WEBVTT 00:00:00.000 --> 00:00:01.960 align:middle line:90% [SQUEAKING] 00:00:01.960 --> 00:00:03.920 align:middle line:90% [RUSTLING] 00:00:03.920 --> 00:00:06.860 align:middle line:90% [CLICKING] 00:00:06.860 --> 00:00:18.150 align:middle line:90% 00:00:18.150 --> 00:00:20.160 align:middle line:90% PROFESSOR: Welcome back. 00:00:20.160 --> 00:00:24.870 align:middle line:84% This is the second lecture on empirical models of demand. 00:00:24.870 --> 00:00:29.460 align:middle line:84% So let me give you a little bit of a roadmap for today. 00:00:29.460 --> 00:00:34.560 align:middle line:84% So I'm going to first talk a little bit about the supply 00:00:34.560 --> 00:00:37.050 align:middle line:84% side of these static models of competition 00:00:37.050 --> 00:00:40.410 align:middle line:84% that we have been discussing, and that's 00:00:40.410 --> 00:00:44.750 align:middle line:84% actually quite useful even for informing the demand side. 00:00:44.750 --> 00:00:50.620 align:middle line:84% So I hope that this will become clear once I go through this. 00:00:50.620 --> 00:00:54.990 align:middle line:84% Then, I'll show you the empirical results in BLP-- 00:00:54.990 --> 00:00:57.300 align:middle line:84% Barry, Levinson, Pakes, the seminal paper 00:00:57.300 --> 00:01:02.790 align:middle line:84% that popularized the random coefficient model. 00:01:02.790 --> 00:01:05.069 align:middle line:84% And the results are actually quite 00:01:05.069 --> 00:01:08.100 align:middle line:84% nice to understand how these models work 00:01:08.100 --> 00:01:10.710 align:middle line:84% and what the case for these models is. 00:01:10.710 --> 00:01:14.970 align:middle line:84% And then I asked you to read Nevo's paper 00:01:14.970 --> 00:01:16.310 align:middle line:90% on the cereal market. 00:01:16.310 --> 00:01:18.370 align:middle line:84% So we'll discuss that, and then we'll 00:01:18.370 --> 00:01:20.080 align:middle line:90% actually stay in this industry. 00:01:20.080 --> 00:01:23.980 align:middle line:84% And we'll talk a little bit about the debate that 00:01:23.980 --> 00:01:27.550 align:middle line:84% has captivated many people in IO, 00:01:27.550 --> 00:01:32.140 align:middle line:84% which is the common ownership hypothesis. 00:01:32.140 --> 00:01:33.450 align:middle line:90% So let's jump in. 00:01:33.450 --> 00:01:36.550 align:middle line:90% 00:01:36.550 --> 00:01:38.740 align:middle line:90% So the supply side. 00:01:38.740 --> 00:01:42.640 align:middle line:84% So here is the objective function 00:01:42.640 --> 00:01:49.180 align:middle line:84% of a firm that owns potentially different products. 00:01:49.180 --> 00:01:53.140 align:middle line:84% This set of products here is calligraphic I, 00:01:53.140 --> 00:02:00.100 align:middle line:84% and it maximizes a sum of these individual level products. 00:02:00.100 --> 00:02:01.580 align:middle line:90% So this here is the margin. 00:02:01.580 --> 00:02:04.210 align:middle line:84% This is a quantity for product J, 00:02:04.210 --> 00:02:06.850 align:middle line:84% and this is a function of the vector of all 00:02:06.850 --> 00:02:08.150 align:middle line:90% the prices in the market. 00:02:08.150 --> 00:02:13.970 align:middle line:84% So think of this q here as coming out of your demand model. 00:02:13.970 --> 00:02:16.870 align:middle line:84% So it's the vector of market shares 00:02:16.870 --> 00:02:18.850 align:middle line:84% that you get out of your demand model times 00:02:18.850 --> 00:02:22.130 align:middle line:84% the number of consumers that are in the market. 00:02:22.130 --> 00:02:24.050 align:middle line:84% And so then there's some fixed cost. 00:02:24.050 --> 00:02:27.730 align:middle line:84% The fixed cost is not going to be relevant for what 00:02:27.730 --> 00:02:30.020 align:middle line:90% we're going to do today. 00:02:30.020 --> 00:02:32.690 align:middle line:84% And later on, you'll learn about models 00:02:32.690 --> 00:02:35.440 align:middle line:84% that actually allow you to get at those fixed costs, as well. 00:02:35.440 --> 00:02:38.025 align:middle line:90% 00:02:38.025 --> 00:02:42.690 align:middle line:84% So now, it's going to be very useful to define what 00:02:42.690 --> 00:02:44.730 align:middle line:90% is called an ownership matrix. 00:02:44.730 --> 00:02:49.080 align:middle line:84% So again, this is potentially a multi-product firm. 00:02:49.080 --> 00:02:53.220 align:middle line:84% The firm owns, let's say, several brands 00:02:53.220 --> 00:02:56.430 align:middle line:84% of cereals in the market, and so each row here 00:02:56.430 --> 00:03:00.270 align:middle line:84% tells us for that firm that indicates that row, which 00:03:00.270 --> 00:03:02.280 align:middle line:90% products does it own. 00:03:02.280 --> 00:03:06.920 align:middle line:84% So this is simply a matrix of binary entries. 00:03:06.920 --> 00:03:08.310 align:middle line:84% And now you can do the following. 00:03:08.310 --> 00:03:13.550 align:middle line:84% You can define a new matrix, omega, 00:03:13.550 --> 00:03:17.540 align:middle line:84% that multiplies this ownership matrix 00:03:17.540 --> 00:03:21.270 align:middle line:84% by the derivatives of each product 00:03:21.270 --> 00:03:24.155 align:middle line:84% with regard to all prices in the market. 00:03:24.155 --> 00:03:29.080 align:middle line:90% 00:03:29.080 --> 00:03:31.750 align:middle line:84% This comes out of the demand model. 00:03:31.750 --> 00:03:34.390 align:middle line:84% You can only construct those derivatives 00:03:34.390 --> 00:03:37.450 align:middle line:90% once you have the demand model. 00:03:37.450 --> 00:03:41.200 align:middle line:84% And so you get this new matrix, omega. 00:03:41.200 --> 00:03:43.360 align:middle line:84% And with this matrix omega, you can 00:03:43.360 --> 00:03:46.060 align:middle line:84% write the sets of first order conditions 00:03:46.060 --> 00:03:52.090 align:middle line:84% of all the firms in the market as this vector of quantities 00:03:52.090 --> 00:03:56.980 align:middle line:84% minus this omega matrix times prices minus marginal costs. 00:03:56.980 --> 00:03:59.500 align:middle line:84% These are all vectors, and this omega is this matrix 00:03:59.500 --> 00:04:01.502 align:middle line:90% that we defined here. 00:04:01.502 --> 00:04:05.980 align:middle line:84% And this allows us to invert those first order conditions 00:04:05.980 --> 00:04:09.360 align:middle line:90% to back out marginal cost. 00:04:09.360 --> 00:04:14.590 align:middle line:84% So this is simply a matrix operation. 00:04:14.590 --> 00:04:21.190 align:middle line:84% So what this shows us is that with demand estimates 00:04:21.190 --> 00:04:24.700 align:middle line:84% and a conduct model that actually allows us to construct 00:04:24.700 --> 00:04:28.060 align:middle line:84% these first order conditions, I need 00:04:28.060 --> 00:04:30.290 align:middle line:84% to know how firms are actually competing. 00:04:30.290 --> 00:04:32.900 align:middle line:90% 00:04:32.900 --> 00:04:36.440 align:middle line:84% As well as observed prices, we can recover 00:04:36.440 --> 00:04:37.810 align:middle line:90% a vector of marginal cost. 00:04:37.810 --> 00:04:40.850 align:middle line:90% 00:04:40.850 --> 00:04:45.470 align:middle line:84% What this also means is that the model 00:04:45.470 --> 00:04:47.750 align:middle line:84% is going to be always rationalized 00:04:47.750 --> 00:04:49.370 align:middle line:84% by a vector of marginal costs and is, 00:04:49.370 --> 00:04:53.270 align:middle line:84% therefore, not testable if we allow the marginal cost 00:04:53.270 --> 00:04:56.810 align:middle line:84% to be completely flexible and have no other way 00:04:56.810 --> 00:05:00.320 align:middle line:84% to discern whether the marginal costs that we back out are 00:05:00.320 --> 00:05:02.180 align:middle line:90% good marginal cost. 00:05:02.180 --> 00:05:05.185 align:middle line:84% So this comes back to the exercise that Nevo is doing. 00:05:05.185 --> 00:05:08.070 align:middle line:90% 00:05:08.070 --> 00:05:12.270 align:middle line:84% So we're always able to rationalize the model, 00:05:12.270 --> 00:05:17.535 align:middle line:84% but we cannot, without further assumptions, test the model. 00:05:17.535 --> 00:05:21.910 align:middle line:84% So again, just to reiterate the point, 00:05:21.910 --> 00:05:27.060 align:middle line:84% depending on what the true ownership structure looks like, 00:05:27.060 --> 00:05:31.710 align:middle line:84% or to what extent firms internalize profits 00:05:31.710 --> 00:05:35.130 align:middle line:84% of competitors, which we'll come to later today, 00:05:35.130 --> 00:05:37.590 align:middle line:84% will depend on how they want to price. 00:05:37.590 --> 00:05:41.400 align:middle line:84% And for any such conduct assumption, 00:05:41.400 --> 00:05:46.870 align:middle line:84% we can back out a vector of marginal cost. 00:05:46.870 --> 00:05:49.360 align:middle line:84% Now, what you can alternatively do, 00:05:49.360 --> 00:05:54.700 align:middle line:84% instead of allowing marginal cost to be completely flexible-- 00:05:54.700 --> 00:05:57.140 align:middle line:84% I mean, we would call this non-parametric 00:05:57.140 --> 00:06:00.710 align:middle line:84% because we're not imposing any assumptions 00:06:00.710 --> 00:06:02.320 align:middle line:90% on those marginal costs. 00:06:02.320 --> 00:06:10.332 align:middle line:84% We can specify some functional form for the marginal cost. 00:06:10.332 --> 00:06:12.040 align:middle line:84% So, for instance, in the case of cereals, 00:06:12.040 --> 00:06:15.670 align:middle line:84% we might know that this varies with the market price of sugar, 00:06:15.670 --> 00:06:21.160 align:middle line:84% wheat, corn, and so on, or advertising prices, shelf 00:06:21.160 --> 00:06:23.140 align:middle line:90% prices, and so on. 00:06:23.140 --> 00:06:26.290 align:middle line:84% And so if I'm pretty confident that I 00:06:26.290 --> 00:06:30.070 align:middle line:84% can write my marginal cost as a function of these things, 00:06:30.070 --> 00:06:33.235 align:middle line:84% I might be willing to impose this. 00:06:33.235 --> 00:06:36.990 align:middle line:90% 00:06:36.990 --> 00:06:40.410 align:middle line:84% So what this means now is that I can come from both sides, 00:06:40.410 --> 00:06:41.280 align:middle line:90% so to say. 00:06:41.280 --> 00:06:47.570 align:middle line:84% I can estimate my demand model with a conduct assumption, 00:06:47.570 --> 00:06:50.425 align:middle line:84% back out the vector of marginal cost. 00:06:50.425 --> 00:06:51.770 align:middle line:90% This is now data. 00:06:51.770 --> 00:06:55.250 align:middle line:84% For any given conduct assumption prices and market shares, 00:06:55.250 --> 00:06:57.750 align:middle line:90% this here is simply data. 00:06:57.750 --> 00:07:02.320 align:middle line:84% And this here is some linear function that I propose. 00:07:02.320 --> 00:07:06.910 align:middle line:90% 00:07:06.910 --> 00:07:10.045 align:middle line:90% Now, why might this be helpful? 00:07:10.045 --> 00:07:13.390 align:middle line:84% So you have read Nevo's paper today 00:07:13.390 --> 00:07:18.370 align:middle line:84% where he first estimated demand, and then tested 00:07:18.370 --> 00:07:22.700 align:middle line:84% the recovered markups against observed markups. 00:07:22.700 --> 00:07:25.850 align:middle line:84% What you can alternatively do, if you're fairly confident 00:07:25.850 --> 00:07:29.720 align:middle line:84% in your conduct assumption-- you believe that these firms are 00:07:29.720 --> 00:07:33.440 align:middle line:84% maximizing their own profits, potentially accounting for these 00:07:33.440 --> 00:07:35.750 align:middle line:90% multi-product effects-- 00:07:35.750 --> 00:07:40.520 align:middle line:84% you can estimate the demand and the supply side jointly. 00:07:40.520 --> 00:07:42.470 align:middle line:84% So you can-- from this equation here, 00:07:42.470 --> 00:07:44.815 align:middle line:84% which again, think of this here as data for a moment. 00:07:44.815 --> 00:07:47.630 align:middle line:90% 00:07:47.630 --> 00:07:49.280 align:middle line:84% For any given demand parameters, you 00:07:49.280 --> 00:07:54.370 align:middle line:84% can think of a vector of marginal cost that you back out. 00:07:54.370 --> 00:07:58.740 align:middle line:84% You could, from this, construct additional moments. 00:07:58.740 --> 00:08:00.510 align:middle line:84% Remember last time, we constructed 00:08:00.510 --> 00:08:06.030 align:middle line:84% moments interacting C with instruments on the demand side. 00:08:06.030 --> 00:08:08.100 align:middle line:84% We can construct a similar set of moments 00:08:08.100 --> 00:08:10.950 align:middle line:84% on the supply side, typically denoted omega, 00:08:10.950 --> 00:08:13.920 align:middle line:84% that we, again, interact with some instruments of things 00:08:13.920 --> 00:08:16.770 align:middle line:90% that are either exogenous, or-- 00:08:16.770 --> 00:08:21.230 align:middle line:84% included or excluded instruments on the supply side. 00:08:21.230 --> 00:08:25.480 align:middle line:84% Now, If I estimate everything jointly, 00:08:25.480 --> 00:08:28.180 align:middle line:84% the supply side will actually also impose restrictions 00:08:28.180 --> 00:08:31.380 align:middle line:90% on the demand side parameters. 00:08:31.380 --> 00:08:37.080 align:middle line:84% So just to give an example, so what happens, 00:08:37.080 --> 00:08:41.034 align:middle line:84% for instance, if you find that firms price 00:08:41.034 --> 00:08:42.909 align:middle line:84% on the inelastic portion of the demand curve? 00:08:42.909 --> 00:08:45.300 align:middle line:84% You may recover negative marginal cost. 00:08:45.300 --> 00:08:47.490 align:middle line:84% I mean, for many strategic models, 00:08:47.490 --> 00:08:49.510 align:middle line:84% you will be recovering negative marginal costs. 00:08:49.510 --> 00:08:53.800 align:middle line:84% You have to somehow rationalize why firms priced it low, 00:08:53.800 --> 00:08:56.520 align:middle line:84% and then it says, if they need some transfers 00:08:56.520 --> 00:08:59.680 align:middle line:84% from somewhere else to rationalize what's happening. 00:08:59.680 --> 00:09:03.960 align:middle line:90% 00:09:03.960 --> 00:09:10.930 align:middle line:84% I can, alternatively, impose that marginal costs are positive 00:09:10.930 --> 00:09:13.845 align:middle line:84% and use the supply side restrictions. 00:09:13.845 --> 00:09:17.800 align:middle line:84% And that means that when I guess a set of demand side parameters 00:09:17.800 --> 00:09:20.590 align:middle line:84% that tell me that the marginal costs are negative, 00:09:20.590 --> 00:09:22.960 align:middle line:84% but I force the marginal cost to be 00:09:22.960 --> 00:09:26.170 align:middle line:84% positive through a reasonable restriction, 00:09:26.170 --> 00:09:29.350 align:middle line:84% that this will lead to a violation of these moment 00:09:29.350 --> 00:09:31.410 align:middle line:90% conditions here. 00:09:31.410 --> 00:09:36.840 align:middle line:84% So it leads to larger deviations on the supply side moment 00:09:36.840 --> 00:09:39.480 align:middle line:84% conditions, and the supply side moments will, therefore, 00:09:39.480 --> 00:09:42.150 align:middle line:84% tell your minimization routine, hey, 00:09:42.150 --> 00:09:44.220 align:middle line:84% the demand side parameters might be such 00:09:44.220 --> 00:09:49.120 align:middle line:84% that I can't recover a good set of marginal cost. 00:09:49.120 --> 00:09:55.210 align:middle line:84% So this can be very helpful to estimate both jointly. 00:09:55.210 --> 00:09:57.940 align:middle line:84% Now in practice, there are not all that many papers 00:09:57.940 --> 00:09:59.320 align:middle line:90% that actually do that. 00:09:59.320 --> 00:10:02.860 align:middle line:84% So I would say the majority of papers 00:10:02.860 --> 00:10:05.360 align:middle line:84% just estimates the demand side separately, 00:10:05.360 --> 00:10:08.020 align:middle line:90% but BLP have done this. 00:10:08.020 --> 00:10:12.610 align:middle line:84% And in part, we credit their good estimates 00:10:12.610 --> 00:10:16.150 align:middle line:84% for the random coefficient parameters to imposing 00:10:16.150 --> 00:10:18.780 align:middle line:90% these supply side restrictions. 00:10:18.780 --> 00:10:19.690 align:middle line:90% Yes? 00:10:19.690 --> 00:10:21.690 align:middle line:84% AUDIENCE: Is this also informative 00:10:21.690 --> 00:10:25.020 align:middle line:84% to the extent to which you're making way too restrictive 00:10:25.020 --> 00:10:27.310 align:middle line:90% assumptions on your model? 00:10:27.310 --> 00:10:30.920 align:middle line:84% Like say you're trying to impose positive marginal costs, 00:10:30.920 --> 00:10:32.650 align:middle line:84% and then in your minimization of nu 00:10:32.650 --> 00:10:35.040 align:middle line:84% you're not getting the results you're expecting. 00:10:35.040 --> 00:10:39.290 align:middle line:84% Is that a sign that your results may be guided by your model 00:10:39.290 --> 00:10:40.780 align:middle line:90% extensions? 00:10:40.780 --> 00:10:43.430 align:middle line:84% PROFESSOR: So that's a great question. 00:10:43.430 --> 00:10:46.450 align:middle line:84% So, in fact, at the very end of the lecture 00:10:46.450 --> 00:10:49.180 align:middle line:84% today, we'll discuss a case where 00:10:49.180 --> 00:10:53.140 align:middle line:84% you can use supply side restrictions as overidentifying 00:10:53.140 --> 00:10:54.940 align:middle line:90% restrictions on the model. 00:10:54.940 --> 00:10:59.590 align:middle line:84% So instead of, let's say, only estimating the demand side 00:10:59.590 --> 00:11:02.170 align:middle line:84% and being just identified, you typically 00:11:02.170 --> 00:11:06.068 align:middle line:84% get a lot of additional moments out of the supply side 00:11:06.068 --> 00:11:08.360 align:middle line:84% because you've not introduced all that many parameters. 00:11:08.360 --> 00:11:11.110 align:middle line:84% But maybe you have the price of wheat, and the price of corn, 00:11:11.110 --> 00:11:14.860 align:middle line:84% and the price of advertising, and all those things vary. 00:11:14.860 --> 00:11:18.170 align:middle line:84% So that leads to overidentifying restrictions. 00:11:18.170 --> 00:11:20.950 align:middle line:84% And indeed, with those overidentifying restrictions, 00:11:20.950 --> 00:11:24.430 align:middle line:84% you can test different conduct models directly. 00:11:24.430 --> 00:11:27.695 align:middle line:90% Yeah, that's right. 00:11:27.695 --> 00:11:29.795 align:middle line:90% And this is somewhat-- 00:11:29.795 --> 00:11:31.170 align:middle line:84% I will comment on this later on-- 00:11:31.170 --> 00:11:34.670 align:middle line:84% it's somewhat different from the exercise that Nevo does. 00:11:34.670 --> 00:11:36.235 align:middle line:90% He directly compares to data. 00:11:36.235 --> 00:11:41.460 align:middle line:90% 00:11:41.460 --> 00:11:45.420 align:middle line:84% So my comments on using the supply side, as I just said, 00:11:45.420 --> 00:11:50.940 align:middle line:84% this oftentimes introduces many new moment conditions. 00:11:50.940 --> 00:11:57.060 align:middle line:84% It's received wisdom that it improves the precision of demand 00:11:57.060 --> 00:12:00.870 align:middle line:84% estimates, and especially the random coefficients, 00:12:00.870 --> 00:12:04.260 align:middle line:84% again, through the logic that I walked you through. 00:12:04.260 --> 00:12:07.410 align:middle line:84% It's econometrically efficient to estimate the demand 00:12:07.410 --> 00:12:10.170 align:middle line:84% and supply side jointly if you're 00:12:10.170 --> 00:12:12.810 align:middle line:84% fairly confident that your conduct assumption is right. 00:12:12.810 --> 00:12:16.410 align:middle line:84% Of course, you may not feel comfortable 00:12:16.410 --> 00:12:20.370 align:middle line:84% making the assumptions that firms maximize their own profits 00:12:20.370 --> 00:12:22.170 align:middle line:90% or collude directly. 00:12:22.170 --> 00:12:25.488 align:middle line:84% So that seems a less innocuous assumption than saying, 00:12:25.488 --> 00:12:27.905 align:middle line:84% well, consumers just pick the best product that they like. 00:12:27.905 --> 00:12:33.160 align:middle line:90% 00:12:33.160 --> 00:12:36.340 align:middle line:84% And again, you can use overidentifying restrictions 00:12:36.340 --> 00:12:40.090 align:middle line:84% to test models of conduct, and we'll 00:12:40.090 --> 00:12:44.080 align:middle line:84% have an example of this at the end of class today. 00:12:44.080 --> 00:12:50.020 align:middle line:84% So let me talk a little bit about the results of the BLP 00:12:50.020 --> 00:12:51.190 align:middle line:90% paper. 00:12:51.190 --> 00:12:54.790 align:middle line:84% I think they show nicely why you would want to favor 00:12:54.790 --> 00:12:56.530 align:middle line:90% random coefficients model. 00:12:56.530 --> 00:13:00.920 align:middle line:84% So they have data from 1971 to 1990. 00:13:00.920 --> 00:13:04.420 align:middle line:84% The market here is the entire US. 00:13:04.420 --> 00:13:09.120 align:middle line:84% They use list prices, so not transaction prices. 00:13:09.120 --> 00:13:11.200 align:middle line:90% You could bicker about this. 00:13:11.200 --> 00:13:14.770 align:middle line:84% And an observation here is a model year, 00:13:14.770 --> 00:13:18.245 align:middle line:90% and they have 2,217 of those. 00:13:18.245 --> 00:13:19.870 align:middle line:84% As characteristics, they use the number 00:13:19.870 --> 00:13:24.040 align:middle line:84% of cylinders, doors, horsepower, length, width, weight, 00:13:24.040 --> 00:13:28.210 align:middle line:84% wheelbase, and measures of fuel efficiency, 00:13:28.210 --> 00:13:29.920 align:middle line:84% dummies for air conditioning, and so on. 00:13:29.920 --> 00:13:32.450 align:middle line:90% 00:13:32.450 --> 00:13:34.880 align:middle line:84% So here are some interesting summary statistics 00:13:34.880 --> 00:13:36.460 align:middle line:90% on the market. 00:13:36.460 --> 00:13:38.500 align:middle line:84% So first, something that's striking 00:13:38.500 --> 00:13:44.560 align:middle line:84% is that the number of products just increased a lot over time. 00:13:44.560 --> 00:13:50.890 align:middle line:84% In fact, there's a recent paper by Grieco, Murry, and Yurukoglu 00:13:50.890 --> 00:13:54.190 align:middle line:84% who estimate mark-ups for the car 00:13:54.190 --> 00:13:56.110 align:middle line:90% market over many, many years. 00:13:56.110 --> 00:13:59.710 align:middle line:84% And the number of products, as you would not be surprised, 00:13:59.710 --> 00:14:01.210 align:middle line:84% goes up even more, and they actually 00:14:01.210 --> 00:14:04.060 align:middle line:84% find that consumer surplus in the car market 00:14:04.060 --> 00:14:08.070 align:middle line:84% has gone up over several decades. 00:14:08.070 --> 00:14:11.460 align:middle line:90% 00:14:11.460 --> 00:14:15.570 align:middle line:84% Prices here go from $8,000 to about $10,000, 00:14:15.570 --> 00:14:18.933 align:middle line:90% they're denoted in 1983 prices. 00:14:18.933 --> 00:14:20.850 align:middle line:84% Something that's interesting is that the share 00:14:20.850 --> 00:14:23.470 align:middle line:84% of domestic brands goes down over time. 00:14:23.470 --> 00:14:27.610 align:middle line:84% This is mostly driven by increasing imports from Japan. 00:14:27.610 --> 00:14:31.830 align:middle line:84% Another interesting observation here is that the size of cars 00:14:31.830 --> 00:14:34.860 align:middle line:84% is going down, very contrary to what we see today, 00:14:34.860 --> 00:14:37.300 align:middle line:84% where cars seem to get bigger and bigger. 00:14:37.300 --> 00:14:40.200 align:middle line:84% And the authors here speculate that this might actually 00:14:40.200 --> 00:14:49.480 align:middle line:84% be driven by the first oil price shock in 1970. 00:14:49.480 --> 00:14:51.270 align:middle line:90% So these are the summary stats. 00:14:51.270 --> 00:14:54.280 align:middle line:84% This is the miles per gallon and miles per dollar. 00:14:54.280 --> 00:14:56.930 align:middle line:84% They use miles per dollar on the demand side. 00:14:56.930 --> 00:15:00.030 align:middle line:84% Consumers will care about how fuel efficient their car is 00:15:00.030 --> 00:15:06.120 align:middle line:84% in terms of the dollar expenses, but producers might not directly 00:15:06.120 --> 00:15:07.410 align:middle line:90% care about that. 00:15:07.410 --> 00:15:10.380 align:middle line:84% They will only care about how much more costly 00:15:10.380 --> 00:15:12.090 align:middle line:84% is it to produce a fuel efficient car, 00:15:12.090 --> 00:15:16.565 align:middle line:84% so they don't care how much the price of gasoline varies. 00:15:16.565 --> 00:15:19.410 align:middle line:90% 00:15:19.410 --> 00:15:22.890 align:middle line:84% Here are some summary statistics where 00:15:22.890 --> 00:15:25.590 align:middle line:84% they denote different percentiles of their variables 00:15:25.590 --> 00:15:27.410 align:middle line:90% with actual models. 00:15:27.410 --> 00:15:34.710 align:middle line:84% So you see prices go from this famously crappy Yugo from $3,000 00:15:34.710 --> 00:15:39.480 align:middle line:84% to almost $70,000, so a factor of 20. 00:15:39.480 --> 00:15:44.530 align:middle line:84% Horsepower is also varying widely. 00:15:44.530 --> 00:15:47.220 align:middle line:84% Somewhat ironically, the Gran Fury 00:15:47.220 --> 00:15:49.800 align:middle line:84% comes in at the lower end of the spectrum in terms 00:15:49.800 --> 00:15:51.510 align:middle line:90% of horsepower per weight. 00:15:51.510 --> 00:15:54.540 align:middle line:84% And you see here different summary statistics-- a lot 00:15:54.540 --> 00:15:55.670 align:middle line:90% of information on cars. 00:15:55.670 --> 00:15:58.250 align:middle line:90% 00:15:58.250 --> 00:16:01.280 align:middle line:84% So here is the first table that I 00:16:01.280 --> 00:16:06.520 align:middle line:84% think is going to be useful to dissect a bit more. 00:16:06.520 --> 00:16:08.830 align:middle line:84% And in part, I want to make a point here 00:16:08.830 --> 00:16:12.580 align:middle line:84% that might be relevant for you practically for those of you who 00:16:12.580 --> 00:16:14.830 align:middle line:84% want to estimate these kinds of models. 00:16:14.830 --> 00:16:21.990 align:middle line:84% So you'll notice that it takes a while to code these models up, 00:16:21.990 --> 00:16:24.030 align:middle line:90% and debug them, and so on. 00:16:24.030 --> 00:16:27.780 align:middle line:84% So it is typically useful to do what these authors are doing 00:16:27.780 --> 00:16:34.380 align:middle line:84% here is before you estimate the big structural model to use 00:16:34.380 --> 00:16:37.043 align:middle line:84% this simple inversion that you get from the logit model 00:16:37.043 --> 00:16:38.210 align:middle line:90% without random coefficients. 00:16:38.210 --> 00:16:41.850 align:middle line:84% And remember, we can write this as a linear estimating equation 00:16:41.850 --> 00:16:44.850 align:middle line:84% and estimate it via OLS or two stage least squares-- 00:16:44.850 --> 00:16:48.545 align:middle line:84% things that you should be well familiar with at this point. 00:16:48.545 --> 00:16:53.730 align:middle line:84% And you can actually see how sensitive your coefficient 00:16:53.730 --> 00:16:55.560 align:middle line:84% estimates are, let's say, to the inclusion 00:16:55.560 --> 00:16:57.500 align:middle line:84% of different instruments and so on. 00:16:57.500 --> 00:16:59.520 align:middle line:84% So I think that's practically something 00:16:59.520 --> 00:17:04.650 align:middle line:84% you should keep in mind that before you code up the big model 00:17:04.650 --> 00:17:08.952 align:middle line:84% and try to probe different specifications 00:17:08.952 --> 00:17:10.410 align:middle line:84% and different kinds of instruments, 00:17:10.410 --> 00:17:14.190 align:middle line:84% you probably want to do this Stata, whatever 00:17:14.190 --> 00:17:17.980 align:middle line:84% your preferred way of doing this is, and just run 00:17:17.980 --> 00:17:19.130 align:middle line:90% simple regressions. 00:17:19.130 --> 00:17:20.869 align:middle line:84% So I found this to be very useful. 00:17:20.869 --> 00:17:24.295 align:middle line:84% And then usually, when you go to the random coefficients 00:17:24.295 --> 00:17:26.920 align:middle line:84% model, of course, you get very different substitution patterns. 00:17:26.920 --> 00:17:28.720 align:middle line:84% But you already have a pretty good idea 00:17:28.720 --> 00:17:30.810 align:middle line:84% what the average coefficient should look like. 00:17:30.810 --> 00:17:34.810 align:middle line:90% 00:17:34.810 --> 00:17:38.600 align:middle line:90% Any questions? 00:17:38.600 --> 00:17:40.610 align:middle line:84% So let me tell you a little bit about what's 00:17:40.610 --> 00:17:41.695 align:middle line:90% going on on this table. 00:17:41.695 --> 00:17:44.230 align:middle line:90% 00:17:44.230 --> 00:17:48.670 align:middle line:84% So here, they have an OLS specification of demand-- 00:17:48.670 --> 00:17:51.670 align:middle line:84% again, based on the simple logit model inversion-- and here, 00:17:51.670 --> 00:17:53.410 align:middle line:90% an IV specification. 00:17:53.410 --> 00:17:56.110 align:middle line:84% What you can see here is something 00:17:56.110 --> 00:18:01.840 align:middle line:84% that you would expect if firms systematically price unobserved 00:18:01.840 --> 00:18:03.670 align:middle line:90% product attributes. 00:18:03.670 --> 00:18:07.620 align:middle line:84% Higher unobserved product quality leads to higher prices. 00:18:07.620 --> 00:18:10.270 align:middle line:84% So you would expect the price coefficient 00:18:10.270 --> 00:18:12.260 align:middle line:90% to be biased towards 0. 00:18:12.260 --> 00:18:15.730 align:middle line:84% And indeed, once we start using instruments, 00:18:15.730 --> 00:18:17.720 align:middle line:84% the price coefficient becomes more negative. 00:18:17.720 --> 00:18:21.200 align:middle line:90% 00:18:21.200 --> 00:18:24.260 align:middle line:84% The other thing that you see, and this is actually 00:18:24.260 --> 00:18:28.917 align:middle line:84% of practical relevance again, that once you have a larger 00:18:28.917 --> 00:18:31.250 align:middle line:84% price coefficient, or a more negative price coefficient, 00:18:31.250 --> 00:18:34.633 align:middle line:84% you have a lot fewer elasticities 00:18:34.633 --> 00:18:36.050 align:middle line:84% that you get out of the model that 00:18:36.050 --> 00:18:38.360 align:middle line:84% are on the inelastic portion of demand, 00:18:38.360 --> 00:18:41.820 align:middle line:84% which would not be consistent with static profit maximization. 00:18:41.820 --> 00:18:46.130 align:middle line:90% 00:18:46.130 --> 00:18:48.930 align:middle line:84% In fact, oftentimes the problem can be even worse than that. 00:18:48.930 --> 00:18:51.380 align:middle line:84% And speaking from my own experience, 00:18:51.380 --> 00:18:55.520 align:middle line:84% you'll oftentimes find that elasticities are positive. 00:18:55.520 --> 00:18:57.560 align:middle line:90% And then you have a problem. 00:18:57.560 --> 00:18:59.387 align:middle line:90% You know you have a problem. 00:18:59.387 --> 00:19:01.970 align:middle line:84% In the other case, you might not know that you have a problem. 00:19:01.970 --> 00:19:03.887 align:middle line:84% But in that case, you know you have a problem. 00:19:03.887 --> 00:19:08.800 align:middle line:90% 00:19:08.800 --> 00:19:11.140 align:middle line:84% And so the last thing that they do here 00:19:11.140 --> 00:19:13.630 align:middle line:84% is they estimate a very simple supply side 00:19:13.630 --> 00:19:17.380 align:middle line:84% model of marginal cost pricing where they simply regress prices 00:19:17.380 --> 00:19:18.530 align:middle line:90% on a bunch of attributes. 00:19:18.530 --> 00:19:21.820 align:middle line:84% This is oftentimes called a hedonic regression. 00:19:21.820 --> 00:19:26.380 align:middle line:84% And to see these here can be interpreted 00:19:26.380 --> 00:19:32.480 align:middle line:84% as elasticities, how cost vary with different attributes. 00:19:32.480 --> 00:19:34.090 align:middle line:84% And so what is surprising here is 00:19:34.090 --> 00:19:37.870 align:middle line:84% that they can explain a large fraction of the variation 00:19:37.870 --> 00:19:40.760 align:middle line:84% with a simple model in terms of these attributes. 00:19:40.760 --> 00:19:43.900 align:middle line:90% 00:19:43.900 --> 00:19:49.640 align:middle line:84% So this here is their main table. 00:19:49.640 --> 00:19:51.630 align:middle line:84% So I just want to point out a few things. 00:19:51.630 --> 00:19:56.210 align:middle line:84% So they have here the coefficient estimates, 00:19:56.210 --> 00:19:58.940 align:middle line:84% here the standard deviations of those coefficients. 00:19:58.940 --> 00:20:00.920 align:middle line:84% So this tells us how heterogeneous taste 00:20:00.920 --> 00:20:03.140 align:middle line:90% is for different attributes. 00:20:03.140 --> 00:20:05.420 align:middle line:84% And they estimate two different models 00:20:05.420 --> 00:20:07.940 align:middle line:84% which change something on the supply side 00:20:07.940 --> 00:20:12.821 align:middle line:84% whether or not there are scale economies in production. 00:20:12.821 --> 00:20:17.690 align:middle line:84% And so a few things to point out here. 00:20:17.690 --> 00:20:20.130 align:middle line:84% They find that-- so first, they have 00:20:20.130 --> 00:20:24.150 align:middle line:84% significant estimates for all their standard deviations. 00:20:24.150 --> 00:20:27.880 align:middle line:84% So this is typically, again, not easy to achieve. 00:20:27.880 --> 00:20:32.310 align:middle line:84% And in part, this is because of the supply side restrictions 00:20:32.310 --> 00:20:36.180 align:middle line:84% that lead to more precise estimates 00:20:36.180 --> 00:20:37.710 align:middle line:90% even on the demand side. 00:20:37.710 --> 00:20:39.990 align:middle line:84% The other thing that's interesting here, 00:20:39.990 --> 00:20:43.120 align:middle line:84% if you look at this, so it's here, 00:20:43.120 --> 00:20:45.540 align:middle line:84% this attribute, miles per dollar, 00:20:45.540 --> 00:20:49.680 align:middle line:90% is valued negative on average. 00:20:49.680 --> 00:20:55.200 align:middle line:84% So the average consumer doesn't seem 00:20:55.200 --> 00:20:57.070 align:middle line:90% to like fuel efficient cars. 00:20:57.070 --> 00:21:01.860 align:middle line:84% However, notice that the standard deviation on this 00:21:01.860 --> 00:21:04.990 align:middle line:90% is also very large. 00:21:04.990 --> 00:21:09.430 align:middle line:84% And so there will certainly be some consumers that value 00:21:09.430 --> 00:21:11.620 align:middle line:90% miles per dollar positively. 00:21:11.620 --> 00:21:15.430 align:middle line:84% So what this model allows us to rationalize, 00:21:15.430 --> 00:21:17.200 align:middle line:84% for instance, is that if you, let's say, 00:21:17.200 --> 00:21:21.550 align:middle line:84% do this for the entire US, as they do, in much of the US, 00:21:21.550 --> 00:21:22.540 align:middle line:90% you see big cars. 00:21:22.540 --> 00:21:26.100 align:middle line:84% People seemingly don't care much about fuel efficiency. 00:21:26.100 --> 00:21:29.640 align:middle line:84% But then in, let's say, other parts of the country, 00:21:29.640 --> 00:21:33.730 align:middle line:84% you see a lot of Toyota Priuses, and Teslas, and so on. 00:21:33.730 --> 00:21:37.380 align:middle line:84% And a model that doesn't account for these heterogeneous 00:21:37.380 --> 00:21:40.350 align:middle line:84% substitution patterns with aggregate data 00:21:40.350 --> 00:21:42.680 align:middle line:84% wouldn't be able to get at this heterogeneity-- 00:21:42.680 --> 00:21:45.360 align:middle line:84% that different people have very different tastes 00:21:45.360 --> 00:21:48.270 align:middle line:84% for these attributes, and as a result of that, 00:21:48.270 --> 00:21:49.540 align:middle line:90% substitute differently. 00:21:49.540 --> 00:21:58.982 align:middle line:90% 00:21:58.982 --> 00:22:04.580 align:middle line:84% So a few more tables that showcase 00:22:04.580 --> 00:22:07.820 align:middle line:84% that they pick up pretty good substitution patterns. 00:22:07.820 --> 00:22:11.990 align:middle line:84% Here, you see a bunch of different models 00:22:11.990 --> 00:22:16.250 align:middle line:84% that have in their estimation, and you can roughly 00:22:16.250 --> 00:22:17.870 align:middle line:84% think of this table as being ordered 00:22:17.870 --> 00:22:20.990 align:middle line:90% by how fancy the cars are. 00:22:20.990 --> 00:22:23.300 align:middle line:84% So this is, of course, not a purely vertical model, 00:22:23.300 --> 00:22:26.060 align:middle line:84% as we discussed last week, but think 00:22:26.060 --> 00:22:29.720 align:middle line:84% of this as these are more expensive, more fancy cars. 00:22:29.720 --> 00:22:33.230 align:middle line:84% And if you pick, let's say, a high end 00:22:33.230 --> 00:22:36.890 align:middle line:84% car like the Lexus, BMW, or Cadillac, 00:22:36.890 --> 00:22:39.350 align:middle line:84% you'll see that if you change the price, 00:22:39.350 --> 00:22:40.710 align:middle line:90% these are semi-elasticities. 00:22:40.710 --> 00:22:43.340 align:middle line:84% That means that how much do market shares change when I 00:22:43.340 --> 00:22:47.690 align:middle line:90% increase my price by $1,000. 00:22:47.690 --> 00:22:51.680 align:middle line:84% You see that if you do this for the BMW, 00:22:51.680 --> 00:22:53.970 align:middle line:84% you increase the price by $1,000, 00:22:53.970 --> 00:22:58.770 align:middle line:84% there's almost no dent in the market share of lower end cars. 00:22:58.770 --> 00:23:02.495 align:middle line:84% And most of the action is here at the higher end models. 00:23:02.495 --> 00:23:06.114 align:middle line:90% 00:23:06.114 --> 00:23:10.260 align:middle line:84% And so similarly for Lexus, and Cadillac, and so on. 00:23:10.260 --> 00:23:13.050 align:middle line:84% The other thing that validates the model 00:23:13.050 --> 00:23:16.620 align:middle line:84% is that since we are looking at the same price 00:23:16.620 --> 00:23:18.750 align:middle line:84% increase for our low end and high end models, 00:23:18.750 --> 00:23:21.570 align:middle line:84% you'll find that the response for low end models 00:23:21.570 --> 00:23:23.950 align:middle line:84% is much, much stronger than for high models, 00:23:23.950 --> 00:23:24.840 align:middle line:90% as you would expect. 00:23:24.840 --> 00:23:29.220 align:middle line:84% If you're buying a $70,000 car, then an extra $1,000 00:23:29.220 --> 00:23:34.815 align:middle line:84% shouldn't be as painful to you as if you're buying the $3,300 00:23:34.815 --> 00:23:35.315 align:middle line:90% Yugo. 00:23:35.315 --> 00:23:44.620 align:middle line:90% 00:23:44.620 --> 00:23:48.030 align:middle line:84% So this here is actually a quite striking table. 00:23:48.030 --> 00:23:55.440 align:middle line:84% This here shows how people substitute to the outside option 00:23:55.440 --> 00:24:00.660 align:middle line:84% if we base these diversion ratios on the logit model 00:24:00.660 --> 00:24:05.125 align:middle line:84% and compare that to the BLP model. 00:24:05.125 --> 00:24:06.900 align:middle line:84% So I'll let you stare at this for a second 00:24:06.900 --> 00:24:13.650 align:middle line:84% and think a little bit about what this comparison tells us, 00:24:13.650 --> 00:24:18.050 align:middle line:84% and in what sense the BLP model might 00:24:18.050 --> 00:24:19.380 align:middle line:90% be doing a better job here. 00:24:19.380 --> 00:24:21.890 align:middle line:90% 00:24:21.890 --> 00:24:26.450 align:middle line:84% So I may just say something which is relevant here. 00:24:26.450 --> 00:24:32.450 align:middle line:84% Again, think of these as ordered in terms of how fancy they are. 00:24:32.450 --> 00:24:35.430 align:middle line:84% These are more expensive cars down here than up here. 00:24:35.430 --> 00:24:37.520 align:middle line:84% AUDIENCE: The BLP is more capturing the fact 00:24:37.520 --> 00:24:39.620 align:middle line:84% that if you're going for an expensive car, 00:24:39.620 --> 00:24:42.020 align:middle line:84% increasing price won't induce you to move away 00:24:42.020 --> 00:24:45.620 align:middle line:84% from buying a car in general, but logit doesn't really 00:24:45.620 --> 00:24:47.618 align:middle line:90% get after that kind of aspect. 00:24:47.618 --> 00:24:48.660 align:middle line:90% PROFESSOR: Exactly right. 00:24:48.660 --> 00:24:52.525 align:middle line:84% So we think that, again, cars are durable goods. 00:24:52.525 --> 00:24:56.240 align:middle line:90% 00:24:56.240 --> 00:24:59.490 align:middle line:84% If I already have a car and I'm considering buying a cheap car, 00:24:59.490 --> 00:25:01.670 align:middle line:84% I might be much closer to the margin 00:25:01.670 --> 00:25:04.700 align:middle line:84% where I drive my car for another year. 00:25:04.700 --> 00:25:08.360 align:middle line:84% Whereas somebody who is buying a $90,000 car, 00:25:08.360 --> 00:25:12.560 align:middle line:84% they'll probably not switch to the outside option if we raise 00:25:12.560 --> 00:25:14.160 align:middle line:84% the price of that car a little bit. 00:25:14.160 --> 00:25:17.420 align:middle line:84% They might instead opt for somewhat cheaper car. 00:25:17.420 --> 00:25:23.170 align:middle line:84% And you see that the logit model cannot rationalize this at all. 00:25:23.170 --> 00:25:28.330 align:middle line:90% So it's quite striking. 00:25:28.330 --> 00:25:30.460 align:middle line:84% And so in the BLP model, you find 00:25:30.460 --> 00:25:32.680 align:middle line:84% that lower end consumers are much more 00:25:32.680 --> 00:25:34.990 align:middle line:84% likely to move towards the outset option. 00:25:34.990 --> 00:25:40.540 align:middle line:90% 00:25:40.540 --> 00:25:45.790 align:middle line:84% So I think this is all I wanted to say about those results. 00:25:45.790 --> 00:25:49.030 align:middle line:84% Let me comment on the BLP paper, and also just give you 00:25:49.030 --> 00:25:52.330 align:middle line:84% some practical leads here on where you can find 00:25:52.330 --> 00:25:53.810 align:middle line:90% additional information on this. 00:25:53.810 --> 00:25:56.650 align:middle line:84% So, you know, as I'm sure you may have noticed, 00:25:56.650 --> 00:25:58.340 align:middle line:84% we spent almost two lectures on this. 00:25:58.340 --> 00:26:00.490 align:middle line:84% This has been a very influential paper, 00:26:00.490 --> 00:26:02.680 align:middle line:84% led to countless empirical studies, 00:26:02.680 --> 00:26:04.640 align:middle line:84% a large methodological literature. 00:26:04.640 --> 00:26:08.890 align:middle line:84% It's one of the bread and butter pieces 00:26:08.890 --> 00:26:11.870 align:middle line:84% of what you do in empirical IO estimating demand, 00:26:11.870 --> 00:26:16.150 align:middle line:84% and this is the standard model to do that. 00:26:16.150 --> 00:26:18.850 align:middle line:84% The empirical results, the papers, of course, 00:26:18.850 --> 00:26:21.140 align:middle line:84% mostly known for its methodological contribution, 00:26:21.140 --> 00:26:26.290 align:middle line:84% but the empirical results make a nice case for this model. 00:26:26.290 --> 00:26:29.260 align:middle line:84% There were many things at the time when this paper came out 00:26:29.260 --> 00:26:31.810 align:middle line:84% that people didn't really know about. 00:26:31.810 --> 00:26:34.810 align:middle line:84% So, for instance, nonparametric identification, 00:26:34.810 --> 00:26:36.730 align:middle line:84% to what extent are random coefficient models 00:26:36.730 --> 00:26:42.290 align:middle line:84% identified if we do not impose a type 1 extreme value assumption? 00:26:42.290 --> 00:26:44.450 align:middle line:84% So people have thought about that, 00:26:44.450 --> 00:26:47.420 align:middle line:84% and they are very general identification results 00:26:47.420 --> 00:26:54.260 align:middle line:90% for general models of this sort. 00:26:54.260 --> 00:26:57.740 align:middle line:84% People have learned a lot about how to estimate this 00:26:57.740 --> 00:26:59.630 align:middle line:90% those things efficiently. 00:26:59.630 --> 00:27:03.320 align:middle line:84% And Aviv Nevo, whose paper we discuss after this, 00:27:03.320 --> 00:27:07.610 align:middle line:84% he has contributed to telling people 00:27:07.610 --> 00:27:09.590 align:middle line:90% how to estimate these models. 00:27:09.590 --> 00:27:15.140 align:middle line:84% There are now many good sources to dive deeper into all of this. 00:27:15.140 --> 00:27:20.150 align:middle line:84% So the new IO handbook chapter just came out in 2021. 00:27:20.150 --> 00:27:23.060 align:middle line:84% If you want to learn about demand, chapter 1 and chapter 2 00:27:23.060 --> 00:27:26.750 align:middle line:90% are good sources. 00:27:26.750 --> 00:27:30.020 align:middle line:84% If you're really interested into identification, 00:27:30.020 --> 00:27:33.290 align:middle line:84% somebody wrote me a question about the type 1 extreme value 00:27:33.290 --> 00:27:35.845 align:middle line:84% error and how restrictive this is, 00:27:35.845 --> 00:27:37.650 align:middle line:90% this would be the place to look. 00:27:37.650 --> 00:27:40.310 align:middle line:84% These are papers that show you how these models are 00:27:40.310 --> 00:27:42.480 align:middle line:90% non-parametrically identified. 00:27:42.480 --> 00:27:45.380 align:middle line:84% And I would probably start with the annual reviews 00:27:45.380 --> 00:27:48.355 align:middle line:84% article, which gives an overview of this econometrica paper. 00:27:48.355 --> 00:27:50.980 align:middle line:90% 00:27:50.980 --> 00:27:58.580 align:middle line:84% There is also now out out-of-the-box estimation with 00:27:58.580 --> 00:28:01.430 align:middle line:90% Python that you can use. 00:28:01.430 --> 00:28:03.710 align:middle line:84% My view on this is this is super useful, 00:28:03.710 --> 00:28:07.200 align:middle line:84% and you may want to check your results with that. 00:28:07.200 --> 00:28:09.140 align:middle line:84% But I also believe that you should 00:28:09.140 --> 00:28:10.610 align:middle line:84% learn how to estimate those models 00:28:10.610 --> 00:28:13.970 align:middle line:84% and do it yourself because that's the real skill. 00:28:13.970 --> 00:28:17.540 align:middle line:84% And especially if you stay in IO, 00:28:17.540 --> 00:28:19.580 align:middle line:84% part of writing an IO job market paper 00:28:19.580 --> 00:28:23.180 align:middle line:84% is to somehow push what you can do with these models, 00:28:23.180 --> 00:28:25.730 align:middle line:84% and show that you understand them deeply as opposed 00:28:25.730 --> 00:28:26.960 align:middle line:90% to just apply them. 00:28:26.960 --> 00:28:29.040 align:middle line:90% So this is a great resource. 00:28:29.040 --> 00:28:32.840 align:middle line:84% But, I hope that once you come to [14.]273, 00:28:32.840 --> 00:28:35.080 align:middle line:84% we teach you how to actually do those things. 00:28:35.080 --> 00:28:39.000 align:middle line:90% 00:28:39.000 --> 00:28:41.280 align:middle line:84% So let me then also just say a few things 00:28:41.280 --> 00:28:43.985 align:middle line:90% empirically on this paper. 00:28:43.985 --> 00:28:45.580 align:middle line:90% We had already discussed this. 00:28:45.580 --> 00:28:49.200 align:middle line:84% The paper ignores the durable good aspect here-- 00:28:49.200 --> 00:28:52.398 align:middle line:84% that people hold cars, and that substitution 00:28:52.398 --> 00:28:54.690 align:middle line:84% to the offset option is, therefore, very heterogeneous, 00:28:54.690 --> 00:28:59.530 align:middle line:84% and potentially endogenous to the counterfactual that you run. 00:28:59.530 --> 00:29:01.950 align:middle line:84% Because if you change prices in a new car market, 00:29:01.950 --> 00:29:04.690 align:middle line:84% prices in the used car market might respond to that. 00:29:04.690 --> 00:29:07.350 align:middle line:84% And so your counterfactual might be misspecified 00:29:07.350 --> 00:29:08.340 align:middle line:90% as a result of that. 00:29:08.340 --> 00:29:11.640 align:middle line:84% But at least, because they have the random coefficient 00:29:11.640 --> 00:29:14.520 align:middle line:84% on the intercept, you get much more heterogeneous substitution 00:29:14.520 --> 00:29:16.860 align:middle line:84% patterns to the outside option than you 00:29:16.860 --> 00:29:18.790 align:middle line:90% get from a simple logit model. 00:29:18.790 --> 00:29:21.680 align:middle line:84% So it captures that to some extent. 00:29:21.680 --> 00:29:24.520 align:middle line:84% Again, there's some great papers to look at this further. 00:29:24.520 --> 00:29:27.500 align:middle line:84% There's a paper by Gowrisankaran and Rysman 00:29:27.500 --> 00:29:31.920 align:middle line:84% that looks at people purchasing durable goods. 00:29:31.920 --> 00:29:34.460 align:middle line:84% This is, I think, for camcorders. 00:29:34.460 --> 00:29:36.440 align:middle line:84% And so they endogenize, basically, 00:29:36.440 --> 00:29:40.050 align:middle line:84% how people update across different generations 00:29:40.050 --> 00:29:42.470 align:middle line:90% of camcorders. 00:29:42.470 --> 00:29:48.620 align:middle line:84% There is a paper on the used car market by Gavazza and Lizzeri 00:29:48.620 --> 00:29:53.290 align:middle line:84% which ignores the product differentiation aspect of cars. 00:29:53.290 --> 00:29:57.515 align:middle line:84% So cars here just basically described by their vintage. 00:29:57.515 --> 00:30:01.480 align:middle line:90% 00:30:01.480 --> 00:30:05.155 align:middle line:84% So thinks about the welfare effects of the used car market. 00:30:05.155 --> 00:30:07.780 align:middle line:84% And then there's a more recent paper that actually puts the two 00:30:07.780 --> 00:30:09.220 align:middle line:90% things together again-- 00:30:09.220 --> 00:30:14.500 align:middle line:84% so the differentiation part, and the fact that cars are aging, 00:30:14.500 --> 00:30:20.410 align:middle line:84% and that different consumers might hold different vintages. 00:30:20.410 --> 00:30:22.790 align:middle line:84% This is a paper that abstracts from dealerships. 00:30:22.790 --> 00:30:25.570 align:middle line:84% This is actually quite problematic 00:30:25.570 --> 00:30:28.660 align:middle line:84% if you think about the US because the US has 00:30:28.660 --> 00:30:34.300 align:middle line:84% state franchise laws that prohibit manufacturers 00:30:34.300 --> 00:30:36.670 align:middle line:90% to sell directly to consumers. 00:30:36.670 --> 00:30:38.530 align:middle line:84% And so you might have heard about this 00:30:38.530 --> 00:30:39.980 align:middle line:90% in the context of Tesla. 00:30:39.980 --> 00:30:43.480 align:middle line:84% They have been rallying against this 00:30:43.480 --> 00:30:45.940 align:middle line:84% because they want to sell directly to consumers, 00:30:45.940 --> 00:30:48.440 align:middle line:84% and do not have large dealership networks. 00:30:48.440 --> 00:30:52.990 align:middle line:84% So that might be important for local availability of brands 00:30:52.990 --> 00:30:55.060 align:middle line:84% and substitution patterns, and could also 00:30:55.060 --> 00:31:00.600 align:middle line:84% lead to other problems such as double marginalization. 00:31:00.600 --> 00:31:04.470 align:middle line:84% The model ignores that car prices are often negotiated. 00:31:04.470 --> 00:31:08.170 align:middle line:84% I think for what they're doing, that's not too problematic. 00:31:08.170 --> 00:31:11.070 align:middle line:84% But we know that in reality, the transaction price 00:31:11.070 --> 00:31:14.190 align:middle line:84% might be very different from the list price. 00:31:14.190 --> 00:31:18.030 align:middle line:84% And I think Glenn has taught you a paper where 00:31:18.030 --> 00:31:20.640 align:middle line:84% that has very different implications for different kinds 00:31:20.640 --> 00:31:21.310 align:middle line:90% of consumers. 00:31:21.310 --> 00:31:24.090 align:middle line:84% So to the extent that you care about these distributional 00:31:24.090 --> 00:31:26.080 align:middle line:90% effects, that might be an issue. 00:31:26.080 --> 00:31:28.560 align:middle line:84% And lastly, it ignores something that 00:31:28.560 --> 00:31:30.960 align:middle line:84% is very important, and increasingly so, 00:31:30.960 --> 00:31:33.380 align:middle line:84% which is the financial transaction that is involved 00:31:33.380 --> 00:31:34.380 align:middle line:90% when you purchase a car. 00:31:34.380 --> 00:31:38.040 align:middle line:84% Because you're typically getting a loan, and maybe insurance, 00:31:38.040 --> 00:31:40.734 align:middle line:90% and other kinds of add-ons. 00:31:40.734 --> 00:31:44.790 align:middle line:84% So obviously, this is an older paper. 00:31:44.790 --> 00:31:49.120 align:middle line:84% There are many papers that now address all of these issues, 00:31:49.120 --> 00:31:53.190 align:middle line:84% so that's a good paper to write if there are many gaps to fill 00:31:53.190 --> 00:31:56.060 align:middle line:84% and people are willing to do that. 00:31:56.060 --> 00:31:57.390 align:middle line:90% Any any questions? 00:31:57.390 --> 00:31:58.991 align:middle line:90% Otherwise, I move on to Nevo. 00:31:58.991 --> 00:32:05.165 align:middle line:90% 00:32:05.165 --> 00:32:05.665 align:middle line:90% OK. 00:32:05.665 --> 00:32:09.110 align:middle line:90% 00:32:09.110 --> 00:32:12.010 align:middle line:84% So the next two papers are in the spirit 00:32:12.010 --> 00:32:16.750 align:middle line:84% of what Glenn has already started talking about-- 00:32:16.750 --> 00:32:19.510 align:middle line:84% Bresnahan in 1987 and Miller and Weinberg 00:32:19.510 --> 00:32:23.150 align:middle line:84% testing different models of conduct against each other. 00:32:23.150 --> 00:32:26.560 align:middle line:84% So the idea is using our estimates 00:32:26.560 --> 00:32:29.570 align:middle line:84% to figure out what game is it that firms are actually playing? 00:32:29.570 --> 00:32:37.180 align:middle line:84% And I think that's a very nice way to use this machinery. 00:32:37.180 --> 00:32:41.150 align:middle line:84% We will look at the ready to eat cereal market. 00:32:41.150 --> 00:32:42.910 align:middle line:84% You have read Nevo's paper, and then we'll 00:32:42.910 --> 00:32:44.590 align:middle line:84% look at Backus, Conlon, and Sinkinson, 00:32:44.590 --> 00:32:47.000 align:middle line:84% which talks about the common ownership hypothesis. 00:32:47.000 --> 00:32:50.180 align:middle line:90% 00:32:50.180 --> 00:32:55.680 align:middle line:84% So the question in Nevo's paper is, 00:32:55.680 --> 00:33:01.050 align:middle line:84% why does the cereal industry sustain such high gross margins 00:33:01.050 --> 00:33:05.222 align:middle line:84% as measured from accounting data and the census of manufacturing? 00:33:05.222 --> 00:33:09.990 align:middle line:84% So they seem to be these very large margins. 00:33:09.990 --> 00:33:12.780 align:middle line:84% There are two competing explanations. 00:33:12.780 --> 00:33:15.630 align:middle line:84% One is that the firms are somehow-- 00:33:15.630 --> 00:33:19.110 align:middle line:84% of which there are few-- are colluding with each other. 00:33:19.110 --> 00:33:20.640 align:middle line:90% And the other one-- 00:33:20.640 --> 00:33:22.260 align:middle line:84% and I'm sure you've noticed this-- 00:33:22.260 --> 00:33:25.560 align:middle line:84% this is a market in which there are a lot of different products. 00:33:25.560 --> 00:33:28.710 align:middle line:84% And it could be that consumers have very heterogeneous tastes 00:33:28.710 --> 00:33:31.710 align:middle line:84% for those products, and somehow, firms 00:33:31.710 --> 00:33:34.770 align:middle line:84% are catering to these very idiosyncratic tastes, 00:33:34.770 --> 00:33:37.110 align:middle line:84% and therefore, are able to sustain 00:33:37.110 --> 00:33:40.920 align:middle line:84% high margins in conjunction with internalizing 00:33:40.920 --> 00:33:43.620 align:middle line:84% their multi-product incentives-- the fact that they do not 00:33:43.620 --> 00:33:48.000 align:middle line:84% want to cannibalize their own products. 00:33:48.000 --> 00:33:55.560 align:middle line:84% The approach is to use demand estimates to recover markups. 00:33:55.560 --> 00:33:58.630 align:middle line:84% A different alternative conduct assumptions, it should say here. 00:33:58.630 --> 00:34:04.260 align:middle line:84% So you're using only your demand estimates, 00:34:04.260 --> 00:34:06.060 align:middle line:84% imposing certain models of conduct, 00:34:06.060 --> 00:34:10.110 align:middle line:84% and then see whether those models of conduct with which you 00:34:10.110 --> 00:34:13.560 align:middle line:84% can recover marginal costs, and therefore compute markups, 00:34:13.560 --> 00:34:15.360 align:middle line:84% line up with the markups that you actually 00:34:15.360 --> 00:34:16.279 align:middle line:90% observe in the data. 00:34:16.279 --> 00:34:22.489 align:middle line:90% 00:34:22.489 --> 00:34:26.840 align:middle line:84% And he uses panel data on demand for cereals 00:34:26.840 --> 00:34:30.110 align:middle line:84% across different geographic markets. 00:34:30.110 --> 00:34:33.080 align:middle line:84% So I thought I'd use this as an opportunity 00:34:33.080 --> 00:34:36.139 align:middle line:84% to tell you about the surprisingly 00:34:36.139 --> 00:34:38.980 align:middle line:84% entertaining origins of the breakfast cereal market. 00:34:38.980 --> 00:34:42.380 align:middle line:90% 00:34:42.380 --> 00:34:46.350 align:middle line:84% So this here is James Caleb Jackson. 00:34:46.350 --> 00:34:50.010 align:middle line:90% He lived in western New York. 00:34:50.010 --> 00:34:53.610 align:middle line:84% He was a very religious vegetarian, 00:34:53.610 --> 00:34:58.230 align:middle line:84% and he stumbled into something that resembled maybe not too 00:34:58.230 --> 00:35:00.220 align:middle line:84% much the modern breakfast cereals. 00:35:00.220 --> 00:35:02.820 align:middle line:84% So it was kind of a hard thing to eat. 00:35:02.820 --> 00:35:03.940 align:middle line:90% Not much fun. 00:35:03.940 --> 00:35:07.890 align:middle line:84% You had to soak it overnight in milk to make it edible at all. 00:35:07.890 --> 00:35:08.900 align:middle line:90% He called it granula. 00:35:08.900 --> 00:35:14.250 align:middle line:90% 00:35:14.250 --> 00:35:17.760 align:middle line:84% Not much later, there's somebody else. 00:35:17.760 --> 00:35:21.120 align:middle line:84% And I should have, said James Caleb Jackson ran 00:35:21.120 --> 00:35:22.740 align:middle line:90% a sanatorium in New York. 00:35:22.740 --> 00:35:26.790 align:middle line:84% So that John Harvey Kellogg in Michigan, his family's 00:35:26.790 --> 00:35:29.250 align:middle line:90% originally from Massachusetts. 00:35:29.250 --> 00:35:33.210 align:middle line:84% He was a surgeon, and he invented something very similar 00:35:33.210 --> 00:35:36.120 align:middle line:84% that he first wanted to call granula. 00:35:36.120 --> 00:35:37.680 align:middle line:84% Then he got into some legal trouble 00:35:37.680 --> 00:35:38.960 align:middle line:90% and he said, call it granola. 00:35:38.960 --> 00:35:43.280 align:middle line:90% 00:35:43.280 --> 00:35:45.800 align:middle line:84% So you compare James Caleb Jackson, Kellogg, 00:35:45.800 --> 00:35:48.080 align:middle line:84% and you compare granula, granola. 00:35:48.080 --> 00:35:52.315 align:middle line:84% So you may already guess whose legacy was more enduring. 00:35:52.315 --> 00:35:55.072 align:middle line:90% 00:35:55.072 --> 00:35:58.980 align:middle line:84% So John Harvey Kellogg had a younger brother, 00:35:58.980 --> 00:36:02.550 align:middle line:84% and he was really onto something. 00:36:02.550 --> 00:36:06.140 align:middle line:84% Because for 10 years, they worked together 00:36:06.140 --> 00:36:09.860 align:middle line:84% and ran a company together until they 00:36:09.860 --> 00:36:14.930 align:middle line:84% had some differences about including sugar into cereals. 00:36:14.930 --> 00:36:17.090 align:middle line:84% So, again, you have to think of the old cereals 00:36:17.090 --> 00:36:20.178 align:middle line:84% as there was a religious impetus, 00:36:20.178 --> 00:36:22.470 align:middle line:84% and there was a whole philosophy around eating healthy. 00:36:22.470 --> 00:36:25.600 align:middle line:84% And then we'll suggest, no, let's put some sugar in it. 00:36:25.600 --> 00:36:26.850 align:middle line:90% We'll make it more successful. 00:36:26.850 --> 00:36:28.565 align:middle line:90% And it did. 00:36:28.565 --> 00:36:31.980 align:middle line:90% 00:36:31.980 --> 00:36:35.880 align:middle line:84% John Harvey Kellogg also had a patient, Charles William, 00:36:35.880 --> 00:36:38.040 align:middle line:90% or CW Post. 00:36:38.040 --> 00:36:40.020 align:middle line:84% So the brothers also had disagreements 00:36:40.020 --> 00:36:43.080 align:middle line:84% about the importance of trade secrets, 00:36:43.080 --> 00:36:45.990 align:middle line:84% and John Harvey Kellogg said, our patients should 00:36:45.990 --> 00:36:47.560 align:middle line:90% learn how to do those things. 00:36:47.560 --> 00:36:50.460 align:middle line:84% And that's what Charles William Post did. 00:36:50.460 --> 00:36:52.890 align:middle line:84% He basically copied the recipe, called it Grape Nuts, 00:36:52.890 --> 00:36:55.985 align:middle line:84% and became a competitor of the Kellogg company. 00:36:55.985 --> 00:36:59.270 align:middle line:90% 00:36:59.270 --> 00:37:03.540 align:middle line:84% So little did they know, so this is what happened to the market. 00:37:03.540 --> 00:37:07.160 align:middle line:84% Now you have hundreds of different products. 00:37:07.160 --> 00:37:09.980 align:middle line:84% It's also known that there's a lot of churn in the market, 00:37:09.980 --> 00:37:13.310 align:middle line:84% and that the market is sustained with very high advertising 00:37:13.310 --> 00:37:14.390 align:middle line:90% expenditures. 00:37:14.390 --> 00:37:18.318 align:middle line:84% And it's just funny to look at those, 00:37:18.318 --> 00:37:20.360 align:middle line:84% but it also illustrates a point that there's just 00:37:20.360 --> 00:37:23.300 align:middle line:90% this massive variety. 00:37:23.300 --> 00:37:26.390 align:middle line:84% You could be, let's say, interested in C-3PO cereals, 00:37:26.390 --> 00:37:30.080 align:middle line:84% or maybe Donkey Kong, or maybe Donkey Kong Junior 00:37:30.080 --> 00:37:34.960 align:middle line:84% if that's not catering precisely enough to your taste. 00:37:34.960 --> 00:37:38.620 align:middle line:90% 00:37:38.620 --> 00:37:43.180 align:middle line:84% And I should say in 1970, there was actually 00:37:43.180 --> 00:37:48.700 align:middle line:84% a complaint by the FTC that the companies in this market 00:37:48.700 --> 00:37:52.480 align:middle line:84% are colluding with each other and proliferating brands 00:37:52.480 --> 00:37:55.000 align:middle line:90% to deter entry from rivals. 00:37:55.000 --> 00:37:59.950 align:middle line:84% And so Schmalensee, who teaches at the Sloan school, 00:37:59.950 --> 00:38:07.160 align:middle line:84% has a paper on this in the Bell Journal of Economics. 00:38:07.160 --> 00:38:10.180 align:middle line:90% 00:38:10.180 --> 00:38:13.840 align:middle line:84% And so what is striking is you see the number of products 00:38:13.840 --> 00:38:17.258 align:middle line:84% in the market, and then you look at how many firms 00:38:17.258 --> 00:38:18.550 align:middle line:90% are controlling those products. 00:38:18.550 --> 00:38:20.180 align:middle line:90% And it's very, very few. 00:38:20.180 --> 00:38:25.150 align:middle line:84% So with the C3 concentration ratio, so the market 00:38:25.150 --> 00:38:30.970 align:middle line:84% share of the top three firms, you get up to more than 70% 00:38:30.970 --> 00:38:33.712 align:middle line:90% of the market. 00:38:33.712 --> 00:38:38.950 align:middle line:84% So what this tells you is this is a market where you want 00:38:38.950 --> 00:38:43.270 align:middle line:84% to take into account these multi-product effects-- 00:38:43.270 --> 00:38:46.120 align:middle line:84% the fact that firms will internalize 00:38:46.120 --> 00:38:49.450 align:middle line:84% how much they will rival their own products when they introduce 00:38:49.450 --> 00:38:52.830 align:middle line:90% new products or price those. 00:38:52.830 --> 00:38:55.350 align:middle line:84% It would not make much sense to not 00:38:55.350 --> 00:38:58.030 align:middle line:84% have such a model when you have four firms and hundreds 00:38:58.030 --> 00:38:59.240 align:middle line:90% of different products. 00:38:59.240 --> 00:39:07.550 align:middle line:90% 00:39:07.550 --> 00:39:11.390 align:middle line:84% So again, to remind you, the exercise that Nevo engages in 00:39:11.390 --> 00:39:17.910 align:middle line:90% is he tries to-- 00:39:17.910 --> 00:39:22.070 align:middle line:84% so he first measures markups directly 00:39:22.070 --> 00:39:27.714 align:middle line:84% in the data based off of accounting estimates. 00:39:27.714 --> 00:39:33.755 align:middle line:84% And the relevant estimate here is the retail margin of 46%. 00:39:33.755 --> 00:39:35.630 align:middle line:84% And then he does some bounding whether or not 00:39:35.630 --> 00:39:38.600 align:middle line:90% you include advertising as a-- 00:39:38.600 --> 00:39:43.040 align:middle line:84% whether advertising is a marginal or variable cost, and-- 00:39:43.040 --> 00:39:46.820 align:middle line:84% sorry, a marginal or a fixed cost. 00:39:46.820 --> 00:39:51.410 align:middle line:84% And in the end uses model estimated markups 00:39:51.410 --> 00:39:55.310 align:middle line:84% to compare it to those directly observed markups in the data. 00:39:55.310 --> 00:39:57.810 align:middle line:84% So this comes from this report here. 00:39:57.810 --> 00:40:00.570 align:middle line:84% But he also has some data from the census of manufacturing. 00:40:00.570 --> 00:40:09.416 align:middle line:90% 00:40:09.416 --> 00:40:15.170 align:middle line:84% So the specification that he uses 00:40:15.170 --> 00:40:19.730 align:middle line:84% is very, very similar to what we have discussed last week. 00:40:19.730 --> 00:40:24.360 align:middle line:84% So I just want to point out a few things here. 00:40:24.360 --> 00:40:28.950 align:middle line:84% So again, you have an indirect utility function 00:40:28.950 --> 00:40:32.170 align:middle line:84% that's linear in terms of product attributes-- 00:40:32.170 --> 00:40:34.590 align:middle line:84% these random coefficients for product 00:40:34.590 --> 00:40:38.700 align:middle line:84% attributes, prices for product J and firm T, 00:40:38.700 --> 00:40:41.580 align:middle line:84% and the random coefficient for prices. 00:40:41.580 --> 00:40:45.510 align:middle line:84% Now, the one difference here is that he decomposes 00:40:45.510 --> 00:40:48.930 align:middle line:84% the unobserved product attributes into a fixed effect 00:40:48.930 --> 00:40:50.345 align:middle line:90% and a deviation from that. 00:40:50.345 --> 00:40:53.860 align:middle line:90% 00:40:53.860 --> 00:40:56.490 align:middle line:84% So I got some questions about this fixed effect. 00:40:56.490 --> 00:41:00.740 align:middle line:84% So why haven't people done this before, 00:41:00.740 --> 00:41:07.210 align:middle line:90% and what is the difficulty here? 00:41:07.210 --> 00:41:13.610 align:middle line:84% So first of all, of course, this came not too much later 00:41:13.610 --> 00:41:15.090 align:middle line:90% after the introduction of BLP. 00:41:15.090 --> 00:41:17.060 align:middle line:84% So there were not that many people 00:41:17.060 --> 00:41:19.710 align:middle line:84% that actually learned how to estimate those models. 00:41:19.710 --> 00:41:24.360 align:middle line:84% That's kind of a perhaps somewhat disappointing answer. 00:41:24.360 --> 00:41:28.010 align:middle line:84% But I think there's one important point here-- 00:41:28.010 --> 00:41:29.760 align:middle line:90% or maybe two points, actually. 00:41:29.760 --> 00:41:34.370 align:middle line:84% The first one is that if you want to include these fixed 00:41:34.370 --> 00:41:37.280 align:middle line:84% effects, you're introducing a lot of new parameters 00:41:37.280 --> 00:41:38.310 align:middle line:90% into your model. 00:41:38.310 --> 00:41:42.950 align:middle line:84% And I walked you last week through the steps 00:41:42.950 --> 00:41:45.950 align:middle line:84% of model estimation, where you had the outer loop 00:41:45.950 --> 00:41:47.990 align:middle line:90% and the inner loop. 00:41:47.990 --> 00:41:52.160 align:middle line:84% Now, what is important here is that if you include 00:41:52.160 --> 00:41:55.010 align:middle line:84% fixed effects, you really want to make sure 00:41:55.010 --> 00:41:59.120 align:middle line:84% that in the inner loop, you're running a regression to back out 00:41:59.120 --> 00:42:00.770 align:middle line:90% those fixed effects. 00:42:00.770 --> 00:42:03.470 align:middle line:84% And you separate the linear parameters 00:42:03.470 --> 00:42:05.330 align:middle line:90% from the nonlinear parameters. 00:42:05.330 --> 00:42:07.670 align:middle line:84% To remind you, the linear parameters 00:42:07.670 --> 00:42:13.370 align:middle line:84% are all the things on observed or unobserved product 00:42:13.370 --> 00:42:15.290 align:middle line:84% attributes, in this case, with fixed effects 00:42:15.290 --> 00:42:18.460 align:middle line:84% that are only J specific and do not depend on the consumer. 00:42:18.460 --> 00:42:21.866 align:middle line:90% 00:42:21.866 --> 00:42:25.300 align:middle line:84% And so, again, fixed effects introduce a lot of parameters. 00:42:25.300 --> 00:42:28.360 align:middle line:84% If you can just optimize over your nonlinear parameters 00:42:28.360 --> 00:42:30.040 align:middle line:84% and then back out the fixed effects 00:42:30.040 --> 00:42:32.650 align:middle line:84% through a regression for any given guess of those nonlinear 00:42:32.650 --> 00:42:37.310 align:middle line:84% parameters, that's going to help you a lot. 00:42:37.310 --> 00:42:39.210 align:middle line:90% So that's one comment on this. 00:42:39.210 --> 00:42:47.410 align:middle line:84% The second comment is that it's a little bit against the spirit 00:42:47.410 --> 00:42:50.230 align:middle line:84% of these characteristic space models 00:42:50.230 --> 00:42:52.370 align:middle line:90% to estimate fixed effects. 00:42:52.370 --> 00:42:53.860 align:middle line:84% Not to say you shouldn't do this, 00:42:53.860 --> 00:42:55.280 align:middle line:90% or that this is not a good idea. 00:42:55.280 --> 00:42:58.010 align:middle line:84% This can be very useful for identification, 00:42:58.010 --> 00:43:00.430 align:middle line:84% but you are introducing new parameters 00:43:00.430 --> 00:43:04.140 align:middle line:84% for each new product, which is something that we try to avoid. 00:43:04.140 --> 00:43:10.530 align:middle line:90% 00:43:10.530 --> 00:43:13.950 align:middle line:84% So one thing that we had discussed last week, 00:43:13.950 --> 00:43:16.890 align:middle line:84% but that's new relative to BLP that I want to point out here, 00:43:16.890 --> 00:43:20.400 align:middle line:84% is that the random coefficients are now not just-- 00:43:20.400 --> 00:43:22.980 align:middle line:84% there's not just the standard deviation 00:43:22.980 --> 00:43:25.390 align:middle line:84% of some mysterious unobserved distribution. 00:43:25.390 --> 00:43:28.590 align:middle line:84% They're now written in terms of demographic variables 00:43:28.590 --> 00:43:30.120 align:middle line:84% of distributions that you actually 00:43:30.120 --> 00:43:32.200 align:middle line:90% observe for each market-- 00:43:32.200 --> 00:43:35.470 align:middle line:84% so the distribution of income, distribution of family 00:43:35.470 --> 00:43:37.230 align:middle line:90% size, and so on. 00:43:37.230 --> 00:43:39.580 align:middle line:84% So I'll come back to this once more when we 00:43:39.580 --> 00:43:44.660 align:middle line:90% discuss the estimates later on. 00:43:44.660 --> 00:43:48.630 align:middle line:84% The market here is defined as a city quarter combination. 00:43:48.630 --> 00:43:52.370 align:middle line:84% I had some questions about market definitions. 00:43:52.370 --> 00:43:55.920 align:middle line:84% Vat, you asked about this, right? 00:43:55.920 --> 00:43:58.620 align:middle line:90% That's a good question. 00:43:58.620 --> 00:44:04.180 align:middle line:84% So the question was, why do we define it this way? 00:44:04.180 --> 00:44:09.900 align:middle line:84% Why don't we take commuting zones, or, let's say, 00:44:09.900 --> 00:44:13.080 align:middle line:84% a quarter, or a half a year, a year? 00:44:13.080 --> 00:44:17.580 align:middle line:84% That's typically a difficult choice to make. 00:44:17.580 --> 00:44:22.740 align:middle line:84% What could happen is that if you don't draw the right boundaries, 00:44:22.740 --> 00:44:25.560 align:middle line:84% that firms at the edge of your market 00:44:25.560 --> 00:44:28.780 align:middle line:84% are actually competing with firms in other markets. 00:44:28.780 --> 00:44:31.710 align:middle line:84% And you mis-specify the number of competitors 00:44:31.710 --> 00:44:33.660 align:middle line:90% that they're responding to. 00:44:33.660 --> 00:44:39.020 align:middle line:84% And so there's some papers that use clustering procedures 00:44:39.020 --> 00:44:42.540 align:middle line:84% to figure out how to separate markets, 00:44:42.540 --> 00:44:45.710 align:middle line:84% but in general, this is a bit more 00:44:45.710 --> 00:44:48.770 align:middle line:84% an art than a science to think about how you how you designate 00:44:48.770 --> 00:44:50.540 align:middle line:90% markets. 00:44:50.540 --> 00:44:52.825 align:middle line:84% But yes, that's an important question. 00:44:52.825 --> 00:45:00.840 align:middle line:90% 00:45:00.840 --> 00:45:04.200 align:middle line:84% So one more issue that comes up with including those fixed 00:45:04.200 --> 00:45:09.330 align:middle line:90% effects is that brands are not-- 00:45:09.330 --> 00:45:13.170 align:middle line:84% or products are not changing their attributes much over time. 00:45:13.170 --> 00:45:15.600 align:middle line:84% Which means that those instruments are completely 00:45:15.600 --> 00:45:18.990 align:middle line:84% absorbing the variation that you would be using when 00:45:18.990 --> 00:45:22.950 align:middle line:90% you build the BLP instruments. 00:45:22.950 --> 00:45:25.410 align:middle line:84% So remember, the BLP instruments, it 00:45:25.410 --> 00:45:30.145 align:middle line:84% was this idea that if I'm offering a very sugary product, 00:45:30.145 --> 00:45:32.520 align:middle line:84% and I'm in a market where all the other products are also 00:45:32.520 --> 00:45:35.500 align:middle line:84% very sugary, then I face more competition. 00:45:35.500 --> 00:45:37.625 align:middle line:84% I'm probably going to respond to that in my prices. 00:45:37.625 --> 00:45:41.750 align:middle line:90% 00:45:41.750 --> 00:45:43.520 align:middle line:84% So what he does to get around this 00:45:43.520 --> 00:45:46.520 align:middle line:84% is to introduce two other instruments. 00:45:46.520 --> 00:45:48.110 align:middle line:84% First, the Hausman instruments, which 00:45:48.110 --> 00:45:50.990 align:middle line:84% we had discussed last week, which 00:45:50.990 --> 00:45:56.360 align:middle line:84% is basically the idea of using average prices 00:45:56.360 --> 00:46:02.530 align:middle line:84% in other markets for the same firm as proxies for costs. 00:46:02.530 --> 00:46:06.280 align:middle line:84% So I've got some questions about this, so I just want to ask, 00:46:06.280 --> 00:46:09.430 align:middle line:84% does somebody want to explain again 00:46:09.430 --> 00:46:12.040 align:middle line:84% what the idea behind those instruments exactly is, 00:46:12.040 --> 00:46:15.310 align:middle line:84% and what are some of the issues that could come up here? 00:46:15.310 --> 00:46:22.310 align:middle line:84% So again, the problem is that in many cases, 00:46:22.310 --> 00:46:23.750 align:middle line:84% typically, costs are not observed, 00:46:23.750 --> 00:46:27.020 align:middle line:90% or it's hard to get at those. 00:46:27.020 --> 00:46:29.360 align:middle line:84% Remember also that-- just picture 00:46:29.360 --> 00:46:32.390 align:middle line:84% a very simple linear supply and demand 00:46:32.390 --> 00:46:34.250 align:middle line:84% picture, and the identification problem 00:46:34.250 --> 00:46:36.080 align:middle line:84% that we had discussed at the very beginning 00:46:36.080 --> 00:46:39.230 align:middle line:84% of the last lecture, that you ideally 00:46:39.230 --> 00:46:42.710 align:middle line:84% want some cost shifters that shift around the supply 00:46:42.710 --> 00:46:45.830 align:middle line:84% function to identify your demand function. 00:46:45.830 --> 00:46:48.892 align:middle line:84% Again, cost side instruments that you 00:46:48.892 --> 00:46:50.600 align:middle line:84% believe are excluded from the demand side 00:46:50.600 --> 00:46:54.610 align:middle line:84% are typically the most convincing types of instruments. 00:46:54.610 --> 00:46:58.408 align:middle line:90% 00:46:58.408 --> 00:47:02.200 align:middle line:84% So instead of using directly observed costs, 00:47:02.200 --> 00:47:04.480 align:middle line:84% the idea, again, behind these Hausman instruments 00:47:04.480 --> 00:47:08.590 align:middle line:84% is to proxy for those costs with average prices of the same firm 00:47:08.590 --> 00:47:10.750 align:middle line:90% in other markets. 00:47:10.750 --> 00:47:15.460 align:middle line:84% And the idea is, if I'm, let's say, the only cereal 00:47:15.460 --> 00:47:18.440 align:middle line:84% producer that uses rice as an ingredient, 00:47:18.440 --> 00:47:20.260 align:middle line:84% and the price of rice goes up, then 00:47:20.260 --> 00:47:23.180 align:middle line:84% that's going to shift my average price. 00:47:23.180 --> 00:47:28.510 align:middle line:84% And to the extent that the shock to the cost of rice 00:47:28.510 --> 00:47:33.220 align:middle line:84% is uncorrelated to demand shocks across different geographic 00:47:33.220 --> 00:47:39.200 align:middle line:84% regions, I can use this as a proxy for a cost shifter. 00:47:39.200 --> 00:47:43.420 align:middle line:84% Now, this already leads to where the error where 00:47:43.420 --> 00:47:47.710 align:middle line:84% this might go wrong is if you have things 00:47:47.710 --> 00:47:51.640 align:middle line:84% in those prices that are actually correlated 00:47:51.640 --> 00:47:53.530 align:middle line:90% across different markets. 00:47:53.530 --> 00:47:56.630 align:middle line:84% So think about advertising campaigns. 00:47:56.630 --> 00:47:58.970 align:middle line:84% If I'm running a big advertising campaign 00:47:58.970 --> 00:48:02.480 align:middle line:84% that leads to more demand for my brands, 00:48:02.480 --> 00:48:05.360 align:middle line:84% and that's going to be potentially reflected 00:48:05.360 --> 00:48:06.950 align:middle line:84% in the prices, I mean, to the extent 00:48:06.950 --> 00:48:08.750 align:middle line:84% that we think that it is a marginal cost. 00:48:08.750 --> 00:48:14.120 align:middle line:84% And then it's as if I'm instrumenting my endogenous-- 00:48:14.120 --> 00:48:16.760 align:middle line:90% 00:48:16.760 --> 00:48:20.660 align:middle line:84% as if I'm instrumenting prices with demand. 00:48:20.660 --> 00:48:22.520 align:middle line:90% That's not a good idea. 00:48:22.520 --> 00:48:28.490 align:middle line:84% So the other instrument that he uses 00:48:28.490 --> 00:48:30.930 align:middle line:84% are what he calls cost side instruments. 00:48:30.930 --> 00:48:33.320 align:middle line:84% They have a somewhat similar logic. 00:48:33.320 --> 00:48:36.320 align:middle line:84% So he uses region dummies to pick up 00:48:36.320 --> 00:48:39.660 align:middle line:84% things like transportation costs, city density. 00:48:39.660 --> 00:48:43.040 align:middle line:84% Again, this is an issue to the extent 00:48:43.040 --> 00:48:46.940 align:middle line:84% that these things might also be picking up systematic variation 00:48:46.940 --> 00:48:50.095 align:middle line:90% in preferences across regions. 00:48:50.095 --> 00:48:53.260 align:middle line:90% 00:48:53.260 --> 00:48:55.550 align:middle line:90% Any other questions on this? 00:48:55.550 --> 00:48:58.290 align:middle line:84% You have read the paper, so you're curious. 00:48:58.290 --> 00:48:58.790 align:middle line:90% Yes? 00:48:58.790 --> 00:49:00.950 align:middle line:84% AUDIENCE: So one of the things that sort of-- 00:49:00.950 --> 00:49:05.660 align:middle line:84% that utility formulation is that the price sensitivity 00:49:05.660 --> 00:49:09.530 align:middle line:84% varies across consumers, but it doesn't vary by type of good. 00:49:09.530 --> 00:49:14.810 align:middle line:84% And so when they present the table with price elasticities 00:49:14.810 --> 00:49:15.590 align:middle line:90% between-- 00:49:15.590 --> 00:49:17.420 align:middle line:84% cross-price elasticity between types 00:49:17.420 --> 00:49:21.740 align:middle line:84% of goods, that's only driven by the type of consumer that's 00:49:21.740 --> 00:49:23.070 align:middle line:90% buying that good, right? 00:49:23.070 --> 00:49:24.890 align:middle line:90% PROFESSOR: OK. 00:49:24.890 --> 00:49:26.960 align:middle line:84% AUDIENCE: So I guess my question is, 00:49:26.960 --> 00:49:28.670 align:middle line:84% how reasonable of an assumption is it 00:49:28.670 --> 00:49:33.140 align:middle line:84% that all of the prices, cross elasticities that we estimate 00:49:33.140 --> 00:49:35.420 align:middle line:84% are only driven by differences in consumers, 00:49:35.420 --> 00:49:38.840 align:middle line:84% and not in types of goods that we're buying? 00:49:38.840 --> 00:49:41.160 align:middle line:84% PROFESSOR: OK that's an interesting question. 00:49:41.160 --> 00:49:44.070 align:middle line:84% So let me try and say a few things. 00:49:44.070 --> 00:49:45.830 align:middle line:84% So first of all, I mean, think first 00:49:45.830 --> 00:49:50.880 align:middle line:84% about the idea behind the characteristic space approach 00:49:50.880 --> 00:49:54.690 align:middle line:84% is that people value characteristics. 00:49:54.690 --> 00:49:56.970 align:middle line:84% And so if you really believe that that's 00:49:56.970 --> 00:49:58.710 align:middle line:84% the way you construct your models, 00:49:58.710 --> 00:50:03.510 align:middle line:84% then you shouldn't actually have random coefficients 00:50:03.510 --> 00:50:05.920 align:middle line:90% be product specific. 00:50:05.920 --> 00:50:09.140 align:middle line:90% So that's one answer to that. 00:50:09.140 --> 00:50:15.880 align:middle line:84% Now, it might be that they're a more flexible, functional form 00:50:15.880 --> 00:50:20.740 align:middle line:84% where, let's say, your arrow is not additively separable leads 00:50:20.740 --> 00:50:23.570 align:middle line:84% you to a specification like this where your random coefficient is 00:50:23.570 --> 00:50:28.060 align:middle line:84% essentially J specific because you're multiplying either alpha 00:50:28.060 --> 00:50:32.040 align:middle line:84% or beta with this unobserved term. 00:50:32.040 --> 00:50:34.976 align:middle line:90% 00:50:34.976 --> 00:50:39.832 align:middle line:84% So I think you would probably need-- 00:50:39.832 --> 00:50:41.790 align:middle line:84% I would have to think about, you would probably 00:50:41.790 --> 00:50:45.360 align:middle line:84% need some, again, variation over time 00:50:45.360 --> 00:50:48.690 align:middle line:84% to estimate such a model because you would supposedly 00:50:48.690 --> 00:50:53.970 align:middle line:84% want to see how does the same consumer vary attributes 00:50:53.970 --> 00:50:55.150 align:middle line:90% across different products. 00:50:55.150 --> 00:51:00.379 align:middle line:84% And you might need microdata for that, but I'm not exactly sure. 00:51:00.379 --> 00:51:01.212 align:middle line:90% AUDIENCE: Thank you. 00:51:01.212 --> 00:51:04.830 align:middle line:90% 00:51:04.830 --> 00:51:07.420 align:middle line:84% PROFESSOR: OK, any other questions? 00:51:07.420 --> 00:51:16.626 align:middle line:90% 00:51:16.626 --> 00:51:22.710 align:middle line:84% So let me briefly talk about his table 00:51:22.710 --> 00:51:28.118 align:middle line:84% where he estimates, again, as a logit model, as I explained 00:51:28.118 --> 00:51:29.410 align:middle line:90% before, and that's a good idea. 00:51:29.410 --> 00:51:31.800 align:middle line:84% And he has a lot of different specifications. 00:51:31.800 --> 00:51:36.040 align:middle line:84% The first three here are without any use of instruments. 00:51:36.040 --> 00:51:39.100 align:middle line:84% These are all instrumental variable specifications. 00:51:39.100 --> 00:51:42.190 align:middle line:84% And I think these two here are including 00:51:42.190 --> 00:51:44.500 align:middle line:84% fixed effects as controls-- so product fixed effects, 00:51:44.500 --> 00:51:46.540 align:middle line:90% whereas the first one does not. 00:51:46.540 --> 00:51:48.400 align:middle line:84% So what you can see is with fixed effects, 00:51:48.400 --> 00:51:50.920 align:middle line:84% you explain a lot more of the variation. 00:51:50.920 --> 00:51:54.490 align:middle line:84% And the price coefficients get more negative. 00:51:54.490 --> 00:51:58.270 align:middle line:84% And even more so when you include consumer demographics. 00:51:58.270 --> 00:52:02.800 align:middle line:84% So this here is an interesting instrumental 00:52:02.800 --> 00:52:06.370 align:middle line:84% variable specification that some of you 00:52:06.370 --> 00:52:09.140 align:middle line:84% pick this up what this is doing-- 00:52:09.140 --> 00:52:13.110 align:middle line:84% the inclusion of brand dummies as instruments. 00:52:13.110 --> 00:52:17.058 align:middle line:90% What is this close to in spirit? 00:52:17.058 --> 00:52:17.993 align:middle line:90% AUDIENCE: BLP. 00:52:17.993 --> 00:52:19.410 align:middle line:84% PROFESSOR: BLP instruments, right? 00:52:19.410 --> 00:52:22.170 align:middle line:90% 00:52:22.170 --> 00:52:27.420 align:middle line:84% Could you think of a story where this is preferable to, 00:52:27.420 --> 00:52:30.650 align:middle line:84% let's say, the simple BLP instruments? 00:52:30.650 --> 00:52:32.750 align:middle line:84% Or why would you want to do this? 00:52:32.750 --> 00:52:34.030 align:middle line:90% OK, so-- 00:52:34.030 --> 00:52:38.390 align:middle line:84% AUDIENCE: Is this what's captured on, like, advertising 00:52:38.390 --> 00:52:40.460 align:middle line:90% effects with a brand? 00:52:40.460 --> 00:52:43.340 align:middle line:90% PROFESSOR: Exactly. 00:52:43.340 --> 00:52:46.040 align:middle line:84% OK, so let me answer this in two ways. 00:52:46.040 --> 00:52:48.920 align:middle line:84% This is exactly right to the extent 00:52:48.920 --> 00:52:51.740 align:middle line:84% that advertising is not captured in the utility specification, 00:52:51.740 --> 00:52:53.910 align:middle line:90% which it isn't here. 00:52:53.910 --> 00:52:59.450 align:middle line:84% You may be worried that the BLP instruments are-- 00:52:59.450 --> 00:53:01.550 align:middle line:84% they're picking up that I'm responding to the fact 00:53:01.550 --> 00:53:06.080 align:middle line:84% that the average sugar that's in my competitor's brands 00:53:06.080 --> 00:53:08.900 align:middle line:84% is very high-- that I'm responding to that in some form, 00:53:08.900 --> 00:53:11.450 align:middle line:90% or the mushiness. 00:53:11.450 --> 00:53:16.430 align:middle line:84% But it might be that I offer the Donkey Kong cereals, 00:53:16.430 --> 00:53:20.940 align:middle line:84% and my competitor offers the Donkey Kong Junior cereals. 00:53:20.940 --> 00:53:24.870 align:middle line:84% And advertising, kids see the ads. 00:53:24.870 --> 00:53:26.870 align:middle line:84% And, of course, in all the markets 00:53:26.870 --> 00:53:29.510 align:middle line:84% where I'm competing with Donkey Kong Junior, 00:53:29.510 --> 00:53:31.350 align:middle line:90% I have to compete much harder. 00:53:31.350 --> 00:53:33.320 align:middle line:84% And those brand dummies would pick that up. 00:53:33.320 --> 00:53:39.240 align:middle line:84% So they capture not just the way that firms are responding 00:53:39.240 --> 00:53:41.340 align:middle line:90% to observed product attributes. 00:53:41.340 --> 00:53:44.070 align:middle line:84% They could potentially also capture other things 00:53:44.070 --> 00:53:46.450 align:middle line:84% that the researcher does not directly observe. 00:53:46.450 --> 00:53:49.870 align:middle line:90% So that can be useful. 00:53:49.870 --> 00:53:53.560 align:middle line:84% But of course, you cannot include those and also directly 00:53:53.560 --> 00:53:54.650 align:middle line:90% control for them. 00:53:54.650 --> 00:53:57.310 align:middle line:84% So that's why he uses here, these 00:53:57.310 --> 00:53:59.630 align:middle line:84% are indicating the Hausman instruments. 00:53:59.630 --> 00:54:02.170 align:middle line:84% So average prices in other markets, 00:54:02.170 --> 00:54:04.190 align:middle line:84% as well as these cost side instruments. 00:54:04.190 --> 00:54:06.820 align:middle line:84% And so if you look at these specifications here, 00:54:06.820 --> 00:54:08.790 align:middle line:84% you get even more negative price coefficient. 00:54:08.790 --> 00:54:11.140 align:middle line:84% I think starting somewhere here, he 00:54:11.140 --> 00:54:14.200 align:middle line:84% finds no longer that there are any elasticities 00:54:14.200 --> 00:54:17.471 align:middle line:84% on the inelastic portion of demand. 00:54:17.471 --> 00:54:20.840 align:middle line:84% So I think, again, this is something useful. 00:54:20.840 --> 00:54:25.240 align:middle line:84% It gives the reader a sense of what is driving your estimates, 00:54:25.240 --> 00:54:28.270 align:middle line:84% and it gives you a sense of what is driving your estimates. 00:54:28.270 --> 00:54:34.560 align:middle line:84% And that's producing these tables can be very useful. 00:54:34.560 --> 00:54:38.240 align:middle line:90% 00:54:38.240 --> 00:54:43.440 align:middle line:84% So let me say something briefly about domains table 00:54:43.440 --> 00:54:46.290 align:middle line:90% of the main specification. 00:54:46.290 --> 00:54:49.530 align:middle line:84% Now, these are the estimates from the BLP model, 00:54:49.530 --> 00:54:52.320 align:middle line:90% what he calls the full model. 00:54:52.320 --> 00:54:56.370 align:middle line:84% So here is a similar observation as before. 00:54:56.370 --> 00:54:59.340 align:middle line:84% Somebody was disappointed that people didn't like mushiness 00:54:59.340 --> 00:55:00.930 align:middle line:90% on average. 00:55:00.930 --> 00:55:02.850 align:middle line:84% I was also a bit disappointed by that 00:55:02.850 --> 00:55:04.650 align:middle line:84% because I also like mushy cereals, 00:55:04.650 --> 00:55:08.860 align:middle line:84% but it seems like the top 15% in the distribution of that. 00:55:08.860 --> 00:55:12.870 align:middle line:84% So that's actually somewhat interesting story here 00:55:12.870 --> 00:55:15.460 align:middle line:84% that he had to actually measure that. 00:55:15.460 --> 00:55:18.600 align:middle line:84% And it was one of his classmates, Sandy Black, 00:55:18.600 --> 00:55:24.600 align:middle line:84% who suggested including this as a characteristic. 00:55:24.600 --> 00:55:29.340 align:middle line:84% But again, the point here-- the point that is important-- 00:55:29.340 --> 00:55:34.050 align:middle line:84% is that you see here the demographic effects on this 00:55:34.050 --> 00:55:35.020 align:middle line:90% are large. 00:55:35.020 --> 00:55:38.220 align:middle line:84% And depending in which demographic you are, 00:55:38.220 --> 00:55:39.870 align:middle line:90% you actually value-- 00:55:39.870 --> 00:55:42.510 align:middle line:84% the sign on this attribute flips. 00:55:42.510 --> 00:55:50.840 align:middle line:84% And again, shows that people can value different attributes 00:55:50.840 --> 00:55:51.710 align:middle line:90% differently. 00:55:51.710 --> 00:55:55.460 align:middle line:84% The one substantive point that I want to make here 00:55:55.460 --> 00:56:00.570 align:middle line:84% is that if you look at what drives the random coefficients, 00:56:00.570 --> 00:56:04.380 align:middle line:84% you see that many of the demographic interaction effects 00:56:04.380 --> 00:56:07.690 align:middle line:90% are significant. 00:56:07.690 --> 00:56:10.730 align:middle line:84% But many of the standard deviations which capture this 00:56:10.730 --> 00:56:15.590 align:middle line:84% unobserved term in the random coefficient-- which, 00:56:15.590 --> 00:56:18.440 align:middle line:84% in the BLP random coefficients, that was the only source that 00:56:18.440 --> 00:56:19.770 align:middle line:90% led to variation-- 00:56:19.770 --> 00:56:22.740 align:middle line:90% are insignificant. 00:56:22.740 --> 00:56:27.960 align:middle line:84% And so this suggests that much of this variation, 00:56:27.960 --> 00:56:31.080 align:middle line:84% which was completely unobserved variation in BLP, 00:56:31.080 --> 00:56:35.280 align:middle line:84% is, in fact, in this market, driven by demographics. 00:56:35.280 --> 00:56:37.170 align:middle line:84% Now, I'll let you be the judge whether you 00:56:37.170 --> 00:56:40.260 align:middle line:84% find these interaction effects convincing. 00:56:40.260 --> 00:56:45.480 align:middle line:84% It seems like higher income people like mushy cereals more. 00:56:45.480 --> 00:56:47.770 align:middle line:84% I don't have very strong priors on that. 00:56:47.770 --> 00:56:52.840 align:middle line:90% 00:56:52.840 --> 00:56:59.530 align:middle line:84% So what does he get out of this at the end of the day? 00:56:59.530 --> 00:57:05.500 align:middle line:84% So let's focus here on his preferred BLP specification. 00:57:05.500 --> 00:57:08.920 align:middle line:84% So to remind you, the exercise that he engages in 00:57:08.920 --> 00:57:14.830 align:middle line:84% is I see some mark ups in accounting data from the census 00:57:14.830 --> 00:57:16.245 align:middle line:90% of manufacturing. 00:57:16.245 --> 00:57:19.180 align:middle line:84% And for a second, let's say these are the true markups 00:57:19.180 --> 00:57:21.250 align:middle line:84% and they do not have the kinds of issues 00:57:21.250 --> 00:57:25.570 align:middle line:84% that we talked about last week that they 00:57:25.570 --> 00:57:29.500 align:middle line:84% don't account for things like opportunity cost and so on. 00:57:29.500 --> 00:57:32.740 align:middle line:84% So these are not necessarily the correct economic costs 00:57:32.740 --> 00:57:35.560 align:middle line:84% that are measured in this accounting data. 00:57:35.560 --> 00:57:38.450 align:middle line:84% But let's say for a moment these are the true markups. 00:57:38.450 --> 00:57:42.850 align:middle line:84% So the idea is I estimate my demand model, 00:57:42.850 --> 00:57:44.290 align:middle line:84% use different conduct assumptions, 00:57:44.290 --> 00:57:47.830 align:middle line:84% and then see which conduct assumption replicates 00:57:47.830 --> 00:57:49.720 align:middle line:90% the true markups in the data. 00:57:49.720 --> 00:57:53.670 align:middle line:84% That's the essence of this exercise. 00:57:53.670 --> 00:57:57.820 align:middle line:84% And so what he finds, so he has this range of markup estimates 00:57:57.820 --> 00:58:00.220 align:middle line:84% depending on whether or not you include advertising 00:58:00.220 --> 00:58:05.090 align:middle line:84% in marginal cost, which ranges from 31% to 46%, 00:58:05.090 --> 00:58:10.930 align:middle line:84% he finds that you can soundly reject this collusive model 00:58:10.930 --> 00:58:15.160 align:middle line:84% where, for the 25 brands that he includes in his estimation, 00:58:15.160 --> 00:58:16.920 align:middle line:90% firms jointly maximize profits. 00:58:16.920 --> 00:58:19.988 align:middle line:90% 00:58:19.988 --> 00:58:23.810 align:middle line:84% And he cannot reject any of the alternative models, 00:58:23.810 --> 00:58:25.790 align:middle line:90% which is either firms-- 00:58:25.790 --> 00:58:29.870 align:middle line:84% or that each brand competes separately. 00:58:29.870 --> 00:58:32.480 align:middle line:84% Prices are set at the brand level 00:58:32.480 --> 00:58:39.330 align:middle line:84% without internalizing these cannibalization effects as well 00:58:39.330 --> 00:58:41.830 align:middle line:84% as the current ownership, where he uses the actual ownership 00:58:41.830 --> 00:58:42.830 align:middle line:90% matrix that he observes. 00:58:42.830 --> 00:58:46.710 align:middle line:90% 00:58:46.710 --> 00:58:50.160 align:middle line:84% One interesting thing here is that with the logit model, 00:58:50.160 --> 00:58:52.770 align:middle line:84% you get, in general, much, much lower markups. 00:58:52.770 --> 00:58:55.780 align:middle line:84% And you would reach a very different conclusion. 00:58:55.780 --> 00:58:58.990 align:middle line:84% You wouldn't reject the collusive model. 00:58:58.990 --> 00:59:02.370 align:middle line:84% So this, again, shows you that for this substantive question, 00:59:02.370 --> 00:59:05.040 align:middle line:84% which is, are firms in this market colluding? 00:59:05.040 --> 00:59:09.600 align:middle line:84% It seems to be important to account for these more realistic 00:59:09.600 --> 00:59:11.790 align:middle line:84% substitution patterns where different people value 00:59:11.790 --> 00:59:13.240 align:middle line:90% product attributes differently. 00:59:13.240 --> 00:59:21.220 align:middle line:90% 00:59:21.220 --> 00:59:25.170 align:middle line:84% So I think this is a well written and explained paper. 00:59:25.170 --> 00:59:28.980 align:middle line:84% In part, I asked you to read it because it provides 00:59:28.980 --> 00:59:32.490 align:middle line:84% a good explanation of the BLP model 00:59:32.490 --> 00:59:34.170 align:middle line:84% without many of the technical aspects 00:59:34.170 --> 00:59:37.800 align:middle line:84% that we will discuss down the line, 00:59:37.800 --> 00:59:45.530 align:middle line:84% and you know, in part, Aviv Nevo is also known in the profession 00:59:45.530 --> 00:59:47.392 align:middle line:84% because he has written this practitioner's 00:59:47.392 --> 00:59:49.100 align:middle line:84% guide at the very beginning that actually 00:59:49.100 --> 00:59:51.900 align:middle line:84% taught graduate students how to estimate these models. 00:59:51.900 --> 00:59:57.170 align:middle line:84% And he put his code out there, so it was very useful for people 00:59:57.170 --> 00:59:58.780 align:middle line:90% to understand what is going on. 00:59:58.780 --> 01:00:01.430 align:middle line:90% 01:00:01.430 --> 01:00:05.420 align:middle line:84% This is, in terms of thinking about picking an industry 01:00:05.420 --> 01:00:07.850 align:middle line:84% and the right model for this industry, what I really 01:00:07.850 --> 01:00:12.050 align:middle line:84% like about the paper is it would be a bit crazy to estimate 01:00:12.050 --> 01:00:14.030 align:middle line:84% a model for this industry without accounting 01:00:14.030 --> 01:00:16.220 align:middle line:84% for these multi-product considerations. 01:00:16.220 --> 01:00:21.830 align:middle line:84% At least as a prior, you'd think that it would matter a lot. 01:00:21.830 --> 01:00:23.870 align:middle line:84% You have four large firms that account 01:00:23.870 --> 01:00:25.820 align:middle line:84% for almost the entire market, and, there are 01:00:25.820 --> 01:00:27.920 align:middle line:90% hundreds of different products. 01:00:27.920 --> 01:00:33.260 align:middle line:84% So you have to do something about that. 01:00:33.260 --> 01:00:37.670 align:middle line:84% Collusion is, obviously, one of the important topics in IO. 01:00:37.670 --> 01:00:41.240 align:middle line:84% So it might not speak to your preferred industry, 01:00:41.240 --> 01:00:44.570 align:middle line:84% but it speaks to a substantive economic question. 01:00:44.570 --> 01:00:47.030 align:middle line:84% You could bicker that advertising is not 01:00:47.030 --> 01:00:48.030 align:middle line:90% a strategic variable. 01:00:48.030 --> 01:00:50.390 align:middle line:84% We know, actually, from the paper that advertising 01:00:50.390 --> 01:00:52.760 align:middle line:84% is very important in this market, 01:00:52.760 --> 01:00:55.050 align:middle line:84% and I imagine it's particularly important 01:00:55.050 --> 01:00:56.300 align:middle line:90% when you introduce new brands. 01:00:56.300 --> 01:00:58.550 align:middle line:84% You want to make them known to kids 01:00:58.550 --> 01:01:01.640 align:middle line:90% that there's a new Donkey Kong-- 01:01:01.640 --> 01:01:03.170 align:middle line:90% the new Donkey Kong cereal. 01:01:03.170 --> 01:01:10.200 align:middle line:84% So I guess another potential criticism is would firms 01:01:10.200 --> 01:01:13.740 align:middle line:84% actually be able to implement this, from their point, 01:01:13.740 --> 01:01:16.170 align:middle line:84% first best solution of full collusion 01:01:16.170 --> 01:01:18.360 align:middle line:84% across all the different products? 01:01:18.360 --> 01:01:21.368 align:middle line:84% And so is that the right comparison point? 01:01:21.368 --> 01:01:23.410 align:middle line:84% And then, as I already said, the testing approach 01:01:23.410 --> 01:01:26.000 align:middle line:84% is predicated on observing the right accounting cost. 01:01:26.000 --> 01:01:28.810 align:middle line:90% And that's not something that-- 01:01:28.810 --> 01:01:31.360 align:middle line:84% you know, once you use that as a benchmark, 01:01:31.360 --> 01:01:34.700 align:middle line:84% you can obviously no longer say anything about that. 01:01:34.700 --> 01:01:38.180 align:middle line:90% 01:01:38.180 --> 01:01:40.770 align:middle line:90% Any other comments on the paper? 01:01:40.770 --> 01:01:41.280 align:middle line:90% Yeah? 01:01:41.280 --> 01:01:42.905 align:middle line:84% AUDIENCE: So in some part of the paper, 01:01:42.905 --> 01:01:46.950 align:middle line:84% he argues that one shortcoming of the BLP literature 01:01:46.950 --> 01:01:51.750 align:middle line:84% is that you take observed product characteristics as given 01:01:51.750 --> 01:01:54.570 align:middle line:84% or as exogenous, and then you should internalize those 01:01:54.570 --> 01:01:58.590 align:middle line:84% because firms actually choose which characteristics 01:01:58.590 --> 01:02:00.335 align:middle line:90% will [INAUDIBLE] target. 01:02:00.335 --> 01:02:02.235 align:middle line:90% [INAUDIBLE] 01:02:02.235 --> 01:02:03.660 align:middle line:90% 01:02:03.660 --> 01:02:07.680 align:middle line:84% So in his model, is he implicitly 01:02:07.680 --> 01:02:12.270 align:middle line:84% allowing for firms to choose certain product characteristics? 01:02:12.270 --> 01:02:14.640 align:middle line:84% Because I don't think he's improving it. 01:02:14.640 --> 01:02:17.850 align:middle line:84% PROFESSOR: No, he's not doing that. 01:02:17.850 --> 01:02:19.530 align:middle line:84% The only endogenous characteristic 01:02:19.530 --> 01:02:21.040 align:middle line:90% is price in this case. 01:02:21.040 --> 01:02:23.520 align:middle line:84% And again, advertising would be a very good candidate 01:02:23.520 --> 01:02:26.580 align:middle line:84% for that because that's something that we potentially 01:02:26.580 --> 01:02:29.220 align:middle line:84% think varies from, let's say, quarter to quarter, 01:02:29.220 --> 01:02:30.963 align:middle line:90% city to city. 01:02:30.963 --> 01:02:34.680 align:middle line:84% And there are papers that look at advertising effects, 01:02:34.680 --> 01:02:39.100 align:middle line:84% and do use advertising as an endogenous variable. 01:02:39.100 --> 01:02:40.750 align:middle line:84% Then, of course, you have to think 01:02:40.750 --> 01:02:42.610 align:middle line:90% about additional instruments. 01:02:42.610 --> 01:02:47.460 align:middle line:90% 01:02:47.460 --> 01:02:51.310 align:middle line:84% But yeah, that's not something that he tackles in this paper. 01:02:51.310 --> 01:02:56.960 align:middle line:90% 01:02:56.960 --> 01:03:03.690 align:middle line:84% So let me use the last 20 minutes to talk about something 01:03:03.690 --> 01:03:04.190 align:middle line:90% that-- 01:03:04.190 --> 01:03:09.080 align:middle line:84% we're staying in this industry, but moving to a new topic. 01:03:09.080 --> 01:03:12.410 align:middle line:84% Which it's a fairly recent debate that 01:03:12.410 --> 01:03:16.200 align:middle line:84% came up in IO, which is called the common ownership hypothesis. 01:03:16.200 --> 01:03:20.840 align:middle line:84% It's something that's fairly controversial and hotly debated 01:03:20.840 --> 01:03:27.320 align:middle line:84% with strong proponents on either side of the argument. 01:03:27.320 --> 01:03:29.540 align:middle line:90% So what's the idea? 01:03:29.540 --> 01:03:32.630 align:middle line:84% The idea is that we have seen this big trend 01:03:32.630 --> 01:03:35.900 align:middle line:84% towards diversified investments where people 01:03:35.900 --> 01:03:39.800 align:middle line:84% hold large portfolios of the entire, let's say, S&P 01:03:39.800 --> 01:03:43.920 align:middle line:90% 500, or stock indices. 01:03:43.920 --> 01:03:46.380 align:middle line:84% And typically, they purchase those 01:03:46.380 --> 01:03:48.870 align:middle line:84% through large institutional investors like Vanguard, 01:03:48.870 --> 01:03:51.690 align:middle line:90% BlackRock, and Fidelity. 01:03:51.690 --> 01:03:55.770 align:middle line:84% And so a lot of people use those to save 01:03:55.770 --> 01:03:59.300 align:middle line:90% for retirement and so on. 01:03:59.300 --> 01:04:04.885 align:middle line:84% And the idea is that these owners, they might hold-- 01:04:04.885 --> 01:04:06.260 align:middle line:84% or these institutional investors, 01:04:06.260 --> 01:04:08.240 align:middle line:84% they might hold significant shares 01:04:08.240 --> 01:04:10.160 align:middle line:84% in a number of different companies 01:04:10.160 --> 01:04:13.170 align:middle line:84% that are actually competing with each other. 01:04:13.170 --> 01:04:16.650 align:middle line:84% So here is an example of the US airline market. 01:04:16.650 --> 01:04:20.340 align:middle line:84% You have American Airlines, Delta, Southwest and United, 01:04:20.340 --> 01:04:22.260 align:middle line:90% the major airlines. 01:04:22.260 --> 01:04:24.450 align:middle line:84% And you look, who are the biggest shareholders 01:04:24.450 --> 01:04:27.610 align:middle line:90% of these companies? 01:04:27.610 --> 01:04:30.450 align:middle line:84% Vanguard holds 6% of American Airlines. 01:04:30.450 --> 01:04:32.370 align:middle line:90% It holds 6% of Delta Airlines. 01:04:32.370 --> 01:04:38.070 align:middle line:84% It holds 6% of Southwest, and 7% of United Airlines. 01:04:38.070 --> 01:04:40.560 align:middle line:84% And so the idea here, this hypothesis, 01:04:40.560 --> 01:04:45.900 align:middle line:84% is that if I'm holding stocks in a bunch of companies 01:04:45.900 --> 01:04:49.500 align:middle line:84% that should be competing with each other, who says that. 01:04:49.500 --> 01:04:51.440 align:middle line:84% I wouldn't be very interested in making 01:04:51.440 --> 01:04:55.520 align:middle line:84% these companies compete a little less hard and raise my profits? 01:04:55.520 --> 01:05:02.580 align:middle line:84% Moving the incentives closer to monopoly incentives. 01:05:02.580 --> 01:05:05.820 align:middle line:84% So this is something that people had 01:05:05.820 --> 01:05:08.640 align:middle line:84% thought about in literature at the intersection 01:05:08.640 --> 01:05:10.170 align:middle line:90% of corporate finance and IO. 01:05:10.170 --> 01:05:12.180 align:middle line:84% There are some theory papers that I'll 01:05:12.180 --> 01:05:15.930 align:middle line:84% talk briefly about just to clarify what this hypothesis is 01:05:15.930 --> 01:05:18.990 align:middle line:84% about, but it wasn't really until very recently 01:05:18.990 --> 01:05:21.850 align:middle line:84% that this was-- that people paid a lot of attention to this. 01:05:21.850 --> 01:05:24.840 align:middle line:84% And of course, in part, this attention is driven by the fact 01:05:24.840 --> 01:05:28.380 align:middle line:84% that we have, now, these large diversified investments that 01:05:28.380 --> 01:05:30.220 align:middle line:90% have been trending up. 01:05:30.220 --> 01:05:34.870 align:middle line:84% So there was a series of papers in 2016, 2018, 01:05:34.870 --> 01:05:37.050 align:middle line:84% now there are many more, that document 01:05:37.050 --> 01:05:42.030 align:middle line:84% that the growth of large, diversified common owners 01:05:42.030 --> 01:05:44.760 align:middle line:84% led to increase in prices in various industries, 01:05:44.760 --> 01:05:48.510 align:middle line:84% for instance, for banking services or for airlines. 01:05:48.510 --> 01:05:52.680 align:middle line:84% So these were actually people that-- 01:05:52.680 --> 01:05:54.840 align:middle line:84% this was first researched in finance. 01:05:54.840 --> 01:05:58.500 align:middle line:84% And so it was a little bit awkward for IO people. 01:05:58.500 --> 01:06:00.150 align:middle line:84% Could we have possibly missed this, 01:06:00.150 --> 01:06:04.990 align:middle line:84% that this big structural force that led the whole economy 01:06:04.990 --> 01:06:08.800 align:middle line:90% to become less competitive? 01:06:08.800 --> 01:06:12.157 align:middle line:84% And do we really think that the price that we 01:06:12.157 --> 01:06:13.990 align:middle line:84% pay for these cheap, diversified investments 01:06:13.990 --> 01:06:16.360 align:middle line:84% is that monopoly power as a whole 01:06:16.360 --> 01:06:20.560 align:middle line:90% has increased in the US economy? 01:06:20.560 --> 01:06:23.420 align:middle line:84% So these papers suggest that it did. 01:06:23.420 --> 01:06:27.868 align:middle line:84% Now, they use an approach which I think Glenn has talked about, 01:06:27.868 --> 01:06:30.160 align:middle line:84% which is a structure conduct performance approach where 01:06:30.160 --> 01:06:33.800 align:middle line:84% you essentially regress prices on measures of concentration. 01:06:33.800 --> 01:06:38.950 align:middle line:84% And we know that this has well-known issues, 01:06:38.950 --> 01:06:42.550 align:middle line:84% in part because these concentration measures are 01:06:42.550 --> 01:06:46.350 align:middle line:84% themselves driven by quantities which are endogenous. 01:06:46.350 --> 01:06:51.670 align:middle line:84% So today, we're going to talk about one paper that 01:06:51.670 --> 01:06:53.950 align:middle line:84% tries to do this in a structural way, 01:06:53.950 --> 01:06:56.440 align:middle line:84% and a companion paper to that that documents 01:06:56.440 --> 01:07:00.400 align:middle line:84% a little bit to what extent this is an actual issue in the US 01:07:00.400 --> 01:07:01.820 align:middle line:90% economy. 01:07:01.820 --> 01:07:07.250 align:middle line:84% So just very briefly, what is this idea here? 01:07:07.250 --> 01:07:09.860 align:middle line:84% Formally, the common ownership hypothesis, 01:07:09.860 --> 01:07:13.190 align:middle line:84% according to this paper by Rotemberg-- and here, 01:07:13.190 --> 01:07:16.430 align:middle line:84% I'm using the notation of O'Brien and Salop-- 01:07:16.430 --> 01:07:20.420 align:middle line:84% is the idea that we have an investor, S, 01:07:20.420 --> 01:07:25.550 align:middle line:84% or different investors, that own a portfolio 01:07:25.550 --> 01:07:28.592 align:middle line:84% of shares in different companies. 01:07:28.592 --> 01:07:30.800 align:middle line:84% And according to the amount of shares that they hold, 01:07:30.800 --> 01:07:33.350 align:middle line:84% they are entitled to some fraction of the profits of each 01:07:33.350 --> 01:07:35.080 align:middle line:90% of the firms that they hold. 01:07:35.080 --> 01:07:38.360 align:middle line:90% 01:07:38.360 --> 01:07:45.870 align:middle line:84% And you have firms that, in this model, 01:07:45.870 --> 01:07:49.220 align:middle line:84% maximizing the value of the shareholders. 01:07:49.220 --> 01:07:52.150 align:middle line:90% So how do they do that? 01:07:52.150 --> 01:07:54.045 align:middle line:84% By virtue of having different portfolios, 01:07:54.045 --> 01:07:55.420 align:middle line:84% different shareholders might have 01:07:55.420 --> 01:07:56.628 align:middle line:90% slightly different interests. 01:07:56.628 --> 01:07:58.600 align:middle line:84% Their interests might not be fully aligned. 01:07:58.600 --> 01:08:00.790 align:middle line:84% So in this model, the way this is handled 01:08:00.790 --> 01:08:04.193 align:middle line:84% is that we're saying the firm places different weights on what 01:08:04.193 --> 01:08:05.360 align:middle line:90% different shareholders want. 01:08:05.360 --> 01:08:07.540 align:middle line:90% So these are pareto weights. 01:08:07.540 --> 01:08:11.110 align:middle line:84% So you can write the firm's objective function, 01:08:11.110 --> 01:08:12.850 align:middle line:84% which is a function of its own prices 01:08:12.850 --> 01:08:18.692 align:middle line:84% and other prices in the market, as a weighted sum over what 01:08:18.692 --> 01:08:20.025 align:middle line:90% its different shareholders want. 01:08:20.025 --> 01:08:23.330 align:middle line:90% 01:08:23.330 --> 01:08:29.830 align:middle line:84% Now, if you plug this in here using this formula from above, 01:08:29.830 --> 01:08:35.380 align:middle line:84% the shareholder objective function, the portfolio, 01:08:35.380 --> 01:08:38.590 align:middle line:84% you find that you're suddenly in a situation 01:08:38.590 --> 01:08:42.040 align:middle line:84% where this firm, F, might be caring 01:08:42.040 --> 01:08:46.000 align:middle line:84% about the profits of some firm, G, in the market 01:08:46.000 --> 01:08:48.800 align:middle line:84% through the incentives of its shareholders. 01:08:48.800 --> 01:08:51.670 align:middle line:90% 01:08:51.670 --> 01:08:53.330 align:middle line:90% So these are the pareto weights. 01:08:53.330 --> 01:08:58.500 align:middle line:84% These are these cash flow shares that the investor, S, 01:08:58.500 --> 01:09:03.270 align:middle line:90% is entitled to terms of firm G. 01:09:03.270 --> 01:09:05.240 align:middle line:90% So you can rewrite this. 01:09:05.240 --> 01:09:11.660 align:middle line:84% You can do some algebra and normalize the share of firm F 01:09:11.660 --> 01:09:14.300 align:middle line:84% to 1, and can write the objective function 01:09:14.300 --> 01:09:19.910 align:middle line:84% of firm F as its own profits plus a weighted sum 01:09:19.910 --> 01:09:23.675 align:middle line:90% of other firms in the market. 01:09:23.675 --> 01:09:27.260 align:middle line:90% 01:09:27.260 --> 01:09:30.109 align:middle line:84% And so we call these profit weights here kappa. 01:09:30.109 --> 01:09:34.970 align:middle line:84% They're a function of these pareto weights and these betas. 01:09:34.970 --> 01:09:37.100 align:middle line:84% So just to link this back to what 01:09:37.100 --> 01:09:40.450 align:middle line:84% I discussed at the very beginning of class today, 01:09:40.450 --> 01:09:44.770 align:middle line:84% this should remind you of something 01:09:44.770 --> 01:09:46.840 align:middle line:84% that we have seen on the first slide. 01:09:46.840 --> 01:09:49.390 align:middle line:84% So here, we have a firm that's maximizing 01:09:49.390 --> 01:09:52.779 align:middle line:84% its own profits plus some weighted sum of other firm's 01:09:52.779 --> 01:09:54.200 align:middle line:90% profits. 01:09:54.200 --> 01:09:58.040 align:middle line:84% This is very similar to this idea of a multi-product firm 01:09:58.040 --> 01:10:03.110 align:middle line:84% that owns different brands, and will take into account that if I 01:10:03.110 --> 01:10:04.850 align:middle line:84% lower my price on one of my brands, 01:10:04.850 --> 01:10:06.290 align:middle line:84% that might potentially cannibalize 01:10:06.290 --> 01:10:14.110 align:middle line:84% one of my other brands and divert some market share. 01:10:14.110 --> 01:10:22.110 align:middle line:84% So here, you get a very similar first order condition for firm J 01:10:22.110 --> 01:10:26.340 align:middle line:84% that places these profit rates that come, again, 01:10:26.340 --> 01:10:32.430 align:middle line:84% through that are a function of the shareholder holdings 01:10:32.430 --> 01:10:36.450 align:middle line:90% on other firms' profits. 01:10:36.450 --> 01:10:40.130 align:middle line:84% So I hope this connection is clear. 01:10:40.130 --> 01:10:42.860 align:middle line:90% 01:10:42.860 --> 01:10:48.350 align:middle line:84% So what can we actually say about those profit weights? 01:10:48.350 --> 01:10:52.940 align:middle line:84% So this is what this companion paper by Backus, Conlon, 01:10:52.940 --> 01:10:54.530 align:middle line:90% and Sinkinson does. 01:10:54.530 --> 01:10:56.210 align:middle line:84% The first document, is it, indeed, 01:10:56.210 --> 01:11:00.290 align:middle line:84% true that, just as a pure matter of accounting, 01:11:00.290 --> 01:11:03.230 align:middle line:84% that firms are placing, or have these incentives 01:11:03.230 --> 01:11:08.750 align:middle line:84% to place more weight on what their competitors are doing, 01:11:08.750 --> 01:11:12.526 align:middle line:84% or any other random firm in the market? 01:11:12.526 --> 01:11:16.700 align:middle line:84% So first, if you look at the time series of what 01:11:16.700 --> 01:11:18.883 align:middle line:84% fraction of the typical firm is actually 01:11:18.883 --> 01:11:21.050 align:middle line:84% held by one of these large institutional investors-- 01:11:21.050 --> 01:11:23.360 align:middle line:84% BlackRock, Vanguard, and State Street-- 01:11:23.360 --> 01:11:28.364 align:middle line:84% we find that this did, indeed, go up a lot. 01:11:28.364 --> 01:11:32.642 align:middle line:84% And, I mean, this is actually not just-- 01:11:32.642 --> 01:11:34.350 align:middle line:84% this trend towards diversified investment 01:11:34.350 --> 01:11:36.880 align:middle line:84% is not just driven by these institutional investors. 01:11:36.880 --> 01:11:40.590 align:middle line:84% This was another result in the paper was already going on even 01:11:40.590 --> 01:11:41.930 align:middle line:90% before that. 01:11:41.930 --> 01:11:45.930 align:middle line:84% So they use some data from SEC filings 01:11:45.930 --> 01:11:52.490 align:middle line:84% that large investment managers have to submit to document this. 01:11:52.490 --> 01:11:55.310 align:middle line:84% So then, they compute these profit weights 01:11:55.310 --> 01:11:59.900 align:middle line:84% that I've just shown you that you can derive from the firm's 01:11:59.900 --> 01:12:01.880 align:middle line:84% profit maximization problem, taking 01:12:01.880 --> 01:12:05.075 align:middle line:84% into account the incentives of the shareholders. 01:12:05.075 --> 01:12:07.057 align:middle line:84% So the one issue that they have here is they 01:12:07.057 --> 01:12:08.640 align:middle line:84% don't actually see this pareto weight. 01:12:08.640 --> 01:12:11.057 align:middle line:84% So they make some assumption that these pareto weights are 01:12:11.057 --> 01:12:14.860 align:middle line:84% a function of these shareholder weights beta, 01:12:14.860 --> 01:12:19.400 align:middle line:84% and so they show how this time series look 01:12:19.400 --> 01:12:22.410 align:middle line:90% for different functional forms. 01:12:22.410 --> 01:12:23.930 align:middle line:90% But the general message is-- 01:12:23.930 --> 01:12:25.910 align:middle line:84% and you can also break this down by industry-- 01:12:25.910 --> 01:12:29.810 align:middle line:84% that if you take a firm in the US, 01:12:29.810 --> 01:12:32.450 align:middle line:84% it's a publicly traded firm in the US economy, 01:12:32.450 --> 01:12:35.330 align:middle line:84% and ask how much weight does that firm place on the profits 01:12:35.330 --> 01:12:38.660 align:middle line:84% according to these diversified investments 01:12:38.660 --> 01:12:41.470 align:middle line:84% on any other random firm, that this has gone up over time. 01:12:41.470 --> 01:12:45.162 align:middle line:90% 01:12:45.162 --> 01:12:49.160 align:middle line:84% So this suggests that in principle, these kinds 01:12:49.160 --> 01:12:52.330 align:middle line:90% of incentives could be at play. 01:12:52.330 --> 01:12:56.050 align:middle line:90% 01:12:56.050 --> 01:12:57.165 align:middle line:90% Any questions here? 01:12:57.165 --> 01:12:59.653 align:middle line:90% 01:12:59.653 --> 01:13:02.320 align:middle line:84% AUDIENCE: Is there any sense of how sensitive these measurements 01:13:02.320 --> 01:13:03.403 align:middle line:90% are to the pareto weights? 01:13:03.403 --> 01:13:06.940 align:middle line:84% Like were there different specifications tested, 01:13:06.940 --> 01:13:08.140 align:middle line:90% or [INAUDIBLE]? 01:13:08.140 --> 01:13:10.820 align:middle line:84% PROFESSOR: This trend is not very sensitive to that. 01:13:10.820 --> 01:13:13.840 align:middle line:84% The levels are-- I mean, in the paper, they show you, 01:13:13.840 --> 01:13:15.880 align:middle line:84% I think, three different types of weights. 01:13:15.880 --> 01:13:19.360 align:middle line:84% But the general message is that this has been upwards 01:13:19.360 --> 01:13:21.070 align:middle line:90% trending regardless of-- 01:13:21.070 --> 01:13:22.930 align:middle line:84% I mean, you can then, of course, argue, 01:13:22.930 --> 01:13:26.800 align:middle line:84% is that an exhaustive search over the potential specification 01:13:26.800 --> 01:13:27.598 align:middle line:90% that you could use? 01:13:27.598 --> 01:13:28.098 align:middle line:90% Yeah. 01:13:28.098 --> 01:13:34.740 align:middle line:90% 01:13:34.740 --> 01:13:37.450 align:middle line:84% So coming back to the cereals market. 01:13:37.450 --> 01:13:39.180 align:middle line:84% So again, what I've just shown you 01:13:39.180 --> 01:13:43.320 align:middle line:84% is from this descriptive paper is that they document 01:13:43.320 --> 01:13:49.757 align:middle line:84% that these incentives, the fact that firms may internalize 01:13:49.757 --> 01:13:51.840 align:middle line:84% profits of other firms that they're competing with 01:13:51.840 --> 01:13:54.990 align:middle line:84% could potentially be an issue based on the time series 01:13:54.990 --> 01:13:58.800 align:middle line:84% that we observe because of this trend towards more 01:13:58.800 --> 01:14:00.120 align:middle line:90% diversified investment. 01:14:00.120 --> 01:14:03.660 align:middle line:84% If you look at the breakfast cereal 01:14:03.660 --> 01:14:05.730 align:middle line:84% market and the different producers here, 01:14:05.730 --> 01:14:11.880 align:middle line:84% you'll find that, indeed, BlackRock also, in 2016, held 01:14:11.880 --> 01:14:17.670 align:middle line:84% a large share of General Mills, of Kellogg, of Quaker Oats, 01:14:17.670 --> 01:14:21.960 align:middle line:84% and Post brands-- so all the big producers. 01:14:21.960 --> 01:14:24.300 align:middle line:84% So in principle, this could be an issue in this market, 01:14:24.300 --> 01:14:24.800 align:middle line:90% as well. 01:14:24.800 --> 01:14:30.850 align:middle line:90% 01:14:30.850 --> 01:14:32.780 align:middle line:84% OK, so what is it that they're doing? 01:14:32.780 --> 01:14:35.080 align:middle line:84% So I'm going to be-- you know, I'll give you 01:14:35.080 --> 01:14:36.560 align:middle line:90% the main economic idea here. 01:14:36.560 --> 01:14:41.680 align:middle line:84% This paper, big part of the contribution 01:14:41.680 --> 01:14:46.240 align:middle line:84% is develop a new and efficient testing procedure for this 01:14:46.240 --> 01:14:50.420 align:middle line:84% and allowing for flexible cost function. 01:14:50.420 --> 01:14:53.260 align:middle line:84% So I'm not going to talk about the econometric details, 01:14:53.260 --> 01:14:59.650 align:middle line:84% but this is very much in the same spirit as the other papers 01:14:59.650 --> 01:15:01.030 align:middle line:90% that we have been discussing-- 01:15:01.030 --> 01:15:03.370 align:middle line:90% Nevo's paper on conduct testing. 01:15:03.370 --> 01:15:09.730 align:middle line:84% So the idea is that with my demand model in hand and conduct 01:15:09.730 --> 01:15:14.200 align:middle line:84% assumption, I can invert my first order conditions 01:15:14.200 --> 01:15:17.170 align:middle line:84% as a function of the observed market shares, prices, 01:15:17.170 --> 01:15:18.550 align:middle line:90% and the demand model. 01:15:18.550 --> 01:15:20.740 align:middle line:84% Again, the demand model, of course, 01:15:20.740 --> 01:15:23.500 align:middle line:84% has some demand parameters, and can back out 01:15:23.500 --> 01:15:26.445 align:middle line:90% my effective marginal cost. 01:15:26.445 --> 01:15:29.120 align:middle line:90% 01:15:29.120 --> 01:15:34.700 align:middle line:84% At the same time, I may be willing to also assume 01:15:34.700 --> 01:15:38.480 align:middle line:84% a functional form for the marginal cost directly. 01:15:38.480 --> 01:15:42.590 align:middle line:84% Again, because I know that the cost of producing cereals 01:15:42.590 --> 01:15:46.200 align:middle line:84% are driven by the cost of sugar, wheat, flour, and so on. 01:15:46.200 --> 01:15:49.665 align:middle line:90% 01:15:49.665 --> 01:15:56.670 align:middle line:84% So what they do is they basically specify this equation 01:15:56.670 --> 01:16:01.830 align:middle line:84% here by allowing for a flexible supply function or marginal cost 01:16:01.830 --> 01:16:02.680 align:middle line:90% function. 01:16:02.680 --> 01:16:06.270 align:middle line:84% And based off this, get additional moment restrictions 01:16:06.270 --> 01:16:09.360 align:middle line:84% that they use for the construction of a GMM objective 01:16:09.360 --> 01:16:13.582 align:middle line:84% function under different conduct models, M. 01:16:13.582 --> 01:16:15.290 align:middle line:84% And of course, in particular, the conduct 01:16:15.290 --> 01:16:17.060 align:middle line:84% model that they are interested in 01:16:17.060 --> 01:16:20.900 align:middle line:84% is the common ownership hypothesis. 01:16:20.900 --> 01:16:23.870 align:middle line:84% And that is, again, very similar to this idea 01:16:23.870 --> 01:16:27.590 align:middle line:84% that I might be internalizing the profits not 01:16:27.590 --> 01:16:32.540 align:middle line:84% of other products that I hold, but of other firms 01:16:32.540 --> 01:16:35.070 align:middle line:84% through the incentives of my diversified investors. 01:16:35.070 --> 01:16:39.020 align:middle line:90% 01:16:39.020 --> 01:16:44.660 align:middle line:84% So the idea here is to construct this equation 01:16:44.660 --> 01:16:46.290 align:middle line:90% for different conduct models. 01:16:46.290 --> 01:16:49.220 align:middle line:84% You get some GMM objective function out of that. 01:16:49.220 --> 01:16:51.170 align:middle line:84% And with this GMM objective function, 01:16:51.170 --> 01:16:53.880 align:middle line:84% you can construct a test statistic. 01:16:53.880 --> 01:16:56.450 align:middle line:90% This here is a test that's-- 01:16:56.450 --> 01:16:59.120 align:middle line:84% it's a non-nested test by Rivers and Vuong. 01:16:59.120 --> 01:17:02.420 align:middle line:84% The idea of this test is that there are two different models 01:17:02.420 --> 01:17:06.350 align:middle line:84% of conduct, and the null hypothesis is that both of them 01:17:06.350 --> 01:17:09.940 align:middle line:84% are equally far from the true model. 01:17:09.940 --> 01:17:12.750 align:middle line:84% But you're always testing the two models against each other. 01:17:12.750 --> 01:17:14.860 align:middle line:84% So what, at the end of the day, you can say is, 01:17:14.860 --> 01:17:17.350 align:middle line:84% is this model better describing the data 01:17:17.350 --> 01:17:20.320 align:middle line:84% that I observed in some other model? 01:17:20.320 --> 01:17:23.020 align:middle line:84% And so, for instance, here, based on, 01:17:23.020 --> 01:17:27.940 align:middle line:84% let's say, common ownership or own profit maximization, 01:17:27.940 --> 01:17:30.940 align:middle line:84% if this test statistic here becomes very large, 01:17:30.940 --> 01:17:35.540 align:middle line:84% then I'm rejecting model 1 in favor of model 2. 01:17:35.540 --> 01:17:38.180 align:middle line:84% Because the violation of the moment 01:17:38.180 --> 01:17:40.590 align:middle line:84% conditions is much larger under model 1. 01:17:40.590 --> 01:17:42.300 align:middle line:90% This is larger. 01:17:42.300 --> 01:17:48.510 align:middle line:84% So it's shown that this is asymptotically normal. 01:17:48.510 --> 01:17:53.510 align:middle line:84% So you can use just standard criteria 01:17:53.510 --> 01:17:58.610 align:middle line:84% for the normal distribution to test with this test statistic. 01:17:58.610 --> 01:18:01.880 align:middle line:84% So just, again, to link this back to what we have already 01:18:01.880 --> 01:18:05.120 align:middle line:84% discussed today, so like Nevo, they estimated the demand model 01:18:05.120 --> 01:18:06.890 align:middle line:90% for cereals. 01:18:06.890 --> 01:18:10.160 align:middle line:84% Do this first order condition inversion. 01:18:10.160 --> 01:18:15.020 align:middle line:84% And like Nevo, they test for different models of conduct 01:18:15.020 --> 01:18:17.060 align:middle line:84% with a focus on common ownership. 01:18:17.060 --> 01:18:21.290 align:middle line:84% But unlike Nevo, what they do is instead of comparing 01:18:21.290 --> 01:18:23.510 align:middle line:84% the markups that I get from this inversion here 01:18:23.510 --> 01:18:26.720 align:middle line:84% of the first order conditions directly to observed markups, 01:18:26.720 --> 01:18:29.750 align:middle line:84% they're seeing whether the marginal costs 01:18:29.750 --> 01:18:32.420 align:middle line:84% that they back out are consistent 01:18:32.420 --> 01:18:37.850 align:middle line:84% with the functional form that I imposed on marginal cost. 01:18:37.850 --> 01:18:41.220 align:middle line:84% And they basically now used demand side variation-- 01:18:41.220 --> 01:18:44.150 align:middle line:84% so things that move around the demand side but I exclude it 01:18:44.150 --> 01:18:45.680 align:middle line:90% from the supply side-- 01:18:45.680 --> 01:18:49.820 align:middle line:84% as overidentifying restrictions to test this model 01:18:49.820 --> 01:18:52.490 align:middle line:84% and construct this test statistic. 01:18:52.490 --> 01:18:56.180 align:middle line:84% And this allows them to compare, let's say, 01:18:56.180 --> 01:19:01.380 align:middle line:84% the common ownership model against other types of models. 01:19:01.380 --> 01:19:04.800 align:middle line:84% So let me tell you what they find from this. 01:19:04.800 --> 01:19:07.050 align:middle line:84% Again, there's some details that I haven't touched on, 01:19:07.050 --> 01:19:09.840 align:middle line:84% which is how do you specify the supply function, 01:19:09.840 --> 01:19:12.150 align:middle line:84% and what are the best instruments? 01:19:12.150 --> 01:19:15.300 align:middle line:84% Their preferred instrument are these optimal demand side 01:19:15.300 --> 01:19:15.850 align:middle line:90% instruments. 01:19:15.850 --> 01:19:18.180 align:middle line:84% Again, now you want to say something 01:19:18.180 --> 01:19:21.600 align:middle line:84% about this flexible supply side, so you use demand side variation 01:19:21.600 --> 01:19:23.670 align:middle line:90% which moves around the markup. 01:19:23.670 --> 01:19:27.720 align:middle line:84% And under the right model, these are 01:19:27.720 --> 01:19:29.250 align:middle line:84% excluded instruments-- instruments 01:19:29.250 --> 01:19:30.875 align:middle line:84% that are excluded from the supply side. 01:19:30.875 --> 01:19:36.160 align:middle line:90% 01:19:36.160 --> 01:19:39.080 align:middle line:84% These should be orthogonal to the variation in omega. 01:19:39.080 --> 01:19:43.390 align:middle line:84% And so what you find here is that using 01:19:43.390 --> 01:19:45.250 align:middle line:84% these optimal instruments, they soundly 01:19:45.250 --> 01:19:49.210 align:middle line:84% reject your different models of common ownership 01:19:49.210 --> 01:19:52.237 align:middle line:90% against own profit maximization. 01:19:52.237 --> 01:19:54.570 align:middle line:84% And so there are different specifications whether or not 01:19:54.570 --> 01:19:59.130 align:middle line:84% you allow for firms responding to a lack of these profit rates. 01:19:59.130 --> 01:20:03.150 align:middle line:84% They also reject perfect competition 01:20:03.150 --> 01:20:07.670 align:middle line:84% using these optimal instruments as well as a full monopoly 01:20:07.670 --> 01:20:09.630 align:middle line:90% model. 01:20:09.630 --> 01:20:15.810 align:middle line:84% So it seems that, against what these time trends suggest, 01:20:15.810 --> 01:20:19.350 align:middle line:84% that firms do not seem to be internalizing 01:20:19.350 --> 01:20:21.100 align:middle line:90% these common ownership effects. 01:20:21.100 --> 01:20:23.550 align:middle line:84% And instead, the data is much more consistent 01:20:23.550 --> 01:20:27.320 align:middle line:90% with own profit maximization. 01:20:27.320 --> 01:20:34.070 align:middle line:84% So this is, of course, just a case study. 01:20:34.070 --> 01:20:41.090 align:middle line:84% So this doesn't speak to the US economy in its entirety. 01:20:41.090 --> 01:20:45.950 align:middle line:84% So again, to wrap up, this is a very interesting hypothesis 01:20:45.950 --> 01:20:48.740 align:middle line:84% that has been very hotly debated in IO, 01:20:48.740 --> 01:20:52.480 align:middle line:84% and people feel very strongly about it. 01:20:52.480 --> 01:20:56.080 align:middle line:90% 01:20:56.080 --> 01:20:59.680 align:middle line:84% You could plausibly conjecture that firms 01:20:59.680 --> 01:21:01.743 align:middle line:84% behave in this way based on these time trends 01:21:01.743 --> 01:21:02.660 align:middle line:90% that we have observed. 01:21:02.660 --> 01:21:08.970 align:middle line:84% It's not crazy to think that they somehow internalize this. 01:21:08.970 --> 01:21:11.550 align:middle line:90% This case study rejects this. 01:21:11.550 --> 01:21:13.170 align:middle line:84% And more generally, people have been 01:21:13.170 --> 01:21:15.570 align:middle line:84% highly skeptical of this hypothesis, 01:21:15.570 --> 01:21:19.740 align:middle line:84% in part due to a lack of a plausible mechanism of how 01:21:19.740 --> 01:21:21.630 align:middle line:90% these firms would actually-- 01:21:21.630 --> 01:21:24.030 align:middle line:84% how these institutional investors would actually 01:21:24.030 --> 01:21:29.610 align:middle line:84% bring the firms to, let's say, soften competition, or engage 01:21:29.610 --> 01:21:31.560 align:middle line:90% in less price competition. 01:21:31.560 --> 01:21:37.740 align:middle line:84% So the mechanism is still a bit elusive, 01:21:37.740 --> 01:21:40.820 align:middle line:84% and, to the best of my knowledge, 01:21:40.820 --> 01:21:42.860 align:middle line:84% there has been no smoking gun-- like a story 01:21:42.860 --> 01:21:46.280 align:middle line:84% where it's clear that they have done this. 01:21:46.280 --> 01:21:48.920 align:middle line:84% Some people argue that managerial incentives 01:21:48.920 --> 01:21:51.880 align:middle line:84% could be a potential explanation for this. 01:21:51.880 --> 01:21:55.640 align:middle line:84% I think this is an exciting topic that you should make up 01:21:55.640 --> 01:21:57.680 align:middle line:90% your own mind on. 01:21:57.680 --> 01:22:01.690 align:middle line:90% And that's all I have for today. 01:22:01.690 --> 01:22:21.000 align:middle line:90%