WEBVTT 00:00:00.000 --> 00:00:01.960 align:middle line:90% [SQUEAKING] 00:00:01.960 --> 00:00:03.430 align:middle line:90% [RUSTLING] 00:00:03.430 --> 00:00:04.900 align:middle line:90% [CLICKING] 00:00:04.900 --> 00:00:18.160 align:middle line:90% 00:00:18.160 --> 00:00:18.930 align:middle line:90% GLENN ELLISON: OK. 00:00:18.930 --> 00:00:21.200 align:middle line:90% Let me go ahead and get started. 00:00:21.200 --> 00:00:23.060 align:middle line:84% So if you remember, last class, I 00:00:23.060 --> 00:00:27.020 align:middle line:84% talked about theory of price discrimination under monopoly 00:00:27.020 --> 00:00:29.930 align:middle line:84% and did first, second, and third degree discrimination. 00:00:29.930 --> 00:00:32.240 align:middle line:84% Today, I'm going to talk about empirical work on price 00:00:32.240 --> 00:00:33.420 align:middle line:90% discrimination. 00:00:33.420 --> 00:00:37.280 align:middle line:84% And just as a reminder, as I said previously, 00:00:37.280 --> 00:00:39.920 align:middle line:84% all the models I talked about in last class 00:00:39.920 --> 00:00:42.330 align:middle line:90% are 100% mathematically correct. 00:00:42.330 --> 00:00:45.240 align:middle line:84% There's nothing we need to test about the theory itself. 00:00:45.240 --> 00:00:48.950 align:middle line:84% But many things that empirical work can do for us. 00:00:48.950 --> 00:00:51.440 align:middle line:84% In some sense, it's asking whether the assumptions are 00:00:51.440 --> 00:00:53.750 align:middle line:84% correct and whether things are working like we think 00:00:53.750 --> 00:00:56.167 align:middle line:84% or whether there should be other assumptions in the model. 00:00:56.167 --> 00:00:58.610 align:middle line:84% So questions, examples we get help with. 00:00:58.610 --> 00:01:01.730 align:middle line:90% Do firms price discriminate? 00:01:01.730 --> 00:01:04.970 align:middle line:84% Is the way that firms are pricing consistent with what 00:01:04.970 --> 00:01:08.990 align:middle line:84% the models say about how prices would be related to elasticities 00:01:08.990 --> 00:01:14.040 align:middle line:84% and other things if firms were profit maximizing? 00:01:14.040 --> 00:01:16.348 align:middle line:90% Magnitudes are important. 00:01:16.348 --> 00:01:18.640 align:middle line:84% Literature makes a big deal about price discrimination. 00:01:18.640 --> 00:01:20.350 align:middle line:84% How important would it be in practice? 00:01:20.350 --> 00:01:23.130 align:middle line:84% How large would the profit gains be? 00:01:23.130 --> 00:01:26.640 align:middle line:84% Would firms really be incentivized to do it? 00:01:26.640 --> 00:01:30.330 align:middle line:84% I discussed Bergemann, Brooks, Morris, whether consumer surplus 00:01:30.330 --> 00:01:32.640 align:middle line:84% goes up or down with price discrimination really 00:01:32.640 --> 00:01:35.260 align:middle line:84% depends on what kind of information the firms have. 00:01:35.260 --> 00:01:37.060 align:middle line:84% What do those look like in practice? 00:01:37.060 --> 00:01:38.640 align:middle line:84% And if you are going to discriminate, 00:01:38.640 --> 00:01:41.410 align:middle line:84% what types of information do we need to discriminate? 00:01:41.410 --> 00:01:44.020 align:middle line:84% And I think this is something that's-- 00:01:44.020 --> 00:01:47.640 align:middle line:84% as I talked about, the interest in online price discrimination 00:01:47.640 --> 00:01:49.620 align:middle line:84% has really increased attention to the field 00:01:49.620 --> 00:01:52.260 align:middle line:84% because it used to be discrimination only occurred 00:01:52.260 --> 00:01:54.540 align:middle line:84% on expensive items where you could pay somebody 00:01:54.540 --> 00:01:56.740 align:middle line:84% to try to discriminate among consumers. 00:01:56.740 --> 00:01:58.948 align:middle line:84% Now, in the online world, where you know so much more 00:01:58.948 --> 00:02:00.323 align:middle line:84% about people, maybe there's going 00:02:00.323 --> 00:02:02.230 align:middle line:84% to be much more scope for discrimination. 00:02:02.230 --> 00:02:04.016 align:middle line:84% At least it's an interesting hypothesis. 00:02:04.016 --> 00:02:08.509 align:middle line:90% 00:02:08.509 --> 00:02:12.080 align:middle line:84% Early papers on price discrimination. 00:02:12.080 --> 00:02:14.480 align:middle line:84% One of the first things that happened in this literature, 00:02:14.480 --> 00:02:17.390 align:middle line:84% after the NEIO revolution, was just people arguing, 00:02:17.390 --> 00:02:19.130 align:middle line:84% do firms actually price discriminate? 00:02:19.130 --> 00:02:21.380 align:middle line:84% How can we tell whether price discrimination is really 00:02:21.380 --> 00:02:26.000 align:middle line:84% occurring or if it's just cost-based price differences? 00:02:26.000 --> 00:02:28.250 align:middle line:90% Two of the early papers on this. 00:02:28.250 --> 00:02:32.065 align:middle line:84% Severin Borenstein had this very nice paper that just argued-- 00:02:32.065 --> 00:02:33.440 align:middle line:84% at the time in the United States, 00:02:33.440 --> 00:02:37.078 align:middle line:84% you could either buy leaded or unleaded gas. 00:02:37.078 --> 00:02:39.620 align:middle line:84% At some point, we switched from having leaded gas to unleaded 00:02:39.620 --> 00:02:43.070 align:middle line:84% gas because leaded gas caused lead poisoning, all kinds 00:02:43.070 --> 00:02:45.890 align:middle line:90% of public health problems. 00:02:45.890 --> 00:02:49.130 align:middle line:84% Because unleaded gas was initially 00:02:49.130 --> 00:02:51.508 align:middle line:84% used by people who had newer cars, 00:02:51.508 --> 00:02:53.300 align:middle line:84% those people were wealthier than the people 00:02:53.300 --> 00:02:56.060 align:middle line:84% who had cars that ran on leaded gas. 00:02:56.060 --> 00:02:58.552 align:middle line:84% Unleaded gas was a lot more expensive than leaded gas. 00:02:58.552 --> 00:03:01.010 align:middle line:84% And so the question was, was that because of the difficulty 00:03:01.010 --> 00:03:04.130 align:middle line:84% of preparing unleaded gas, or was that due to price 00:03:04.130 --> 00:03:04.990 align:middle line:90% discrimination? 00:03:04.990 --> 00:03:07.790 align:middle line:84% Borenstein has this paper, just basically 00:03:07.790 --> 00:03:09.880 align:middle line:84% doing very detailed cost accounting 00:03:09.880 --> 00:03:11.630 align:middle line:84% and arguing that, no, the price difference 00:03:11.630 --> 00:03:13.463 align:middle line:84% between leaded and unleaded gas was much too 00:03:13.463 --> 00:03:15.990 align:middle line:84% large to be anything other than price discrimination. 00:03:15.990 --> 00:03:18.110 align:middle line:84% And then Andrea Shepard has this paper, 00:03:18.110 --> 00:03:20.660 align:middle line:84% which would be totally self-evident to those of you who 00:03:20.660 --> 00:03:21.902 align:middle line:90% are from California. 00:03:21.902 --> 00:03:23.360 align:middle line:84% I know in California, you go there, 00:03:23.360 --> 00:03:27.020 align:middle line:84% and it's like self-serve gas is $5.99 a gallon 00:03:27.020 --> 00:03:29.370 align:middle line:84% and full-serve gas is $7.59 a gallon. 00:03:29.370 --> 00:03:31.010 align:middle line:84% And you just do the math in your head. 00:03:31.010 --> 00:03:33.200 align:middle line:84% I'm paying someone $1.50 a gallon to pump my gas. 00:03:33.200 --> 00:03:34.130 align:middle line:90% That's $2. 00:03:34.130 --> 00:03:36.060 align:middle line:84% That person, they only make $15 an hour. 00:03:36.060 --> 00:03:37.560 align:middle line:84% There's no way that's what it costs. 00:03:37.560 --> 00:03:40.160 align:middle line:84% But Andrea Shepard has a paper that 00:03:40.160 --> 00:03:44.920 align:middle line:84% looks at stations actually in Massachusetts, 00:03:44.920 --> 00:03:47.110 align:middle line:84% comparing stations that offer both self 00:03:47.110 --> 00:03:49.690 align:middle line:84% and full-service with stations that offer self-service 00:03:49.690 --> 00:03:51.550 align:middle line:84% only and full-service only, and using 00:03:51.550 --> 00:03:55.000 align:middle line:84% the gap between the self and full-serve only stations 00:03:55.000 --> 00:03:57.440 align:middle line:84% to estimate what the cost differences could be. 00:03:57.440 --> 00:04:00.240 align:middle line:84% And then looking at the mark-ups between the dual product 00:04:00.240 --> 00:04:02.740 align:middle line:84% stations and arguing that that's clear evidence that there's 00:04:02.740 --> 00:04:04.810 align:middle line:84% second degree price discrimination going on, 00:04:04.810 --> 00:04:06.910 align:middle line:84% where you're charging the wealthier people more 00:04:06.910 --> 00:04:10.030 align:middle line:84% for the high-quality product as the gas that someone else 00:04:10.030 --> 00:04:13.060 align:middle line:90% pumps for you. 00:04:13.060 --> 00:04:13.960 align:middle line:90% It's instructive. 00:04:13.960 --> 00:04:16.480 align:middle line:84% If you want to go back and get a sense for the history 00:04:16.480 --> 00:04:19.540 align:middle line:84% of the economics profession and how much the required 00:04:19.540 --> 00:04:22.160 align:middle line:84% level of statistics and data has changed, 00:04:22.160 --> 00:04:25.820 align:middle line:84% if you compare Shepard's paper in the JPE 00:04:25.820 --> 00:04:28.600 align:middle line:84% 30 years ago to the most recent paper, last paper I'm 00:04:28.600 --> 00:04:30.710 align:middle line:84% going to cover in this lecture, you 00:04:30.710 --> 00:04:32.590 align:middle line:84% get the sense of how the economics 00:04:32.590 --> 00:04:37.090 align:middle line:84% profession has changed in standards for empirical work. 00:04:37.090 --> 00:04:38.680 align:middle line:84% Second topic I wanted to discuss, 00:04:38.680 --> 00:04:42.470 align:middle line:84% though, is racial and gender discrimination in retail prices. 00:04:42.470 --> 00:04:44.650 align:middle line:84% This is another place where IO is definitely 00:04:44.650 --> 00:04:48.080 align:middle line:84% overlapping with the public policy question. 00:04:48.080 --> 00:04:53.290 align:middle line:84% The public policy question is, do prices 00:04:53.290 --> 00:04:55.040 align:middle line:90% differ by race and gender? 00:04:55.040 --> 00:04:58.510 align:middle line:84% And what's the effect of race and gender-based price 00:04:58.510 --> 00:05:02.530 align:middle line:84% discrimination on people in disadvantaged groups 00:05:02.530 --> 00:05:04.590 align:middle line:84% or potentially people in advantaged groups? 00:05:04.590 --> 00:05:07.120 align:middle line:90% 00:05:07.120 --> 00:05:09.220 align:middle line:84% The early papers in this literature 00:05:09.220 --> 00:05:11.390 align:middle line:90% reported very different results. 00:05:11.390 --> 00:05:13.750 align:middle line:90% I think one of the-- 00:05:13.750 --> 00:05:16.450 align:middle line:84% really one of the first audit studies in economics, Ayres 00:05:16.450 --> 00:05:19.840 align:middle line:90% and Siegelman. 00:05:19.840 --> 00:05:22.420 align:middle line:84% Ayres and Siegelman, basically the idea was, they just 00:05:22.420 --> 00:05:25.810 align:middle line:84% trained 38 undergraduate students 00:05:25.810 --> 00:05:30.290 align:middle line:84% who are like juniors or seniors to bargain for car prices. 00:05:30.290 --> 00:05:33.530 align:middle line:84% And they train them with a very specific script. 00:05:33.530 --> 00:05:36.160 align:middle line:84% So they were supposed to walk into a car dealer and say, 00:05:36.160 --> 00:05:40.580 align:middle line:84% I would like the Toyota Corolla, four cylinder engine, base trim. 00:05:40.580 --> 00:05:43.190 align:middle line:90% I'm willing to pay $10,200. 00:05:43.190 --> 00:05:44.770 align:middle line:90% I don't need financing. 00:05:44.770 --> 00:05:46.520 align:middle line:90% Can I have it at that price? 00:05:46.520 --> 00:05:51.490 align:middle line:84% And then they just recorded what did the dealer say back to them. 00:05:51.490 --> 00:05:55.930 align:middle line:84% And they took these 38 trained dealers-- of the 38 students 00:05:55.930 --> 00:05:58.510 align:middle line:84% they trained, 19 were white males, 00:05:58.510 --> 00:06:00.980 align:middle line:84% and the other 19 were not white males. 00:06:00.980 --> 00:06:04.660 align:middle line:84% And they sent pairs of them to the same dealer 00:06:04.660 --> 00:06:09.190 align:middle line:84% to ask for the same car, at times far enough apart that it 00:06:09.190 --> 00:06:11.080 align:middle line:90% wouldn't seem suspicious. 00:06:11.080 --> 00:06:14.500 align:middle line:84% And then they report back, what were the dealers' initial 00:06:14.500 --> 00:06:17.200 align:middle line:84% counteroffers, what was the end of the price they got after they 00:06:17.200 --> 00:06:20.120 align:middle line:84% had a long bargaining script they were supposed to follow, 00:06:20.120 --> 00:06:23.390 align:middle line:84% which included things like, if you say $10,000, 00:06:23.390 --> 00:06:25.740 align:middle line:84% he says $12,000, say, how about $11,000? 00:06:25.740 --> 00:06:27.710 align:middle line:84% Or if he says $10,000, you say $12,000, 00:06:27.710 --> 00:06:29.540 align:middle line:90% how about $10,200 or whatever? 00:06:29.540 --> 00:06:31.070 align:middle line:84% So they gave them specific scripts 00:06:31.070 --> 00:06:33.740 align:middle line:90% they were supposed to follow. 00:06:33.740 --> 00:06:35.870 align:middle line:84% If you read the paper, they even said, 00:06:35.870 --> 00:06:38.780 align:middle line:84% they rated all the students as being of average attractiveness 00:06:38.780 --> 00:06:41.810 align:middle line:84% because they were concerned that better looking people 00:06:41.810 --> 00:06:42.810 align:middle line:90% will get better deals. 00:06:42.810 --> 00:06:46.320 align:middle line:84% So they had white and Black students, men and women. 00:06:46.320 --> 00:06:48.800 align:middle line:84% They rated them all as being of average attractiveness 00:06:48.800 --> 00:06:50.957 align:middle line:84% in their hiring process, which I don't 00:06:50.957 --> 00:06:52.790 align:middle line:84% know whether hiring people by attractiveness 00:06:52.790 --> 00:06:53.707 align:middle line:90% is allowed these days. 00:06:53.707 --> 00:06:54.980 align:middle line:90% But anyway, they did that. 00:06:54.980 --> 00:06:57.560 align:middle line:84% And that was-- results were that there 00:06:57.560 --> 00:07:01.170 align:middle line:84% was tremendous discrimination against women and minorities. 00:07:01.170 --> 00:07:05.250 align:middle line:84% So the Black males were the ones who were treated worse. 00:07:05.250 --> 00:07:08.360 align:middle line:84% They were charged $1,100 more than white males for a car 00:07:08.360 --> 00:07:09.660 align:middle line:90% in the initial offer. 00:07:09.660 --> 00:07:12.180 align:middle line:84% And cars at the time were about $10,000. 00:07:12.180 --> 00:07:16.450 align:middle line:84% That's like charging 10% more on the price of a car. 00:07:16.450 --> 00:07:21.950 align:middle line:90% 00:07:21.950 --> 00:07:24.222 align:middle line:84% Black females were $410 more than white females. 00:07:24.222 --> 00:07:26.930 align:middle line:84% And I forget what the number is, but white females were also more 00:07:26.930 --> 00:07:28.490 align:middle line:90% than white males. 00:07:28.490 --> 00:07:32.250 align:middle line:84% So they reported there was this substantial discrimination, 5% 00:07:32.250 --> 00:07:33.820 align:middle line:90% to 10%, in the price of cars. 00:07:33.820 --> 00:07:36.510 align:middle line:90% 00:07:36.510 --> 00:07:38.250 align:middle line:84% About the same time, Penny Goldberg 00:07:38.250 --> 00:07:42.060 align:middle line:84% had a follow-up article, looking at exactly the same topic, 00:07:42.060 --> 00:07:45.340 align:middle line:84% using data from the Consumer Expenditure Survey. 00:07:45.340 --> 00:07:47.730 align:middle line:84% And one thing you worry about with the audit study 00:07:47.730 --> 00:07:50.970 align:middle line:84% is that these people were not actually buying cars 00:07:50.970 --> 00:07:55.090 align:middle line:84% and were the people who were not actually buying cars, 00:07:55.090 --> 00:07:57.280 align:middle line:84% would the prices they're bargaining 00:07:57.280 --> 00:07:59.830 align:middle line:84% end up being the same for real people who buy cars, 00:07:59.830 --> 00:08:02.273 align:middle line:84% and maybe the way that the dealers react to these college 00:08:02.273 --> 00:08:03.940 align:middle line:84% students was different from the way they 00:08:03.940 --> 00:08:06.160 align:middle line:90% would react to other people. 00:08:06.160 --> 00:08:09.430 align:middle line:84% So Goldberg used car purchases from the Consumer Expenditure 00:08:09.430 --> 00:08:12.250 align:middle line:84% Survey and did the same racial test. 00:08:12.250 --> 00:08:16.660 align:middle line:84% And she found no difference at all in prices paid for cars. 00:08:16.660 --> 00:08:20.050 align:middle line:84% Again, though, data limitation of Goldberg's paper 00:08:20.050 --> 00:08:23.920 align:middle line:84% was that, the way the Consumer Expenditure Survey is, is it's 00:08:23.920 --> 00:08:25.090 align:middle line:90% an annual survey. 00:08:25.090 --> 00:08:27.940 align:middle line:84% And you go to people like me and say, Glenn, 00:08:27.940 --> 00:08:30.130 align:middle line:84% did you buy a car in the last year? 00:08:30.130 --> 00:08:31.510 align:middle line:90% And I'd say, yes. 00:08:31.510 --> 00:08:32.770 align:middle line:90% What kind of car is it? 00:08:32.770 --> 00:08:36.400 align:middle line:90% It's a Hyundai Kona SUV. 00:08:36.400 --> 00:08:38.450 align:middle line:84% And then what did you pay for it? 00:08:38.450 --> 00:08:42.280 align:middle line:84% And so you're relying on both people describing the car to you 00:08:42.280 --> 00:08:46.210 align:middle line:84% and their memory 0 to 12 months later of, 00:08:46.210 --> 00:08:47.890 align:middle line:84% what price did you pay for your car? 00:08:47.890 --> 00:08:51.475 align:middle line:84% If you asked me what I paid for my last car that I bought, 00:08:51.475 --> 00:08:53.100 align:middle line:84% I have no idea within several thousand. 00:08:53.100 --> 00:08:55.920 align:middle line:84% At the time, when I was bargaining for the car, 00:08:55.920 --> 00:08:58.048 align:middle line:84% I knew very well what I should pay for the car 00:08:58.048 --> 00:08:59.590 align:middle line:84% and what a good deal was or whatever. 00:08:59.590 --> 00:09:02.610 align:middle line:84% But six months later, do I have any recollection 00:09:02.610 --> 00:09:04.710 align:middle line:90% of what the numbers were? 00:09:04.710 --> 00:09:06.480 align:middle line:90% I would say I don't. 00:09:06.480 --> 00:09:07.620 align:middle line:90% So anyway. 00:09:07.620 --> 00:09:10.350 align:middle line:84% But what Goldberg found was that there was no difference 00:09:10.350 --> 00:09:13.080 align:middle line:90% across racial groups. 00:09:13.080 --> 00:09:17.430 align:middle line:84% One other study about the same time, Katy Graddy 00:09:17.430 --> 00:09:20.490 align:middle line:84% had a paper where she looked at purchases of fish 00:09:20.490 --> 00:09:22.500 align:middle line:84% at the New York Fulton Street Fish Market 00:09:22.500 --> 00:09:24.480 align:middle line:90% and looked at the same question. 00:09:24.480 --> 00:09:27.240 align:middle line:84% There, she did it by just going to the Fulton Street Fish 00:09:27.240 --> 00:09:30.180 align:middle line:84% Market, watching people buying fish, 00:09:30.180 --> 00:09:34.860 align:middle line:84% and recording what price they agreed to buy the fish at. 00:09:34.860 --> 00:09:37.750 align:middle line:84% And she found that actually here, discrimination 00:09:37.750 --> 00:09:39.750 align:middle line:84% against whites at the Fulton Street Fish Market, 00:09:39.750 --> 00:09:45.530 align:middle line:84% that Asians paid 7% less than whites for the same fish. 00:09:45.530 --> 00:09:48.050 align:middle line:84% But again, these are limited number of cars. 00:09:48.050 --> 00:09:51.540 align:middle line:84% Ayres and Siegelman, it's 500 observations. 00:09:51.540 --> 00:09:53.390 align:middle line:90% Goldberg, it's not so many. 00:09:53.390 --> 00:09:57.680 align:middle line:84% It's more observations, but much more heterogeneity in the cars. 00:09:57.680 --> 00:10:00.750 align:middle line:84% And hard to know what the prices are in Graddy's. 00:10:00.750 --> 00:10:01.990 align:middle line:90% Again, it's a limited sample. 00:10:01.990 --> 00:10:05.570 align:middle line:84% What could one graduate student do hanging out in New York Fish 00:10:05.570 --> 00:10:08.520 align:middle line:90% Market for a month? 00:10:08.520 --> 00:10:09.020 align:middle line:90% OK. 00:10:09.020 --> 00:10:11.330 align:middle line:84% So first bit I wanted to cover was this one 00:10:11.330 --> 00:10:16.550 align:middle line:84% by Scott Morton, Zettelmeyer, Silva-Risso. 00:10:16.550 --> 00:10:18.330 align:middle line:90% Multiple motivations. 00:10:18.330 --> 00:10:21.620 align:middle line:84% One of them is just to re-examine the earlier 00:10:21.620 --> 00:10:24.560 align:middle line:90% findings with much better data. 00:10:24.560 --> 00:10:27.500 align:middle line:84% And then besides me wanting to ask the applied question 00:10:27.500 --> 00:10:31.460 align:middle line:84% of, Is there racial and gender discrimination in car prices?, 00:10:31.460 --> 00:10:33.200 align:middle line:84% it's a little bit of trying to think 00:10:33.200 --> 00:10:35.480 align:middle line:84% about, how would we think about this from an IO 00:10:35.480 --> 00:10:38.712 align:middle line:84% perspective of IO says that people with different price 00:10:38.712 --> 00:10:40.670 align:middle line:84% elasticities are going to pay different prices, 00:10:40.670 --> 00:10:43.190 align:middle line:84% and do we see evidence that these price patterns look 00:10:43.190 --> 00:10:46.920 align:middle line:84% like they follow price elasticity and other models? 00:10:46.920 --> 00:10:50.930 align:middle line:84% And then the other motivation was, 00:10:50.930 --> 00:10:54.140 align:middle line:84% look at this shift from online sale-- 00:10:54.140 --> 00:10:56.030 align:middle line:84% from offline sales to online sales 00:10:56.030 --> 00:10:57.560 align:middle line:84% and whether online sales are going 00:10:57.560 --> 00:11:00.560 align:middle line:84% to change whether we get discrimination, which you could 00:11:00.560 --> 00:11:03.590 align:middle line:84% imagine-- you could imagine that in online sales people, 00:11:03.590 --> 00:11:06.260 align:middle line:84% the dealers don't know the race and gender of the person buying 00:11:06.260 --> 00:11:08.730 align:middle line:84% the car, and therefore they quote different prices. 00:11:08.730 --> 00:11:12.500 align:middle line:84% Or you could also get the signaling of things differ-- 00:11:12.500 --> 00:11:16.110 align:middle line:84% signaling of knowledge and other things differs. 00:11:16.110 --> 00:11:19.810 align:middle line:84% So she's got the cartoon from the paper on the internet. 00:11:19.810 --> 00:11:20.920 align:middle line:90% Nobody knows you're a dog. 00:11:20.920 --> 00:11:22.795 align:middle line:84% And so dogs can do things different from what 00:11:22.795 --> 00:11:27.000 align:middle line:90% they can do in real life. 00:11:27.000 --> 00:11:27.660 align:middle line:90% OK. 00:11:27.660 --> 00:11:29.680 align:middle line:90% This is early internet. 00:11:29.680 --> 00:11:33.000 align:middle line:84% So this is data from 1999 to 2000. 00:11:33.000 --> 00:11:37.950 align:middle line:84% Autobytel was the leading online car purchasing site in 1999. 00:11:37.950 --> 00:11:41.470 align:middle line:84% The way Autobytel worked was you would go to Autobytel, 00:11:41.470 --> 00:11:43.350 align:middle line:84% you would fill out a form telling it, 00:11:43.350 --> 00:11:46.710 align:middle line:90% I want to buy a Toyota Corolla. 00:11:46.710 --> 00:11:51.270 align:middle line:84% And then Autobytel would simply send your name and email address 00:11:51.270 --> 00:11:53.050 align:middle line:90% to one dealer. 00:11:53.050 --> 00:11:57.330 align:middle line:84% And that one dealer was supposed to quote you a good, fair price. 00:11:57.330 --> 00:11:59.530 align:middle line:84% So it was supposed to be a no haggle experience. 00:11:59.530 --> 00:12:01.420 align:middle line:84% You just say, I want a Toyota Corolla. 00:12:01.420 --> 00:12:03.360 align:middle line:84% They pass your name on to exactly one dealer, 00:12:03.360 --> 00:12:06.000 align:middle line:84% and that dealer emails you half an hour later and says, 00:12:06.000 --> 00:12:11.130 align:middle line:84% I could sell you that Corolla for $11,300. 00:12:11.130 --> 00:12:12.210 align:middle line:90% OK. 00:12:12.210 --> 00:12:14.310 align:middle line:84% The nice thing about this paper, though, 00:12:14.310 --> 00:12:16.740 align:middle line:84% when you compare it to the earlier ones on discrimination 00:12:16.740 --> 00:12:20.190 align:middle line:84% is, the data from autobytel.com, they 00:12:20.190 --> 00:12:22.060 align:middle line:84% have the complete Autobytel database. 00:12:22.060 --> 00:12:25.330 align:middle line:84% So 2 million people sent in such requests to Autobytel. 00:12:25.330 --> 00:12:29.650 align:middle line:84% They can track whether each of those-- what dealer each 00:12:29.650 --> 00:12:32.080 align:middle line:90% of those requests was sent to. 00:12:32.080 --> 00:12:34.510 align:middle line:84% And then the more impressive data set, 00:12:34.510 --> 00:12:36.400 align:middle line:84% there's a firm called JD Power that 00:12:36.400 --> 00:12:38.390 align:middle line:90% gathers data from car dealers. 00:12:38.390 --> 00:12:40.610 align:middle line:84% And JD Power is basically-- it's a data provider. 00:12:40.610 --> 00:12:42.610 align:middle line:84% They gather information from car dealers about the sales, 00:12:42.610 --> 00:12:44.560 align:middle line:84% and then they share that information with other car 00:12:44.560 --> 00:12:47.018 align:middle line:84% dealers who want to know what other car dealers are selling 00:12:47.018 --> 00:12:48.880 align:middle line:84% things for and with the car manufacturers. 00:12:48.880 --> 00:12:52.090 align:middle line:84% And basically, the JD Power data, 00:12:52.090 --> 00:12:54.520 align:middle line:84% it's a complete listing of all cars 00:12:54.520 --> 00:12:58.180 align:middle line:84% purchased in the United States from thousands of dealerships 00:12:58.180 --> 00:13:01.350 align:middle line:84% and has a tremendous amount of information about those cars. 00:13:01.350 --> 00:13:04.060 align:middle line:90% So it's transaction-level data. 00:13:04.060 --> 00:13:11.710 align:middle line:84% It has 671,000 car purchases from 3,562 dealerships. 00:13:11.710 --> 00:13:15.860 align:middle line:84% On every car, you not just get, it's a Toyota Corolla. 00:13:15.860 --> 00:13:21.110 align:middle line:84% You get, it's the Toyota Corolla LX with a 1.4 liter engine. 00:13:21.110 --> 00:13:23.830 align:middle line:84% Here are the 14 accessories on the window sticker 00:13:23.830 --> 00:13:24.950 align:middle line:90% that it had installed. 00:13:24.950 --> 00:13:26.030 align:middle line:90% It had the cup holder. 00:13:26.030 --> 00:13:27.650 align:middle line:90% It had this radio. 00:13:27.650 --> 00:13:30.430 align:middle line:84% It had, I don't know, the sunroof. 00:13:30.430 --> 00:13:33.580 align:middle line:84% All the accessories that are listed on the window sticker. 00:13:33.580 --> 00:13:38.630 align:middle line:84% It has the dealer cost of all of those accessories. 00:13:38.630 --> 00:13:41.560 align:middle line:84% So it knows that if you go into the car dealer 00:13:41.560 --> 00:13:44.080 align:middle line:84% and the upgraded radio is $399, it 00:13:44.080 --> 00:13:47.020 align:middle line:84% knows that radio cost the dealer $127. 00:13:47.020 --> 00:13:48.970 align:middle line:90% And so it records all that. 00:13:48.970 --> 00:13:52.570 align:middle line:84% It also records the trade-in in overallowance. 00:13:52.570 --> 00:13:56.240 align:middle line:84% People often are selling a car when they're buying a car. 00:13:56.240 --> 00:13:58.240 align:middle line:84% And you bring your old car, and the dealer says, 00:13:58.240 --> 00:14:01.550 align:middle line:84% I'll give you $3,500 for your 9-year-old car. 00:14:01.550 --> 00:14:03.490 align:middle line:84% They actually record the dealer's notation 00:14:03.490 --> 00:14:05.440 align:middle line:84% of both how much the dealer paid for the car 00:14:05.440 --> 00:14:07.860 align:middle line:84% and what the dealer thought the car was actually worth. 00:14:07.860 --> 00:14:11.170 align:middle line:84% So the dealer puts in, I gave a $3,500 trade-in allowance 00:14:11.170 --> 00:14:16.630 align:middle line:84% for a car that was worth $3,750 or a car that was worth $2,850. 00:14:16.630 --> 00:14:21.640 align:middle line:84% So it's tremendous price data compared to what you 00:14:21.640 --> 00:14:25.000 align:middle line:90% got in Goldberg or whatever. 00:14:25.000 --> 00:14:29.400 align:middle line:84% It also contains-- this is not privacy. 00:14:29.400 --> 00:14:32.030 align:middle line:84% I'm sure they had to go a while through the IRB for this. 00:14:32.030 --> 00:14:34.480 align:middle line:84% The data set has the buyer's home address. 00:14:34.480 --> 00:14:36.420 align:middle line:90% It has the buyer's name. 00:14:36.420 --> 00:14:38.460 align:middle line:84% And based on having the buyer's name, 00:14:38.460 --> 00:14:41.760 align:middle line:84% they make guesses of the buyer's gender 00:14:41.760 --> 00:14:47.340 align:middle line:84% and whether the buyer is Latinx or Asian ethnicity. 00:14:47.340 --> 00:14:49.660 align:middle line:90% OK. 00:14:49.660 --> 00:14:50.160 align:middle line:90% Anyway. 00:14:50.160 --> 00:14:52.285 align:middle line:84% And then, yeah, the transaction data from Autobytel 00:14:52.285 --> 00:14:56.100 align:middle line:90% has 2 million requests. 00:14:56.100 --> 00:14:59.280 align:middle line:84% Because the Autobytel data also has the buyer's home address 00:14:59.280 --> 00:15:01.168 align:middle line:84% and name, they can match it to this data 00:15:01.168 --> 00:15:03.210 align:middle line:84% and see whether the person bought from the dealer 00:15:03.210 --> 00:15:06.060 align:middle line:84% Autobytel sent them to or whether the person bought a car 00:15:06.060 --> 00:15:07.655 align:middle line:90% from a different dealer later. 00:15:07.655 --> 00:15:12.150 align:middle line:90% 00:15:12.150 --> 00:15:15.980 align:middle line:84% So initial analysis of this data is very, very simple. 00:15:15.980 --> 00:15:17.240 align:middle line:90% It's just a regression. 00:15:17.240 --> 00:15:20.140 align:middle line:84% They're just regressing the log of the price paid 00:15:20.140 --> 00:15:25.540 align:middle line:84% on demographics of the buyer and a set of controls x. 00:15:25.540 --> 00:15:29.170 align:middle line:84% The set of controls has car fixed effects, 00:15:29.170 --> 00:15:32.120 align:middle line:84% controls for the cost of all the included accessories. 00:15:32.120 --> 00:15:36.160 align:middle line:84% It's got month, region, weekend, end-of-month dummies, 00:15:36.160 --> 00:15:38.325 align:middle line:90% a whole bunch of fixed effects. 00:15:38.325 --> 00:15:39.700 align:middle line:84% And the idea is that you can just 00:15:39.700 --> 00:15:42.550 align:middle line:84% pick up the demographic differences in prices 00:15:42.550 --> 00:15:45.220 align:middle line:84% from those demographic coefficients. 00:15:45.220 --> 00:15:48.550 align:middle line:84% And most of this is just playing off the JD Power data, not 00:15:48.550 --> 00:15:49.560 align:middle line:90% the Autobytel data. 00:15:49.560 --> 00:15:53.070 align:middle line:90% 00:15:53.070 --> 00:15:53.760 align:middle line:90% OK. 00:15:53.760 --> 00:15:54.640 align:middle line:90% What do they find? 00:15:54.640 --> 00:15:58.680 align:middle line:84% Well, they're getting much more precise estimates 00:15:58.680 --> 00:16:02.170 align:middle line:84% than previous papers because they have so much more data. 00:16:02.170 --> 00:16:03.840 align:middle line:84% And basically what these data say 00:16:03.840 --> 00:16:07.500 align:middle line:84% is that there is discrimination in pricing, 00:16:07.500 --> 00:16:10.650 align:middle line:84% but that it's a lot smaller than Ayres and Siegelman had 00:16:10.650 --> 00:16:13.950 align:middle line:84% suggested discrimination might be in equilibrium. 00:16:13.950 --> 00:16:14.680 align:middle line:90% So examples. 00:16:14.680 --> 00:16:17.220 align:middle line:90% 00:16:17.220 --> 00:16:19.680 align:middle line:84% They don't know who is Black, but they 00:16:19.680 --> 00:16:25.590 align:middle line:84% do know the percentage Black in the census block group 00:16:25.590 --> 00:16:26.830 align:middle line:90% where you live. 00:16:26.830 --> 00:16:30.210 align:middle line:84% And they know whether the census block group where you live 00:16:30.210 --> 00:16:31.830 align:middle line:90% is primarily Black. 00:16:31.830 --> 00:16:36.000 align:middle line:84% And they find that this suggests that those in Black areas 00:16:36.000 --> 00:16:40.050 align:middle line:84% pay roughly 1.5% more than those living in white areas. 00:16:40.050 --> 00:16:45.812 align:middle line:84% Those in Hispanic areas pay 1.1% more. 00:16:45.812 --> 00:16:47.270 align:middle line:84% All these coefficients are circled. 00:16:47.270 --> 00:16:48.687 align:middle line:84% I don't know if that really helps. 00:16:48.687 --> 00:16:53.920 align:middle line:84% Those in Asian areas are paying 4/10 of a percent less. 00:16:53.920 --> 00:16:58.870 align:middle line:84% The Hispanic and Asian effects, you can also see them-- 00:16:58.870 --> 00:17:02.710 align:middle line:84% people who live in more heavily Hispanic-- more heavily 00:17:02.710 --> 00:17:08.089 align:middle line:90% Asian census blocks pay less. 00:17:08.089 --> 00:17:14.140 align:middle line:84% Those who themselves have Asian sounding names also pay less. 00:17:14.140 --> 00:17:17.119 align:middle line:90% And women pay more for cars. 00:17:17.119 --> 00:17:19.160 align:middle line:90% This one is small. 00:17:19.160 --> 00:17:23.450 align:middle line:84% Women pay 2/10 of a percent more for cars on average. 00:17:23.450 --> 00:17:27.160 align:middle line:90% 00:17:27.160 --> 00:17:31.270 align:middle line:84% In addition to these demographic differences, 00:17:31.270 --> 00:17:33.430 align:middle line:90% you also see differences-- 00:17:33.430 --> 00:17:36.430 align:middle line:84% I guess, given the 670,000 car sample, 00:17:36.430 --> 00:17:38.810 align:middle line:84% just about everything is significant. 00:17:38.810 --> 00:17:40.780 align:middle line:84% And so you get, for instance, those 00:17:40.780 --> 00:17:43.670 align:middle line:84% who live in more college educated areas pay less. 00:17:43.670 --> 00:17:46.360 align:middle line:84% Those in areas with more home owners pay less. 00:17:46.360 --> 00:17:48.360 align:middle line:90% So you get other-- 00:17:48.360 --> 00:17:50.670 align:middle line:84% you might have thought the discrimination would 00:17:50.670 --> 00:17:54.150 align:middle line:84% go in the sense of, like with the leaded/unleaded gasoline, 00:17:54.150 --> 00:17:57.060 align:middle line:84% high income people end up being charged higher prices. 00:17:57.060 --> 00:17:59.402 align:middle line:84% Here, the messages sound like the discrimination, 00:17:59.402 --> 00:18:01.110 align:middle line:84% to the extent it's there, looks like it's 00:18:01.110 --> 00:18:04.660 align:middle line:84% discrimination in favor of more sophisticated buyers. 00:18:04.660 --> 00:18:06.510 align:middle line:84% So the people who have college degrees, 00:18:06.510 --> 00:18:09.390 align:middle line:84% own their homes, whatever, they end up paying less 00:18:09.390 --> 00:18:12.060 align:middle line:84% for their cars, perhaps because as if they're 00:18:12.060 --> 00:18:15.150 align:middle line:84% better at bargaining or have more information or more ability 00:18:15.150 --> 00:18:17.790 align:middle line:84% to comparison shop at multiple dealers 00:18:17.790 --> 00:18:19.260 align:middle line:84% or are better using the internet, 00:18:19.260 --> 00:18:22.940 align:middle line:90% whatever the factor is. 00:18:22.940 --> 00:18:25.515 align:middle line:84% But again, these effects tend to be relatively small. 00:18:25.515 --> 00:18:27.140 align:middle line:84% They're all in the orders of a fraction 00:18:27.140 --> 00:18:29.125 align:middle line:90% of a percent in the price. 00:18:29.125 --> 00:18:34.890 align:middle line:90% 00:18:34.890 --> 00:18:36.450 align:middle line:90% OK. 00:18:36.450 --> 00:18:37.998 align:middle line:84% Why are these prices differences? 00:18:37.998 --> 00:18:39.540 align:middle line:84% You can think of a number of reasons. 00:18:39.540 --> 00:18:42.810 align:middle line:84% If you are a Chicago School economist, you can come up with, 00:18:42.810 --> 00:18:45.480 align:middle line:84% no, these are all just picking up cost differences 00:18:45.480 --> 00:18:49.020 align:middle line:84% because my cost of selling someone a car, 00:18:49.020 --> 00:18:51.660 align:middle line:84% there is this cost of, I have a salesperson's time 00:18:51.660 --> 00:18:53.640 align:middle line:84% talking to them that I have to recoup, 00:18:53.640 --> 00:18:55.660 align:middle line:90% even if the person doesn't buy. 00:18:55.660 --> 00:18:59.412 align:middle line:84% And so people less likely to buy might get quoted higher prices. 00:18:59.412 --> 00:19:00.870 align:middle line:84% You also have these effects of, who 00:19:00.870 --> 00:19:03.690 align:middle line:84% is going to bring the car into the dealer for service? 00:19:03.690 --> 00:19:06.762 align:middle line:84% And perhaps the wealthier people are 00:19:06.762 --> 00:19:08.970 align:middle line:84% going to bring the car back to the dealer for service 00:19:08.970 --> 00:19:11.680 align:middle line:84% and get charged $199 for an oil change 00:19:11.680 --> 00:19:14.560 align:middle line:84% instead of going to the Jiffy Lube and paying $19.99. 00:19:14.560 --> 00:19:17.430 align:middle line:84% Therefore, the car dealers are more anxious to sell to them 00:19:17.430 --> 00:19:19.972 align:middle line:84% because they know they're going to make the money back later. 00:19:19.972 --> 00:19:23.080 align:middle line:90% 00:19:23.080 --> 00:19:25.120 align:middle line:84% Obviously, the price discrimination theory 00:19:25.120 --> 00:19:28.750 align:middle line:84% suggests that this could just be profit maximizing 00:19:28.750 --> 00:19:32.560 align:middle line:84% against differences, either differences in intensity 00:19:32.560 --> 00:19:36.940 align:middle line:84% of preferences or search costs, ability to comparison shop, 00:19:36.940 --> 00:19:38.950 align:middle line:84% differences in bargaining skills. 00:19:38.950 --> 00:19:42.760 align:middle line:84% Or there could be racial or gender bias. 00:19:42.760 --> 00:19:43.970 align:middle line:90% I dislike women. 00:19:43.970 --> 00:19:46.090 align:middle line:84% I'm going to charge women more for cars 00:19:46.090 --> 00:19:49.150 align:middle line:90% just because I don't like them. 00:19:49.150 --> 00:19:55.120 align:middle line:84% So the data set, unfortunately, while it's very rich, 00:19:55.120 --> 00:19:56.530 align:middle line:90% it doesn't have-- 00:19:56.530 --> 00:19:59.180 align:middle line:84% there's no exogenous price variation in this data. 00:19:59.180 --> 00:20:02.160 align:middle line:84% So there's nothing really to estimate price elasticities 00:20:02.160 --> 00:20:02.660 align:middle line:90% here. 00:20:02.660 --> 00:20:05.480 align:middle line:84% This is just equilibrium prices paid. 00:20:05.480 --> 00:20:08.470 align:middle line:84% So that really knocks out a lot of the opportunities 00:20:08.470 --> 00:20:11.110 align:middle line:84% you would like to see of, is the price discrimination model 00:20:11.110 --> 00:20:12.200 align:middle line:90% really explaining this? 00:20:12.200 --> 00:20:14.170 align:middle line:84% You'd really like to say, do women 00:20:14.170 --> 00:20:16.042 align:middle line:84% have different demand elasticities than men? 00:20:16.042 --> 00:20:18.250 align:middle line:84% And are those demand elasticities such that you would 00:20:18.250 --> 00:20:22.100 align:middle line:84% price 2/10 of a percent different because of them? 00:20:22.100 --> 00:20:26.890 align:middle line:84% So instead, they just have some auxiliary regressions, 00:20:26.890 --> 00:20:30.310 align:middle line:84% trying to make the case that they think-- 00:20:30.310 --> 00:20:33.237 align:middle line:84% so I guess, I don't know if they're trying to make a case. 00:20:33.237 --> 00:20:33.820 align:middle line:90% They come now. 00:20:33.820 --> 00:20:35.237 align:middle line:84% Their bottom line is they're going 00:20:35.237 --> 00:20:38.470 align:middle line:84% to argue that this looks a lot like profit maximizing, price 00:20:38.470 --> 00:20:42.460 align:middle line:84% discrimination, and exploitation of consumers with higher search 00:20:42.460 --> 00:20:45.730 align:middle line:84% costs or who are less sophisticated at shopping. 00:20:45.730 --> 00:20:49.633 align:middle line:84% So what are some of the things that do this? 00:20:49.633 --> 00:20:50.133 align:middle line:90% OK. 00:20:50.133 --> 00:20:55.850 align:middle line:84% So doo doo doo doo doo doo doo doo doo doo doo doo. 00:20:55.850 --> 00:20:56.390 align:middle line:90% Yeah. 00:20:56.390 --> 00:21:00.470 align:middle line:84% So, for instance, the race and gender effects 00:21:00.470 --> 00:21:02.300 align:middle line:84% are only a little smaller in the sample 00:21:02.300 --> 00:21:04.320 align:middle line:90% that doesn't require financing. 00:21:04.320 --> 00:21:07.490 align:middle line:84% And so they argue that in the sample that comes in and tells 00:21:07.490 --> 00:21:10.160 align:middle line:84% you, I have cash, I don't need dealer financing, 00:21:10.160 --> 00:21:13.640 align:middle line:84% you're more sure that the sale will go through. 00:21:13.640 --> 00:21:15.260 align:middle line:84% The cost base difference of I'm not 00:21:15.260 --> 00:21:17.927 align:middle line:84% worried this person isn't going to be able to buy the car anyway 00:21:17.927 --> 00:21:18.810 align:middle line:90% isn't there. 00:21:18.810 --> 00:21:21.170 align:middle line:84% They notice that the effects are really only slightly 00:21:21.170 --> 00:21:23.630 align:middle line:90% different in that effect. 00:21:23.630 --> 00:21:26.180 align:middle line:84% Another one is they look at the subsample of people 00:21:26.180 --> 00:21:28.280 align:middle line:90% who are trading in cars. 00:21:28.280 --> 00:21:31.970 align:middle line:84% The subsample of people who are trading in cars, 00:21:31.970 --> 00:21:35.173 align:middle line:84% if someone is trading in a car, they have a car, 00:21:35.173 --> 00:21:36.590 align:middle line:84% and they ought to be just about as 00:21:36.590 --> 00:21:39.380 align:middle line:84% good at driving around to other dealers and doing comparison. 00:21:39.380 --> 00:21:42.290 align:middle line:84% They have similar opportunities to drive in and do 00:21:42.290 --> 00:21:45.090 align:middle line:84% comparison shopping because you know they already own a car. 00:21:45.090 --> 00:21:48.080 align:middle line:84% And what they find is that in this restricting the sample 00:21:48.080 --> 00:21:53.820 align:middle line:84% to the people who are trading in cars, 00:21:53.820 --> 00:21:57.720 align:middle line:84% the race premiums are about half as big 00:21:57.720 --> 00:21:59.700 align:middle line:90% as they are in the full sample. 00:21:59.700 --> 00:22:03.150 align:middle line:84% And so they say that that suggests that the differential 00:22:03.150 --> 00:22:06.780 align:middle line:84% ability to shop, perhaps because you don't have a car, 00:22:06.780 --> 00:22:09.450 align:middle line:84% are less able to shop around, would 00:22:09.450 --> 00:22:11.850 align:middle line:84% be explaining at least half of the premium 00:22:11.850 --> 00:22:14.910 align:middle line:84% that they're finding to begin with. 00:22:14.910 --> 00:22:17.795 align:middle line:84% And obviously, I think part of-- from this comparison, 00:22:17.795 --> 00:22:19.920 align:middle line:84% that the effect is only so big if you don't control 00:22:19.920 --> 00:22:22.380 align:middle line:84% for anything, they would argue that, look, 00:22:22.380 --> 00:22:24.690 align:middle line:84% if it's down to 0.8 and 0.6 here, 00:22:24.690 --> 00:22:27.452 align:middle line:84% if we controlled for additional demographic variables, 00:22:27.452 --> 00:22:29.285 align:middle line:84% we would expect the gap to get even smaller. 00:22:29.285 --> 00:22:33.320 align:middle line:90% 00:22:33.320 --> 00:22:35.700 align:middle line:90% Questions on that? 00:22:35.700 --> 00:22:36.200 align:middle line:90% OK. 00:22:36.200 --> 00:22:40.940 align:middle line:84% Then the final part of the paper is 00:22:40.940 --> 00:22:44.300 align:middle line:84% they also look at people who bought 00:22:44.300 --> 00:22:47.030 align:middle line:90% the cars through Autobytel. 00:22:47.030 --> 00:22:49.790 align:middle line:84% And the Autobytel sample, the main sample they look for 00:22:49.790 --> 00:22:52.970 align:middle line:84% is people who shop through Autobytel 00:22:52.970 --> 00:22:56.620 align:middle line:84% and buy from the dealer that Autobytel assigns them to. 00:22:56.620 --> 00:22:57.762 align:middle line:90% All right. 00:22:57.762 --> 00:22:59.720 align:middle line:84% I forget whether it was buy through that dealer 00:22:59.720 --> 00:23:00.900 align:middle line:90% or buy through-- 00:23:00.900 --> 00:23:01.400 align:middle line:90% no. 00:23:01.400 --> 00:23:06.302 align:middle line:84% I guess it's eventually bought on autobytel.com 00:23:06.302 --> 00:23:08.510 align:middle line:84% the same car that they searched for, not a different. 00:23:08.510 --> 00:23:10.552 align:middle line:84% So you throw out people who searched on Autobytel 00:23:10.552 --> 00:23:12.480 align:middle line:84% for a Corolla and then buy a Honda Civic, 00:23:12.480 --> 00:23:13.970 align:middle line:84% but they keep everyone who searched 00:23:13.970 --> 00:23:16.490 align:middle line:84% on Autobytel for a Corolla and buys a Corolla. 00:23:16.490 --> 00:23:18.470 align:middle line:90% What they find is that-- 00:23:18.470 --> 00:23:19.440 align:middle line:90% two main results here. 00:23:19.440 --> 00:23:20.300 align:middle line:90% Do I have arrows? 00:23:20.300 --> 00:23:21.800 align:middle line:90% I do not have arrows. 00:23:21.800 --> 00:23:25.330 align:middle line:84% One is that Autobytel buyers pay less. 00:23:25.330 --> 00:23:27.830 align:middle line:84% They pay about 9/10 less than the people searching for a car 00:23:27.830 --> 00:23:29.360 align:middle line:90% through traditional means. 00:23:29.360 --> 00:23:31.010 align:middle line:84% So this is, again, arguing something 00:23:31.010 --> 00:23:34.800 align:middle line:84% about either the Autobytel company or sophistication 00:23:34.800 --> 00:23:36.010 align:middle line:90% of people using the internet. 00:23:36.010 --> 00:23:37.740 align:middle line:84% The fact that they're using an internet to buy a car 00:23:37.740 --> 00:23:39.720 align:middle line:84% and getting a price quote, they end up 00:23:39.720 --> 00:23:41.340 align:middle line:90% paying less for their cars. 00:23:41.340 --> 00:23:45.090 align:middle line:84% And then the interactions with the demographics, 00:23:45.090 --> 00:23:49.810 align:middle line:84% you'll see that the Black coefficient here was 1.5% 00:23:49.810 --> 00:23:53.370 align:middle line:84% but that Autobytel times Black is minus 1.2. 00:23:53.370 --> 00:23:58.870 align:middle line:84% Hispanic was 0.7 plus 0.5, is like 1.2. 00:23:58.870 --> 00:24:01.840 align:middle line:84% You get a minus 2% for buying it online. 00:24:01.840 --> 00:24:09.450 align:middle line:84% So that the-- it seems like the people buying through Autobytel, 00:24:09.450 --> 00:24:15.950 align:middle line:84% the race and gender effects have largely been eliminated. 00:24:15.950 --> 00:24:18.750 align:middle line:84% And obviously, in many cases, the dealers 00:24:18.750 --> 00:24:21.180 align:middle line:84% are still going-- at least the gender, 00:24:21.180 --> 00:24:22.800 align:middle line:84% the Asian/Hispanic things, dealers 00:24:22.800 --> 00:24:24.790 align:middle line:84% are going to see the name of the person buying the car. 00:24:24.790 --> 00:24:26.332 align:middle line:84% So they would be able to discriminate 00:24:26.332 --> 00:24:27.720 align:middle line:90% based on those factors. 00:24:27.720 --> 00:24:31.782 align:middle line:84% But they appear not to in the Autobytel data. 00:24:31.782 --> 00:24:33.490 align:middle line:84% And they argue that this, again, suggests 00:24:33.490 --> 00:24:36.070 align:middle line:84% that a lot of what we're seeing is discrimination 00:24:36.070 --> 00:24:39.190 align:middle line:84% against people who are less able to shop 00:24:39.190 --> 00:24:40.845 align:middle line:90% or less savvy at shopping. 00:24:40.845 --> 00:24:48.120 align:middle line:90% 00:24:48.120 --> 00:24:51.380 align:middle line:90% Any questions on that paper? 00:24:51.380 --> 00:24:52.960 align:middle line:90% No? 00:24:52.960 --> 00:24:55.300 align:middle line:90% OK. 00:24:55.300 --> 00:25:00.340 align:middle line:90% So yeah. 00:25:00.340 --> 00:25:02.190 align:middle line:84% I guess one final comment I wanted to make 00:25:02.190 --> 00:25:07.350 align:middle line:84% is that if you were to ask, What would be the effect of banning 00:25:07.350 --> 00:25:11.220 align:middle line:84% discrimination here, at least banning mean discrimination?, 00:25:11.220 --> 00:25:14.490 align:middle line:84% this paper essentially makes it suggest 00:25:14.490 --> 00:25:16.890 align:middle line:84% that banning discrimination would not do 00:25:16.890 --> 00:25:18.580 align:middle line:90% much at all to dealer profits. 00:25:18.580 --> 00:25:24.080 align:middle line:84% Because if you think about it, a feature of monopoly pricing is 00:25:24.080 --> 00:25:29.050 align:middle line:84% that d(pi)/d(p) evaluated at pm equals 0. 00:25:29.050 --> 00:25:32.960 align:middle line:84% So a change in the price has only a second order effect 00:25:32.960 --> 00:25:34.650 align:middle line:90% on the dealer's profits. 00:25:34.650 --> 00:25:38.987 align:middle line:84% And so if you're only changing-- if d(pi)/d(p)-- 00:25:38.987 --> 00:25:41.570 align:middle line:84% if you have only a second order effect on profits from a price 00:25:41.570 --> 00:25:42.945 align:middle line:84% change and you're banning dealers 00:25:42.945 --> 00:25:45.703 align:middle line:84% from having 1% price differences, that 00:25:45.703 --> 00:25:47.870 align:middle line:84% seems like that would have a very, very minor effect 00:25:47.870 --> 00:25:49.170 align:middle line:90% on dealer profits. 00:25:49.170 --> 00:25:53.568 align:middle line:84% Now, this is banning mean price discrimination. 00:25:53.568 --> 00:25:55.110 align:middle line:84% It is the case that in the data there 00:25:55.110 --> 00:25:57.210 align:middle line:84% are some people who pay 5% or 10% 00:25:57.210 --> 00:25:59.890 align:middle line:84% more and some people who pay 5% or 10% less than others. 00:25:59.890 --> 00:26:03.180 align:middle line:84% So it could be that the dealers are making a lot of profits 00:26:03.180 --> 00:26:05.670 align:middle line:84% from being allowed to discriminate, 00:26:05.670 --> 00:26:07.950 align:middle line:84% but being allowed to discriminate in a way that 00:26:07.950 --> 00:26:12.210 align:middle line:84% systematically disadvantages demographic groups that can't 00:26:12.210 --> 00:26:13.415 align:middle line:90% be important to the dealers. 00:26:13.415 --> 00:26:18.540 align:middle line:90% 00:26:18.540 --> 00:26:19.350 align:middle line:90% Questions? 00:26:19.350 --> 00:26:23.190 align:middle line:90% 00:26:23.190 --> 00:26:23.690 align:middle line:90% OK. 00:26:23.690 --> 00:26:26.397 align:middle line:84% So then the second topic I figured 00:26:26.397 --> 00:26:27.980 align:middle line:84% I'd focus on with the other two papers 00:26:27.980 --> 00:26:32.560 align:middle line:84% was this question of, how large are 00:26:32.560 --> 00:26:35.080 align:middle line:84% the gains from price discrimination online going 00:26:35.080 --> 00:26:36.070 align:middle line:90% to be? 00:26:36.070 --> 00:26:39.580 align:middle line:84% And what would that price discrimination 00:26:39.580 --> 00:26:41.800 align:middle line:84% do to consumer surplus and welfare? 00:26:41.800 --> 00:26:43.960 align:middle line:84% And again, as I said at the start of this lecture, 00:26:43.960 --> 00:26:47.047 align:middle line:84% in the previous time, we know that price discrimination 00:26:47.047 --> 00:26:48.880 align:middle line:84% is always good for firms because they always 00:26:48.880 --> 00:26:51.310 align:middle line:90% have the option not to do it. 00:26:51.310 --> 00:26:58.090 align:middle line:84% But there was that sort of-- we had the Bergemann-Brooks-Morris 00:26:58.090 --> 00:27:01.330 align:middle line:84% triangle that said, monopoly pricing is here. 00:27:01.330 --> 00:27:03.710 align:middle line:84% Perfect price discrimination is here. 00:27:03.710 --> 00:27:07.810 align:middle line:84% If you allow firms to start price discriminating, 00:27:07.810 --> 00:27:09.400 align:middle line:84% we don't know whether you go this way 00:27:09.400 --> 00:27:11.600 align:middle line:90% or this way or this way. 00:27:11.600 --> 00:27:16.712 align:middle line:84% And this was consumer surplus on this axis, profit on this axis. 00:27:16.712 --> 00:27:18.170 align:middle line:84% You let firms start discriminating. 00:27:18.170 --> 00:27:20.380 align:middle line:84% You don't know-- you know profits are going up, 00:27:20.380 --> 00:27:23.230 align:middle line:84% but you don't know whether consumer surplus is going up, 00:27:23.230 --> 00:27:26.020 align:middle line:84% is being flat, or is going down, or is going down 00:27:26.020 --> 00:27:28.990 align:middle line:84% by enough to be below the 45 degree line 00:27:28.990 --> 00:27:31.783 align:middle line:90% and make social welfare go down. 00:27:31.783 --> 00:27:33.700 align:middle line:84% So in some sense, what these papers are trying 00:27:33.700 --> 00:27:37.900 align:middle line:90% to answer is, which direction-- 00:27:37.900 --> 00:27:39.490 align:middle line:90% how big are these arrows? 00:27:39.490 --> 00:27:43.370 align:middle line:84% Do profits move a lot, or do profits move a little? 00:27:43.370 --> 00:27:47.440 align:middle line:84% And in what direction does the consumer surplus change 00:27:47.440 --> 00:27:49.730 align:middle line:90% when the profits are changing? 00:27:49.730 --> 00:27:52.040 align:middle line:90% OK. 00:27:52.040 --> 00:27:55.580 align:middle line:84% So first, I'm going to cover the paper written 00:27:55.580 --> 00:27:58.610 align:middle line:84% by someone who was a graduate student at the time using 00:27:58.610 --> 00:28:02.970 align:middle line:84% graduate student accessible data. 00:28:02.970 --> 00:28:08.930 align:middle line:84% So Shiller is using information from web browsing activities 00:28:08.930 --> 00:28:10.250 align:middle line:90% from ComScore. 00:28:10.250 --> 00:28:17.460 align:middle line:84% So ComScore data, it's not free, but it's purchasable. 00:28:17.460 --> 00:28:18.500 align:middle line:90% You can access it. 00:28:18.500 --> 00:28:22.850 align:middle line:84% He accessed it through Wharton Research Data Services. 00:28:22.850 --> 00:28:25.710 align:middle line:84% ComScore data, it's modified web logs. 00:28:25.710 --> 00:28:26.840 align:middle line:90% So he has this. 00:28:26.840 --> 00:28:29.240 align:middle line:84% ComScore is a company that, basically, it's 00:28:29.240 --> 00:28:34.210 align:middle line:84% an information broker that sells information to firms. 00:28:34.210 --> 00:28:36.580 align:middle line:90% What they do is you allow-- 00:28:36.580 --> 00:28:37.837 align:middle line:90% ComScore will pay you money. 00:28:37.837 --> 00:28:39.670 align:middle line:84% I don't know if it's like they pay you $100. 00:28:39.670 --> 00:28:44.260 align:middle line:84% And they will say, can you install this browser add-on? 00:28:44.260 --> 00:28:46.120 align:middle line:84% And this browser add-on is going to track 00:28:46.120 --> 00:28:49.160 align:middle line:84% everything you do online and send it up to ComScore. 00:28:49.160 --> 00:28:52.090 align:middle line:84% So you have to be fairly confident 00:28:52.090 --> 00:28:53.840 align:middle line:90% in the anonymity of that. 00:28:53.840 --> 00:28:57.160 align:middle line:84% And obviously, we all know it's very hard to have anonymous-- 00:28:57.160 --> 00:28:59.290 align:middle line:84% given that people tend to look at their own page, 00:28:59.290 --> 00:29:01.840 align:middle line:84% their own Facebook, or their own whatever, 00:29:01.840 --> 00:29:04.552 align:middle line:84% how anonymous is your entire stream of web usage? 00:29:04.552 --> 00:29:06.010 align:middle line:84% Can someone figure out who you were 00:29:06.010 --> 00:29:08.410 align:middle line:84% and what else you were looking at online? 00:29:08.410 --> 00:29:13.150 align:middle line:84% But anyway, the ComScore panel at the time had 61,000 users who 00:29:13.150 --> 00:29:15.400 align:middle line:84% were allowing ComScore to track everything they did 00:29:15.400 --> 00:29:16.790 align:middle line:90% and reporting that. 00:29:16.790 --> 00:29:19.743 align:middle line:84% And then firms interested in tracking consumers 00:29:19.743 --> 00:29:21.160 align:middle line:84% or understanding consumer behavior 00:29:21.160 --> 00:29:23.140 align:middle line:90% could buy ComScore data. 00:29:23.140 --> 00:29:26.950 align:middle line:90% So Ben bought the ComScore data. 00:29:26.950 --> 00:29:30.010 align:middle line:84% The ComScore data that you get in this public way, 00:29:30.010 --> 00:29:32.300 align:middle line:90% it's not the full browsing logs. 00:29:32.300 --> 00:29:34.270 align:middle line:84% They realized that was a little sensitive, 00:29:34.270 --> 00:29:36.020 align:middle line:84% but what they do provide you with is still 00:29:36.020 --> 00:29:41.180 align:middle line:84% a lot, which is timestamped lists of all the domains 00:29:41.180 --> 00:29:42.060 align:middle line:90% that you go to. 00:29:42.060 --> 00:29:45.020 align:middle line:84% So for instance, this person went to facebook.com. 00:29:45.020 --> 00:29:47.660 align:middle line:84% They viewed 83 pages on facebook.com 00:29:47.660 --> 00:29:49.650 align:middle line:90% over a period of 57 minutes. 00:29:49.650 --> 00:29:52.280 align:middle line:84% Then they went to ebay.com, they did this. 00:29:52.280 --> 00:29:53.960 align:middle line:84% Then they went to newyorktimes.com. 00:29:53.960 --> 00:29:56.910 align:middle line:84% They viewed 12 pages there, spent 15 minutes. 00:29:56.910 --> 00:29:59.570 align:middle line:84% And it just gives you the entire history. 00:29:59.570 --> 00:30:01.048 align:middle line:90% It does also contain-- 00:30:01.048 --> 00:30:02.840 align:middle line:84% because it's meant for commercial purposes, 00:30:02.840 --> 00:30:07.310 align:middle line:84% it does also contain information on things that they purchased. 00:30:07.310 --> 00:30:11.000 align:middle line:84% So if you went to Netflix and signed up for Netflix, 00:30:11.000 --> 00:30:13.310 align:middle line:84% that would be recorded in the ComScore data. 00:30:13.310 --> 00:30:15.920 align:middle line:84% This person went to Netflix, viewed 12 pages, 00:30:15.920 --> 00:30:17.900 align:middle line:90% and purchased a subscription. 00:30:17.900 --> 00:30:18.400 align:middle line:90% OK? 00:30:18.400 --> 00:30:20.960 align:middle line:90% 00:30:20.960 --> 00:30:25.550 align:middle line:84% The paper focuses on demand for Netflix at the time. 00:30:25.550 --> 00:30:28.290 align:middle line:84% Maybe you guys are too young to remember this world. 00:30:28.290 --> 00:30:30.620 align:middle line:84% Netflix was not an online service 00:30:30.620 --> 00:30:32.360 align:middle line:84% where you clicked on movies and could 00:30:32.360 --> 00:30:33.930 align:middle line:90% download them and watch them. 00:30:33.930 --> 00:30:36.680 align:middle line:84% Netflix was a service that mailed you DVDs. 00:30:36.680 --> 00:30:38.360 align:middle line:84% So you would have a list on Netflix of, 00:30:38.360 --> 00:30:40.760 align:middle line:84% these are the movies that I want to watch. 00:30:40.760 --> 00:30:44.300 align:middle line:84% Netflix would mail you a movie in an envelope. 00:30:44.300 --> 00:30:46.460 align:middle line:84% And then, when you were done watching your movie, 00:30:46.460 --> 00:30:47.930 align:middle line:84% you would put the envelope in the mail, 00:30:47.930 --> 00:30:49.347 align:middle line:84% you would mail it back to Netflix, 00:30:49.347 --> 00:30:51.620 align:middle line:84% and the next movie on your queue would arrive. 00:30:51.620 --> 00:30:53.900 align:middle line:84% So it was kind of like a video rental store, 00:30:53.900 --> 00:30:56.410 align:middle line:84% except you didn't have to walk to the video rental store. 00:30:56.410 --> 00:30:58.160 align:middle line:84% And you could have different subscriptions 00:30:58.160 --> 00:31:00.110 align:middle line:84% like a subscription where you only 00:31:00.110 --> 00:31:01.880 align:middle line:84% have one movie at a time or a subscription 00:31:01.880 --> 00:31:03.360 align:middle line:84% where you got two movies at a time 00:31:03.360 --> 00:31:05.642 align:middle line:90% so that you didn't have to wait. 00:31:05.642 --> 00:31:07.100 align:middle line:84% The two-movie subscription, as soon 00:31:07.100 --> 00:31:09.050 align:middle line:84% as you watch one of the two, you mail it back. 00:31:09.050 --> 00:31:11.090 align:middle line:84% And then the second one can arrive in the mail 00:31:11.090 --> 00:31:13.173 align:middle line:84% two days later, giving you time to watch it. 00:31:13.173 --> 00:31:14.840 align:middle line:84% You still have the second movie to watch 00:31:14.840 --> 00:31:17.660 align:middle line:90% while you're doing that. 00:31:17.660 --> 00:31:19.140 align:middle line:90% Why does he do this? 00:31:19.140 --> 00:31:21.980 align:middle line:84% In some ways, this is a very clever design. 00:31:21.980 --> 00:31:23.480 align:middle line:84% I think, if you think about somebody 00:31:23.480 --> 00:31:26.210 align:middle line:84% who has access to this limited data, 00:31:26.210 --> 00:31:29.270 align:middle line:84% and he's trying to understand price discrimination, 00:31:29.270 --> 00:31:32.520 align:middle line:84% there are several nice things about Netflix. 00:31:32.520 --> 00:31:37.160 align:middle line:84% One is the number of people who purchase something in a year 00:31:37.160 --> 00:31:38.160 align:middle line:90% is limited. 00:31:38.160 --> 00:31:40.580 align:middle line:84% But he can get everyone who's a Netflix subscriber 00:31:40.580 --> 00:31:42.500 align:middle line:84% because if you're a Netflix subscriber, 00:31:42.500 --> 00:31:45.680 align:middle line:84% you regularly go to netflix.com and view multiple pages 00:31:45.680 --> 00:31:47.880 align:middle line:84% and spend a little time there choosing your movies. 00:31:47.880 --> 00:31:50.840 align:middle line:84% And so by seeing whether someone is a regular visitor 00:31:50.840 --> 00:31:53.360 align:middle line:84% to netflix.com, you'll know whether they're 00:31:53.360 --> 00:31:55.936 align:middle line:84% a Netflix subscriber or not a Netflix subscriber. 00:31:55.936 --> 00:31:58.910 align:middle line:90% 00:31:58.910 --> 00:32:01.760 align:middle line:84% You can reasonably treat Netflix as a monopolist. 00:32:01.760 --> 00:32:04.880 align:middle line:84% Blockbuster video also had a DVD by mail program, 00:32:04.880 --> 00:32:07.290 align:middle line:84% but Netflix had a dominant market share. 00:32:07.290 --> 00:32:10.880 align:middle line:84% And you can think of them as monopoly pricing. 00:32:10.880 --> 00:32:13.940 align:middle line:84% The fact that 16% of all Americans 00:32:13.940 --> 00:32:16.310 align:middle line:84% had Netflix subscriptions at the time 00:32:16.310 --> 00:32:18.710 align:middle line:84% means if you're trying to think about estimating demand 00:32:18.710 --> 00:32:20.425 align:middle line:90% for something, you need the-- 00:32:20.425 --> 00:32:22.930 align:middle line:90% 00:32:22.930 --> 00:32:27.130 align:middle line:84% you have 60,000 users, but if only 50 of them have bought 00:32:27.130 --> 00:32:30.350 align:middle line:84% something, you really only have 50 observations in some sense. 00:32:30.350 --> 00:32:34.690 align:middle line:84% So the fact that you have 60,000 observations and 16% have 00:32:34.690 --> 00:32:37.840 align:middle line:84% Netflix subscriptions mean you've got 10,000 subscribers 00:32:37.840 --> 00:32:39.370 align:middle line:90% and 50,000 non-subscribers. 00:32:39.370 --> 00:32:42.550 align:middle line:84% And comparing the subscribers and non-subscribers is much more 00:32:42.550 --> 00:32:49.780 align:middle line:84% powerful than if you're comparing 60 and 59,940. 00:32:49.780 --> 00:32:52.990 align:middle line:84% And then another aspect is, like I said, 00:32:52.990 --> 00:32:55.340 align:middle line:84% people do studies of cardboard boxes. 00:32:55.340 --> 00:32:58.030 align:middle line:84% Netflix is a very simple business. 00:32:58.030 --> 00:33:00.550 align:middle line:84% You can estimate very well Netflix's cost, which 00:33:00.550 --> 00:33:03.170 align:middle line:90% is Netflix has to buy DVDs. 00:33:03.170 --> 00:33:05.800 align:middle line:84% They have to have people put them in envelopes, 00:33:05.800 --> 00:33:08.170 align:middle line:84% send them in the mail, and then get them back 00:33:08.170 --> 00:33:09.760 align:middle line:90% and sort them again. 00:33:09.760 --> 00:33:11.260 align:middle line:90% Netflix also had just-- 00:33:11.260 --> 00:33:13.210 align:middle line:84% so if you look at Netflix's accounting data 00:33:13.210 --> 00:33:15.730 align:middle line:84% and you look at how much they spend buying DVDs, 00:33:15.730 --> 00:33:17.890 align:middle line:84% how much do they spend on their mail rooms, 00:33:17.890 --> 00:33:20.740 align:middle line:84% you can get very good accounting estimates of, 00:33:20.740 --> 00:33:24.620 align:middle line:84% what is Netflix's cost of having a subscriber? 00:33:24.620 --> 00:33:28.250 align:middle line:84% So he treats this as this is a product where 00:33:28.250 --> 00:33:31.940 align:middle line:84% we can observe demand well, and we know the cost from accounting 00:33:31.940 --> 00:33:34.640 align:middle line:90% data. 00:33:34.640 --> 00:33:40.130 align:middle line:84% Big limitation, there is no price variation. 00:33:40.130 --> 00:33:42.500 align:middle line:84% Netflix did have some coupons or whatever, 00:33:42.500 --> 00:33:44.057 align:middle line:90% but he doesn't know that. 00:33:44.057 --> 00:33:45.890 align:middle line:84% He doesn't know what most of his subscribers 00:33:45.890 --> 00:33:47.348 align:middle line:84% paid for Netflix because he doesn't 00:33:47.348 --> 00:33:48.643 align:middle line:90% observe them buying Netflix. 00:33:48.643 --> 00:33:50.810 align:middle line:84% He doesn't know what kind of subscription they have, 00:33:50.810 --> 00:33:53.030 align:middle line:84% whether it's a two-disc subscription 00:33:53.030 --> 00:33:55.680 align:middle line:84% or a three-disc subscription or something like that. 00:33:55.680 --> 00:34:00.800 align:middle line:84% So normally you think of, I need exogenous price variation 00:34:00.800 --> 00:34:02.960 align:middle line:90% to estimate demand elasticities. 00:34:02.960 --> 00:34:07.730 align:middle line:84% So I guess the second clever insight 00:34:07.730 --> 00:34:09.770 align:middle line:84% that he makes in the paper, besides that Netflix 00:34:09.770 --> 00:34:11.540 align:middle line:84% is a good experimental subject, is 00:34:11.540 --> 00:34:16.310 align:middle line:84% that you can try to estimate price elasticities here 00:34:16.310 --> 00:34:18.179 align:middle line:90% without any price variation. 00:34:18.179 --> 00:34:22.370 align:middle line:84% So Tobias is going to talk about demand estimation 00:34:22.370 --> 00:34:24.500 align:middle line:84% in a couple of weeks, which can't come too soon, 00:34:24.500 --> 00:34:25.917 align:middle line:84% it seems, the papers I'm covering. 00:34:25.917 --> 00:34:28.133 align:middle line:90% 00:34:28.133 --> 00:34:29.550 align:middle line:84% A basic insight of that literature 00:34:29.550 --> 00:34:31.440 align:middle line:84% is, if you have exogenous price variation, 00:34:31.440 --> 00:34:33.840 align:middle line:84% you can estimate demand elasticities. 00:34:33.840 --> 00:34:37.350 align:middle line:84% And then once you've estimated the demand elasticities, 00:34:37.350 --> 00:34:40.469 align:middle line:84% if you're willing to assume that firms are profit maximizing, 00:34:40.469 --> 00:34:44.580 align:middle line:84% then you've got this condition that the firm is maximizing 00:34:44.580 --> 00:34:50.070 align:middle line:84% over p, p minus c D of p, or the firm 00:34:50.070 --> 00:34:52.469 align:middle line:84% is going to be doing p minus c over p 00:34:52.469 --> 00:34:55.250 align:middle line:84% equals 1 over the elasticity of demand. 00:34:55.250 --> 00:34:57.270 align:middle line:84% So if I observe the elasticity of demand 00:34:57.270 --> 00:35:00.000 align:middle line:84% and I observe the prices, I can figure out 00:35:00.000 --> 00:35:02.460 align:middle line:84% what the firm's costs are without any cost data 00:35:02.460 --> 00:35:08.220 align:middle line:84% if I assume that firms are profit maximizing monopolists. 00:35:08.220 --> 00:35:11.970 align:middle line:84% What Shiller notes is that if you think about this equation, 00:35:11.970 --> 00:35:15.060 align:middle line:84% let's suppose that we know that costs are known 00:35:15.060 --> 00:35:18.240 align:middle line:84% and Netflix is profit maximizing. 00:35:18.240 --> 00:35:22.050 align:middle line:84% If costs are known and Netflix is profit maximizing, 00:35:22.050 --> 00:35:25.530 align:middle line:84% then I know p and I know c and I know 00:35:25.530 --> 00:35:29.800 align:middle line:84% p, which means I can infer what the epsilon is. 00:35:29.800 --> 00:35:35.350 align:middle line:84% So having the knowledge of what fraction of people buy 00:35:35.350 --> 00:35:40.540 align:middle line:84% lets me get dq / dp because epsilon is dq / dp 00:35:40.540 --> 00:35:42.110 align:middle line:90% times p over q or whatever. 00:35:42.110 --> 00:35:46.480 align:middle line:84% So if I know p and I know q, then I can also get dq / dp. 00:35:46.480 --> 00:35:48.855 align:middle line:90% So that gives me dq / dp. 00:35:48.855 --> 00:35:53.300 align:middle line:90% 00:35:53.300 --> 00:35:57.500 align:middle line:84% So I can really do what we often want 00:35:57.500 --> 00:36:01.580 align:middle line:84% to get out of demand estimation, even without any exogenous price 00:36:01.580 --> 00:36:04.100 align:middle line:84% variation, if I have cost variation and profit 00:36:04.100 --> 00:36:05.990 align:middle line:90% maximization. 00:36:05.990 --> 00:36:08.790 align:middle line:90% So anyway, what does he do? 00:36:08.790 --> 00:36:12.590 align:middle line:84% So again, it looks like a lot of the logit-like demand structures 00:36:12.590 --> 00:36:13.850 align:middle line:90% I talked about earlier. 00:36:13.850 --> 00:36:15.920 align:middle line:84% He assumes that consumer i's utility 00:36:15.920 --> 00:36:24.170 align:middle line:84% from buying Netflix package j is some utility from Netflix 00:36:24.170 --> 00:36:28.620 align:middle line:84% minus alpha times the price of the package. 00:36:28.620 --> 00:36:30.980 align:middle line:84% So alpha is this price sensitivity. 00:36:30.980 --> 00:36:35.870 align:middle line:84% Plus xi j, which is the quality of package j, 00:36:35.870 --> 00:36:38.120 align:middle line:84% plus this random error term that makes 00:36:38.120 --> 00:36:43.740 align:middle line:84% some people who are similar buy and some people not buy. 00:36:43.740 --> 00:36:46.830 align:middle line:84% And he assumes that this average willingness 00:36:46.830 --> 00:36:51.930 align:middle line:84% to pay for Netflix among people with demographics X 00:36:51.930 --> 00:36:54.110 align:middle line:90% is Xi times beta. 00:36:54.110 --> 00:36:57.120 align:middle line:84% And he's going to think about that Netflix could discriminate 00:36:57.120 --> 00:36:58.710 align:middle line:84% by charging different prices to people 00:36:58.710 --> 00:37:01.020 align:middle line:90% with different observables Xi. 00:37:01.020 --> 00:37:04.720 align:middle line:84% And in his data set, there are all kinds of observables. 00:37:04.720 --> 00:37:08.550 align:middle line:84% So he has 18 demographic variables, 00:37:08.550 --> 00:37:13.170 align:middle line:84% 15 variables that relate to the quantity and timing of when you 00:37:13.170 --> 00:37:18.930 align:middle line:84% are on the internet, but then he puts in 4,600 other variables, 00:37:18.930 --> 00:37:21.480 align:middle line:84% recording things like, how much time do you spend 00:37:21.480 --> 00:37:23.528 align:middle line:90% on microsoftnetwork.com? 00:37:23.528 --> 00:37:25.570 align:middle line:84% How much time do you spend on the New York Times? 00:37:25.570 --> 00:37:27.370 align:middle line:84% How much time do you spend on Gamefly? 00:37:27.370 --> 00:37:30.340 align:middle line:84% How much time do you spend on many other websites? 00:37:30.340 --> 00:37:34.860 align:middle line:84% And so he basically takes the time spent on the 4,600 most 00:37:34.860 --> 00:37:40.290 align:middle line:84% popular websites as a 4,600 dimensional description of you. 00:37:40.290 --> 00:37:48.050 align:middle line:84% And then, in his model, there's only a single alpha. 00:37:48.050 --> 00:37:51.310 align:middle line:90% So dq / dp is the same for-- 00:37:51.310 --> 00:37:53.020 align:middle line:84% well, not dq / dp, but d (utility) 00:37:53.020 --> 00:37:55.990 align:middle line:84% / d(p) is the same for everybody. 00:37:55.990 --> 00:38:04.247 align:middle line:84% But because the mean utilities differ, 00:38:04.247 --> 00:38:06.580 align:middle line:84% people who have a very high v are going to buy the thing 00:38:06.580 --> 00:38:08.630 align:middle line:84% basically almost regardless of what the price is. 00:38:08.630 --> 00:38:11.720 align:middle line:84% People with a very low v are highly unlikely to buy it. 00:38:11.720 --> 00:38:14.650 align:middle line:84% And so the price sensitivity does differ across people 00:38:14.650 --> 00:38:17.770 align:middle line:84% because you have these nonlinearities in how much 00:38:17.770 --> 00:38:19.270 align:middle line:84% it's worth to you, and that affects 00:38:19.270 --> 00:38:21.170 align:middle line:90% how much you're willing to pay. 00:38:21.170 --> 00:38:29.430 align:middle line:84% And so then he's going to estimate these beta parameters. 00:38:29.430 --> 00:38:31.180 align:middle line:84% He's going to estimate the beta parameters 00:38:31.180 --> 00:38:36.050 align:middle line:84% via a Lasso-like maximum likelihood estimation. 00:38:36.050 --> 00:38:38.480 align:middle line:84% Do people here know what Lasso models are? 00:38:38.480 --> 00:38:38.980 align:middle line:90% OK. 00:38:38.980 --> 00:38:45.880 align:middle line:84% So Lasso model, you're used to seeing simple regression. 00:38:45.880 --> 00:38:47.630 align:middle line:90% How does OLS regression work? 00:38:47.630 --> 00:38:53.740 align:middle line:84% I'm just trying to maximize the sum of y minus yi minus beta Xi 00:38:53.740 --> 00:38:55.030 align:middle line:90% squared. 00:38:55.030 --> 00:38:58.120 align:middle line:84% And this is my OLS estimator of beta. 00:38:58.120 --> 00:39:03.730 align:middle line:84% Sorry, I minimize over beta y minus x beta squared. 00:39:03.730 --> 00:39:05.980 align:middle line:90% OK? 00:39:05.980 --> 00:39:09.995 align:middle line:84% The lasso model is, suppose you have-- 00:39:09.995 --> 00:39:12.370 align:middle line:84% and something you're taught in undergraduate econometrics 00:39:12.370 --> 00:39:14.380 align:middle line:84% is if I'm going to do this, I need 00:39:14.380 --> 00:39:17.270 align:middle line:84% to have many fewer x's than I have observations. 00:39:17.270 --> 00:39:20.290 align:middle line:90% And so I need the-- 00:39:20.290 --> 00:39:23.260 align:middle line:84% if the number of variables is N and the number of observations 00:39:23.260 --> 00:39:25.400 align:middle line:84% is T, I need N small relative to T. 00:39:25.400 --> 00:39:28.510 align:middle line:84% I need T going to infinity relative to the number 00:39:28.510 --> 00:39:33.130 align:middle line:84% of right-hand side variables in order to estimate this equation. 00:39:33.130 --> 00:39:37.480 align:middle line:84% What econometricians or computer scientists and econometricians 00:39:37.480 --> 00:39:39.397 align:middle line:84% have noticed is that you don't actually 00:39:39.397 --> 00:39:41.230 align:middle line:84% need that to estimate a regression if you're 00:39:41.230 --> 00:39:44.982 align:middle line:84% willing to make the assumption that most of the betas are 0. 00:39:44.982 --> 00:39:46.690 align:middle line:84% And if you're willing to-- the assumption 00:39:46.690 --> 00:39:48.107 align:middle line:84% is you have a regression where you 00:39:48.107 --> 00:39:50.180 align:middle line:84% have many, many right-hand side variables. 00:39:50.180 --> 00:39:52.210 align:middle line:90% Most of the betas are 0. 00:39:52.210 --> 00:39:54.470 align:middle line:84% The only thing is, you don't know which ones are 0. 00:39:54.470 --> 00:40:00.250 align:middle line:84% Then what you can do is instead minimize over beta 00:40:00.250 --> 00:40:03.760 align:middle line:90% the sum across observations i. 00:40:03.760 --> 00:40:09.010 align:middle line:90% yi minus Xi beta squared. 00:40:09.010 --> 00:40:18.995 align:middle line:84% And then subtract off the sum over all j of lambda j. 00:40:18.995 --> 00:40:20.120 align:middle line:90% I actually want to do this. 00:40:20.120 --> 00:40:25.700 align:middle line:84% Minus lambda the sum over j absolute value beta j. 00:40:25.700 --> 00:40:29.390 align:middle line:84% So I'm using i to index people, j to index 00:40:29.390 --> 00:40:31.700 align:middle line:90% variables in the equation. 00:40:31.700 --> 00:40:34.640 align:middle line:84% So if I instead minimize an objective 00:40:34.640 --> 00:40:36.320 align:middle line:84% function like this, where I'm minimizing 00:40:36.320 --> 00:40:39.380 align:middle line:84% the sum of squared residuals minus the sum 00:40:39.380 --> 00:40:42.020 align:middle line:84% of the absolute values of the coefficients. 00:40:42.020 --> 00:40:45.020 align:middle line:84% And because you're minimizing the sum of the absolute value 00:40:45.020 --> 00:40:46.580 align:middle line:84% of the coefficients, it's like you're 00:40:46.580 --> 00:40:49.880 align:middle line:84% putting a penalty function that looks like this if you 00:40:49.880 --> 00:40:52.230 align:middle line:90% use a nonzero beta j. 00:40:52.230 --> 00:40:55.940 align:middle line:84% And so anything where the beta j looks like it 00:40:55.940 --> 00:40:57.832 align:middle line:84% has a very, very small effect, you 00:40:57.832 --> 00:40:59.540 align:middle line:84% minimize a function like this with a kink 00:40:59.540 --> 00:41:02.810 align:middle line:84% by just setting beta j equal to 0. 00:41:02.810 --> 00:41:05.570 align:middle line:84% And so what you do in a Lasso model is 00:41:05.570 --> 00:41:07.460 align:middle line:84% you estimate a regression like this, 00:41:07.460 --> 00:41:10.070 align:middle line:84% and you put in these penalty functions. 00:41:10.070 --> 00:41:11.540 align:middle line:84% And you put those penalty functions 00:41:11.540 --> 00:41:14.600 align:middle line:84% in in a way that sets most of the coefficients equal to zero. 00:41:14.600 --> 00:41:18.135 align:middle line:84% And that is asymptotically a valid estimate 00:41:18.135 --> 00:41:19.760 align:middle line:84% under some set of assumptions about how 00:41:19.760 --> 00:41:25.070 align:middle line:84% fast the number of nonzero coefficients increases with t. 00:41:25.070 --> 00:41:26.060 align:middle line:90% OK? 00:41:26.060 --> 00:41:28.670 align:middle line:84% So he's doing something very much 00:41:28.670 --> 00:41:31.010 align:middle line:84% like that in maximum likelihood estimation, 00:41:31.010 --> 00:41:36.080 align:middle line:84% where what he's doing is he's maximizing 00:41:36.080 --> 00:41:41.510 align:middle line:84% the sum of the likelihood of yi given Xi. 00:41:41.510 --> 00:41:45.800 align:middle line:84% And then also subtracting lambda times the sum of the absolute 00:41:45.800 --> 00:41:47.480 align:middle line:90% value of the beta j's. 00:41:47.480 --> 00:41:49.760 align:middle line:84% So he's doing a maximum likelihood version 00:41:49.760 --> 00:41:52.640 align:middle line:84% of that, where you're maximizing the sum of the log likelihoods 00:41:52.640 --> 00:41:55.040 align:middle line:84% and subtracting this penalty function for every nonzero 00:41:55.040 --> 00:41:56.320 align:middle line:90% coefficient. 00:41:56.320 --> 00:41:59.630 align:middle line:84% And so what his model does in practice, I mean, 00:41:59.630 --> 00:42:03.420 align:middle line:84% I'm a little surprised at how many variables he keeps, 00:42:03.420 --> 00:42:06.570 align:middle line:84% but roughly he keeps 900 of the right-hand side 00:42:06.570 --> 00:42:08.230 align:middle line:90% variables in the equation. 00:42:08.230 --> 00:42:11.550 align:middle line:84% But then it's as if the other 3,700 variables get dropped from 00:42:11.550 --> 00:42:14.490 align:middle line:84% the equation because the beta estimates are so small that you 00:42:14.490 --> 00:42:15.455 align:middle line:90% just set them to 0. 00:42:15.455 --> 00:42:19.240 align:middle line:90% 00:42:19.240 --> 00:42:22.300 align:middle line:84% Just to give a sense of what his data looks like, 00:42:22.300 --> 00:42:25.390 align:middle line:84% the demographics are things like the age of the people. 00:42:25.390 --> 00:42:29.140 align:middle line:84% Is the oldest person in the household, 21 to 24, 25 to 29, 00:42:29.140 --> 00:42:30.940 align:middle line:90% 30 to 34? 00:42:30.940 --> 00:42:32.450 align:middle line:90% He's got race in there. 00:42:32.450 --> 00:42:34.600 align:middle line:90% He's got income in there. 00:42:34.600 --> 00:42:37.570 align:middle line:84% And then the browsing habits, he's got things like, 00:42:37.570 --> 00:42:40.810 align:middle line:84% what fraction of your internet use is midnight to 6:00 AM? 00:42:40.810 --> 00:42:43.100 align:middle line:84% What fraction is 6:00 AM to 9:00 AM? 00:42:43.100 --> 00:42:45.145 align:middle line:84% What fraction is 6:00 PM to midnight? 00:42:45.145 --> 00:42:47.770 align:middle line:84% How much do you use the internet on Monday, Tuesday, Wednesday, 00:42:47.770 --> 00:42:49.490 align:middle line:90% Thursday, Friday, Saturday? 00:42:49.490 --> 00:42:54.190 align:middle line:84% So he has this set of minimal demographics 00:42:54.190 --> 00:42:57.200 align:middle line:84% on people, although more than you would often see. 00:42:57.200 --> 00:42:59.690 align:middle line:84% At least he does have some set of income brackets. 00:42:59.690 --> 00:43:04.030 align:middle line:84% And then he's got all of your 4,600 website visits. 00:43:04.030 --> 00:43:07.862 align:middle line:84% And the results come out that the most powerful predictor 00:43:07.862 --> 00:43:09.820 align:middle line:84% of whether you're going to subscribe to Netflix 00:43:09.820 --> 00:43:13.210 align:middle line:84% is whether you're a frequent visitor to Gamefly.com. 00:43:13.210 --> 00:43:14.990 align:middle line:84% I don't know if Gamefly still exists. 00:43:14.990 --> 00:43:17.110 align:middle line:84% Gamefly was a company like Netflix, 00:43:17.110 --> 00:43:21.340 align:middle line:84% but that mailed you video game CD-ROMs. 00:43:21.340 --> 00:43:24.940 align:middle line:84% So you would just have a CD-ROM subscription to Gamefly. 00:43:24.940 --> 00:43:26.693 align:middle line:90% You're playing Call of Duty 3. 00:43:26.693 --> 00:43:28.610 align:middle line:84% You get tired of it, you mail it back to them, 00:43:28.610 --> 00:43:30.580 align:middle line:84% and they send you the next video game on your list. 00:43:30.580 --> 00:43:32.020 align:middle line:84% You play that for a month, you send it back, 00:43:32.020 --> 00:43:33.830 align:middle line:84% they send you the next one on your list. 00:43:33.830 --> 00:43:34.330 align:middle line:90% But anyway. 00:43:34.330 --> 00:43:38.020 align:middle line:84% So subscription to gamefly.com was the most powerful 00:43:38.020 --> 00:43:38.920 align:middle line:90% positive predictor. 00:43:38.920 --> 00:43:40.420 align:middle line:84% The most powerful negative predictor 00:43:40.420 --> 00:43:46.540 align:middle line:84% of having a Netflix subscription was, I don't know, ameblo.jp. 00:43:46.540 --> 00:43:50.560 align:middle line:84% Some kind of Japanese blogging site. 00:43:50.560 --> 00:43:52.220 align:middle line:84% People, I guess, who apparently visit 00:43:52.220 --> 00:43:55.190 align:middle line:84% Japanese blogging sites don't do a lot of watching 00:43:55.190 --> 00:43:56.160 align:middle line:90% Netflix movies. 00:43:56.160 --> 00:43:58.070 align:middle line:90% I don't know. 00:43:58.070 --> 00:44:01.700 align:middle line:84% People who visit audible.com and rent music 00:44:01.700 --> 00:44:03.675 align:middle line:90% also tend to rent DVDs. 00:44:03.675 --> 00:44:06.560 align:middle line:90% 00:44:06.560 --> 00:44:10.550 align:middle line:84% People who do things like visit jacksonville.com, 00:44:10.550 --> 00:44:15.350 align:middle line:84% they don't watch a lot of Netflix. 00:44:15.350 --> 00:44:17.000 align:middle line:90% You get the sense it's sort of-- 00:44:17.000 --> 00:44:19.100 align:middle line:84% obviously he has the age in there, 00:44:19.100 --> 00:44:21.860 align:middle line:84% but there are people who sit around and read 00:44:21.860 --> 00:44:27.350 align:middle line:84% local news on the internet and people who watch entertainment. 00:44:27.350 --> 00:44:29.360 align:middle line:90% And it is-- 00:44:29.360 --> 00:44:30.540 align:middle line:90% I don't know. 00:44:30.540 --> 00:44:33.770 align:middle line:84% I find this interesting, that, yes, 00:44:33.770 --> 00:44:36.353 align:middle line:84% once you think about it, yes, many, many, many of these things 00:44:36.353 --> 00:44:38.603 align:middle line:84% make sense if I'm trying to predict whether someone is 00:44:38.603 --> 00:44:39.890 align:middle line:90% likely to buy Netflix or not. 00:44:39.890 --> 00:44:42.390 align:middle line:84% There's a tremendous amount of information in their website. 00:44:42.390 --> 00:44:43.920 align:middle line:84% And people are very, very different, 00:44:43.920 --> 00:44:46.580 align:middle line:84% and it could be revealed through these many different channels 00:44:46.580 --> 00:44:47.486 align:middle line:90% simultaneously. 00:44:47.486 --> 00:44:51.600 align:middle line:90% 00:44:51.600 --> 00:44:52.100 align:middle line:90% OK. 00:44:52.100 --> 00:44:53.750 align:middle line:90% What does he find? 00:44:53.750 --> 00:44:59.360 align:middle line:84% First finding here is that demographic targeting-- 00:44:59.360 --> 00:45:01.765 align:middle line:84% and again, remember, he had the same demographics 00:45:01.765 --> 00:45:03.390 align:middle line:84% we were talking about in the car paper. 00:45:03.390 --> 00:45:08.120 align:middle line:84% He does have-- he has race, he has Hispanic. 00:45:08.120 --> 00:45:10.080 align:middle line:90% He's actually got income. 00:45:10.080 --> 00:45:12.440 align:middle line:84% All of those things, he says, are basically 00:45:12.440 --> 00:45:14.750 align:middle line:90% useless for pricing Netflix. 00:45:14.750 --> 00:45:17.870 align:middle line:90% He said that if you estimate-- 00:45:17.870 --> 00:45:19.910 align:middle line:84% I'm going to estimate demand elasticities 00:45:19.910 --> 00:45:21.950 align:middle line:84% as a function of the demographic group 00:45:21.950 --> 00:45:24.500 align:middle line:84% and think about switching from the optimal uniform price 00:45:24.500 --> 00:45:27.930 align:middle line:84% to the optimal demographic targeted price. 00:45:27.930 --> 00:45:30.560 align:middle line:84% And he finds that targeting the demographics 00:45:30.560 --> 00:45:33.060 align:middle line:84% would let Netflix increase its profits by 0.3%. 00:45:33.060 --> 00:45:35.690 align:middle line:90% 00:45:35.690 --> 00:45:37.310 align:middle line:84% I think, though, in some sense, this 00:45:37.310 --> 00:45:41.030 align:middle line:84% is consistent with what I said before in that, 00:45:41.030 --> 00:45:42.200 align:middle line:90% again, with that sort of-- 00:45:42.200 --> 00:45:45.380 align:middle line:84% if profits are quadratic in the extent to which you 00:45:45.380 --> 00:45:47.150 align:middle line:90% want to change the price-- 00:45:47.150 --> 00:45:50.580 align:middle line:84% and profit, it's a second order change. 00:45:50.580 --> 00:45:53.710 align:middle line:84% You only want to make changes on the order of a few percent, 00:45:53.710 --> 00:45:56.130 align:middle line:84% then the profit effect from that has to be small. 00:45:56.130 --> 00:45:57.420 align:middle line:90% And that's what he says. 00:45:57.420 --> 00:46:01.270 align:middle line:90% The profit impact is 0.3%. 00:46:01.270 --> 00:46:05.140 align:middle line:84% He argues that targeting on the web, though, in some sense, 00:46:05.140 --> 00:46:09.070 align:middle line:84% may be much more common than targeting in the offline world 00:46:09.070 --> 00:46:11.920 align:middle line:84% would be because these browsing variables are much more 00:46:11.920 --> 00:46:12.950 align:middle line:90% powerful. 00:46:12.950 --> 00:46:15.700 align:middle line:84% He estimates that they would let you increase Netflix's profits 00:46:15.700 --> 00:46:18.000 align:middle line:90% by 13%. 00:46:18.000 --> 00:46:20.500 align:middle line:84% It's still not a huge number when you think about you've got 00:46:20.500 --> 00:46:22.670 align:middle line:90% those 4,600 variables in there. 00:46:22.670 --> 00:46:24.508 align:middle line:84% What could you do with all of those? 00:46:24.508 --> 00:46:26.050 align:middle line:84% He says it's much, much more than you 00:46:26.050 --> 00:46:28.870 align:middle line:84% can do with demographics, but it's 13%. 00:46:28.870 --> 00:46:33.360 align:middle line:90% 00:46:33.360 --> 00:46:36.200 align:middle line:90% What are the welfare effects? 00:46:36.200 --> 00:46:37.230 align:middle line:90% Oh, do I have it? 00:46:37.230 --> 00:46:37.980 align:middle line:90% OK. 00:46:37.980 --> 00:46:42.540 align:middle line:90% So welfare effects are-- 00:46:42.540 --> 00:46:46.050 align:middle line:84% here he's arguing that with demographic targeting, 00:46:46.050 --> 00:46:47.970 align:middle line:90% we're actually-- 00:46:47.970 --> 00:46:52.620 align:middle line:84% in this diagram, it's the tiniest arrow in the world, 00:46:52.620 --> 00:46:54.370 align:middle line:90% but it's an arrow to the right. 00:46:54.370 --> 00:46:58.360 align:middle line:84% That is, the profits are going up by 0.3%. 00:46:58.360 --> 00:47:04.230 align:middle line:84% The consumer surplus is going up by 0.05%. 00:47:04.230 --> 00:47:08.230 align:middle line:84% Demographic targeting, it's like this kind of arrow. 00:47:08.230 --> 00:47:11.310 align:middle line:84% So it's a bigger arrow, going up 13%. 00:47:11.310 --> 00:47:17.700 align:middle line:84% And it's going slightly down on the consumer surplus side. 00:47:17.700 --> 00:47:19.770 align:middle line:84% Total social welfare, I assume, has 00:47:19.770 --> 00:47:22.560 align:middle line:84% got to be going up because you've 00:47:22.560 --> 00:47:24.450 align:middle line:90% got the 13% increase in profit. 00:47:24.450 --> 00:47:26.920 align:middle line:84% The only negative, a half percent on consumer surplus. 00:47:26.920 --> 00:47:28.890 align:middle line:84% I don't know what the mean consumer surplus is 00:47:28.890 --> 00:47:33.610 align:middle line:84% versus profits, but assuming that they're roughly comparable, 00:47:33.610 --> 00:47:36.713 align:middle line:84% social welfare has got to be going up in both cases. 00:47:36.713 --> 00:47:39.650 align:middle line:90% 00:47:39.650 --> 00:47:43.250 align:middle line:84% He does not have nearly as much data to do this. 00:47:43.250 --> 00:47:46.390 align:middle line:84% But he says, if Netflix were to use second degree discrimination 00:47:46.390 --> 00:47:52.300 align:middle line:84% instead, discriminating across people by offering the one CD 00:47:52.300 --> 00:47:55.970 align:middle line:84% plan, the two DVD plan, and the three DVD plan, 00:47:55.970 --> 00:47:58.090 align:middle line:84% there, he argues that second degree discrimination 00:47:58.090 --> 00:48:00.210 align:middle line:90% would be much more powerful. 00:48:00.210 --> 00:48:02.468 align:middle line:84% And he claims, based on a conversation with a Netflix 00:48:02.468 --> 00:48:04.510 align:middle line:84% executive, that they were not doing second degree 00:48:04.510 --> 00:48:06.350 align:middle line:90% discrimination at the time. 00:48:06.350 --> 00:48:09.130 align:middle line:90% He said that they were doing-- 00:48:09.130 --> 00:48:12.580 align:middle line:84% plans are priced proportional to cost with a single multiplier, 00:48:12.580 --> 00:48:14.200 align:middle line:84% whereas secondary discrimination model 00:48:14.200 --> 00:48:16.930 align:middle line:84% says you should be charging different markups 00:48:16.930 --> 00:48:19.390 align:middle line:90% to different groups. 00:48:19.390 --> 00:48:20.050 align:middle line:90% But anyway. 00:48:20.050 --> 00:48:22.180 align:middle line:84% He says that secondary discrimination could raise 00:48:22.180 --> 00:48:24.970 align:middle line:84% profits by much more and is a much bigger deal 00:48:24.970 --> 00:48:29.560 align:middle line:84% than the third degree discrimination they can do. 00:48:29.560 --> 00:48:33.580 align:middle line:84% And here's an example of what he's estimating, 00:48:33.580 --> 00:48:35.860 align:middle line:90% the consumer types. 00:48:35.860 --> 00:48:39.173 align:middle line:84% If you have no demographics at all, 00:48:39.173 --> 00:48:40.840 align:middle line:84% it's as if you're assuming that everyone 00:48:40.840 --> 00:48:43.750 align:middle line:84% has the constant mean willingness 00:48:43.750 --> 00:48:45.860 align:middle line:90% to pay v for the product. 00:48:45.860 --> 00:48:47.620 align:middle line:84% If you put in the demographics, it's 00:48:47.620 --> 00:48:51.520 align:middle line:84% basically saying, instead of having a spike here, almost 00:48:51.520 --> 00:48:54.280 align:middle line:84% everybody has one of these two willingness 00:48:54.280 --> 00:48:56.350 align:middle line:84% to pay that are very close to each other. 00:48:56.350 --> 00:48:59.320 align:middle line:84% And then when you look at the full model with web browsing 00:48:59.320 --> 00:49:04.690 align:middle line:84% data, you get this bigger distribution like that. 00:49:04.690 --> 00:49:07.690 align:middle line:84% And if you remember back to what I said about Bergemann, Brooks, 00:49:07.690 --> 00:49:10.880 align:middle line:84% Morris, it's the shape of this distribution that's 00:49:10.880 --> 00:49:13.190 align:middle line:84% going to tend to determine whether consumer surplus is 00:49:13.190 --> 00:49:14.640 align:middle line:90% going up or down. 00:49:14.640 --> 00:49:17.810 align:middle line:84% And it seems like the shape of this distribution 00:49:17.810 --> 00:49:20.330 align:middle line:84% is something to the fact that it keeps them from raising 00:49:20.330 --> 00:49:21.990 align:middle line:90% their prices by too much. 00:49:21.990 --> 00:49:25.370 align:middle line:84% And that's why that would have to be 00:49:25.370 --> 00:49:30.200 align:middle line:84% involved in consumer surplus not going down a lot 00:49:30.200 --> 00:49:30.910 align:middle line:90% when you do this. 00:49:30.910 --> 00:49:36.350 align:middle line:90% 00:49:36.350 --> 00:49:38.070 align:middle line:90% Questions on Shiller? 00:49:38.070 --> 00:49:38.632 align:middle line:90% No? 00:49:38.632 --> 00:49:39.840 align:middle line:90% I think I had one more slide. 00:49:39.840 --> 00:49:41.210 align:middle line:90% So some comments. 00:49:41.210 --> 00:49:44.360 align:middle line:90% 00:49:44.360 --> 00:49:47.330 align:middle line:84% Again, I think this is a nice, creative paper 00:49:47.330 --> 00:49:50.470 align:middle line:84% if you think about what it takes to write papers 00:49:50.470 --> 00:49:51.640 align:middle line:90% as graduate students. 00:49:51.640 --> 00:49:53.320 align:middle line:90% He had this sort of-- 00:49:53.320 --> 00:49:55.330 align:middle line:84% just uses an off-the-shelf data set, 00:49:55.330 --> 00:49:57.730 align:middle line:84% and he has these two different ideas. 00:49:57.730 --> 00:50:00.550 align:middle line:84% One is this identification idea that, if I 00:50:00.550 --> 00:50:02.770 align:middle line:84% can find a product where I think cost is known, 00:50:02.770 --> 00:50:04.360 align:middle line:84% I can estimate demand elasticities 00:50:04.360 --> 00:50:07.430 align:middle line:84% without having any price variation in the data. 00:50:07.430 --> 00:50:11.500 align:middle line:84% And Netflix is a product that's both very popular 00:50:11.500 --> 00:50:14.200 align:middle line:84% and where I can track usage and where I can 00:50:14.200 --> 00:50:16.215 align:middle line:90% get the observable cost data. 00:50:16.215 --> 00:50:19.440 align:middle line:90% 00:50:19.440 --> 00:50:21.210 align:middle line:84% He also has this other nice observation 00:50:21.210 --> 00:50:24.360 align:middle line:84% where I can do this comparison of demographic price 00:50:24.360 --> 00:50:29.430 align:middle line:84% discrimination versus web browsing price discrimination. 00:50:29.430 --> 00:50:32.100 align:middle line:84% I think these are important applied observations 00:50:32.100 --> 00:50:36.300 align:middle line:84% to the extent that, obviously, the limitations of the data 00:50:36.300 --> 00:50:38.640 align:middle line:90% may make it-- 00:50:38.640 --> 00:50:41.820 align:middle line:84% it's a good data point, but I don't know how much we believe. 00:50:41.820 --> 00:50:44.610 align:middle line:84% But to the extent we believe that, it suggests 00:50:44.610 --> 00:50:47.070 align:middle line:84% that demographics are of very little value for price 00:50:47.070 --> 00:50:48.450 align:middle line:90% discrimination. 00:50:48.450 --> 00:50:49.980 align:middle line:84% So maybe we really shouldn't expect 00:50:49.980 --> 00:50:52.800 align:middle line:84% to see much demographic price discrimination. 00:50:52.800 --> 00:50:56.980 align:middle line:84% It shows that web browsing can reveal a lot of information. 00:50:56.980 --> 00:51:00.390 align:middle line:84% And something else I would say is that it also shows that web 00:51:00.390 --> 00:51:02.820 align:middle line:84% browsing-based discrimination is-- 00:51:02.820 --> 00:51:04.560 align:middle line:84% one obstacle to discrimination is always 00:51:04.560 --> 00:51:07.260 align:middle line:90% people hiding who they are. 00:51:07.260 --> 00:51:08.850 align:middle line:90% If you try to have a student-- 00:51:08.850 --> 00:51:11.613 align:middle line:84% imagine you try to have a student surcharge at ART. 00:51:11.613 --> 00:51:13.530 align:middle line:84% If you go there and you're a graduate student, 00:51:13.530 --> 00:51:17.370 align:middle line:84% they charge you 20% more than someone who's not a student. 00:51:17.370 --> 00:51:19.510 align:middle line:84% It would be very hard to get someone 00:51:19.510 --> 00:51:21.933 align:middle line:84% to prove they're not a student to you. 00:51:21.933 --> 00:51:23.350 align:middle line:84% That price discrimination wouldn't 00:51:23.350 --> 00:51:25.750 align:middle line:84% work because it's hard to prove that you're not actually 00:51:25.750 --> 00:51:28.540 align:middle line:90% in school somewhere. 00:51:28.540 --> 00:51:34.120 align:middle line:84% Here, it would be very hard to distort your web browsing 00:51:34.120 --> 00:51:37.240 align:middle line:84% behavior in a way that prevented Netflix from discriminating 00:51:37.240 --> 00:51:40.990 align:middle line:84% against you because Netflix can discriminate against you 00:51:40.990 --> 00:51:45.350 align:middle line:84% by looking at your visits to 4,000 different websites. 00:51:45.350 --> 00:51:49.120 align:middle line:84% And OK, maybe I'm going to stay off gamefly.com in the month 00:51:49.120 --> 00:51:51.620 align:middle line:84% before I'm trying to buy Netflix to get my price to go down. 00:51:51.620 --> 00:51:54.162 align:middle line:84% But that's just-- they're only putting three times the weight 00:51:54.162 --> 00:51:59.460 align:middle line:84% on that as they use on these thousands of other websites. 00:51:59.460 --> 00:52:02.250 align:middle line:84% And so you really would have to do a tremendous amount 00:52:02.250 --> 00:52:04.930 align:middle line:84% of distorting your behavior to try to save money on Netflix. 00:52:04.930 --> 00:52:07.650 align:middle line:84% And given that it's distorting your behavior on 4,000 other 00:52:07.650 --> 00:52:09.870 align:middle line:84% websites to get whatever discount they're giving you, 00:52:09.870 --> 00:52:13.590 align:middle line:84% it's just going to be impossible for consumers to hide who they 00:52:13.590 --> 00:52:17.340 align:middle line:84% are, other than consumers hiding who they are via some 00:52:17.340 --> 00:52:21.387 align:middle line:84% technological means, like some tracking blocker software. 00:52:21.387 --> 00:52:23.220 align:middle line:84% But then again, obviously what we don't know 00:52:23.220 --> 00:52:25.170 align:middle line:84% is, what's the mean willingness to pay in people who 00:52:25.170 --> 00:52:26.483 align:middle line:90% use tracker blocking software? 00:52:26.483 --> 00:52:27.650 align:middle line:90% Do they pay even more money? 00:52:27.650 --> 00:52:31.080 align:middle line:90% 00:52:31.080 --> 00:52:34.020 align:middle line:84% AUDIENCE: Was Netflix able to see people's browser activity 00:52:34.020 --> 00:52:36.000 align:middle line:90% back during this time? 00:52:36.000 --> 00:52:40.460 align:middle line:90% GLENN ELLISON: So yeah. 00:52:40.460 --> 00:52:44.120 align:middle line:84% Obviously, there are many data brokers out there 00:52:44.120 --> 00:52:48.710 align:middle line:84% who stick cookies on your computer 00:52:48.710 --> 00:52:50.190 align:middle line:90% when you visit websites. 00:52:50.190 --> 00:52:53.300 align:middle line:84% And Netflix could have made agreements 00:52:53.300 --> 00:52:55.670 align:middle line:84% with some of these firms to buy their data 00:52:55.670 --> 00:52:57.730 align:middle line:84% and get the tracking data from you. 00:52:57.730 --> 00:53:00.470 align:middle line:90% 00:53:00.470 --> 00:53:04.310 align:middle line:84% I think in 2006, we know Netflix was-- 00:53:04.310 --> 00:53:07.760 align:middle line:84% I mean, they're just basically doing $6.99 if you want one DVD, 00:53:07.760 --> 00:53:08.420 align:middle line:90% $8.99. 00:53:08.420 --> 00:53:12.140 align:middle line:84% They weren't doing any discrimination whatsoever. 00:53:12.140 --> 00:53:16.010 align:middle line:84% I don't know when those data markets developed, but it 00:53:16.010 --> 00:53:17.720 align:middle line:84% certainly wasn't long after this that you 00:53:17.720 --> 00:53:22.363 align:middle line:84% could purchase an awful lot of data about users from many-- 00:53:22.363 --> 00:53:23.780 align:middle line:84% there are many businesses who have 00:53:23.780 --> 00:53:25.880 align:middle line:84% this model of sticking cookies on your machine 00:53:25.880 --> 00:53:27.965 align:middle line:84% and then selling the data about you every time 00:53:27.965 --> 00:53:28.840 align:middle line:90% you visit their site. 00:53:28.840 --> 00:53:29.465 align:middle line:90% AUDIENCE: Sure. 00:53:29.465 --> 00:53:34.860 align:middle line:90% 00:53:34.860 --> 00:53:37.745 align:middle line:84% GLENN ELLISON: Obviously, critical aspects of the paper. 00:53:37.745 --> 00:53:41.368 align:middle line:90% 00:53:41.368 --> 00:53:43.410 align:middle line:84% It's a very nice observation that you can do this 00:53:43.410 --> 00:53:46.440 align:middle line:84% without any price variation, but it takes some strong assumptions 00:53:46.440 --> 00:53:48.270 align:middle line:90% to do this. 00:53:48.270 --> 00:53:50.310 align:middle line:90% I really would like to have-- 00:53:50.310 --> 00:54:00.930 align:middle line:84% he's got-- he had vi minus alpha pj plus xi j plus epsilon ij. 00:54:00.930 --> 00:54:03.810 align:middle line:90% He's got the vi is Xi beta. 00:54:03.810 --> 00:54:06.030 align:middle line:84% You would really like to have the alpha also 00:54:06.030 --> 00:54:08.940 align:middle line:90% be Xi times gamma. 00:54:08.940 --> 00:54:11.880 align:middle line:84% You would really like to be able to get at variation in the alpha 00:54:11.880 --> 00:54:13.505 align:middle line:84% separate from the variation in the v's. 00:54:13.505 --> 00:54:17.954 align:middle line:84% And maybe we would also like a coefficient here too. 00:54:17.954 --> 00:54:20.940 align:middle line:90% Xi times eta times the xi j. 00:54:20.940 --> 00:54:23.520 align:middle line:84% Like do people-- also, does their valuation 00:54:23.520 --> 00:54:25.170 align:middle line:84% for high quality versus low quality 00:54:25.170 --> 00:54:26.920 align:middle line:84% also differ with their demographics? 00:54:26.920 --> 00:54:32.560 align:middle line:84% So you've just got limited things that you can do. 00:54:32.560 --> 00:54:34.740 align:middle line:84% And another thing we have is that, 00:54:34.740 --> 00:54:36.740 align:middle line:84% what are the properties of the utility function? 00:54:36.740 --> 00:54:39.490 align:middle line:84% If you remember, I said, with linear demand curves, 00:54:39.490 --> 00:54:41.410 align:middle line:84% we know that price discrimination is always 00:54:41.410 --> 00:54:44.020 align:middle line:90% bad for welfare. 00:54:44.020 --> 00:54:49.060 align:middle line:84% And so does this model with variation 00:54:49.060 --> 00:54:51.670 align:middle line:84% in v's, does it have properties that 00:54:51.670 --> 00:54:54.722 align:middle line:84% make it so that the consumer surplus doesn't go down much? 00:54:54.722 --> 00:54:56.680 align:middle line:84% Or is that something that he's estimated, or is 00:54:56.680 --> 00:54:59.177 align:middle line:90% that something he's assumed? 00:54:59.177 --> 00:55:01.510 align:middle line:84% I don't know so much about the properties of this demand 00:55:01.510 --> 00:55:03.010 align:middle line:84% system, but in some sense you'd like 00:55:03.010 --> 00:55:05.830 align:middle line:84% to try to get at, what are the properties of the demand system 00:55:05.830 --> 00:55:08.770 align:middle line:90% that can do that or not? 00:55:08.770 --> 00:55:10.597 align:middle line:90% OK. 00:55:10.597 --> 00:55:11.555 align:middle line:90% Questions? 00:55:11.555 --> 00:55:14.420 align:middle line:90% 00:55:14.420 --> 00:55:14.920 align:middle line:90% OK. 00:55:14.920 --> 00:55:20.410 align:middle line:84% So then the final paper I was going to talk about today, Dubé 00:55:20.410 --> 00:55:24.730 align:middle line:90% and Misra, it's an example. 00:55:24.730 --> 00:55:28.750 align:middle line:84% This is sort of very high quality data set, 00:55:28.750 --> 00:55:31.540 align:middle line:90% very high quality econometrics. 00:55:31.540 --> 00:55:35.680 align:middle line:84% Again, the paper you would love to have as a job market paper. 00:55:35.680 --> 00:55:38.800 align:middle line:84% Obviously, a big difficulty of this job market paper-- 00:55:38.800 --> 00:55:41.710 align:middle line:84% this kind of paper is it shows there 00:55:41.710 --> 00:55:45.190 align:middle line:84% is this huge benefit towards data in this world 00:55:45.190 --> 00:55:47.680 align:middle line:84% and having connections with firms 00:55:47.680 --> 00:55:49.930 align:middle line:84% that can generate such data sets for you 00:55:49.930 --> 00:55:53.410 align:middle line:84% and let you be the person who can analyze such data sets. 00:55:53.410 --> 00:55:57.730 align:middle line:84% So the particular application is-- 00:55:57.730 --> 00:56:00.020 align:middle line:84% I don't know if people have heard of-- 00:56:00.020 --> 00:56:01.700 align:middle line:84% we are back in the modern world now. 00:56:01.700 --> 00:56:03.940 align:middle line:90% So this is ZipRecruiter. 00:56:03.940 --> 00:56:07.720 align:middle line:84% ZipRecruiter is an employment website. 00:56:07.720 --> 00:56:09.550 align:middle line:84% If you're a small firm or a big firm 00:56:09.550 --> 00:56:11.800 align:middle line:84% and you want to hire employees, you 00:56:11.800 --> 00:56:14.920 align:middle line:84% can sign up for an account with ZipRecruiter, 00:56:14.920 --> 00:56:17.690 align:middle line:84% and then ask ZipRecruiter to send you 00:56:17.690 --> 00:56:20.300 align:middle line:84% resumes of people who have signed up 00:56:20.300 --> 00:56:24.000 align:middle line:84% on ZipRecruiter saying, I would like to be a software engineer. 00:56:24.000 --> 00:56:25.350 align:middle line:90% I have these qualifications. 00:56:25.350 --> 00:56:28.070 align:middle line:84% You can have ZipRecruiter send you resumes 00:56:28.070 --> 00:56:32.420 align:middle line:84% of people who are currently looking for jobs, who meet 00:56:32.420 --> 00:56:36.170 align:middle line:90% the description for your job. 00:56:36.170 --> 00:56:39.650 align:middle line:84% The way ZipRecruiter works is it's 00:56:39.650 --> 00:56:42.068 align:middle line:90% free to use for job searchers. 00:56:42.068 --> 00:56:44.360 align:middle line:84% If you're a job searcher, you can just post your resume 00:56:44.360 --> 00:56:48.500 align:middle line:90% on ZipRecruiter.com for free. 00:56:48.500 --> 00:56:51.320 align:middle line:84% But then the firms who are looking to hire people 00:56:51.320 --> 00:56:54.980 align:middle line:84% have to pay a monthly fee to be on ZipRecruiter. 00:56:54.980 --> 00:56:59.340 align:middle line:84% The paper is dealing with small employers. 00:56:59.340 --> 00:57:02.690 align:middle line:84% The small firms are firms that had 50 or fewer employees, 00:57:02.690 --> 00:57:05.330 align:middle line:84% reported they were going to search for fewer than one 00:57:05.330 --> 00:57:08.120 align:middle line:84% to three employees in the near future. 00:57:08.120 --> 00:57:10.700 align:middle line:84% And what ZipRecruiter would do with such new firms 00:57:10.700 --> 00:57:15.210 align:middle line:84% when they sign up is charge them a $99 a month subscription fee. 00:57:15.210 --> 00:57:16.980 align:middle line:84% And so you can sign up to ZipRecruiter. 00:57:16.980 --> 00:57:18.170 align:middle line:90% You pay $99 a month. 00:57:18.170 --> 00:57:21.575 align:middle line:84% And then you post the job that you want to get, 00:57:21.575 --> 00:57:23.700 align:middle line:84% and they send you a bunch of resumes that match it. 00:57:23.700 --> 00:57:26.000 align:middle line:84% And then you can try to interview or whatever, 00:57:26.000 --> 00:57:29.100 align:middle line:90% look over those resumes. 00:57:29.100 --> 00:57:32.940 align:middle line:84% This paper, they had internal access to ZipRecruiter. 00:57:32.940 --> 00:57:36.180 align:middle line:84% And actually, ZipRecruiter was willing to run 00:57:36.180 --> 00:57:37.630 align:middle line:90% experiments for them. 00:57:37.630 --> 00:57:42.930 align:middle line:84% You get the sense that this is, in some sense, a joint research 00:57:42.930 --> 00:57:45.990 align:middle line:84% consulting kind of relationship with ZipRecruiter, 00:57:45.990 --> 00:57:50.070 align:middle line:84% where they get ZipRecruiter to let them run experiments 00:57:50.070 --> 00:57:52.892 align:middle line:84% on ZipRecruiter's pricing, and then they report the results 00:57:52.892 --> 00:57:54.600 align:middle line:84% back to ZipRecruiter because ZipRecruiter 00:57:54.600 --> 00:57:57.540 align:middle line:84% was interested in the question of, what price should we 00:57:57.540 --> 00:57:58.040 align:middle line:90% be charging? 00:57:58.040 --> 00:57:59.748 align:middle line:84% Should we be discriminating against firms 00:57:59.748 --> 00:58:02.350 align:middle line:84% in charging some firms more than we're charging other firms? 00:58:02.350 --> 00:58:06.300 align:middle line:84% And can we make more money that way? 00:58:06.300 --> 00:58:11.850 align:middle line:84% So the way it worked is they had this experiment where 00:58:11.850 --> 00:58:14.640 align:middle line:84% if you go to ZipRecruiter on the firm side, 00:58:14.640 --> 00:58:16.150 align:middle line:90% you can't just sign up. 00:58:16.150 --> 00:58:20.310 align:middle line:84% You have to enter in information about your firm. 00:58:20.310 --> 00:58:24.420 align:middle line:84% So what are the things about the firm you have to enter? 00:58:24.420 --> 00:58:27.600 align:middle line:84% One-- oh, I guess not here-- is the number of employees that 00:58:27.600 --> 00:58:29.003 align:middle line:90% you-- this sort of basic-- 00:58:29.003 --> 00:58:30.420 align:middle line:84% to get into the bucket at all, you 00:58:30.420 --> 00:58:32.810 align:middle line:84% had to be the under 50 employees and looking 00:58:32.810 --> 00:58:35.100 align:middle line:90% to hire one to three people. 00:58:35.100 --> 00:58:37.170 align:middle line:84% And then you have lots of features of the jobs, 00:58:37.170 --> 00:58:39.160 align:middle line:84% like where is the job going to be located? 00:58:39.160 --> 00:58:40.960 align:middle line:90% What type of company are you? 00:58:40.960 --> 00:58:42.760 align:middle line:84% How many people do you need to hire? 00:58:42.760 --> 00:58:44.642 align:middle line:90% Does the job provide benefits? 00:58:44.642 --> 00:58:45.850 align:middle line:90% Do you need to have a resume? 00:58:45.850 --> 00:58:49.320 align:middle line:84% Do you provide mental benefits, dental benefits? 00:58:49.320 --> 00:58:51.670 align:middle line:84% What category does the job fall in? 00:58:51.670 --> 00:58:53.610 align:middle line:84% Is it an administrative job, a technical job, 00:58:53.610 --> 00:58:57.150 align:middle line:90% a programming job, whatever? 00:58:57.150 --> 00:59:00.220 align:middle line:84% It's fairly minimal evidence from a menu. 00:59:00.220 --> 00:59:03.060 align:middle line:84% But you've got to do all this before ZipRecruiter will 00:59:03.060 --> 00:59:04.600 align:middle line:90% tell you what the price is. 00:59:04.600 --> 00:59:07.540 align:middle line:84% And obviously, if you told ZipRecruiter I'm a 50,000-person 00:59:07.540 --> 00:59:09.822 align:middle line:84% firm, I'm going to hire hundreds of people a month, 00:59:09.822 --> 00:59:12.280 align:middle line:84% they were already discriminating based on that information. 00:59:12.280 --> 00:59:13.988 align:middle line:84% And they would say, if you want to hire-- 00:59:13.988 --> 00:59:16.330 align:middle line:84% if you're a 10,000-person-- you're MIT, 00:59:16.330 --> 00:59:19.180 align:middle line:84% you've got 15,000 employees, you're going to use ZipRecruiter 00:59:19.180 --> 00:59:22.420 align:middle line:84% to hire 1,000 people a year, you'll get charged many, 00:59:22.420 --> 00:59:26.890 align:middle line:84% many thousands of dollars a month, not $99 a month. 00:59:26.890 --> 00:59:29.620 align:middle line:84% But what they did in the experiment was, 00:59:29.620 --> 00:59:32.020 align:middle line:84% they had a month-long experiment where 00:59:32.020 --> 00:59:34.540 align:middle line:84% when you signed up, if you answered everything that put you 00:59:34.540 --> 00:59:37.390 align:middle line:84% in the low category, the under 50 employee category, 00:59:37.390 --> 00:59:40.010 align:middle line:84% they randomized what they told you the monthly price was 00:59:40.010 --> 00:59:40.510 align:middle line:90% going to be. 00:59:40.510 --> 00:59:41.920 align:middle line:90% And they did it-- 00:59:41.920 --> 00:59:44.570 align:middle line:84% $99 was the price they were currently charging. 00:59:44.570 --> 00:59:46.570 align:middle line:84% But then, in the randomized sample, 00:59:46.570 --> 00:59:49.480 align:middle line:84% some people were charged $19, $39, $59, $79, 00:59:49.480 --> 00:59:52.330 align:middle line:84% going all the way up to $399 a month. 00:59:52.330 --> 00:59:55.900 align:middle line:84% So they have this explicitly randomized experiment 00:59:55.900 --> 00:59:58.360 align:middle line:84% and an explicitly randomized experiment 00:59:58.360 --> 01:00:00.670 align:middle line:84% where they can think about discriminating 01:00:00.670 --> 01:00:04.570 align:middle line:84% based on the 133 features they get out of this. 01:00:04.570 --> 01:00:06.580 align:middle line:90% What are the 133 features? 01:00:06.580 --> 01:00:09.040 align:middle line:84% You get a whole bunch of them out of state dummies. 01:00:09.040 --> 01:00:12.430 align:middle line:84% So you get 50 state dummies and some Canadian province dummies. 01:00:12.430 --> 01:00:14.330 align:middle line:90% You do get job categories. 01:00:14.330 --> 01:00:15.935 align:middle line:90% You get benefits, yes or no. 01:00:15.935 --> 01:00:17.560 align:middle line:84% Some of these are binary, some are not. 01:00:17.560 --> 01:00:19.607 align:middle line:84% Anyway, they had 133, quote, "features" 01:00:19.607 --> 01:00:21.690 align:middle line:84% they could then think about discriminating versus. 01:00:21.690 --> 01:00:26.080 align:middle line:90% 01:00:26.080 --> 01:00:29.000 align:middle line:84% So the experiment, it's a big experiment. 01:00:29.000 --> 01:00:31.450 align:middle line:84% I think the experiment was everyone who came 01:00:31.450 --> 01:00:34.870 align:middle line:84% to ZipRecruiter that month, but that's still a sample of just 01:00:34.870 --> 01:00:40.510 align:middle line:90% 7,867 firms. 01:00:40.510 --> 01:00:42.970 align:middle line:84% Utility model is very similar to what 01:00:42.970 --> 01:00:44.300 align:middle line:90% you saw in the previous paper. 01:00:44.300 --> 01:00:46.690 align:middle line:84% It's simpler here because Netflix 01:00:46.690 --> 01:00:50.980 align:middle line:84% had many different products that needed xi j's to determine which 01:00:50.980 --> 01:00:53.480 align:middle line:90% Netflix DVD level you bought. 01:00:53.480 --> 01:00:54.648 align:middle line:90% Here, there's just one. 01:00:54.648 --> 01:00:56.440 align:middle line:84% It's just like you go through this process, 01:00:56.440 --> 01:00:58.040 align:middle line:90% and it's either 1 or 0. 01:00:58.040 --> 01:01:01.870 align:middle line:84% You click, yes, I'll give you my credit card and sign up for $99 01:01:01.870 --> 01:01:04.930 align:middle line:84% a month, or no, I don't want to sign up now. 01:01:04.930 --> 01:01:11.530 align:middle line:84% So the utility is again, Xi beta is your mean utility 01:01:11.530 --> 01:01:16.720 align:middle line:84% given your characteristics, Xi, for signing up for ZipRecruiter. 01:01:16.720 --> 01:01:20.750 align:middle line:84% They then have the alpha coefficient vary across firms. 01:01:20.750 --> 01:01:24.350 align:middle line:84% So you take the Xi multiplied by another coefficient vector 01:01:24.350 --> 01:01:26.450 align:middle line:84% gives you your price sensitivity. 01:01:26.450 --> 01:01:29.090 align:middle line:84% And then you have an epsilon i again 01:01:29.090 --> 01:01:32.000 align:middle line:84% logit error term at the end, where it's just 01:01:32.000 --> 01:01:34.130 align:middle line:84% some consumers with the same characteristics 01:01:34.130 --> 01:01:37.970 align:middle line:84% are willing to pay more or less to be on ZipRecruiter. 01:01:37.970 --> 01:01:43.310 align:middle line:84% And they estimate this by three different procedures. 01:01:43.310 --> 01:01:47.870 align:middle line:84% One is a pure maximum likelihood estimation to estimate the-- 01:01:47.870 --> 01:01:52.200 align:middle line:84% again, this is 133 parameters here, 133 parameters here. 01:01:52.200 --> 01:01:55.850 align:middle line:84% So 266 parameters to be estimated. 01:01:55.850 --> 01:01:58.680 align:middle line:90% 7,867 customers. 01:01:58.680 --> 01:02:02.620 align:middle line:84% One thing they do is just pure maximum likelihood estimation. 01:02:02.620 --> 01:02:05.990 align:middle line:84% Again, empirically, people worry about over-- 01:02:05.990 --> 01:02:11.030 align:middle line:84% 200 and whatever, 266, is too many parameters to estimate with 01:02:11.030 --> 01:02:12.320 align:middle line:90% 7,000 observations. 01:02:12.320 --> 01:02:14.410 align:middle line:90% You end up overfitting the data. 01:02:14.410 --> 01:02:19.070 align:middle line:84% So they use a Lasso-like MLE procedure again 01:02:19.070 --> 01:02:24.480 align:middle line:84% to estimate this, or maybe it's a pure Lasso procedure. 01:02:24.480 --> 01:02:24.980 align:middle line:90% I forget. 01:02:24.980 --> 01:02:28.040 align:middle line:84% They use a Lasso procedure also to estimate this, 01:02:28.040 --> 01:02:30.500 align:middle line:84% setting more coefficients equal to 0. 01:02:30.500 --> 01:02:35.120 align:middle line:84% And then they also do a Weighted Likelihood Bootstrap Lasso, 01:02:35.120 --> 01:02:36.540 align:middle line:90% which you can look in the paper. 01:02:36.540 --> 01:02:38.490 align:middle line:84% It takes up multiple pages describing 01:02:38.490 --> 01:02:40.320 align:middle line:90% what this procedure is. 01:02:40.320 --> 01:02:43.800 align:middle line:84% The Weighted Likelihood Bootstrap procedure, 01:02:43.800 --> 01:02:47.490 align:middle line:84% the Lasso model essentially sets some number of coefficients 01:02:47.490 --> 01:02:48.790 align:middle line:90% equal to 0. 01:02:48.790 --> 01:02:50.310 align:middle line:84% And all the inference on it is based 01:02:50.310 --> 01:02:51.555 align:middle line:90% on assuming those are all 0. 01:02:51.555 --> 01:02:53.430 align:middle line:84% And then what's the uncertainty in estimating 01:02:53.430 --> 01:02:56.010 align:middle line:90% the rest of the coefficients? 01:02:56.010 --> 01:02:58.950 align:middle line:84% This Weighted Likelihood Bootstrap model 01:02:58.950 --> 01:03:01.470 align:middle line:84% is keeping track of both uncertainty over which 01:03:01.470 --> 01:03:03.510 align:middle line:84% coefficients are nonzero and uncertainty 01:03:03.510 --> 01:03:04.960 align:middle line:90% over the coefficient values. 01:03:04.960 --> 01:03:07.530 align:middle line:84% So it's like giving you a posterior over all 01:03:07.530 --> 01:03:09.600 align:middle line:84% the different models that might be true 01:03:09.600 --> 01:03:12.970 align:middle line:84% and reflecting that you don't know what the correct model is. 01:03:12.970 --> 01:03:14.280 align:middle line:90% You have a posterior over that. 01:03:14.280 --> 01:03:15.697 align:middle line:84% And then you also have a posterior 01:03:15.697 --> 01:03:17.610 align:middle line:84% over coefficients within the correct model 01:03:17.610 --> 01:03:21.383 align:middle line:84% or within each correct and incorrect model. 01:03:21.383 --> 01:03:22.800 align:middle line:84% Typical way these things are done, 01:03:22.800 --> 01:03:24.730 align:middle line:90% you have a held-out sample. 01:03:24.730 --> 01:03:27.250 align:middle line:84% You estimate the model on 90% of the data, 01:03:27.250 --> 01:03:28.680 align:middle line:84% and then see how well it predicts 01:03:28.680 --> 01:03:30.490 align:middle line:90% on the other 10% of the data. 01:03:30.490 --> 01:03:33.630 align:middle line:84% And they say that from that validation kind of exercise, 01:03:33.630 --> 01:03:37.560 align:middle line:84% they find that the Weighted Likelihood Bootstrap 01:03:37.560 --> 01:03:40.300 align:middle line:84% estimates are the best estimates relative to the other two. 01:03:40.300 --> 01:03:42.090 align:middle line:90% That is, they do a-- 01:03:42.090 --> 01:03:45.510 align:middle line:84% the maximum likelihood always fits the data best 01:03:45.510 --> 01:03:46.690 align:middle line:90% on the in-sample part. 01:03:46.690 --> 01:03:48.572 align:middle line:84% But then when you go to the out-sample data, 01:03:48.572 --> 01:03:50.280 align:middle line:84% this one is doing a better job predicting 01:03:50.280 --> 01:03:54.210 align:middle line:84% the out-of-sample data than are the MLE estimates. 01:03:54.210 --> 01:03:58.200 align:middle line:90% So that's what they recommend. 01:03:58.200 --> 01:03:59.270 align:middle line:90% What happens? 01:03:59.270 --> 01:04:03.770 align:middle line:84% So first, this is really a monopoly pricing problem-- 01:04:03.770 --> 01:04:08.198 align:middle line:84% a monopoly pricing observation, not a price 01:04:08.198 --> 01:04:09.240 align:middle line:90% discriminate observation. 01:04:09.240 --> 01:04:13.340 align:middle line:84% But what they find is that ZipRecruiter was not 01:04:13.340 --> 01:04:16.190 align:middle line:84% short-run profit maximizing to these consumers. 01:04:16.190 --> 01:04:19.610 align:middle line:84% This chart here at the bottom is, how much money would they 01:04:19.610 --> 01:04:20.903 align:middle line:90% make? 01:04:20.903 --> 01:04:22.820 align:middle line:84% How much money do they make in the first month 01:04:22.820 --> 01:04:26.030 align:middle line:84% off the customers who got different prices shown to them? 01:04:26.030 --> 01:04:30.500 align:middle line:84% And you show them $19 a month, still only half of them 01:04:30.500 --> 01:04:31.820 align:middle line:90% buy at $19 a month. 01:04:31.820 --> 01:04:35.000 align:middle line:84% So you're only getting like $8 a month in revenue. 01:04:35.000 --> 01:04:39.200 align:middle line:84% As you go up, $39, $59, $79, $99, you 01:04:39.200 --> 01:04:42.600 align:middle line:84% keep making more money because fewer people sign up, 01:04:42.600 --> 01:04:45.420 align:middle line:84% but this price is five times that price. 01:04:45.420 --> 01:04:47.300 align:middle line:84% It's not one fifth the number of customers. 01:04:47.300 --> 01:04:48.680 align:middle line:90% You make more money. 01:04:48.680 --> 01:04:51.380 align:middle line:84% But then you go to the right of where ZipRecruiter 01:04:51.380 --> 01:04:55.280 align:middle line:84% was at the time to $159, $199 to $249, 01:04:55.280 --> 01:04:59.910 align:middle line:84% and it looks like the profits just keep marching up. 01:04:59.910 --> 01:05:04.110 align:middle line:84% So the result is that ZipRecruiter was just 01:05:04.110 --> 01:05:05.610 align:middle line:84% charging too little for its service 01:05:05.610 --> 01:05:08.470 align:middle line:84% from the short-run profit maximization perspective. 01:05:08.470 --> 01:05:12.000 align:middle line:84% And the standard error bars, they're not 01:05:12.000 --> 01:05:14.020 align:middle line:90% insignificant on each of these. 01:05:14.020 --> 01:05:16.920 align:middle line:84% But clearly, if you're fitting a curve through that, 01:05:16.920 --> 01:05:20.520 align:middle line:84% it looks like it goes like this and profits are maximized up 01:05:20.520 --> 01:05:23.240 align:middle line:90% here somewhere. 01:05:23.240 --> 01:05:26.180 align:middle line:84% So anyway, they estimate the optimal price would be 01:05:26.180 --> 01:05:30.440 align:middle line:90% $327 a month, not $99 a month. 01:05:30.440 --> 01:05:32.160 align:middle line:90% OK. 01:05:32.160 --> 01:05:32.660 align:middle line:90% Anyway. 01:05:32.660 --> 01:05:34.680 align:middle line:84% So that's a test of monopoly pricing, 01:05:34.680 --> 01:05:37.290 align:middle line:84% is that ZipRecruiter is not static monopoly pricing 01:05:37.290 --> 01:05:39.360 align:middle line:90% before this experiment is run. 01:05:39.360 --> 01:05:41.820 align:middle line:84% Incidentally, one of the postscripts of the paper 01:05:41.820 --> 01:05:45.630 align:middle line:84% is that seeing these results, ZipRecruiter switched to $249 01:05:45.630 --> 01:05:48.090 align:middle line:90% a month. 01:05:48.090 --> 01:05:50.460 align:middle line:84% So then maybe short-run profit maximization model 01:05:50.460 --> 01:05:52.740 align:middle line:84% does work once you've hired consultants 01:05:52.740 --> 01:05:55.590 align:middle line:90% from Chicago Business School. 01:05:55.590 --> 01:05:58.620 align:middle line:84% So second observation is, there's substantial 01:05:58.620 --> 01:06:00.660 align:middle line:84% heterogeneity, both in price sensitivity 01:06:00.660 --> 01:06:02.500 align:middle line:90% and in the consumer surplus. 01:06:02.500 --> 01:06:04.770 align:middle line:84% So as I said, they've got heterogeneity 01:06:04.770 --> 01:06:07.740 align:middle line:84% both in the v's and in the alphas. 01:06:07.740 --> 01:06:10.270 align:middle line:84% And the paper says there's a lot of both. 01:06:10.270 --> 01:06:15.110 align:middle line:84% So this is the fitted price coefficient, 01:06:15.110 --> 01:06:22.290 align:middle line:84% what I was calling the Xi times theta alpha, the predicted price 01:06:22.290 --> 01:06:23.640 align:middle line:90% coefficient here. 01:06:23.640 --> 01:06:28.180 align:middle line:84% And you can see, on an absolute scale, this is 0.02. 01:06:28.180 --> 01:06:29.770 align:middle line:90% So this is like 0. 01:06:29.770 --> 01:06:34.520 align:middle line:84% So this thing is double the price sensitivity of this. 01:06:34.520 --> 01:06:38.050 align:middle line:84% There's just a great deal of heterogeneity 01:06:38.050 --> 01:06:39.970 align:middle line:90% in that price coefficient. 01:06:39.970 --> 01:06:42.130 align:middle line:84% So there is substantial number of people 01:06:42.130 --> 01:06:43.850 align:middle line:90% who are-- it's around 0.3. 01:06:43.850 --> 01:06:46.703 align:middle line:84% There's a lot who it's around 0.6. 01:06:46.703 --> 01:06:49.120 align:middle line:84% There's a lot of difference in what that price coefficient 01:06:49.120 --> 01:06:50.210 align:middle line:90% seems to be. 01:06:50.210 --> 01:06:53.062 align:middle line:84% And again, this exogenous variation in price 01:06:53.062 --> 01:06:55.020 align:middle line:84% means we can estimate these things pretty well. 01:06:55.020 --> 01:06:57.610 align:middle line:90% 01:06:57.610 --> 01:06:59.410 align:middle line:84% What is the heterogeneity in how much 01:06:59.410 --> 01:07:02.908 align:middle line:84% consumer surplus those who are purchasing are getting? 01:07:02.908 --> 01:07:04.450 align:middle line:84% Again, their estimate is that there's 01:07:04.450 --> 01:07:10.380 align:middle line:84% this big upper tail of values and that some people are getting 01:07:10.380 --> 01:07:12.880 align:middle line:84% very little surplus and are on the margin to buying and not, 01:07:12.880 --> 01:07:16.220 align:middle line:84% but most people are getting at least $50 worth of surplus. 01:07:16.220 --> 01:07:18.880 align:middle line:84% And then there's this thick fat tail 01:07:18.880 --> 01:07:20.590 align:middle line:84% of people getting very high consumer 01:07:20.590 --> 01:07:22.083 align:middle line:90% surplus from this product. 01:07:22.083 --> 01:07:23.500 align:middle line:84% So they're estimating that there's 01:07:23.500 --> 01:07:26.920 align:middle line:84% a tremendous amount of consumer surplus, where 01:07:26.920 --> 01:07:27.880 align:middle line:90% consumer means firm. 01:07:27.880 --> 01:07:33.310 align:middle line:84% But firm surplus from using ZipRecruiter at the time. 01:07:33.310 --> 01:07:34.880 align:middle line:90% How much surplus is there? 01:07:34.880 --> 01:07:37.780 align:middle line:84% Well, they estimate that if you look at what optimal 01:07:37.780 --> 01:07:40.480 align:middle line:84% personalized prices would be, the optimal personalized prices 01:07:40.480 --> 01:07:47.440 align:middle line:84% would range from $126 to $6,292 a month. 01:07:47.440 --> 01:07:50.433 align:middle line:84% The $126 is noteworthy because they were only charging $99. 01:07:50.433 --> 01:07:51.850 align:middle line:84% So this would say, even if someone 01:07:51.850 --> 01:07:53.470 align:middle line:84% has all of the attributes that make 01:07:53.470 --> 01:07:55.990 align:middle line:84% you think they're the most price-sensitive consumer 01:07:55.990 --> 01:07:58.650 align:middle line:84% with the lowest value, what would you charge? 01:07:58.650 --> 01:08:01.330 align:middle line:90% The answer is $126. 01:08:01.330 --> 01:08:05.140 align:middle line:84% $6,292, it's a little heroic there because they 01:08:05.140 --> 01:08:07.760 align:middle line:84% didn't experiment with any price above $399. 01:08:07.760 --> 01:08:11.290 align:middle line:84% So having an experiment where the maximum price is $399 01:08:11.290 --> 01:08:13.180 align:middle line:84% and then you say, OK, given that experiment, 01:08:13.180 --> 01:08:15.580 align:middle line:84% $6,000 is what you should charge that guy, 01:08:15.580 --> 01:08:18.819 align:middle line:84% obviously you've got all kinds of functional form assumptions 01:08:18.819 --> 01:08:21.160 align:middle line:84% for what happens out of sample, and is that price really 01:08:21.160 --> 01:08:22.010 align:middle line:90% correct? 01:08:22.010 --> 01:08:25.750 align:middle line:84% So I think, in fact, when they see things-- 01:08:25.750 --> 01:08:30.010 align:middle line:84% one thing they do say is, the mean price would only be $277, 01:08:30.010 --> 01:08:32.510 align:middle line:84% whereas its $327 with a uniform price. 01:08:32.510 --> 01:08:35.470 align:middle line:84% So the skew is such that you do charge a lot of people 01:08:35.470 --> 01:08:36.580 align:middle line:90% low prices. 01:08:36.580 --> 01:08:41.290 align:middle line:84% And even with the $6,000s mixed in, you do a lower mean price 01:08:41.290 --> 01:08:45.850 align:middle line:84% with discriminating than you do without discriminating. 01:08:45.850 --> 01:08:50.620 align:middle line:84% And the I think more reasonable treatment of this data 01:08:50.620 --> 01:08:53.979 align:middle line:84% is, let's look at-- suppose we look at the personalized prices 01:08:53.979 --> 01:08:56.120 align:middle line:90% and just cap them all at $499. 01:08:56.120 --> 01:08:57.670 align:middle line:90% And when they actually-- 01:08:57.670 --> 01:08:59.410 align:middle line:84% ZipRecruiter said that to them, is 01:08:59.410 --> 01:09:02.170 align:middle line:84% that we're not going above $499 no matter what you say. 01:09:02.170 --> 01:09:03.670 align:middle line:84% So they have this counterfactual of, 01:09:03.670 --> 01:09:05.620 align:middle line:84% let's suppose they discriminated, 01:09:05.620 --> 01:09:10.750 align:middle line:84% but they let prices go up to $499 based on this prediction. 01:09:10.750 --> 01:09:13.840 align:middle line:84% The model estimates were that discriminating 01:09:13.840 --> 01:09:19.029 align:middle line:84% would increase profits by 8.2% relative to the $327 price. 01:09:19.029 --> 01:09:23.720 align:middle line:84% So if you discriminate, that's the magnitude of discrimination. 01:09:23.720 --> 01:09:28.720 align:middle line:84% So obviously it's much bigger than Shiller's number. 01:09:28.720 --> 01:09:33.300 align:middle line:84% But I think of it as still relatively small. 01:09:33.300 --> 01:09:37.140 align:middle line:90% 01:09:37.140 --> 01:09:39.955 align:middle line:84% What would the optimal personalized prices do? 01:09:39.955 --> 01:09:42.330 align:middle line:84% The optimal personalized prices are the ones that include 01:09:42.330 --> 01:09:44.609 align:middle line:90% the $6,000 numbers in them. 01:09:44.609 --> 01:09:46.620 align:middle line:84% They say the optimal personalized prices 01:09:46.620 --> 01:09:51.060 align:middle line:84% would reduce consumer surplus by about 25% 01:09:51.060 --> 01:09:55.410 align:middle line:84% and would reduce social welfare relative to the uniform pricing. 01:09:55.410 --> 01:09:56.260 align:middle line:90% Why is that? 01:09:56.260 --> 01:09:56.935 align:middle line:90% Well, one thing. 01:09:56.935 --> 01:09:58.560 align:middle line:84% The summary, essentially, of the report 01:09:58.560 --> 01:10:01.200 align:middle line:84% is, under these discriminatory prices, 01:10:01.200 --> 01:10:04.710 align:middle line:84% even though the mean is $277 instead of $327, 01:10:04.710 --> 01:10:06.930 align:middle line:90% demand goes down. 01:10:06.930 --> 01:10:08.400 align:middle line:84% And the reason is, they're taking 01:10:08.400 --> 01:10:12.270 align:middle line:84% all of these super high value people and monopoly 01:10:12.270 --> 01:10:13.570 align:middle line:90% pricing against them. 01:10:13.570 --> 01:10:18.840 align:middle line:84% And as I said, when you do monopoly pricing, 01:10:18.840 --> 01:10:21.360 align:middle line:84% no matter how high demand is, monopoly pricing, 01:10:21.360 --> 01:10:25.075 align:middle line:84% you're always, in some sense, selling to half the consumers. 01:10:25.075 --> 01:10:27.450 align:middle line:84% And so if there are all these high value people out there 01:10:27.450 --> 01:10:29.242 align:middle line:84% willing to pay thousands of dollars a month 01:10:29.242 --> 01:10:31.140 align:middle line:84% and you still only sell to half of them, 01:10:31.140 --> 01:10:33.270 align:middle line:84% there's a huge amount of deadweight loss 01:10:33.270 --> 01:10:34.980 align:middle line:90% that comes from not selling to-- 01:10:34.980 --> 01:10:38.910 align:middle line:84% taking every one of those really high value populations 01:10:38.910 --> 01:10:43.170 align:middle line:84% and not selling to a lot of people in those. 01:10:43.170 --> 01:10:45.960 align:middle line:84% So anyway, the estimate is that consumer surplus goes down 01:10:45.960 --> 01:10:48.210 align:middle line:90% by 25% here. 01:10:48.210 --> 01:10:53.240 align:middle line:84% And so total social welfare has gone down. 01:10:53.240 --> 01:10:56.880 align:middle line:90% Again, profits only go up by 8%. 01:10:56.880 --> 01:10:58.653 align:middle line:84% Consumer surplus goes down by 25%. 01:10:58.653 --> 01:11:01.070 align:middle line:84% And then how big was consumer surplus relative to profits? 01:11:01.070 --> 01:11:03.200 align:middle line:84% Well, it looks like consumer surplus was actually 01:11:03.200 --> 01:11:04.780 align:middle line:90% quite big relative to profits. 01:11:04.780 --> 01:11:09.770 align:middle line:90% 01:11:09.770 --> 01:11:11.480 align:middle line:90% OK. 01:11:11.480 --> 01:11:13.850 align:middle line:84% It's a nice paper in that they actually 01:11:13.850 --> 01:11:17.060 align:middle line:84% include my Bergemann-Brooks-Morris triangle 01:11:17.060 --> 01:11:18.330 align:middle line:90% themselves. 01:11:18.330 --> 01:11:19.980 align:middle line:84% So I don't have to draw in the arrow. 01:11:19.980 --> 01:11:23.010 align:middle line:84% So here's what they say personalization does. 01:11:23.010 --> 01:11:24.680 align:middle line:84% So personalization takes you from here, 01:11:24.680 --> 01:11:27.210 align:middle line:84% and it's headed in that direction. 01:11:27.210 --> 01:11:29.000 align:middle line:84% So it's going to the direction where 01:11:29.000 --> 01:11:30.770 align:middle line:90% you're really getting only-- 01:11:30.770 --> 01:11:32.840 align:middle line:90% if this is the gap in there-- 01:11:32.840 --> 01:11:34.230 align:middle line:84% they've done this for their data. 01:11:34.230 --> 01:11:36.597 align:middle line:84% So this is the gap between uninformed monopoly prices 01:11:36.597 --> 01:11:38.930 align:middle line:84% and what they think would happen with first degree price 01:11:38.930 --> 01:11:40.012 align:middle line:90% discrimination. 01:11:40.012 --> 01:11:41.720 align:middle line:84% And so first degree price discrimination, 01:11:41.720 --> 01:11:43.950 align:middle line:84% you'd make a huge increase in profits. 01:11:43.950 --> 01:11:46.280 align:middle line:84% You're only getting a little bit off the floor 01:11:46.280 --> 01:11:49.730 align:middle line:84% by doing this personalization with 133 variables. 01:11:49.730 --> 01:11:52.280 align:middle line:84% And you're getting it in this arrow that's 01:11:52.280 --> 01:11:54.860 align:middle line:84% going very much horizontally, is their claim, 01:11:54.860 --> 01:11:58.280 align:middle line:84% that consumer surplus is just going down rapidly, 01:11:58.280 --> 01:12:02.430 align:middle line:84% and profits is only going up a little in this application. 01:12:02.430 --> 01:12:05.800 align:middle line:90% 01:12:05.800 --> 01:12:08.500 align:middle line:84% So in some sense, this is one where this is-- 01:12:08.500 --> 01:12:12.790 align:middle line:84% it's a positive and negative message from social perspective. 01:12:12.790 --> 01:12:16.520 align:middle line:84% One is that it's saying socially, 01:12:16.520 --> 01:12:19.970 align:middle line:84% price discrimination is really bad in this application. 01:12:19.970 --> 01:12:22.190 align:middle line:84% It's destroying an awful lot of social surplus 01:12:22.190 --> 01:12:24.210 align:middle line:84% just to raise a tiny bit of profit. 01:12:24.210 --> 01:12:26.120 align:middle line:84% But then the positive counterpoint to that 01:12:26.120 --> 01:12:28.562 align:middle line:84% is that given the profit gain is so little, 01:12:28.562 --> 01:12:30.020 align:middle line:84% maybe the firm wouldn't even bother 01:12:30.020 --> 01:12:33.120 align:middle line:84% to do it because it's only an 8% increase in profits. 01:12:33.120 --> 01:12:35.850 align:middle line:84% And what are the other downsides of doing it? 01:12:35.850 --> 01:12:39.230 align:middle line:84% Are you going to pay the consultants to do this and take 01:12:39.230 --> 01:12:42.110 align:middle line:84% whatever political pressure, bad news you might get, 01:12:42.110 --> 01:12:43.110 align:middle line:90% or whatever to do it? 01:12:43.110 --> 01:12:44.990 align:middle line:84% And in fact, follow-up experiment 01:12:44.990 --> 01:12:47.840 align:middle line:84% is, ZipRecruiter, in fact, seeing these results, 01:12:47.840 --> 01:12:49.080 align:middle line:90% did not do it. 01:12:49.080 --> 01:12:51.420 align:middle line:84% ZipRecruiter said, thank you very much. 01:12:51.420 --> 01:12:53.120 align:middle line:84% Good to know our prices are too low. 01:12:53.120 --> 01:12:55.520 align:middle line:90% We'll start charging $249. 01:12:55.520 --> 01:12:57.530 align:middle line:84% And actually, a very nice part of the paper 01:12:57.530 --> 01:12:59.660 align:middle line:84% is that ZipRecruiter then let them come in and do 01:12:59.660 --> 01:13:03.590 align:middle line:84% a follow-on study, judging, you have 01:13:03.590 --> 01:13:05.300 align:middle line:84% all these very out-of-sample predictions 01:13:05.300 --> 01:13:06.660 align:middle line:90% of what things would do. 01:13:06.660 --> 01:13:08.678 align:middle line:84% Let's see you discriminate in practice. 01:13:08.678 --> 01:13:09.720 align:middle line:90% Again, a held-out sample. 01:13:09.720 --> 01:13:12.380 align:middle line:84% So they use experiment number one to predict-- 01:13:12.380 --> 01:13:19.170 align:middle line:84% to find the beta hat or theta alpha hat. 01:13:19.170 --> 01:13:21.720 align:middle line:84% And then that lets them get, what 01:13:21.720 --> 01:13:24.180 align:middle line:90% is pi hat for every consumer i? 01:13:24.180 --> 01:13:27.840 align:middle line:84% Let's see how well does your price discrimination scheme 01:13:27.840 --> 01:13:30.600 align:middle line:84% work, reflecting the fact that there 01:13:30.600 --> 01:13:32.220 align:middle line:84% is estimation error in these things 01:13:32.220 --> 01:13:34.262 align:middle line:84% and you're trying it some number of months later. 01:13:34.262 --> 01:13:36.030 align:middle line:84% Does your model still tell us who's 01:13:36.030 --> 01:13:38.100 align:middle line:90% price sensitive or whatever? 01:13:38.100 --> 01:13:40.360 align:middle line:84% So anyway, they ran a second experiment. 01:13:40.360 --> 01:13:45.100 align:middle line:84% In the second experiment, $249 had been the standard price 01:13:45.100 --> 01:13:46.710 align:middle line:90% before the experiment was run. 01:13:46.710 --> 01:13:49.470 align:middle line:84% They experimented with 25% of the subjects 01:13:49.470 --> 01:13:55.560 align:middle line:84% going back to the $99 price, 25% keeping the $249 price, 01:13:55.560 --> 01:14:00.090 align:middle line:84% and then 50% of the sample being on the personalized prices that 01:14:00.090 --> 01:14:05.670 align:middle line:90% varied from $129 up to $499. 01:14:05.670 --> 01:14:08.370 align:middle line:84% They rounded everything to the nearest 9. 01:14:08.370 --> 01:14:12.130 align:middle line:84% And what you find is that, really, it's 01:14:12.130 --> 01:14:16.205 align:middle line:84% impressive in showing that their model is pretty 01:14:16.205 --> 01:14:18.580 align:middle line:84% accurately predicting what would happen under these three 01:14:18.580 --> 01:14:19.640 align:middle line:90% policies. 01:14:19.640 --> 01:14:22.990 align:middle line:84% So what their prediction was was that the mean probability 01:14:22.990 --> 01:14:28.540 align:middle line:84% of signing up would be 26% in the $99 group, 01:14:28.540 --> 01:14:34.390 align:middle line:84% 15% in the $249 group, and only 14% in the personalized group 01:14:34.390 --> 01:14:36.550 align:middle line:84% where they're charging high prices to some 01:14:36.550 --> 01:14:38.770 align:middle line:90% of these high value populations. 01:14:38.770 --> 01:14:41.920 align:middle line:84% And what they find in the data, just a simple comparison 01:14:41.920 --> 01:14:44.500 align:middle line:84% of means, it was actually only 23% 01:14:44.500 --> 01:14:48.190 align:middle line:84% instead of 26% in the $99 group, it matched 15% 01:14:48.190 --> 01:14:50.590 align:middle line:84% in the $249 group, and it was 15% 01:14:50.590 --> 01:14:53.280 align:middle line:84% versus 14% in the personalized group. 01:14:53.280 --> 01:14:55.230 align:middle line:84% If you look at the revenue predictions, 01:14:55.230 --> 01:14:58.470 align:middle line:84% they were predicting the revenues, per consumer, 01:14:58.470 --> 01:15:01.360 align:middle line:90% would be $25, $38, $41. 01:15:01.360 --> 01:15:06.420 align:middle line:90% It's $22, $38, $41.50. 01:15:06.420 --> 01:15:09.570 align:middle line:84% So I think this experiment gives you a lot of confidence 01:15:09.570 --> 01:15:13.980 align:middle line:84% that, at least in the short run, yes, price discrimination 01:15:13.980 --> 01:15:15.750 align:middle line:84% would raise ZipRecruiter's profits 01:15:15.750 --> 01:15:19.373 align:middle line:84% by 8% using this feasible price discrimination scheme. 01:15:19.373 --> 01:15:22.080 align:middle line:90% 01:15:22.080 --> 01:15:25.387 align:middle line:84% And obviously, ZipRecruiter, I believe, 01:15:25.387 --> 01:15:26.970 align:middle line:84% having seen the results of this, still 01:15:26.970 --> 01:15:29.730 align:middle line:84% just stuck with the $249 price and didn't 01:15:29.730 --> 01:15:32.220 align:middle line:84% go ahead and change and adopt this change 01:15:32.220 --> 01:15:34.145 align:middle line:84% to try to get that 8% out of people. 01:15:34.145 --> 01:15:37.240 align:middle line:90% 01:15:37.240 --> 01:15:38.770 align:middle line:90% OK, some comments. 01:15:38.770 --> 01:15:43.540 align:middle line:84% Again, benefits of working collaboratively with firms 01:15:43.540 --> 01:15:45.400 align:middle line:90% is just huge here. 01:15:45.400 --> 01:15:49.090 align:middle line:84% It's very nice that Shiller could be clever and think 01:15:49.090 --> 01:15:51.903 align:middle line:84% about his identification ability and the ability 01:15:51.903 --> 01:15:53.320 align:middle line:84% to do this without any price data, 01:15:53.320 --> 01:15:57.040 align:middle line:90% but price variation is good. 01:15:57.040 --> 01:15:59.740 align:middle line:84% So obviously, can you walk into ZipRecruiter 01:15:59.740 --> 01:16:03.190 align:middle line:84% and get them to run this experiment for you? 01:16:03.190 --> 01:16:03.920 align:middle line:90% Probably not. 01:16:03.920 --> 01:16:07.240 align:middle line:84% But if you have friends who have internet startups that are still 01:16:07.240 --> 01:16:09.490 align:middle line:84% small enough they would rather hire a graduate student 01:16:09.490 --> 01:16:12.460 align:middle line:84% than some Chicago Business School marketing professors 01:16:12.460 --> 01:16:15.070 align:middle line:84% to do their experiments for them, 01:16:15.070 --> 01:16:18.370 align:middle line:84% finding those opportunities to get data sets like this 01:16:18.370 --> 01:16:20.080 align:middle line:90% and let someone run experiments. 01:16:20.080 --> 01:16:24.160 align:middle line:84% And often, actually, you can go to MIT alums or people 01:16:24.160 --> 01:16:30.700 align:middle line:84% who work at not quite so fancy companies may help out. 01:16:30.700 --> 01:16:32.585 align:middle line:90% Wayfair is in Boston. 01:16:32.585 --> 01:16:34.460 align:middle line:84% Maybe you have a friend who works at Wayfair. 01:16:34.460 --> 01:16:37.760 align:middle line:84% They can get you an introduction to their data science team. 01:16:37.760 --> 01:16:41.780 align:middle line:84% Could run experiments through firms like that. 01:16:41.780 --> 01:16:46.220 align:middle line:84% I have to say that segmentation analysis here, 01:16:46.220 --> 01:16:48.140 align:middle line:84% that second experiment really shows you 01:16:48.140 --> 01:16:52.050 align:middle line:84% that the Weighted Likelihood Bootstrap Lasso 01:16:52.050 --> 01:16:54.450 align:middle line:84% thing that they're doing seems like that is doing 01:16:54.450 --> 01:16:56.040 align:middle line:84% a good job of segmenting the consumers 01:16:56.040 --> 01:16:57.248 align:middle line:90% and predicting their demands. 01:16:57.248 --> 01:17:00.360 align:middle line:84% And at least they don't give you so much sense of like, 01:17:00.360 --> 01:17:01.920 align:middle line:84% are the price elasticities estimated 01:17:01.920 --> 01:17:04.503 align:middle line:84% from the new data the same as they would be from the old data? 01:17:04.503 --> 01:17:07.050 align:middle line:84% But at least it looks like this is a very good method 01:17:07.050 --> 01:17:08.470 align:middle line:90% for segmenting people. 01:17:08.470 --> 01:17:13.700 align:middle line:84% And if you wanted to think about doing another study, 01:17:13.700 --> 01:17:16.880 align:middle line:84% I think of the applied results as being similar to Shiller's 01:17:16.880 --> 01:17:18.980 align:middle line:90% in that demographics are just-- 01:17:18.980 --> 01:17:21.200 align:middle line:84% they have more demographics than Shiller did. 01:17:21.200 --> 01:17:23.030 align:middle line:84% They're using 133 features, whereas he 01:17:23.030 --> 01:17:25.400 align:middle line:90% was using 33 features. 01:17:25.400 --> 01:17:27.192 align:middle line:84% They're getting more powerful results. 01:17:27.192 --> 01:17:29.150 align:middle line:84% Maybe there's a sense of having done the demand 01:17:29.150 --> 01:17:30.470 align:middle line:90% estimation better. 01:17:30.470 --> 01:17:32.993 align:middle line:90% But it's still not all that big. 01:17:32.993 --> 01:17:36.123 align:middle line:90% 01:17:36.123 --> 01:17:38.540 align:middle line:84% Another concern, as I said, going back to Shiller's paper, 01:17:38.540 --> 01:17:42.530 align:middle line:84% is it's a very nice study, but it is-- 01:17:42.530 --> 01:17:45.740 align:middle line:84% got this maintained assumptions that people 01:17:45.740 --> 01:17:48.770 align:middle line:84% have these logit demands with the logit demand errors, 01:17:48.770 --> 01:17:50.810 align:middle line:84% and it's only heterogeneity in this coefficient 01:17:50.810 --> 01:17:53.690 align:middle line:90% and this coefficient. 01:17:53.690 --> 01:17:56.720 align:middle line:84% Again, we would like to know more about, 01:17:56.720 --> 01:17:58.820 align:middle line:84% to what extent are these predictions 01:17:58.820 --> 01:18:04.050 align:middle line:84% of the model versus results estimated from the model? 01:18:04.050 --> 01:18:06.050 align:middle line:84% I mean, I think it's nice as an empirical result 01:18:06.050 --> 01:18:14.947 align:middle line:84% that Shiller was finding this and Dubé and Misra find this. 01:18:14.947 --> 01:18:17.030 align:middle line:84% It makes you think that maybe this model is fairly 01:18:17.030 --> 01:18:19.460 align:middle line:84% flexible in which direction those arrows could end up 01:18:19.460 --> 01:18:21.170 align:middle line:90% pointing. 01:18:21.170 --> 01:18:25.580 align:middle line:84% But it would be nice to know more about that. 01:18:25.580 --> 01:18:27.590 align:middle line:84% One other thing that's important-- maybe this 01:18:27.590 --> 01:18:30.860 align:middle line:84% is more of a practical consulting point than a paper. 01:18:30.860 --> 01:18:37.120 align:middle line:84% But these experiments only identify 01:18:37.120 --> 01:18:39.440 align:middle line:90% short-run individual effects. 01:18:39.440 --> 01:18:41.440 align:middle line:84% They have a bit in the paper where they look at, 01:18:41.440 --> 01:18:43.910 align:middle line:84% do people persist on ZipRecruiter for more than just 01:18:43.910 --> 01:18:47.560 align:middle line:84% the sign-up month if you charge them $249 instead of $99? 01:18:47.560 --> 01:18:49.630 align:middle line:84% But there are all kinds of things 01:18:49.630 --> 01:18:51.220 align:middle line:84% that these estimates are never going 01:18:51.220 --> 01:18:55.990 align:middle line:84% to-- these sort of what we call A/B testing in models like this 01:18:55.990 --> 01:18:57.700 align:middle line:84% are going to not tell you, which is, 01:18:57.700 --> 01:19:01.030 align:middle line:84% what are the general equilibrium or systemic effects? 01:19:01.030 --> 01:19:03.550 align:middle line:84% Is there a word of mouth customer acquisition effect, 01:19:03.550 --> 01:19:06.550 align:middle line:84% where someone signs up for ZipRecruiter at $99, 01:19:06.550 --> 01:19:08.050 align:middle line:84% they're satisfied with ZipRecruiter, 01:19:08.050 --> 01:19:09.890 align:middle line:84% they recommend ZipRecruiter to someone else. 01:19:09.890 --> 01:19:12.610 align:middle line:84% Whereas the $249 people are signing up for ZipRecruiter, 01:19:12.610 --> 01:19:16.660 align:middle line:84% but they're not telling their friends to sign up for it. 01:19:16.660 --> 01:19:20.170 align:middle line:84% Does the price per quality affect 01:19:20.170 --> 01:19:22.060 align:middle line:84% when the experts are writing articles 01:19:22.060 --> 01:19:27.610 align:middle line:84% in the newspaper about which employment site should your firm 01:19:27.610 --> 01:19:29.350 align:middle line:90% work with. 01:19:29.350 --> 01:19:32.710 align:middle line:84% Is there some possibility that there will be popular outrage? 01:19:32.710 --> 01:19:37.950 align:middle line:84% Also, we've dropped the number of firms here. 01:19:37.950 --> 01:19:45.080 align:middle line:84% It was 23% going down to 15% using the site. 01:19:45.080 --> 01:19:47.330 align:middle line:84% In some sense, that means there are one third fewer 01:19:47.330 --> 01:19:50.960 align:middle line:84% jobs on the site from small new employers. 01:19:50.960 --> 01:19:53.780 align:middle line:84% One third fewer jobs is going to mean fewer resumes 01:19:53.780 --> 01:19:54.840 align:middle line:90% posted to the site. 01:19:54.840 --> 01:19:56.520 align:middle line:84% Fewer resumes posted to the site, 01:19:56.520 --> 01:19:59.030 align:middle line:84% is the thing less valuable to employers? 01:19:59.030 --> 01:20:04.250 align:middle line:84% Is there a spiral down where fewer people sign up for it, 01:20:04.250 --> 01:20:07.340 align:middle line:84% therefore fewer people use the site? 01:20:07.340 --> 01:20:08.870 align:middle line:90% Thinking about it. 01:20:08.870 --> 01:20:10.250 align:middle line:84% Also, I think John is going to do 01:20:10.250 --> 01:20:13.290 align:middle line:84% in a few weeks two-sided market models. 01:20:13.290 --> 01:20:15.110 align:middle line:84% But this is a two-sided market model 01:20:15.110 --> 01:20:17.610 align:middle line:84% where you shouldn't be doing the one-sided monopoly pricing. 01:20:17.610 --> 01:20:20.172 align:middle line:84% You should think about, we're setting our price. 01:20:20.172 --> 01:20:22.880 align:middle line:84% It's a multi-product thing where I might price to firms and price 01:20:22.880 --> 01:20:24.090 align:middle line:90% to consumers. 01:20:24.090 --> 01:20:25.728 align:middle line:84% My price to consumers is fixed at zero, 01:20:25.728 --> 01:20:27.770 align:middle line:84% but I still want more of them to be there to make 01:20:27.770 --> 01:20:29.400 align:middle line:90% it more valuable for the firms. 01:20:29.400 --> 01:20:35.290 align:middle line:84% I should be thinking that multi-side thing into effect. 01:20:35.290 --> 01:20:38.530 align:middle line:84% And actually, final thing is, this 01:20:38.530 --> 01:20:40.570 align:middle line:84% is a feature from several years ago 01:20:40.570 --> 01:20:45.700 align:middle line:84% where Jeff Bezos, back in the early days of Amazon, Amazon 01:20:45.700 --> 01:20:47.920 align:middle line:84% appeared to be price discriminating against consumers 01:20:47.920 --> 01:20:50.560 align:middle line:84% and charging different people different amounts, 01:20:50.560 --> 01:20:52.540 align:middle line:84% just like there's outrage against price 01:20:52.540 --> 01:20:56.980 align:middle line:84% gouging for gasoline during bad weather. 01:20:56.980 --> 01:20:59.685 align:middle line:84% Anyway, Amazon get into a tremendous bad press 01:20:59.685 --> 01:21:01.060 align:middle line:84% for discriminating against people 01:21:01.060 --> 01:21:04.150 align:middle line:84% and charging some people more than others. 01:21:04.150 --> 01:21:07.000 align:middle line:84% And Amazon had to come out and say, OK, we 01:21:07.000 --> 01:21:08.770 align:middle line:84% promise we're never going to do this. 01:21:08.770 --> 01:21:12.460 align:middle line:84% Again, we don't know, are there going to be similar effects-- 01:21:12.460 --> 01:21:15.053 align:middle line:84% is this part of why ZipRecruiter said 8% is not enough, 01:21:15.053 --> 01:21:17.470 align:middle line:84% because they don't want to get caught price discriminating 01:21:17.470 --> 01:21:19.750 align:middle line:90% against people? 01:21:19.750 --> 01:21:20.350 align:middle line:90% OK. 01:21:20.350 --> 01:21:21.980 align:middle line:90% That's what I've got for today. 01:21:21.980 --> 01:21:26.140 align:middle line:84% So Wednesday, I'm back to oligopoly competition. 01:21:26.140 --> 01:21:28.570 align:middle line:90% It's also going to be a very-- 01:21:28.570 --> 01:21:31.140 align:middle line:84% back to a mostly textbook chapter where a lot of it 01:21:31.140 --> 01:21:31.640 align:middle line:90% is old. 01:21:31.640 --> 01:21:33.223 align:middle line:84% I'll get to some new stuff at the end. 01:21:33.223 --> 01:21:34.900 align:middle line:84% But I think actually the new theory 01:21:34.900 --> 01:21:36.800 align:middle line:84% gets pushed into next Monday's lecture 01:21:36.800 --> 01:21:39.860 align:middle line:84% because I'm doing less empirical and more theory 01:21:39.860 --> 01:21:42.010 align:middle line:90% on oligopoly pricing. 01:21:42.010 --> 01:22:00.000 align:middle line:90%