 00:24:46.130 --> 00:24:49.010 align:middle line:84% Anyway, so regressions here, there's 00:24:49.010 --> 00:24:51.950 align:middle line:84% a data set the Census of Manufacturers-- 00:24:51.950 --> 00:24:54.745 align:middle line:84% Census of Manufacturers, it's a census of manufacturing firms 00:24:54.745 --> 00:24:55.620 align:middle line:90% in the United States. 00:24:55.620 --> 00:24:58.895 align:middle line:84% So it has data on every individual firm in the United 00:24:58.895 --> 00:25:00.770 align:middle line:84% States, and they send them a big, long survey 00:25:00.770 --> 00:25:03.050 align:middle line:90% with hundreds of questions. 00:25:03.050 --> 00:25:04.910 align:middle line:84% If you ever want to write papers using 00:25:04.910 --> 00:25:06.410 align:middle line:84% the Census of Manufacturers, you can 00:25:06.410 --> 00:25:08.840 align:middle line:84% access the Census of Manufacturers individual data 00:25:08.840 --> 00:25:13.100 align:middle line:84% from the NBER, from a small, windowless, unpleasant room 00:25:13.100 --> 00:25:13.920 align:middle line:90% at the NBER. 00:25:13.920 --> 00:25:16.370 align:middle line:84% But there is a windowless room where 00:25:16.370 --> 00:25:19.010 align:middle line:84% you can access the firm-level data on every manufacturing 00:25:19.010 --> 00:25:20.870 align:middle line:90% firm in the US. 00:25:20.870 --> 00:25:24.590 align:middle line:84% People in that era didn't really access the census microdata. 00:25:24.590 --> 00:25:27.590 align:middle line:84% They just got the published census data of means of these 00:25:27.590 --> 00:25:31.250 align:middle line:84% variables for each of the 153-digit manufacturing 00:25:31.250 --> 00:25:34.513 align:middle line:84% industries or 454-digit manufacturing industries. 00:25:34.513 --> 00:25:36.180 align:middle line:84% They'd run these regressions, and they'd 00:25:36.180 --> 00:25:38.610 align:middle line:90% talk about what determined-- 00:25:38.610 --> 00:25:41.370 align:middle line:84% how many firms did you need for a markup to be competitive, 00:25:41.370 --> 00:25:43.740 align:middle line:84% or how did competitiveness vary with the number of firms 00:25:43.740 --> 00:25:44.407 align:middle line:90% in the industry? 00:25:44.407 --> 00:25:49.400 align:middle line:90% 00:25:49.400 --> 00:25:53.090 align:middle line:84% This was popular in the '60s and '70s, and then it started 00:25:53.090 --> 00:25:54.080 align:middle line:90% to become-- 00:25:54.080 --> 00:25:57.110 align:middle line:84% a number of critiques of this approach 00:25:57.110 --> 00:25:59.510 align:middle line:90% became very well accepted. 00:25:59.510 --> 00:26:03.920 align:middle line:84% One critique was that just all these studies 00:26:03.920 --> 00:26:07.203 align:middle line:84% are leaving out very important features of the industry. 00:26:07.203 --> 00:26:09.620 align:middle line:84% So for instance, if I'm going to study the steel industry, 00:26:09.620 --> 00:26:12.620 align:middle line:84% I can't talk about prices in the steel industry 00:26:12.620 --> 00:26:15.980 align:middle line:84% or profits in the steel industry without having price of iron ore 00:26:15.980 --> 00:26:17.280 align:middle line:90% on the right side. 00:26:17.280 --> 00:26:19.280 align:middle line:84% I can't think about prices in the steel industry 00:26:19.280 --> 00:26:21.420 align:middle line:84% without price of coke and coal on the right side. 00:26:21.420 --> 00:26:22.837 align:middle line:84% I can't think about steel industry 00:26:22.837 --> 00:26:25.850 align:middle line:84% without strength of union labor in the Midwestern United 00:26:25.850 --> 00:26:27.210 align:middle line:90% States on the right side. 00:26:27.210 --> 00:26:29.258 align:middle line:84% So for every industry I had, I would 00:26:29.258 --> 00:26:30.800 align:middle line:84% think of three or four or five things 00:26:30.800 --> 00:26:32.342 align:middle line:84% that just had to be on the right side 00:26:32.342 --> 00:26:33.740 align:middle line:90% to talk about that industry. 00:26:33.740 --> 00:26:35.748 align:middle line:84% But if you've got 150 industries, 00:26:35.748 --> 00:26:37.790 align:middle line:84% and each industry brings in three or four or five 00:26:37.790 --> 00:26:39.792 align:middle line:84% unique things to put on the right side, 00:26:39.792 --> 00:26:41.750 align:middle line:84% and then you put them all in on the right side, 00:26:41.750 --> 00:26:43.850 align:middle line:84% you end up with a regression with 750 things 00:26:43.850 --> 00:26:46.660 align:middle line:84% on the right side and 150 observations. 00:26:46.660 --> 00:26:49.722 align:middle line:84% And obviously, now machine learning, 00:26:49.722 --> 00:26:51.430 align:middle line:84% people will tell you that's totally fine. 00:26:51.430 --> 00:26:55.050 align:middle line:84% 750 explanatory variables, 150 observations, it all works fine. 00:26:55.050 --> 00:26:57.850 align:middle line:84% But at least in the 1980s, that did not work. 00:26:57.850 --> 00:27:00.990 align:middle line:84% You couldn't run regressions like that. 00:27:00.990 --> 00:27:03.510 align:middle line:84% The other problem that came to be accepted 00:27:03.510 --> 00:27:09.480 align:middle line:84% is that the census data, like when you're studying markups, 00:27:09.480 --> 00:27:11.760 align:middle line:84% markup in industry i was really price 00:27:11.760 --> 00:27:18.150 align:middle line:84% in industry i minus the cost of production in industry i. 00:27:18.150 --> 00:27:22.830 align:middle line:84% Or maybe it was price minus cost over price. 00:27:22.830 --> 00:27:26.370 align:middle line:84% But you would need to use this data on what's 00:27:26.370 --> 00:27:28.200 align:middle line:90% the cost of production. 00:27:28.200 --> 00:27:30.360 align:middle line:84% And the cost of production that you 00:27:30.360 --> 00:27:34.980 align:middle line:84% get when managers put in for the cost what their accountant told 00:27:34.980 --> 00:27:37.020 align:middle line:84% them the cost is, is very different from what 00:27:37.020 --> 00:27:40.320 align:middle line:84% you would get if an economist asked what the cost is. 00:27:40.320 --> 00:27:43.980 align:middle line:84% For an economist, you really want the marginal cost, not 00:27:43.980 --> 00:27:45.540 align:middle line:90% the average cost. 00:27:45.540 --> 00:27:47.520 align:middle line:84% But then even though when you think about-- 00:27:47.520 --> 00:27:50.910 align:middle line:84% even getting the marginal versus average, accounting costs 00:27:50.910 --> 00:27:53.190 align:middle line:84% of all these conventions in of how 00:27:53.190 --> 00:27:54.930 align:middle line:84% do you treat the cost of capital, 00:27:54.930 --> 00:27:56.790 align:middle line:84% and do you amortize it over time, 00:27:56.790 --> 00:27:59.640 align:middle line:84% and you put in a depreciation adjustment. 00:27:59.640 --> 00:28:01.890 align:middle line:84% And what do you do when the firm spends a lot of money 00:28:01.890 --> 00:28:02.860 align:middle line:90% on advertising? 00:28:02.860 --> 00:28:04.500 align:middle line:84% Is that a one-period cost, or is that 00:28:04.500 --> 00:28:07.890 align:middle line:84% a flow cost that depreciates over the number of times 00:28:07.890 --> 00:28:10.060 align:middle line:90% that the goodwill stock decays? 00:28:10.060 --> 00:28:12.390 align:middle line:90% And so the costs-- 00:28:12.390 --> 00:28:14.045 align:middle line:84% a lot of the elements of the cost, 00:28:14.045 --> 00:28:16.170 align:middle line:84% how do you put the firm overhead and all the people 00:28:16.170 --> 00:28:19.080 align:middle line:84% who work in the management office or whatever? 00:28:19.080 --> 00:28:21.600 align:middle line:84% The costs have a lot of assumptions 00:28:21.600 --> 00:28:23.520 align:middle line:90% that go into defining them. 00:28:23.520 --> 00:28:27.510 align:middle line:84% And so the difference between the accounting cost 00:28:27.510 --> 00:28:30.210 align:middle line:90% and the true cost is large. 00:28:30.210 --> 00:28:33.570 align:middle line:84% And it's probably systematically related to things like 00:28:33.570 --> 00:28:36.030 align:middle line:84% whether the firm is doing a lot of advertising or not, 00:28:36.030 --> 00:28:39.060 align:middle line:84% whether it's doing a lot of R&D, how old 00:28:39.060 --> 00:28:41.220 align:middle line:84% the product is, whether it's produced 00:28:41.220 --> 00:28:42.820 align:middle line:84% in big factories or little factories. 00:28:42.820 --> 00:28:45.180 align:middle line:84% So all the stuff that you're putting on the right side 00:28:45.180 --> 00:28:48.040 align:middle line:84% is correlated with the measurement error that you want. 00:28:48.040 --> 00:28:50.340 align:middle line:84% And so then are we really just finding correlations 00:28:50.340 --> 00:28:52.590 align:middle line:84% with measurement error, or are we finding correlations 00:28:52.590 --> 00:28:55.380 align:middle line:90% with actual firm conduct? 00:28:55.380 --> 00:28:59.670 align:middle line:90% And so that was a big issue. 00:28:59.670 --> 00:29:01.840 align:middle line:90% Endogeneity is also a big issue. 00:29:01.840 --> 00:29:05.230 align:middle line:84% You're putting the number of firms on the right side. 00:29:05.230 --> 00:29:07.690 align:middle line:84% Something you get taught over and over again in classes 00:29:07.690 --> 00:29:10.570 align:middle line:84% like this is, well, if you have a business that's 00:29:10.570 --> 00:29:13.000 align:middle line:84% highly profitable, then more firms are going to enter, 00:29:13.000 --> 00:29:15.010 align:middle line:84% and that's going to change the profits. 00:29:15.010 --> 00:29:18.415 align:middle line:84% So this four-firm concentration ratio 00:29:18.415 --> 00:29:20.290 align:middle line:84% is also going to be endogenous, or the number 00:29:20.290 --> 00:29:22.810 align:middle line:84% of firms on the right side is going to be endogenous. 00:29:22.810 --> 00:29:24.970 align:middle line:84% There really wasn't much in the way of IV 00:29:24.970 --> 00:29:28.900 align:middle line:84% estimation in the structure conduct performance literature. 00:29:28.900 --> 00:29:34.540 align:middle line:84% And then there was also interest in IO theory. 00:29:34.540 --> 00:29:39.710 align:middle line:84% This led to what was called the NEIO revolution. 00:29:39.710 --> 00:29:42.800 align:middle line:84% So NEIO stands for New Empirical Industrial Organization. 00:29:42.800 --> 00:29:44.950 align:middle line:84% I guess lots of fields run into this problem. 00:29:44.950 --> 00:29:48.077 align:middle line:84% You have the sort of new something, 00:29:48.077 --> 00:29:49.910 align:middle line:84% and then there's something comes after that. 00:29:49.910 --> 00:29:51.110 align:middle line:84% Like you have the new trade theory, 00:29:51.110 --> 00:29:52.943 align:middle line:84% and then you have to have the new, new trade 00:29:52.943 --> 00:29:54.260 align:middle line:90% theory or whatever, because-- 00:29:54.260 --> 00:29:57.920 align:middle line:84% anyway, so in the 19 early '80s, people started to call this 00:29:57.920 --> 00:30:00.000 align:middle line:84% the new empirical industrial organization. 00:30:00.000 --> 00:30:01.040 align:middle line:90% It's no longer very new. 00:30:01.040 --> 00:30:04.550 align:middle line:90% But what happened in NEIO-- 00:30:04.550 --> 00:30:07.400 align:middle line:84% so one is, it went back to the '60s. 00:30:07.400 --> 00:30:10.760 align:middle line:84% It went back to people writing papers about single industries. 00:30:10.760 --> 00:30:13.460 align:middle line:84% Everyone thinks that there are a lot of specific things 00:30:13.460 --> 00:30:14.940 align:middle line:84% about every industry that matters, 00:30:14.940 --> 00:30:17.000 align:middle line:84% so we need to study it one industry at a time 00:30:17.000 --> 00:30:18.740 align:middle line:84% and control for all those factors that 00:30:18.740 --> 00:30:21.620 align:middle line:90% should be on the right side. 00:30:21.620 --> 00:30:25.080 align:middle line:84% There was an emphasis on using what was well measured. 00:30:25.080 --> 00:30:26.570 align:middle line:90% So it's sort of funny. 00:30:26.570 --> 00:30:29.477 align:middle line:84% In macro, they think cost is well measured, 00:30:29.477 --> 00:30:31.310 align:middle line:84% and price is hard to measure, so they always 00:30:31.310 --> 00:30:32.660 align:middle line:90% want to use cost data. 00:30:32.660 --> 00:30:35.480 align:middle line:84% In IO, people think that price is well measured, 00:30:35.480 --> 00:30:38.900 align:middle line:84% cost is hard to measure, so we always use price data. 00:30:38.900 --> 00:30:43.100 align:middle line:84% So there's this emphasis on using quantities, prices, 00:30:43.100 --> 00:30:45.920 align:middle line:84% number of firms, market shares, those things 00:30:45.920 --> 00:30:49.200 align:middle line:84% as the explanatory variables-- as the variables you use 00:30:49.200 --> 00:30:52.950 align:middle line:84% rather than accounting variables. 00:30:52.950 --> 00:30:55.410 align:middle line:84% Great deal of attention paid now to identification 00:30:55.410 --> 00:30:56.760 align:middle line:90% and endogeneity. 00:30:56.760 --> 00:31:01.570 align:middle line:84% And theory also came to take on a big role in empirical IO. 00:31:01.570 --> 00:31:04.300 align:middle line:84% And it's often that the-- two different reasons. 00:31:04.300 --> 00:31:08.820 align:middle line:84% One is that the empirical work is often designed to improve 00:31:08.820 --> 00:31:11.190 align:middle line:84% our understanding of some theoretical model, 00:31:11.190 --> 00:31:13.720 align:middle line:84% which were being recently developed in the '80s. 00:31:13.720 --> 00:31:16.568 align:middle line:84% And then the second is because even if you're not 00:31:16.568 --> 00:31:18.360 align:middle line:84% trying to comment on the theory, the theory 00:31:18.360 --> 00:31:21.722 align:middle line:84% is going to give you thoughts about what 00:31:21.722 --> 00:31:23.430 align:middle line:84% the functional form should be or what you 00:31:23.430 --> 00:31:25.050 align:middle line:90% should be regressing on what. 00:31:25.050 --> 00:31:27.180 align:middle line:84% And the papers are often designed 00:31:27.180 --> 00:31:29.218 align:middle line:84% to test some particular-- or enlightened 00:31:29.218 --> 00:31:31.260 align:middle line:84% by some particular theory when they're setting up 00:31:31.260 --> 00:31:32.630 align:middle line:90% what analyses they're doing. 00:31:32.630 --> 00:31:36.060 align:middle line:90% 00:31:36.060 --> 00:31:37.920 align:middle line:84% More recently-- so at the beginning 00:31:37.920 --> 00:31:40.890 align:middle line:84% of the NEIO revolution, people would sometimes 00:31:40.890 --> 00:31:42.133 align:middle line:90% describe work as structural. 00:31:42.133 --> 00:31:43.800 align:middle line:84% And they would use the word "structural" 00:31:43.800 --> 00:31:47.160 align:middle line:84% in a way that's different from what structural means today. 00:31:47.160 --> 00:31:50.910 align:middle line:84% So in the '80s, structural really meant instead 00:31:50.910 --> 00:31:54.540 align:middle line:84% of estimating equilibrium objects like understanding price 00:31:54.540 --> 00:31:57.600 align:middle line:84% as a function of things, you're understanding supply 00:31:57.600 --> 00:32:00.090 align:middle line:84% as a function of things and demand as a function of things. 00:32:00.090 --> 00:32:01.950 align:middle line:84% And the underlying equations that 00:32:01.950 --> 00:32:03.840 align:middle line:84% determine the equilibrium instead 00:32:03.840 --> 00:32:06.030 align:middle line:90% of analyzing the equilibrium. 00:32:06.030 --> 00:32:08.370 align:middle line:84% I think, over time, structural work 00:32:08.370 --> 00:32:12.090 align:middle line:84% has increasingly come to be used to describe papers that 00:32:12.090 --> 00:32:15.390 align:middle line:84% assume that some theoretical model is correct 00:32:15.390 --> 00:32:18.300 align:middle line:84% and then focus on estimating the parameters 00:32:18.300 --> 00:32:20.080 align:middle line:90% of that theoretical model. 00:32:20.080 --> 00:32:22.680 align:middle line:84% And the word "structural" is also increasingly used to talk 00:32:22.680 --> 00:32:26.110 align:middle line:84% about work that goes all the way to estimating the primitives. 00:32:26.110 --> 00:32:28.740 align:middle line:84% In some ways, every industry, the outcomes 00:32:28.740 --> 00:32:33.720 align:middle line:84% depend on the production functions of the firms, 00:32:33.720 --> 00:32:36.360 align:middle line:84% the utilities of the consumers, and then something 00:32:36.360 --> 00:32:39.312 align:middle line:84% about the strategies that are available to the firms. 00:32:39.312 --> 00:32:40.770 align:middle line:84% And so structural work increasingly 00:32:40.770 --> 00:32:43.920 align:middle line:84% is trying to go back to those primitives. 00:32:43.920 --> 00:32:45.955 align:middle line:84% Instead of just estimating supply and demand, 00:32:45.955 --> 00:32:47.580 align:middle line:84% you're trying to estimate distributions 00:32:47.580 --> 00:32:50.250 align:middle line:90% of consumer utility functions. 00:32:50.250 --> 00:32:54.710 align:middle line:84% Something that's very popular in IO papers is that you-- 00:32:54.710 --> 00:32:57.420 align:middle line:84% a reason for getting all the way to the primitives 00:32:57.420 --> 00:33:00.510 align:middle line:84% is then you have these counterfactual simulations where 00:33:00.510 --> 00:33:03.790 align:middle line:84% you say, I observe this industry in this state. 00:33:03.790 --> 00:33:06.190 align:middle line:84% What would happen if there was entry in that industry? 00:33:06.190 --> 00:33:08.530 align:middle line:84% What would happen if we broke up these firms? 00:33:08.530 --> 00:33:11.760 align:middle line:84% What would happen if we put in a price cap regulation? 00:33:11.760 --> 00:33:13.710 align:middle line:84% Once you know these primitives, you 00:33:13.710 --> 00:33:16.560 align:middle line:84% can then talk about here are all these alternate policies we 00:33:16.560 --> 00:33:19.840 align:middle line:84% might implement, and then what would that do. 00:33:19.840 --> 00:33:21.480 align:middle line:84% And in some sense, this is what's 00:33:21.480 --> 00:33:25.110 align:middle line:84% really become very popular in many fields like education 00:33:25.110 --> 00:33:28.340 align:middle line:84% and development, where you estimate a fundamental model, 00:33:28.340 --> 00:33:30.930 align:middle line:84% and then you talk about I did a trial of this one, 00:33:30.930 --> 00:33:33.130 align:middle line:84% I did a randomized trial of this policy. 00:33:33.130 --> 00:33:34.740 align:middle line:84% I can also do an empirical estimate 00:33:34.740 --> 00:33:37.530 align:middle line:84% of what other policies would have done because I've 00:33:37.530 --> 00:33:40.656 align:middle line:90% uncovered the primitives. 00:33:40.656 --> 00:33:43.170 align:middle line:84% And I would say what's also happening at the same time-- 00:33:43.170 --> 00:33:46.110 align:middle line:84% while there's a lot of good structural work, over time, 00:33:46.110 --> 00:33:49.260 align:middle line:84% structural-empirical IO has become less and less connected 00:33:49.260 --> 00:33:50.020 align:middle line:90% to theory. 00:33:50.020 --> 00:33:52.440 align:middle line:84% And in some sense, structural-empirical IO models 00:33:52.440 --> 00:33:55.080 align:middle line:84% these days tend to use structures 00:33:55.080 --> 00:33:57.090 align:middle line:84% that previous structural-empirical papers have 00:33:57.090 --> 00:34:00.570 align:middle line:84% used rather than the structures of the theoretical models 00:34:00.570 --> 00:34:03.310 align:middle line:84% that theorists would use to describe the same situation. 00:34:03.310 --> 00:34:06.780 align:middle line:84% And part of that is because repeated games really 00:34:06.780 --> 00:34:08.813 align:middle line:84% don't work very well for empirical work. 00:34:08.813 --> 00:34:10.980 align:middle line:84% Incomplete information doesn't really work very well 00:34:10.980 --> 00:34:11.860 align:middle line:90% for empirical work. 00:34:11.860 --> 00:34:14.159 align:middle line:84% And so empirical work tends to have these models that 00:34:14.159 --> 00:34:18.060 align:middle line:84% have this very limited sense of dynamic competition 00:34:18.060 --> 00:34:20.370 align:middle line:84% and of incomplete information and other factors 00:34:20.370 --> 00:34:25.440 align:middle line:84% because it's not what you can estimate well structurally-- 00:34:25.440 --> 00:34:27.190 align:middle line:84% although, obviously, there's a lot of work 00:34:27.190 --> 00:34:29.230 align:middle line:90% trying to do that better. 00:34:29.230 --> 00:34:33.940 align:middle line:84% So as far as how do you come up with a topic for an empirical IO 00:34:33.940 --> 00:34:37.965 align:middle line:84% paper, people do multiple different things. 00:34:37.965 --> 00:34:39.340 align:middle line:84% One thing that you can always do, 00:34:39.340 --> 00:34:41.170 align:middle line:84% just like in any other applied field, 00:34:41.170 --> 00:34:44.050 align:middle line:84% is just come up with some applied or policy question 00:34:44.050 --> 00:34:49.239 align:middle line:84% and say, what can I teach people about that. 00:34:49.239 --> 00:34:51.880 align:middle line:84% A second thing one can do is improve understanding 00:34:51.880 --> 00:34:53.270 align:middle line:90% of some theoretical model. 00:34:53.270 --> 00:34:58.010 align:middle line:84% There's recent work on this monopoly information design. 00:34:58.010 --> 00:35:00.470 align:middle line:84% Can I do some empirical work to understand, 00:35:00.470 --> 00:35:02.560 align:middle line:84% does information design work in practice, 00:35:02.560 --> 00:35:05.050 align:middle line:84% or how well does it work or does it not work? 00:35:05.050 --> 00:35:06.910 align:middle line:84% The other third thing you can do is also do 00:35:06.910 --> 00:35:09.140 align:middle line:84% an improvement on existing empirical technique. 00:35:09.140 --> 00:35:11.140 align:middle line:84% If there are many previous papers trying 00:35:11.140 --> 00:35:14.540 align:middle line:84% to estimate empirical models of auctions, 00:35:14.540 --> 00:35:16.810 align:middle line:84% I have some better way to estimate 00:35:16.810 --> 00:35:18.102 align:middle line:90% empirical model of an auction. 00:35:18.102 --> 00:35:19.810 align:middle line:84% And you try to find a data set to improve 00:35:19.810 --> 00:35:21.186 align:middle line:90% on that empirical technique. 00:35:21.186 --> 00:35:24.340 align:middle line:90% 00:35:24.340 --> 00:35:26.298 align:middle line:90% How do you find your topic? 00:35:26.298 --> 00:35:28.090 align:middle line:84% Obviously, if you have an applied question, 00:35:28.090 --> 00:35:30.940 align:middle line:84% you look for the data to find it. 00:35:30.940 --> 00:35:34.090 align:middle line:84% If you're trying to do something improving understanding 00:35:34.090 --> 00:35:36.490 align:middle line:84% of a theoretical model, this is where 00:35:36.490 --> 00:35:38.480 align:middle line:84% we get papers on ketchup and cardboard. 00:35:38.480 --> 00:35:41.680 align:middle line:84% It's like, it's very complicated to estimate anything 00:35:41.680 --> 00:35:44.635 align:middle line:84% in the mobile phone environment or the computer environment 00:35:44.635 --> 00:35:46.510 align:middle line:84% because there's so many different firms doing 00:35:46.510 --> 00:35:48.532 align:middle line:90% so many different things. 00:35:48.532 --> 00:35:50.740 align:middle line:84% If you're trying to study some topic and you're like, 00:35:50.740 --> 00:35:53.330 align:middle line:84% I could study this in the cardboard box industry, 00:35:53.330 --> 00:35:56.080 align:middle line:84% the cardboard box industry is an incredibly simple industry. 00:35:56.080 --> 00:36:01.210 align:middle line:84% Basically, you take paper or you take leftover waste 00:36:01.210 --> 00:36:07.060 align:middle line:84% from wood mills, you mush it together, you add water, 00:36:07.060 --> 00:36:09.400 align:middle line:90% you dry it. 00:36:09.400 --> 00:36:11.750 align:middle line:84% There's just, like, three ingredients. 00:36:11.750 --> 00:36:13.360 align:middle line:90% There's nothing much there. 00:36:13.360 --> 00:36:14.748 align:middle line:90% You have shipping costs. 00:36:14.748 --> 00:36:17.290 align:middle line:84% You can estimate what it costs to ship cardboard from point A 00:36:17.290 --> 00:36:18.380 align:middle line:90% to point B. 00:36:18.380 --> 00:36:20.800 align:middle line:84% If you find an industry that strips away 00:36:20.800 --> 00:36:23.270 align:middle line:84% all the complicated stuff going on, 00:36:23.270 --> 00:36:25.190 align:middle line:84% there aren't agreements with Google doing 00:36:25.190 --> 00:36:26.430 align:middle line:90% this for some other reason. 00:36:26.430 --> 00:36:27.870 align:middle line:90% It's just cardboard. 00:36:27.870 --> 00:36:31.370 align:middle line:84% So sometimes you can pick a simple industry 00:36:31.370 --> 00:36:34.130 align:middle line:84% in order to study a complicated model because you 00:36:34.130 --> 00:36:36.890 align:middle line:84% want to isolate that theoretical factor that you're 00:36:36.890 --> 00:36:38.720 align:middle line:90% interested in. 00:36:38.720 --> 00:36:44.480 align:middle line:84% Third thing you can always do is find data. 00:36:44.480 --> 00:36:46.230 align:middle line:90% Data is always in short supply. 00:36:46.230 --> 00:36:48.570 align:middle line:84% IO papers tend to collect their own data. 00:36:48.570 --> 00:36:50.930 align:middle line:84% So you have a friend who works at a startup 00:36:50.930 --> 00:36:53.300 align:middle line:84% and is willing to give you individual-level consumer 00:36:53.300 --> 00:36:55.690 align:middle line:90% data on something or other. 00:36:55.690 --> 00:36:58.190 align:middle line:84% I tend to think, you look at all the topics in this course-- 00:36:58.190 --> 00:36:59.090 align:middle line:84% I'm going to talk about monopoly, 00:36:59.090 --> 00:37:00.480 align:middle line:84% I'm going to talk about price discrimination, 00:37:00.480 --> 00:37:01.340 align:middle line:84% I'm going to talk about entry, and I'm 00:37:01.340 --> 00:37:02.715 align:middle line:84% going to talk about search costs. 00:37:02.715 --> 00:37:05.600 align:middle line:84% And there must be some of those things that are involved 00:37:05.600 --> 00:37:06.810 align:middle line:90% in your friend's business. 00:37:06.810 --> 00:37:09.260 align:middle line:84% So if you can get the really, really rich microdata 00:37:09.260 --> 00:37:12.547 align:middle line:84% that no one has, there must be some interesting comment 00:37:12.547 --> 00:37:14.130 align:middle line:84% about that industry that you can make. 00:37:14.130 --> 00:37:15.950 align:middle line:84% There must be something that that data 00:37:15.950 --> 00:37:17.340 align:middle line:90% set is good for studying. 00:37:17.340 --> 00:37:20.810 align:middle line:84% And so that's another way to think about paper topics. 00:37:20.810 --> 00:37:25.520 align:middle line:84% So I would say that somehow, back to the 1970s, 00:37:25.520 --> 00:37:29.000 align:middle line:84% there has also been this recent revival of interest in IO in big 00:37:29.000 --> 00:37:33.200 align:middle line:84% questions, in part driven by macroeconomists asking big IO 00:37:33.200 --> 00:37:36.080 align:middle line:84% questions because the IO people haven't been doing it. 00:37:36.080 --> 00:37:38.480 align:middle line:84% And so there has been a big question in these questions 00:37:38.480 --> 00:37:44.360 align:middle line:84% like, Has competition declined in the US economy as a whole? 00:37:44.360 --> 00:37:46.220 align:middle line:84% And do we believe that there's been 00:37:46.220 --> 00:37:48.830 align:middle line:84% this increase in mark-ups or drop 00:37:48.830 --> 00:37:52.940 align:middle line:84% in concentration over the last decades? 00:37:52.940 --> 00:37:56.000 align:middle line:84% So two papers are both-- these are both going 00:37:56.000 --> 00:37:57.860 align:middle line:90% to be classic NEIO papers. 00:37:57.860 --> 00:38:00.440 align:middle line:84% So they're going to have industry-specific questions 00:38:00.440 --> 00:38:01.260 align:middle line:90% of interest. 00:38:01.260 --> 00:38:03.890 align:middle line:84% And they're also going to have theoretical models 00:38:03.890 --> 00:38:08.840 align:middle line:84% that they're either trying to comment on or use as guidance. 00:38:08.840 --> 00:38:11.840 align:middle line:84% The first one is, I would say, really-- 00:38:11.840 --> 00:38:15.080 align:middle line:84% it's an industry-specific question of interest. 00:38:15.080 --> 00:38:18.080 align:middle line:84% It's also, more than that, a paper that's 00:38:18.080 --> 00:38:20.240 align:middle line:84% about understanding theoretical models 00:38:20.240 --> 00:38:21.900 align:middle line:84% and whether they really apply or not. 00:38:21.900 --> 00:38:24.260 align:middle line:84% So the paper's, "Are durable goods consumers 00:38:24.260 --> 00:38:27.320 align:middle line:84% forward-looking-- evidence from college textbooks." 00:38:27.320 --> 00:38:29.390 align:middle line:84% To the extent there are policy questions here, 00:38:29.390 --> 00:38:30.860 align:middle line:84% these should be familiar with you. 00:38:30.860 --> 00:38:36.740 align:middle line:84% One policy question is, Why is it that textbooks cost $300? 00:38:36.740 --> 00:38:39.210 align:middle line:84% We know that you can print a book for 20 bucks. 00:38:39.210 --> 00:38:40.570 align:middle line:90% Why do they cost 300? 00:38:40.570 --> 00:38:42.040 align:middle line:84% Is there something wrong with that? 00:38:42.040 --> 00:38:43.470 align:middle line:84% And the other thing I don't know if you experienced 00:38:43.470 --> 00:38:46.440 align:middle line:84% as an undergraduate is, Why is it that this book that was only 00:38:46.440 --> 00:38:49.980 align:middle line:84% 2-and-1/2 years old is now already having a new edition 00:38:49.980 --> 00:38:53.100 align:middle line:84% come out, and I can't sell my old book back to the bookstore, 00:38:53.100 --> 00:38:55.860 align:middle line:84% which I had bought for $300, and it's now worthless 00:38:55.860 --> 00:38:59.370 align:middle line:84% because the 14th edition of somebody's book just came out 00:38:59.370 --> 00:39:01.900 align:middle line:84% replacing the 13th edition three years later? 00:39:01.900 --> 00:39:04.530 align:middle line:84% Is this some kind of scam or whatever 00:39:04.530 --> 00:39:06.340 align:middle line:84% the firms are doing that there's something 00:39:06.340 --> 00:39:08.010 align:middle line:84% about the durable goods sales where they bring out 00:39:08.010 --> 00:39:10.140 align:middle line:84% these new books over and over and over again 00:39:10.140 --> 00:39:11.910 align:middle line:84% to earn greater profits than they could 00:39:11.910 --> 00:39:14.130 align:middle line:84% if they just brought out books at an efficient time 00:39:14.130 --> 00:39:15.620 align:middle line:90% scale, whatever that means? 00:39:15.620 --> 00:39:22.325 align:middle line:90% 00:39:22.325 --> 00:39:24.520 align:middle line:84% So I think I said most of these things. 00:39:24.520 --> 00:39:27.430 align:middle line:84% Yeah, so obviously, the theory-related question 00:39:27.430 --> 00:39:32.460 align:middle line:84% is, Are these durable goods models relevant in practice? 00:39:32.460 --> 00:39:35.530 align:middle line:84% In particular, are consumers forward-looking and 00:39:35.530 --> 00:39:40.158 align:middle line:84% well-designed and solving dynamic optimization problems 00:39:40.158 --> 00:39:41.950 align:middle line:84% when they're doing things, which always you 00:39:41.950 --> 00:39:45.130 align:middle line:84% think about students are so bad at answering questions 00:39:45.130 --> 00:39:47.690 align:middle line:84% on my exams about dynamic durable good pricing. 00:39:47.690 --> 00:39:50.290 align:middle line:84% But somehow, we always assume that they always 00:39:50.290 --> 00:39:52.480 align:middle line:84% know how to do that for themselves in real life 00:39:52.480 --> 00:39:54.910 align:middle line:84% and think about the future discounted value 00:39:54.910 --> 00:39:58.300 align:middle line:84% when they're making any purchases that they're making. 00:39:58.300 --> 00:40:00.276 align:middle line:90% Is that a good assumption? 00:40:00.276 --> 00:40:02.890 align:middle line:90% 00:40:02.890 --> 00:40:04.780 align:middle line:84% The choice to study this, if you're 00:40:04.780 --> 00:40:12.160 align:middle line:84% trying to study durable goods and forward-looking consumers, 00:40:12.160 --> 00:40:14.740 align:middle line:84% this paper, you can see why they do what they do. 00:40:14.740 --> 00:40:17.860 align:middle line:84% First is, in addition to studying durable goods, 00:40:17.860 --> 00:40:20.570 align:middle line:84% they do get an industry-specific question of interest 00:40:20.570 --> 00:40:21.280 align:middle line:90% at the same time. 00:40:21.280 --> 00:40:25.060 align:middle line:84% It's good to do two things instead of one in a paper. 00:40:25.060 --> 00:40:28.420 align:middle line:84% There are aspects of this textbook example 00:40:28.420 --> 00:40:30.010 align:middle line:84% where the simplicity of the business 00:40:30.010 --> 00:40:31.280 align:middle line:90% makes the modeling easier. 00:40:31.280 --> 00:40:34.945 align:middle line:84% So, for instance, all consumers have to either buy the textbook 00:40:34.945 --> 00:40:36.820 align:middle line:84% in the first week of the semester or the week 00:40:36.820 --> 00:40:39.590 align:middle line:84% before the semester starts or not at all. 00:40:39.590 --> 00:40:42.760 align:middle line:84% There's none of this "I can wait till 7 weeks into the semester 00:40:42.760 --> 00:40:46.060 align:middle line:84% and buy the book or buy it a semester after I take the course 00:40:46.060 --> 00:40:48.920 align:middle line:84% to take advantage of the price declines or things like that." 00:40:48.920 --> 00:40:51.670 align:middle line:84% It just-- it's a static decision. 00:40:51.670 --> 00:40:53.770 align:middle line:84% The other thing is that they only buy the textbook 00:40:53.770 --> 00:40:54.850 align:middle line:90% that you're assigned. 00:40:54.850 --> 00:40:57.580 align:middle line:84% If you're assigned by Pindyck and Rubinfeld's 00:40:57.580 --> 00:41:00.250 align:middle line:84% intermediate micro textbook, you don't think about, 00:41:00.250 --> 00:41:04.630 align:middle line:84% OK, instead of buying Pindyck and Rubinfeld, 00:41:04.630 --> 00:41:06.910 align:middle line:84% I could buy Bernheim and Winston, 00:41:06.910 --> 00:41:09.715 align:middle line:84% or I could buy whoever and Snyder. 00:41:09.715 --> 00:41:12.900 align:middle line:90% 00:41:12.900 --> 00:41:15.360 align:middle line:84% In some sense, students only make 00:41:15.360 --> 00:41:17.920 align:middle line:84% the decision between two products or three products-- 00:41:17.920 --> 00:41:19.930 align:middle line:84% buy the new copy of Pindyck and Rubinfeld, 00:41:19.930 --> 00:41:22.420 align:middle line:84% buy the used copy of Pindyck and Rubinfeld, buy nothing. 00:41:22.420 --> 00:41:24.450 align:middle line:84% And so you get this much simpler demand system 00:41:24.450 --> 00:41:26.400 align:middle line:84% than you would get if students were 00:41:26.400 --> 00:41:32.030 align:middle line:84% choosing from this menu of 57 different things they could buy. 00:41:32.030 --> 00:41:34.160 align:middle line:84% We also can think about things we can ignore, 00:41:34.160 --> 00:41:35.780 align:middle line:84% like in many used good markets, you'd 00:41:35.780 --> 00:41:37.880 align:middle line:90% want to study a lemons problem. 00:41:37.880 --> 00:41:39.830 align:middle line:84% What do consumers think about the quality 00:41:39.830 --> 00:41:41.600 align:middle line:84% of the used car versus the new car, 00:41:41.600 --> 00:41:43.642 align:middle line:84% and how is that impacting them, and how does that 00:41:43.642 --> 00:41:44.610 align:middle line:90% change over time? 00:41:44.610 --> 00:41:46.490 align:middle line:84% Used textbooks, especially if you buy them 00:41:46.490 --> 00:41:48.073 align:middle line:84% in the college bookstore, you can just 00:41:48.073 --> 00:41:51.440 align:middle line:84% open them up, look at them, see how much highlighting there is 00:41:51.440 --> 00:41:53.370 align:middle line:90% in it, and then buy it or not. 00:41:53.370 --> 00:41:56.060 align:middle line:84% There isn't this asymmetric information 00:41:56.060 --> 00:41:57.535 align:middle line:90% we would want to model. 00:41:57.535 --> 00:41:59.660 align:middle line:84% And then the other thing you're always trying to do 00:41:59.660 --> 00:42:07.140 align:middle line:84% is get enough observations to estimate your model. 00:42:07.140 --> 00:42:09.980 align:middle line:84% And here, what's really nice is that there's a very, very 00:42:09.980 --> 00:42:11.790 align:middle line:90% large number of textbooks. 00:42:11.790 --> 00:42:16.250 align:middle line:84% And so they have three subjects-- 00:42:16.250 --> 00:42:16.820 align:middle line:90% what is it? 00:42:16.820 --> 00:42:19.770 align:middle line:84% It's economics, psychology, and biology. 00:42:19.770 --> 00:42:22.232 align:middle line:84% They have 10 semesters in which they can study it, 00:42:22.232 --> 00:42:24.440 align:middle line:84% and then they have hundreds and hundreds and hundreds 00:42:24.440 --> 00:42:25.980 align:middle line:90% and hundreds of textbooks. 00:42:25.980 --> 00:42:29.870 align:middle line:84% And so that combination of semesters times subjects times 00:42:29.870 --> 00:42:33.490 align:middle line:84% the textbooks is going to give you many, many observations. 00:42:33.490 --> 00:42:35.280 align:middle line:84% And you can-- there's always this thing 00:42:35.280 --> 00:42:37.337 align:middle line:84% of if you study 100 different products, 00:42:37.337 --> 00:42:39.420 align:middle line:84% are those 100 different products really identical? 00:42:39.420 --> 00:42:41.470 align:middle line:84% Can you use the same parameters for them? 00:42:41.470 --> 00:42:43.920 align:middle line:84% You can think textbooks are similar enough to each other 00:42:43.920 --> 00:42:46.860 align:middle line:84% that the elasticity of demand for Pindyck and Rubinfeld 00:42:46.860 --> 00:42:49.320 align:middle line:84% and the elasticity of demand for Bernheim and Winston 00:42:49.320 --> 00:42:50.903 align:middle line:84% are probably the same, and we can just 00:42:50.903 --> 00:42:52.800 align:middle line:84% put one parameter in there instead of having 00:42:52.800 --> 00:42:53.906 align:middle line:90% hundreds of parameters. 00:42:53.906 --> 00:42:58.230 align:middle line:90% 00:42:58.230 --> 00:43:01.512 align:middle line:84% Some background facts before you start modeling-- 00:43:01.512 --> 00:43:03.720 align:middle line:84% it's always good to think about things that are true. 00:43:03.720 --> 00:43:06.000 align:middle line:84% Well, one is that new editions come out 00:43:06.000 --> 00:43:08.850 align:middle line:90% every three to five years. 00:43:08.850 --> 00:43:12.820 align:middle line:84% Second fact is that textbook prices seem to be fixed. 00:43:12.820 --> 00:43:15.615 align:middle line:84% Why this is true is kind of a puzzle. 00:43:15.615 --> 00:43:17.490 align:middle line:84% Maybe you have to have some reputation model, 00:43:17.490 --> 00:43:19.800 align:middle line:84% but you might have thought, OK, here's 00:43:19.800 --> 00:43:21.510 align:middle line:90% how I'm going to do this test. 00:43:21.510 --> 00:43:24.463 align:middle line:84% The book is going to come out the first-- 00:43:24.463 --> 00:43:25.880 align:middle line:84% if I buy the book when it's new, I 00:43:25.880 --> 00:43:28.620 align:middle line:84% know I could rent it to five future students. 00:43:28.620 --> 00:43:30.225 align:middle line:84% If I buy the book one semester, well, 00:43:30.225 --> 00:43:32.100 align:middle line:84% I'm going to rent it to four future students. 00:43:32.100 --> 00:43:33.642 align:middle line:84% If I buy it, like, four semesters in, 00:43:33.642 --> 00:43:35.550 align:middle line:84% I'm going to rent it once to somebody. 00:43:35.550 --> 00:43:38.250 align:middle line:84% So there ought to be this real decline in the prices of a book. 00:43:38.250 --> 00:43:40.520 align:middle line:84% And a book that's five semesters old 00:43:40.520 --> 00:43:42.590 align:middle line:84% ought to be worth one-fifth of a book that's 00:43:42.590 --> 00:43:45.800 align:middle line:84% brand new because you can't keep renting it over and over again. 00:43:45.800 --> 00:43:47.720 align:middle line:90% Well, that just doesn't happen. 00:43:47.720 --> 00:43:50.160 align:middle line:90% Textbook prices do not drop. 00:43:50.160 --> 00:43:52.860 align:middle line:90% The book comes out, it's $350. 00:43:52.860 --> 00:43:55.590 align:middle line:90% Five semesters later, it's $350. 00:43:55.590 --> 00:43:58.340 align:middle line:84% So for whatever reason, we're not 00:43:58.340 --> 00:44:01.310 align:middle line:84% going to think of optimal pricing. 00:44:01.310 --> 00:44:04.520 align:middle line:84% Maybe it's a pricing puzzle, but textbook prices just 00:44:04.520 --> 00:44:06.475 align:middle line:84% don't change in between editions. 00:44:06.475 --> 00:44:11.440 align:middle line:90% 00:44:11.440 --> 00:44:13.330 align:middle line:84% I don't know to what extent these are still 00:44:13.330 --> 00:44:14.380 align:middle line:90% true in the internet era. 00:44:14.380 --> 00:44:17.500 align:middle line:84% But at the time, the used-to-new ratio 00:44:17.500 --> 00:44:22.990 align:middle line:90% was standardized at 50% to 75%. 00:44:22.990 --> 00:44:24.190 align:middle line:90% Does the Coop even sell-- 00:44:24.190 --> 00:44:26.440 align:middle line:84% you can still go to the Coop and buy all the textbooks 00:44:26.440 --> 00:44:27.107 align:middle line:90% for the courses? 00:44:27.107 --> 00:44:28.060 align:middle line:90% Is that right? 00:44:28.060 --> 00:44:30.765 align:middle line:90% So maybe no one does that, OK. 00:44:30.765 --> 00:44:33.070 align:middle line:84% So anyway, if you wanted to buy your books, 00:44:33.070 --> 00:44:34.390 align:middle line:90% you would go to the MIT Coop. 00:44:34.390 --> 00:44:35.932 align:middle line:84% Down in the basement of the MIT Coop, 00:44:35.932 --> 00:44:37.510 align:middle line:90% there'd be a bunch of textbooks. 00:44:37.510 --> 00:44:40.550 align:middle line:84% MIT would have an agreement with the bookstore. 00:44:40.550 --> 00:44:42.610 align:middle line:84% In this case, we used to own the bookstore, 00:44:42.610 --> 00:44:44.810 align:middle line:84% but it's now Barnes and Noble who owns it. 00:44:44.810 --> 00:44:46.930 align:middle line:84% But they would sell the used textbook 00:44:46.930 --> 00:44:50.680 align:middle line:84% at, at most, 75% of the price of the new textbook. 00:44:50.680 --> 00:44:52.210 align:middle line:84% And then you would also often have 00:44:52.210 --> 00:44:53.710 align:middle line:84% an agreement with the company that was selling 00:44:53.710 --> 00:44:54.710 align:middle line:90% books in your bookstore. 00:44:54.710 --> 00:44:58.030 align:middle line:84% If you want to buy back used books at the MIT Coop, 00:44:58.030 --> 00:45:01.810 align:middle line:84% you must pay 50% of the cover price when you buy it back. 00:45:01.810 --> 00:45:05.082 align:middle line:84% And so there are these norms of people 00:45:05.082 --> 00:45:06.790 align:middle line:84% would show up in the bookstore at the end 00:45:06.790 --> 00:45:08.707 align:middle line:84% of the semester willing to buy back your books 00:45:08.707 --> 00:45:10.460 align:middle line:90% at 50% of the cover price. 00:45:10.460 --> 00:45:12.950 align:middle line:84% If you wanted to buy a used book instead of a new book, 00:45:12.950 --> 00:45:17.790 align:middle line:84% it would be sold at 75% of the cover price. 00:45:17.790 --> 00:45:23.280 align:middle line:84% And from aggregate numbers, roughly 30% 00:45:23.280 --> 00:45:26.510 align:middle line:84% of the students in a course at the time would buy a textbook. 00:45:26.510 --> 00:45:29.310 align:middle line:84% About 20% would buy a used textbook. 00:45:29.310 --> 00:45:31.380 align:middle line:84% And then the other 50%, we hope, make 00:45:31.380 --> 00:45:34.530 align:middle line:84% extensive use of the library or some other resource 00:45:34.530 --> 00:45:36.725 align:middle line:84% because they don't seem to buy books at all. 00:45:36.725 --> 00:45:40.910 align:middle line:90% 00:45:40.910 --> 00:45:43.820 align:middle line:84% The hazard rate for revision seems 00:45:43.820 --> 00:45:46.550 align:middle line:84% to peak in about the third year, that many books 00:45:46.550 --> 00:45:49.110 align:middle line:84% have a new version that comes out every three years. 00:45:49.110 --> 00:45:51.380 align:middle line:84% So if you're looking at John Gruber's public finance 00:45:51.380 --> 00:45:54.750 align:middle line:84% textbook, if it's been two years since it's come out, 00:45:54.750 --> 00:45:57.650 align:middle line:84% he's probably going to have a new edition next year. 00:45:57.650 --> 00:45:59.930 align:middle line:84% Once you've gone a bit past three, 00:45:59.930 --> 00:46:02.510 align:middle line:84% the probability that it gets revised goes down and down 00:46:02.510 --> 00:46:06.380 align:middle line:84% and down, because it's probably the faculty member has just 00:46:06.380 --> 00:46:09.560 align:middle line:84% gotten exhausted with the 11th edition or the 12th or the 13th 00:46:09.560 --> 00:46:12.590 align:middle line:84% and has decided they're not making enough money off the new 00:46:12.590 --> 00:46:14.093 align:middle line:90% editions to write the new ones. 00:46:14.093 --> 00:46:16.260 align:middle line:84% They haven't got a new person to take over the book. 00:46:16.260 --> 00:46:19.730 align:middle line:84% But if the thing hasn't come out and it's a few years late, 00:46:19.730 --> 00:46:21.235 align:middle line:90% it's probably not coming out. 00:46:21.235 --> 00:46:24.460 align:middle line:90% 00:46:24.460 --> 00:46:27.760 align:middle line:84% So papers, empirical approach is a two-step strategy. 00:46:27.760 --> 00:46:29.320 align:middle line:84% I'm going to estimate the probability 00:46:29.320 --> 00:46:30.910 align:middle line:84% that the book is going to be revised 00:46:30.910 --> 00:46:32.830 align:middle line:90% before the next semester. 00:46:32.830 --> 00:46:34.930 align:middle line:84% And then I'm going to see whether-- 00:46:34.930 --> 00:46:37.720 align:middle line:84% because prices don't change, if you're a student 00:46:37.720 --> 00:46:40.060 align:middle line:84% and you know a new book is coming out next semester, 00:46:40.060 --> 00:46:41.938 align:middle line:84% you shouldn't buy, or you should be 00:46:41.938 --> 00:46:43.480 align:middle line:84% willing to pay much less for the book 00:46:43.480 --> 00:46:48.940 align:middle line:84% because you know that the cost is really $350, not $175, 00:46:48.940 --> 00:46:52.510 align:middle line:84% so that, in some sense, the price is much, much 00:46:52.510 --> 00:46:54.910 align:middle line:84% higher in the semester when a book is 00:46:54.910 --> 00:46:58.420 align:middle line:84% about to be rendered obsolete than it is in other semesters. 00:46:58.420 --> 00:47:01.450 align:middle line:84% And so what we ought to see is a drop in price, 00:47:01.450 --> 00:47:04.030 align:middle line:84% a drop in demand in semesters when 00:47:04.030 --> 00:47:05.655 align:middle line:84% books are going to be obsolete or when 00:47:05.655 --> 00:47:07.030 align:middle line:84% people are going to think there's 00:47:07.030 --> 00:47:10.840 align:middle line:84% a higher probability that they will become obsolete. 00:47:10.840 --> 00:47:14.690 align:middle line:84% And if consumers are rational and forward-looking, 00:47:14.690 --> 00:47:16.850 align:middle line:90% we can see-- 00:47:16.850 --> 00:47:19.400 align:middle line:84% we can figure out-- if we know the elasticity of demand 00:47:19.400 --> 00:47:22.340 align:middle line:84% for books, we know how much drop demand would go down 00:47:22.340 --> 00:47:23.872 align:middle line:90% with such a price increase. 00:47:23.872 --> 00:47:25.580 align:middle line:84% We can almost estimate how many consumers 00:47:25.580 --> 00:47:27.330 align:middle line:84% are rational and forward-looking by seeing 00:47:27.330 --> 00:47:30.265 align:middle line:84% how much does demand drop when that new book is coming out. 00:47:30.265 --> 00:47:33.330 align:middle line:90% 00:47:33.330 --> 00:47:39.420 align:middle line:84% So step 1 is a very simple exercise. 00:47:39.420 --> 00:47:45.000 align:middle line:90% This is just regress-- 00:47:45.000 --> 00:47:47.760 align:middle line:84% you just estimate a hazard model where you just 00:47:47.760 --> 00:47:50.010 align:middle line:84% have the probability that the book is replaced 00:47:50.010 --> 00:47:53.910 align:middle line:84% is a function of the age of the book in semesters, 00:47:53.910 --> 00:47:57.570 align:middle line:84% whether it's the biology or psychology or economics book, 00:47:57.570 --> 00:48:03.603 align:middle line:84% whether it's an introductory textbook like for economics I 00:48:03.603 --> 00:48:05.520 align:middle line:84% or whether it's an intermediate micro textbook 00:48:05.520 --> 00:48:08.430 align:middle line:90% or intermediate macro textbook. 00:48:08.430 --> 00:48:12.510 align:middle line:84% And here's the graph they get for the introductory books. 00:48:12.510 --> 00:48:15.300 align:middle line:84% So you can see in economics books, 00:48:15.300 --> 00:48:17.110 align:middle line:90% your book is five semesters old. 00:48:17.110 --> 00:48:19.500 align:middle line:84% There's, like, a 35% chance your book will be rendered 00:48:19.500 --> 00:48:21.960 align:middle line:90% obsolete one semester from now. 00:48:21.960 --> 00:48:25.122 align:middle line:84% If your book is six semesters old-- 00:48:25.122 --> 00:48:26.950 align:middle line:84% oh, actually, this is six semesters old. 00:48:26.950 --> 00:48:30.365 align:middle line:84% So there's actually only an 8% chance 00:48:30.365 --> 00:48:32.990 align:middle line:84% that your book is going to get rendered obsolete next semester. 00:48:32.990 --> 00:48:34.150 align:middle line:90% Or is it here? 00:48:34.150 --> 00:48:37.690 align:middle line:84% Anyway, seven semesters old, it's a 5% chance. 00:48:37.690 --> 00:48:40.420 align:middle line:84% But there's this big peak centered 00:48:40.420 --> 00:48:44.160 align:middle line:84% around buying five-semester-old books is a bad idea. 00:48:44.160 --> 00:48:47.770 align:middle line:84% And the thing is less consistent in some 00:48:47.770 --> 00:48:49.160 align:middle line:90% of the other disciplines. 00:48:49.160 --> 00:48:52.150 align:middle line:84% Some of the biology books are being on a two-year cycle 00:48:52.150 --> 00:48:53.860 align:middle line:90% instead of a three-year cycle. 00:48:53.860 --> 00:48:57.040 align:middle line:84% Some of the psychology books are on a 2-and-1/2 year cycle, 00:48:57.040 --> 00:48:57.900 align:middle line:90% something like that. 00:48:57.900 --> 00:49:00.426 align:middle line:90% 00:49:00.426 --> 00:49:03.630 align:middle line:90% And again, your book is-- 00:49:03.630 --> 00:49:06.210 align:middle line:84% your economics book is eight years old, or any of your books 00:49:06.210 --> 00:49:08.310 align:middle line:84% are eight years since the edition has changed, 00:49:08.310 --> 00:49:10.080 align:middle line:84% they're not going to have a new one. 00:49:10.080 --> 00:49:11.660 align:middle line:84% So that book is very safe to buy. 00:49:11.660 --> 00:49:17.860 align:middle line:90% 00:49:17.860 --> 00:49:21.070 align:middle line:90% How do we do demand? 00:49:21.070 --> 00:49:22.578 align:middle line:90% So can I ask-- 00:49:22.578 --> 00:49:24.620 align:middle line:84% well, actually, some of you haven't taken it yet. 00:49:24.620 --> 00:49:27.610 align:middle line:84% So how much discrete choice estimation is there left 00:49:27.610 --> 00:49:30.260 align:middle line:90% in [14.]381? 00:49:30.260 --> 00:49:32.475 align:middle line:84% Or has it all gone to causal inference these days? 00:49:32.475 --> 00:49:35.558 align:middle line:90% 00:49:35.558 --> 00:49:36.850 align:middle line:90% A lot of discrete choice or no? 00:49:36.850 --> 00:49:37.640 align:middle line:90% AUDIENCE: Causal inference. 00:49:37.640 --> 00:49:39.000 align:middle line:84% GLENN ELLISON: It's all causal inference. 00:49:39.000 --> 00:49:41.390 align:middle line:84% So is there any place where people are doing logit models 00:49:41.390 --> 00:49:44.630 align:middle line:90% and whatever? 00:49:44.630 --> 00:49:45.578 align:middle line:90% OK. 00:49:45.578 --> 00:49:47.210 align:middle line:90% [LAUGHS] 00:49:47.210 --> 00:49:51.530 align:middle line:84% OK, anyway, let me say, Tobias has a full week of lectures 00:49:51.530 --> 00:49:53.547 align:middle line:84% planned on discrete choice estimation, 00:49:53.547 --> 00:49:55.380 align:middle line:84% which obviously, a full week is not so much. 00:49:55.380 --> 00:49:57.350 align:middle line:90% There's much more in [14.]273. 00:49:57.350 --> 00:49:59.910 align:middle line:84% But anyway, so it's week four of this course, 00:49:59.910 --> 00:50:03.950 align:middle line:84% Tobias is going to spend a week on discrete choice estimation. 00:50:03.950 --> 00:50:06.770 align:middle line:84% For now, let me just say this is how everybody always 00:50:06.770 --> 00:50:08.920 align:middle line:84% seems to do demand estimation in IO these days. 00:50:08.920 --> 00:50:10.587 align:middle line:84% There are a lot of funny elements of it. 00:50:10.587 --> 00:50:13.850 align:middle line:84% But so what we're doing here is we have this model of consumers, 00:50:13.850 --> 00:50:15.482 align:middle line:84% and we specify a utility function 00:50:15.482 --> 00:50:17.690 align:middle line:84% that they're going to get from buying three objects-- 00:50:17.690 --> 00:50:23.390 align:middle line:84% a new book, a used book, and no book at all. 00:50:23.390 --> 00:50:26.570 align:middle line:84% And what utility do you get if you buy a new book? 00:50:26.570 --> 00:50:29.330 align:middle line:84% First, there's a Xjt beta, which is 00:50:29.330 --> 00:50:31.250 align:middle line:84% something on book characteristics 00:50:31.250 --> 00:50:35.110 align:middle line:90% times some coefficients. 00:50:35.110 --> 00:50:38.180 align:middle line:84% Then there's the disutility of paying for the book. 00:50:38.180 --> 00:50:41.940 align:middle line:84% So alpha is your price sensitivity. 00:50:41.940 --> 00:50:45.870 align:middle line:84% And then we multiply alpha by the net price 00:50:45.870 --> 00:50:47.490 align:middle line:90% that you would pay. 00:50:47.490 --> 00:50:50.940 align:middle line:84% The net price that you would pay is the full price 00:50:50.940 --> 00:50:55.450 align:middle line:84% minus a discount that you apply to money that's 00:50:55.450 --> 00:50:59.010 align:middle line:84% going to come to you in the future times 00:50:59.010 --> 00:51:02.740 align:middle line:84% the money you expect to get back when you resell the book. 00:51:02.740 --> 00:51:07.120 align:middle line:84% So this is if the book doesn't die, I can sell it back. 00:51:07.120 --> 00:51:10.650 align:middle line:84% And when I sell it back, I get mu times Pjt, 00:51:10.650 --> 00:51:12.930 align:middle line:90% where mu is maybe 50%. 00:51:12.930 --> 00:51:15.630 align:middle line:84% And so this discount factor delta 00:51:15.630 --> 00:51:19.140 align:middle line:84% is both I'm getting the money six months from now when I 00:51:19.140 --> 00:51:21.370 align:middle line:84% sell it back instead of getting it today. 00:51:21.370 --> 00:51:23.950 align:middle line:84% And it's also, am I thinking about it. 00:51:23.950 --> 00:51:27.000 align:middle line:84% Like, you could imagine totally myopic consumers are not 00:51:27.000 --> 00:51:29.790 align:middle line:84% thinking about selling it back, and they're just 00:51:29.790 --> 00:51:31.740 align:middle line:90% treating delta as 0. 00:51:31.740 --> 00:51:34.740 align:middle line:84% So they're just, Do I want to pay $350 today, 00:51:34.740 --> 00:51:36.065 align:middle line:90% not thinking about the future? 00:51:36.065 --> 00:51:40.374 align:middle line:90% 00:51:40.374 --> 00:51:43.640 align:middle line:84% In order to explain the data, there are always-- 00:51:43.640 --> 00:51:45.690 align:middle line:84% for some reason, [INAUDIBLE] started 00:51:45.690 --> 00:51:48.720 align:middle line:90% this use of the letter xi. 00:51:48.720 --> 00:51:52.230 align:middle line:84% So this is the unobserved quality of the book. 00:51:52.230 --> 00:51:56.040 align:middle line:84% What is it that makes Pindyck and Rubinfeld sell many more 00:51:56.040 --> 00:51:59.370 align:middle line:84% copies than many other intermediate micro textbooks? 00:51:59.370 --> 00:52:02.760 align:middle line:84% The characteristics in terms of how many pages long it is 00:52:02.760 --> 00:52:05.010 align:middle line:84% and whether it's hardcover and whether it has a CD-ROM 00:52:05.010 --> 00:52:06.460 align:middle line:90% shrink-wrapped to it-- 00:52:06.460 --> 00:52:08.400 align:middle line:90% those, it's not very different. 00:52:08.400 --> 00:52:11.190 align:middle line:84% But it must have some unobserved-- 00:52:11.190 --> 00:52:14.280 align:middle line:84% or it may have some unobserved quality that makes it a higher 00:52:14.280 --> 00:52:16.710 align:middle line:90% fraction of students buy it. 00:52:16.710 --> 00:52:18.240 align:middle line:90% And then you have this-- 00:52:18.240 --> 00:52:20.430 align:middle line:84% consumers have this sort of a type. 00:52:20.430 --> 00:52:22.680 align:middle line:84% In my previous lectures, consumers always 00:52:22.680 --> 00:52:24.930 align:middle line:84% had this type theta that determined 00:52:24.930 --> 00:52:26.700 align:middle line:90% like their taste for goods. 00:52:26.700 --> 00:52:29.530 align:middle line:84% Here, consumers have a three-dimensional type. 00:52:29.530 --> 00:52:33.630 align:middle line:84% They have a three-dimensional type epsilon 00:52:33.630 --> 00:52:42.790 align:middle line:84% ijt new, epsilon ijt used, and epsilon ijt none. 00:52:42.790 --> 00:52:44.380 align:middle line:84% So I have this three-dimensional type 00:52:44.380 --> 00:52:47.050 align:middle line:84% describing how much I like buying new books, 00:52:47.050 --> 00:52:48.910 align:middle line:84% how much I like buying used books, 00:52:48.910 --> 00:52:51.010 align:middle line:84% and how much I like buying no books at all 00:52:51.010 --> 00:52:52.930 align:middle line:90% for this particular book. 00:52:52.930 --> 00:52:57.850 align:middle line:84% And we model this as being three independent extreme value 00:52:57.850 --> 00:53:00.100 align:middle line:90% errors, typically. 00:53:00.100 --> 00:53:02.750 align:middle line:84% So this would be my utility for buying a new book. 00:53:02.750 --> 00:53:04.990 align:middle line:84% Again, if my utility for buying a used book 00:53:04.990 --> 00:53:09.130 align:middle line:84% is going to be some function of the book characteristics 00:53:09.130 --> 00:53:12.370 align:middle line:84% plus some unobserved quality of the used book 00:53:12.370 --> 00:53:14.080 align:middle line:84% plus this idiosyncratic reaction, 00:53:14.080 --> 00:53:17.220 align:middle line:84% then I have an idiosyncratic utility for buying no book. 00:53:17.220 --> 00:53:21.340 align:middle line:84% And so what we model many consumers doing is having these 00:53:21.340 --> 00:53:24.100 align:middle line:90% three choices, and then based-- 00:53:24.100 --> 00:53:25.930 align:middle line:84% different people make different choices 00:53:25.930 --> 00:53:30.630 align:middle line:84% because different people have different epsilons. 00:53:30.630 --> 00:53:32.310 align:middle line:84% So we have this utility specification, 00:53:32.310 --> 00:53:36.000 align:middle line:84% but then given any set of parameters-- 00:53:36.000 --> 00:53:39.780 align:middle line:84% beta, alpha, mu, et cetera-- and given assumptions 00:53:39.780 --> 00:53:42.960 align:middle line:84% about the joint distributions of these objects, 00:53:42.960 --> 00:53:44.460 align:middle line:84% we're going to be able to get demand 00:53:44.460 --> 00:53:46.300 align:middle line:84% as a function of the economic primitives. 00:53:46.300 --> 00:53:46.800 align:middle line:90% Yeah? 00:53:46.800 --> 00:53:48.967 align:middle line:84% AUDIENCE: So just do the characteristics of the book 00:53:48.967 --> 00:53:53.560 align:middle line:84% enter differently, and your user is just yf or just another beta? 00:53:53.560 --> 00:53:55.440 align:middle line:84% GLENN ELLISON: Yeah, no, so it potentially 00:53:55.440 --> 00:53:58.020 align:middle line:84% could be xjt times beta, or it could 00:53:58.020 --> 00:53:59.970 align:middle line:90% be some other function of xjt. 00:53:59.970 --> 00:54:02.940 align:middle line:84% In their estimation, it turns out it's-- 00:54:02.940 --> 00:54:06.030 align:middle line:84% under their estimation assumptions, it doesn't matter. 00:54:06.030 --> 00:54:09.250 align:middle line:84% In some sense, any function here, the estimation would work. 00:54:09.250 --> 00:54:11.185 align:middle line:84% So it's like we estimate this robustly. 00:54:11.185 --> 00:54:13.685 align:middle line:84% We don't need to assume this equals this, and so they don't. 00:54:13.685 --> 00:54:17.370 align:middle line:90% 00:54:17.370 --> 00:54:22.410 align:middle line:84% So theorem-- and this is the magic of logit models. 00:54:22.410 --> 00:54:24.223 align:middle line:90% The magic of logit models-- 00:54:24.223 --> 00:54:26.640 align:middle line:84% and, you know, this may seem skeptical at first-- is like, 00:54:26.640 --> 00:54:29.970 align:middle line:84% basically, you have these quantities, 00:54:29.970 --> 00:54:32.040 align:middle line:84% and then you take the log of the quantity 00:54:32.040 --> 00:54:36.240 align:middle line:84% and you subtract something, and you have converted from quantity 00:54:36.240 --> 00:54:38.670 align:middle line:90% space to utility space. 00:54:38.670 --> 00:54:40.890 align:middle line:90% And again, Tobias explain more-- 00:54:40.890 --> 00:54:43.300 align:middle line:84% make this stronger argument for why that's not-- 00:54:43.300 --> 00:54:44.830 align:middle line:90% doesn't seem so crazy. 00:54:44.830 --> 00:54:47.820 align:middle line:84% Maybe it is crazy, but people do it all the time. 00:54:47.820 --> 00:54:51.120 align:middle line:84% So anyway, the theorem is, if the epsilons are i.i.d. 00:54:51.120 --> 00:54:55.920 align:middle line:84% across consumers and products, and each of those epsilons 00:54:55.920 --> 00:54:58.890 align:middle line:84% has a type 1 extreme value distribution, 00:54:58.890 --> 00:55:03.510 align:middle line:84% then the log of the ratio between the number of people 00:55:03.510 --> 00:55:06.930 align:middle line:84% buying new books and the people buying no books 00:55:06.930 --> 00:55:13.060 align:middle line:90% is that utility specification. 00:55:13.060 --> 00:55:16.210 align:middle line:84% So it is book characteristics times effect 00:55:16.210 --> 00:55:19.240 align:middle line:84% of book characteristics on utility minus alpha times 00:55:19.240 --> 00:55:23.170 align:middle line:84% the price, plus alpha delta mu times the probability 00:55:23.170 --> 00:55:25.210 align:middle line:84% that it doesn't die times the price, 00:55:25.210 --> 00:55:29.770 align:middle line:84% plus this thing that was the unobserved product quality 00:55:29.770 --> 00:55:34.900 align:middle line:84% becomes the error term if you're regressing new book 00:55:34.900 --> 00:55:39.560 align:middle line:84% sales on book characteristics and prices. 00:55:39.560 --> 00:55:42.950 align:middle line:84% So we make this functional form transformation, 00:55:42.950 --> 00:55:46.760 align:middle line:84% and we get to this thing that we could estimate these parameters, 00:55:46.760 --> 00:55:50.400 align:middle line:84% the utility function, just by running a regression. 00:55:50.400 --> 00:55:54.565 align:middle line:84% So we're going to regress log of the share-- 00:55:54.565 --> 00:55:55.940 align:middle line:84% log of the fraction of people who 00:55:55.940 --> 00:56:01.010 align:middle line:84% buy the book on book characteristics, 00:56:01.010 --> 00:56:03.730 align:middle line:84% prices, probability of death, and so on. 00:56:03.730 --> 00:56:09.820 align:middle line:90% 00:56:09.820 --> 00:56:12.430 align:middle line:84% One thing that this specification points out 00:56:12.430 --> 00:56:16.242 align:middle line:84% is, the coefficient on the book characteristics 00:56:16.242 --> 00:56:17.950 align:middle line:84% are going to tell us how much people like 00:56:17.950 --> 00:56:20.773 align:middle line:90% big books versus small books. 00:56:20.773 --> 00:56:23.440 align:middle line:84% Coefficient on price is going to tell us how much people dislike 00:56:23.440 --> 00:56:26.350 align:middle line:90% buying books. 00:56:26.350 --> 00:56:29.290 align:middle line:84% The coefficient on the probability 00:56:29.290 --> 00:56:35.080 align:middle line:84% of survival times the price is going to give us alpha times 00:56:35.080 --> 00:56:36.550 align:middle line:90% delta times mu. 00:56:36.550 --> 00:56:40.390 align:middle line:84% So delta and mu are not separately identified. 00:56:40.390 --> 00:56:43.690 align:middle line:84% So if I said that delta was the key parameter of are people 00:56:43.690 --> 00:56:46.300 align:middle line:84% forward-looking, the mu was how much money 00:56:46.300 --> 00:56:48.880 align:middle line:84% do they expect to get back when they resell the book, 00:56:48.880 --> 00:56:51.400 align:middle line:90% if they're allowed to resell it. 00:56:51.400 --> 00:56:53.630 align:middle line:84% We can't separately identify this product. 00:56:53.630 --> 00:56:56.580 align:middle line:84% All we can do is estimate those two things multiplied together. 00:56:56.580 --> 00:57:03.600 align:middle line:90% 00:57:03.600 --> 00:57:06.270 align:middle line:84% So another observation, we're going 00:57:06.270 --> 00:57:08.578 align:middle line:90% to need an instrument for price. 00:57:08.578 --> 00:57:10.620 align:middle line:84% Why are we going to need an instrument for price? 00:57:10.620 --> 00:57:12.245 align:middle line:84% We always need an instrument for price. 00:57:12.245 --> 00:57:17.730 align:middle line:84% But if the firm has a book that has a high unobserved quality, 00:57:17.730 --> 00:57:20.190 align:middle line:84% this is a book that many people like-- 00:57:20.190 --> 00:57:23.220 align:middle line:84% the same reason that just like when demand is high, 00:57:23.220 --> 00:57:25.420 align:middle line:84% price is high, to the extent that's true. 00:57:25.420 --> 00:57:29.790 align:middle line:84% But if the book has an epsilon new that 00:57:29.790 --> 00:57:33.600 align:middle line:84% makes the firm want to set a high price, 00:57:33.600 --> 00:57:37.390 align:middle line:84% then price is going to be correlated with this error term, 00:57:37.390 --> 00:57:39.000 align:middle line:90% so price is endogenous. 00:57:39.000 --> 00:57:42.870 align:middle line:84% So we need an instrument for price. 00:57:42.870 --> 00:57:44.360 align:middle line:90% What do they do? 00:57:44.360 --> 00:57:45.840 align:middle line:84% The instruments they use for price 00:57:45.840 --> 00:57:48.340 align:middle line:84% are having a nonprofit publisher. 00:57:48.340 --> 00:57:50.430 align:middle line:84% So, for instance, Tirole's textbook 00:57:50.430 --> 00:57:52.890 align:middle line:84% is published by MIT Press rather than 00:57:52.890 --> 00:57:55.770 align:middle line:90% by Macmillan or somebody else. 00:57:55.770 --> 00:57:59.610 align:middle line:84% Tirole's book is cheaper than many other books 00:57:59.610 --> 00:58:04.540 align:middle line:84% because nonprofit publishers tend to set lower prices. 00:58:04.540 --> 00:58:07.330 align:middle line:84% Other instruments they use are, What 00:58:07.330 --> 00:58:09.760 align:middle line:84% fraction of your close competitors 00:58:09.760 --> 00:58:11.930 align:middle line:84% are published by nonprofit publishers? 00:58:11.930 --> 00:58:15.190 align:middle line:84% So if intermediate macro books, a lot of them 00:58:15.190 --> 00:58:17.050 align:middle line:84% are sold by nonprofit publishers, 00:58:17.050 --> 00:58:18.730 align:middle line:84% then maybe the for-profit publishers 00:58:18.730 --> 00:58:21.340 align:middle line:90% also have to set lower prices. 00:58:21.340 --> 00:58:23.800 align:middle line:84% They also have a measure, which is what's 00:58:23.800 --> 00:58:25.250 align:middle line:90% the concentration of ownership? 00:58:25.250 --> 00:58:27.040 align:middle line:84% Are there 10 different major publishers 00:58:27.040 --> 00:58:30.700 align:middle line:84% who have intermediate machine learning books, 00:58:30.700 --> 00:58:33.250 align:middle line:84% or are there only two such publishers 00:58:33.250 --> 00:58:34.500 align:middle line:90% who have dominated the market? 00:58:34.500 --> 00:58:38.910 align:middle line:90% 00:58:38.910 --> 00:58:43.160 align:middle line:84% Anyway, so this is basically the estimating equation. 00:58:43.160 --> 00:58:45.018 align:middle line:90% So it's just an IV regression. 00:58:45.018 --> 00:58:47.560 align:middle line:84% And it's an IV regression that, under the assumptions they've 00:58:47.560 --> 00:58:49.990 align:middle line:84% made, should let us estimate delta mu, which 00:58:49.990 --> 00:58:50.910 align:middle line:90% is what we care about. 00:58:50.910 --> 00:58:54.070 align:middle line:90% 00:58:54.070 --> 00:58:55.530 align:middle line:90% And what do they find? 00:58:55.530 --> 00:58:58.670 align:middle line:90% 00:58:58.670 --> 00:59:03.600 align:middle line:84% So they define lambda equals delta times mu. 00:59:03.600 --> 00:59:05.310 align:middle line:90% What did we think mu is? 00:59:05.310 --> 00:59:08.460 align:middle line:84% We think that mu, you've got to be able to get at least-- 00:59:08.460 --> 00:59:11.170 align:middle line:84% normally, you'd get 50% back at the bookstore. 00:59:11.170 --> 00:59:13.170 align:middle line:84% If you're going to sell it back to your friends, 00:59:13.170 --> 00:59:15.930 align:middle line:84% maybe you could go to your-- well, not your closest friends-- 00:59:15.930 --> 00:59:18.738 align:middle line:84% and say, I'll sell my book to you for only 75% 00:59:18.738 --> 00:59:20.280 align:middle line:84% of the cover price, because if people 00:59:20.280 --> 00:59:22.860 align:middle line:84% want to go to the campus bookstore, 00:59:22.860 --> 00:59:25.248 align:middle line:84% they can buy the used book at 75% of the cover price. 00:59:25.248 --> 00:59:27.540 align:middle line:84% You're not going be able to sell it for more than that. 00:59:27.540 --> 00:59:29.640 align:middle line:84% Just posting a sign in the dorm saying, 00:59:29.640 --> 00:59:31.740 align:middle line:90% I've got a used copy to sell. 00:59:31.740 --> 00:59:36.240 align:middle line:84% But so if we think mu is between 0.4 and 0.75, 00:59:36.240 --> 00:59:40.980 align:middle line:84% they estimate is-- they estimate lambda is 0.61. 00:59:40.980 --> 00:59:44.460 align:middle line:84% And so if you estimate that lambda is 0.61 00:59:44.460 --> 00:59:48.390 align:middle line:84% and you think that mu is in this range, 00:59:48.390 --> 00:59:50.790 align:middle line:84% that's got to tell you that mu is not 0.4. 00:59:50.790 --> 00:59:55.030 align:middle line:84% Mu is bigger than 0.4, and delta is pretty close to 1. 00:59:55.030 --> 00:59:58.630 align:middle line:84% So the estimates are indicating that consumers 00:59:58.630 --> 01:00:02.980 align:middle line:84% must be fairly close to all being forward-looking 01:00:02.980 --> 01:00:07.240 align:middle line:84% and putting a very large weight on that revision probability 01:00:07.240 --> 01:00:11.620 align:middle line:84% because they're reacting a great deal to books 01:00:11.620 --> 01:00:13.280 align:middle line:90% that are about to die. 01:00:13.280 --> 01:00:15.520 align:middle line:84% And they could only react that much 01:00:15.520 --> 01:00:19.900 align:middle line:84% to books that are about to die if they recognize 01:00:19.900 --> 01:00:22.810 align:middle line:84% they're going to die, or there's a high probability of death 01:00:22.810 --> 01:00:24.953 align:middle line:90% and put a lot of weight on that. 01:00:24.953 --> 01:00:26.870 align:middle line:84% So in some sense, we're estimating two things. 01:00:26.870 --> 01:00:29.420 align:middle line:90% We get this-- 01:00:29.420 --> 01:00:33.430 align:middle line:90% I have alpha and alpha delta mu. 01:00:33.430 --> 01:00:36.470 align:middle line:84% So the coefficient on the probability of death 01:00:36.470 --> 01:00:40.680 align:middle line:84% is giving me this, the coefficient on the price 01:00:40.680 --> 01:00:43.035 align:middle line:84% is giving me this, and so then I divide them, 01:00:43.035 --> 01:00:44.660 align:middle line:84% and that's where I get the lambda from. 01:00:44.660 --> 01:00:53.560 align:middle line:90% 01:00:53.560 --> 01:00:55.125 align:middle line:90% Any questions on that? 01:00:55.125 --> 01:00:59.290 align:middle line:90% 01:00:59.290 --> 01:01:01.890 align:middle line:84% So, obviously, as with any empirical paper, 01:01:01.890 --> 01:01:05.550 align:middle line:90% I can think of lots of concerns. 01:01:05.550 --> 01:01:08.010 align:middle line:84% And I think one of the hurdles to getting research done 01:01:08.010 --> 01:01:09.210 align:middle line:84% is you just-- there are always going 01:01:09.210 --> 01:01:10.502 align:middle line:90% to be problems with your paper. 01:01:10.502 --> 01:01:12.638 align:middle line:84% You just go ahead and write them. 01:01:12.638 --> 01:01:14.430 align:middle line:84% In our profession, the people who get ahead 01:01:14.430 --> 01:01:15.630 align:middle line:84% are the people who write the papers, 01:01:15.630 --> 01:01:17.172 align:middle line:84% and then other people criticize them. 01:01:17.172 --> 01:01:18.090 align:middle line:90% That's OK. 01:01:18.090 --> 01:01:20.250 align:middle line:84% Just writing the papers and having 01:01:20.250 --> 01:01:21.750 align:middle line:84% people talk about your paper is what 01:01:21.750 --> 01:01:23.500 align:middle line:84% you have to do to get ahead in this world. 01:01:23.500 --> 01:01:25.800 align:middle line:84% And so you write the paper, there may be flaws, 01:01:25.800 --> 01:01:27.400 align:middle line:84% there are things people can criticize. 01:01:27.400 --> 01:01:30.550 align:middle line:84% But as with any paper, there can be lots of things to criticize, 01:01:30.550 --> 01:01:33.000 align:middle line:84% and that's OK as long as you get it out there. 01:01:33.000 --> 01:01:36.090 align:middle line:84% So what can I criticize about this one? 01:01:36.090 --> 01:01:39.100 align:middle line:84% They have no data on the non-bookstore sales. 01:01:39.100 --> 01:01:41.440 align:middle line:84% They have very good data on the bookstore sales 01:01:41.440 --> 01:01:43.900 align:middle line:84% they get through the bookstores themselves. 01:01:43.900 --> 01:01:46.832 align:middle line:84% They really don't know much about the used market. 01:01:46.832 --> 01:01:49.290 align:middle line:84% They have a couple different treatments of the used market, 01:01:49.290 --> 01:01:53.460 align:middle line:84% but, in some sense, they're kind of making up the used book sales 01:01:53.460 --> 01:01:57.940 align:middle line:90% and just going with that. 01:01:57.940 --> 01:02:00.220 align:middle line:90% Instruments are important. 01:02:00.220 --> 01:02:04.150 align:middle line:84% I think the non-profit publisher instrument is a good instrument. 01:02:04.150 --> 01:02:06.880 align:middle line:84% But let's suppose you haven't fully instrumented 01:02:06.880 --> 01:02:10.570 align:middle line:90% for this thing, the lambda is-- 01:02:10.570 --> 01:02:13.450 align:middle line:84% lambda hat is going to be like alpha delta mu 01:02:13.450 --> 01:02:16.120 align:middle line:90% hat over alpha hat. 01:02:16.120 --> 01:02:18.760 align:middle line:84% If the alpha hat ends up biased toward zero, 01:02:18.760 --> 01:02:21.050 align:middle line:84% you're going to be inflating the lambda hat. 01:02:21.050 --> 01:02:24.820 align:middle line:84% So imperfect job of instrumenting for alpha 01:02:24.820 --> 01:02:27.500 align:middle line:84% is going to make you think that lambda is bigger than it is. 01:02:27.500 --> 01:02:29.590 align:middle line:84% You worry, could that be part of what's going on? 01:02:29.590 --> 01:02:32.260 align:middle line:90% 01:02:32.260 --> 01:02:35.990 align:middle line:84% Distribution of errors-- again, we have that theorem. 01:02:35.990 --> 01:02:38.530 align:middle line:90% If the epsilons are i.i.d. 01:02:38.530 --> 01:02:42.310 align:middle line:84% extreme value-type errors, then log of s over log of s 01:02:42.310 --> 01:02:44.562 align:middle line:90% equals this. 01:02:44.562 --> 01:02:46.270 align:middle line:84% There's a converse of that theorem, which 01:02:46.270 --> 01:02:49.150 align:middle line:84% says if the errors are not i.i.d. extreme value type one, 01:02:49.150 --> 01:02:54.160 align:middle line:84% then log of s new over s0 not equal to x times beta 01:02:54.160 --> 01:02:56.980 align:middle line:84% minus alpha times p minus alpha times p1 minus die. 01:02:56.980 --> 01:03:00.430 align:middle line:84% So obviously, we're relying on this functional form 01:03:00.430 --> 01:03:02.890 align:middle line:90% identification in that sense. 01:03:02.890 --> 01:03:05.410 align:middle line:84% We're also relying on functional form identification 01:03:05.410 --> 01:03:08.860 align:middle line:84% in the second sense, which is the probability that a book will 01:03:08.860 --> 01:03:14.080 align:middle line:84% die is a function of its age and the field it's in 01:03:14.080 --> 01:03:16.060 align:middle line:90% and other things. 01:03:16.060 --> 01:03:20.470 align:middle line:84% So by using that probability of death as an instrument 01:03:20.470 --> 01:03:21.790 align:middle line:90% on the right side-- 01:03:21.790 --> 01:03:25.540 align:middle line:84% so the way the estimation is actually done 01:03:25.540 --> 01:03:29.848 align:middle line:84% is they put 1 minus whether it actually died here 01:03:29.848 --> 01:03:32.140 align:middle line:84% on the right side and then use the probability of death 01:03:32.140 --> 01:03:34.790 align:middle line:84% as an instrument for whether it actually died. 01:03:34.790 --> 01:03:37.565 align:middle line:90% But still, using that-- 01:03:37.565 --> 01:03:39.190 align:middle line:84% you're using as an instrument something 01:03:39.190 --> 01:03:42.405 align:middle line:84% that's a function of variables that are all in your equation. 01:03:42.405 --> 01:03:43.780 align:middle line:84% And so it has to be that somehow, 01:03:43.780 --> 01:03:46.300 align:middle line:84% that nonlinear function of semesters 01:03:46.300 --> 01:03:49.480 align:middle line:84% and whatever is uncorrelated with the errors in the demand 01:03:49.480 --> 01:03:52.900 align:middle line:84% equation, even though you put the variables themselves also 01:03:52.900 --> 01:03:55.420 align:middle line:90% on the right side just linearly. 01:03:55.420 --> 01:03:58.360 align:middle line:84% So there's a use of functional forms to identify that. 01:03:58.360 --> 01:04:00.490 align:middle line:84% I should say, one nice thing that they 01:04:00.490 --> 01:04:03.970 align:middle line:84% did at the back of the paper to try to make it more convincing 01:04:03.970 --> 01:04:07.940 align:middle line:84% is this paper was written before books were sold on Amazon. 01:04:07.940 --> 01:04:09.850 align:middle line:84% They then got some more recent data 01:04:09.850 --> 01:04:12.020 align:middle line:90% of used books sold on Amazon. 01:04:12.020 --> 01:04:15.430 align:middle line:84% And they also find this effect there on the price side 01:04:15.430 --> 01:04:17.560 align:middle line:84% that if you look at, at what price can you 01:04:17.560 --> 01:04:21.760 align:middle line:84% sell your used book on Amazon, it goes down dramatically 01:04:21.760 --> 01:04:23.800 align:middle line:90% in semester 5. 01:04:23.800 --> 01:04:28.330 align:middle line:84% And sometimes, you go on Amazon, and you see the 12th edition is 01:04:28.330 --> 01:04:30.760 align:middle line:84% already available, not yet available to buy, 01:04:30.760 --> 01:04:31.930 align:middle line:90% but it's there. 01:04:31.930 --> 01:04:33.280 align:middle line:90% I own the 11th-- 01:04:33.280 --> 01:04:35.060 align:middle line:84% I'm being asked to buy the 11th edition. 01:04:35.060 --> 01:04:39.800 align:middle line:84% The 12th edition is listed on Amazon as available October 15. 01:04:39.800 --> 01:04:42.670 align:middle line:84% In that situation, price is really-- 01:04:42.670 --> 01:04:44.820 align:middle line:84% price of used books really, really drop. 01:04:44.820 --> 01:04:51.790 align:middle line:90% 01:04:51.790 --> 01:04:53.715 align:middle line:84% Any further questions on this paper? 01:04:53.715 --> 01:04:59.910 align:middle line:90% 01:04:59.910 --> 01:05:02.995 align:middle line:84% So let me go to final paper-- and again, this is a-- 01:05:02.995 --> 01:05:04.990 align:middle line:84% the final paper I thought I'd talk about is-- 01:05:04.990 --> 01:05:06.290 align:middle line:90% it's 10 years old. 01:05:06.290 --> 01:05:07.130 align:middle line:90% It's 10 years old. 01:05:07.130 --> 01:05:09.170 align:middle line:84% It almost seems-- or 12 years old. 01:05:09.170 --> 01:05:12.010 align:middle line:84% It almost seems more topical than it did when it was written. 01:05:12.010 --> 01:05:14.620 align:middle line:90% So this is about media slant. 01:05:14.620 --> 01:05:23.290 align:middle line:84% This one, the question is, why do we see biased news media? 01:05:23.290 --> 01:05:25.960 align:middle line:84% One reason why you might think we see biased news media 01:05:25.960 --> 01:05:29.590 align:middle line:84% is maybe Rupert Murdoch is very conservative 01:05:29.590 --> 01:05:33.410 align:middle line:84% and wants to spread conservative values across the United States. 01:05:33.410 --> 01:05:38.620 align:middle line:84% Or maybe Newsmax is secretly controlled by Russia, 01:05:38.620 --> 01:05:41.110 align:middle line:84% and they're out to display misinformation that will 01:05:41.110 --> 01:05:43.220 align:middle line:90% destroy American democracy. 01:05:43.220 --> 01:05:47.200 align:middle line:84% But another explanation would simply be that running Fox News 01:05:47.200 --> 01:05:48.850 align:middle line:84% and getting people angry about things 01:05:48.850 --> 01:05:50.417 align:middle line:84% is a very, very profitable business. 01:05:50.417 --> 01:05:51.250 align:middle line:90% And you don't even-- 01:05:51.250 --> 01:05:52.833 align:middle line:84% Rupert Murdoch doesn't even have to be 01:05:52.833 --> 01:05:55.030 align:middle line:84% conservative to want to do this because he's just 01:05:55.030 --> 01:05:58.345 align:middle line:84% making lots and lots of money because people watch Fox News. 01:05:58.345 --> 01:06:01.542 align:middle line:90% 01:06:01.542 --> 01:06:03.500 align:middle line:84% In this case, I said in Chevalier and Goolsbee, 01:06:03.500 --> 01:06:06.080 align:middle line:84% it was like there was a public policy question about books 01:06:06.080 --> 01:06:08.330 align:middle line:84% and why they're so expensive and why they get revised. 01:06:08.330 --> 01:06:10.695 align:middle line:84% But then there's an understanding the theory better. 01:06:10.695 --> 01:06:13.070 align:middle line:84% Here, I think it's less of an understanding of the theory 01:06:13.070 --> 01:06:15.170 align:middle line:84% better than just recognizing the theory, which 01:06:15.170 --> 01:06:17.300 align:middle line:84% is the theory is the optimal product quality 01:06:17.300 --> 01:06:20.810 align:middle line:84% theorem from monopoly pricing, that we think monopolists ought 01:06:20.810 --> 01:06:22.970 align:middle line:84% to choose the quality of their product 01:06:22.970 --> 01:06:25.170 align:middle line:84% that is optimal to maximize profits. 01:06:25.170 --> 01:06:27.680 align:middle line:84% And so the question here is, does that optimal quality 01:06:27.680 --> 01:06:30.560 align:middle line:84% theorem apply to help us understand Fox News, 01:06:30.560 --> 01:06:34.820 align:middle line:84% and does it look like Fox News is profit maximizing, 01:06:34.820 --> 01:06:37.850 align:middle line:84% and do we understand things are profit maximizing? 01:06:37.850 --> 01:06:40.890 align:middle line:84% Now, turns out this paper is not about Fox News. 01:06:40.890 --> 01:06:42.560 align:middle line:90% This is about newspapers. 01:06:42.560 --> 01:06:47.340 align:middle line:90% And again, things do age. 01:06:47.340 --> 01:06:49.230 align:middle line:84% There was a time in the United States 01:06:49.230 --> 01:06:52.500 align:middle line:84% not so long ago when people read newspapers. 01:06:52.500 --> 01:06:55.650 align:middle line:84% Newspapers were things printed on physical paper, delivered 01:06:55.650 --> 01:06:57.772 align:middle line:90% to people's houses by carriers. 01:06:57.772 --> 01:07:00.480 align:middle line:84% And then you would get up in the morning, pick up your newspaper, 01:07:00.480 --> 01:07:04.380 align:middle line:84% and read it while drinking your coffee and eating breakfast. 01:07:04.380 --> 01:07:07.540 align:middle line:84% There were many, many local newspapers in the United States. 01:07:07.540 --> 01:07:09.580 align:middle line:84% Every city had multiple newspapers. 01:07:09.580 --> 01:07:11.940 align:middle line:84% So in Boston, instead of just having the Boston Globe, 01:07:11.940 --> 01:07:14.232 align:middle line:84% we had the Boston Globe and something called the Boston 01:07:14.232 --> 01:07:14.820 align:middle line:90% Herald. 01:07:14.820 --> 01:07:16.530 align:middle line:84% Boston Herald was smaller and easier 01:07:16.530 --> 01:07:18.930 align:middle line:84% to read while riding on the subway. 01:07:18.930 --> 01:07:22.090 align:middle line:84% But many towns across the United States had newspapers. 01:07:22.090 --> 01:07:24.330 align:middle line:84% Those newspapers would convey both local news 01:07:24.330 --> 01:07:25.520 align:middle line:90% and national news. 01:07:25.520 --> 01:07:30.000 align:middle line:90% 01:07:30.000 --> 01:07:35.650 align:middle line:84% So let's see, motivation-- so anyway, media slant 01:07:35.650 --> 01:07:37.640 align:middle line:90% remains a very big topic. 01:07:37.640 --> 01:07:38.680 align:middle line:90% I should say that-- 01:07:38.680 --> 01:07:41.390 align:middle line:84% I mentioned Fox News, I think another very interesting topic, 01:07:41.390 --> 01:07:44.750 align:middle line:84% there's a company flying more under the radar, Sinclair Media. 01:07:44.750 --> 01:07:48.350 align:middle line:84% Sinclair Media is basically the Fox News of local television. 01:07:48.350 --> 01:07:52.600 align:middle line:84% So Sinclair has bought hundreds and hundreds of local stations 01:07:52.600 --> 01:07:55.780 align:middle line:84% across the United States and made the 10:00 and 11:00 news 01:07:55.780 --> 01:07:59.220 align:middle line:84% that you receive free over the air more Fox News-like. 01:07:59.220 --> 01:08:02.200 align:middle line:90% 01:08:02.200 --> 01:08:04.700 align:middle line:84% Anyway, what Gentzkow and Shapiro 01:08:04.700 --> 01:08:07.060 align:middle line:84% were trying to do is ideally, they 01:08:07.060 --> 01:08:10.510 align:middle line:84% would like to have this two-step investigation of understanding 01:08:10.510 --> 01:08:12.970 align:middle line:90% motivations for biased news. 01:08:12.970 --> 01:08:17.600 align:middle line:84% Step one is estimate the slant that would be profit maximizing. 01:08:17.600 --> 01:08:21.490 align:middle line:84% And then step two is compare the actual and profit-maximizing 01:08:21.490 --> 01:08:24.199 align:middle line:90% slant and see whether there-- 01:08:24.199 --> 01:08:26.103 align:middle line:84% are there differences between the two? 01:08:26.103 --> 01:08:27.520 align:middle line:84% If there are differences, are they 01:08:27.520 --> 01:08:31.370 align:middle line:84% related to the ideology of the person who owns the newspaper? 01:08:31.370 --> 01:08:34.310 align:middle line:90% 01:08:34.310 --> 01:08:39.649 align:middle line:84% I think the paper falls a little short of the self-description, 01:08:39.649 --> 01:08:41.698 align:middle line:84% or it's a little heroic in the assumptions 01:08:41.698 --> 01:08:43.490 align:middle line:84% that you need to think that it's estimating 01:08:43.490 --> 01:08:45.710 align:middle line:90% profit-maximizing slants. 01:08:45.710 --> 01:08:50.880 align:middle line:84% But it has been very influential for the questions it raised 01:08:50.880 --> 01:08:54.090 align:middle line:84% and also, actually, for the ways it's quantifying media bias. 01:08:54.090 --> 01:08:56.250 align:middle line:84% And techniques developed here get 01:08:56.250 --> 01:08:58.999 align:middle line:84% used in many other papers on media bias. 01:08:58.999 --> 01:09:02.210 align:middle line:90% 01:09:02.210 --> 01:09:03.859 align:middle line:90% OK, so quantifying media slant-- 01:09:03.859 --> 01:09:06.470 align:middle line:90% 01:09:06.470 --> 01:09:08.967 align:middle line:84% this was really one of the first papers to use-- 01:09:08.967 --> 01:09:10.550 align:middle line:84% you wouldn't even these days call them 01:09:10.550 --> 01:09:12.680 align:middle line:84% machine learning techniques, but automated text 01:09:12.680 --> 01:09:15.590 align:middle line:90% processing as a way to get data. 01:09:15.590 --> 01:09:18.500 align:middle line:84% And so what they're trying to do is they're trying to figure out, 01:09:18.500 --> 01:09:23.800 align:middle line:84% is a newspaper right skewed or is a newspaper left skewed? 01:09:23.800 --> 01:09:26.680 align:middle line:84% How do you make a newspaper skewed one way or the other? 01:09:26.680 --> 01:09:29.490 align:middle line:84% Well, one thing you do is just with the words you use. 01:09:29.490 --> 01:09:32.180 align:middle line:84% And in the early 2000s, they give examples like, 01:09:32.180 --> 01:09:35.569 align:middle line:84% do you call something the Iraq War or the War on Terror? 01:09:35.569 --> 01:09:37.760 align:middle line:84% And generally, people who favored the Iraq War 01:09:37.760 --> 01:09:39.130 align:middle line:90% would call it the War on Terror. 01:09:39.130 --> 01:09:43.100 align:middle line:84% People who opposed the War in Iraq would call it the Iraq War. 01:09:43.100 --> 01:09:46.729 align:middle line:84% Estate tax and death tax is a very nice linguistic thing. 01:09:46.729 --> 01:09:50.240 align:middle line:84% In the United States, if you have less than $5 million 01:09:50.240 --> 01:09:53.720 align:middle line:84% when you die, you can pass it on to your heirs tax free. 01:09:53.720 --> 01:09:55.822 align:middle line:84% If you have billions of dollars, it 01:09:55.822 --> 01:09:58.280 align:middle line:84% seems like you can often pass it on to your heirs tax free, 01:09:58.280 --> 01:10:00.660 align:middle line:84% never having paid the capital gains taxes either. 01:10:00.660 --> 01:10:03.060 align:middle line:84% But there are taxes on the books that you're 01:10:03.060 --> 01:10:06.930 align:middle line:84% supposed to pay when you pass your billions of dollars 01:10:06.930 --> 01:10:09.600 align:middle line:90% on to your children. 01:10:09.600 --> 01:10:13.530 align:middle line:84% But anyway, the Republicans tend to refer to it as the death tax, 01:10:13.530 --> 01:10:16.283 align:middle line:84% because who wants to pile on someone's misery 01:10:16.283 --> 01:10:17.700 align:middle line:84% when someone has died and tax them 01:10:17.700 --> 01:10:19.860 align:middle line:90% for it, versus the estate tax. 01:10:19.860 --> 01:10:21.550 align:middle line:84% Everyone can agree on taxing estates. 01:10:21.550 --> 01:10:23.280 align:middle line:84% If you have wealthy people, estate-- 01:10:23.280 --> 01:10:25.062 align:middle line:84% wealthy estates should pay taxes. 01:10:25.062 --> 01:10:26.520 align:middle line:84% So there's a question of what words 01:10:26.520 --> 01:10:28.210 align:middle line:90% do you use in your newspaper? 01:10:28.210 --> 01:10:29.760 align:middle line:90% But then much more-- 01:10:29.760 --> 01:10:31.860 align:middle line:84% if you look at what Fox News does, some of it 01:10:31.860 --> 01:10:34.658 align:middle line:84% may be outright lying about voting machine tampering 01:10:34.658 --> 01:10:36.450 align:middle line:84% or something like that to get people angry. 01:10:36.450 --> 01:10:38.085 align:middle line:84% But much more, it's a question of, 01:10:38.085 --> 01:10:39.840 align:middle line:84% do you have a lot of news stories 01:10:39.840 --> 01:10:43.680 align:middle line:84% about undocumented immigrants assaulting and killing people? 01:10:43.680 --> 01:10:48.460 align:middle line:84% Or do you have a lot of stories about people 01:10:48.460 --> 01:10:51.160 align:middle line:84% invading the US capitol and beating police officers 01:10:51.160 --> 01:10:52.120 align:middle line:90% or whatever? 01:10:52.120 --> 01:10:57.090 align:middle line:84% So in some sense, the extent to which you 01:10:57.090 --> 01:10:59.490 align:middle line:84% talk about tax breaks being given to corporations 01:10:59.490 --> 01:11:01.590 align:middle line:84% or the extent you talk about illegal immigrants 01:11:01.590 --> 01:11:04.980 align:middle line:84% committing crimes is going to be one way in which you can 01:11:04.980 --> 01:11:06.720 align:middle line:84% slant your news coverage and change 01:11:06.720 --> 01:11:09.330 align:middle line:84% political opinions of the people who live in the area 01:11:09.330 --> 01:11:11.070 align:middle line:90% that you're serving. 01:11:11.070 --> 01:11:15.600 align:middle line:84% And so what they try to do is-- the question is, how do you 01:11:15.600 --> 01:11:19.560 align:middle line:84% pick out a Democratic from a Republican issue or a Democratic 01:11:19.560 --> 01:11:21.785 align:middle line:84% from a Republican talking point or Democratic 01:11:21.785 --> 01:11:23.380 align:middle line:90% from Republican phrasing? 01:11:23.380 --> 01:11:26.880 align:middle line:84% And what they do is they go through the full text 01:11:26.880 --> 01:11:29.723 align:middle line:84% of all speeches given on the floor 01:11:29.723 --> 01:11:31.140 align:middle line:84% of the US House of Representatives 01:11:31.140 --> 01:11:32.340 align:middle line:90% and the House of Senate. 01:11:32.340 --> 01:11:34.710 align:middle line:84% And they pick up two- and three-word phrases 01:11:34.710 --> 01:11:38.190 align:middle line:84% that are being used by Democrats and Republicans. 01:11:38.190 --> 01:11:42.360 align:middle line:84% And they actually classify each congressperson as a 01:11:42.360 --> 01:11:45.090 align:middle line:84% where they are on a continuum spectrum 01:11:45.090 --> 01:11:49.050 align:middle line:84% by what's the vote share for George W. 01:11:49.050 --> 01:11:51.820 align:middle line:84% Bush in their congressional district. 01:11:51.820 --> 01:11:53.700 align:middle line:84% And so they use this for what is being 01:11:53.700 --> 01:11:55.570 align:middle line:84% spoken by far left people, what's being 01:11:55.570 --> 01:11:57.290 align:middle line:90% spoken by far right people. 01:11:57.290 --> 01:11:58.990 align:middle line:90% And so they identify basically-- 01:11:58.990 --> 01:12:05.200 align:middle line:90% 01:12:05.200 --> 01:12:08.980 align:middle line:84% they select 1,000 two- and three-word phrases that are used 01:12:08.980 --> 01:12:11.260 align:middle line:84% significantly different frequencies by people 01:12:11.260 --> 01:12:13.960 align:middle line:84% on the left versus the right of the political spectrum. 01:12:13.960 --> 01:12:17.200 align:middle line:84% And then for each newspaper, they 01:12:17.200 --> 01:12:19.330 align:middle line:84% regress the excess frequency with which 01:12:19.330 --> 01:12:23.230 align:middle line:84% it uses each phrase on that phrase's Republican-ness. 01:12:23.230 --> 01:12:28.630 align:middle line:84% And they define the slant of the newspaper as the coefficient 01:12:28.630 --> 01:12:32.050 align:middle line:90% on that variable. 01:12:32.050 --> 01:12:35.800 align:middle line:84% So if it's really using almost exclusively right 01:12:35.800 --> 01:12:38.500 align:middle line:84% Republican phrases, it's going to be a very high coefficient. 01:12:38.500 --> 01:12:41.000 align:middle line:84% If it's using mostly Democratic phrases-- 01:12:41.000 --> 01:12:45.820 align:middle line:84% so if the distribution looks like this, 01:12:45.820 --> 01:12:47.882 align:middle line:84% it's going to be a very left-skewed newspaper. 01:12:47.882 --> 01:12:49.840 align:middle line:84% If the distribution of phrases looks like that, 01:12:49.840 --> 01:12:51.798 align:middle line:84% it's going to be a very right-skewed newspaper. 01:12:51.798 --> 01:12:55.230 align:middle line:84% And they measure the skew by the slope of those estimated lines. 01:12:55.230 --> 01:12:58.470 align:middle line:90% 01:12:58.470 --> 01:13:02.070 align:middle line:84% Now, obviously, part of that in these examples sound 01:13:02.070 --> 01:13:04.257 align:middle line:84% like things where you're doing media slant. 01:13:04.257 --> 01:13:06.840 align:middle line:84% There are also going to be all kinds of other things picked up 01:13:06.840 --> 01:13:07.340 align:middle line:90% in this. 01:13:07.340 --> 01:13:10.290 align:middle line:84% So, for instance, some Congresspeople 01:13:10.290 --> 01:13:12.540 align:middle line:84% talk about agriculture policy, and some 01:13:12.540 --> 01:13:17.100 align:middle line:84% talk about guns in their urban neighborhoods. 01:13:17.100 --> 01:13:19.290 align:middle line:84% Talking about agriculture policy may just 01:13:19.290 --> 01:13:21.480 align:middle line:84% mean that you're talking about agricultural policy. 01:13:21.480 --> 01:13:24.720 align:middle line:84% But generally, Republicans tend to represent 01:13:24.720 --> 01:13:27.420 align:middle line:84% agricultural states, and so you may get corn 01:13:27.420 --> 01:13:30.960 align:middle line:84% as being a Republican phrase when it's not-- 01:13:30.960 --> 01:13:36.810 align:middle line:84% talking about corn prices is not a political statement. 01:13:36.810 --> 01:13:39.190 align:middle line:84% They didn't pick the phrases to have political content. 01:13:39.190 --> 01:13:41.200 align:middle line:84% They just picked phrases automatically. 01:13:41.200 --> 01:13:44.340 align:middle line:84% And so there's going to be some contamination of political slant 01:13:44.340 --> 01:13:45.526 align:middle line:90% with other differences. 01:13:45.526 --> 01:13:50.780 align:middle line:90% 01:13:50.780 --> 01:13:52.940 align:middle line:84% Consumer demand, again, is going to be 01:13:52.940 --> 01:13:55.190 align:middle line:90% derived from utility function. 01:13:55.190 --> 01:14:01.010 align:middle line:84% We assume that the utility that consumer I in ZIP code Z 01:14:01.010 --> 01:14:04.520 align:middle line:84% gets from buying newspaper N is some function 01:14:04.520 --> 01:14:12.430 align:middle line:84% of the average quality of the newspaper minus disutility 01:14:12.430 --> 01:14:14.980 align:middle line:84% that the consumer has from consuming something that's 01:14:14.980 --> 01:14:17.700 align:middle line:90% different from the ideal slant. 01:14:17.700 --> 01:14:21.450 align:middle line:84% So Iz is the ideal slant of all people in ZIP code 01:14:21.450 --> 01:14:25.570 align:middle line:84% Z. yn is the actual slant of newspaper n. 01:14:25.570 --> 01:14:28.990 align:middle line:84% So as the slant of the newspaper differs 01:14:28.990 --> 01:14:31.730 align:middle line:84% from what's ideal for people in that ZIP code, 01:14:31.730 --> 01:14:34.960 align:middle line:84% utility of people in that ZIP code goes down. 01:14:34.960 --> 01:14:36.900 align:middle line:84% And then there's a random error term. 01:14:36.900 --> 01:14:41.610 align:middle line:90% 01:14:41.610 --> 01:14:44.160 align:middle line:84% They assume that the utility-maximizing slant 01:14:44.160 --> 01:14:49.770 align:middle line:84% in ZIP code z is just alpha plus beta times rz, where 01:14:49.770 --> 01:14:53.580 align:middle line:84% rz is the fraction of people in ZIP code z who are Republican. 01:14:53.580 --> 01:14:58.900 align:middle line:84% So it's a very specific assumption in several ways. 01:14:58.900 --> 01:15:01.280 align:middle line:90% So there's no heterogeneity. 01:15:01.280 --> 01:15:03.510 align:middle line:84% All people in ZIP code z have exactly 01:15:03.510 --> 01:15:06.790 align:middle line:90% the same preferred slant. 01:15:06.790 --> 01:15:11.190 align:middle line:84% And there is no error in this equation. 01:15:11.190 --> 01:15:15.340 align:middle line:84% All people in-- preferred slant in ZIP code z 01:15:15.340 --> 01:15:18.250 align:middle line:84% is this linear function of the fraction Republican, 01:15:18.250 --> 01:15:20.560 align:middle line:84% with no heterogeneity across ZIP codes. 01:15:20.560 --> 01:15:23.825 align:middle line:84% There are no ZIP codes where a left slant is better 01:15:23.825 --> 01:15:26.325 align:middle line:84% than you would think, given how many Republicans live there. 01:15:26.325 --> 01:15:30.330 align:middle line:90% 01:15:30.330 --> 01:15:33.810 align:middle line:84% Once again, they also assume now that these epsilon izn's 01:15:33.810 --> 01:15:34.500 align:middle line:90% are i.i.d. 01:15:34.500 --> 01:15:37.740 align:middle line:84% extreme value, and then they make this magical transformation 01:15:37.740 --> 01:15:40.470 align:middle line:84% from sales of the newspaper to utility 01:15:40.470 --> 01:15:42.280 align:middle line:90% of the purchasing consumers. 01:15:42.280 --> 01:15:48.720 align:middle line:84% And they get that the log of the market share of purchasers 01:15:48.720 --> 01:15:53.430 align:middle line:84% versus nonpurchasers is the average utility minus gamma 01:15:53.430 --> 01:15:56.130 align:middle line:84% times the difference between the newspaper slant 01:15:56.130 --> 01:16:03.270 align:middle line:84% and that linear function of the fraction Republican in the ZIP 01:16:03.270 --> 01:16:04.050 align:middle line:90% code. 01:16:04.050 --> 01:16:07.680 align:middle line:84% And obviously, we can just expand this quadratic. 01:16:07.680 --> 01:16:09.390 align:middle line:84% And when you expand this quadratic, 01:16:09.390 --> 01:16:11.593 align:middle line:84% you're just going to get things like-- 01:16:11.593 --> 01:16:13.260 align:middle line:84% you're going to get y on the right side. 01:16:13.260 --> 01:16:14.760 align:middle line:90% You're going to get y squared. 01:16:14.760 --> 01:16:16.468 align:middle line:84% You're going to get r squared, and you're 01:16:16.468 --> 01:16:17.730 align:middle line:90% going to get y times r. 01:16:17.730 --> 01:16:19.647 align:middle line:84% So you're just going to have this thing that's 01:16:19.647 --> 01:16:23.610 align:middle line:84% this right-hand side term that's got y's and r's and y squareds 01:16:23.610 --> 01:16:26.800 align:middle line:90% and r squareds and y times r. 01:16:26.800 --> 01:16:33.500 align:middle line:84% And in this model, if this model is correct, 01:16:33.500 --> 01:16:37.970 align:middle line:84% the utility-maximizing slant is alpha plus beta rz. 01:16:37.970 --> 01:16:42.910 align:middle line:84% And we can estimate how the utility-maximizing slant varies 01:16:42.910 --> 01:16:46.390 align:middle line:84% with rz by just looking at the coefficient of-- the coefficient 01:16:46.390 --> 01:16:50.403 align:middle line:90% of yn times rz is going to be-- 01:16:50.403 --> 01:16:55.420 align:middle line:90% 01:16:55.420 --> 01:17:01.510 align:middle line:84% it's going to be gamma minus 2 times gamma times beta. 01:17:01.510 --> 01:17:04.605 align:middle line:84% So anyway, we're going to get that-- 01:17:04.605 --> 01:17:08.160 align:middle line:84% yeah, the coefficient on yn times 01:17:08.160 --> 01:17:11.208 align:middle line:84% rz is going to be minus 2 times gamma times beta. 01:17:11.208 --> 01:17:12.750 align:middle line:84% And so we're going to be estimating-- 01:17:12.750 --> 01:17:15.750 align:middle line:84% we're going to get gamma times-- identify gamma times 01:17:15.750 --> 01:17:18.900 align:middle line:84% beta by just looking at the coefficient of this 01:17:18.900 --> 01:17:22.980 align:middle line:84% when I regress log of share of purchases 01:17:22.980 --> 01:17:26.410 align:middle line:84% of the newspaper on the right-hand side variables. 01:17:26.410 --> 01:17:32.340 align:middle line:90% 01:17:32.340 --> 01:17:35.160 align:middle line:90% What did they find? 01:17:35.160 --> 01:17:37.290 align:middle line:84% The clear finding is that this ZIP 01:17:37.290 --> 01:17:44.980 align:middle line:84% share Republican times slant has a clear, positive, significant 01:17:44.980 --> 01:17:45.890 align:middle line:90% coefficient. 01:17:45.890 --> 01:17:51.070 align:middle line:84% So this is evidence that the ideal slant 01:17:51.070 --> 01:17:56.140 align:middle line:84% is to put more Republican news in a Republican ZIP code 01:17:56.140 --> 01:18:00.160 align:middle line:90% from a profit-maximizing sense. 01:18:00.160 --> 01:18:04.910 align:middle line:84% Let me say that my view on this is-- 01:18:04.910 --> 01:18:06.470 align:middle line:84% it's a very, very strong assumption 01:18:06.470 --> 01:18:08.512 align:middle line:84% to assume that there's a profit-maximizing slant, 01:18:08.512 --> 01:18:11.480 align:middle line:84% and it doesn't differ across consumers within a ZIP code, 01:18:11.480 --> 01:18:14.900 align:middle line:84% and it has no error term across ZIP codes. 01:18:14.900 --> 01:18:18.473 align:middle line:84% I find it convincing evidence that in more 01:18:18.473 --> 01:18:20.390 align:middle line:84% right-leaning areas, you do better with a more 01:18:20.390 --> 01:18:22.950 align:middle line:90% right-leaning newspaper. 01:18:22.950 --> 01:18:27.350 align:middle line:84% The fact that this slant is the profit-maximizing slant-- 01:18:27.350 --> 01:18:31.020 align:middle line:84% that relies on those no errors, homogeneity assumptions. 01:18:31.020 --> 01:18:32.900 align:middle line:84% So it takes pretty heroic assumptions 01:18:32.900 --> 01:18:35.690 align:middle line:84% to say that this really is the profit-maximizing slant. 01:18:35.690 --> 01:18:37.760 align:middle line:84% But I do think it's pretty clear evidence 01:18:37.760 --> 01:18:40.580 align:middle line:84% that right-slanted news does sell better 01:18:40.580 --> 01:18:43.310 align:middle line:84% in right-slanted areas, which is fundamentally-- 01:18:43.310 --> 01:18:46.940 align:middle line:84% we're regressing log of q on the share Republican times 01:18:46.940 --> 01:18:48.950 align:middle line:84% slant interaction and getting a positive effect, 01:18:48.950 --> 01:18:51.860 align:middle line:84% saying in more Republican areas, more right-slanted news 01:18:51.860 --> 01:18:52.945 align:middle line:90% is selling more copies. 01:18:52.945 --> 01:18:58.940 align:middle line:90% 01:18:58.940 --> 01:19:05.760 align:middle line:84% And then is slant profit maximizing? 01:19:05.760 --> 01:19:08.100 align:middle line:90% They show this graph. 01:19:08.100 --> 01:19:11.220 align:middle line:84% And this graph is supposed to show you, 01:19:11.220 --> 01:19:15.150 align:middle line:84% are newspapers more right slanted in more 01:19:15.150 --> 01:19:16.270 align:middle line:90% Republican areas? 01:19:16.270 --> 01:19:19.290 align:middle line:84% And the answer, unsurprisingly, is yes. 01:19:19.290 --> 01:19:22.530 align:middle line:90% So we get this positive slope. 01:19:22.530 --> 01:19:25.020 align:middle line:84% The more Republican is the area in which your newspaper is 01:19:25.020 --> 01:19:28.680 align:middle line:84% being sold, the more right you do make 01:19:28.680 --> 01:19:31.710 align:middle line:90% the slant of your newspaper. 01:19:31.710 --> 01:19:35.170 align:middle line:84% The demand estimates aren't separately 01:19:35.170 --> 01:19:37.540 align:middle line:84% identifying gamma and beta, so it's a little hard 01:19:37.540 --> 01:19:40.450 align:middle line:84% to say whether this slope is the slope we would expect 01:19:40.450 --> 01:19:44.093 align:middle line:90% given the profit maximization. 01:19:44.093 --> 01:19:45.510 align:middle line:84% So then the final thing they do is 01:19:45.510 --> 01:19:48.040 align:middle line:90% they say, well, let's look at-- 01:19:48.040 --> 01:19:52.330 align:middle line:84% regress the slant on the share Republican in the market 01:19:52.330 --> 01:19:58.443 align:middle line:84% and then ownership fixed effects to see whether it looks like-- 01:19:58.443 --> 01:19:59.360 align:middle line:90% they do this two ways. 01:19:59.360 --> 01:20:01.540 align:middle line:84% One is they look at owner political contributions 01:20:01.540 --> 01:20:04.100 align:middle line:84% and put owner political contributions on the right side. 01:20:04.100 --> 01:20:06.050 align:middle line:84% For some reason, it's not in the same table. 01:20:06.050 --> 01:20:08.140 align:middle line:84% And the other thing they just do is put in ownership fixed 01:20:08.140 --> 01:20:09.515 align:middle line:84% effects and see whether the owner 01:20:09.515 --> 01:20:13.180 align:middle line:84% fixed effects increase the r squared of the regression. 01:20:13.180 --> 01:20:15.730 align:middle line:84% And if you put the owner fixed effects in, 01:20:15.730 --> 01:20:18.650 align:middle line:84% it does make a big increase in the r squared of the regression. 01:20:18.650 --> 01:20:22.177 align:middle line:84% It more than doubles the r squared from 0.18 to 0.44. 01:20:22.177 --> 01:20:24.760 align:middle line:84% But obviously, you're putting in an awful lot of fixed effects 01:20:24.760 --> 01:20:26.000 align:middle line:90% on the right side. 01:20:26.000 --> 01:20:30.520 align:middle line:84% And so what they note is that if you put in state fixed effects 01:20:30.520 --> 01:20:33.310 align:middle line:84% instead of owner fixed effects and then put in state fixed 01:20:33.310 --> 01:20:36.280 align:middle line:84% effects and then an ownership effect, 01:20:36.280 --> 01:20:39.490 align:middle line:84% then the ownership effect is not significant. 01:20:39.490 --> 01:20:44.080 align:middle line:84% So while the ownership fixed effects do increase the fit, 01:20:44.080 --> 01:20:46.637 align:middle line:84% you put in owner fixed effects and state fixed effects, 01:20:46.637 --> 01:20:49.220 align:middle line:84% suddenly you can't tell whether the owner effects are actually 01:20:49.220 --> 01:20:50.900 align:middle line:90% there or not at all. 01:20:50.900 --> 01:20:54.740 align:middle line:84% So they take this to argue that it 01:20:54.740 --> 01:20:57.500 align:middle line:84% looks like there isn't much evidence that owner slant-- 01:20:57.500 --> 01:21:01.110 align:middle line:84% it looks like the "Rupert Murdoch is just making money" 01:21:01.110 --> 01:21:05.030 align:middle line:84% story works pretty well here, and it's hard to say that 01:21:05.030 --> 01:21:07.690 align:middle line:84% there's a lot of media bias other than that. 01:21:07.690 --> 01:21:09.440 align:middle line:84% They also put in-- they do a version where 01:21:09.440 --> 01:21:14.810 align:middle line:84% they put in owners' political contributions on the right side. 01:21:14.810 --> 01:21:17.840 align:middle line:84% And it's like, well, this share Republican in the ZIP code 01:21:17.840 --> 01:21:19.640 align:middle line:84% explains 20% of the variance, it's 01:21:19.640 --> 01:21:22.220 align:middle line:84% like the ownership political contributions explain 01:21:22.220 --> 01:21:23.675 align:middle line:90% 5% of the variance. 01:21:23.675 --> 01:21:26.300 align:middle line:84% So it looks like the ownership political contributions are just 01:21:26.300 --> 01:21:29.240 align:middle line:84% not a powerful explanatory variable compared 01:21:29.240 --> 01:21:33.100 align:middle line:84% to the Republican-ness of the ZIP code that they're serving. 01:21:33.100 --> 01:21:51.000 align:middle line:90%