WEBVTT 00:00:00.000 --> 00:00:02.455 align:middle line:90% [SQUEAKING] 00:00:02.455 --> 00:00:04.419 align:middle line:90% [RUSTLING] 00:00:04.419 --> 00:00:09.329 align:middle line:90% [CLICKING] 00:00:09.329 --> 00:00:16.610 align:middle line:90% 00:00:16.610 --> 00:00:17.360 align:middle line:90% GLENN ELLISON: OK. 00:00:17.360 --> 00:00:18.330 align:middle line:90% Welcome back. 00:00:18.330 --> 00:00:20.270 align:middle line:84% So if you remember, last class, I 00:00:20.270 --> 00:00:25.460 align:middle line:84% was talking about search theory, and discussed two main things. 00:00:25.460 --> 00:00:29.750 align:middle line:84% One is that search costs can complement 00:00:29.750 --> 00:00:32.720 align:middle line:84% product differentiation and raise prices in markets. 00:00:32.720 --> 00:00:35.510 align:middle line:84% And the other is that in Bertrand-like markets, 00:00:35.510 --> 00:00:38.270 align:middle line:84% search costs can also lead to price dispersion. 00:00:38.270 --> 00:00:41.885 align:middle line:84% In most of this class, we've had a pattern of two lectures, 00:00:41.885 --> 00:00:43.760 align:middle line:84% one where I do the theory, and then one where 00:00:43.760 --> 00:00:45.410 align:middle line:90% I talk about the empirics. 00:00:45.410 --> 00:00:47.250 align:middle line:90% On search, it's a big topic. 00:00:47.250 --> 00:00:48.450 align:middle line:90% We decided to do three. 00:00:48.450 --> 00:00:53.630 align:middle line:90% So I did the theory last time. 00:00:53.630 --> 00:00:56.030 align:middle line:84% Today I'm going to talk about reduced form 00:00:56.030 --> 00:00:57.938 align:middle line:84% evidence on empirical evidence on search, 00:00:57.938 --> 00:01:00.230 align:middle line:84% and then we're going to have a third lecture on search, 00:01:00.230 --> 00:01:02.772 align:middle line:84% where Tobias is going to talk about the more structural work. 00:01:02.772 --> 00:01:05.600 align:middle line:90% And that will come later. 00:01:05.600 --> 00:01:10.160 align:middle line:84% So anyway, empirical work on search-- 00:01:10.160 --> 00:01:12.770 align:middle line:84% again, why do we do empirical work in IO? 00:01:12.770 --> 00:01:14.390 align:middle line:84% We're often generally trying to do 00:01:14.390 --> 00:01:17.140 align:middle line:84% two different things, or maybe a combination of two things. 00:01:17.140 --> 00:01:19.620 align:middle line:84% One is we're trying to get evidence on the theories, 00:01:19.620 --> 00:01:22.200 align:middle line:84% whether they're applicable, whether they're 00:01:22.200 --> 00:01:24.930 align:middle line:84% what might be correct or incorrect, 00:01:24.930 --> 00:01:26.650 align:middle line:90% or all theories are correct. 00:01:26.650 --> 00:01:28.090 align:middle line:84% But what is it about the theories 00:01:28.090 --> 00:01:30.363 align:middle line:84% we might want to modify to have it fit applications? 00:01:30.363 --> 00:01:33.030 align:middle line:84% And the other thing we're trying to do is just take the theories 00:01:33.030 --> 00:01:35.580 align:middle line:84% and apply them to particular empirical examples 00:01:35.580 --> 00:01:37.950 align:middle line:84% that we care about and estimate parameters of interest, 00:01:37.950 --> 00:01:40.080 align:middle line:90% how big or small effects are. 00:01:40.080 --> 00:01:42.430 align:middle line:84% So I'm going to talk about four papers today. 00:01:42.430 --> 00:01:44.790 align:middle line:84% The first of these is of the just 00:01:44.790 --> 00:01:47.840 align:middle line:84% gaining insights in the theory literature. 00:01:47.840 --> 00:01:50.340 align:middle line:84% Generally, when we're trying to gain insights in the theory, 00:01:50.340 --> 00:01:53.730 align:middle line:84% we're often looking at unusual situations 00:01:53.730 --> 00:01:56.413 align:middle line:84% because we want to look at markets that are particularly 00:01:56.413 --> 00:01:58.080 align:middle line:84% simple, where the theory is particularly 00:01:58.080 --> 00:02:02.490 align:middle line:84% stark to try to understand its predictions, whether they apply. 00:02:02.490 --> 00:02:06.390 align:middle line:84% Particularly, Alan in his paper is talking about-- 00:02:06.390 --> 00:02:11.190 align:middle line:84% trying to understand, do we see the kinds of price dispersion 00:02:11.190 --> 00:02:16.410 align:middle line:84% predicted by Stahl-like models of heterogeneous consumer 00:02:16.410 --> 00:02:17.170 align:middle line:90% search? 00:02:17.170 --> 00:02:22.080 align:middle line:84% And is it true that frequent violations we see of the quote, 00:02:22.080 --> 00:02:25.200 align:middle line:84% "law of one price" are because of consumer search costs leading 00:02:25.200 --> 00:02:26.610 align:middle line:90% to heterogeneity? 00:02:26.610 --> 00:02:28.920 align:middle line:84% And so in this paper, what Alan was trying to do 00:02:28.920 --> 00:02:33.600 align:middle line:84% is think of an application where he could argue that the Bertrand 00:02:33.600 --> 00:02:38.520 align:middle line:84% assumptions fit very, very well and examine in a model 00:02:38.520 --> 00:02:40.900 align:middle line:84% where Bertrand assumptions fit very, very well, 00:02:40.900 --> 00:02:42.390 align:middle line:90% do we see price dispersion? 00:02:42.390 --> 00:02:44.550 align:middle line:84% And can we provide evidence that price dispersion 00:02:44.550 --> 00:02:46.917 align:middle line:84% seems to be due to consumer search costs? 00:02:46.917 --> 00:02:48.750 align:middle line:84% You'll see that this is a paper written when 00:02:48.750 --> 00:02:51.420 align:middle line:90% Alan was a PhD student at MIT. 00:02:51.420 --> 00:02:54.810 align:middle line:84% Very much it's a paper that has sort of minimal data, 00:02:54.810 --> 00:02:57.880 align:middle line:84% self-collected that anyone could write. 00:02:57.880 --> 00:03:01.620 align:middle line:84% So his idea was that we could get evidence-- a good example 00:03:01.620 --> 00:03:04.800 align:middle line:84% to think about Bertrand-like competition 00:03:04.800 --> 00:03:06.832 align:middle line:90% is prescription drugs. 00:03:06.832 --> 00:03:08.790 align:middle line:84% One of the nice things about prescription drugs 00:03:08.790 --> 00:03:12.090 align:middle line:84% is that a bottle of amoxicillin tablets bought 00:03:12.090 --> 00:03:15.450 align:middle line:84% from pharmacy A and a bottle of amoxicillin tablets bought 00:03:15.450 --> 00:03:18.570 align:middle line:84% from pharmacy B should be completely identical products. 00:03:18.570 --> 00:03:19.970 align:middle line:84% People should not-- people should 00:03:19.970 --> 00:03:22.940 align:middle line:84% know that these things are exactly the same product, 00:03:22.940 --> 00:03:25.028 align:middle line:84% that we're just buying it from different stores. 00:03:25.028 --> 00:03:26.570 align:middle line:84% So we could think of people as having 00:03:26.570 --> 00:03:28.472 align:middle line:90% Bertrand-like preferences. 00:03:28.472 --> 00:03:30.680 align:middle line:84% One of the other nice things about prescription drugs 00:03:30.680 --> 00:03:33.300 align:middle line:84% is you're always struggling in empirical papers. 00:03:33.300 --> 00:03:35.690 align:middle line:84% How do I get the number of degrees of freedom 00:03:35.690 --> 00:03:38.220 align:middle line:90% to actually estimate something? 00:03:38.220 --> 00:03:39.890 align:middle line:84% The great thing about prescription drugs 00:03:39.890 --> 00:03:41.510 align:middle line:84% is there are hundreds of different prescription 00:03:41.510 --> 00:03:43.790 align:middle line:84% drugs, hundreds of different prescription drugs given 00:03:43.790 --> 00:03:46.640 align:middle line:84% hundreds of different observations. 00:03:46.640 --> 00:03:49.820 align:middle line:84% So what Alan did for this paper is 00:03:49.820 --> 00:03:52.910 align:middle line:84% if you think about the Northeastern United States, 00:03:52.910 --> 00:03:58.790 align:middle line:84% so we're in Massachusetts, which looks something like this. 00:03:58.790 --> 00:04:02.900 align:middle line:84% You've got Connecticut here, Rhode Island, New York State. 00:04:02.900 --> 00:04:08.990 align:middle line:84% Alan noticed that New York in the late 1990s had a law that 00:04:08.990 --> 00:04:14.870 align:middle line:84% pharmacies had to post all of their retail prices for the 200 00:04:14.870 --> 00:04:18.360 align:middle line:84% most popular prescription drugs on a poster that was 00:04:18.360 --> 00:04:20.980 align:middle line:84% in displayed in large format in the stores. 00:04:20.980 --> 00:04:24.150 align:middle line:84% Therefore, what one could do is go to a pharmacy in New York 00:04:24.150 --> 00:04:27.360 align:middle line:84% and very quickly find out what prices they were charging 00:04:27.360 --> 00:04:29.940 align:middle line:90% for every different drug. 00:04:29.940 --> 00:04:32.880 align:middle line:84% In the 1990s, people did not have phones in their pockets 00:04:32.880 --> 00:04:35.340 align:middle line:90% with high resolution cameras. 00:04:35.340 --> 00:04:37.310 align:middle line:84% So you could not go there and just take 00:04:37.310 --> 00:04:38.560 align:middle line:90% a picture of the price photos. 00:04:38.560 --> 00:04:42.450 align:middle line:84% So what Alan did was go to New York with a clipboard and just 00:04:42.450 --> 00:04:44.250 align:middle line:84% copy down off these price posters 00:04:44.250 --> 00:04:46.930 align:middle line:84% what prices the pharmacies were charging for all the drugs. 00:04:46.930 --> 00:04:51.150 align:middle line:84% And in particular, he does this in two small towns, Middletown 00:04:51.150 --> 00:04:52.960 align:middle line:90% and Newburgh, New York. 00:04:52.960 --> 00:04:54.540 align:middle line:84% And these happen to be located where 00:04:54.540 --> 00:04:57.900 align:middle line:84% if you drive from Boston along the Mass Pike, 00:04:57.900 --> 00:05:00.210 align:middle line:84% and then get on 84, and then go to New York, 00:05:00.210 --> 00:05:04.980 align:middle line:84% these are two of the first towns you come to in New York State. 00:05:04.980 --> 00:05:07.020 align:middle line:84% One thing that was nice about them is he 00:05:07.020 --> 00:05:09.930 align:middle line:84% wanted to get all the prices in a market. 00:05:09.930 --> 00:05:12.910 align:middle line:84% Middletown and Newburgh are both small towns. 00:05:12.910 --> 00:05:15.630 align:middle line:84% They have 10 or 12 pharmacies in them, 00:05:15.630 --> 00:05:18.820 align:middle line:84% and then there are no other pharmacies within a 15-minute 00:05:18.820 --> 00:05:21.460 align:middle line:90% drive or 20-minute drive away. 00:05:21.460 --> 00:05:23.200 align:middle line:84% And so you could think of anyone who 00:05:23.200 --> 00:05:25.300 align:middle line:84% lives in Middletown as shopping-- 00:05:25.300 --> 00:05:27.795 align:middle line:84% having the options to be buying prescription drugs 00:05:27.795 --> 00:05:29.170 align:middle line:84% all the pharmacies in Middletown, 00:05:29.170 --> 00:05:32.060 align:middle line:84% and in Newburgh, they can buy the things in Newburgh. 00:05:32.060 --> 00:05:35.770 align:middle line:84% And I think I've said all these things. 00:05:35.770 --> 00:05:40.420 align:middle line:84% Oh, I guess another thing about how the world was different-- 00:05:40.420 --> 00:05:42.820 align:middle line:84% another thing about how the world was different in late 00:05:42.820 --> 00:05:46.900 align:middle line:84% 1990s is in the late 1990s, most people did not have prescription 00:05:46.900 --> 00:05:47.930 align:middle line:90% drug insurance. 00:05:47.930 --> 00:05:51.670 align:middle line:84% And so at the time, the prices that the pharmacies charged 00:05:51.670 --> 00:05:54.130 align:middle line:84% for a prescription drug was what people-- customers 00:05:54.130 --> 00:05:55.310 align:middle line:90% would be paying. 00:05:55.310 --> 00:05:57.130 align:middle line:84% Customers would have the same incentive 00:05:57.130 --> 00:06:00.123 align:middle line:84% to shop for drugs they would have for any other product. 00:06:00.123 --> 00:06:01.540 align:middle line:84% Obviously, today, many more people 00:06:01.540 --> 00:06:03.440 align:middle line:84% have prescription drug insurance. 00:06:03.440 --> 00:06:05.990 align:middle line:84% The prices at pharmacies would not be what they're paying. 00:06:05.990 --> 00:06:09.340 align:middle line:84% But at the time, this thing had this advantage. 00:06:09.340 --> 00:06:13.070 align:middle line:84% So anyway, here's an example of the study design. 00:06:13.070 --> 00:06:16.220 align:middle line:84% So you can see that-- this is not the actual poster, 00:06:16.220 --> 00:06:19.290 align:middle line:84% but there was a poster that was very much like this. 00:06:19.290 --> 00:06:22.820 align:middle line:84% I think it was roughly 150 drugs, but then many of them 00:06:22.820 --> 00:06:25.560 align:middle line:84% had generic as well as branded versions. 00:06:25.560 --> 00:06:29.077 align:middle line:84% And so you were getting over 200 prescription drugs. 00:06:29.077 --> 00:06:30.660 align:middle line:84% And so there would be this big poster. 00:06:30.660 --> 00:06:33.650 align:middle line:84% It would list Accupril, 10 milligram tablets, Accupril, 20 00:06:33.650 --> 00:06:36.290 align:middle line:90% milligram tablets, and so on. 00:06:36.290 --> 00:06:38.130 align:middle line:84% And then it would have the prices there. 00:06:38.130 --> 00:06:40.460 align:middle line:84% And so Alan could just go into the pharmacies, 00:06:40.460 --> 00:06:42.890 align:middle line:84% copy down the prices listed on this poster, 00:06:42.890 --> 00:06:46.160 align:middle line:84% and get an example of what they cost. 00:06:46.160 --> 00:06:48.885 align:middle line:84% And different drugs have different characteristics. 00:06:48.885 --> 00:06:50.510 align:middle line:84% And so what he's going to be able to do 00:06:50.510 --> 00:06:53.510 align:middle line:84% is not only say how much dispersion is there 00:06:53.510 --> 00:06:57.560 align:middle line:84% within a drug, but then how much dispersion is there 00:06:57.560 --> 00:06:58.818 align:middle line:90% across drugs? 00:06:58.818 --> 00:07:00.860 align:middle line:84% Sorry-- how does the dispersion vary across drugs 00:07:00.860 --> 00:07:03.170 align:middle line:90% with drug characteristics? 00:07:03.170 --> 00:07:05.690 align:middle line:84% One thing that's also nice about this application is 00:07:05.690 --> 00:07:08.450 align:middle line:84% you could say one reason why drug prices should 00:07:08.450 --> 00:07:10.370 align:middle line:84% be different across pharmacies is 00:07:10.370 --> 00:07:14.010 align:middle line:84% because their drugs are bundled with pharmacy amenities. 00:07:14.010 --> 00:07:16.460 align:middle line:84% Some pharmacy has better customer service, 00:07:16.460 --> 00:07:19.500 align:middle line:84% is cleaner, and might be quicker to get in and out of. 00:07:19.500 --> 00:07:22.050 align:middle line:84% People would pay extra to go to that pharmacy. 00:07:22.050 --> 00:07:26.670 align:middle line:84% But the idea is having 200 different drugs, 00:07:26.670 --> 00:07:29.030 align:middle line:84% you can put in pharmacy fixed effects, 00:07:29.030 --> 00:07:31.970 align:middle line:84% and say, the pharmacy may be nicer, 00:07:31.970 --> 00:07:34.760 align:middle line:84% but there's not something like, this pharmacy is a nicer place 00:07:34.760 --> 00:07:37.010 align:middle line:84% to get amoxicillin, and that pharmacy is a nicer place 00:07:37.010 --> 00:07:38.805 align:middle line:90% to get cephalexin. 00:07:38.805 --> 00:07:40.430 align:middle line:84% Pharmacy fixed effects ought to account 00:07:40.430 --> 00:07:42.805 align:middle line:84% for most of the amenities, and lets him control for that. 00:07:42.805 --> 00:07:46.003 align:middle line:90% 00:07:46.003 --> 00:07:47.670 align:middle line:84% Not much sophistication in the analysis, 00:07:47.670 --> 00:07:49.270 align:middle line:84% so I won't go through how it's really done. 00:07:49.270 --> 00:07:51.395 align:middle line:84% But let me just talk about some of the basic facts. 00:07:51.395 --> 00:07:55.560 align:middle line:84% So first basic fact, there's substantial price dispersion. 00:07:55.560 --> 00:07:57.990 align:middle line:84% The difference between the highest and lowest price 00:07:57.990 --> 00:07:59.970 align:middle line:84% for drug across the drugstores in a town 00:07:59.970 --> 00:08:03.060 align:middle line:90% is $13 for the average drug. 00:08:03.060 --> 00:08:05.320 align:middle line:84% Some drugs have less price dispersion. 00:08:05.320 --> 00:08:07.890 align:middle line:90% 10th percentile range is $5. 00:08:07.890 --> 00:08:11.560 align:middle line:90% 95th percentile range is $25. 00:08:11.560 --> 00:08:15.270 align:middle line:84% So substantial price dispersion, obviously, it's not enormous, 00:08:15.270 --> 00:08:20.400 align:middle line:84% but $13 is about what you can think of. 00:08:20.400 --> 00:08:23.510 align:middle line:84% The second finding is that the price dispersion 00:08:23.510 --> 00:08:27.320 align:middle line:84% doesn't appear to be due to pharmacy fixed effects. 00:08:27.320 --> 00:08:29.490 align:middle line:84% He argues this in a couple of ways. 00:08:29.490 --> 00:08:32.539 align:middle line:84% First is just this simple table where 00:08:32.539 --> 00:08:35.669 align:middle line:84% he puts for every pharmacy in these towns, 00:08:35.669 --> 00:08:38.090 align:middle line:84% how often is their price in the bottom third 00:08:38.090 --> 00:08:39.000 align:middle line:90% of the distribution? 00:08:39.000 --> 00:08:40.559 align:middle line:84% Here in Middletown, there are 10. 00:08:40.559 --> 00:08:45.030 align:middle line:84% So how often is their price one of the three lowest? 00:08:45.030 --> 00:08:46.590 align:middle line:90% How often is it in the middle? 00:08:46.590 --> 00:08:48.515 align:middle line:84% How often is it in the three highest? 00:08:48.515 --> 00:08:50.390 align:middle line:84% And for many of the pharmacies, like the ones 00:08:50.390 --> 00:08:52.730 align:middle line:84% I've highlighted, Eckerd, Immediate, and K-Mart, 00:08:52.730 --> 00:08:54.397 align:middle line:84% there's a substantial number of drugs 00:08:54.397 --> 00:08:56.480 align:middle line:84% where they're one of the lowest priced pharmacies, 00:08:56.480 --> 00:08:58.022 align:middle line:84% a substantial number of drugs for one 00:08:58.022 --> 00:08:59.600 align:middle line:84% of the highest priced pharmacies. 00:08:59.600 --> 00:09:01.710 align:middle line:90% There are a couple exceptions. 00:09:01.710 --> 00:09:03.500 align:middle line:84% Rite Aid seems to be a terrible place 00:09:03.500 --> 00:09:05.390 align:middle line:84% to buy drugs if you live in Middletown. 00:09:05.390 --> 00:09:08.330 align:middle line:84% It's almost always one of the highest three prices-- 00:09:08.330 --> 00:09:11.720 align:middle line:84% also, almost one of the highest prices in Newburgh. 00:09:11.720 --> 00:09:13.820 align:middle line:84% Price Chopper and Walmart seem to have 00:09:13.820 --> 00:09:16.330 align:middle line:90% lower prices than the others. 00:09:16.330 --> 00:09:21.580 align:middle line:84% But if you think about how important are the pharmacy fixed 00:09:21.580 --> 00:09:23.920 align:middle line:84% effects, he has a simple table where 00:09:23.920 --> 00:09:29.890 align:middle line:84% you regress the prices of drugs in a pharmacy on drug fixed 00:09:29.890 --> 00:09:30.670 align:middle line:90% effects. 00:09:30.670 --> 00:09:33.620 align:middle line:90% That has an R squared of 0.907. 00:09:33.620 --> 00:09:35.890 align:middle line:84% You include drug and pharmacy effects 00:09:35.890 --> 00:09:38.040 align:middle line:90% as an R squared of 0.938. 00:09:38.040 --> 00:09:41.770 align:middle line:84% So what that's saying is the pharmacy fixed effects 00:09:41.770 --> 00:09:44.410 align:middle line:84% or the amenities are accounting for about 1/3 of the price 00:09:44.410 --> 00:09:48.110 align:middle line:84% dispersion, which leaves 2/3 of the remaining price dispersion. 00:09:48.110 --> 00:09:51.370 align:middle line:84% So 2/3 of the $13 appears to be pure, 00:09:51.370 --> 00:09:54.790 align:middle line:84% idiosyncratic variation across pharmacies, 00:09:54.790 --> 00:09:57.320 align:middle line:84% not correlated with pharmacy fixed effects. 00:09:57.320 --> 00:10:01.690 align:middle line:84% And then Alan also argues in the paper 00:10:01.690 --> 00:10:04.308 align:middle line:84% even that may be an underestimate of how much of it 00:10:04.308 --> 00:10:06.100 align:middle line:84% is pharmacy fixed effects, because he looks 00:10:06.100 --> 00:10:09.280 align:middle line:84% at the expensive pharmacies, like goes to the Rite Aids, 00:10:09.280 --> 00:10:11.260 align:middle line:90% and goes to the Price Chopper. 00:10:11.260 --> 00:10:15.590 align:middle line:84% And his informal evidence of, how friendly is the pharmacist? 00:10:15.590 --> 00:10:16.580 align:middle line:90% How clean is the place? 00:10:16.580 --> 00:10:17.870 align:middle line:90% How long is the line? 00:10:17.870 --> 00:10:20.270 align:middle line:84% He would rather go to Price Chopper than go to Rite Aid. 00:10:20.270 --> 00:10:23.380 align:middle line:84% So in some sense, this high price at Rite Aid 00:10:23.380 --> 00:10:26.380 align:middle line:84% may just be part also of the idiosyncratic price 00:10:26.380 --> 00:10:30.980 align:middle line:84% dispersion, not part of the bundling with amenities. 00:10:30.980 --> 00:10:34.025 align:middle line:90% 00:10:34.025 --> 00:10:37.910 align:middle line:84% The main thing that Alan does to then argue 00:10:37.910 --> 00:10:40.910 align:middle line:84% that this price dispersion is related to search costs 00:10:40.910 --> 00:10:43.610 align:middle line:90% is to sort of-- 00:10:43.610 --> 00:10:46.110 align:middle line:84% you want to think about-- well, let's go back to the theory. 00:10:46.110 --> 00:10:51.320 align:middle line:84% If you remember, what did I teach about price dispersion? 00:10:51.320 --> 00:10:57.440 align:middle line:84% If you have a market and you have-- 00:10:57.440 --> 00:10:59.600 align:middle line:84% if we have very low search costs, 00:10:59.600 --> 00:11:03.140 align:middle line:84% and this is price equals cost, what you would see 00:11:03.140 --> 00:11:05.450 align:middle line:84% is with very low search costs, we 00:11:05.450 --> 00:11:11.750 align:middle line:84% get a distribution that's fairly tightly distributed around-- 00:11:11.750 --> 00:11:13.610 align:middle line:84% we get a distribution of prices that's 00:11:13.610 --> 00:11:18.740 align:middle line:84% fairly tightly distributed around costs, because if anyone 00:11:18.740 --> 00:11:20.750 align:middle line:84% sees a high price, they're going to be tempted 00:11:20.750 --> 00:11:22.520 align:middle line:90% to go back and search again. 00:11:22.520 --> 00:11:26.510 align:middle line:84% That gives us a search cost distribution that's low. 00:11:26.510 --> 00:11:30.380 align:middle line:84% Prices are close to cost, and then not very 00:11:30.380 --> 00:11:32.422 align:middle line:84% dispersed, because if prices were very dispersed, 00:11:32.422 --> 00:11:34.963 align:middle line:84% anyone seeing a high price would just go search again and try 00:11:34.963 --> 00:11:35.770 align:middle line:90% to get a low price. 00:11:35.770 --> 00:11:38.100 align:middle line:84% So we get very little price dispersion. 00:11:38.100 --> 00:11:43.350 align:middle line:84% When we have high search costs, what you get instead 00:11:43.350 --> 00:11:46.650 align:middle line:84% is more dispersed distribution that 00:11:46.650 --> 00:11:48.630 align:middle line:84% may look something like this, where 00:11:48.630 --> 00:11:50.610 align:middle line:90% a few firms set low prices. 00:11:50.610 --> 00:11:52.900 align:middle line:90% Many more firms set high prices. 00:11:52.900 --> 00:11:54.750 align:middle line:84% Even the consumers who see the high prices 00:11:54.750 --> 00:11:56.230 align:middle line:90% don't want to search again. 00:11:56.230 --> 00:11:57.700 align:middle line:90% So we get two different things. 00:11:57.700 --> 00:12:01.110 align:middle line:84% We get a higher average price, and we get more dispersion, 00:12:01.110 --> 00:12:02.580 align:middle line:84% whereas here we get a lower average 00:12:02.580 --> 00:12:05.250 align:middle line:90% price and less dispersion. 00:12:05.250 --> 00:12:07.800 align:middle line:84% So that's what Alan would like to say, is, 00:12:07.800 --> 00:12:11.190 align:middle line:84% can we observe drugs that have lower search costs versus higher 00:12:11.190 --> 00:12:14.220 align:middle line:84% search costs and find out do the ones that 00:12:14.220 --> 00:12:18.485 align:middle line:84% have lower search costs have lower and less dispersed prices? 00:12:18.485 --> 00:12:22.400 align:middle line:90% 00:12:22.400 --> 00:12:24.290 align:middle line:90% It's hard to think-- 00:12:24.290 --> 00:12:26.630 align:middle line:84% you can think of reasons why some drugs may have 00:12:26.630 --> 00:12:27.870 align:middle line:90% higher or lower search costs. 00:12:27.870 --> 00:12:31.250 align:middle line:84% So for instance, drugs that treat conditions 00:12:31.250 --> 00:12:34.970 align:middle line:84% that are acute and very painful or make it difficult 00:12:34.970 --> 00:12:38.120 align:middle line:84% for you to shop could have high search costs. 00:12:38.120 --> 00:12:40.970 align:middle line:84% Drugs that treat sexually transmitted diseases 00:12:40.970 --> 00:12:44.317 align:middle line:84% might have high search costs because you're embarrassed 00:12:44.317 --> 00:12:45.150 align:middle line:90% about the condition. 00:12:45.150 --> 00:12:47.358 align:middle line:84% You don't want to go tell every pharmacy in the town, 00:12:47.358 --> 00:12:49.490 align:middle line:84% call them up, and say, what would your cost be 00:12:49.490 --> 00:12:51.680 align:middle line:90% for buying this prescription? 00:12:51.680 --> 00:12:54.950 align:middle line:84% But what Alan comes up with is this idea 00:12:54.950 --> 00:12:56.610 align:middle line:90% of frequently purchased drugs. 00:12:56.610 --> 00:12:59.030 align:middle line:84% So you have some drugs, like antibiotics, 00:12:59.030 --> 00:13:01.670 align:middle line:84% that you're typically buying and you're going to use once. 00:13:01.670 --> 00:13:05.985 align:middle line:84% You have other drugs, like high blood pressure medications. 00:13:05.985 --> 00:13:07.610 align:middle line:84% High blood pressure medications, almost 00:13:07.610 --> 00:13:10.340 align:middle line:84% everyone who's taking them is going to take them-- 00:13:10.340 --> 00:13:12.900 align:middle line:84% buy them every month for the rest of their lives. 00:13:12.900 --> 00:13:14.930 align:middle line:84% And so if you think about shopping 00:13:14.930 --> 00:13:17.330 align:middle line:84% for an antibiotic versus shopping for high blood pressure 00:13:17.330 --> 00:13:19.973 align:middle line:84% medication, it pays to shop around for the high blood 00:13:19.973 --> 00:13:21.390 align:middle line:84% pressure medication because you're 00:13:21.390 --> 00:13:23.348 align:middle line:84% going to be able to amortize those search costs 00:13:23.348 --> 00:13:24.630 align:middle line:90% over the rest of your life. 00:13:24.630 --> 00:13:26.190 align:middle line:84% Whereas the antibiotics, you're just 00:13:26.190 --> 00:13:28.010 align:middle line:90% going to buy this thing once. 00:13:28.010 --> 00:13:29.650 align:middle line:84% You're never going to buy it again. 00:13:29.650 --> 00:13:31.860 align:middle line:84% The benefit of search is smaller. 00:13:31.860 --> 00:13:36.570 align:middle line:84% So what Alan does is think about frequently purchased drugs 00:13:36.570 --> 00:13:39.780 align:middle line:84% as drugs that have effectively lower search costs 00:13:39.780 --> 00:13:45.030 align:middle line:84% and examine whether the range in prices and the level of prices 00:13:45.030 --> 00:13:48.600 align:middle line:84% are lower for drugs that are purchased more frequently. 00:13:48.600 --> 00:13:50.800 align:middle line:90% The range is what's done here. 00:13:50.800 --> 00:13:53.400 align:middle line:84% So he has four different models-- 00:13:53.400 --> 00:13:56.370 align:middle line:84% four different measures of the range of prices, 00:13:56.370 --> 00:14:00.270 align:middle line:84% the range, standard deviation, the residual range 00:14:00.270 --> 00:14:01.890 align:middle line:84% after controlling for some pharmacy 00:14:01.890 --> 00:14:05.250 align:middle line:84% fixed effects and other things, residual standard deviation. 00:14:05.250 --> 00:14:08.460 align:middle line:84% In all four cases, he shows that the more frequently purchased 00:14:08.460 --> 00:14:12.330 align:middle line:84% drugs do have less price dispersion than the frequently 00:14:12.330 --> 00:14:15.330 align:middle line:90% purchased drugs. 00:14:15.330 --> 00:14:17.380 align:middle line:84% He also chose to have a table for this. 00:14:17.380 --> 00:14:18.330 align:middle line:90% I do not. 00:14:18.330 --> 00:14:22.800 align:middle line:84% He also shows that the average markups for these drugs 00:14:22.800 --> 00:14:25.620 align:middle line:84% relative to published wholesale prices 00:14:25.620 --> 00:14:28.300 align:middle line:84% are lower for the frequently purchased drugs. 00:14:28.300 --> 00:14:30.360 align:middle line:84% And then a third fact, as I said, 00:14:30.360 --> 00:14:33.240 align:middle line:84% we expect for idiosyncratic reasons, some drugs 00:14:33.240 --> 00:14:35.100 align:middle line:84% to have lower or higher search costs 00:14:35.100 --> 00:14:37.570 align:middle line:84% apart from this sort of purchase frequency. 00:14:37.570 --> 00:14:39.210 align:middle line:84% And what he shows is that drugs that 00:14:39.210 --> 00:14:41.460 align:middle line:84% have unexpectedly high average markups 00:14:41.460 --> 00:14:43.920 align:middle line:84% also have unexpectedly high dispersion, 00:14:43.920 --> 00:14:46.950 align:middle line:84% suggesting that there's also unobserved variation in search 00:14:46.950 --> 00:14:49.230 align:middle line:84% costs, and that unobserved variation in search costs 00:14:49.230 --> 00:14:53.280 align:middle line:84% is driving some of the price dispersion and price levels. 00:14:53.280 --> 00:14:57.190 align:middle line:84% Anyway, so I think it's a very nice paper. 00:14:57.190 --> 00:14:59.100 align:middle line:84% It's providing very simple evidence 00:14:59.100 --> 00:15:01.410 align:middle line:84% saying that these theories we have, 00:15:01.410 --> 00:15:04.260 align:middle line:84% like Stahl's model of price dispersion, do 00:15:04.260 --> 00:15:06.030 align:middle line:84% appear to be applicable, and do appear 00:15:06.030 --> 00:15:09.793 align:middle line:84% that the price dispersion they predict is there. 00:15:09.793 --> 00:15:11.460 align:middle line:84% And it appears that the price dispersion 00:15:11.460 --> 00:15:12.575 align:middle line:90% is search cost related. 00:15:12.575 --> 00:15:15.220 align:middle line:90% 00:15:15.220 --> 00:15:18.070 align:middle line:84% Stango and Zinman are on the opposite side 00:15:18.070 --> 00:15:20.890 align:middle line:84% of the primary second motivation, 00:15:20.890 --> 00:15:24.550 align:middle line:84% trying to look at a question that we really care about 00:15:24.550 --> 00:15:26.440 align:middle line:90% and understand. 00:15:26.440 --> 00:15:29.260 align:middle line:84% Does-- do models of price dispersion 00:15:29.260 --> 00:15:31.030 align:middle line:84% help us understand what's going on? 00:15:31.030 --> 00:15:33.370 align:middle line:84% And how big are the magnitudes of price dispersion 00:15:33.370 --> 00:15:35.325 align:middle line:90% in this application? 00:15:35.325 --> 00:15:37.690 align:middle line:84% The particular thing they're talking about 00:15:37.690 --> 00:15:40.150 align:middle line:90% is credit card debt. 00:15:40.150 --> 00:15:44.230 align:middle line:84% Credit card debt is something that many people 00:15:44.230 --> 00:15:45.340 align:middle line:90% feel is unfortunate. 00:15:45.340 --> 00:15:46.570 align:middle line:84% There are many people in the United States 00:15:46.570 --> 00:15:48.278 align:middle line:84% who pay much more on credit card interest 00:15:48.278 --> 00:15:52.330 align:middle line:84% and have gotten themselves into substantial trouble. 00:15:52.330 --> 00:15:57.490 align:middle line:84% There was a set of papers in the 1990s set off by this AER paper 00:15:57.490 --> 00:16:01.340 align:middle line:84% by Larry Ausubel raising what he called the credit card puzzle. 00:16:01.340 --> 00:16:03.370 align:middle line:84% And the question is, why is it that credit card 00:16:03.370 --> 00:16:05.650 align:middle line:84% interest rates are so high, given 00:16:05.650 --> 00:16:08.650 align:middle line:84% that there are a million banks in the United States? 00:16:08.650 --> 00:16:12.400 align:middle line:84% Why is it that this competition between the huge number of banks 00:16:12.400 --> 00:16:15.660 align:middle line:84% does not dissipate profits in the credit card industry? 00:16:15.660 --> 00:16:18.600 align:middle line:84% And how do we get profits that are so high for the firms? 00:16:18.600 --> 00:16:20.660 align:middle line:84% The other thing that Ausubel raised 00:16:20.660 --> 00:16:24.590 align:middle line:84% was actually non-dispersion intuition, 00:16:24.590 --> 00:16:27.290 align:middle line:84% which is that Ausubel reported that not only are credit card 00:16:27.290 --> 00:16:29.493 align:middle line:84% rates very high, but every big bank 00:16:29.493 --> 00:16:31.410 align:middle line:84% charges very high rates on their credit cards. 00:16:31.410 --> 00:16:33.262 align:middle line:90% So there's not price dispersion. 00:16:33.262 --> 00:16:34.970 align:middle line:84% He would have thought-- he'd sort of say, 00:16:34.970 --> 00:16:38.120 align:middle line:84% why is it that all these banks are charging very high interest 00:16:38.120 --> 00:16:38.850 align:middle line:90% rates? 00:16:38.850 --> 00:16:40.880 align:middle line:84% Why doesn't one of them undercut the others 00:16:40.880 --> 00:16:44.960 align:middle line:84% and try to compete with them, bring interest rates down? 00:16:44.960 --> 00:16:46.880 align:middle line:84% Ausubel also reported that rates were fairly 00:16:46.880 --> 00:16:48.170 align:middle line:90% insensitive to interest rates. 00:16:48.170 --> 00:16:50.270 align:middle line:90% So his data covers times when-- 00:16:50.270 --> 00:16:54.260 align:middle line:84% there are times when interest rates on mortgage loans 00:16:54.260 --> 00:16:56.940 align:middle line:84% are 10%, times when those rates are 2%. 00:16:56.940 --> 00:16:59.750 align:middle line:84% When interest rates on Treasury bills and mortgages 00:16:59.750 --> 00:17:02.619 align:middle line:84% are changing, credit card rates seem to be fairly insensitive. 00:17:02.619 --> 00:17:04.369 align:middle line:84% And there was another puzzle he raised is, 00:17:04.369 --> 00:17:06.150 align:middle line:90% why is it these rates are not-- 00:17:06.150 --> 00:17:08.970 align:middle line:84% why is it they're so fixed, and there's so little dispersion? 00:17:08.970 --> 00:17:13.589 align:middle line:84% They don't vary over time or across banks. 00:17:13.589 --> 00:17:16.170 align:middle line:84% So Stango and Zinman are actually 00:17:16.170 --> 00:17:19.530 align:middle line:84% going to argue that the basic facts that Ausubel 00:17:19.530 --> 00:17:22.500 align:middle line:90% had us think about were wrong. 00:17:22.500 --> 00:17:24.359 align:middle line:90% It's not the case-- 00:17:24.359 --> 00:17:28.980 align:middle line:84% it is true that rates across banks tend to be fairly similar, 00:17:28.980 --> 00:17:31.680 align:middle line:84% but what's not true is that rates across individuals 00:17:31.680 --> 00:17:32.350 align:middle line:90% are true. 00:17:32.350 --> 00:17:36.270 align:middle line:84% And so there's this hidden thing that Ausubel was missing, 00:17:36.270 --> 00:17:38.197 align:middle line:84% which was that all of these banks 00:17:38.197 --> 00:17:40.530 align:middle line:84% have high average interest rates, but all of these banks 00:17:40.530 --> 00:17:43.020 align:middle line:84% are also offering many different interest rates 00:17:43.020 --> 00:17:44.303 align:middle line:90% to different clients. 00:17:44.303 --> 00:17:46.470 align:middle line:84% Some of their clients are paying low interest rates, 00:17:46.470 --> 00:17:48.940 align:middle line:84% some of their clients are paying high interest rates. 00:17:48.940 --> 00:17:51.420 align:middle line:84% And so what Stango and Zinman are trying to think about 00:17:51.420 --> 00:17:56.550 align:middle line:84% is document that there's substantial price dispersion, 00:17:56.550 --> 00:18:01.300 align:middle line:84% and think about search costs as a potential explanation for why 00:18:01.300 --> 00:18:03.940 align:middle line:84% it is that we see these high interest rates. 00:18:03.940 --> 00:18:06.940 align:middle line:84% So anyway, some basic facts in their paper, and I 00:18:06.940 --> 00:18:08.990 align:middle line:84% should say, where does their data come from? 00:18:08.990 --> 00:18:11.800 align:middle line:84% So this is not the sort of amateur data collection 00:18:11.800 --> 00:18:13.600 align:middle line:90% we see in Sorensen's paper. 00:18:13.600 --> 00:18:16.510 align:middle line:84% This is the data collection where you make connections 00:18:16.510 --> 00:18:20.570 align:middle line:84% with some firm that has access to very high quality data. 00:18:20.570 --> 00:18:23.200 align:middle line:84% So in particular, there's a company called Lightspeed. 00:18:23.200 --> 00:18:25.190 align:middle line:84% And there are a few companies like this. 00:18:25.190 --> 00:18:32.350 align:middle line:84% What Lightspeed does is it has a panel of consumers who basically 00:18:32.350 --> 00:18:34.390 align:middle line:84% open their lives to Lightspeed in exchange 00:18:34.390 --> 00:18:35.780 align:middle line:90% for some type of payment. 00:18:35.780 --> 00:18:40.970 align:middle line:84% So in particular, there are 4,312 customers in their data. 00:18:40.970 --> 00:18:44.500 align:middle line:84% What these customers do is they give Lightspeed access 00:18:44.500 --> 00:18:46.190 align:middle line:90% to their credit card statements. 00:18:46.190 --> 00:18:50.500 align:middle line:84% So every month, Lightspeed gets their complete credit card 00:18:50.500 --> 00:18:53.140 align:middle line:84% statement, everything they bought, 00:18:53.140 --> 00:18:57.170 align:middle line:84% how much, what they paid, how much interest they paid. 00:18:57.170 --> 00:18:58.275 align:middle line:90% Did they pay a late fee? 00:18:58.275 --> 00:19:00.650 align:middle line:84% It's basically like they've got the credit card statement 00:19:00.650 --> 00:19:04.010 align:middle line:90% for 4,312 people. 00:19:04.010 --> 00:19:07.500 align:middle line:84% They also have information on the people in their panel. 00:19:07.500 --> 00:19:10.555 align:middle line:84% So when people join their panel, they 00:19:10.555 --> 00:19:11.930 align:middle line:84% complete a survey that gives them 00:19:11.930 --> 00:19:13.980 align:middle line:84% a lot of demographic information. 00:19:13.980 --> 00:19:16.402 align:middle line:84% They also know their consumers' credit scores. 00:19:16.402 --> 00:19:17.360 align:middle line:90% They've got this from-- 00:19:17.360 --> 00:19:21.470 align:middle line:84% I don't know-- Experian or some other credit reporting agency. 00:19:21.470 --> 00:19:28.450 align:middle line:84% So they have some measure of the credit worthiness of people, 00:19:28.450 --> 00:19:30.950 align:middle line:84% which you would think would be a driver of the heterogeneity 00:19:30.950 --> 00:19:33.230 align:middle line:90% in interest rates being paid. 00:19:33.230 --> 00:19:36.930 align:middle line:84% So let me start from some basic facts. 00:19:36.930 --> 00:19:40.520 align:middle line:84% The first fact is there's tremendous heterogeneity 00:19:40.520 --> 00:19:43.610 align:middle line:84% in interest rates across consumers. 00:19:43.610 --> 00:19:45.410 align:middle line:84% If you look at the interquartile range, 00:19:45.410 --> 00:19:49.100 align:middle line:84% just between the 75th percentile and the 25th percentile, 00:19:49.100 --> 00:19:53.300 align:middle line:84% it's 800 basis points, or 8 percentage points per year. 00:19:53.300 --> 00:19:55.520 align:middle line:84% This is omitting anyone who has teaser rates 00:19:55.520 --> 00:19:56.490 align:middle line:90% and who pays in full. 00:19:56.490 --> 00:19:59.030 align:middle line:84% So these are people who are paying interest on their credit 00:19:59.030 --> 00:20:00.250 align:middle line:90% cards every month. 00:20:00.250 --> 00:20:02.300 align:middle line:84% There's just tremendous heterogeneity there. 00:20:02.300 --> 00:20:04.942 align:middle line:90% 00:20:04.942 --> 00:20:07.400 align:middle line:84% Of course, there's going to be heterogeneity in credit card 00:20:07.400 --> 00:20:09.200 align:middle line:84% rates, because some people are bad credit risks. 00:20:09.200 --> 00:20:10.650 align:middle line:84% Some people are good credit risks. 00:20:10.650 --> 00:20:12.440 align:middle line:84% You would think if you were running a credit card company, 00:20:12.440 --> 00:20:13.670 align:middle line:84% you would give lower interest rates 00:20:13.670 --> 00:20:15.440 align:middle line:84% to people who are going to pay you back, 00:20:15.440 --> 00:20:17.107 align:middle line:84% a higher interest rate to the people who 00:20:17.107 --> 00:20:18.530 align:middle line:90% you think might default. 00:20:18.530 --> 00:20:22.220 align:middle line:84% What they find is that default risk explains about 40% 00:20:22.220 --> 00:20:25.100 align:middle line:90% of the interest rate variation. 00:20:25.100 --> 00:20:28.220 align:middle line:84% Other factors, like offsetting rewards, 00:20:28.220 --> 00:20:30.990 align:middle line:84% demographics, explain very, very little. 00:20:30.990 --> 00:20:36.500 align:middle line:84% So roughly 60% of this gap is due to what 00:20:36.500 --> 00:20:38.670 align:middle line:84% appears to be pure, idiosyncratic heterogeneity. 00:20:38.670 --> 00:20:40.520 align:middle line:84% People who are similarly situated, equally 00:20:40.520 --> 00:20:42.350 align:middle line:84% likely to pay the bank back, some 00:20:42.350 --> 00:20:44.690 align:middle line:84% are paying much more than others. 00:20:44.690 --> 00:20:46.820 align:middle line:84% And a third fact I'll talk about later 00:20:46.820 --> 00:20:50.060 align:middle line:84% is that there's also substantial within consumer variation 00:20:50.060 --> 00:20:51.000 align:middle line:90% offers received. 00:20:51.000 --> 00:20:52.790 align:middle line:84% They have a separate second data set 00:20:52.790 --> 00:20:55.070 align:middle line:90% where they have people open-- 00:20:55.070 --> 00:20:56.750 align:middle line:84% they know all of the credit card mailers 00:20:56.750 --> 00:20:59.810 align:middle line:84% that are going to particular individuals. 00:20:59.810 --> 00:21:01.040 align:middle line:90% People in their data-- 00:21:01.040 --> 00:21:02.900 align:middle line:84% I don't know if this applies to you-- 00:21:02.900 --> 00:21:05.560 align:middle line:84% get multiple credit card offers per month. 00:21:05.560 --> 00:21:08.242 align:middle line:84% They look within one month what credit card offers people get. 00:21:08.242 --> 00:21:10.200 align:middle line:84% And there's substantial variation in the offers 00:21:10.200 --> 00:21:11.910 align:middle line:90% if you open them up. 00:21:11.910 --> 00:21:13.430 align:middle line:90% So let me just sort of-- 00:21:13.430 --> 00:21:15.420 align:middle line:84% some summary statistics, and this 00:21:15.420 --> 00:21:18.690 align:middle line:84% is a case where I think part of the interest in this paper 00:21:18.690 --> 00:21:22.890 align:middle line:84% is this Lightspeed data gives us this view of what 00:21:22.890 --> 00:21:24.360 align:middle line:90% is it that consumers are doing. 00:21:24.360 --> 00:21:26.670 align:middle line:84% And you have sort of shocking statistics 00:21:26.670 --> 00:21:28.060 align:middle line:90% about problems people have. 00:21:28.060 --> 00:21:34.800 align:middle line:84% So they divide the data into the quartiles of the average balance 00:21:34.800 --> 00:21:38.160 align:middle line:84% that people have on their credit card. 00:21:38.160 --> 00:21:40.500 align:middle line:84% You were probably told that it's good practice 00:21:40.500 --> 00:21:43.560 align:middle line:84% to pay off your credit card bills every month, 00:21:43.560 --> 00:21:46.160 align:middle line:90% and simply not pay any interest. 00:21:46.160 --> 00:21:51.960 align:middle line:84% And what they find is that the 25th percentile balance is $499, 00:21:51.960 --> 00:21:55.590 align:middle line:84% which means the fraction of people in their data who pay off 00:21:55.590 --> 00:21:59.860 align:middle line:84% their bill every month is less than 25% So it's close to 25%, 00:21:59.860 --> 00:22:02.850 align:middle line:84% but most people have a balance on their credit cards, 00:22:02.850 --> 00:22:05.790 align:middle line:84% and run a balance on their credit cards every month 00:22:05.790 --> 00:22:07.920 align:middle line:90% and never pay it off. 00:22:07.920 --> 00:22:09.820 align:middle line:90% What's the upper quartile? 00:22:09.820 --> 00:22:12.390 align:middle line:84% The upper quartile is people with credit card balances 00:22:12.390 --> 00:22:17.010 align:middle line:90% between $4,586 and $62,000. 00:22:17.010 --> 00:22:18.600 align:middle line:84% People in this quartile, on average, 00:22:18.600 --> 00:22:22.440 align:middle line:84% have a revolving balance of $11,000 on their credit cards. 00:22:22.440 --> 00:22:25.300 align:middle line:84% They're paying $2,000 in interest. 00:22:25.300 --> 00:22:28.230 align:middle line:84% And obviously, these are the people that people really 00:22:28.230 --> 00:22:31.778 align:middle line:84% worry about and people who are making big financial mistakes. 00:22:31.778 --> 00:22:34.320 align:middle line:84% And is there something we can do to help people avoid getting 00:22:34.320 --> 00:22:38.970 align:middle line:84% stuck in this situation where you're spending $2,000 a month 00:22:38.970 --> 00:22:39.820 align:middle line:90% on interest? 00:22:39.820 --> 00:22:42.510 align:middle line:84% You may also notice that the people who are spending $2,000 00:22:42.510 --> 00:22:45.180 align:middle line:84% a month in interest, what's their income? 00:22:45.180 --> 00:22:48.280 align:middle line:84% A quarter of them have incomes below $45,000. 00:22:48.280 --> 00:22:52.290 align:middle line:84% So being income below $45,000, paying $2,000 a month 00:22:52.290 --> 00:22:55.200 align:middle line:84% on your credit card, these are people who've made very big 00:22:55.200 --> 00:22:55.950 align:middle line:90% mistakes. 00:22:55.950 --> 00:22:58.740 align:middle line:84% Some of them have incomes of above $100,000, 00:22:58.740 --> 00:23:02.100 align:middle line:90% but it's still substantial. 00:23:02.100 --> 00:23:04.980 align:middle line:84% Also, noteworthy is in some sense, 00:23:04.980 --> 00:23:07.270 align:middle line:84% some of the U-shaped relationships here. 00:23:07.270 --> 00:23:10.020 align:middle line:84% It's actually people in the bottom quartile 00:23:10.020 --> 00:23:14.070 align:middle line:84% and people in the top quartile who have the best credit scores 00:23:14.070 --> 00:23:16.620 align:middle line:84% and who are charging the most every month. 00:23:16.620 --> 00:23:18.720 align:middle line:84% Some people in the second quartile-- 00:23:18.720 --> 00:23:21.240 align:middle line:84% people in second quartile are spending less per month, 00:23:21.240 --> 00:23:23.898 align:middle line:84% have lower credit reports, lower credit scores. 00:23:23.898 --> 00:23:25.440 align:middle line:84% These people are probably constrained 00:23:25.440 --> 00:23:26.942 align:middle line:90% in how much they can borrow. 00:23:26.942 --> 00:23:28.650 align:middle line:84% But it's the people at the top, actually, 00:23:28.650 --> 00:23:31.980 align:middle line:84% have fairly good credit scores, and the credit card companies 00:23:31.980 --> 00:23:38.240 align:middle line:84% enjoy being able to charge them $2,000 a month in interest. 00:23:38.240 --> 00:23:40.730 align:middle line:84% So this fact I told you about, there's 00:23:40.730 --> 00:23:43.450 align:middle line:84% tremendous dispersion in credit card interest rates. 00:23:43.450 --> 00:23:46.370 align:middle line:90% 00:23:46.370 --> 00:23:51.570 align:middle line:84% Here's the basic raw information on what people are paying. 00:23:51.570 --> 00:23:56.070 align:middle line:84% And you can see it's not based on your revolving balance. 00:23:56.070 --> 00:23:57.830 align:middle line:84% So people in the lowest quartile, 00:23:57.830 --> 00:23:59.450 align:middle line:84% there are people paying between 12%-- 00:23:59.450 --> 00:24:01.888 align:middle line:84% there are people paying 12%, and people paying 26%. 00:24:01.888 --> 00:24:03.180 align:middle line:90% So this is the 10th percentile. 00:24:03.180 --> 00:24:04.680 align:middle line:90% This is the 90th percentile. 00:24:04.680 --> 00:24:07.220 align:middle line:84% People in the upper quartile are people paying 11%, 00:24:07.220 --> 00:24:09.150 align:middle line:90% people paying 26%. 00:24:09.150 --> 00:24:14.270 align:middle line:84% It seems to be very similar across those four groups. 00:24:14.270 --> 00:24:16.640 align:middle line:84% For every level of revolving balance, 00:24:16.640 --> 00:24:20.150 align:middle line:84% there are people paying very low interest rates, people paying 00:24:20.150 --> 00:24:21.670 align:middle line:90% very, very high interest rates. 00:24:21.670 --> 00:24:26.290 align:middle line:90% 00:24:26.290 --> 00:24:27.790 align:middle line:84% As I said, the second fact they have 00:24:27.790 --> 00:24:30.160 align:middle line:84% is, are people paying very different interest rates 00:24:30.160 --> 00:24:32.680 align:middle line:84% because they have very different credit scores? 00:24:32.680 --> 00:24:36.400 align:middle line:84% We see here is the blue dashed line 00:24:36.400 --> 00:24:39.530 align:middle line:84% is the distribution of interest rates 00:24:39.530 --> 00:24:42.710 align:middle line:84% people are being paid relative to the average rate. 00:24:42.710 --> 00:24:44.000 align:middle line:90% So 0 is here. 00:24:44.000 --> 00:24:47.470 align:middle line:84% So there are a number of people paying interest rates 5% 00:24:47.470 --> 00:24:48.410 align:middle line:90% below average. 00:24:48.410 --> 00:24:50.800 align:middle line:84% There are a number of people paying 10% below average. 00:24:50.800 --> 00:24:54.250 align:middle line:84% Also, many people paying 0 to 10% above average. 00:24:54.250 --> 00:24:56.930 align:middle line:84% Not many people pay more than 10% or 12% above average. 00:24:56.930 --> 00:24:59.810 align:middle line:84% But the blue is a distribution of interest rates 00:24:59.810 --> 00:25:03.220 align:middle line:84% people are paying relative to the average. 00:25:03.220 --> 00:25:05.920 align:middle line:84% The red is what you get if you say, 00:25:05.920 --> 00:25:07.960 align:middle line:84% now I'm going to control for your credit score 00:25:07.960 --> 00:25:10.558 align:middle line:84% and ask, once I've controlled for your credit score, 00:25:10.558 --> 00:25:12.850 align:middle line:84% what are you paying relative to the average for someone 00:25:12.850 --> 00:25:15.770 align:middle line:84% with your credit score and other attributes? 00:25:15.770 --> 00:25:20.780 align:middle line:84% And what they find is that this explains some of the variation. 00:25:20.780 --> 00:25:24.500 align:middle line:84% The red distribution is more tightly 00:25:24.500 --> 00:25:27.140 align:middle line:84% distributed than the blue distribution, 00:25:27.140 --> 00:25:31.920 align:middle line:84% but it's explaining perhaps 40% of the variation. 00:25:31.920 --> 00:25:34.940 align:middle line:84% So some of these tails go away, but we still 00:25:34.940 --> 00:25:38.300 align:middle line:84% have very substantial differences in interest rates 00:25:38.300 --> 00:25:42.800 align:middle line:84% paid by people who appear to be similar on all attributes. 00:25:42.800 --> 00:25:47.830 align:middle line:84% So then the next thing the paper does is try to say, 00:25:47.830 --> 00:25:51.560 align:middle line:84% can we tie this to search intensity? 00:25:51.560 --> 00:25:55.400 align:middle line:84% And can we find that people with lower search costs-- 00:25:55.400 --> 00:25:58.790 align:middle line:84% the people who are more likely to search are the ones paying 00:25:58.790 --> 00:26:02.600 align:middle line:84% interest rates down here, and the people who are less-- 00:26:02.600 --> 00:26:05.180 align:middle line:84% have higher search costs end up getting stuck and paying 00:26:05.180 --> 00:26:07.235 align:middle line:90% high interest rates up there? 00:26:07.235 --> 00:26:09.980 align:middle line:90% 00:26:09.980 --> 00:26:12.800 align:middle line:84% It's not as clean as Alan's paper, where 00:26:12.800 --> 00:26:15.710 align:middle line:84% Alan has this exogenous characteristic of a drug that 00:26:15.710 --> 00:26:17.390 align:middle line:90% shifts its search costs. 00:26:17.390 --> 00:26:19.550 align:middle line:84% But what they have as a measure of search costs 00:26:19.550 --> 00:26:23.060 align:middle line:84% is just a self answer to a survey question. 00:26:23.060 --> 00:26:27.812 align:middle line:84% They ask consumers-- so they not only have access 00:26:27.812 --> 00:26:29.020 align:middle line:90% to the Lightspeed panel data. 00:26:29.020 --> 00:26:31.145 align:middle line:84% They have the ability to ask questions of consumers 00:26:31.145 --> 00:26:32.560 align:middle line:90% in the Lightspeed panel data. 00:26:32.560 --> 00:26:34.565 align:middle line:84% And they ask those consumers, how 00:26:34.565 --> 00:26:36.190 align:middle line:84% likely are you to look at a credit card 00:26:36.190 --> 00:26:38.230 align:middle line:90% offer you get in the mail? 00:26:38.230 --> 00:26:39.760 align:middle line:84% And you might think that that would 00:26:39.760 --> 00:26:41.000 align:middle line:90% be a measure of search costs. 00:26:41.000 --> 00:26:43.180 align:middle line:84% People who have low search costs, people 00:26:43.180 --> 00:26:44.890 align:middle line:84% who like opening credit card envelopes, 00:26:44.890 --> 00:26:46.802 align:middle line:84% and look, thinking about them are 00:26:46.802 --> 00:26:48.010 align:middle line:90% people with low search costs. 00:26:48.010 --> 00:26:50.657 align:middle line:84% People who dislike opening their credit card 00:26:50.657 --> 00:26:53.240 align:middle line:84% offers they get in the mail are people with high search costs. 00:26:53.240 --> 00:26:54.792 align:middle line:90% Is that going to explain things? 00:26:54.792 --> 00:26:57.250 align:middle line:84% When you think about that, there's this obvious endogeneity 00:26:57.250 --> 00:26:57.750 align:middle line:90% concern. 00:26:57.750 --> 00:27:01.090 align:middle line:84% You ask people, how likely are you to open a credit card offer 00:27:01.090 --> 00:27:02.260 align:middle line:90% you get in the mail? 00:27:02.260 --> 00:27:04.810 align:middle line:84% Well, if you know you have a bad credit card 00:27:04.810 --> 00:27:06.460 align:middle line:84% and are paying a high interest rate, 00:27:06.460 --> 00:27:08.110 align:middle line:84% you may be more likely to say, yes, 00:27:08.110 --> 00:27:09.460 align:middle line:84% I'm going to open up things, because you're 00:27:09.460 --> 00:27:10.750 align:middle line:90% looking for a new credit card. 00:27:10.750 --> 00:27:13.067 align:middle line:84% Whereas if you already have a really good credit card, 00:27:13.067 --> 00:27:14.650 align:middle line:84% you may say, no, I don't open them up. 00:27:14.650 --> 00:27:17.317 align:middle line:84% And you're not opening them up, not because you have high search 00:27:17.317 --> 00:27:20.650 align:middle line:84% costs, but because you know you already have a very good deal. 00:27:20.650 --> 00:27:22.045 align:middle line:90% So what they do is-- 00:27:22.045 --> 00:27:26.005 align:middle line:90% 00:27:26.005 --> 00:27:29.120 align:middle line:84% the first thing they do is they just run this OLS regression. 00:27:29.120 --> 00:27:31.580 align:middle line:84% And they regress the APR that you end up 00:27:31.580 --> 00:27:36.600 align:middle line:84% paying on your self-reported search intensity. 00:27:36.600 --> 00:27:38.780 align:middle line:84% And we get a small negative coefficient, saying, 00:27:38.780 --> 00:27:43.100 align:middle line:84% people who search more actively are paying slightly lower rates, 00:27:43.100 --> 00:27:44.890 align:middle line:90% but then they sort of-- 00:27:44.890 --> 00:27:49.250 align:middle line:84% it has that endogeneity, so they run this IV regression. 00:27:49.250 --> 00:27:51.890 align:middle line:84% The IV regression says that we can 00:27:51.890 --> 00:27:54.890 align:middle line:84% think about gender and marital status as instruments for search 00:27:54.890 --> 00:28:00.140 align:middle line:84% intensity because the offers that you get in the mail 00:28:00.140 --> 00:28:03.410 align:middle line:84% should not differ across gender, across marital status once you 00:28:03.410 --> 00:28:06.230 align:middle line:84% control for credit score and other factors, 00:28:06.230 --> 00:28:09.800 align:middle line:84% because it's illegal to discriminate against women 00:28:09.800 --> 00:28:12.650 align:middle line:84% or against married or unmarried people. 00:28:12.650 --> 00:28:16.670 align:middle line:84% Therefore, those things should relate to your search intensity, 00:28:16.670 --> 00:28:19.880 align:middle line:84% but be unrelated to the set of raw set of offers 00:28:19.880 --> 00:28:22.430 align:middle line:84% that you get in the mail for credit cards. 00:28:22.430 --> 00:28:25.260 align:middle line:84% When they put that IV in, now they 00:28:25.260 --> 00:28:28.170 align:middle line:84% get a very large negative coefficient, saying, 00:28:28.170 --> 00:28:31.530 align:middle line:84% people who search more intensively 00:28:31.530 --> 00:28:33.520 align:middle line:90% are paying lower interest rates. 00:28:33.520 --> 00:28:38.280 align:middle line:84% I mean, it's a very large gap between the OLS and the IV 00:28:38.280 --> 00:28:38.920 align:middle line:90% estimates. 00:28:38.920 --> 00:28:40.530 align:middle line:84% And then the standard error really 00:28:40.530 --> 00:28:41.890 align:middle line:90% blows up relative to this. 00:28:41.890 --> 00:28:47.920 align:middle line:84% So you wonder-_ again, it's only a slightly significant result. 00:28:47.920 --> 00:28:48.810 align:middle line:90% Why is that? 00:28:48.810 --> 00:28:52.450 align:middle line:84% Is the reverse causality really so large that it causes this? 00:28:52.450 --> 00:28:54.780 align:middle line:84% But anyway, that's the evidence they provide, 00:28:54.780 --> 00:28:57.390 align:middle line:84% that it does seem to be the dispersion 00:28:57.390 --> 00:28:59.250 align:middle line:84% that people are paying is related 00:28:59.250 --> 00:29:01.772 align:middle line:84% to some characteristic that does their search intensity. 00:29:01.772 --> 00:29:03.480 align:middle line:84% So you'll see at the end of this lecture, 00:29:03.480 --> 00:29:05.910 align:middle line:84% I'm going to come back to another paper that further 00:29:05.910 --> 00:29:07.620 align:middle line:84% examines this question, but I think 00:29:07.620 --> 00:29:11.310 align:middle line:84% it does clearly reframe the debate relative to what 00:29:11.310 --> 00:29:14.130 align:middle line:84% you thought from Ausubel and other early papers 00:29:14.130 --> 00:29:17.440 align:middle line:84% to say that there is just substantial price 00:29:17.440 --> 00:29:19.030 align:middle line:90% dispersion in credit cards. 00:29:19.030 --> 00:29:21.160 align:middle line:84% Search costs seem to be a part of why 00:29:21.160 --> 00:29:22.810 align:middle line:84% it is that it's so hard for firms 00:29:22.810 --> 00:29:25.773 align:middle line:84% to compete in this market is if you send out lots of credit card 00:29:25.773 --> 00:29:27.190 align:middle line:84% offers and people don't open them, 00:29:27.190 --> 00:29:29.410 align:middle line:84% people don't consider buying them, 00:29:29.410 --> 00:29:32.110 align:middle line:84% that may make it one of the big factors that's 00:29:32.110 --> 00:29:34.980 align:middle line:84% obstacles to eliminating the high markups that we see. 00:29:34.980 --> 00:29:41.130 align:middle line:90% 00:29:41.130 --> 00:29:43.410 align:middle line:84% So moving on, I'm going to spend much longer 00:29:43.410 --> 00:29:45.510 align:middle line:90% talking about a paper of mine. 00:29:45.510 --> 00:29:47.850 align:middle line:84% I talk about it in part because I know it better. 00:29:47.850 --> 00:29:50.730 align:middle line:84% It's one of my favorite of my papers, 00:29:50.730 --> 00:29:53.910 align:middle line:90% or papers, mine with Sara. 00:29:53.910 --> 00:29:57.270 align:middle line:84% This is a paper where we're both trying to-- 00:29:57.270 --> 00:29:59.460 align:middle line:90% we are trying to-- 00:29:59.460 --> 00:30:01.770 align:middle line:84% it's a more talking about the theory paper 00:30:01.770 --> 00:30:04.950 align:middle line:84% where we are trying to understand what causes 00:30:04.950 --> 00:30:07.680 align:middle line:90% price dispersions and markups. 00:30:07.680 --> 00:30:09.720 align:middle line:84% I see this-- we saw it at the time 00:30:09.720 --> 00:30:12.480 align:middle line:84% as a paper that was looking forward. 00:30:12.480 --> 00:30:16.650 align:middle line:84% The question was that in the early 2000s, search engines, 00:30:16.650 --> 00:30:19.560 align:middle line:84% price search engines had recently been invented. 00:30:19.560 --> 00:30:23.040 align:middle line:84% There were a few markets, like airplane flights, hotels, 00:30:23.040 --> 00:30:27.270 align:middle line:84% rental cars, where many people use price search engines. 00:30:27.270 --> 00:30:31.270 align:middle line:84% Most other markets, people don't still use price search engines. 00:30:31.270 --> 00:30:34.830 align:middle line:84% But our question was, what's going to happen in the world 00:30:34.830 --> 00:30:38.040 align:middle line:84% if price search engines become much better? 00:30:38.040 --> 00:30:41.140 align:middle line:84% First of all, how are the airplane flights, and the hotel, 00:30:41.140 --> 00:30:43.550 align:middle line:84% and rental car markets going to adjust when flights get-- 00:30:43.550 --> 00:30:45.772 align:middle line:84% when search technologies get better? 00:30:45.772 --> 00:30:47.980 align:middle line:84% And what's going to happen to the rest of the economy 00:30:47.980 --> 00:30:49.620 align:middle line:84% when search technologies get better? 00:30:49.620 --> 00:30:52.210 align:middle line:84% I mean, I think somehow this still seems topical 20 years 00:30:52.210 --> 00:30:57.340 align:middle line:84% later, when we see these amazing AI advances and chat robots, 00:30:57.340 --> 00:30:57.910 align:middle line:90% and whatever. 00:30:57.910 --> 00:30:59.620 align:middle line:84% You wonder, am I going be able to-- 00:30:59.620 --> 00:31:01.578 align:middle line:84% are we going to get to a point where I can just 00:31:01.578 --> 00:31:04.870 align:middle line:84% type into Google, what's the best place for me to buy 00:31:04.870 --> 00:31:05.770 align:middle line:90% amoxicillin? 00:31:05.770 --> 00:31:09.020 align:middle line:84% And have amox-- and have Google say, you're living-- 00:31:09.020 --> 00:31:11.180 align:middle line:84% I can see that you live in Newton, Massachusetts. 00:31:11.180 --> 00:31:13.510 align:middle line:84% For you to buy amoxicillin, the best pharmacy to go to 00:31:13.510 --> 00:31:16.762 align:middle line:84% is this one that's on your drive home, 00:31:16.762 --> 00:31:19.220 align:middle line:84% and it costs this much there, and this much somewhere else. 00:31:19.220 --> 00:31:21.490 align:middle line:84% Or are we going to get to the point where I can say, 00:31:21.490 --> 00:31:25.240 align:middle line:84% I want to buy a tennis racket, and then Google just tells me, 00:31:25.240 --> 00:31:26.950 align:middle line:84% given what I know about you, this 00:31:26.950 --> 00:31:28.460 align:middle line:84% would be the best tennis racket for you to buy. 00:31:28.460 --> 00:31:29.900 align:middle line:84% And this store has it at this price. 00:31:29.900 --> 00:31:32.608 align:middle line:84% This is where you should buy it, and just get it shipped directly 00:31:32.608 --> 00:31:33.220 align:middle line:90% to your house. 00:31:33.220 --> 00:31:36.410 align:middle line:84% So who knows whether we will ever get to that point. 00:31:36.410 --> 00:31:38.500 align:middle line:84% But the question, I think, still remains, 00:31:38.500 --> 00:31:41.830 align:middle line:84% how do we expect markets to change when search costs-- 00:31:41.830 --> 00:31:44.680 align:middle line:84% when internet technologies improve. 00:31:44.680 --> 00:31:46.840 align:middle line:90% Search costs go down. 00:31:46.840 --> 00:31:51.040 align:middle line:84% Are we going to see a collapse in price dispersion, where 00:31:51.040 --> 00:31:55.970 align:middle line:84% formerly high dispersed prices become very low tight prices? 00:31:55.970 --> 00:32:00.400 align:middle line:84% And we mentioned in the paper the potential Bertrand paradox. 00:32:00.400 --> 00:32:03.460 align:middle line:84% Real retailers have substantial fixed costs. 00:32:03.460 --> 00:32:06.640 align:middle line:84% The fixed costs of running a store are something like 20%. 00:32:06.640 --> 00:32:10.420 align:middle line:84% So if you can't earn 20% markups over your wholesale costs 00:32:10.420 --> 00:32:12.500 align:middle line:84% on goods, you're going to go out of business. 00:32:12.500 --> 00:32:14.900 align:middle line:84% Even if you're Walmart, you're super efficient, 00:32:14.900 --> 00:32:17.470 align:middle line:84% you need to have 20% markups of your wholesale cost 00:32:17.470 --> 00:32:20.770 align:middle line:84% to cover all the fixed costs of having a physical store, 00:32:20.770 --> 00:32:22.040 align:middle line:90% and things like that. 00:32:22.040 --> 00:32:26.020 align:middle line:84% And so the question is, how is retail and other business going 00:32:26.020 --> 00:32:28.840 align:middle line:84% to survive when the internet makes 00:32:28.840 --> 00:32:30.073 align:middle line:90% price search more efficient? 00:32:30.073 --> 00:32:32.740 align:middle line:84% And I guess, one of the answers we're going to say in this paper 00:32:32.740 --> 00:32:36.310 align:middle line:84% is that, models of search traditionally 00:32:36.310 --> 00:32:38.563 align:middle line:84% thought of consumer search as a one-sided problem, 00:32:38.563 --> 00:32:40.480 align:middle line:84% where these consumers have these search costs. 00:32:40.480 --> 00:32:42.370 align:middle line:90% The search costs are fixed. 00:32:42.370 --> 00:32:44.910 align:middle line:84% And then they search, and that fixed search cost 00:32:44.910 --> 00:32:46.282 align:middle line:90% determines the markups. 00:32:46.282 --> 00:32:48.740 align:middle line:84% The thing we're going to think about is actually two-sided. 00:32:48.740 --> 00:32:50.420 align:middle line:84% You have the search engine is trying 00:32:50.420 --> 00:32:52.420 align:middle line:84% to make search more efficient, and the consumers 00:32:52.420 --> 00:32:54.378 align:middle line:84% are trying to make their search more efficient. 00:32:54.378 --> 00:32:58.090 align:middle line:84% And at the same time, the stores are on the other side, 00:32:58.090 --> 00:32:59.780 align:middle line:84% trying to make search less efficient. 00:32:59.780 --> 00:33:03.792 align:middle line:84% And so we can think of search costs as not being exogenous, 00:33:03.792 --> 00:33:05.500 align:middle line:84% or not being just a result of technology, 00:33:05.500 --> 00:33:08.410 align:middle line:84% but being an outcome of a two-side-- equilibrium 00:33:08.410 --> 00:33:10.092 align:middle line:84% of a two-sided game, where some people 00:33:10.092 --> 00:33:11.800 align:middle line:84% are trying to make search more efficient, 00:33:11.800 --> 00:33:14.420 align:middle line:84% other people are trying to make search less efficient, 00:33:14.420 --> 00:33:16.300 align:middle line:84% and this balance between search and what 00:33:16.300 --> 00:33:19.540 align:middle line:84% we call obfuscation ends up determining price levels 00:33:19.540 --> 00:33:21.360 align:middle line:90% and markets. 00:33:21.360 --> 00:33:24.720 align:middle line:84% So as I said, this is a paper where 00:33:24.720 --> 00:33:28.770 align:middle line:84% we're focused on this theory of, how would obfuscation work? 00:33:28.770 --> 00:33:32.550 align:middle line:84% We go to a particular narrow corner of the internet, 00:33:32.550 --> 00:33:34.793 align:middle line:84% where everything is much simpler, 00:33:34.793 --> 00:33:36.210 align:middle line:84% and it makes it much simpler to do 00:33:36.210 --> 00:33:39.730 align:middle line:84% the analysis than it would if we look at a more popular product. 00:33:39.730 --> 00:33:43.050 align:middle line:84% And so in particular, we look at a price search 00:33:43.050 --> 00:33:45.520 align:middle line:90% engine called pricewatch.com. 00:33:45.520 --> 00:33:48.420 align:middle line:84% I think it still exists, although it's long 00:33:48.420 --> 00:33:49.300 align:middle line:90% past its heyday. 00:33:49.300 --> 00:33:52.800 align:middle line:84% But in the early 2000s, Pricewatch was this simple 00:33:52.800 --> 00:33:58.620 align:middle line:84% database search engine that people who were sophisticated 00:33:58.620 --> 00:34:00.580 align:middle line:84% would use to find things that they needed. 00:34:00.580 --> 00:34:03.000 align:middle line:84% So for instance, if you were to go to the computer 00:34:03.000 --> 00:34:06.060 align:middle line:84% guys in the MIT Economics department and say, 00:34:06.060 --> 00:34:07.410 align:middle line:90% I have a computer. 00:34:07.410 --> 00:34:09.449 align:middle line:90% I'm getting low memory errors. 00:34:09.449 --> 00:34:13.350 align:middle line:84% Could I just buy some extra memory and plug it into the-- 00:34:13.350 --> 00:34:16.020 align:middle line:84% plug some extra memory modules into this part of my circuit 00:34:16.020 --> 00:34:20.610 align:middle line:84% board and increase the memory capacity of my computer? 00:34:20.610 --> 00:34:24.310 align:middle line:84% People who worked in IT offices might use pricewatch.com 00:34:24.310 --> 00:34:27.100 align:middle line:84% to find an inexpensive place to buy memory that you 00:34:27.100 --> 00:34:29.710 align:middle line:90% could add to your computer. 00:34:29.710 --> 00:34:30.670 align:middle line:90% What did it look like? 00:34:30.670 --> 00:34:31.929 align:middle line:90% And again, this is-- 00:34:31.929 --> 00:34:32.900 align:middle line:90% think 2000. 00:34:32.900 --> 00:34:34.960 align:middle line:84% The world was a much more primitive place. 00:34:34.960 --> 00:34:38.980 align:middle line:84% People had lower bandwidth on their home internet. 00:34:38.980 --> 00:34:41.330 align:middle line:84% Things just had simple text-based designs. 00:34:41.330 --> 00:34:42.889 align:middle line:90% So this was Pricewatch-- 00:34:42.889 --> 00:34:45.679 align:middle line:84% very proud that it was established in 1995. 00:34:45.679 --> 00:34:49.150 align:middle line:84% And you can see the pixelated graphics there. 00:34:49.150 --> 00:34:51.429 align:middle line:84% But anyway, you'd go to pricewatch.com. 00:34:51.429 --> 00:34:53.710 align:middle line:84% And it would say, these are the things 00:34:53.710 --> 00:34:54.949 align:middle line:90% you can buy through our site. 00:34:54.949 --> 00:34:58.220 align:middle line:84% You can buy computers, PCs with or without an operating system. 00:34:58.220 --> 00:34:59.110 align:middle line:90% You can buy CPUs. 00:34:59.110 --> 00:35:01.210 align:middle line:90% You can buy motherboards. 00:35:01.210 --> 00:35:03.170 align:middle line:90% We're going to be buying memory. 00:35:03.170 --> 00:35:05.140 align:middle line:84% So RAM memory versus flash cards, 00:35:05.140 --> 00:35:06.560 align:middle line:90% we're going to be buying RAM. 00:35:06.560 --> 00:35:08.230 align:middle line:90% So anyway, I go to this page. 00:35:08.230 --> 00:35:10.480 align:middle line:90% I click on RAM. 00:35:10.480 --> 00:35:12.340 align:middle line:84% When I click on RAM, it brings up 00:35:12.340 --> 00:35:14.860 align:middle line:84% this page, which shows me all the different kinds of RAM 00:35:14.860 --> 00:35:16.300 align:middle line:90% I can buy. 00:35:16.300 --> 00:35:18.730 align:middle line:84% Again, you need to be somewhat sophisticated. 00:35:18.730 --> 00:35:20.450 align:middle line:90% Do you need DDR memory? 00:35:20.450 --> 00:35:23.350 align:middle line:84% Do you need notebook DDR2 memory? 00:35:23.350 --> 00:35:25.318 align:middle line:84% These are all the memory modules you can buy. 00:35:25.318 --> 00:35:27.110 align:middle line:84% They're going to have various descriptions. 00:35:27.110 --> 00:35:30.850 align:middle line:84% So the 512 megabytes, 4 gigabytes, 2 gigabytes, 00:35:30.850 --> 00:35:33.890 align:middle line:84% that's the storage capacity of the memory module. 00:35:33.890 --> 00:35:40.300 align:middle line:84% This 4000, 3500, 3200, those are describing 00:35:40.300 --> 00:35:44.290 align:middle line:84% the speed with which the memory communicates 00:35:44.290 --> 00:35:45.760 align:middle line:90% with the motherboard. 00:35:45.760 --> 00:35:50.080 align:middle line:84% The DDR, DDR2, those are the other aspect of the thing. 00:35:50.080 --> 00:35:53.680 align:middle line:84% Basically, you need to know what type of memory 00:35:53.680 --> 00:35:55.150 align:middle line:90% you need for your computer. 00:35:55.150 --> 00:35:58.490 align:middle line:84% My computer needs DDR PC3200 memory. 00:35:58.490 --> 00:35:59.680 align:middle line:90% It is a choice then. 00:35:59.680 --> 00:36:02.623 align:middle line:84% Do I buy 4 gigabytes, 2 gigabytes, 00:36:02.623 --> 00:36:04.540 align:middle line:84% 1 gigabyte, based on how much I want to spend, 00:36:04.540 --> 00:36:05.990 align:middle line:90% how much memory I want to add? 00:36:05.990 --> 00:36:08.380 align:middle line:84% But the site is designed for people 00:36:08.380 --> 00:36:11.280 align:middle line:84% who are sophisticated in knowing, 00:36:11.280 --> 00:36:15.650 align:middle line:84% what memory do I need for my motherboard? 00:36:15.650 --> 00:36:17.970 align:middle line:84% This is now getting to actual data. 00:36:17.970 --> 00:36:21.110 align:middle line:84% So this was taken a bit later, when 00:36:21.110 --> 00:36:24.050 align:middle line:84% you can see people buying 4 megabyte and 2 megabyte-- 00:36:24.050 --> 00:36:25.670 align:middle line:90% or gigabyte modules. 00:36:25.670 --> 00:36:28.160 align:middle line:84% Back in 1999, what people would buy 00:36:28.160 --> 00:36:30.800 align:middle line:84% is either 128 megabytes of memory 00:36:30.800 --> 00:36:33.900 align:middle line:90% or 256 megabytes of memory. 00:36:33.900 --> 00:36:37.730 align:middle line:84% So I clicked on one particular type of memory, 128 megabyte 00:36:37.730 --> 00:36:39.620 align:middle line:90% modules, that follow the-- 00:36:39.620 --> 00:36:41.480 align:middle line:90% PC100 would be the-- 00:36:41.480 --> 00:36:45.960 align:middle line:84% 100 would be the speed which it communicates with your computer. 00:36:45.960 --> 00:36:47.600 align:middle line:84% So you know you need PC100 memory. 00:36:47.600 --> 00:36:49.370 align:middle line:90% You want to buy 128 megabytes. 00:36:49.370 --> 00:36:50.570 align:middle line:90% You click on this link. 00:36:50.570 --> 00:36:52.490 align:middle line:90% And it brings up this page. 00:36:52.490 --> 00:36:55.370 align:middle line:84% This page just has a set of offers for where 00:36:55.370 --> 00:36:56.790 align:middle line:90% you could buy the memory from. 00:36:56.790 --> 00:36:58.940 align:middle line:84% I can buy it from Computer Craft, Inc, 00:36:58.940 --> 00:37:01.650 align:middle line:84% from Connect Computers, from First Choice Memory. 00:37:01.650 --> 00:37:03.530 align:middle line:84% This retailer is located in Florida. 00:37:03.530 --> 00:37:05.702 align:middle line:84% This one is in California, and so on. 00:37:05.702 --> 00:37:07.160 align:middle line:84% And then the price at which they'll 00:37:07.160 --> 00:37:12.630 align:middle line:84% sell me the memory is 68, 69, 70, 72, 74, 74, 74, 75. 00:37:12.630 --> 00:37:15.630 align:middle line:84% And then how much they'll charge me for shipping, 00:37:15.630 --> 00:37:18.150 align:middle line:84% it's usually-- at the time, Pricewatch had a cap. 00:37:18.150 --> 00:37:20.190 align:middle line:84% You couldn't charge more than $11. 00:37:20.190 --> 00:37:22.100 align:middle line:84% And so many of the firms are some number 00:37:22.100 --> 00:37:25.670 align:middle line:90% that's slightly less than $11. 00:37:25.670 --> 00:37:27.350 align:middle line:84% Again, it was a primitive period, 00:37:27.350 --> 00:37:29.420 align:middle line:84% where if I wanted to buy the memory, 00:37:29.420 --> 00:37:33.110 align:middle line:84% I could then click on Computer Craft Inc.'s website. 00:37:33.110 --> 00:37:35.780 align:middle line:84% It would take me not to-- it's not 00:37:35.780 --> 00:37:38.160 align:middle line:84% like this would add the memory to my cart. 00:37:38.160 --> 00:37:39.450 align:middle line:90% I would click on that button. 00:37:39.450 --> 00:37:41.607 align:middle line:84% I would typically get to the Computer Craft site. 00:37:41.607 --> 00:37:43.940 align:middle line:84% And then I'd have to search around on the Computer Craft 00:37:43.940 --> 00:37:48.240 align:middle line:84% site to find where is their 128 megabyte PC100 memory. 00:37:48.240 --> 00:37:50.760 align:middle line:84% Can I find it at that price of $68 00:37:50.760 --> 00:37:53.780 align:middle line:84% that Pricewatch is telling me that they offer it at. 00:37:53.780 --> 00:37:56.370 align:middle line:84% Pricewatch was not scraping these numbers. 00:37:56.370 --> 00:37:58.640 align:middle line:84% The retailers were having to enter these numbers 00:37:58.640 --> 00:38:00.300 align:middle line:90% into a database at Pricewatch. 00:38:00.300 --> 00:38:01.880 align:middle line:84% So if you're one of these retailers, 00:38:01.880 --> 00:38:04.490 align:middle line:84% you would get up in the morning, decide what your price is, 00:38:04.490 --> 00:38:06.105 align:middle line:84% upload the new price to Pricewatch, 00:38:06.105 --> 00:38:08.605 align:middle line:84% change it on your website, and that's what you would charge. 00:38:08.605 --> 00:38:12.760 align:middle line:90% 00:38:12.760 --> 00:38:15.820 align:middle line:84% So basic observations on Pricewatch 00:38:15.820 --> 00:38:21.010 align:middle line:84% is there was a tremendous amount of competition on Pricewatch. 00:38:21.010 --> 00:38:22.630 align:middle line:90% Prices were very low. 00:38:22.630 --> 00:38:27.110 align:middle line:84% At this time when this memory module cost $68, 00:38:27.110 --> 00:38:29.350 align:middle line:84% if I'd gone to Best Buy to buy the same thing, 00:38:29.350 --> 00:38:30.920 align:middle line:90% I might have paid $100. 00:38:30.920 --> 00:38:34.480 align:middle line:84% So markups were very low on Pricewatch. 00:38:34.480 --> 00:38:36.790 align:middle line:90% But there was some dispersion. 00:38:36.790 --> 00:38:40.990 align:middle line:84% And it was clearly not at all the frictionless ideal of, 00:38:40.990 --> 00:38:42.370 align:middle line:90% I click there. 00:38:42.370 --> 00:38:43.940 align:middle line:84% Everyone's in Bertrand competition. 00:38:43.940 --> 00:38:45.430 align:middle line:90% I get exactly what I want. 00:38:45.430 --> 00:38:48.460 align:middle line:84% Two experiences were common when you used it. 00:38:48.460 --> 00:38:52.650 align:middle line:84% First, as I said, you didn't go directly to the landing page. 00:38:52.650 --> 00:38:55.710 align:middle line:84% Even when you did go directly to the page where you clicked on, 00:38:55.710 --> 00:38:59.310 align:middle line:84% you would find these pages that would be annoying 00:38:59.310 --> 00:39:01.990 align:middle line:84% and time consuming to buy what you wanted to buy. 00:39:01.990 --> 00:39:04.410 align:middle line:84% So this was an example from a bit later, where 00:39:04.410 --> 00:39:07.890 align:middle line:84% I clicked on a offer from topshop.com 00:39:07.890 --> 00:39:12.670 align:middle line:84% to buy some memory module that was supposed to cost $53.81. 00:39:12.670 --> 00:39:16.500 align:middle line:84% I clicked on a link, saying, buy this module for $53.81. 00:39:16.500 --> 00:39:19.350 align:middle line:84% And they give me this page that tells me-- 00:39:19.350 --> 00:39:21.250 align:middle line:84% and there's actually more than this page. 00:39:21.250 --> 00:39:26.440 align:middle line:84% This page saying, price with all the selected options is $90.36. 00:39:26.440 --> 00:39:29.920 align:middle line:84% And so, why is it that it's $90.36? 00:39:29.920 --> 00:39:32.022 align:middle line:84% What they've done is this page, it's 00:39:32.022 --> 00:39:33.480 align:middle line:84% got all these other things that you 00:39:33.480 --> 00:39:36.850 align:middle line:84% can click on to add to the cost of your module. 00:39:36.850 --> 00:39:39.840 align:middle line:84% And some of them are just pure annoyances. 00:39:39.840 --> 00:39:42.030 align:middle line:84% Somehow, they've pre-checked for me "Bonus buy, 00:39:42.030 --> 00:39:44.710 align:middle line:84% 10 pack of hand thumbscrews for my case." 00:39:44.710 --> 00:39:48.030 align:middle line:84% So if I had hand thumbscrews to open up the computer 00:39:48.030 --> 00:39:50.770 align:middle line:84% case in which I wanted to change my memory, 00:39:50.770 --> 00:39:53.600 align:middle line:84% why I would need new thumbscrews, I can't imagine. 00:39:53.600 --> 00:39:56.090 align:middle line:84% But anyway, for $4.95, I can get those 10 thumbscrews. 00:39:56.090 --> 00:39:57.910 align:middle line:84% I just have to go through, read that, 00:39:57.910 --> 00:39:59.210 align:middle line:90% and say, no, I don't want that. 00:39:59.210 --> 00:40:02.680 align:middle line:84% Uncheck that box, and it takes $4.95 off the price. 00:40:02.680 --> 00:40:04.900 align:middle line:84% Others of these are things that it's less clear 00:40:04.900 --> 00:40:05.960 align:middle line:90% that I should uncheck. 00:40:05.960 --> 00:40:09.730 align:middle line:84% So this is-- take advantage of these special offers-- memory 00:40:09.730 --> 00:40:11.650 align:middle line:90% upgrade, CAS 2.5 upgrade. 00:40:11.650 --> 00:40:13.690 align:middle line:84% Improve performance and help things. 00:40:13.690 --> 00:40:16.900 align:middle line:84% They're charging me $6.35 from that. 00:40:16.900 --> 00:40:18.500 align:middle line:90% This is another example-- 00:40:18.500 --> 00:40:21.700 align:middle line:84% six layer stability, more layer, better design-- 00:40:21.700 --> 00:40:23.950 align:middle line:90% not in perfect English, $8.37. 00:40:23.950 --> 00:40:28.070 align:middle line:84% But then they also have these others that are more concerning. 00:40:28.070 --> 00:40:31.950 align:middle line:84% Pre-test, standard pre-test, avoid costly return agreements, 00:40:31.950 --> 00:40:33.170 align:middle line:90% $6.97. 00:40:33.170 --> 00:40:35.390 align:middle line:84% So they're telling me is that if I 00:40:35.390 --> 00:40:39.290 align:middle line:84% want them to test the product-- they're shipping it to me. 00:40:39.290 --> 00:40:41.850 align:middle line:84% Before they ship it to me and make sure it works, 00:40:41.850 --> 00:40:43.555 align:middle line:84% I have to pay $7 to get them to test 00:40:43.555 --> 00:40:45.680 align:middle line:84% the product they're going to sell me, and make sure 00:40:45.680 --> 00:40:46.560 align:middle line:90% that it works. 00:40:46.560 --> 00:40:48.410 align:middle line:84% And if I don't ask them to test it for me 00:40:48.410 --> 00:40:49.970 align:middle line:84% and make sure that it works, they're 00:40:49.970 --> 00:40:52.940 align:middle line:84% going to charge me to ship it back to them. 00:40:52.940 --> 00:40:55.760 align:middle line:84% And so it may be that you look at this offer 00:40:55.760 --> 00:40:58.350 align:middle line:84% and realize this offer is not what I thought this offer was. 00:40:58.350 --> 00:40:59.892 align:middle line:84% I should think of this offer as being 00:40:59.892 --> 00:41:02.690 align:middle line:90% $7 higher than this offer was. 00:41:02.690 --> 00:41:06.710 align:middle line:84% And in some sense, I'm going to have to just spend a long time 00:41:06.710 --> 00:41:09.440 align:middle line:84% reading the fine print of Tough Shop's offers, where I trust, 00:41:09.440 --> 00:41:11.097 align:middle line:90% do I want to buy this product? 00:41:11.097 --> 00:41:12.180 align:middle line:90% Do I not want to buy this? 00:41:12.180 --> 00:41:16.265 align:middle line:84% And it's just going to take me a while to shop. 00:41:16.265 --> 00:41:20.570 align:middle line:84% The second thing that was common was 00:41:20.570 --> 00:41:24.200 align:middle line:84% offers that were not designed to just delay and annoy you, 00:41:24.200 --> 00:41:26.540 align:middle line:84% but ones where it was clearly aimed 00:41:26.540 --> 00:41:29.330 align:middle line:84% to get you to shift away from buying something 00:41:29.330 --> 00:41:30.880 align:middle line:90% you bought into something else. 00:41:30.880 --> 00:41:37.520 align:middle line:84% If you remember back to when I talked about competitive price 00:41:37.520 --> 00:41:39.840 align:middle line:84% dispersion, I talked about add-on pricing. 00:41:39.840 --> 00:41:44.870 align:middle line:84% That is, if you have two firms selling two products, 00:41:44.870 --> 00:41:47.900 align:middle line:84% models of price dispersion were totally well-- 00:41:47.900 --> 00:41:49.730 align:middle line:84% sorry-- models of price discrimination 00:41:49.730 --> 00:41:52.190 align:middle line:84% were totally well with differentiated products 00:41:52.190 --> 00:41:53.280 align:middle line:90% and multiple firms. 00:41:53.280 --> 00:41:56.370 align:middle line:84% And you can have two firms competing against each other. 00:41:56.370 --> 00:41:58.140 align:middle line:90% They advertise all their prices. 00:41:58.140 --> 00:42:01.760 align:middle line:84% They sell low quality goods at low prices, high quality 00:42:01.760 --> 00:42:03.120 align:middle line:90% goods at high prices. 00:42:03.120 --> 00:42:07.460 align:middle line:84% And that can be rational if the consumers for the low quality 00:42:07.460 --> 00:42:09.600 align:middle line:84% goods are more price sensitive across firms. 00:42:09.600 --> 00:42:12.020 align:middle line:84% And so you have an equilibrium where 00:42:12.020 --> 00:42:15.500 align:middle line:84% it's like the competition on two-line model. 00:42:15.500 --> 00:42:18.680 align:middle line:84% For low quality goods, firms are very undifferentiated. 00:42:18.680 --> 00:42:20.820 align:middle line:84% For the high quality goods, they're differentiated. 00:42:20.820 --> 00:42:22.680 align:middle line:84% These goods are sold at a low markup. 00:42:22.680 --> 00:42:25.700 align:middle line:84% These goods are sold at a high markup. 00:42:25.700 --> 00:42:27.680 align:middle line:84% And one thing I said when I talked 00:42:27.680 --> 00:42:30.710 align:middle line:84% about that is one way in which firms in this model 00:42:30.710 --> 00:42:34.430 align:middle line:84% can raise prices is you can think of another game 00:42:34.430 --> 00:42:37.700 align:middle line:84% where you advertise your low quality product, 00:42:37.700 --> 00:42:39.590 align:middle line:84% and then you have an unadvertised price 00:42:39.590 --> 00:42:42.290 align:middle line:90% for a higher quality product. 00:42:42.290 --> 00:42:44.227 align:middle line:84% In that game where you only show people 00:42:44.227 --> 00:42:45.810 align:middle line:84% the price for the high quality product 00:42:45.810 --> 00:42:50.670 align:middle line:84% once they show up at the store, I gave this model that said, 00:42:50.670 --> 00:42:53.380 align:middle line:84% markups are going to be even higher in equilibrium. 00:42:53.380 --> 00:42:55.860 align:middle line:84% And the reason is that firms are going 00:42:55.860 --> 00:42:58.650 align:middle line:84% to charge the ex post monopoly price for the upgrades. 00:42:58.650 --> 00:43:01.170 align:middle line:90% 00:43:01.170 --> 00:43:04.020 align:middle line:84% That causes a wedge between-- when you force the prices 00:43:04.020 --> 00:43:07.080 align:middle line:84% of the low and high quality goods further apart, 00:43:07.080 --> 00:43:08.760 align:middle line:84% there's this adverse selection effect, 00:43:08.760 --> 00:43:10.960 align:middle line:90% where you don't want to get-- 00:43:10.960 --> 00:43:12.900 align:middle line:84% you don't want to sell to the cheapskates who 00:43:12.900 --> 00:43:15.390 align:middle line:84% are going to buy your rental car for $19.95 a day, 00:43:15.390 --> 00:43:17.980 align:middle line:84% and not get the insurance, and not get the car seat. 00:43:17.980 --> 00:43:20.550 align:middle line:84% And because there's this adverse selection effect where 00:43:20.550 --> 00:43:22.560 align:middle line:84% it's the cheapskates who don't buy the insurance 00:43:22.560 --> 00:43:26.510 align:middle line:84% and the rental-- and the car seats, you then want to dump-- 00:43:26.510 --> 00:43:29.820 align:middle line:84% you go from wanting to undercut the rival to get consumers 00:43:29.820 --> 00:43:33.700 align:middle line:84% to wanting to overcut the rival to dump the cheapskates on them. 00:43:33.700 --> 00:43:36.180 align:middle line:84% Markups for both products can end up-- markups 00:43:36.180 --> 00:43:37.800 align:middle line:84% for the sort of-- average markups 00:43:37.800 --> 00:43:39.540 align:middle line:90% can go up in equilibrium. 00:43:39.540 --> 00:43:42.970 align:middle line:84% We see what looks like an add-on strategy here. 00:43:42.970 --> 00:43:46.260 align:middle line:84% So you've clicked on this product I want to buy. 00:43:46.260 --> 00:43:49.200 align:middle line:84% I clicked on, I want to buy OEM 512 megabyte memory. 00:43:49.200 --> 00:43:51.540 align:middle line:84% They actually put the check mark in the correct box. 00:43:51.540 --> 00:43:53.490 align:middle line:84% This is what I said I wanted to buy. 00:43:53.490 --> 00:43:57.020 align:middle line:84% But then they tell me that I can spend $15 more 00:43:57.020 --> 00:44:00.140 align:middle line:84% and get this product, or I can spend $25 more 00:44:00.140 --> 00:44:03.710 align:middle line:84% and get this product, which I'm going to think of as add-ons. 00:44:03.710 --> 00:44:04.740 align:middle line:90% What do they do? 00:44:04.740 --> 00:44:09.050 align:middle line:84% They tell me a lot about why the better products are better. 00:44:09.050 --> 00:44:11.610 align:middle line:84% So for instance, this one is CAS 3 latency. 00:44:11.610 --> 00:44:13.470 align:middle line:90% This is CAS 2.5 latency. 00:44:13.470 --> 00:44:15.750 align:middle line:90% This is CAS 2.5 latency. 00:44:15.750 --> 00:44:17.580 align:middle line:90% This one has a four-layer board. 00:44:17.580 --> 00:44:20.750 align:middle line:84% These each have six-layer boards. 00:44:20.750 --> 00:44:24.958 align:middle line:84% These have OEM DRAM downgrade chips, 00:44:24.958 --> 00:44:26.750 align:middle line:84% whereas these have industry standard chips, 00:44:26.750 --> 00:44:29.180 align:middle line:84% whereas these have even better brand name chips. 00:44:29.180 --> 00:44:31.030 align:middle line:84% Samsung and Micron are major brands. 00:44:31.030 --> 00:44:34.730 align:middle line:90% 00:44:34.730 --> 00:44:36.430 align:middle line:90% These have a restocking fee. 00:44:36.430 --> 00:44:39.360 align:middle line:84% These have no restocking fee-- satisfaction and compatibility 00:44:39.360 --> 00:44:41.370 align:middle line:90% fully guaranteed. 00:44:41.370 --> 00:44:42.500 align:middle line:90% In fact, one thing that-- 00:44:42.500 --> 00:44:44.870 align:middle line:84% for fun, I clicked on some of these things. 00:44:44.870 --> 00:44:46.700 align:middle line:84% This one has verify compatibility 00:44:46.700 --> 00:44:48.740 align:middle line:90% with your memory configurator. 00:44:48.740 --> 00:44:51.960 align:middle line:84% I clicked on this, and entered a zillion different computers. 00:44:51.960 --> 00:44:54.980 align:middle line:84% Every consumer said this seemed to be a button that just says, 00:44:54.980 --> 00:44:56.303 align:middle line:90% can't verify compatibility. 00:44:56.303 --> 00:44:58.220 align:middle line:84% So no matter what you typed in, in some sense, 00:44:58.220 --> 00:45:00.530 align:middle line:84% it would just say always is not compatible. 00:45:00.530 --> 00:45:03.290 align:middle line:84% So clearly, this page is designed 00:45:03.290 --> 00:45:05.738 align:middle line:84% to get us to buy this instead of buy that, 00:45:05.738 --> 00:45:07.280 align:middle line:84% and just telling you all the reasons. 00:45:07.280 --> 00:45:09.073 align:middle line:84% These words like OEM downgrade chips. 00:45:09.073 --> 00:45:10.490 align:middle line:84% I don't even know what that means. 00:45:10.490 --> 00:45:12.000 align:middle line:90% But I know it's bad. 00:45:12.000 --> 00:45:14.550 align:middle line:84% It's not trying to teach us about the products. 00:45:14.550 --> 00:45:17.120 align:middle line:84% It's not trying to tell me why is CAS 2.5 latency 00:45:17.120 --> 00:45:18.540 align:middle line:90% better than CAS 3 latency. 00:45:18.540 --> 00:45:21.210 align:middle line:84% Or why is a six-layer board better than a four-layer board? 00:45:21.210 --> 00:45:22.610 align:middle line:90% Should you care about that? 00:45:22.610 --> 00:45:23.990 align:middle line:84% But it does seem like it's trying 00:45:23.990 --> 00:45:26.570 align:middle line:84% to convince us to buy this product 00:45:26.570 --> 00:45:29.215 align:middle line:84% and pay $25 extra instead of buying that product. 00:45:29.215 --> 00:45:32.040 align:middle line:90% 00:45:32.040 --> 00:45:35.280 align:middle line:84% So what we're going to argue in this paper 00:45:35.280 --> 00:45:39.992 align:middle line:84% is that if you really had Bertrand competition for memory 00:45:39.992 --> 00:45:41.700 align:middle line:84% modules and they were being sold at cost, 00:45:41.700 --> 00:45:43.780 align:middle line:84% the retailers couldn't cover their fixed costs. 00:45:43.780 --> 00:45:45.490 align:middle line:90% The market wouldn't exist. 00:45:45.490 --> 00:45:46.840 align:middle line:90% The market does exist. 00:45:46.840 --> 00:45:47.620 align:middle line:90% Why is that? 00:45:47.620 --> 00:45:50.970 align:middle line:84% Well, our thought is that firms are engaged. 00:45:50.970 --> 00:45:54.720 align:middle line:84% While Pricewatch has invested in technology 00:45:54.720 --> 00:45:57.120 align:middle line:84% to reduce search costs, the firms 00:45:57.120 --> 00:45:59.520 align:middle line:84% are fighting back with obfuscation 00:45:59.520 --> 00:46:02.550 align:middle line:84% to try to raise search costs and use those higher search costs 00:46:02.550 --> 00:46:06.080 align:middle line:90% to keep the increased markups. 00:46:06.080 --> 00:46:08.940 align:middle line:84% We can think of two different ways you could do that. 00:46:08.940 --> 00:46:11.910 align:middle line:84% One is-- the simplest theory of obfuscation 00:46:11.910 --> 00:46:17.190 align:middle line:84% is we know that you get higher markups and more dispersion 00:46:17.190 --> 00:46:20.260 align:middle line:84% when s is higher than when s is lower. 00:46:20.260 --> 00:46:23.730 align:middle line:84% So maybe what firms are doing is just making s higher. 00:46:23.730 --> 00:46:27.420 align:middle line:84% If I make you spend 5 minutes clicking-- reading through-- 00:46:27.420 --> 00:46:30.250 align:middle line:84% unclicking the thumbscrews, unclicking other things, 00:46:30.250 --> 00:46:31.968 align:middle line:84% I spend 5 minutes clicking through doing 00:46:31.968 --> 00:46:34.510 align:middle line:84% all the things where I finally found out the price, I'm like, 00:46:34.510 --> 00:46:37.010 align:middle line:84% do I really want to go back and go through another 5 minutes 00:46:37.010 --> 00:46:39.820 align:middle line:84% with the next website, and find out what their real price is. 00:46:39.820 --> 00:46:43.000 align:middle line:84% So by just making shopping more time consuming, 00:46:43.000 --> 00:46:44.200 align:middle line:90% I've raised the s. 00:46:44.200 --> 00:46:46.600 align:middle line:84% Raising the s is going to raise the equilibrium price 00:46:46.600 --> 00:46:48.150 align:middle line:90% distribution. 00:46:48.150 --> 00:46:50.700 align:middle line:84% And then the second thought would be-- 00:46:50.700 --> 00:46:53.160 align:middle line:84% this could be what I mentioned before, this add-on pricing 00:46:53.160 --> 00:46:56.490 align:middle line:84% situation where we know that equilibrium 00:46:56.490 --> 00:46:59.400 align:middle line:84% prices with add-on pricing are higher than equilibrium prices 00:46:59.400 --> 00:47:01.780 align:middle line:84% when all goods prices are advertised. 00:47:01.780 --> 00:47:04.410 align:middle line:84% Maybe what firms are trying to do is shift the equilibrium-- 00:47:04.410 --> 00:47:05.280 align:middle line:90% shift the-- 00:47:05.280 --> 00:47:09.150 align:middle line:84% change the game from a game in which firms advertise all prices 00:47:09.150 --> 00:47:13.200 align:middle line:84% to games in which firms advertise base goods, 00:47:13.200 --> 00:47:15.430 align:middle line:90% and then invent add-ons. 00:47:15.430 --> 00:47:17.790 align:middle line:84% And that model, the larger is the add-on 00:47:17.790 --> 00:47:19.755 align:middle line:84% as a fraction of the price, the higher 00:47:19.755 --> 00:47:21.460 align:middle line:84% are the equilibrium markups going to be. 00:47:21.460 --> 00:47:24.180 align:middle line:84% So firms are in some sense inventing add-ons 00:47:24.180 --> 00:47:26.340 align:middle line:84% or inventing inferior products that the add-ons 00:47:26.340 --> 00:47:27.510 align:middle line:90% can be even larger. 00:47:27.510 --> 00:47:30.160 align:middle line:84% And that can be a way to raise equilibrium prices. 00:47:30.160 --> 00:47:34.080 align:middle line:84% And what we're going to try to do in this paper is look 00:47:34.080 --> 00:47:36.330 align:middle line:84% at demand and look at what demand looks like, and see, 00:47:36.330 --> 00:47:40.590 align:middle line:84% can this model add-on pricing explain the equilibrium markups 00:47:40.590 --> 00:47:43.690 align:middle line:90% that we end up with? 00:47:43.690 --> 00:47:48.470 align:middle line:84% So the paper, we're going to look at four different products. 00:47:48.470 --> 00:47:52.810 align:middle line:84% The four different products are 128 megabyte PC100 memory 00:47:52.810 --> 00:47:55.720 align:middle line:84% modules, 100 megabyte PC133 modules, 00:47:55.720 --> 00:47:58.660 align:middle line:90% and 256 megabyte versions. 00:47:58.660 --> 00:48:01.545 align:middle line:84% We're going to treat these four categories differently, and just 00:48:01.545 --> 00:48:02.920 align:middle line:84% basically do it as an opportunity 00:48:02.920 --> 00:48:05.148 align:middle line:90% to have four separate products. 00:48:05.148 --> 00:48:06.940 align:middle line:84% We'll analyze each one separately, and just 00:48:06.940 --> 00:48:08.648 align:middle line:84% get some reinforcement of the conclusions 00:48:08.648 --> 00:48:10.000 align:middle line:90% we would get from one. 00:48:10.000 --> 00:48:11.590 align:middle line:84% But we do spend most of the time-- 00:48:11.590 --> 00:48:15.290 align:middle line:84% I'll focus today on 128 megabyte PC100 memory modules. 00:48:15.290 --> 00:48:18.790 align:middle line:84% That was the product that was most popular and stayed alive 00:48:18.790 --> 00:48:23.320 align:middle line:84% for longest over the one-year data collection. 00:48:23.320 --> 00:48:25.870 align:middle line:90% So what is our data collection? 00:48:25.870 --> 00:48:29.080 align:middle line:84% Within each of these categories, there's 00:48:29.080 --> 00:48:33.790 align:middle line:84% one retailer that we got internal data from. 00:48:33.790 --> 00:48:35.770 align:middle line:84% And then we have data, external data 00:48:35.770 --> 00:48:38.320 align:middle line:84% from all the other retailers from Pricewatch. 00:48:38.320 --> 00:48:41.140 align:middle line:84% The one retailer that we had this special data connection 00:48:41.140 --> 00:48:43.850 align:middle line:84% with, that retailer sold three prices-- 00:48:43.850 --> 00:48:45.810 align:middle line:84% three products within each category. 00:48:45.810 --> 00:48:48.440 align:middle line:84% We'll call them the low quality product, the medium quality 00:48:48.440 --> 00:48:50.490 align:middle line:84% product, the high quality product. 00:48:50.490 --> 00:48:53.300 align:middle line:84% You can roughly think of the low quality product 00:48:53.300 --> 00:48:55.020 align:middle line:84% as something that looks like this. 00:48:55.020 --> 00:48:57.170 align:middle line:84% The medium quality product looks like this. 00:48:57.170 --> 00:48:59.750 align:middle line:84% The high quality product looks like this. 00:48:59.750 --> 00:49:05.030 align:middle line:84% Something to know about them is in this market, the medium 00:49:05.030 --> 00:49:07.550 align:middle line:84% quality products really didn't cost much more 00:49:07.550 --> 00:49:09.320 align:middle line:90% than the low quality products. 00:49:09.320 --> 00:49:12.410 align:middle line:84% Often if these cost $60 wholesale, 00:49:12.410 --> 00:49:15.020 align:middle line:90% these would cost $61 wholesale. 00:49:15.020 --> 00:49:17.840 align:middle line:84% The retailer we got the data from said, in fact, sometimes 00:49:17.840 --> 00:49:20.990 align:middle line:84% he would even sell someone the $60 product, 00:49:20.990 --> 00:49:23.630 align:middle line:84% and then ship the $61 product, just 00:49:23.630 --> 00:49:26.540 align:middle line:84% because dealing with the hassles and the customer-- 00:49:26.540 --> 00:49:30.260 align:middle line:84% customer disappointment when they got the low quality 00:49:30.260 --> 00:49:31.920 align:middle line:90% product, then it was broken. 00:49:31.920 --> 00:49:32.673 align:middle line:90% They returned it. 00:49:32.673 --> 00:49:33.840 align:middle line:90% It wasn't worth his hassles. 00:49:33.840 --> 00:49:36.423 align:middle line:84% And you sometimes sell them a better product than they bought, 00:49:36.423 --> 00:49:39.950 align:middle line:84% just because these really didn't cost much more than these ones. 00:49:39.950 --> 00:49:44.070 align:middle line:84% But anyway, but he had the three products available. 00:49:44.070 --> 00:49:46.560 align:middle line:84% The high quality products that have the name brand 00:49:46.560 --> 00:49:49.843 align:middle line:84% chips in them did cost substantially more. 00:49:49.843 --> 00:49:51.510 align:middle line:84% When you're thinking about this, though, 00:49:51.510 --> 00:49:54.450 align:middle line:84% note that high quality versus low quality, 00:49:54.450 --> 00:49:56.010 align:middle line:90% it's really a bundle. 00:49:56.010 --> 00:50:00.870 align:middle line:84% So it's not the case that you can compare my medium quality 00:50:00.870 --> 00:50:03.600 align:middle line:84% product from firm A and medium quality product from firm B, 00:50:03.600 --> 00:50:07.240 align:middle line:84% because quality is this many, many dimensional thing. 00:50:07.240 --> 00:50:08.890 align:middle line:84% What's the restocking fee policy? 00:50:08.890 --> 00:50:10.290 align:middle line:90% What's the shipping policy? 00:50:10.290 --> 00:50:13.150 align:middle line:84% How many layers does the board have? 00:50:13.150 --> 00:50:15.030 align:middle line:90% What's the CAS latency? 00:50:15.030 --> 00:50:19.240 align:middle line:84% These are not something that's comparable across retailers. 00:50:19.240 --> 00:50:21.690 align:middle line:84% So we can really-- every different retailer 00:50:21.690 --> 00:50:23.640 align:middle line:84% may have its own low, medium, and high quality 00:50:23.640 --> 00:50:25.390 align:middle line:84% that don't correspond with the other ones. 00:50:25.390 --> 00:50:28.080 align:middle line:84% But so anyway, our retailer had these three different quality 00:50:28.080 --> 00:50:29.710 align:middle line:90% levels that were consistent. 00:50:29.710 --> 00:50:32.360 align:middle line:90% 00:50:32.360 --> 00:50:36.920 align:middle line:84% The price data that we get is we went to the pricewatch.com 00:50:36.920 --> 00:50:37.490 align:middle line:90% website. 00:50:37.490 --> 00:50:39.950 align:middle line:84% And we just wrote an early web scraping program, 00:50:39.950 --> 00:50:43.490 align:middle line:84% and just basically downloaded the raw HTML files, 00:50:43.490 --> 00:50:47.120 align:middle line:84% and got the 12 or 24 lowest prices in each category. 00:50:47.120 --> 00:50:49.370 align:middle line:84% We did that hourly, and we did that hourly 00:50:49.370 --> 00:50:51.440 align:middle line:90% over the course of a year. 00:50:51.440 --> 00:50:53.720 align:middle line:84% Presumably, all the prices in that data 00:50:53.720 --> 00:50:55.490 align:middle line:90% are for low quality products. 00:50:55.490 --> 00:50:58.140 align:middle line:84% If the high quality products cost $15 more, 00:50:58.140 --> 00:51:03.640 align:middle line:84% they would never show up in the lowest 12 or 24 lowest prices. 00:51:03.640 --> 00:51:06.680 align:middle line:84% We have quantity data from one retailer. 00:51:06.680 --> 00:51:10.210 align:middle line:84% So that one retailer we had contact with owned websites A 00:51:10.210 --> 00:51:14.020 align:middle line:84% and B. For those websites, we have the medium and high quality 00:51:14.020 --> 00:51:15.370 align:middle line:90% prices available. 00:51:15.370 --> 00:51:19.480 align:middle line:84% And we have the number of modules of each type 00:51:19.480 --> 00:51:20.630 align:middle line:90% that it sold. 00:51:20.630 --> 00:51:22.850 align:middle line:84% Typically, we're going to use this at a daily level. 00:51:22.850 --> 00:51:24.250 align:middle line:84% We're going to aggregate up to-- we 00:51:24.250 --> 00:51:25.870 align:middle line:84% have the real-time sales data, but we're 00:51:25.870 --> 00:51:27.745 align:middle line:84% going to aggregate up to the real-time sales, 00:51:27.745 --> 00:51:29.360 align:middle line:90% to the daily sales. 00:51:29.360 --> 00:51:31.810 align:middle line:84% And then, also, something that's very rare 00:51:31.810 --> 00:51:34.850 align:middle line:84% is we have tremendous cost data here. 00:51:34.850 --> 00:51:37.420 align:middle line:90% So we actually have the raw-- 00:51:37.420 --> 00:51:40.840 align:middle line:84% we have all of the invoices that these-- 00:51:40.840 --> 00:51:42.110 align:middle line:90% from this retailer. 00:51:42.110 --> 00:51:45.160 align:middle line:84% So we know every time that it bought modules wholesale, 00:51:45.160 --> 00:51:46.450 align:middle line:90% what did it buy. 00:51:46.450 --> 00:51:48.760 align:middle line:84% And what's interesting about this business 00:51:48.760 --> 00:51:51.740 align:middle line:84% is this was basically a negative inventory business. 00:51:51.740 --> 00:51:54.320 align:middle line:84% This is a business where prices were falling over time. 00:51:54.320 --> 00:51:56.055 align:middle line:84% You didn't want to hold inventory. 00:51:56.055 --> 00:51:57.430 align:middle line:84% And so what the retailer would do 00:51:57.430 --> 00:51:59.590 align:middle line:84% is basically every day would turn on its computers, 00:51:59.590 --> 00:52:01.580 align:middle line:84% and would start selling memory modules. 00:52:01.580 --> 00:52:04.430 align:middle line:84% And at the end of the day, or at 3:00 in the afternoon, 00:52:04.430 --> 00:52:07.610 align:middle line:84% it would say, I've already sold 12 of this module, 00:52:07.610 --> 00:52:11.130 align:middle line:84% and 24 of this one, and 36 of this one, and 14 of this one. 00:52:11.130 --> 00:52:15.890 align:middle line:84% And then it would buy them wholesale that afternoon, 00:52:15.890 --> 00:52:18.890 align:middle line:84% drive a truck up to city of industry in California, 00:52:18.890 --> 00:52:23.060 align:middle line:84% go to a wholesale distributor, and buy the 340 memory modules 00:52:23.060 --> 00:52:25.170 align:middle line:84% that it had sold that day online. 00:52:25.170 --> 00:52:30.870 align:middle line:84% Drive the truck back to the firm's retail office. 00:52:30.870 --> 00:52:32.900 align:middle line:84% And then the warehouse workers would then 00:52:32.900 --> 00:52:34.460 align:middle line:84% start putting those things in boxes, 00:52:34.460 --> 00:52:36.260 align:middle line:84% and shipping them out the next day. 00:52:36.260 --> 00:52:38.900 align:middle line:84% So basically, as a result, it was 00:52:38.900 --> 00:52:42.390 align:middle line:84% buying things wholesale every day, or every two or three days. 00:52:42.390 --> 00:52:45.260 align:middle line:84% And we have very high frequency cost data, 00:52:45.260 --> 00:52:48.650 align:middle line:84% which is rare in understanding how the cost to the retailer 00:52:48.650 --> 00:52:49.150 align:middle line:90% varied. 00:52:49.150 --> 00:52:51.820 align:middle line:90% 00:52:51.820 --> 00:52:56.590 align:middle line:84% Just for some basic picture of the industry, 00:52:56.590 --> 00:52:59.980 align:middle line:84% this shows the price of our 128-megabyte modules over 00:52:59.980 --> 00:53:02.490 align:middle line:84% the course of the year from which we collected data. 00:53:02.490 --> 00:53:05.110 align:middle line:90% So they start out at about $80. 00:53:05.110 --> 00:53:07.760 align:middle line:90% Prices rose to about $120. 00:53:07.760 --> 00:53:10.040 align:middle line:84% And then with technological progress, they drop. 00:53:10.040 --> 00:53:12.040 align:middle line:84% And by the end of the sample, those same modules 00:53:12.040 --> 00:53:14.500 align:middle line:84% that were $20 at the beginning or like $80 at the beginning 00:53:14.500 --> 00:53:15.890 align:middle line:90% are $25. 00:53:15.890 --> 00:53:17.840 align:middle line:84% This is actually an interesting event here. 00:53:17.840 --> 00:53:22.330 align:middle line:84% This price spike was due to collusion 00:53:22.330 --> 00:53:24.740 align:middle line:90% by the memory manufacturers. 00:53:24.740 --> 00:53:30.190 align:middle line:84% So there was actually an antitrust case or legal case 00:53:30.190 --> 00:53:32.128 align:middle line:84% against these companies for price fixing. 00:53:32.128 --> 00:53:33.670 align:middle line:84% And companies were actually, I think, 00:53:33.670 --> 00:53:35.860 align:middle line:90% found guilty for price fixing. 00:53:35.860 --> 00:53:38.200 align:middle line:84% And they managed to raise prices dramatically, 00:53:38.200 --> 00:53:43.060 align:middle line:84% but only for a short time before the collusive agreement 00:53:43.060 --> 00:53:44.770 align:middle line:90% fell apart. 00:53:44.770 --> 00:53:47.560 align:middle line:84% And actually, this is one of the cases, actually, one 00:53:47.560 --> 00:53:49.900 align:middle line:84% of the rare times I made some money on my research 00:53:49.900 --> 00:53:52.520 align:middle line:84% where I was contacted by a law firm, 00:53:52.520 --> 00:53:54.870 align:middle line:84% looking for high frequency data on these memory prices. 00:53:54.870 --> 00:53:57.500 align:middle line:84% And I just got this phone call out of the blue from someone who 00:53:57.500 --> 00:54:00.830 align:middle line:84% said, we're representing some company that's 00:54:00.830 --> 00:54:03.050 align:middle line:84% suing Micron or whomever about these things, 00:54:03.050 --> 00:54:05.150 align:middle line:84% and we wanted to get access to your data. 00:54:05.150 --> 00:54:08.040 align:middle line:84% How much will you charge for your memory data? 00:54:08.040 --> 00:54:10.550 align:middle line:84% And I was like, charge for my memory data? 00:54:10.550 --> 00:54:13.370 align:middle line:84% I just give it away to everybody all the time. 00:54:13.370 --> 00:54:15.560 align:middle line:84% I haven't put it on my website at the time, 00:54:15.560 --> 00:54:17.780 align:middle line:84% but I normally just give it away for free. 00:54:17.780 --> 00:54:20.750 align:middle line:84% And they're like, how much do I charge? 00:54:20.750 --> 00:54:21.965 align:middle line:90% I don't know. $5,000? 00:54:21.965 --> 00:54:23.340 align:middle line:84% And the guy was like, yeah, sure. 00:54:23.340 --> 00:54:23.910 align:middle line:90% OK, thanks. 00:54:23.910 --> 00:54:24.720 align:middle line:90% We'll send you the check. 00:54:24.720 --> 00:54:26.887 align:middle line:84% And I was like, oh, I guess maybe I should have said 00:54:26.887 --> 00:54:27.490 align:middle line:90% $50,000. 00:54:27.490 --> 00:54:28.260 align:middle line:90% So I don't know. 00:54:28.260 --> 00:54:31.262 align:middle line:84% But anyway, these guys were price fixing. 00:54:31.262 --> 00:54:32.720 align:middle line:84% The price fixing scheme fell apart, 00:54:32.720 --> 00:54:36.170 align:middle line:84% and then memory prices continued on that downward trend. 00:54:36.170 --> 00:54:38.000 align:middle line:84% What's notable about the downward trend 00:54:38.000 --> 00:54:40.223 align:middle line:90% for this paper is-- 00:54:40.223 --> 00:54:41.890 align:middle line:84% this is showing you, I think, the 12th-- 00:54:41.890 --> 00:54:44.740 align:middle line:84% the lowest price and the 12th lowest price for the memory over 00:54:44.740 --> 00:54:45.570 align:middle line:90% time. 00:54:45.570 --> 00:54:47.960 align:middle line:84% You can see that prices are changing dramatically. 00:54:47.960 --> 00:54:51.040 align:middle line:84% This graph here is our retailer's price. 00:54:51.040 --> 00:54:54.460 align:middle line:84% And our retailer changes its price occasionally, 00:54:54.460 --> 00:54:58.253 align:middle line:84% but as a result of changing its price occasionally, every time 00:54:58.253 --> 00:54:59.920 align:middle line:84% it changes its price, it may drop and be 00:54:59.920 --> 00:55:01.212 align:middle line:90% the lowest price in the market. 00:55:01.212 --> 00:55:02.830 align:middle line:84% And then by leaving its price fixed, 00:55:02.830 --> 00:55:05.290 align:middle line:84% it's drifting, you can think, downward 00:55:05.290 --> 00:55:08.680 align:middle line:84% on that price watch list, where I leave my price fixed. 00:55:08.680 --> 00:55:11.640 align:middle line:84% And I go from having the lowest price to having the 10th lowest 00:55:11.640 --> 00:55:12.140 align:middle line:90% price. 00:55:12.140 --> 00:55:14.182 align:middle line:84% I lower it again to have the second lowest price, 00:55:14.182 --> 00:55:15.910 align:middle line:90% and then my price drifts across. 00:55:15.910 --> 00:55:20.080 align:middle line:84% And every time it drifts across as prices elsewhere are falling, 00:55:20.080 --> 00:55:23.390 align:middle line:84% I end up getting bumped up and down on that Pricewatch list. 00:55:23.390 --> 00:55:27.520 align:middle line:84% So maybe I've just reset my price. 00:55:27.520 --> 00:55:28.540 align:middle line:90% And I set my price. 00:55:28.540 --> 00:55:30.560 align:middle line:90% And I'm here at $69. 00:55:30.560 --> 00:55:33.550 align:middle line:84% But then as I stay at $69, and this guy jumps to $68, 00:55:33.550 --> 00:55:36.520 align:middle line:84% and this one goes to $67, and this one goes to $66, 00:55:36.520 --> 00:55:39.010 align:middle line:84% I fall down the list, and then I reset my price. 00:55:39.010 --> 00:55:40.500 align:middle line:90% And I jump back up. 00:55:40.500 --> 00:55:43.015 align:middle line:84% And we think that's very nice for identification 00:55:43.015 --> 00:55:44.473 align:middle line:84% if we're trying to estimate demand. 00:55:44.473 --> 00:55:46.980 align:middle line:90% 00:55:46.980 --> 00:55:50.190 align:middle line:84% We're basically regressing demand on your price. 00:55:50.190 --> 00:55:53.760 align:middle line:84% And we're going to think of the price, your relative price 00:55:53.760 --> 00:55:55.240 align:middle line:90% as exogenous variation. 00:55:55.240 --> 00:55:58.713 align:middle line:84% It's not that I'm deciding here to be the lowest price, 00:55:58.713 --> 00:56:00.880 align:middle line:84% and I'm deciding here to be the second lowest price. 00:56:00.880 --> 00:56:02.280 align:middle line:84% I'm trying here to be the eighth lowest price, 00:56:02.280 --> 00:56:03.940 align:middle line:84% and here to be the 10th lowest price. 00:56:03.940 --> 00:56:06.060 align:middle line:84% It's basically the firm's getting buffeted around 00:56:06.060 --> 00:56:07.480 align:middle line:90% by the randomness in prices. 00:56:07.480 --> 00:56:09.420 align:middle line:84% And we're going to think of that with where 00:56:09.420 --> 00:56:12.000 align:middle line:84% you are within the distribution as being largely random 00:56:12.000 --> 00:56:15.870 align:middle line:90% from the firm's perspective. 00:56:15.870 --> 00:56:23.990 align:middle line:84% So some summary statistics, so for the 128-megabyte PC modules, 00:56:23.990 --> 00:56:25.010 align:middle line:90% again, we have two-- 00:56:25.010 --> 00:56:26.560 align:middle line:84% we have one year of data, but then we 00:56:26.560 --> 00:56:28.430 align:middle line:90% have two different websites. 00:56:28.430 --> 00:56:35.590 align:middle line:84% So days times websites, 365, we end up with 680 website day 00:56:35.590 --> 00:56:37.120 align:middle line:90% observations. 00:56:37.120 --> 00:56:39.790 align:middle line:90% What do the memory modules cost? 00:56:39.790 --> 00:56:44.380 align:middle line:84% So the low quality modules generally cost, on average, $67. 00:56:44.380 --> 00:56:47.870 align:middle line:84% But obviously, remember, it was a big change over time. 00:56:47.870 --> 00:56:51.280 align:middle line:84% Medium quality modules are about $25 more. 00:56:51.280 --> 00:56:53.540 align:middle line:84% High quality modules are about $50 more. 00:56:53.540 --> 00:56:58.570 align:middle line:84% So very, very substantial add-on strategy here. 00:56:58.570 --> 00:57:00.850 align:middle line:84% The low quality modules are usually not 00:57:00.850 --> 00:57:03.500 align:middle line:84% very different from the lowest price in the market. 00:57:03.500 --> 00:57:05.600 align:middle line:84% The lowest price in the market is $63. 00:57:05.600 --> 00:57:08.440 align:middle line:90% This firm is charging $66. 00:57:08.440 --> 00:57:10.790 align:middle line:84% What do people do when faced with these prices? 00:57:10.790 --> 00:57:13.780 align:middle line:84% Well, given that they are charging an awful lot of money 00:57:13.780 --> 00:57:16.120 align:middle line:84% for those add-ons, most people are choosing 00:57:16.120 --> 00:57:17.830 align:middle line:90% to buy the low quality product. 00:57:17.830 --> 00:57:22.730 align:middle line:84% They sell on average 12.8 per website per day. 00:57:22.730 --> 00:57:25.520 align:middle line:84% Medium quality modules are selling 2.4. 00:57:25.520 --> 00:57:27.840 align:middle line:84% High quality modules are selling 2.0. 00:57:27.840 --> 00:57:30.720 align:middle line:84% So most people buy the low quality modules. 00:57:30.720 --> 00:57:34.303 align:middle line:84% But a number of people buy medium and high quality modules. 00:57:34.303 --> 00:57:35.720 align:middle line:84% Some people have also asked about, 00:57:35.720 --> 00:57:37.440 align:middle line:84% is this just some really tiny market? 00:57:37.440 --> 00:57:39.173 align:middle line:84% Well, what's going on is this firm 00:57:39.173 --> 00:57:41.090 align:middle line:84% has a website that has hundreds, and hundreds, 00:57:41.090 --> 00:57:42.132 align:middle line:90% and hundreds of products. 00:57:42.132 --> 00:57:52.400 align:middle line:84% So it's selling 17 PC100 128-megabyte modules per website 00:57:52.400 --> 00:57:53.220 align:middle line:90% per day. 00:57:53.220 --> 00:57:56.090 align:middle line:84% So it's selling 17 on website A. It's selling 17 on website B. 00:57:56.090 --> 00:57:59.870 align:middle line:84% But then it's selling many 256-megabyte modules, 00:57:59.870 --> 00:58:03.930 align:middle line:84% and many PC133 modules, and many PC whatever modules. 00:58:03.930 --> 00:58:06.440 align:middle line:84% So this is one of many, many products, 00:58:06.440 --> 00:58:08.510 align:middle line:84% but we just pick out this daily sales data, 00:58:08.510 --> 00:58:11.880 align:middle line:84% because this is the one that had things bought consistently. 00:58:11.880 --> 00:58:14.532 align:middle line:90% 00:58:14.532 --> 00:58:16.240 align:middle line:84% The main thing we want to do in the paper 00:58:16.240 --> 00:58:20.230 align:middle line:84% is just analyze what is the demand curve. 00:58:20.230 --> 00:58:23.140 align:middle line:84% With all the obfuscation out there, competition 00:58:23.140 --> 00:58:27.730 align:middle line:84% between the sites, the attempt of this website 00:58:27.730 --> 00:58:30.130 align:middle line:84% to convince people to buy upgrades, 00:58:30.130 --> 00:58:32.740 align:middle line:84% what does demand look like and what do all the cross-price 00:58:32.740 --> 00:58:34.720 align:middle line:90% elasticities look like? 00:58:34.720 --> 00:58:36.850 align:middle line:84% We do this in a very simple thing. 00:58:36.850 --> 00:58:41.230 align:middle line:84% I think what we would like to do is, in some sense, 00:58:41.230 --> 00:58:47.650 align:middle line:84% we want to just regress log Q equals alpha 0 minus alpha 1 log 00:58:47.650 --> 00:58:51.790 align:middle line:90% P and so on. 00:58:51.790 --> 00:58:54.280 align:middle line:84% You can't run a regression of log Q 00:58:54.280 --> 00:58:58.300 align:middle line:84% on log P when the Q's are discrete things that are often 00:58:58.300 --> 00:58:59.290 align:middle line:90% 0. 00:58:59.290 --> 00:59:01.725 align:middle line:84% So what we do instead is, in some sense, the-- 00:59:01.725 --> 00:59:05.040 align:middle line:90% 00:59:05.040 --> 00:59:09.810 align:middle line:84% you can think of it as log Q equals x beta plus epsilon. 00:59:09.810 --> 00:59:13.020 align:middle line:84% Instead of running log Q equals x beta plus epsilon, 00:59:13.020 --> 00:59:15.160 align:middle line:90% we just exponentiate both sides. 00:59:15.160 --> 00:59:17.970 align:middle line:84% We'll have Q equals e to the x beta times 00:59:17.970 --> 00:59:21.750 align:middle line:84% u, where u is a multiplicative shock with expectation 00:59:21.750 --> 00:59:25.230 align:middle line:90% of u given x equals 1. 00:59:25.230 --> 00:59:27.210 align:middle line:90% What is on the right side? 00:59:27.210 --> 00:59:31.440 align:middle line:90% So we're regressing log Q on-- 00:59:31.440 --> 00:59:33.407 align:middle line:84% the log of the price of low quality memory, 00:59:33.407 --> 00:59:35.490 align:middle line:84% the log of the price of the medium quality memory, 00:59:35.490 --> 00:59:38.660 align:middle line:84% the log of the price of high quality memory, 00:59:38.660 --> 00:59:40.167 align:middle line:90% other covariates. 00:59:40.167 --> 00:59:41.750 align:middle line:84% Also, particularly what we've included 00:59:41.750 --> 00:59:44.270 align:middle line:84% to control for the sales is, what 00:59:44.270 --> 00:59:46.980 align:middle line:84% is the rank of your low quality memory on the screen? 00:59:46.980 --> 00:59:51.240 align:middle line:84% This turns out to be a very powerful predictor. 00:59:51.240 --> 00:59:53.090 align:middle line:84% If you're in the first spot on that screen, 00:59:53.090 --> 00:59:54.440 align:middle line:84% you sell much more memory than if you're 00:59:54.440 --> 00:59:56.880 align:middle line:84% in the second, or third, or fourth spot on the screen. 00:59:56.880 --> 01:00:01.130 align:middle line:84% This is, in some sense, how we control for rivals' prices is 01:00:01.130 --> 01:00:07.070 align:middle line:84% I'm regressing sales at website A of medium quality memory 01:00:07.070 --> 01:00:10.070 align:middle line:84% on the prices at website A for low, medium, and high quality 01:00:10.070 --> 01:00:10.920 align:middle line:90% memory. 01:00:10.920 --> 01:00:13.550 align:middle line:84% And then I'm controlling for competitors' prices 01:00:13.550 --> 01:00:14.887 align:middle line:90% through these two variables. 01:00:14.887 --> 01:00:16.470 align:middle line:84% What's the lowest price on the screen? 01:00:16.470 --> 01:00:18.080 align:middle line:90% How high am I relative to that? 01:00:18.080 --> 01:00:21.630 align:middle line:84% And what's my rank on that screen? 01:00:21.630 --> 01:00:24.170 align:middle line:84% So in some sense, you might want to include 01:00:24.170 --> 01:00:25.337 align:middle line:90% all the competitors' prices. 01:00:25.337 --> 01:00:26.878 align:middle line:84% That would, obviously, be infeasible. 01:00:26.878 --> 01:00:29.150 align:middle line:84% So I include, what's your rank on the screen, as well 01:00:29.150 --> 01:00:30.690 align:middle line:90% as my prices here. 01:00:30.690 --> 01:00:33.860 align:middle line:84% And notice that I'm putting low, medium, and high quality prices 01:00:33.860 --> 01:00:35.280 align:middle line:90% on the right side. 01:00:35.280 --> 01:00:37.790 align:middle line:84% Whether the left side has low quality memory 01:00:37.790 --> 01:00:40.680 align:middle line:84% as a dependent variable, medium quality 01:00:40.680 --> 01:00:43.270 align:middle line:84% memory, or high quality memory as a dependent variable. 01:00:43.270 --> 01:00:45.930 align:middle line:84% And I'm doing that because people 01:00:45.930 --> 01:00:48.880 align:middle line:84% are going to substitute between the firm's offerings. 01:00:48.880 --> 01:00:50.670 align:middle line:84% If it's charging $25, the upgrade 01:00:50.670 --> 01:00:52.503 align:middle line:84% is going to sell fewer upgrades than if it's 01:00:52.503 --> 01:00:53.850 align:middle line:90% charging $10 for the upgrade. 01:00:53.850 --> 01:00:57.990 align:middle line:84% It'll sell many fewer if it's charging $50 for the upgrade. 01:00:57.990 --> 01:01:01.080 align:middle line:84% Our primary specification is just 01:01:01.080 --> 01:01:04.140 align:middle line:84% basically doing the OLS equivalent, 01:01:04.140 --> 01:01:06.030 align:middle line:84% treating the prices on the right side 01:01:06.030 --> 01:01:09.490 align:middle line:84% as exogenous for the reasons I explained earlier. 01:01:09.490 --> 01:01:13.140 align:middle line:84% Just expectation of u given the x is equal to 1. 01:01:13.140 --> 01:01:17.820 align:middle line:84% We do also in the paper have a couple different IV strategies. 01:01:17.820 --> 01:01:22.170 align:middle line:84% One IV strategy is using the costs as instrument for prices. 01:01:22.170 --> 01:01:25.230 align:middle line:84% Costs are a classic instrument for prices. 01:01:25.230 --> 01:01:29.580 align:middle line:84% So anyway, we use that wholesale cost data to get an IV. 01:01:29.580 --> 01:01:32.430 align:middle line:84% The other thing we do is also use the prices 01:01:32.430 --> 01:01:34.650 align:middle line:84% for other products as instrument IV. 01:01:34.650 --> 01:01:37.530 align:middle line:84% The explanation for that one is if you imagine 01:01:37.530 --> 01:01:39.180 align:middle line:90% the firm comes in for the day. 01:01:39.180 --> 01:01:42.480 align:middle line:84% It sees that two of its warehouse employees 01:01:42.480 --> 01:01:45.420 align:middle line:84% call in sick, which is not at all an uncommon thing 01:01:45.420 --> 01:01:47.640 align:middle line:84% when you're paying minimum wage workers. 01:01:47.640 --> 01:01:50.430 align:middle line:84% If two of your workers are sick, what you do is you know, 01:01:50.430 --> 01:01:52.380 align:middle line:90% you can't package as many boxes. 01:01:52.380 --> 01:01:55.440 align:middle line:84% So you just try to sell fewer goods that day 01:01:55.440 --> 01:01:57.280 align:middle line:90% by raising your prices. 01:01:57.280 --> 01:02:00.330 align:middle line:84% You raise your prices, sell a few items at a high markup 01:02:00.330 --> 01:02:02.440 align:middle line:84% rather than a lot of items at a low markup. 01:02:02.440 --> 01:02:05.970 align:middle line:84% And so I can use the PC133 prices 01:02:05.970 --> 01:02:08.230 align:middle line:84% as an instrument for PC100 prices, 01:02:08.230 --> 01:02:09.737 align:middle line:84% because if your employees are sick, 01:02:09.737 --> 01:02:11.320 align:middle line:84% you would probably raise both of them. 01:02:11.320 --> 01:02:13.830 align:middle line:84% And obviously, the PC133 prices aren't 01:02:13.830 --> 01:02:16.230 align:middle line:84% going to reflect demand for PC100, which is 01:02:16.230 --> 01:02:18.127 align:middle line:90% bought by separate consumers. 01:02:18.127 --> 01:02:19.710 align:middle line:84% But anyway, I think mostly we're going 01:02:19.710 --> 01:02:23.220 align:middle line:84% to say we think this is a case where within the distribution 01:02:23.220 --> 01:02:25.810 align:middle line:84% price variation really is as good as random. 01:02:25.810 --> 01:02:29.580 align:middle line:84% And we're mostly going to estimate this by, well, its GMM, 01:02:29.580 --> 01:02:31.530 align:middle line:84% but it's like OLS, where we're just treating 01:02:31.530 --> 01:02:34.830 align:middle line:90% those prices as exogenous. 01:02:34.830 --> 01:02:36.080 align:middle line:90% What are some of the findings? 01:02:36.080 --> 01:02:40.420 align:middle line:84% So first, demand for low quality memory 01:02:40.420 --> 01:02:42.940 align:middle line:90% is very, very price sensitive. 01:02:42.940 --> 01:02:46.840 align:middle line:84% If your rank changes from 1 to 2 to 3, 01:02:46.840 --> 01:02:48.880 align:middle line:84% every time you drop down on that list, 01:02:48.880 --> 01:02:52.380 align:middle line:90% you sell many, many fewer units. 01:02:52.380 --> 01:02:56.630 align:middle line:84% We're going to report that price elasticities are as high as 33. 01:02:56.630 --> 01:03:00.820 align:middle line:84% So you raise your prices by 1% from $100 to $101, 01:03:00.820 --> 01:03:02.600 align:middle line:84% a third of your demand goes away. 01:03:02.600 --> 01:03:04.420 align:middle line:84% You raise price from $101 to $102, 01:03:04.420 --> 01:03:06.760 align:middle line:84% another third of your demand goes away again. 01:03:06.760 --> 01:03:08.780 align:middle line:84% $102 to $103, a third of your demand goes away. 01:03:08.780 --> 01:03:11.950 align:middle line:84% So it's just incredibly price sensitive demand for these firms 01:03:11.950 --> 01:03:13.610 align:middle line:90% caused by this rank effect. 01:03:13.610 --> 01:03:17.906 align:middle line:84% As I go from 1 to 2 to 3 to 4, I just keep losing customers. 01:03:17.906 --> 01:03:20.460 align:middle line:90% 01:03:20.460 --> 01:03:23.690 align:middle line:84% The second thing is that-- interesting finding is there's 01:03:23.690 --> 01:03:25.440 align:middle line:90% a big loss leader effect here. 01:03:25.440 --> 01:03:29.810 align:middle line:84% So I'm selling this low quality memory basically at cost. 01:03:29.810 --> 01:03:33.680 align:middle line:84% But when I'm-- by selling this low quality memory, 01:03:33.680 --> 01:03:35.360 align:middle line:90% I get to the top of the list. 01:03:35.360 --> 01:03:36.470 align:middle line:90% And being at the top-- 01:03:36.470 --> 01:03:38.262 align:middle line:90% moving down the list-- 01:03:38.262 --> 01:03:39.720 align:middle line:84% when I move to the top of the list, 01:03:39.720 --> 01:03:42.150 align:middle line:84% I not only sell much more low quality memory. 01:03:42.150 --> 01:03:44.190 align:middle line:84% I also am able to talk some of those people 01:03:44.190 --> 01:03:46.607 align:middle line:84% into buying medium quality memory and some of those people 01:03:46.607 --> 01:03:47.690 align:middle line:90% into high quality memory. 01:03:47.690 --> 01:03:50.400 align:middle line:84% So being high on the screen, I sell 01:03:50.400 --> 01:03:53.070 align:middle line:84% more of the product that makes my price high, 01:03:53.070 --> 01:03:55.750 align:middle line:84% but holding the price of the medium quality memory fixed, 01:03:55.750 --> 01:03:58.170 align:middle line:84% having a low price for low quality memory 01:03:58.170 --> 01:03:59.850 align:middle line:90% has this big loss leader effect. 01:03:59.850 --> 01:04:02.220 align:middle line:84% And I sell much more medium quality memory, 01:04:02.220 --> 01:04:04.080 align:middle line:84% and much more high quality memory 01:04:04.080 --> 01:04:07.310 align:middle line:84% when I'm up at the top of that list. 01:04:07.310 --> 01:04:09.050 align:middle line:84% And this, again, it's controlling 01:04:09.050 --> 01:04:11.493 align:middle line:84% for the price of my medium quality memory itself. 01:04:11.493 --> 01:04:13.910 align:middle line:84% So it's holding that price of medium quality memory fixed. 01:04:13.910 --> 01:04:15.785 align:middle line:84% When I set a low price on low quality memory, 01:04:15.785 --> 01:04:17.868 align:middle line:84% I just sell much more of the medium quality memory 01:04:17.868 --> 01:04:19.540 align:middle line:84% because I'm drawing people to my site. 01:04:19.540 --> 01:04:22.310 align:middle line:90% 01:04:22.310 --> 01:04:26.600 align:middle line:84% I mentioned that the reason add-on pricing can raise profits 01:04:26.600 --> 01:04:29.180 align:middle line:84% and equilibrium a few lectures ago 01:04:29.180 --> 01:04:32.280 align:middle line:84% is because of this adverse selection effect. 01:04:32.280 --> 01:04:35.100 align:middle line:84% When the cheapskates who come to your website 01:04:35.100 --> 01:04:37.920 align:middle line:84% to buy the low quality product won't buy the upgrades, 01:04:37.920 --> 01:04:41.170 align:middle line:84% you don't want those cheapskates buying stuff at a loss. 01:04:41.170 --> 01:04:43.920 align:middle line:84% And so that causes firms to have this incentive, flipping 01:04:43.920 --> 01:04:45.870 align:middle line:84% from undercutting to overcutting, 01:04:45.870 --> 01:04:48.400 align:middle line:84% and raises equilibrium price levels. 01:04:48.400 --> 01:04:51.370 align:middle line:84% Here we're finding evidence of that adverse selection effect. 01:04:51.370 --> 01:04:54.570 align:middle line:84% So increasing this variable by one 01:04:54.570 --> 01:04:58.830 align:middle line:84% raises low quality sales by 129%, medium quality sales 01:04:58.830 --> 01:05:03.120 align:middle line:84% by 77%, high quality sales by 51%. 01:05:03.120 --> 01:05:06.280 align:middle line:84% So I'm selling more medium and high quality memory. 01:05:06.280 --> 01:05:09.090 align:middle line:84% My product mix is getting worse, and worse, and worse, 01:05:09.090 --> 01:05:12.540 align:middle line:84% and a larger and larger fraction of my sales 01:05:12.540 --> 01:05:17.080 align:middle line:84% are the low quality, low profit goods as I move up on the list. 01:05:17.080 --> 01:05:19.500 align:middle line:84% So this is the adverse selection effect 01:05:19.500 --> 01:05:23.070 align:middle line:84% we talked about, that it is clear evidence 01:05:23.070 --> 01:05:25.280 align:middle line:84% that adverse selection effect exists in this market. 01:05:25.280 --> 01:05:28.268 align:middle line:90% 01:05:28.268 --> 01:05:30.800 align:middle line:84% In addition to these rank effects, 01:05:30.800 --> 01:05:33.440 align:middle line:84% the own price effects are significant as well. 01:05:33.440 --> 01:05:36.200 align:middle line:84% So when I raise my price for medium quality memory, 01:05:36.200 --> 01:05:37.980 align:middle line:84% I do sell less medium quality memory. 01:05:37.980 --> 01:05:41.300 align:middle line:84% When I raise my price for high quality memory, I do sell less. 01:05:41.300 --> 01:05:43.530 align:middle line:84% Those elasticities are like 5 or 6. 01:05:43.530 --> 01:05:48.090 align:middle line:84% So if I raise my medium quality price by $1, I sell 5% less. 01:05:48.090 --> 01:05:49.250 align:middle line:90% I raise it by $5. 01:05:49.250 --> 01:05:53.610 align:middle line:84% I sell-- sorry-- by 5%, I sell 25% less. 01:05:53.610 --> 01:05:55.520 align:middle line:84% So this does appear to be substitution 01:05:55.520 --> 01:05:59.870 align:middle line:84% within the website, where if those upgrades are 15, 30, 01:05:59.870 --> 01:06:02.510 align:middle line:84% or those upgrades are 30, 50, I sell much fewer when 01:06:02.510 --> 01:06:06.620 align:middle line:90% I have higher prices. 01:06:06.620 --> 01:06:13.370 align:middle line:84% So when we add together these two effects, the rank effects 01:06:13.370 --> 01:06:16.383 align:middle line:84% and the own price effects, this is giving us the effects 01:06:16.383 --> 01:06:17.550 align:middle line:90% of the different categories. 01:06:17.550 --> 01:06:20.220 align:middle line:84% But I'll focus on the 128 megabyte PC100. 01:06:20.220 --> 01:06:23.180 align:middle line:84% What I get is this incredibly elastic demand 01:06:23.180 --> 01:06:27.810 align:middle line:84% for the low quality memory, and then somewhat elastic cost 01:06:27.810 --> 01:06:29.730 align:middle line:90% for the other memory. 01:06:29.730 --> 01:06:32.280 align:middle line:84% I get these two notable off-diagonal elements 01:06:32.280 --> 01:06:34.470 align:middle line:84% saying that there's that loss leader effect. 01:06:34.470 --> 01:06:36.180 align:middle line:84% The other off-diagonal ones down here 01:06:36.180 --> 01:06:37.920 align:middle line:84% are positive, as you would expect. 01:06:37.920 --> 01:06:40.980 align:middle line:84% When I raise my price on medium quality memory, 01:06:40.980 --> 01:06:43.308 align:middle line:84% I sell a little bit more low quality memory, 01:06:43.308 --> 01:06:45.850 align:middle line:84% because we'll substitute away from the medium quality memory. 01:06:45.850 --> 01:06:47.850 align:middle line:84% I'll sell a little bit more high quality memory. 01:06:47.850 --> 01:06:49.505 align:middle line:84% But again, these are small effects. 01:06:49.505 --> 01:06:54.150 align:middle line:90% 01:06:54.150 --> 01:06:56.280 align:middle line:90% I said that. 01:06:56.280 --> 01:06:58.830 align:middle line:84% Those are the loss leader effects. 01:06:58.830 --> 01:06:59.945 align:middle line:90% I also said that. 01:06:59.945 --> 01:07:03.570 align:middle line:90% 01:07:03.570 --> 01:07:05.060 align:middle line:84% Anyway, so we've got this-- again, 01:07:05.060 --> 01:07:07.730 align:middle line:84% just summarizing this super elastic demand 01:07:07.730 --> 01:07:11.150 align:middle line:84% for the low quality memory and price sensitive, but not 01:07:11.150 --> 01:07:14.833 align:middle line:84% nearly as elastic demand for the medium and high quality memory. 01:07:14.833 --> 01:07:16.250 align:middle line:84% Naively, this is going to make you 01:07:16.250 --> 01:07:17.360 align:middle line:84% think that these products are going 01:07:17.360 --> 01:07:19.850 align:middle line:84% to be sold at higher markups because demand is less price 01:07:19.850 --> 01:07:23.140 align:middle line:90% sensitive. 01:07:23.140 --> 01:07:26.890 align:middle line:84% Question now-- I think the paper's biggest contribution is 01:07:26.890 --> 01:07:32.570 align:middle line:84% saying, we know that this add-on pricing theory is correct. 01:07:32.570 --> 01:07:34.360 align:middle line:84% If you sell products, if you have-- 01:07:34.360 --> 01:07:38.740 align:middle line:84% get people to compete by only advertising low quality 01:07:38.740 --> 01:07:40.210 align:middle line:90% products, and then having to-- 01:07:40.210 --> 01:07:43.027 align:middle line:84% selling the add-ons at high monopoly prices, 01:07:43.027 --> 01:07:44.860 align:middle line:84% we're going to get higher equilibrium prices 01:07:44.860 --> 01:07:47.352 align:middle line:90% than we would get without that. 01:07:47.352 --> 01:07:52.490 align:middle line:84% The question is, how large is that effect quantitatively? 01:07:52.490 --> 01:07:55.420 align:middle line:84% Does that affect-- is that effect large 01:07:55.420 --> 01:07:57.560 align:middle line:84% enough quantitatively to explain what's going on? 01:07:57.560 --> 01:07:59.570 align:middle line:84% And in fact, how big are the markups here? 01:07:59.570 --> 01:08:02.710 align:middle line:84% Are they explained by this add-on pricing theory? 01:08:02.710 --> 01:08:06.130 align:middle line:84% To do that, the paper has a generalization of what was 01:08:06.130 --> 01:08:10.930 align:middle line:84% in the QJE paper, noting that a formula for price cost margins-- 01:08:10.930 --> 01:08:14.920 align:middle line:84% if you remember, standard single good price-- 01:08:14.920 --> 01:08:19.359 align:middle line:84% standard single good monopoly pricing 01:08:19.359 --> 01:08:23.050 align:middle line:84% is price minus cost over price equals minus 1 01:08:23.050 --> 01:08:26.109 align:middle line:90% over the elasticity of demand. 01:08:26.109 --> 01:08:30.979 align:middle line:84% What I note here is that there's a generalization of that model, 01:08:30.979 --> 01:08:33.859 align:middle line:84% of that formula to a situation where you have a firm selling 01:08:33.859 --> 01:08:36.529 align:middle line:90% both a base good and an add-on. 01:08:36.529 --> 01:08:40.399 align:middle line:84% And what this is, so this is a generalization 01:08:40.399 --> 01:08:41.700 align:middle line:90% of the price cost margin. 01:08:41.700 --> 01:08:44.540 align:middle line:84% So this is the price for low quality memory 01:08:44.540 --> 01:08:47.750 align:middle line:84% plus the x is going to be the fraction who 01:08:47.750 --> 01:08:50.930 align:middle line:84% upgrade when you're competing against another firm, 01:08:50.930 --> 01:08:52.189 align:middle line:90% and your price is p1L. 01:08:52.189 --> 01:08:53.899 align:middle line:90% The other firm's price is p2L. 01:08:53.899 --> 01:08:58.010 align:middle line:84% So this is price plus fractional upgrade times the margin 01:08:58.010 --> 01:08:59.430 align:middle line:90% times the upgrade price. 01:08:59.430 --> 01:09:02.720 align:middle line:84% You can think of this as what's your equilibrium price 01:09:02.720 --> 01:09:04.573 align:middle line:84% per customer, taking into account 01:09:04.573 --> 01:09:05.990 align:middle line:84% that some pay more than they think 01:09:05.990 --> 01:09:08.420 align:middle line:84% they're going to pay when they click on that link, 01:09:08.420 --> 01:09:10.729 align:middle line:84% minus the cost of providing that good, 01:09:10.729 --> 01:09:12.643 align:middle line:84% which is the cost of low quality memory, 01:09:12.643 --> 01:09:14.810 align:middle line:84% plus the fractional upgrade times the cost of buying 01:09:14.810 --> 01:09:15.899 align:middle line:90% the upgrade. 01:09:15.899 --> 01:09:17.850 align:middle line:84% And then, again, this is the price. 01:09:17.850 --> 01:09:20.300 align:middle line:84% So it's just price minus cost over price, 01:09:20.300 --> 01:09:23.060 align:middle line:84% where in price, I've included the fraction upgrading 01:09:23.060 --> 01:09:24.210 align:middle line:90% times the upgrade price. 01:09:24.210 --> 01:09:26.760 align:middle line:84% So this is per consumer amount paid. 01:09:26.760 --> 01:09:31.580 align:middle line:84% What you get is that price minus cost over price is minus 1 01:09:31.580 --> 01:09:35.250 align:middle line:84% over epsilon times 1 plus something. 01:09:35.250 --> 01:09:37.550 align:middle line:84% And I sometimes refer to this as the adverse selection 01:09:37.550 --> 01:09:38.729 align:middle line:90% multiplier. 01:09:38.729 --> 01:09:41.640 align:middle line:84% The worse is the adverse selection multiplier, 01:09:41.640 --> 01:09:46.410 align:middle line:84% the higher is your markup relative to 1 over epsilon. 01:09:46.410 --> 01:09:49.580 align:middle line:84% The key term here is this one, which 01:09:49.580 --> 01:09:53.779 align:middle line:84% is, what's the upgrade profit times how 01:09:53.779 --> 01:09:58.280 align:middle line:84% does the fraction who choose to upgrade increase 01:09:58.280 --> 01:10:00.150 align:middle line:90% when your price increases? 01:10:00.150 --> 01:10:02.210 align:middle line:84% So as you get further down on the screen, 01:10:02.210 --> 01:10:03.950 align:middle line:84% have less of an adverse selection, 01:10:03.950 --> 01:10:06.780 align:middle line:84% how much more successful are you at selling upgrades? 01:10:06.780 --> 01:10:10.820 align:middle line:84% And this formula predicts that as the adverse selection 01:10:10.820 --> 01:10:13.280 align:middle line:84% problem gets worse, the dx / dp is bigger. 01:10:13.280 --> 01:10:15.720 align:middle line:84% And the p upgrade minus cu is bigger. 01:10:15.720 --> 01:10:18.500 align:middle line:84% You're going to have a larger margin relative to the minus 1 01:10:18.500 --> 01:10:22.010 align:middle line:84% over epsilon, which is the naive level you would expect 01:10:22.010 --> 01:10:24.260 align:middle line:90% from the single good intuition. 01:10:24.260 --> 01:10:27.260 align:middle line:84% Anyway, so the paper shows that this is true 01:10:27.260 --> 01:10:30.320 align:middle line:84% if you define epsilon to be this thing we call 01:10:30.320 --> 01:10:32.390 align:middle line:84% the overall elasticity, which looks 01:10:32.390 --> 01:10:36.140 align:middle line:84% an awful lot like a derivative of low quality 01:10:36.140 --> 01:10:38.185 align:middle line:84% demand with respect to low quality price. 01:10:38.185 --> 01:10:42.950 align:middle line:90% 01:10:42.950 --> 01:10:44.630 align:middle line:84% What we do is-- again, this paper 01:10:44.630 --> 01:10:47.600 align:middle line:84% is unusual in that we have this cost data. 01:10:47.600 --> 01:10:50.900 align:middle line:84% So we have the daily cost for each of these modules. 01:10:50.900 --> 01:10:53.750 align:middle line:84% We also know just from talking to the proprietor of the firm, 01:10:53.750 --> 01:10:55.190 align:middle line:90% what does it cost-- 01:10:55.190 --> 01:10:56.673 align:middle line:90% it's not rocket science. 01:10:56.673 --> 01:10:58.340 align:middle line:84% What does it cost to hire a minimum wage 01:10:58.340 --> 01:11:01.130 align:middle line:84% worker to take the memory module off the shelf, 01:11:01.130 --> 01:11:04.460 align:middle line:84% put it in a cardboard box, put one piece of tape on it, 01:11:04.460 --> 01:11:06.230 align:middle line:84% run it through the postage meter, 01:11:06.230 --> 01:11:07.682 align:middle line:90% and send it to the customer? 01:11:07.682 --> 01:11:09.140 align:middle line:84% And so the firm was able to give us 01:11:09.140 --> 01:11:12.740 align:middle line:84% very good estimates for what is the cost of sending a memory 01:11:12.740 --> 01:11:13.340 align:middle line:90% module-- 01:11:13.340 --> 01:11:17.120 align:middle line:84% of packing a memory module into a cardboard box and sending it. 01:11:17.120 --> 01:11:19.470 align:middle line:84% So we're able to estimate markups here. 01:11:19.470 --> 01:11:23.540 align:middle line:84% And what we find is that, qualitatively, these loss 01:11:23.540 --> 01:11:25.970 align:middle line:84% leaders, the low quality memory, it's being-- it actually 01:11:25.970 --> 01:11:28.140 align:middle line:84% is being sold at a loss most of the time. 01:11:28.140 --> 01:11:32.210 align:middle line:84% So here it's approximately zero markups. 01:11:32.210 --> 01:11:36.630 align:middle line:84% But low quality memory's average markup is minus 0.7%. 01:11:36.630 --> 01:11:39.270 align:middle line:84% Medium quality memory has got a 17% markup. 01:11:39.270 --> 01:11:44.050 align:middle line:84% High quality memory has got a 27% markup. 01:11:44.050 --> 01:11:47.340 align:middle line:84% So overall, when you add up the 12 units of this, two of this, 01:11:47.340 --> 01:11:52.360 align:middle line:84% two of this, the average mark up is 7%. 01:11:52.360 --> 01:11:53.830 align:middle line:84% How does that compare with what you 01:11:53.830 --> 01:11:57.220 align:middle line:84% would have expected had this add-on strategy not 01:11:57.220 --> 01:11:58.070 align:middle line:90% been going on? 01:11:58.070 --> 01:12:04.010 align:middle line:84% Well, the overall elasticity is minus 23.9. 01:12:04.010 --> 01:12:07.180 align:middle line:84% So if you had the simple price minus cost 01:12:07.180 --> 01:12:09.400 align:middle line:84% over price is 1 over epsilon pricing, 01:12:09.400 --> 01:12:13.660 align:middle line:90% you'd expect a 4.2% markup. 01:12:13.660 --> 01:12:17.080 align:middle line:90% And what we see is 7.7. 01:12:17.080 --> 01:12:20.750 align:middle line:84% Can we explain why we're seeing 7.7 instead of 4.2? 01:12:20.750 --> 01:12:24.250 align:middle line:84% Well, if each of the things in this adverse selection 01:12:24.250 --> 01:12:26.740 align:middle line:84% multiplier are also things that we're 01:12:26.740 --> 01:12:28.660 align:middle line:90% estimating our demand system. 01:12:28.660 --> 01:12:32.740 align:middle line:84% So when I estimate those, I find that my estimate 01:12:32.740 --> 01:12:35.300 align:middle line:84% for the adverse selection multiplier is two. 01:12:35.300 --> 01:12:38.770 align:middle line:84% So when we estimated this demand system 01:12:38.770 --> 01:12:41.020 align:middle line:84% without using any supply side information, 01:12:41.020 --> 01:12:43.990 align:middle line:84% without imposing that the firms were maximizing profits, 01:12:43.990 --> 01:12:46.450 align:middle line:84% and yet what we find is that when we look at the estimated 01:12:46.450 --> 01:12:49.890 align:middle line:84% parameters, with the estimated parameters, 01:12:49.890 --> 01:12:52.380 align:middle line:84% the equilibrium markup prediction 01:12:52.380 --> 01:12:54.745 align:middle line:84% is almost exactly the true markup prediction. 01:12:54.745 --> 01:12:57.120 align:middle line:84% So this is saying that, we estimate the adverse selection 01:12:57.120 --> 01:12:58.530 align:middle line:90% multiplier is two. 01:12:58.530 --> 01:13:01.860 align:middle line:84% And that does seem to explain why profits margins are 01:13:01.860 --> 01:13:04.650 align:middle line:90% 7% instead of 4.2%. 01:13:04.650 --> 01:13:07.180 align:middle line:84% We do this across several different categories. 01:13:07.180 --> 01:13:09.720 align:middle line:84% It predicts that the adverse selection is even more 01:13:09.720 --> 01:13:12.160 align:middle line:90% important for PC133 memory. 01:13:12.160 --> 01:13:15.230 align:middle line:84% It predicts that it's about the same for the 256 megabyte 01:13:15.230 --> 01:13:15.730 align:middle line:90% memory. 01:13:15.730 --> 01:13:20.730 align:middle line:84% And in each case, we find that the 1 over epsilon formula 01:13:20.730 --> 01:13:22.920 align:middle line:90% would say you'd get 3.6% markup. 01:13:22.920 --> 01:13:25.480 align:middle line:90% We actually get an 11.5% markup. 01:13:25.480 --> 01:13:28.170 align:middle line:84% But that's perfectly explained by this adverse selection 01:13:28.170 --> 01:13:32.400 align:middle line:84% multiplier that says, yes, it should multiply markups by 3.5%. 01:13:32.400 --> 01:13:34.650 align:middle line:84% So I think the basic takeaway from this paper 01:13:34.650 --> 01:13:39.680 align:middle line:84% is that one technique that firms are using for obfuscation 01:13:39.680 --> 01:13:44.100 align:middle line:84% here is, well, one technique is making things annoying. 01:13:44.100 --> 01:13:47.960 align:middle line:84% That's going into what determines these-- 01:13:47.960 --> 01:13:49.940 align:middle line:84% what determines the sensitivity to rank 01:13:49.940 --> 01:13:51.295 align:middle line:90% and how quickly people give up. 01:13:51.295 --> 01:13:53.480 align:middle line:84% The second thing that's going into margins 01:13:53.480 --> 01:13:57.890 align:middle line:84% is creating this adverse selection effect 01:13:57.890 --> 01:13:59.390 align:middle line:90% by offering add-ons. 01:13:59.390 --> 01:14:02.180 align:middle line:84% And the amount by which markups go 01:14:02.180 --> 01:14:04.490 align:middle line:84% up relative to the single good model 01:14:04.490 --> 01:14:06.203 align:middle line:84% seems to be almost exactly in line 01:14:06.203 --> 01:14:07.620 align:middle line:84% with the adverse selection effect. 01:14:07.620 --> 01:14:08.990 align:middle line:84% So it seems like, in some sense, we 01:14:08.990 --> 01:14:11.157 align:middle line:84% don't need anything other than the adverse selection 01:14:11.157 --> 01:14:12.890 align:middle line:84% effect and the level of search costs 01:14:12.890 --> 01:14:14.760 align:middle line:90% to explain what's going on here. 01:14:14.760 --> 01:14:18.360 align:middle line:84% It does seem like the adverse selection has an effect. 01:14:18.360 --> 01:14:21.300 align:middle line:84% It has an effect as big as we would expect it to be. 01:14:21.300 --> 01:14:25.280 align:middle line:84% And it's not a puzzle why markups in this market 01:14:25.280 --> 01:14:30.290 align:middle line:84% end up what they are, which is markups are quite low, but not 01:14:30.290 --> 01:14:33.080 align:middle line:84% so low that these firms, which are bare bones 01:14:33.080 --> 01:14:35.320 align:middle line:84% firms with no advertising, can't survive. 01:14:35.320 --> 01:14:40.350 align:middle line:90% 01:14:40.350 --> 01:14:41.900 align:middle line:84% And I think the rest of these slides 01:14:41.900 --> 01:14:43.560 align:middle line:84% just have the same things I just said. 01:14:43.560 --> 01:14:46.250 align:middle line:84% So yeah, here's where you see the low quality 01:14:46.250 --> 01:14:49.520 align:middle line:84% memory has got negative, and the others have positive. 01:14:49.520 --> 01:14:51.750 align:middle line:84% There's the adverse selection effect. 01:14:51.750 --> 01:14:55.370 align:middle line:90% There are the predictions. 01:14:55.370 --> 01:14:58.490 align:middle line:84% And then this summary of the conclusions 01:14:58.490 --> 01:15:00.793 align:middle line:84% that I told you that observed markups 01:15:00.793 --> 01:15:02.960 align:middle line:84% are very close to ones you'd predict from the demand 01:15:02.960 --> 01:15:04.580 align:middle line:90% estimates. 01:15:04.580 --> 01:15:07.760 align:middle line:84% Adverse selection is roughly doubling average markups. 01:15:07.760 --> 01:15:11.150 align:middle line:84% And markups for low quality products are very low. 01:15:11.150 --> 01:15:15.710 align:middle line:84% So in my final 15 minutes, I'm going to give my final paper, 01:15:15.710 --> 01:15:19.093 align:middle line:84% which is Galenianos and Gavazza's paper about 01:15:19.093 --> 01:15:21.260 align:middle line:84% "Regulatory Intervention in Consumer Search Markets, 01:15:21.260 --> 01:15:23.000 align:middle line:90% The Case of Credit Cards." 01:15:23.000 --> 01:15:24.740 align:middle line:84% One thing that I often do in this course 01:15:24.740 --> 01:15:27.200 align:middle line:84% is tell students that it's a mistake 01:15:27.200 --> 01:15:29.420 align:middle line:84% to be too ambitious when you're trying to come up 01:15:29.420 --> 01:15:31.100 align:middle line:84% with paper topics, or you don't need 01:15:31.100 --> 01:15:33.650 align:middle line:90% to be so original and creative. 01:15:33.650 --> 01:15:35.810 align:middle line:84% It's a great idea to just take someone else's paper 01:15:35.810 --> 01:15:37.460 align:middle line:84% and improve it, because if you take someone who's 01:15:37.460 --> 01:15:39.790 align:middle line:84% at the state of the art, you improve what they're doing, 01:15:39.790 --> 01:15:41.450 align:middle line:84% you know you have a state of the art paper, 01:15:41.450 --> 01:15:42.800 align:middle line:84% whereas if you start from scratch, 01:15:42.800 --> 01:15:44.758 align:middle line:84% you're going to end up-- you're going to end up 01:15:44.758 --> 01:15:47.300 align:middle line:90% very much behind the frontier. 01:15:47.300 --> 01:15:50.180 align:middle line:84% Normally, this is advice for writing theoretical papers, 01:15:50.180 --> 01:15:51.885 align:middle line:84% because you take a theoretical paper. 01:15:51.885 --> 01:15:53.510 align:middle line:84% You can just-- it's a well-known thing. 01:15:53.510 --> 01:15:54.870 align:middle line:90% You take a theoretical model. 01:15:54.870 --> 01:15:56.010 align:middle line:90% You improve on that. 01:15:56.010 --> 01:15:59.633 align:middle line:84% You've got a new and better model. 01:15:59.633 --> 01:16:01.050 align:middle line:84% The thing I want to point out here 01:16:01.050 --> 01:16:03.600 align:middle line:84% is one can do that in empirical work as well. 01:16:03.600 --> 01:16:06.390 align:middle line:84% One of my earliest published papers 01:16:06.390 --> 01:16:08.160 align:middle line:90% was a paper where I took-- 01:16:08.160 --> 01:16:11.070 align:middle line:84% you'll see later in the semester where I took a published 01:16:11.070 --> 01:16:12.520 align:middle line:90% empirical paper by Rob Porter. 01:16:12.520 --> 01:16:15.490 align:middle line:84% I just did some additional things on it. 01:16:15.490 --> 01:16:18.240 align:middle line:84% That was one of my first published papers. 01:16:18.240 --> 01:16:20.910 align:middle line:84% Galenianos and Gavazza decades later are actually 01:16:20.910 --> 01:16:23.192 align:middle line:84% doing the same thing here, with the idea that they 01:16:23.192 --> 01:16:24.900 align:middle line:84% must have read Stango and Zinman's paper, 01:16:24.900 --> 01:16:26.317 align:middle line:84% and they thought Stango and Zinman 01:16:26.317 --> 01:16:29.400 align:middle line:84% is a very nice paper, showing that search costs seem 01:16:29.400 --> 01:16:31.060 align:middle line:90% to be related to markups. 01:16:31.060 --> 01:16:34.140 align:middle line:84% And what they're trying to do is, say, 01:16:34.140 --> 01:16:36.180 align:middle line:84% Stango and Zinman said that search costs 01:16:36.180 --> 01:16:37.480 align:middle line:90% are related to markups. 01:16:37.480 --> 01:16:40.200 align:middle line:84% Can we actually turn their paper into 01:16:40.200 --> 01:16:43.290 align:middle line:84% a structural, quantitative paper, 01:16:43.290 --> 01:16:46.952 align:middle line:84% where we say, can search costs explain 01:16:46.952 --> 01:16:48.660 align:middle line:84% the magnitudes of everything that they're 01:16:48.660 --> 01:16:49.577 align:middle line:90% finding in their data? 01:16:49.577 --> 01:16:52.650 align:middle line:84% Or how would we say, is their data really consistent 01:16:52.650 --> 01:16:56.340 align:middle line:84% with the search cost explanation versus another explanation? 01:16:56.340 --> 01:16:58.990 align:middle line:84% It's also, I think, a very nice paper 01:16:58.990 --> 01:17:02.560 align:middle line:84% in that it's doing something we don't do often. 01:17:02.560 --> 01:17:04.540 align:middle line:84% It's doing something different in IO. 01:17:04.540 --> 01:17:07.060 align:middle line:84% Like many people have this view that every IO paper has 01:17:07.060 --> 01:17:09.460 align:middle line:84% to have this really, really complicated individual level 01:17:09.460 --> 01:17:13.570 align:middle line:84% data set and estimate this sophisticated model observation 01:17:13.570 --> 01:17:14.780 align:middle line:90% by observation. 01:17:14.780 --> 01:17:17.770 align:middle line:84% And what Galenianos and Gavazza do in this paper 01:17:17.770 --> 01:17:21.470 align:middle line:84% is they're basically doing a macro style calibration. 01:17:21.470 --> 01:17:24.730 align:middle line:84% And I think it's not what other people would 01:17:24.730 --> 01:17:27.250 align:middle line:84% have done with this data, but I think that it really 01:17:27.250 --> 01:17:29.500 align:middle line:84% makes the case that doing a calibration like this 01:17:29.500 --> 01:17:32.020 align:middle line:84% can be a very interesting way to think about 01:17:32.020 --> 01:17:35.120 align:middle line:84% whether a theory really explains the data or not. 01:17:35.120 --> 01:17:36.880 align:middle line:84% And do we always need to do the same kind 01:17:36.880 --> 01:17:38.330 align:middle line:90% of structural estimations we do? 01:17:38.330 --> 01:17:42.830 align:middle line:84% And I think it's a very elegant paper. 01:17:42.830 --> 01:17:46.150 align:middle line:84% It also makes some nice technical observations. 01:17:46.150 --> 01:17:48.018 align:middle line:84% And so anyway, I'll do it quickly here, 01:17:48.018 --> 01:17:49.310 align:middle line:90% because I don't have much time. 01:17:49.310 --> 01:17:55.920 align:middle line:84% So let me say, the basic thing is what they show in the paper-- 01:17:55.920 --> 01:17:57.760 align:middle line:84% think back to what Stango and Zinman showed. 01:17:57.760 --> 01:17:59.718 align:middle line:84% What Stango and Zinman showed is that consumers 01:17:59.718 --> 01:18:03.750 align:middle line:84% get very different offers in their offer database, 01:18:03.750 --> 01:18:06.810 align:middle line:84% and that consumers end up paying very different amounts. 01:18:06.810 --> 01:18:11.240 align:middle line:84% And so we have these sort of two data sets. 01:18:11.240 --> 01:18:14.120 align:middle line:84% We have this offered data set, and then we 01:18:14.120 --> 01:18:17.000 align:middle line:84% have the sort of what consumers-- 01:18:17.000 --> 01:18:22.070 align:middle line:84% the accepted offers, what consumers end up paying. 01:18:22.070 --> 01:18:25.970 align:middle line:84% And the starting point for Galenianos and Gavazza 01:18:25.970 --> 01:18:32.030 align:middle line:84% is think about we have these two separate data sets. 01:18:32.030 --> 01:18:36.110 align:middle line:84% The offers are going to lead to accepted offers 01:18:36.110 --> 01:18:37.740 align:middle line:90% based on consumer preferences. 01:18:37.740 --> 01:18:40.400 align:middle line:84% So if you think about what consumer preferences look like, 01:18:40.400 --> 01:18:43.310 align:middle line:84% what are their search costs, how much they view these offers as 01:18:43.310 --> 01:18:47.000 align:middle line:84% differentiated, that's going to lead us to some mapping 01:18:47.000 --> 01:18:49.010 align:middle line:84% that maps offers that they receive 01:18:49.010 --> 01:18:50.630 align:middle line:90% into what offers they accept. 01:18:50.630 --> 01:18:52.880 align:middle line:84% And notice that you can think about two different ways 01:18:52.880 --> 01:18:54.680 align:middle line:90% to explain the mapping. 01:18:54.680 --> 01:18:57.680 align:middle line:84% We know that these offers are quite dispersed. 01:18:57.680 --> 01:19:00.350 align:middle line:84% And we know that these accepted offers are fairly dispersed 01:19:00.350 --> 01:19:01.820 align:middle line:90% and quite high. 01:19:01.820 --> 01:19:04.290 align:middle line:84% You could explain this mapping in two different ways. 01:19:04.290 --> 01:19:06.437 align:middle line:84% One way would be that consumers have search costs. 01:19:06.437 --> 01:19:08.520 align:middle line:84% Therefore, they don't open a lot of the envelopes. 01:19:08.520 --> 01:19:10.520 align:middle line:84% Therefore, they don't all take the lowest offer. 01:19:10.520 --> 01:19:12.262 align:middle line:84% And therefore, they pay high prices, 01:19:12.262 --> 01:19:14.720 align:middle line:84% or they pay somewhat dispersed prices, because some of them 01:19:14.720 --> 01:19:18.620 align:middle line:84% open the low envelope first and pay that one. 01:19:18.620 --> 01:19:21.350 align:middle line:84% The other way to explain this same fact would 01:19:21.350 --> 01:19:23.857 align:middle line:84% be with consumers just have very idiosyncratic preferences. 01:19:23.857 --> 01:19:25.190 align:middle line:90% They get these different offers. 01:19:25.190 --> 01:19:26.630 align:middle line:90% They open all of them. 01:19:26.630 --> 01:19:28.550 align:middle line:84% But they care a lot whether their credit card 01:19:28.550 --> 01:19:31.020 align:middle line:84% says Bank of America or Capital One on it. 01:19:31.020 --> 01:19:33.210 align:middle line:84% Therefore, some of them pay much, 01:19:33.210 --> 01:19:35.690 align:middle line:84% much more than they need to, even after having opened 01:19:35.690 --> 01:19:38.190 align:middle line:84% all the envelopes, because they really like that credit card 01:19:38.190 --> 01:19:40.545 align:middle line:90% company. 01:19:40.545 --> 01:19:43.170 align:middle line:84% [INAUDIBLE] note is that we really 01:19:43.170 --> 01:19:45.120 align:middle line:84% can't just from thinking about this mapping 01:19:45.120 --> 01:19:48.750 align:middle line:84% tell the offer, the differentiation explanation, 01:19:48.750 --> 01:19:50.920 align:middle line:84% apart from the search cost explanation. 01:19:50.920 --> 01:19:54.690 align:middle line:84% But there's also a mapping that goes in this direction, where 01:19:54.690 --> 01:19:59.310 align:middle line:84% if you know what consumers are going to accept, 01:19:59.310 --> 01:20:01.620 align:middle line:84% the firms who are profit maximizing 01:20:01.620 --> 01:20:04.180 align:middle line:84% should be choosing their offers optimally. 01:20:04.180 --> 01:20:09.420 align:middle line:84% And if consumers had these very strong idiosyncratic preferences 01:20:09.420 --> 01:20:11.490 align:middle line:84% and viewed the offers as very differentiated, 01:20:11.490 --> 01:20:14.170 align:middle line:84% then all the firms would be sending high offers. 01:20:14.170 --> 01:20:16.920 align:middle line:84% Whereas if it's the search cost explanation, 01:20:16.920 --> 01:20:19.020 align:middle line:84% the search cost explanation could perhaps 01:20:19.020 --> 01:20:22.110 align:middle line:84% better explain why consumers are sending low offers. 01:20:22.110 --> 01:20:24.793 align:middle line:84% That's sort of intuitive, that we 01:20:24.793 --> 01:20:26.460 align:middle line:84% need the search cost to get the firms be 01:20:26.460 --> 01:20:30.580 align:middle line:84% setting these low and heterogeneous offers. 01:20:30.580 --> 01:20:34.480 align:middle line:84% But Galenianos and Gavazza are trying to say, quantitatively, 01:20:34.480 --> 01:20:35.330 align:middle line:90% is this true? 01:20:35.330 --> 01:20:41.440 align:middle line:84% Can we say that a model with search costs can explain 01:20:41.440 --> 01:20:44.650 align:middle line:84% the data quantitatively, whereas a model with differentiated 01:20:44.650 --> 01:20:46.210 align:middle line:90% products cannot? 01:20:46.210 --> 01:20:47.740 align:middle line:84% So what they do is they write down 01:20:47.740 --> 01:20:49.810 align:middle line:84% a model that's simple enough that they 01:20:49.810 --> 01:20:52.300 align:middle line:84% can estimate what the equilibrium is 01:20:52.300 --> 01:20:53.600 align:middle line:90% for a set of parameters. 01:20:53.600 --> 01:20:55.540 align:middle line:84% And then they say, we're going to calibrate 01:20:55.540 --> 01:20:58.550 align:middle line:84% the parameters to match some macro moments in the data. 01:20:58.550 --> 01:21:01.250 align:middle line:84% And then see if we can fit the data. 01:21:01.250 --> 01:21:03.000 align:middle line:84% So what does the model look like? 01:21:03.000 --> 01:21:07.950 align:middle line:84% So anyway, the model has buyers who have value z. 01:21:07.950 --> 01:21:10.930 align:middle line:84% z is distributed according to some distribution 01:21:10.930 --> 01:21:12.600 align:middle line:90% m on some interval. 01:21:12.600 --> 01:21:16.240 align:middle line:84% So this is your value of having a credit card from any firm 01:21:16.240 --> 01:21:17.170 align:middle line:90% whatsoever. 01:21:17.170 --> 01:21:18.700 align:middle line:84% In their model, some customers will 01:21:18.700 --> 01:21:21.070 align:middle line:84% have z relatively low-- won't get a credit card, 01:21:21.070 --> 01:21:24.070 align:middle line:90% or won't buy any offer. 01:21:24.070 --> 01:21:26.660 align:middle line:84% There's some lenders who have mass l. 01:21:26.660 --> 01:21:31.010 align:middle line:84% These lenders have heterogeneous costs of supplying funds, k, 01:21:31.010 --> 01:21:33.580 align:middle line:84% distributed g on k lower bar, k upper bar. 01:21:33.580 --> 01:21:36.230 align:middle line:90% 01:21:36.230 --> 01:21:40.302 align:middle line:84% As I said, these heterogeneous costs 01:21:40.302 --> 01:21:42.260 align:middle line:84% are going to be the way they generate the price 01:21:42.260 --> 01:21:44.060 align:middle line:90% dispersion in the data. 01:21:44.060 --> 01:21:48.440 align:middle line:84% In the Stahl model, you can generate price dispersion by-- 01:21:48.440 --> 01:21:51.320 align:middle line:84% generate price dispersion by all the firms are exactly identical, 01:21:51.320 --> 01:21:53.400 align:middle line:84% and they just randomize over prices. 01:21:53.400 --> 01:21:57.207 align:middle line:84% But there's an earlier literature on price dispersion, 01:21:57.207 --> 01:21:59.540 align:middle line:84% going back to Jennifer Reinganum has one of the earliest 01:21:59.540 --> 01:22:03.170 align:middle line:84% models of this variety, where you can also 01:22:03.170 --> 01:22:06.050 align:middle line:84% get price dispersion as a pure strategy phenomenon 01:22:06.050 --> 01:22:09.870 align:middle line:84% by simply allowing the firms to have asymmetric different costs. 01:22:09.870 --> 01:22:13.250 align:middle line:84% And then where the firms with identical costs randomize, 01:22:13.250 --> 01:22:14.900 align:middle line:84% you can have a purification of that, 01:22:14.900 --> 01:22:17.180 align:middle line:84% where the firms with low costs set low prices. 01:22:17.180 --> 01:22:19.110 align:middle line:84% Firms with high costs set high prices. 01:22:19.110 --> 01:22:22.860 align:middle line:84% So they're going to add this lender heterogeneity. 01:22:22.860 --> 01:22:25.520 align:middle line:84% And then, interesting model here is 01:22:25.520 --> 01:22:28.340 align:middle line:84% that rather than having buyers choose an equilibrium, how 01:22:28.340 --> 01:22:31.430 align:middle line:84% many offers to get, the buyers are going 01:22:31.430 --> 01:22:34.250 align:middle line:90% to exert an effort level s. 01:22:34.250 --> 01:22:36.560 align:middle line:84% And then if you exert effort level s-- you 01:22:36.560 --> 01:22:39.620 align:middle line:84% can think about I'm going to spend one hour on my computer, 01:22:39.620 --> 01:22:43.850 align:middle line:84% looking up credit cards online, and see what offers I can find. 01:22:43.850 --> 01:22:46.890 align:middle line:84% They have a Poisson random number of offers. 01:22:46.890 --> 01:22:54.680 align:middle line:84% So if I spend time s searching, I get s of-- 01:22:54.680 --> 01:22:57.980 align:middle line:84% in expectation, I get s times L offers, 01:22:57.980 --> 01:23:01.610 align:middle line:84% but I get a Poisson random number of offers. 01:23:01.610 --> 01:23:03.385 align:middle line:84% So I may actually find zero offers. 01:23:03.385 --> 01:23:04.010 align:middle line:90% I may find one. 01:23:04.010 --> 01:23:04.730 align:middle line:90% I may find two. 01:23:04.730 --> 01:23:07.190 align:middle line:90% I may find three. 01:23:07.190 --> 01:23:10.190 align:middle line:84% Doing this effort has a nice feature 01:23:10.190 --> 01:23:12.058 align:middle line:90% that s is a continuous variable. 01:23:12.058 --> 01:23:13.850 align:middle line:84% I can do a first order condition for what's 01:23:13.850 --> 01:23:17.210 align:middle line:84% the optimal amount of search effort s to make. 01:23:17.210 --> 01:23:19.820 align:middle line:84% You might think, well, yes, you made s continuous, 01:23:19.820 --> 01:23:21.290 align:middle line:84% but now you've got this nightmare, 01:23:21.290 --> 01:23:22.880 align:middle line:90% where instead of people having-- 01:23:22.880 --> 01:23:26.570 align:middle line:84% all having one offer, or two, or three, you've got-- 01:23:26.570 --> 01:23:28.860 align:middle line:84% I'm getting this Poisson random number of offers. 01:23:28.860 --> 01:23:30.560 align:middle line:84% How am I going to do that at all? 01:23:30.560 --> 01:23:32.930 align:middle line:84% An observation in this paper is that-- 01:23:32.930 --> 01:23:35.540 align:middle line:84% and there also been some others who have recently have 01:23:35.540 --> 01:23:38.900 align:middle line:84% done the same thing, is that this Poisson random number 01:23:38.900 --> 01:23:41.690 align:middle line:84% of quotes can actually be a more tractable model 01:23:41.690 --> 01:23:44.040 align:middle line:84% than having a deterministic number of quotes. 01:23:44.040 --> 01:23:46.550 align:middle line:84% And so because it's got this feature that effort 01:23:46.550 --> 01:23:49.460 align:middle line:84% is continuous, giving us a continuous first order 01:23:49.460 --> 01:23:53.570 align:middle line:84% condition, and the expected value to search 01:23:53.570 --> 01:23:56.260 align:middle line:84% is actually tractable, this can be a very nice framework. 01:23:56.260 --> 01:23:58.010 align:middle line:84% And I think other people are going to want 01:23:58.010 --> 01:23:59.490 align:middle line:90% to do this in the future. 01:23:59.490 --> 01:24:01.280 align:middle line:84% And let me say, also, if you remember back 01:24:01.280 --> 01:24:04.190 align:middle line:84% to intuition for Stahl, to get heterogeneity 01:24:04.190 --> 01:24:08.060 align:middle line:84% in accepted offers, what Stahl did is 01:24:08.060 --> 01:24:09.710 align:middle line:90% have some people who get very-- 01:24:09.710 --> 01:24:11.990 align:middle line:84% some people who have negative or low search 01:24:11.990 --> 01:24:13.410 align:middle line:90% costs, who search every firm. 01:24:13.410 --> 01:24:16.820 align:middle line:84% Some people with high search costs search all the firms. 01:24:16.820 --> 01:24:18.950 align:middle line:84% In Galenianos and Gavazza, they don't 01:24:18.950 --> 01:24:21.380 align:middle line:84% have randomness in the search costs, 01:24:21.380 --> 01:24:24.740 align:middle line:84% but what they have is this randomness in the value z. 01:24:24.740 --> 01:24:26.600 align:middle line:84% If you have a high value for a credit card, 01:24:26.600 --> 01:24:29.000 align:middle line:84% you're going to want to get a lot of quotes, 01:24:29.000 --> 01:24:31.980 align:middle line:84% because you know you need to buy at least one. 01:24:31.980 --> 01:24:33.980 align:middle line:84% If you have a low value z, you're 01:24:33.980 --> 01:24:35.570 align:middle line:84% going to do a little search effort, 01:24:35.570 --> 01:24:37.760 align:middle line:84% because you're going to be perfectly happy to not 01:24:37.760 --> 01:24:39.030 align:middle line:90% buy from anyone. 01:24:39.030 --> 01:24:43.290 align:middle line:84% And it's the heterogeneity in z that's 01:24:43.290 --> 01:24:45.290 align:middle line:84% going to lead to heterogeneity in search effort. 01:24:45.290 --> 01:24:47.930 align:middle line:84% That's going to lead to equilibrium price dispersion. 01:24:47.930 --> 01:24:50.990 align:middle line:84% Anyway, so going over the model quickly, 01:24:50.990 --> 01:24:53.360 align:middle line:84% buyers have this payoff from getting a card, which 01:24:53.360 --> 01:24:57.410 align:middle line:84% is their raw value from getting a card minus the disutility 01:24:57.410 --> 01:25:00.620 align:middle line:84% of the interest rate minus an epsilon, which allows 01:25:00.620 --> 01:25:02.690 align:middle line:84% for this sort of we're going to have the horse 01:25:02.690 --> 01:25:06.420 align:middle line:84% race between search costs versus product differentiation. 01:25:06.420 --> 01:25:09.650 align:middle line:84% So they allow for product differentiation. 01:25:09.650 --> 01:25:14.875 align:middle line:84% They write Fc for the CDF of a random draw of this plus this. 01:25:14.875 --> 01:25:16.250 align:middle line:84% That's obviously something that's 01:25:16.250 --> 01:25:17.900 align:middle line:90% determined in equilibrium. 01:25:17.900 --> 01:25:24.470 align:middle line:84% And then the seller's payoff is going to be the probability of-- 01:25:24.470 --> 01:25:26.690 align:middle line:84% the probability that someone buys from you 01:25:26.690 --> 01:25:29.480 align:middle line:84% if your interest rate is R times the payoff 01:25:29.480 --> 01:25:31.730 align:middle line:84% you get if your interest rate is R, which is R times 1 01:25:31.730 --> 01:25:34.890 align:middle line:84% minus rho, minus k, which is the cost of capital. 01:25:34.890 --> 01:25:37.310 align:middle line:84% And I see I've got a misprint here. 01:25:37.310 --> 01:25:39.680 align:middle line:84% R-- rho would be the default probability here, 01:25:39.680 --> 01:25:42.590 align:middle line:84% because your payoff is R times the probability 01:25:42.590 --> 01:25:44.970 align:middle line:84% the customer pays you back, which is 1 minus rho. 01:25:44.970 --> 01:25:49.340 align:middle line:84% So rho would be the default probability. 01:25:49.340 --> 01:25:52.070 align:middle line:84% Anyway, as I said, the Poisson search model 01:25:52.070 --> 01:25:54.770 align:middle line:90% is surprisingly tractable. 01:25:54.770 --> 01:25:56.550 align:middle line:90% Effort is a continuous variable. 01:25:56.550 --> 01:25:58.310 align:middle line:84% And then when you think about what's 01:25:58.310 --> 01:26:02.660 align:middle line:84% the probability that a consumer accepts an offer with total cost 01:26:02.660 --> 01:26:06.680 align:middle line:84% c, the probability that a consumer accepts 01:26:06.680 --> 01:26:09.440 align:middle line:84% an offer with total cost to them of c, 01:26:09.440 --> 01:26:12.500 align:middle line:84% you're summing over the number of offers they receive. 01:26:12.500 --> 01:26:14.210 align:middle line:90% n equals 0 to infinity. 01:26:14.210 --> 01:26:16.700 align:middle line:84% So n is the number of offers they receive. 01:26:16.700 --> 01:26:21.620 align:middle line:84% If they receive n offers other than yours, 01:26:21.620 --> 01:26:24.710 align:middle line:84% they're going to buy your product with probability 1 minus 01:26:24.710 --> 01:26:26.430 align:middle line:90% F of c to the n-th power. 01:26:26.430 --> 01:26:30.470 align:middle line:84% They only buy your product if every other product they get 01:26:30.470 --> 01:26:31.740 align:middle line:90% has higher cost to them. 01:26:31.740 --> 01:26:34.430 align:middle line:84% Or again, cost includes the financial cost 01:26:34.430 --> 01:26:36.870 align:middle line:84% and the idiosyncratic preference cost. 01:26:36.870 --> 01:26:39.860 align:middle line:84% So it's the sum of e to the minus-- 01:26:39.860 --> 01:26:41.720 align:middle line:84% so this is the Poisson probability 01:26:41.720 --> 01:26:45.560 align:middle line:84% that if they do effort, I guess if they do effort, a sub 01:26:45.560 --> 01:26:49.350 align:middle line:84% z, alpha sub z, they get n costs that's 01:26:49.350 --> 01:26:52.870 align:middle line:84% e to the minus alpha z, alpha z to the n over n factorial. 01:26:52.870 --> 01:26:55.200 align:middle line:84% If you look at this, it's this sum. 01:26:55.200 --> 01:26:58.080 align:middle line:84% And you notice it's got e to the minus alpha z over n factorial, 01:26:58.080 --> 01:27:00.880 align:middle line:84% alpha z to the n, 1 minus F of c to the n. 01:27:00.880 --> 01:27:04.180 align:middle line:84% So it's just like this and this go together. 01:27:04.180 --> 01:27:06.600 align:middle line:84% And so the probability they buy from you is just 01:27:06.600 --> 01:27:10.070 align:middle line:84% e to the minus alpha z times F of c. 01:27:10.070 --> 01:27:15.530 align:middle line:84% And then when I set a price, I'm setting an interest rate of-- 01:27:15.530 --> 01:27:17.600 align:middle line:84% I'm not setting a cost c to them. 01:27:17.600 --> 01:27:20.100 align:middle line:84% I'm setting an interest rate, and then they have an epsilon. 01:27:20.100 --> 01:27:22.235 align:middle line:84% So there is an integral over the epsilon 01:27:22.235 --> 01:27:25.010 align:middle line:84% to determine the probability of purchase. 01:27:25.010 --> 01:27:27.380 align:middle line:84% But anyway, we get this sort of nice 01:27:27.380 --> 01:27:29.780 align:middle line:84% closed form for the probability of purchase 01:27:29.780 --> 01:27:33.710 align:middle line:84% as a function of the interest rate 01:27:33.710 --> 01:27:35.940 align:middle line:84% that I do charge on a credit card. 01:27:35.940 --> 01:27:37.760 align:middle line:90% And so this sort of like-- 01:27:37.760 --> 01:27:39.700 align:middle line:84% in some sense, this tractability of the model 01:27:39.700 --> 01:27:42.200 align:middle line:84% is just a key step that makes it possible to solve the model 01:27:42.200 --> 01:27:44.450 align:middle line:90% given any set of parameters. 01:27:44.450 --> 01:27:45.600 align:middle line:90% What do they do? 01:27:45.600 --> 01:27:49.660 align:middle line:84% Well, they make parametric forms for many of the primitives. 01:27:49.660 --> 01:27:52.160 align:middle line:84% For example, assuming that the buyer's value distribution is 01:27:52.160 --> 01:27:54.830 align:middle line:84% log normal with parameters to be estimated. 01:27:54.830 --> 01:27:58.310 align:middle line:84% I guess this is off a working paper version that had 11 model 01:27:58.310 --> 01:28:00.030 align:middle line:90% parameters and 15 moments. 01:28:00.030 --> 01:28:02.042 align:middle line:84% I think the published version of the paper has-- 01:28:02.042 --> 01:28:03.500 align:middle line:84% referees always ask you to do more. 01:28:03.500 --> 01:28:05.540 align:middle line:84% I think the published paper has 24 model-- 01:28:05.540 --> 01:28:08.490 align:middle line:84% 28 model parameters, and 36 moments to be matched. 01:28:08.490 --> 01:28:13.140 align:middle line:84% But basically, you're setting model parameters-- 01:28:13.140 --> 01:28:16.170 align:middle line:84% parameters of the various distributions that are there 01:28:16.170 --> 01:28:21.070 align:middle line:84% to match a bunch of moments in the data, like, 01:28:21.070 --> 01:28:25.730 align:middle line:84% what is the 10th percentile of the accepted offers, 01:28:25.730 --> 01:28:28.240 align:middle line:84% median accepted offer, 90% offer? 01:28:28.240 --> 01:28:30.783 align:middle line:84% What fraction of consumers are getting at least two offers? 01:28:30.783 --> 01:28:32.200 align:middle line:84% What's the median number of offers 01:28:32.200 --> 01:28:33.650 align:middle line:90% conditional on getting two? 01:28:33.650 --> 01:28:36.790 align:middle line:84% So you're trying to explain simultaneously 01:28:36.790 --> 01:28:40.660 align:middle line:84% what are the accepted offers here 01:28:40.660 --> 01:28:43.040 align:middle line:84% and what are the offers being sent. 01:28:43.040 --> 01:28:45.440 align:middle line:84% We're not trying to explain them customer by customer. 01:28:45.440 --> 01:28:47.230 align:middle line:84% We're just trying to say, I want a model 01:28:47.230 --> 01:28:50.200 align:middle line:84% that gives this amount of distribution that's 01:28:50.200 --> 01:28:53.890 align:middle line:84% observed in the offers and the amount of dispersion that's 01:28:53.890 --> 01:28:55.760 align:middle line:90% observed in the accepted offers. 01:28:55.760 --> 01:28:59.050 align:middle line:84% And I need to match other parameters 01:28:59.050 --> 01:29:00.490 align:middle line:90% that I know from the data. 01:29:00.490 --> 01:29:03.640 align:middle line:84% Anyway, so what they do is they estimate these things. 01:29:03.640 --> 01:29:05.560 align:middle line:84% And then the-- I won't go through it, 01:29:05.560 --> 01:29:07.570 align:middle line:84% but the main thing they then have 01:29:07.570 --> 01:29:09.370 align:middle line:84% is with this calibrated model, they 01:29:09.370 --> 01:29:11.020 align:middle line:90% talk about a counterfactual. 01:29:11.020 --> 01:29:14.050 align:middle line:84% Counterfactual is, what would be the effect of capping 01:29:14.050 --> 01:29:15.320 align:middle line:90% credit card interest rates? 01:29:15.320 --> 01:29:18.800 align:middle line:84% So many people pay 35% interest on their credit cards. 01:29:18.800 --> 01:29:22.310 align:middle line:84% You could ask, suppose in instead of charging everybody 01:29:22.310 --> 01:29:24.710 align:middle line:84% 35% on the credit cards, you put a law 01:29:24.710 --> 01:29:30.860 align:middle line:84% that says you can't charge more than 27.5%, or 25%, or 22%? 01:29:30.860 --> 01:29:32.550 align:middle line:84% That's going to have two effects. 01:29:32.550 --> 01:29:34.040 align:middle line:84% One is that some people are going 01:29:34.040 --> 01:29:35.450 align:middle line:84% to pay much less on their credit cards, 01:29:35.450 --> 01:29:36.560 align:middle line:84% because they're going to get an offer that 01:29:36.560 --> 01:29:37.680 align:middle line:90% has a lower rate in the mail. 01:29:37.680 --> 01:29:39.660 align:middle line:84% They're going to buy an offer that is a lower rate. 01:29:39.660 --> 01:29:42.350 align:middle line:84% Some people are going to be hurt by that, because they can't get 01:29:42.350 --> 01:29:45.050 align:middle line:84% a credit card, because people would not 01:29:45.050 --> 01:29:47.640 align:middle line:84% be willing to lend to them at the lower 25% rate. 01:29:47.640 --> 01:29:50.510 align:middle line:84% They're just going to get no credit card at all. 01:29:50.510 --> 01:29:52.130 align:middle line:84% What they find-- do I have things 01:29:52.130 --> 01:29:55.280 align:middle line:84% here-- is that they find that actually here, 01:29:55.280 --> 01:29:59.030 align:middle line:84% their model predicts that laws capping credit card interest 01:29:59.030 --> 01:30:00.530 align:middle line:90% rates should work well. 01:30:00.530 --> 01:30:03.440 align:middle line:84% What we get is that when you cap interest rates 01:30:03.440 --> 01:30:11.244 align:middle line:84% at 27.5% with these parameters, that's good for consumers. 01:30:11.244 --> 01:30:13.930 align:middle line:84% Yes, it is true that some consumers are locked out 01:30:13.930 --> 01:30:16.540 align:middle line:84% of the market, but the number of consumers who are locked out 01:30:16.540 --> 01:30:19.300 align:middle line:84% of the market is fairly small, and many other consumers 01:30:19.300 --> 01:30:21.980 align:middle line:84% benefit from the drop in interest rates. 01:30:21.980 --> 01:30:23.740 align:middle line:90% They do notice that-- 01:30:23.740 --> 01:30:26.950 align:middle line:84% they do have results showing that these are not 01:30:26.950 --> 01:30:27.940 align:middle line:90% monotone effects. 01:30:27.940 --> 01:30:31.720 align:middle line:90% 01:30:31.720 --> 01:30:34.690 align:middle line:84% Lowering the-- what's best for consumers 01:30:34.690 --> 01:30:36.440 align:middle line:84% is not to have a cap, obviously, of zero, 01:30:36.440 --> 01:30:37.940 align:middle line:84% because then no one would be served. 01:30:37.940 --> 01:30:39.398 align:middle line:84% And in fact, in their data, I think 01:30:39.398 --> 01:30:42.640 align:middle line:84% they show that consumers are better off with a 27.5% cap 01:30:42.640 --> 01:30:43.360 align:middle line:90% than with no cap. 01:30:43.360 --> 01:30:45.740 align:middle line:84% They're better off still with a 25% cap. 01:30:45.740 --> 01:30:49.030 align:middle line:84% When you go to 22.5%, consumers start to be worse off. 01:30:49.030 --> 01:30:54.010 align:middle line:84% So anyway, I think it's a very nice paper in two dimensions. 01:30:54.010 --> 01:30:55.510 align:middle line:90% It's got this sort of-- 01:30:55.510 --> 01:30:58.420 align:middle line:84% it's got originality, and thinking of doing a calibration, 01:30:58.420 --> 01:31:01.060 align:middle line:84% rather than doing what we normally do in IO. 01:31:01.060 --> 01:31:04.330 align:middle line:84% It really makes the case that this model quantitatively 01:31:04.330 --> 01:31:05.870 align:middle line:90% can explain what goes on. 01:31:05.870 --> 01:31:08.810 align:middle line:84% And several of the key moments can be explained simultaneously 01:31:08.810 --> 01:31:09.677 align:middle line:90% by this model. 01:31:09.677 --> 01:31:11.510 align:middle line:84% And then it has this sort of counterfactual, 01:31:11.510 --> 01:31:14.297 align:middle line:84% saying that, here's a policy relevant question we care about. 01:31:14.297 --> 01:31:16.130 align:middle line:84% What would happen if you cap interest rates? 01:31:16.130 --> 01:31:17.960 align:middle line:84% And with their model, they predict 01:31:17.960 --> 01:31:21.920 align:middle line:84% that capping interest rates at still high, but moderate rates 01:31:21.920 --> 01:31:23.810 align:middle line:90% would be good for consumers. 01:31:23.810 --> 01:31:26.370 align:middle line:90% So that's all I've got for you. 01:31:26.370 --> 01:31:28.730 align:middle line:84% So I guess what we're going to do, 01:31:28.730 --> 01:31:30.680 align:middle line:84% we have a bit of a non-standard timing 01:31:30.680 --> 01:31:33.350 align:middle line:90% in the physical class at MIT. 01:31:33.350 --> 01:31:35.970 align:middle line:84% Jean Tirole is going to come give guest lectures next week. 01:31:35.970 --> 01:31:37.820 align:middle line:84% So we'll have two lectures from Jean, 01:31:37.820 --> 01:31:39.920 align:middle line:84% and then Tobias will come back the week after 01:31:39.920 --> 01:31:42.170 align:middle line:84% and talk about structural work on search. 01:31:42.170 --> 01:31:44.000 align:middle line:84% I guess in the OCW course, for those of you 01:31:44.000 --> 01:31:45.500 align:middle line:84% who are watching this there, I think 01:31:45.500 --> 01:31:48.530 align:middle line:84% we may flip back to the natural order, and have-- 01:31:48.530 --> 01:31:50.750 align:middle line:84% next lecture will be the third lecture on search, 01:31:50.750 --> 01:31:52.370 align:middle line:84% and then we'll have the guest lectures 01:31:52.370 --> 01:31:54.850 align:middle line:90% by Jean on platform markets. 01:31:54.850 --> 01:32:13.000 align:middle line:90%