WEBVTT 00:00:00.000 --> 00:00:01.992 align:middle line:90% [SQUEAKING] 00:00:01.992 --> 00:00:03.486 align:middle line:90% [RUSTLING] 00:00:03.486 --> 00:00:05.478 align:middle line:90% [CLICKING] 00:00:05.478 --> 00:00:17.730 align:middle line:90% 00:00:17.730 --> 00:00:19.390 align:middle line:90% TOBIAS SALZ: OK. 00:00:19.390 --> 00:00:20.990 align:middle line:90% Thanks so much for being here. 00:00:20.990 --> 00:00:24.190 align:middle line:84% I know it's a very busy time of the semester. 00:00:24.190 --> 00:00:28.550 align:middle line:84% So I'm going to talk a little bit more about search models. 00:00:28.550 --> 00:00:30.910 align:middle line:84% And I know that Glenn has already 00:00:30.910 --> 00:00:33.850 align:middle line:84% introduced you both to the theory side and some empirics. 00:00:33.850 --> 00:00:40.420 align:middle line:84% So I'm going to be brief and motivating the topic. 00:00:40.420 --> 00:00:43.690 align:middle line:84% So there was a famous article published 00:00:43.690 --> 00:00:48.400 align:middle line:84% by George Stigler in 1961 where he made the observation 00:00:48.400 --> 00:00:52.060 align:middle line:84% that in many markets, it's not the case that there's 00:00:52.060 --> 00:00:53.830 align:middle line:84% a central marketplace where everybody 00:00:53.830 --> 00:00:55.250 align:middle line:90% observes the relevant prices. 00:00:55.250 --> 00:00:59.800 align:middle line:84% But oftentimes you actually have to move around, 00:00:59.800 --> 00:01:03.100 align:middle line:84% go to different places, incur physical search costs 00:01:03.100 --> 00:01:05.290 align:middle line:90% to figure out what prices are. 00:01:05.290 --> 00:01:09.100 align:middle line:84% Or maybe it's mentally costly to figure out 00:01:09.100 --> 00:01:10.550 align:middle line:90% what the right price is. 00:01:10.550 --> 00:01:13.900 align:middle line:84% And so that introduces frictions, 00:01:13.900 --> 00:01:16.200 align:middle line:90% which he termed search. 00:01:16.200 --> 00:01:19.400 align:middle line:84% And so in this quote here, he says 00:01:19.400 --> 00:01:21.990 align:middle line:84% that we can look at the dispersion in prices. 00:01:21.990 --> 00:01:23.720 align:middle line:84% And in some sense, they're a measure 00:01:23.720 --> 00:01:26.365 align:middle line:84% of the amount of ignorance in the market. 00:01:26.365 --> 00:01:29.230 align:middle line:90% 00:01:29.230 --> 00:01:34.530 align:middle line:84% So why do we want to look at economic settings 00:01:34.530 --> 00:01:36.280 align:middle line:84% through the lens of a search model? 00:01:36.280 --> 00:01:39.960 align:middle line:84% Well, so first of all, there's several positive arguments 00:01:39.960 --> 00:01:41.140 align:middle line:90% why we may want to do that. 00:01:41.140 --> 00:01:42.810 align:middle line:84% So first of all, in many settings, 00:01:42.810 --> 00:01:47.190 align:middle line:84% we do see that products that are very similar or essentially 00:01:47.190 --> 00:01:50.860 align:middle line:84% identical are sold for very different prices. 00:01:50.860 --> 00:01:53.040 align:middle line:84% And so as researchers, we would like 00:01:53.040 --> 00:01:56.865 align:middle line:84% to make sense of that observation how can this be. 00:01:56.865 --> 00:01:59.796 align:middle line:90% 00:01:59.796 --> 00:02:04.500 align:middle line:84% And again, so one potential explanation 00:02:04.500 --> 00:02:06.330 align:middle line:84% which we will explore today empirically 00:02:06.330 --> 00:02:08.699 align:middle line:84% is that with search frictions, it 00:02:08.699 --> 00:02:10.860 align:middle line:84% could be that you get a mixed strategy pricing 00:02:10.860 --> 00:02:14.700 align:middle line:84% equilibrium, which leads to equilibrium price dispersion. 00:02:14.700 --> 00:02:19.050 align:middle line:84% But then within the theme of empirical IO, 00:02:19.050 --> 00:02:22.000 align:middle line:90% it's also normatively important. 00:02:22.000 --> 00:02:25.770 align:middle line:84% So when we quantify, let's say, welfare in markets, 00:02:25.770 --> 00:02:29.610 align:middle line:84% we might want to take search frictions into account 00:02:29.610 --> 00:02:30.990 align:middle line:84% foremost because they're actually 00:02:30.990 --> 00:02:32.850 align:middle line:90% a source of market power. 00:02:32.850 --> 00:02:37.050 align:middle line:84% So in IO, we are predominantly concerned with various ways 00:02:37.050 --> 00:02:38.920 align:middle line:90% that firms gain market power. 00:02:38.920 --> 00:02:43.890 align:middle line:84% Such frictions are one potential source of market power. 00:02:43.890 --> 00:02:46.290 align:middle line:84% There are also, depending on how you think 00:02:46.290 --> 00:02:48.010 align:middle line:90% about them, a social cost. 00:02:48.010 --> 00:02:52.170 align:middle line:84% So if people have to engage in wasteful effort 00:02:52.170 --> 00:02:54.900 align:middle line:84% to drive around to various car dealerships 00:02:54.900 --> 00:02:57.190 align:middle line:84% to figure out what the right car prices are, 00:02:57.190 --> 00:02:59.260 align:middle line:90% that's socially costly. 00:02:59.260 --> 00:03:01.860 align:middle line:84% And if we do welfare calculations, 00:03:01.860 --> 00:03:05.520 align:middle line:84% we may perhaps want to account for those costs. 00:03:05.520 --> 00:03:08.370 align:middle line:84% Such costs can also lead to misallocation. 00:03:08.370 --> 00:03:10.410 align:middle line:84% So if you think about you're looking for, 00:03:10.410 --> 00:03:12.870 align:middle line:84% let's say, a hotel for your vacation 00:03:12.870 --> 00:03:15.360 align:middle line:84% or, let's say, a car or something else, 00:03:15.360 --> 00:03:17.460 align:middle line:84% you might not, in fact, find the product 00:03:17.460 --> 00:03:21.270 align:middle line:84% that you would like to purchase the most at a given price point. 00:03:21.270 --> 00:03:24.150 align:middle line:84% And similarly, if you think about a firm, let's say, 00:03:24.150 --> 00:03:26.070 align:middle line:84% purchasing inputs from a supplier, 00:03:26.070 --> 00:03:27.900 align:middle line:84% the firm might not find a supplier 00:03:27.900 --> 00:03:31.410 align:middle line:84% that can offer the specific product at the lowest cost. 00:03:31.410 --> 00:03:34.770 align:middle line:84% And so that leads to welfare losses 00:03:34.770 --> 00:03:37.350 align:middle line:84% that, again, we might want to quantify. 00:03:37.350 --> 00:03:40.782 align:middle line:90% 00:03:40.782 --> 00:03:42.770 align:middle line:90% So this is general. 00:03:42.770 --> 00:03:47.090 align:middle line:84% For the second paper that I'm going to discuss today, 00:03:47.090 --> 00:03:49.940 align:middle line:84% I will look at markets in which prices 00:03:49.940 --> 00:03:52.550 align:middle line:90% are not posted but negotiated. 00:03:52.550 --> 00:03:54.690 align:middle line:84% And these are very interesting markets. 00:03:54.690 --> 00:03:57.320 align:middle line:84% And we have recently much more data 00:03:57.320 --> 00:04:00.800 align:middle line:84% which allow us to make some empirical progress 00:04:00.800 --> 00:04:02.280 align:middle line:90% on these kinds of markets. 00:04:02.280 --> 00:04:05.030 align:middle line:84% So if you think in many settings, 00:04:05.030 --> 00:04:08.700 align:middle line:84% people negotiate prices with the seller. 00:04:08.700 --> 00:04:12.590 align:middle line:84% So if you buy a car, a house, a mortgage, 00:04:12.590 --> 00:04:16.190 align:middle line:84% so many kinds of consumer financial markets 00:04:16.190 --> 00:04:19.700 align:middle line:84% have this feature that prices are somehow negotiated 00:04:19.700 --> 00:04:21.510 align:middle line:90% and various kinds of services. 00:04:21.510 --> 00:04:26.750 align:middle line:84% So if you get a contractor for your home or, let's say, 00:04:26.750 --> 00:04:28.950 align:middle line:84% a plumber or something like that. 00:04:28.950 --> 00:04:31.280 align:middle line:84% So in these markets, it's oftentimes 00:04:31.280 --> 00:04:33.230 align:middle line:84% particularly hard for consumers to know 00:04:33.230 --> 00:04:36.230 align:middle line:84% what is the fair or the right price, 00:04:36.230 --> 00:04:38.030 align:middle line:84% because searching is particularly costly 00:04:38.030 --> 00:04:40.380 align:middle line:84% and you have to go through some process. 00:04:40.380 --> 00:04:44.570 align:middle line:84% And there is oftentimes this element of idiosyncratic pricing 00:04:44.570 --> 00:04:47.970 align:middle line:84% that makes that even more difficult. 00:04:47.970 --> 00:04:49.588 align:middle line:84% What makes this also very interesting, 00:04:49.588 --> 00:04:51.630 align:middle line:84% and this is something that we will explore today, 00:04:51.630 --> 00:04:55.260 align:middle line:84% is that there are potential distributional implications 00:04:55.260 --> 00:04:58.650 align:middle line:84% of pricing And policies such as mergers that we 00:04:58.650 --> 00:05:01.226 align:middle line:90% think about as economists. 00:05:01.226 --> 00:05:04.000 align:middle line:90% 00:05:04.000 --> 00:05:06.070 align:middle line:84% An empirical challenge that arises 00:05:06.070 --> 00:05:10.780 align:middle line:84% when looking at those markets is that if sellers are quoting 00:05:10.780 --> 00:05:14.380 align:middle line:84% different prices to buyers, but we only 00:05:14.380 --> 00:05:16.240 align:middle line:84% observe the transaction prices, this 00:05:16.240 --> 00:05:18.610 align:middle line:90% is a selected set of prices. 00:05:18.610 --> 00:05:21.090 align:middle line:84% And so we need to know how we can 00:05:21.090 --> 00:05:23.340 align:middle line:84% map this back to the prices that are actually offered. 00:05:23.340 --> 00:05:28.410 align:middle line:90% 00:05:28.410 --> 00:05:31.620 align:middle line:90% So here's the roadmap for today. 00:05:31.620 --> 00:05:35.700 align:middle line:84% I'm going to talk first about a paper that's 00:05:35.700 --> 00:05:37.860 align:middle line:84% kind of conceptually very interesting 00:05:37.860 --> 00:05:40.860 align:middle line:84% and ties back to the theory lecture 00:05:40.860 --> 00:05:44.670 align:middle line:90% that Glenn held on search. 00:05:44.670 --> 00:05:48.300 align:middle line:84% And then a more empirically oriented applications 00:05:48.300 --> 00:05:50.160 align:middle line:90% on mergers and search markets. 00:05:50.160 --> 00:05:52.960 align:middle line:84% And if I have time, I'll probably not have too much time, 00:05:52.960 --> 00:05:54.960 align:middle line:84% I'm also going to talk about my job market paper 00:05:54.960 --> 00:05:59.065 align:middle line:84% for a little bit, which is in this space. 00:05:59.065 --> 00:06:02.580 align:middle line:90% 00:06:02.580 --> 00:06:08.870 align:middle line:84% So let me talk about this paper by Hong and Shum. 00:06:08.870 --> 00:06:13.070 align:middle line:84% So it's really an interesting exercise. 00:06:13.070 --> 00:06:16.460 align:middle line:84% And one thing that makes it interesting 00:06:16.460 --> 00:06:20.570 align:middle line:84% is it takes really this idea that was in Stigler's quote 00:06:20.570 --> 00:06:21.770 align:middle line:90% to the extreme. 00:06:21.770 --> 00:06:29.000 align:middle line:84% So it asks what can we learn from a price distribution? 00:06:29.000 --> 00:06:33.410 align:middle line:84% And it imposes a specific equilibrium, namely 00:06:33.410 --> 00:06:36.470 align:middle line:84% that firms are playing mixed strategy, mixed strategy 00:06:36.470 --> 00:06:38.150 align:middle line:90% pricing. 00:06:38.150 --> 00:06:40.280 align:middle line:84% The model that it builds on is Burdett and Judd. 00:06:40.280 --> 00:06:43.460 align:middle line:84% This is conceptually quite close to the style model 00:06:43.460 --> 00:06:45.385 align:middle line:84% that you had seen in Glenn's lecture. 00:06:45.385 --> 00:06:48.660 align:middle line:90% 00:06:48.660 --> 00:06:55.100 align:middle line:84% This data only-- this paper only uses data on prices. 00:06:55.100 --> 00:06:57.500 align:middle line:90% No quantity data. 00:06:57.500 --> 00:07:04.100 align:middle line:84% And so, again, the idea here is we take these mixed strategy 00:07:04.100 --> 00:07:06.770 align:middle line:84% pricing equilibria very, very seriously 00:07:06.770 --> 00:07:09.030 align:middle line:84% and see how far this gets us empirically. 00:07:09.030 --> 00:07:10.940 align:middle line:84% And at the end, we can discuss a little bit 00:07:10.940 --> 00:07:15.235 align:middle line:84% whether this is a sensible way to go about this. 00:07:15.235 --> 00:07:17.758 align:middle line:90% 00:07:17.758 --> 00:07:19.800 align:middle line:84% The other interesting thing that this paper does, 00:07:19.800 --> 00:07:22.440 align:middle line:84% it looks at basically the same problem 00:07:22.440 --> 00:07:24.660 align:middle line:90% with two different models. 00:07:24.660 --> 00:07:27.580 align:middle line:84% A sequential and a non-sequential search model. 00:07:27.580 --> 00:07:30.420 align:middle line:84% And it turns out that empirically, it 00:07:30.420 --> 00:07:32.170 align:middle line:90% matters which one you use. 00:07:32.170 --> 00:07:33.750 align:middle line:84% And I think that's a good exercise 00:07:33.750 --> 00:07:37.950 align:middle line:84% to go through, to think a little bit about you're conceptually 00:07:37.950 --> 00:07:39.070 align:middle line:90% interested in something. 00:07:39.070 --> 00:07:44.550 align:middle line:84% There are different kinds of models, slight variations that 00:07:44.550 --> 00:07:48.270 align:middle line:84% might allow you to in some cases recover primitives 00:07:48.270 --> 00:07:49.810 align:middle line:90% better than in other cases. 00:07:49.810 --> 00:07:52.260 align:middle line:84% And so I think in that, the paper 00:07:52.260 --> 00:07:55.970 align:middle line:84% is quite nice that it helps you sort through these things. 00:07:55.970 --> 00:08:00.010 align:middle line:90% 00:08:00.010 --> 00:08:08.860 align:middle line:84% So in this setting, consumers are taking non-sequent-- so 00:08:08.860 --> 00:08:11.020 align:middle line:84% in the first version of the paper, 00:08:11.020 --> 00:08:15.710 align:middle line:84% non-sequential draws from the price distribution. 00:08:15.710 --> 00:08:22.260 align:middle line:84% So they're basically committing ex ante to l number of draws. 00:08:22.260 --> 00:08:28.080 align:middle line:84% And the way they do this is by minimizing their total cost, 00:08:28.080 --> 00:08:29.890 align:middle line:90% which is their search cost. 00:08:29.890 --> 00:08:33.000 align:middle line:84% Each consumer has an idiosyncratic cost 00:08:33.000 --> 00:08:36.840 align:middle line:84% for searching that's constant across searches. 00:08:36.840 --> 00:08:42.030 align:middle line:84% And then there's an expected price that results from taking l 00:08:42.030 --> 00:08:44.440 align:middle line:84% draws from the price distribution. 00:08:44.440 --> 00:08:48.440 align:middle line:84% And one of the draws is for free. 00:08:48.440 --> 00:08:52.280 align:middle line:84% So one thing I should have mentioned 00:08:52.280 --> 00:08:54.770 align:middle line:84% is that consumers in this market, 00:08:54.770 --> 00:08:58.640 align:middle line:84% they have a unit inelastic demand. 00:08:58.640 --> 00:09:01.535 align:middle line:84% That means they purchase one product for sure, but not more 00:09:01.535 --> 00:09:02.035 align:middle line:90% than that. 00:09:02.035 --> 00:09:10.070 align:middle line:90% 00:09:10.070 --> 00:09:12.250 align:middle line:84% And so this here is the price distribution 00:09:12.250 --> 00:09:16.130 align:middle line:84% that comes out of the mixed strategy pricing equilibrium. 00:09:16.130 --> 00:09:24.560 align:middle line:84% So again, this expression here is just the expected lowest 00:09:24.560 --> 00:09:28.070 align:middle line:90% price out of taking l draws. 00:09:28.070 --> 00:09:32.770 align:middle line:84% Any questions about what consumers are doing here? 00:09:32.770 --> 00:09:36.240 align:middle line:84% OK, so how can we think about this? 00:09:36.240 --> 00:09:41.530 align:middle line:84% So consumers are comparing the marginal cost of searching, 00:09:41.530 --> 00:09:47.690 align:middle line:84% which is ci, to the marginal benefit of searching. 00:09:47.690 --> 00:09:49.670 align:middle line:84% What is the marginal benefit of searching? 00:09:49.670 --> 00:09:54.860 align:middle line:84% Well, if I think about going from i to i plus 1 draws, 00:09:54.860 --> 00:09:59.510 align:middle line:84% it's how much lower the price is if I take an extra draw. 00:09:59.510 --> 00:10:01.580 align:middle line:84% So the marginal benefit of searching 00:10:01.580 --> 00:10:04.400 align:middle line:84% is the difference in expected prices from taking another draw 00:10:04.400 --> 00:10:05.850 align:middle line:90% from the price distribution. 00:10:05.850 --> 00:10:09.374 align:middle line:90% 00:10:09.374 --> 00:10:12.790 align:middle line:84% And so they're comparing marginal cost 00:10:12.790 --> 00:10:14.193 align:middle line:90% and marginal benefit. 00:10:14.193 --> 00:10:15.610 align:middle line:84% As long as the marginal benefit is 00:10:15.610 --> 00:10:17.735 align:middle line:84% higher than the marginal cost, the price reduction, 00:10:17.735 --> 00:10:20.145 align:middle line:84% they're going to increase the number of draws by 1. 00:10:20.145 --> 00:10:23.210 align:middle line:90% 00:10:23.210 --> 00:10:26.230 align:middle line:84% It's going to be helpful to keep track of these types here, 00:10:26.230 --> 00:10:29.720 align:middle line:84% the delta i's, because these are consumers. 00:10:29.720 --> 00:10:32.330 align:middle line:84% Consumers, again, have heterogeneous search costs 00:10:32.330 --> 00:10:37.130 align:middle line:84% who are just indifferent between drawing i and i plus 1 times 00:10:37.130 --> 00:10:39.320 align:middle line:90% from the price distribution. 00:10:39.320 --> 00:10:42.840 align:middle line:84% So we get this kind of sequence of marginal benefits. 00:10:42.840 --> 00:10:44.630 align:middle line:84% Now, with this in mind, there's actually 00:10:44.630 --> 00:10:50.270 align:middle line:84% a really nice way of displaying what consumers 00:10:50.270 --> 00:10:51.840 align:middle line:90% are doing in this market. 00:10:51.840 --> 00:10:56.540 align:middle line:90% So this here shows the CDF of-- 00:10:56.540 --> 00:10:59.900 align:middle line:84% sorry, the density of the search cost distribution. 00:10:59.900 --> 00:11:03.380 align:middle line:84% And consumers are ordered in terms of how many draws 00:11:03.380 --> 00:11:05.870 align:middle line:84% they're taking from the price distribution. 00:11:05.870 --> 00:11:08.420 align:middle line:84% Consumers with a very high search cost, 00:11:08.420 --> 00:11:13.260 align:middle line:84% with a very high search cost, they only take their free draw. 00:11:13.260 --> 00:11:16.313 align:middle line:84% And as you lower search costs that take more and more 00:11:16.313 --> 00:11:17.730 align:middle line:84% draws from the price distribution. 00:11:17.730 --> 00:11:20.370 align:middle line:84% And again, these are the difference types 00:11:20.370 --> 00:11:24.598 align:middle line:84% that segment these different bins of the distribution 00:11:24.598 --> 00:11:25.265 align:middle line:90% of search costs. 00:11:25.265 --> 00:11:29.450 align:middle line:90% 00:11:29.450 --> 00:11:35.180 align:middle line:84% So we get these kind of mass points of consumers that take 00:11:35.180 --> 00:11:39.970 align:middle line:90% i draws from the distribution. 00:11:39.970 --> 00:11:44.170 align:middle line:90% And we call those q tilde. 00:11:44.170 --> 00:11:46.090 align:middle line:90% Any questions on this? 00:11:46.090 --> 00:11:48.788 align:middle line:84% In this direction, search costs are increasing. 00:11:48.788 --> 00:11:50.830 align:middle line:84% With lower search costs, you're taking more draws 00:11:50.830 --> 00:11:54.280 align:middle line:90% from the price distribution. 00:11:54.280 --> 00:11:58.000 align:middle line:84% OK, so let's now think about what firms are doing. 00:11:58.000 --> 00:12:03.160 align:middle line:84% So this is, again, based on a model by Burdett and Judd 00:12:03.160 --> 00:12:06.460 align:middle line:90% from 1983. 00:12:06.460 --> 00:12:10.200 align:middle line:90% Firms are posting prices here. 00:12:10.200 --> 00:12:17.310 align:middle line:84% And they're doing so maximizing the margin times the probability 00:12:17.310 --> 00:12:18.695 align:middle line:90% of selling to a consumer. 00:12:18.695 --> 00:12:22.300 align:middle line:90% 00:12:22.300 --> 00:12:25.030 align:middle line:84% These q tildes, remember, were how many 00:12:25.030 --> 00:12:28.010 align:middle line:84% consumers are taking k draws from the price distribution. 00:12:28.010 --> 00:12:30.590 align:middle line:90% 00:12:30.590 --> 00:12:35.630 align:middle line:84% So the q tilde k of them, they take k draws. 00:12:35.630 --> 00:12:39.595 align:middle line:84% And I'm selling to them if I'm offering them the lowest price. 00:12:39.595 --> 00:12:44.400 align:middle line:90% 00:12:44.400 --> 00:12:47.850 align:middle line:84% And so, again, you have to think of this here as the price 00:12:47.850 --> 00:12:51.300 align:middle line:84% distribution that comes out of the mixed strategy pricing 00:12:51.300 --> 00:12:52.090 align:middle line:90% equilibrium. 00:12:52.090 --> 00:12:55.590 align:middle line:84% This is a smooth, atomless distribution 00:12:55.590 --> 00:12:58.505 align:middle line:84% has some support points that will be important in a second. 00:12:58.505 --> 00:13:02.252 align:middle line:90% 00:13:02.252 --> 00:13:09.300 align:middle line:84% So how can we get some traction on this empirically? 00:13:09.300 --> 00:13:13.200 align:middle line:84% The restriction that we're going to use here 00:13:13.200 --> 00:13:18.690 align:middle line:84% is the fact that firms have to be indifferent between offering 00:13:18.690 --> 00:13:20.850 align:middle line:84% all the prices that are in the support of the price 00:13:20.850 --> 00:13:24.120 align:middle line:90% distribution. 00:13:24.120 --> 00:13:28.200 align:middle line:84% As is the standard condition in a mixed strategy pricing 00:13:28.200 --> 00:13:28.950 align:middle line:90% equilibrium. 00:13:28.950 --> 00:13:31.170 align:middle line:84% Everything that's in the support, 00:13:31.170 --> 00:13:35.640 align:middle line:84% you have to be indifferent between those strategies. 00:13:35.640 --> 00:13:39.060 align:middle line:84% And so in particular, firms have to be 00:13:39.060 --> 00:13:42.780 align:middle line:84% indifferent between offering any price p 00:13:42.780 --> 00:13:46.320 align:middle line:84% and offering the highest price and only 00:13:46.320 --> 00:13:48.640 align:middle line:84% sell to consumers that are not searching. 00:13:48.640 --> 00:13:52.640 align:middle line:90% 00:13:52.640 --> 00:13:54.460 align:middle line:84% So now let's stare at this for a second. 00:13:54.460 --> 00:13:57.110 align:middle line:90% 00:13:57.110 --> 00:13:59.270 align:middle line:84% So what are the things that we do know 00:13:59.270 --> 00:14:02.950 align:middle line:84% and what are the things that we do not know in this equation? 00:14:02.950 --> 00:14:06.030 align:middle line:90% We do know the prices. 00:14:06.030 --> 00:14:09.020 align:middle line:84% Again, as I told you earlier, they're 00:14:09.020 --> 00:14:12.140 align:middle line:84% starting from the assumption that the researcher observes 00:14:12.140 --> 00:14:13.585 align:middle line:90% prices in the market. 00:14:13.585 --> 00:14:16.370 align:middle line:90% 00:14:16.370 --> 00:14:19.520 align:middle line:84% But we do not know how many consumers 00:14:19.520 --> 00:14:23.060 align:middle line:84% are searching how many times, as well as what 00:14:23.060 --> 00:14:25.375 align:middle line:90% the marginal cost is of firms. 00:14:25.375 --> 00:14:32.596 align:middle line:90% 00:14:32.596 --> 00:14:38.270 align:middle line:90% So this gives us conditions. 00:14:38.270 --> 00:14:40.610 align:middle line:84% Linear equations, in fact, that allow us 00:14:40.610 --> 00:14:44.360 align:middle line:90% to figure out what q and r are. 00:14:44.360 --> 00:14:46.340 align:middle line:84% Again, because we observe prices, 00:14:46.340 --> 00:14:49.300 align:middle line:84% we can construct this CDF of prices here. 00:14:49.300 --> 00:14:56.830 align:middle line:90% 00:14:56.830 --> 00:15:01.690 align:middle line:84% So as long as we observe more prices 00:15:01.690 --> 00:15:06.010 align:middle line:84% than the maximum number of draws that consumers 00:15:06.010 --> 00:15:10.030 align:middle line:84% can take in this market, or we have 00:15:10.030 --> 00:15:14.650 align:middle line:84% to observe more prices than k plus 1 of that, 00:15:14.650 --> 00:15:20.350 align:middle line:84% we can build moment conditions or equations 00:15:20.350 --> 00:15:25.510 align:middle line:84% from which we simply back out these unknown q's and r. 00:15:25.510 --> 00:15:28.000 align:middle line:84% So this just amounts to basically counting 00:15:28.000 --> 00:15:32.170 align:middle line:84% the number of equations relative to the number of unknowns. 00:15:32.170 --> 00:15:34.640 align:middle line:84% So these are sets of linear equations. 00:15:34.640 --> 00:15:37.120 align:middle line:84% The way this actually works in the paper 00:15:37.120 --> 00:15:43.030 align:middle line:84% is that you typically are overidentified because n minus 1 00:15:43.030 --> 00:15:46.000 align:middle line:84% is much larger than k, and so you use, let's say, 00:15:46.000 --> 00:15:48.010 align:middle line:84% a moment based method or, in fact, in the paper, 00:15:48.010 --> 00:15:54.220 align:middle line:84% they use empirical likelihood to back out these unknowns. 00:15:54.220 --> 00:15:55.770 align:middle line:90% Again, so what are the unknowns? 00:15:55.770 --> 00:16:00.360 align:middle line:84% I have k minus 1 of these unknown q tildes. 00:16:00.360 --> 00:16:02.630 align:middle line:84% Remember, this was this partition of the search cost 00:16:02.630 --> 00:16:04.490 align:middle line:90% distribution. 00:16:04.490 --> 00:16:08.600 align:middle line:84% And you have the unobserved marginal cost 00:16:08.600 --> 00:16:10.640 align:middle line:84% of firms, which is r, and assumed 00:16:10.640 --> 00:16:11.935 align:middle line:90% to be constant across firms. 00:16:11.935 --> 00:16:18.260 align:middle line:90% 00:16:18.260 --> 00:16:20.870 align:middle line:84% STUDENT: So just real quick, a place where Hong and Shum are 00:16:20.870 --> 00:16:22.400 align:middle line:84% cheating a little bit is they've got 00:16:22.400 --> 00:16:27.560 align:middle line:84% that f hat of pi, the CDF evaluated at pi, 00:16:27.560 --> 00:16:30.080 align:middle line:84% and they're setting that CDF evaluated 00:16:30.080 --> 00:16:33.090 align:middle line:84% pi in the empirical CDF evaluated at pi. 00:16:33.090 --> 00:16:34.490 align:middle line:90% So it's as if-- 00:16:34.490 --> 00:16:37.470 align:middle line:84% it's not the prices are drawn from a distribution. 00:16:37.470 --> 00:16:40.160 align:middle line:84% They're assuming that is the true CDF 00:16:40.160 --> 00:16:41.810 align:middle line:90% at every observed price. 00:16:41.810 --> 00:16:44.390 align:middle line:84% The number of prices is less than the pi. 00:16:44.390 --> 00:16:46.940 align:middle line:84% And that's where their equations come from. 00:16:46.940 --> 00:16:49.190 align:middle line:84% Because at every observed price, they 00:16:49.190 --> 00:16:53.512 align:middle line:90% assume they know f [INAUDIBLE]. 00:16:53.512 --> 00:16:55.220 align:middle line:84% TOBIAS SALZ: Yeah, I'll come back to this 00:16:55.220 --> 00:16:59.030 align:middle line:84% because if you, if you actually look at the price distributions, 00:16:59.030 --> 00:17:03.000 align:middle line:90% they look a little bit funny. 00:17:03.000 --> 00:17:03.960 align:middle line:90% They have mass points. 00:17:03.960 --> 00:17:10.339 align:middle line:84% And so it doesn't look super consistent with an atomless kind 00:17:10.339 --> 00:17:11.970 align:middle line:90% of smooth price distribution. 00:17:11.970 --> 00:17:20.220 align:middle line:90% 00:17:20.220 --> 00:17:27.089 align:middle line:84% So the bottom line here is that as long 00:17:27.089 --> 00:17:30.090 align:middle line:84% as we have more observed prices than k, 00:17:30.090 --> 00:17:32.650 align:middle line:84% and they haven't really told us where k is coming from. 00:17:32.650 --> 00:17:35.140 align:middle line:84% I'm going to talk about this later on. 00:17:35.140 --> 00:17:38.700 align:middle line:84% But as long as n minus 1 is larger than the maximum number 00:17:38.700 --> 00:17:40.380 align:middle line:84% of draws that consumers can take, 00:17:40.380 --> 00:17:45.540 align:middle line:84% you can use these equations here and simply solve 00:17:45.540 --> 00:17:47.680 align:middle line:84% a system of equations for these unknowns. 00:17:47.680 --> 00:17:51.330 align:middle line:84% Then again, practically, this is going to boil down 00:17:51.330 --> 00:17:52.540 align:middle line:90% to a moment condition. 00:17:52.540 --> 00:17:56.860 align:middle line:90% 00:17:56.860 --> 00:18:02.530 align:middle line:84% OK, so let's now move to the sequential search model. 00:18:02.530 --> 00:18:09.460 align:middle line:84% The sequential search model is using a solution 00:18:09.460 --> 00:18:12.920 align:middle line:84% that you might have seen from, let's say, the McCall model. 00:18:12.920 --> 00:18:16.450 align:middle line:84% I'm assuming you have seen the labor search McCall 00:18:16.450 --> 00:18:18.110 align:middle line:90% model in macro or somewhere. 00:18:18.110 --> 00:18:20.320 align:middle line:84% But let me remind you how this works. 00:18:20.320 --> 00:18:24.100 align:middle line:84% So instead of committing ex ante to a certain number of draws 00:18:24.100 --> 00:18:29.440 align:middle line:84% from the price distribution, you're going to draw a price 00:18:29.440 --> 00:18:32.470 align:middle line:84% and decide whether or not to draw again. 00:18:32.470 --> 00:18:36.140 align:middle line:84% And how do you make this decision? 00:18:36.140 --> 00:18:40.420 align:middle line:84% Well, the typical solution to this kind of problem 00:18:40.420 --> 00:18:42.670 align:middle line:90% is a cut off rule. 00:18:42.670 --> 00:18:44.117 align:middle line:90% So there's a reservation price. 00:18:44.117 --> 00:18:46.450 align:middle line:84% As long as the prices that you're drawing from the price 00:18:46.450 --> 00:18:49.120 align:middle line:84% distribution are higher than the reservation price, 00:18:49.120 --> 00:18:50.870 align:middle line:90% you keep on searching. 00:18:50.870 --> 00:18:56.730 align:middle line:84% But once it falls below that, you stop. 00:18:56.730 --> 00:18:59.850 align:middle line:84% And so now the question is, how do you actually 00:18:59.850 --> 00:19:04.560 align:middle line:90% set this reservation value? 00:19:04.560 --> 00:19:08.298 align:middle line:84% And so here, again, the same idea as before. 00:19:08.298 --> 00:19:10.840 align:middle line:84% You're thinking about what is the marginal cost of searching. 00:19:10.840 --> 00:19:14.310 align:middle line:84% Again, this is constant and equal to ci. 00:19:14.310 --> 00:19:18.780 align:middle line:84% And what is the marginal benefit of searching? 00:19:18.780 --> 00:19:24.240 align:middle line:84% Which is basically the weighted average of price improvements 00:19:24.240 --> 00:19:28.150 align:middle line:84% that I could get where z is my reservation value. 00:19:28.150 --> 00:19:33.330 align:middle line:84% Again, this is implicitly a function of my search cost. 00:19:33.330 --> 00:19:37.690 align:middle line:84% And so you get this weighted average of price improvements. 00:19:37.690 --> 00:19:40.320 align:middle line:84% This here is the marginal benefit of searching 00:19:40.320 --> 00:19:44.300 align:middle line:90% and ci is the marginal cost. 00:19:44.300 --> 00:19:48.580 align:middle line:84% So z is set such that these two are equal to each other. 00:19:48.580 --> 00:19:51.440 align:middle line:90% 00:19:51.440 --> 00:19:58.010 align:middle line:84% Does everybody-- may I ask, have you seen the McCall model? 00:19:58.010 --> 00:20:01.420 align:middle line:90% 00:20:01.420 --> 00:20:02.535 align:middle line:90% So some of you. 00:20:02.535 --> 00:20:06.450 align:middle line:90% 00:20:06.450 --> 00:20:17.310 align:middle line:84% So with this solution, it turns out 00:20:17.310 --> 00:20:23.370 align:middle line:84% that empirically, you can make much less progress, and in fact, 00:20:23.370 --> 00:20:25.990 align:middle line:84% not rely on the same argument as we did before. 00:20:25.990 --> 00:20:29.350 align:middle line:90% So why is that the case? 00:20:29.350 --> 00:20:33.940 align:middle line:84% So first, let's think about the reservation prices. 00:20:33.940 --> 00:20:35.700 align:middle line:84% So one thing that's different here 00:20:35.700 --> 00:20:37.800 align:middle line:84% is that beforehand, we had this sort 00:20:37.800 --> 00:20:41.130 align:middle line:84% of discrete number of things that consumers could do. 00:20:41.130 --> 00:20:45.630 align:middle line:84% So they were taking a continuous draw from the search cost 00:20:45.630 --> 00:20:46.300 align:middle line:90% distribution. 00:20:46.300 --> 00:20:50.310 align:middle line:84% But still, in the end, there are only finitely many things 00:20:50.310 --> 00:20:51.420 align:middle line:90% that they could do. 00:20:51.420 --> 00:20:55.620 align:middle line:84% So you get these kind of bins of the search cost distribution. 00:20:55.620 --> 00:20:58.800 align:middle line:84% And we were identifying these bins. 00:20:58.800 --> 00:21:05.100 align:middle line:84% Here what is different is that the optimal strategy 00:21:05.100 --> 00:21:09.250 align:middle line:84% is a function of ci and every unique search cost 00:21:09.250 --> 00:21:14.062 align:middle line:84% type has a unique reservation value strategy. 00:21:14.062 --> 00:21:16.980 align:middle line:84% Except for those that potentially bunch at the top. 00:21:16.980 --> 00:21:19.740 align:middle line:84% So if you have a higher reservation 00:21:19.740 --> 00:21:21.960 align:middle line:84% value, then the highest possible price 00:21:21.960 --> 00:21:24.540 align:middle line:84% in the support of the price distribution, then of course, 00:21:24.540 --> 00:21:28.220 align:middle line:84% you stop searching when you reach-- when you drop p bar. 00:21:28.220 --> 00:21:31.460 align:middle line:90% 00:21:31.460 --> 00:21:34.340 align:middle line:84% So now let's think about the demand of firms 00:21:34.340 --> 00:21:36.253 align:middle line:90% with this kind of strategy. 00:21:36.253 --> 00:21:39.414 align:middle line:90% 00:21:39.414 --> 00:21:44.010 align:middle line:84% So as before, we can construct indifference conditions. 00:21:44.010 --> 00:21:46.740 align:middle line:90% So what is the demand of a firm? 00:21:46.740 --> 00:21:48.990 align:middle line:84% I'm going to sell only to consumers 00:21:48.990 --> 00:21:51.527 align:middle line:84% that have a reservation value higher than the price p 00:21:51.527 --> 00:21:52.235 align:middle line:90% that I'm setting. 00:21:52.235 --> 00:21:57.474 align:middle line:90% 00:21:57.474 --> 00:22:00.020 align:middle line:84% So the demand is proportional to the mass 00:22:00.020 --> 00:22:07.350 align:middle line:84% of consumers above the price p that I include, that I charge. 00:22:07.350 --> 00:22:11.270 align:middle line:84% So now again, we can try and construct 00:22:11.270 --> 00:22:17.970 align:middle line:84% these indifference conditions for any given price pi. 00:22:17.970 --> 00:22:21.500 align:middle line:84% Firms need to be indifferent between charging that price, 00:22:21.500 --> 00:22:24.590 align:middle line:84% reaching consumers only above that have a reservation 00:22:24.590 --> 00:22:31.670 align:middle line:84% value above that price and the maximum possible price 00:22:31.670 --> 00:22:34.020 align:middle line:84% in the support of the price distribution. 00:22:34.020 --> 00:22:36.820 align:middle line:90% 00:22:36.820 --> 00:22:39.510 align:middle line:84% Can you see what the issue here is? 00:22:39.510 --> 00:22:42.140 align:middle line:90% 00:22:42.140 --> 00:22:46.940 align:middle line:84% So the problem is we have a incidental parameter problem. 00:22:46.940 --> 00:22:53.850 align:middle line:84% For any nu of those indifference conditions that are introduced, 00:22:53.850 --> 00:22:56.600 align:middle line:84% I'm also introducing a nu unknown, 00:22:56.600 --> 00:22:58.915 align:middle line:84% which is the mass of consumers above that point. 00:22:58.915 --> 00:23:01.934 align:middle line:90% 00:23:01.934 --> 00:23:09.080 align:middle line:84% So in other words, this simple moment counting versus unknowns 00:23:09.080 --> 00:23:13.550 align:middle line:84% no longer works because I have n plus 1 unknowns. 00:23:13.550 --> 00:23:17.630 align:middle line:84% The mass of consumers below pi and also 00:23:17.630 --> 00:23:19.910 align:middle line:90% the unknown cost of the firm. 00:23:19.910 --> 00:23:23.630 align:middle line:84% But I have n minus 1 indifference conditions. 00:23:23.630 --> 00:23:29.540 align:middle line:84% So basically we can no longer use this argument. 00:23:29.540 --> 00:23:32.630 align:middle line:84% So in the paper, they say to non-parametrically identify 00:23:32.630 --> 00:23:35.840 align:middle line:84% the search cost distribution, in the previous case, 00:23:35.840 --> 00:23:38.240 align:middle line:84% we also didn't quite non-parametrically identify 00:23:38.240 --> 00:23:40.640 align:middle line:84% the search cost distribution because we only got 00:23:40.640 --> 00:23:46.160 align:middle line:90% this partition into these bins. 00:23:46.160 --> 00:23:49.820 align:middle line:84% But here we basically with this strategy 00:23:49.820 --> 00:23:52.040 align:middle line:84% can learn anything about the search cost distribution 00:23:52.040 --> 00:23:56.270 align:middle line:84% unless we're willing to impose some parametric form. 00:23:56.270 --> 00:23:58.470 align:middle line:84% And so that's what they're going to do in the end. 00:23:58.470 --> 00:24:02.520 align:middle line:84% They basically assume that the search cost distribution 00:24:02.520 --> 00:24:05.740 align:middle line:84% takes on a specific parametric distribution 00:24:05.740 --> 00:24:08.555 align:middle line:84% and then simply estimate it based off of that. 00:24:08.555 --> 00:24:14.280 align:middle line:90% 00:24:14.280 --> 00:24:17.452 align:middle line:84% So this comes back to Glenn's point. 00:24:17.452 --> 00:24:18.660 align:middle line:90% What is the application here? 00:24:18.660 --> 00:24:23.820 align:middle line:84% The application is perhaps not the most inspiring, but maybe 00:24:23.820 --> 00:24:29.580 align:middle line:84% in this crowd, which is the different sets of econ graduate 00:24:29.580 --> 00:24:30.520 align:middle line:90% textbooks. 00:24:30.520 --> 00:24:34.230 align:middle line:84% So Stokey-Lucas, Lazear, Billingsley-- 00:24:34.230 --> 00:24:35.500 align:middle line:90% Probability and Measure. 00:24:35.500 --> 00:24:38.400 align:middle line:84% So if you stare at these price distributions, 00:24:38.400 --> 00:24:41.910 align:middle line:84% you can ask yourself whether you think it's plausible 00:24:41.910 --> 00:24:45.390 align:middle line:84% that these are coming out of this mixed strategy pricing 00:24:45.390 --> 00:24:46.090 align:middle line:90% equilibrium. 00:24:46.090 --> 00:24:49.020 align:middle line:84% So one thing that I, for example, find somewhat curious 00:24:49.020 --> 00:24:53.250 align:middle line:84% is that you get these very large mass points here. 00:24:53.250 --> 00:24:59.800 align:middle line:84% And remember from Stahl but also from Burdett and Judd, 00:24:59.800 --> 00:25:03.215 align:middle line:84% we would expect an atomless smooth distribution of prices. 00:25:03.215 --> 00:25:10.510 align:middle line:90% 00:25:10.510 --> 00:25:12.123 align:middle line:84% Let me show you some of the results. 00:25:12.123 --> 00:25:14.750 align:middle line:90% 00:25:14.750 --> 00:25:18.930 align:middle line:84% So first, they have to figure out what k is. 00:25:18.930 --> 00:25:20.870 align:middle line:84% Remember, k was the number of draws 00:25:20.870 --> 00:25:23.360 align:middle line:84% from the maximum number of draws that consumers take 00:25:23.360 --> 00:25:24.980 align:middle line:90% from the price distribution. 00:25:24.980 --> 00:25:27.810 align:middle line:84% The way they do this is somewhat hacky. 00:25:27.810 --> 00:25:31.670 align:middle line:84% They basically just look at what is the highest 00:25:31.670 --> 00:25:34.730 align:middle line:90% k for which my model converges. 00:25:34.730 --> 00:25:38.150 align:middle line:90% So estimate the model with that. 00:25:38.150 --> 00:25:40.400 align:middle line:84% And so one striking thing that they find 00:25:40.400 --> 00:25:47.000 align:middle line:84% is that for most of these different markets, 00:25:47.000 --> 00:25:50.150 align:middle line:84% the average consumer only searches once. 00:25:50.150 --> 00:25:52.940 align:middle line:84% So about 50% of consumers only take one draw 00:25:52.940 --> 00:25:55.190 align:middle line:90% from the price distribution. 00:25:55.190 --> 00:25:59.870 align:middle line:84% And in fact, only for one of the textbooks, Lazear, 00:25:59.870 --> 00:26:04.490 align:middle line:84% you actually have consumers that take three draws from the price 00:26:04.490 --> 00:26:07.310 align:middle line:90% distribution. 00:26:07.310 --> 00:26:14.530 align:middle line:84% So if you think about this here, one potential issue 00:26:14.530 --> 00:26:18.070 align:middle line:84% with having 50% of consumers only 00:26:18.070 --> 00:26:22.030 align:middle line:84% searching once, which remember you get these cut off 00:26:22.030 --> 00:26:23.650 align:middle line:90% points for consumers. 00:26:23.650 --> 00:26:26.440 align:middle line:84% Let me just go back to this briefly 00:26:26.440 --> 00:26:28.270 align:middle line:90% to illustrate this point. 00:26:28.270 --> 00:26:35.120 align:middle line:84% So what I'm identifying here are these partitions. 00:26:35.120 --> 00:26:39.110 align:middle line:84% So I'm getting out of the model a partition of the search cost 00:26:39.110 --> 00:26:40.180 align:middle line:90% distribution. 00:26:40.180 --> 00:26:44.810 align:middle line:84% Now, if let's say delta 1 lies very, very low 00:26:44.810 --> 00:26:48.440 align:middle line:84% and most consumers only search once, 00:26:48.440 --> 00:26:51.890 align:middle line:84% then I don't actually learn all that much 00:26:51.890 --> 00:26:53.870 align:middle line:84% about the price distribution above that point-- 00:26:53.870 --> 00:26:56.037 align:middle line:84% about the search cost distribution above that point. 00:26:56.037 --> 00:27:00.380 align:middle line:90% 00:27:00.380 --> 00:27:05.840 align:middle line:84% So the problem is that what we're really recovering here 00:27:05.840 --> 00:27:12.080 align:middle line:84% is not a primitive but an endogenous object, which 00:27:12.080 --> 00:27:16.060 align:middle line:84% is the search strategy of consumers. 00:27:16.060 --> 00:27:18.970 align:middle line:84% And so we have to, in some sense, 00:27:18.970 --> 00:27:22.660 align:middle line:84% cross our fingers and hope that how consumers are searching 00:27:22.660 --> 00:27:25.870 align:middle line:84% in the market tells us enough about the search cost 00:27:25.870 --> 00:27:26.750 align:middle line:90% distribution. 00:27:26.750 --> 00:27:30.860 align:middle line:84% But it could just be that most consumers don't search at all. 00:27:30.860 --> 00:27:33.520 align:middle line:84% And then I don't really learn anything. 00:27:33.520 --> 00:27:37.600 align:middle line:84% So that's kind of one potential shortcoming of this approach 00:27:37.600 --> 00:27:41.890 align:middle line:84% is that this identification argument doesn't directly 00:27:41.890 --> 00:27:44.410 align:middle line:84% map to the distribution, but it maps 00:27:44.410 --> 00:27:47.410 align:middle line:84% to what consumers are doing in the market. 00:27:47.410 --> 00:27:51.460 align:middle line:84% If you think about this paper and let me comment a little bit 00:27:51.460 --> 00:27:58.480 align:middle line:84% on this, if you think about the exercise, just using price data 00:27:58.480 --> 00:28:03.490 align:middle line:90% is a somewhat heroic exercise. 00:28:03.490 --> 00:28:05.830 align:middle line:84% And you have to think about this as being this 00:28:05.830 --> 00:28:08.320 align:middle line:84% was one of the very first empirical papers on search 00:28:08.320 --> 00:28:11.680 align:middle line:84% or papers that use the equilibrium conditions that you 00:28:11.680 --> 00:28:14.620 align:middle line:84% get out of a search model to say something empirically. 00:28:14.620 --> 00:28:17.380 align:middle line:90% 00:28:17.380 --> 00:28:19.240 align:middle line:84% So one thing you may want to ask yourself 00:28:19.240 --> 00:28:23.500 align:middle line:84% is how plausible you find it that firms are really 00:28:23.500 --> 00:28:25.305 align:middle line:90% setting prices by mixing. 00:28:25.305 --> 00:28:29.773 align:middle line:90% 00:28:29.773 --> 00:28:31.690 align:middle line:84% And whether there are alternative explanations 00:28:31.690 --> 00:28:32.857 align:middle line:90% why we see different prices. 00:28:32.857 --> 00:28:36.670 align:middle line:90% 00:28:36.670 --> 00:28:39.410 align:middle line:84% This wouldn't be on the top of my list of complaints, 00:28:39.410 --> 00:28:43.730 align:middle line:84% but the marginal cost of firms are assumed to be constant here. 00:28:43.730 --> 00:28:46.570 align:middle line:84% So you could take some issue with that. 00:28:46.570 --> 00:28:49.870 align:middle line:84% And then I want to point out something 00:28:49.870 --> 00:28:54.430 align:middle line:84% that is more of a common critique of many papers 00:28:54.430 --> 00:28:56.680 align:middle line:90% in the search literature. 00:28:56.680 --> 00:29:02.230 align:middle line:84% So first, most papers find very large search costs. 00:29:02.230 --> 00:29:08.620 align:middle line:84% And one issue is that you look at the dispersion of prices 00:29:08.620 --> 00:29:11.170 align:middle line:84% and you say that all of this dispersion, all 00:29:11.170 --> 00:29:13.420 align:middle line:84% those differences across prices is really 00:29:13.420 --> 00:29:15.790 align:middle line:84% kind of structural variation in a sense 00:29:15.790 --> 00:29:20.680 align:middle line:84% that all of that variation has an economic interpretation. 00:29:20.680 --> 00:29:23.260 align:middle line:84% As opposed to things that happen with data, 00:29:23.260 --> 00:29:25.660 align:middle line:84% it could be that there's, let's say, measurement error 00:29:25.660 --> 00:29:30.800 align:middle line:84% in prices or other reasons why the prices that you observe 00:29:30.800 --> 00:29:32.520 align:middle line:84% are not exactly equal to each other. 00:29:32.520 --> 00:29:37.460 align:middle line:84% And so in these models, you load typically 00:29:37.460 --> 00:29:40.880 align:middle line:84% almost the entire variation of prices 00:29:40.880 --> 00:29:44.660 align:middle line:84% onto this search cost distribution. 00:29:44.660 --> 00:29:50.090 align:middle line:84% And so that's part of the reason why in many of these papers, 00:29:50.090 --> 00:29:53.180 align:middle line:84% you find somewhat excessively large search costs. 00:29:53.180 --> 00:29:57.352 align:middle line:84% And so I think that's something that this literature has 00:29:57.352 --> 00:29:58.560 align:middle line:90% to grapple with a little bit. 00:29:58.560 --> 00:30:01.190 align:middle line:84% But it's actually common to other literatures that 00:30:01.190 --> 00:30:03.710 align:middle line:84% interpret all price variation as structural, 00:30:03.710 --> 00:30:06.640 align:middle line:84% which is, for example, also true in auctions. 00:30:06.640 --> 00:30:09.610 align:middle line:84% So the other thing that you could potentially 00:30:09.610 --> 00:30:13.600 align:middle line:84% complain about here is that these sellers of textbooks, they 00:30:13.600 --> 00:30:16.090 align:middle line:84% might be differentiated in other ways. 00:30:16.090 --> 00:30:17.680 align:middle line:84% Some of them might have funny names 00:30:17.680 --> 00:30:21.160 align:middle line:84% or located in places that people, I don't know, 00:30:21.160 --> 00:30:22.160 align:middle line:90% don't like that much. 00:30:22.160 --> 00:30:25.090 align:middle line:84% And so for that reason, they have to charge different prices 00:30:25.090 --> 00:30:26.560 align:middle line:90% for the same kind of product. 00:30:26.560 --> 00:30:29.062 align:middle line:84% So although these are in principle very homogeneous 00:30:29.062 --> 00:30:30.520 align:middle line:84% products, you could think there are 00:30:30.520 --> 00:30:34.360 align:middle line:84% other unobserved differentiating factors that 00:30:34.360 --> 00:30:35.940 align:middle line:90% lead to differences in prices. 00:30:35.940 --> 00:30:40.350 align:middle line:90% 00:30:40.350 --> 00:30:44.350 align:middle line:84% So again, this is the way you have to think about this paper. 00:30:44.350 --> 00:30:46.380 align:middle line:84% This is, I would say, mostly kind 00:30:46.380 --> 00:30:48.220 align:middle line:84% of a conceptually interesting exercise. 00:30:48.220 --> 00:30:50.520 align:middle line:84% It teaches us something about using these equilibrium 00:30:50.520 --> 00:30:51.750 align:middle line:90% restrictions. 00:30:51.750 --> 00:30:58.860 align:middle line:84% It's perhaps not a paper that you would use nowadays 00:30:58.860 --> 00:31:01.230 align:middle line:84% to really think about such frictions, 00:31:01.230 --> 00:31:03.480 align:middle line:84% in part because of this very strong equilibrium 00:31:03.480 --> 00:31:07.140 align:middle line:84% restriction that firms are playing mixed strategies, which 00:31:07.140 --> 00:31:11.070 align:middle line:84% in some settings we know is sensible. 00:31:11.070 --> 00:31:16.050 align:middle line:84% Let's think about, for instance, sports and other places 00:31:16.050 --> 00:31:18.010 align:middle line:84% where we know that mixing makes sense. 00:31:18.010 --> 00:31:20.920 align:middle line:84% But here, you can do some introspection. 00:31:20.920 --> 00:31:22.420 align:middle line:84% I find it somewhat less plausible. 00:31:22.420 --> 00:31:25.680 align:middle line:90% 00:31:25.680 --> 00:31:28.500 align:middle line:84% So what I'm going to talk about next 00:31:28.500 --> 00:31:34.980 align:middle line:84% is a paper that is much more empirically motivated 00:31:34.980 --> 00:31:38.710 align:middle line:84% by Jason Allen, Rob Clarke, and Jean-Francois Houde. 00:31:38.710 --> 00:31:42.420 align:middle line:90% It came out in the AR in 2013. 00:31:42.420 --> 00:31:47.850 align:middle line:84% And it basically asks, what are the implications 00:31:47.850 --> 00:31:53.520 align:middle line:84% of a merger in a market where prices are negotiated? 00:31:53.520 --> 00:31:59.820 align:middle line:84% So just to remind you a little bit, so in IO, 00:31:59.820 --> 00:32:01.950 align:middle line:84% we have a very developed machinery 00:32:01.950 --> 00:32:05.910 align:middle line:84% for thinking about mergers in posted price markets. 00:32:05.910 --> 00:32:08.230 align:middle line:90% So we estimate demand. 00:32:08.230 --> 00:32:11.280 align:middle line:84% We have some ownership matrix with demand. 00:32:11.280 --> 00:32:15.600 align:middle line:84% And the ownership matrix we can compute first order conditions 00:32:15.600 --> 00:32:17.090 align:middle line:90% and can recover costs. 00:32:17.090 --> 00:32:18.840 align:middle line:84% And once we have costs, we can, let's say, 00:32:18.840 --> 00:32:24.900 align:middle line:84% change the ownership matrix and recompute prices. 00:32:24.900 --> 00:32:28.377 align:middle line:84% Up until very recently, there was not a lot of work 00:32:28.377 --> 00:32:30.960 align:middle line:84% that thinks about what are some of the implications of mergers 00:32:30.960 --> 00:32:33.377 align:middle line:84% in markets where different consumers pay different prices. 00:32:33.377 --> 00:32:38.300 align:middle line:90% 00:32:38.300 --> 00:32:43.590 align:middle line:84% So they're going to look at this in the Canadian mortgage market. 00:32:43.590 --> 00:32:47.820 align:middle line:84% They have some great data on insured mortgage contracts, 00:32:47.820 --> 00:32:52.440 align:middle line:84% which are collected directly from the insurer. 00:32:52.440 --> 00:32:55.440 align:middle line:84% And what is nice about the fact that these mortgages are insured 00:32:55.440 --> 00:33:01.050 align:middle line:84% is that they can claim that these contracts are very, very 00:33:01.050 --> 00:33:03.990 align:middle line:84% homogeneous, that consumer risk is not 00:33:03.990 --> 00:33:05.460 align:middle line:84% playing a large role in explaining 00:33:05.460 --> 00:33:06.990 align:middle line:90% price differentiation. 00:33:06.990 --> 00:33:11.010 align:middle line:84% So they claim that this is a very, very homogeneous product 00:33:11.010 --> 00:33:13.690 align:middle line:84% because idiosyncratic risk here is insured away. 00:33:13.690 --> 00:33:17.890 align:middle line:90% 00:33:17.890 --> 00:33:21.820 align:middle line:84% So part of why I like this paper is it 00:33:21.820 --> 00:33:26.110 align:middle line:84% provides a very nice combination of treatment effect analysis 00:33:26.110 --> 00:33:29.230 align:middle line:84% of an actual merger that happened in the data 00:33:29.230 --> 00:33:37.690 align:middle line:84% and interpreting this variation through a structural model. 00:33:37.690 --> 00:33:40.330 align:middle line:84% And so I'll try to make this clear today, 00:33:40.330 --> 00:33:42.190 align:middle line:84% but there's basically a very nice connection 00:33:42.190 --> 00:33:44.170 align:middle line:84% between the treatment effect analysis 00:33:44.170 --> 00:33:47.934 align:middle line:90% and the structural analysis. 00:33:47.934 --> 00:33:51.630 align:middle line:84% So there's a merger between two large lenders that 00:33:51.630 --> 00:33:55.800 align:middle line:84% happens in the data, which has implications 00:33:55.800 --> 00:33:57.730 align:middle line:90% for different markets. 00:33:57.730 --> 00:34:02.060 align:middle line:84% In some markets, you have branches from both lenders. 00:34:02.060 --> 00:34:05.560 align:middle line:84% In some other markets, none of the lenders have branches. 00:34:05.560 --> 00:34:08.205 align:middle line:84% And then that changes the choice set of consumers 00:34:08.205 --> 00:34:09.455 align:middle line:90% differentially across markets. 00:34:09.455 --> 00:34:14.540 align:middle line:90% 00:34:14.540 --> 00:34:19.159 align:middle line:84% So here's a picture of what happened in this market 00:34:19.159 --> 00:34:21.159 align:middle line:90% over time. 00:34:21.159 --> 00:34:26.000 align:middle line:84% So one thing that's pretty clear and visible 00:34:26.000 --> 00:34:29.120 align:middle line:84% is that over time, this market has become much, much 00:34:29.120 --> 00:34:31.400 align:middle line:90% more consolidated. 00:34:31.400 --> 00:34:36.330 align:middle line:84% So initially, there were banks and credit unions, trust shares. 00:34:36.330 --> 00:34:41.330 align:middle line:84% So these are basically lenders that are owned by their members. 00:34:41.330 --> 00:34:43.830 align:middle line:84% And then there is what they refer to as the big eight. 00:34:43.830 --> 00:34:49.760 align:middle line:84% These are the largest banks such as TD Bank, Nova Scotia, 00:34:49.760 --> 00:34:52.489 align:middle line:90% The Royal Bank and others. 00:34:52.489 --> 00:34:56.540 align:middle line:84% And these have been subsequently overtaking trusts. 00:34:56.540 --> 00:34:58.910 align:middle line:84% There were also some regulatory changes, 00:34:58.910 --> 00:35:02.815 align:middle line:84% and so the market has consolidated over time. 00:35:02.815 --> 00:35:08.770 align:middle line:90% 00:35:08.770 --> 00:35:13.840 align:middle line:84% So some institutional details that are going to be relevant. 00:35:13.840 --> 00:35:17.440 align:middle line:84% So consumers here have to go to a lender 00:35:17.440 --> 00:35:22.080 align:middle line:84% to figure out what the price of their mortgage is. 00:35:22.080 --> 00:35:25.420 align:middle line:84% So they have to talk to the local branch managers. 00:35:25.420 --> 00:35:28.060 align:middle line:84% And the local branches, they're posting prices, 00:35:28.060 --> 00:35:34.130 align:middle line:84% but the consumers can negotiate discounts from those prices. 00:35:34.130 --> 00:35:36.140 align:middle line:84% And they have explicit incentives 00:35:36.140 --> 00:35:41.000 align:middle line:84% to not give too many discounts, because their compensation 00:35:41.000 --> 00:35:43.220 align:middle line:84% is affected by the level of discounts. 00:35:43.220 --> 00:35:47.270 align:middle line:84% And as they claim in the paper, they also 00:35:47.270 --> 00:35:51.767 align:middle line:84% are explicitly told not to compete with other branches that 00:35:51.767 --> 00:35:52.975 align:middle line:90% are owned by the same lender. 00:35:52.975 --> 00:35:59.596 align:middle line:90% 00:35:59.596 --> 00:36:08.640 align:middle line:84% So because the data here has essentially administrative data 00:36:08.640 --> 00:36:10.620 align:middle line:84% that comes from these insurers, they 00:36:10.620 --> 00:36:14.340 align:middle line:84% see very, very rich information about both the consumers 00:36:14.340 --> 00:36:16.650 align:middle line:84% and the type of contract that they get. 00:36:16.650 --> 00:36:20.490 align:middle line:84% So they see, for instance, the consumer's home location, 00:36:20.490 --> 00:36:26.850 align:middle line:84% the loan size, house price, things like loan to value, 00:36:26.850 --> 00:36:30.392 align:middle line:84% and demographics such as income and even 00:36:30.392 --> 00:36:32.975 align:middle line:84% whether or not they had a prior relationship with this lender. 00:36:32.975 --> 00:36:36.932 align:middle line:90% 00:36:36.932 --> 00:36:38.390 align:middle line:84% Another nice thing that they can do 00:36:38.390 --> 00:36:40.432 align:middle line:84% is because they see the consumer's home location, 00:36:40.432 --> 00:36:46.320 align:middle line:84% they can actually define markets around consumer's home address. 00:36:46.320 --> 00:36:50.840 align:middle line:84% And that's in terms of market definition usually pretty good, 00:36:50.840 --> 00:36:53.480 align:middle line:84% because then you don't get these weird cases where somebody 00:36:53.480 --> 00:36:57.680 align:middle line:84% lives at the border of a market and would potentially 00:36:57.680 --> 00:37:00.590 align:middle line:84% shop from a market that's defined 00:37:00.590 --> 00:37:03.140 align:middle line:90% as a separate one in your data. 00:37:03.140 --> 00:37:05.788 align:middle line:84% And so because the Canadians do this in kilometers, 00:37:05.788 --> 00:37:06.955 align:middle line:90% so of course, most sensible. 00:37:06.955 --> 00:37:10.090 align:middle line:90% 00:37:10.090 --> 00:37:13.490 align:middle line:84% So what are the treatment and control groups here? 00:37:13.490 --> 00:37:15.970 align:middle line:84% So they're going to only focus on the short run 00:37:15.970 --> 00:37:17.650 align:middle line:90% effect of the merger. 00:37:17.650 --> 00:37:21.970 align:middle line:84% So that means that they're not going to look at entry and exit, 00:37:21.970 --> 00:37:25.300 align:middle line:84% but only at the immediate change in prices 00:37:25.300 --> 00:37:29.220 align:middle line:84% that the year after the merger happened. 00:37:29.220 --> 00:37:32.370 align:middle line:84% And so I'm only talking about the merger here because I cannot 00:37:32.370 --> 00:37:35.550 align:middle line:84% actually disclose who the lenders were that merged. 00:37:35.550 --> 00:37:39.830 align:middle line:84% So these were two larger lenders. 00:37:39.830 --> 00:37:45.260 align:middle line:84% And so what are what are the control markets? 00:37:45.260 --> 00:37:49.640 align:middle line:84% There are some markets in which neither of the lenders 00:37:49.640 --> 00:37:51.620 align:middle line:84% or only one of them were present. 00:37:51.620 --> 00:37:52.892 align:middle line:90% These are the control markets. 00:37:52.892 --> 00:37:54.350 align:middle line:84% And then there are markets in which 00:37:54.350 --> 00:37:59.000 align:middle line:84% both of the merging lenders had a branch. 00:37:59.000 --> 00:38:02.510 align:middle line:84% And in those markets, competition is now reduced. 00:38:02.510 --> 00:38:06.110 align:middle line:84% Because, again, according to this story 00:38:06.110 --> 00:38:08.240 align:middle line:84% that lenders are internalizing and not 00:38:08.240 --> 00:38:11.370 align:middle line:84% competing against branches in their own network. 00:38:11.370 --> 00:38:13.640 align:middle line:84% The choice set has now been reduced. 00:38:13.640 --> 00:38:16.580 align:middle line:84% So consumers cannot compare as many prices as before. 00:38:16.580 --> 00:38:21.210 align:middle line:84% Any questions about the institutional details here? 00:38:21.210 --> 00:38:23.200 align:middle line:90% So here's what I do. 00:38:23.200 --> 00:38:25.350 align:middle line:84% I first have a descriptive exercise 00:38:25.350 --> 00:38:30.510 align:middle line:84% where they look at the price variation 00:38:30.510 --> 00:38:32.083 align:middle line:90% that they observe in the data. 00:38:32.083 --> 00:38:34.500 align:middle line:84% And again, what's important here is that they can credibly 00:38:34.500 --> 00:38:37.710 align:middle line:84% claim that most of that is not because of risk 00:38:37.710 --> 00:38:41.370 align:middle line:90% variation across consumers. 00:38:41.370 --> 00:38:47.550 align:middle line:84% So they first run a regression of the margin 00:38:47.550 --> 00:38:51.330 align:middle line:84% of a consumer, that's the rate that they observe 00:38:51.330 --> 00:38:56.500 align:middle line:84% in the data minus the bond rate, on a set of control variables. 00:38:56.500 --> 00:38:59.340 align:middle line:84% So how long the mortgage was, how large 00:38:59.340 --> 00:39:01.090 align:middle line:90% the size of the mortgage was. 00:39:01.090 --> 00:39:03.300 align:middle line:84% So the idea here of this regression 00:39:03.300 --> 00:39:05.430 align:middle line:84% is that you really want to get at the unobserved 00:39:05.430 --> 00:39:07.270 align:middle line:90% component of the price. 00:39:07.270 --> 00:39:09.930 align:middle line:84% So there could still be differences across consumers, 00:39:09.930 --> 00:39:13.590 align:middle line:84% because again, one consumer gets a larger mortgage than another. 00:39:13.590 --> 00:39:16.090 align:middle line:84% Or there could be differences in FICO scores. 00:39:16.090 --> 00:39:18.990 align:middle line:84% So there could be some sort of perceived residual risk 00:39:18.990 --> 00:39:20.080 align:middle line:90% of a consumer. 00:39:20.080 --> 00:39:22.620 align:middle line:84% But they want to look at once we control 00:39:22.620 --> 00:39:24.960 align:middle line:84% for all these things, how much price variation is 00:39:24.960 --> 00:39:26.580 align:middle line:90% left after that? 00:39:26.580 --> 00:39:28.350 align:middle line:84% And so the key object of interest 00:39:28.350 --> 00:39:30.910 align:middle line:84% here is this what they call negotiated margin. 00:39:30.910 --> 00:39:32.710 align:middle line:84% It's the residual from this regression. 00:39:32.710 --> 00:39:37.020 align:middle line:90% 00:39:37.020 --> 00:39:40.410 align:middle line:84% So this is what they mostly analyze. 00:39:40.410 --> 00:39:44.010 align:middle line:84% So again, control variables here also account for time variation 00:39:44.010 --> 00:39:44.770 align:middle line:90% and so on. 00:39:44.770 --> 00:39:50.020 align:middle line:90% 00:39:50.020 --> 00:39:52.550 align:middle line:90% So here are some summary tables. 00:39:52.550 --> 00:39:55.990 align:middle line:84% There's a lot going on, but I want to point out a few things. 00:39:55.990 --> 00:39:59.830 align:middle line:84% So up here we have control markets. 00:39:59.830 --> 00:40:04.330 align:middle line:84% And down here we have treatment markets. 00:40:04.330 --> 00:40:07.300 align:middle line:90% And so both before and after. 00:40:07.300 --> 00:40:12.010 align:middle line:84% So one thing that's true is that the control markets 00:40:12.010 --> 00:40:14.710 align:middle line:84% have somewhat different consumers than treatment 00:40:14.710 --> 00:40:15.620 align:middle line:90% markets. 00:40:15.620 --> 00:40:17.620 align:middle line:84% So for instance, in the control markets, 00:40:17.620 --> 00:40:19.990 align:middle line:84% consumers' house value is somewhat lower. 00:40:19.990 --> 00:40:22.150 align:middle line:84% Consumers also have somewhat lower income 00:40:22.150 --> 00:40:23.830 align:middle line:90% than in the treatment markets. 00:40:23.830 --> 00:40:26.000 align:middle line:84% But on other variables, they're fairly balanced. 00:40:26.000 --> 00:40:30.490 align:middle line:84% So for instance, loan to value is almost identical. 00:40:30.490 --> 00:40:33.550 align:middle line:90% FICO scores are quite similar. 00:40:33.550 --> 00:40:36.640 align:middle line:84% The number of renters is pretty similar and so on. 00:40:36.640 --> 00:40:40.850 align:middle line:84% So they appear to be quite balanced on observables. 00:40:40.850 --> 00:40:44.210 align:middle line:90% 00:40:44.210 --> 00:40:46.100 align:middle line:84% Again, the key object of interest 00:40:46.100 --> 00:40:49.310 align:middle line:90% is this negotiated margin here. 00:40:49.310 --> 00:40:52.700 align:middle line:84% And the initial negotiated margin 00:40:52.700 --> 00:40:55.670 align:middle line:84% is very similar across both markets. 00:40:55.670 --> 00:40:56.600 align:middle line:90% So this is here. 00:40:56.600 --> 00:40:59.060 align:middle line:84% Again, control and treatment in the before period. 00:40:59.060 --> 00:41:02.970 align:middle line:84% And it's 106 and 107 basis points respectively. 00:41:02.970 --> 00:41:06.430 align:middle line:90% 00:41:06.430 --> 00:41:11.560 align:middle line:84% So now we can look at this negotiated margin in the after 00:41:11.560 --> 00:41:12.860 align:middle line:90% period. 00:41:12.860 --> 00:41:16.680 align:middle line:84% And in both markets, it goes up to 167 00:41:16.680 --> 00:41:19.720 align:middle line:84% in the control markets and a bit more 00:41:19.720 --> 00:41:29.040 align:middle line:84% than that to 172 in the treatment markets, 00:41:29.040 --> 00:41:33.022 align:middle line:84% where both lenders, both merging lenders had a branch, 00:41:33.022 --> 00:41:34.730 align:middle line:84% and therefore the choice set was reduced. 00:41:34.730 --> 00:41:39.000 align:middle line:90% 00:41:39.000 --> 00:41:41.270 align:middle line:84% So one thing that is interesting, 00:41:41.270 --> 00:41:44.510 align:middle line:84% you see it directly in the summary statistics here 00:41:44.510 --> 00:41:49.223 align:middle line:84% is the following, which is that if you look at this 00:41:49.223 --> 00:41:51.140 align:middle line:84% was the average effect that I just pointed out 00:41:51.140 --> 00:41:52.320 align:middle line:90% in the negotiated margin. 00:41:52.320 --> 00:41:56.420 align:middle line:84% But if you look at across the distribution of prices, 00:41:56.420 --> 00:41:59.990 align:middle line:84% you see that the 75th percentile of the negotiated margin is 00:41:59.990 --> 00:42:01.980 align:middle line:84% again very, very close to each other. 00:42:01.980 --> 00:42:04.160 align:middle line:84% They're very close in the after period. 00:42:04.160 --> 00:42:09.050 align:middle line:84% But there's larger difference, about eight basis points, 00:42:09.050 --> 00:42:12.630 align:middle line:90% in the 25th percentile. 00:42:12.630 --> 00:42:18.480 align:middle line:84% So it appears that most of this effect of higher average 00:42:18.480 --> 00:42:20.340 align:middle line:84% negotiated margin is coming from the fact 00:42:20.340 --> 00:42:23.337 align:middle line:84% that consumers at the low end of the price distribution 00:42:23.337 --> 00:42:24.545 align:middle line:90% are now paying higher prices. 00:42:24.545 --> 00:42:27.290 align:middle line:90% 00:42:27.290 --> 00:42:30.450 align:middle line:84% And so this is kind of a theme that will run through this. 00:42:30.450 --> 00:42:33.112 align:middle line:84% I told you at the beginning that what 00:42:33.112 --> 00:42:34.820 align:middle line:84% makes this paper interesting is that they 00:42:34.820 --> 00:42:36.620 align:middle line:84% want to look at the distributional effects 00:42:36.620 --> 00:42:37.410 align:middle line:90% of mergers. 00:42:37.410 --> 00:42:41.900 align:middle line:84% And it's tying this back to the theme of this class, which 00:42:41.900 --> 00:42:44.010 align:middle line:90% are search costs. 00:42:44.010 --> 00:42:48.950 align:middle line:84% The idea is that different consumers search differently 00:42:48.950 --> 00:42:50.790 align:middle line:84% as a function of their search costs. 00:42:50.790 --> 00:42:53.060 align:middle line:84% And what this implies for their search strategy, 00:42:53.060 --> 00:42:55.670 align:middle line:84% they get different kinds of prices. 00:42:55.670 --> 00:42:58.490 align:middle line:84% And if I'm a consumer who basically never searches, 00:42:58.490 --> 00:43:00.890 align:middle line:84% then the merger will probably not 00:43:00.890 --> 00:43:03.410 align:middle line:90% have as big of an effect on me. 00:43:03.410 --> 00:43:05.180 align:middle line:84% If I'm a consumer that searches a lot, 00:43:05.180 --> 00:43:10.160 align:middle line:84% then having an extra lender that I can go to 00:43:10.160 --> 00:43:12.840 align:middle line:84% is going to potentially affect my price. 00:43:12.840 --> 00:43:15.680 align:middle line:84% And so what is nice that is this idea 00:43:15.680 --> 00:43:19.626 align:middle line:84% is already visible here directly in the summary statistics. 00:43:19.626 --> 00:43:27.320 align:middle line:90% 00:43:27.320 --> 00:43:29.240 align:middle line:84% So this would normally be now a great table 00:43:29.240 --> 00:43:33.230 align:middle line:84% to ask my 14.03 undergrads to compute 00:43:33.230 --> 00:43:37.230 align:middle line:84% the difference in difference treatment effect from this. 00:43:37.230 --> 00:43:40.160 align:middle line:84% But I'm going to just show it to you. 00:43:40.160 --> 00:43:44.270 align:middle line:90% So it's not very large. 00:43:44.270 --> 00:43:47.660 align:middle line:84% Depending on how you look at it, whether you account for trends 00:43:47.660 --> 00:43:52.040 align:middle line:84% or not, you can do this either with a matching estimator 00:43:52.040 --> 00:43:54.570 align:middle line:84% or not, but the results are fairly consistent. 00:43:54.570 --> 00:43:55.840 align:middle line:90% It's about seven basis points. 00:43:55.840 --> 00:43:58.620 align:middle line:90% 00:43:58.620 --> 00:44:01.620 align:middle line:84% And so I think in the paper, they 00:44:01.620 --> 00:44:07.200 align:middle line:84% say for the average mortgage in this data, 00:44:07.200 --> 00:44:12.930 align:middle line:84% this implies about an additional $5 of extra payments per month. 00:44:12.930 --> 00:44:16.390 align:middle line:84% So the average effect is not large. 00:44:16.390 --> 00:44:21.780 align:middle line:84% And it suggests that this is a fairly, to begin with, 00:44:21.780 --> 00:44:23.000 align:middle line:90% a fairly competitive market. 00:44:23.000 --> 00:44:25.640 align:middle line:90% 00:44:25.640 --> 00:44:34.380 align:middle line:84% But what it tells us is that before IO people started 00:44:34.380 --> 00:44:37.140 align:middle line:84% looking at these markets, one view of the world 00:44:37.140 --> 00:44:39.120 align:middle line:84% could be that most of the variation that we 00:44:39.120 --> 00:44:42.355 align:middle line:84% see in these kinds of markets is due to risk or cost differences. 00:44:42.355 --> 00:44:43.980 align:middle line:84% And what this shows is that potentially 00:44:43.980 --> 00:44:45.995 align:middle line:84% there's a market power explanation here. 00:44:45.995 --> 00:44:49.680 align:middle line:90% 00:44:49.680 --> 00:44:52.450 align:middle line:90% I think I said this. 00:44:52.450 --> 00:44:54.270 align:middle line:84% And again, for people who search a lot, 00:44:54.270 --> 00:44:58.290 align:middle line:84% you would expect this to be a fairly competitive market. 00:44:58.290 --> 00:45:00.990 align:middle line:84% So this was the average treatment effect. 00:45:00.990 --> 00:45:04.680 align:middle line:84% So next we're going to look at the treatment effects 00:45:04.680 --> 00:45:07.980 align:middle line:84% at different points of the price distribution. 00:45:07.980 --> 00:45:15.510 align:middle line:84% So for that I have to introduce some additional notation here. 00:45:15.510 --> 00:45:18.630 align:middle line:84% So again, we're interested in the distribution 00:45:18.630 --> 00:45:21.900 align:middle line:90% of negotiated margins. 00:45:21.900 --> 00:45:24.660 align:middle line:84% And they're going to index those here with g and t, 00:45:24.660 --> 00:45:26.790 align:middle line:84% where g stands for whether or not 00:45:26.790 --> 00:45:30.900 align:middle line:84% you're part of the treatment or the control market. 00:45:30.900 --> 00:45:35.010 align:middle line:84% And t stands for whether or not it was before or after 00:45:35.010 --> 00:45:36.035 align:middle line:90% the merger happened. 00:45:36.035 --> 00:45:40.330 align:middle line:90% 00:45:40.330 --> 00:45:44.320 align:middle line:84% What we want to make this comparison 00:45:44.320 --> 00:45:48.820 align:middle line:84% is a counterfactual distribution of negotiated margins 00:45:48.820 --> 00:45:50.830 align:middle line:90% for the treated markets. 00:45:50.830 --> 00:45:53.725 align:middle line:84% So we would like to know what, just as you would 00:45:53.725 --> 00:45:54.850 align:middle line:90% in your typical difference. 00:45:54.850 --> 00:45:56.892 align:middle line:84% In difference analysis, you make some assumptions 00:45:56.892 --> 00:45:58.960 align:middle line:84% to know what would the counterfactual market have 00:45:58.960 --> 00:45:59.820 align:middle line:90% looked like. 00:45:59.820 --> 00:46:03.310 align:middle line:84% So similarly here, we want to construct 00:46:03.310 --> 00:46:05.680 align:middle line:84% a full counterfactual distribution of prices 00:46:05.680 --> 00:46:11.500 align:middle line:84% for markets that were treated for the counterfactual where 00:46:11.500 --> 00:46:14.320 align:middle line:90% the merger didn't happen. 00:46:14.320 --> 00:46:17.800 align:middle line:84% So what they're going to use is something 00:46:17.800 --> 00:46:20.000 align:middle line:84% that Athey and Imbens proposed in 2006. 00:46:20.000 --> 00:46:21.890 align:middle line:84% It's called a changes-in-changes estimator. 00:46:21.890 --> 00:46:25.000 align:middle line:84% It's basically a generalization of difference in differences 00:46:25.000 --> 00:46:27.110 align:middle line:84% where you basically look at treatment effects. 00:46:27.110 --> 00:46:30.040 align:middle line:84% We can look at treatment effects at different points 00:46:30.040 --> 00:46:31.940 align:middle line:90% of the distribution. 00:46:31.940 --> 00:46:38.050 align:middle line:90% 00:46:38.050 --> 00:46:40.900 align:middle line:84% So the way this works, always have to write it like this 00:46:40.900 --> 00:46:42.220 align:middle line:90% to remind myself. 00:46:42.220 --> 00:46:48.910 align:middle line:84% So the way this works is we're asking take some margin 00:46:48.910 --> 00:46:52.870 align:middle line:84% and look at basically remember here 00:46:52.870 --> 00:46:57.220 align:middle line:84% t stands for before and after. g stands for whether or not 00:46:57.220 --> 00:46:59.660 align:middle line:90% this is a treated market. 00:46:59.660 --> 00:47:05.980 align:middle line:84% So we want to look for those markets that were not treated. 00:47:05.980 --> 00:47:09.400 align:middle line:90% 00:47:09.400 --> 00:47:13.930 align:middle line:84% How large was the change in the margin at a specific point 00:47:13.930 --> 00:47:17.064 align:middle line:90% of the distribution? 00:47:17.064 --> 00:47:20.080 align:middle line:84% And we basically want that before for margin 00:47:20.080 --> 00:47:26.060 align:middle line:84% that gives rise to this across both markets 00:47:26.060 --> 00:47:29.000 align:middle line:84% is the same, because that defines 00:47:29.000 --> 00:47:31.070 align:middle line:90% the counterfactual distribution. 00:47:31.070 --> 00:47:33.980 align:middle line:84% So across all points of the distribution, 00:47:33.980 --> 00:47:39.500 align:middle line:84% we want to map from prices after the merger has 00:47:39.500 --> 00:47:43.370 align:middle line:84% happened to a set of prices before the merger. 00:47:43.370 --> 00:47:45.607 align:middle line:84% And under the counterfactual distribution, 00:47:45.607 --> 00:47:46.690 align:middle line:90% those have to be the same. 00:47:46.690 --> 00:47:51.080 align:middle line:90% 00:47:51.080 --> 00:47:56.690 align:middle line:84% STUDENT: Why is the first subscript a 0, 1 and not a 1, 0? 00:47:56.690 --> 00:47:57.690 align:middle line:90% TOBIAS SALZ: Let me see. 00:47:57.690 --> 00:47:58.713 align:middle line:90% So this here is-- 00:47:58.713 --> 00:48:03.040 align:middle line:90% 00:48:03.040 --> 00:48:04.610 align:middle line:90% yeah, this has to be 1, 0. 00:48:04.610 --> 00:48:05.900 align:middle line:90% Good catch. 00:48:05.900 --> 00:48:06.870 align:middle line:90% I apologize for that. 00:48:06.870 --> 00:48:14.470 align:middle line:90% 00:48:14.470 --> 00:48:15.550 align:middle line:90% OK. 00:48:15.550 --> 00:48:16.760 align:middle line:90% Sorry about that. 00:48:16.760 --> 00:48:19.740 align:middle line:84% So again, we're looking on both sides here at the change. 00:48:19.740 --> 00:48:24.248 align:middle line:90% 00:48:24.248 --> 00:48:26.140 align:middle line:84% And again, the counterfactual distribution 00:48:26.140 --> 00:48:28.720 align:middle line:84% is defined such that from any point 00:48:28.720 --> 00:48:32.410 align:middle line:84% that I look in the after distribution, the margin 00:48:32.410 --> 00:48:37.560 align:middle line:84% that maps to this point has to be the same. 00:48:37.560 --> 00:48:41.700 align:middle line:84% So now I can use this condition here. 00:48:41.700 --> 00:48:43.410 align:middle line:84% This is the object that I do not know, 00:48:43.410 --> 00:48:46.350 align:middle line:84% which is the counterfactual distribution of prices. 00:48:46.350 --> 00:48:54.455 align:middle line:84% And I can apply, basically, and yeah, here it's correct. 00:48:54.455 --> 00:48:57.660 align:middle line:90% 00:48:57.660 --> 00:49:00.550 align:middle line:84% So I can apply f 1, 0 to both sides. 00:49:00.550 --> 00:49:06.480 align:middle line:84% And I back out the counterfactual distribution 00:49:06.480 --> 00:49:06.995 align:middle line:90% of prices. 00:49:06.995 --> 00:49:11.600 align:middle line:90% 00:49:11.600 --> 00:49:17.250 align:middle line:84% So what is nice about this is once I have this object here, 00:49:17.250 --> 00:49:19.050 align:middle line:84% which is, again, constructed from three 00:49:19.050 --> 00:49:22.380 align:middle line:84% observed distributions, I can look 00:49:22.380 --> 00:49:25.830 align:middle line:84% at different points of the distribution 00:49:25.830 --> 00:49:29.170 align:middle line:90% how are consumers treated. 00:49:29.170 --> 00:49:31.710 align:middle line:84% And in particular, I can come back to this idea 00:49:31.710 --> 00:49:34.680 align:middle line:84% that consumers that I expect to search a lot 00:49:34.680 --> 00:49:38.700 align:middle line:84% should be hurt more by the merger than consumers 00:49:38.700 --> 00:49:39.275 align:middle line:90% who do not. 00:49:39.275 --> 00:49:48.080 align:middle line:90% 00:49:48.080 --> 00:49:52.220 align:middle line:90% Any questions on this idea here? 00:49:52.220 --> 00:49:56.030 align:middle line:84% So again, conceptually, it's like difference in differences, 00:49:56.030 --> 00:49:58.950 align:middle line:84% but you look at different points of the distribution. 00:49:58.950 --> 00:50:01.460 align:middle line:90% 00:50:01.460 --> 00:50:07.100 align:middle line:84% So here are the results that we get out of this. 00:50:07.100 --> 00:50:10.250 align:middle line:84% Again, you can look at those either what 00:50:10.250 --> 00:50:14.990 align:middle line:84% they call baseline or adjusting for trends. 00:50:14.990 --> 00:50:17.930 align:middle line:84% And what you find here is that if you 00:50:17.930 --> 00:50:24.320 align:middle line:84% look at the top end of the distribution, the confidence 00:50:24.320 --> 00:50:26.310 align:middle line:90% bands, they cross 0. 00:50:26.310 --> 00:50:28.700 align:middle line:84% So these are all not significant. 00:50:28.700 --> 00:50:30.920 align:middle line:84% But at the lower end of the price distribution, 00:50:30.920 --> 00:50:32.660 align:middle line:84% you get significant effects at the order 00:50:32.660 --> 00:50:34.870 align:middle line:90% of 7 to 10 basis points. 00:50:34.870 --> 00:50:38.250 align:middle line:90% 00:50:38.250 --> 00:50:40.680 align:middle line:84% So again, what is the interpretation here? 00:50:40.680 --> 00:50:46.320 align:middle line:84% There's some consumers in the market that are searching a lot. 00:50:46.320 --> 00:50:50.820 align:middle line:84% And once their choice set is reduced, their prices go up. 00:50:50.820 --> 00:50:53.440 align:middle line:84% But consumers who were not shopping around to begin with, 00:50:53.440 --> 00:50:55.775 align:middle line:84% they're unaffected by this price change. 00:50:55.775 --> 00:51:00.510 align:middle line:90% 00:51:00.510 --> 00:51:06.600 align:middle line:84% So this leads to dispersion, to reduction and dispersion 00:51:06.600 --> 00:51:07.560 align:middle line:90% and prices. 00:51:07.560 --> 00:51:10.350 align:middle line:84% So different measures, different dispersion measures 00:51:10.350 --> 00:51:13.020 align:middle line:84% such as standard deviation, coefficient of variation, 00:51:13.020 --> 00:51:15.743 align:middle line:84% and interquartile range all tell a consistent story 00:51:15.743 --> 00:51:17.160 align:middle line:84% that prices are getting compressed 00:51:17.160 --> 00:51:19.740 align:middle line:90% once the merger happens. 00:51:19.740 --> 00:51:22.230 align:middle line:84% What they also do is they are looking 00:51:22.230 --> 00:51:24.840 align:middle line:84% at whether or not this counterfactual distribution 00:51:24.840 --> 00:51:29.910 align:middle line:84% of prices that they back out is equal 00:51:29.910 --> 00:51:32.810 align:middle line:84% or not to the actual price distribution that they observe. 00:51:32.810 --> 00:51:36.230 align:middle line:90% 00:51:36.230 --> 00:51:40.180 align:middle line:84% And they can reject this with a Kolmogorov-Smirnov test. 00:51:40.180 --> 00:51:44.290 align:middle line:84% So basically, looking over the full distribution of prices, 00:51:44.290 --> 00:51:47.950 align:middle line:84% reject the null hypothesis that the price distribution didn't 00:51:47.950 --> 00:51:50.440 align:middle line:84% change relative to the counterfactual price 00:51:50.440 --> 00:51:53.490 align:middle line:90% distribution. 00:51:53.490 --> 00:51:57.300 align:middle line:84% So I think this is nice because, again, it 00:51:57.300 --> 00:52:02.190 align:middle line:84% tells a story that's quite consistent with the idea 00:52:02.190 --> 00:52:06.310 align:middle line:84% of different consumers faring very different in these markets. 00:52:06.310 --> 00:52:10.340 align:middle line:84% And you can observe it in those heterogeneous treatment effects. 00:52:10.340 --> 00:52:13.600 align:middle line:90% 00:52:13.600 --> 00:52:17.230 align:middle line:84% So they now construct a search model 00:52:17.230 --> 00:52:21.940 align:middle line:90% to make sense of those results. 00:52:21.940 --> 00:52:24.880 align:middle line:84% And the model is constructed so that in the end, 00:52:24.880 --> 00:52:29.800 align:middle line:84% it maps back to these negotiated margins in the after period. 00:52:29.800 --> 00:52:35.410 align:middle line:84% So we have n plus 1 lenders in this market. 00:52:35.410 --> 00:52:38.590 align:middle line:84% And there are three stages to the model. 00:52:38.590 --> 00:52:42.100 align:middle line:84% So initially, as in Hong and Shum, 00:52:42.100 --> 00:52:45.120 align:middle line:84% consumers get a free price quote. 00:52:45.120 --> 00:52:48.310 align:middle line:84% So you can think of this maybe from their home lender. 00:52:48.310 --> 00:52:50.740 align:middle line:90% I get a free quote. 00:52:50.740 --> 00:52:55.210 align:middle line:84% I can either accept this quote or I can reject the quote 00:52:55.210 --> 00:53:01.450 align:middle line:84% and engage in costly effort to gain additional quotes. 00:53:01.450 --> 00:53:08.710 align:middle line:84% And so they're using a search technology 00:53:08.710 --> 00:53:11.270 align:middle line:84% that's somewhat different from what you have seen before. 00:53:11.270 --> 00:53:14.700 align:middle line:84% So basically, there's a continuous effort function. 00:53:14.700 --> 00:53:18.610 align:middle line:84% And as I increase my continuous effort, 00:53:18.610 --> 00:53:21.820 align:middle line:84% I get a higher probability that I'm getting more quotes 00:53:21.820 --> 00:53:23.710 align:middle line:90% from the price distribution. 00:53:23.710 --> 00:53:29.230 align:middle line:84% And the way that they construct this is that with probability s, 00:53:29.230 --> 00:53:32.530 align:middle line:84% I either get quotes from all the firms in the market 00:53:32.530 --> 00:53:36.102 align:middle line:84% and with probability 1 minus s, I only get two quotes. 00:53:36.102 --> 00:53:37.810 align:middle line:84% So this turns out to be quite convenient, 00:53:37.810 --> 00:53:42.610 align:middle line:84% because they can basically map this just in continuous effort 00:53:42.610 --> 00:53:48.190 align:middle line:84% problem instead of dealing with something like a reservation 00:53:48.190 --> 00:53:50.650 align:middle line:84% price strategy in the sequential search model 00:53:50.650 --> 00:53:54.790 align:middle line:84% or having to compare different bins 00:53:54.790 --> 00:53:56.890 align:middle line:84% and where consumers sort themselves into. 00:53:56.890 --> 00:54:02.780 align:middle line:84% So this is somewhat similar to what people do in labor search. 00:54:02.780 --> 00:54:05.510 align:middle line:84% So I'll come back to that comparison later on. 00:54:05.510 --> 00:54:09.820 align:middle line:84% And it just makes the model slightly easier to handle. 00:54:09.820 --> 00:54:13.280 align:middle line:90% 00:54:13.280 --> 00:54:16.850 align:middle line:90% So these are the three stages. 00:54:16.850 --> 00:54:18.810 align:middle line:90% Consumers get an initial quote. 00:54:18.810 --> 00:54:21.120 align:middle line:90% They either search or not. 00:54:21.120 --> 00:54:24.320 align:middle line:84% And when they search, consumer, the lenders 00:54:24.320 --> 00:54:28.530 align:middle line:90% compete for their business. 00:54:28.530 --> 00:54:29.820 align:middle line:90% So let's work backwards. 00:54:29.820 --> 00:54:31.630 align:middle line:90% Let's first go to stage three. 00:54:31.630 --> 00:54:34.730 align:middle line:84% So this is where consumers went out and get additional quotes 00:54:34.730 --> 00:54:36.920 align:middle line:90% from the market. 00:54:36.920 --> 00:54:38.820 align:middle line:84% They modeled this as an English auction. 00:54:38.820 --> 00:54:40.610 align:middle line:84% Again, this is a fairly strong assumption, 00:54:40.610 --> 00:54:47.060 align:middle line:84% but it makes dealing with the model somewhat easier. 00:54:47.060 --> 00:54:50.420 align:middle line:84% So lenders here have a common cost component 00:54:50.420 --> 00:54:55.960 align:middle line:84% c and an idiosyncratic component epsilon j. 00:54:55.960 --> 00:55:00.550 align:middle line:84% And as you know, there is a weakly dominant strategy 00:55:00.550 --> 00:55:04.450 align:middle line:84% in English auctions, which is to bid your cost. 00:55:04.450 --> 00:55:07.480 align:middle line:84% So the lowest cost lender wins and is 00:55:07.480 --> 00:55:09.585 align:middle line:84% paid the price of the second lowest cost lender. 00:55:09.585 --> 00:55:15.220 align:middle line:90% 00:55:15.220 --> 00:55:18.970 align:middle line:84% So with this, they can then define 00:55:18.970 --> 00:55:23.950 align:middle line:84% an expected price, or an expected negotiated margin, 00:55:23.950 --> 00:55:27.580 align:middle line:84% which is, again, a function of this common part here, 00:55:27.580 --> 00:55:30.790 align:middle line:84% as well as the expected second order statistic out of taking 00:55:30.790 --> 00:55:33.800 align:middle line:90% n draws from the distribution. 00:55:33.800 --> 00:55:38.090 align:middle line:90% 00:55:38.090 --> 00:55:40.920 align:middle line:84% And since they make this assumption that consumers either 00:55:40.920 --> 00:55:45.300 align:middle line:84% take two draws or n draws, depending on whether or not 00:55:45.300 --> 00:55:48.450 align:middle line:84% their search is successful, you get this difference 00:55:48.450 --> 00:55:53.815 align:middle line:84% in expected prices or delta n as a function of how many lenders 00:55:53.815 --> 00:55:54.815 align:middle line:90% there are in the market. 00:55:54.815 --> 00:55:57.620 align:middle line:90% 00:55:57.620 --> 00:56:01.790 align:middle line:84% So thinking about this, you can already 00:56:01.790 --> 00:56:04.220 align:middle line:84% see how this maps back to the empirics. 00:56:04.220 --> 00:56:07.640 align:middle line:84% So in some markets, the total number 00:56:07.640 --> 00:56:13.490 align:middle line:84% of firms that you can search from is reduced by 1. 00:56:13.490 --> 00:56:18.560 align:middle line:84% And of course, under this specific search technology, 00:56:18.560 --> 00:56:21.440 align:middle line:84% some consumers who would have taken usually 00:56:21.440 --> 00:56:23.420 align:middle line:84% n plus 1 draws are now only taking n 00:56:23.420 --> 00:56:25.230 align:middle line:84% draws when their search is successful. 00:56:25.230 --> 00:56:33.090 align:middle line:90% 00:56:33.090 --> 00:56:36.840 align:middle line:84% So this was the last stage where the lenders are 00:56:36.840 --> 00:56:38.470 align:middle line:90% competing against each other. 00:56:38.470 --> 00:56:43.080 align:middle line:84% We now move to stage two where consumers are 00:56:43.080 --> 00:56:47.040 align:middle line:90% deciding how much to search. 00:56:47.040 --> 00:56:52.050 align:middle line:84% And so their total search cost here, which they call kappa i, 00:56:52.050 --> 00:56:55.410 align:middle line:84% is a function of how much effort they exert. 00:56:55.410 --> 00:56:58.320 align:middle line:84% And in particular, it depends on an idiosyncratic cost 00:56:58.320 --> 00:57:00.120 align:middle line:90% for each consumer UI. 00:57:00.120 --> 00:57:01.910 align:middle line:84% So this is the consumer's marginal cost 00:57:01.910 --> 00:57:04.950 align:middle line:84% of searching, exerting more effort, 00:57:04.950 --> 00:57:08.640 align:middle line:90% plus a fixed component eta i. 00:57:08.640 --> 00:57:11.510 align:middle line:84% So if you reject your initial quote and you go out to search, 00:57:11.510 --> 00:57:13.370 align:middle line:90% you pay this fixed cost. 00:57:13.370 --> 00:57:16.060 align:middle line:84% It will turn out, I'll come back to this, that they basically 00:57:16.060 --> 00:57:17.810 align:middle line:84% dropped this in the empirical application, 00:57:17.810 --> 00:57:19.227 align:middle line:84% but I'll leave it in here for now. 00:57:19.227 --> 00:57:23.320 align:middle line:90% 00:57:23.320 --> 00:57:27.430 align:middle line:84% So an important assumption and restriction in this model 00:57:27.430 --> 00:57:31.660 align:middle line:84% is that they assume that the marginal cost of searching 00:57:31.660 --> 00:57:34.900 align:middle line:84% of the consumer is publicly observed. 00:57:34.900 --> 00:57:40.810 align:middle line:84% So in other words, the lender, once you talk to the lender, 00:57:40.810 --> 00:57:42.610 align:middle line:84% you walk into the branch, they can actually 00:57:42.610 --> 00:57:45.430 align:middle line:84% somehow learn enough about you to know exactly 00:57:45.430 --> 00:57:47.500 align:middle line:90% what your cost of searching is. 00:57:47.500 --> 00:57:49.640 align:middle line:84% And this has quite strong implications. 00:57:49.640 --> 00:57:50.680 align:middle line:90% I'll come back to this. 00:57:50.680 --> 00:57:53.260 align:middle line:84% But this is something important to keep in mind here. 00:57:53.260 --> 00:57:57.075 align:middle line:84% So the marginal cost of searching is publicly observed. 00:57:57.075 --> 00:58:00.330 align:middle line:90% 00:58:00.330 --> 00:58:02.708 align:middle line:84% They assume that there's matching probability. 00:58:02.708 --> 00:58:04.500 align:middle line:84% So the chance that the search is successful 00:58:04.500 --> 00:58:06.480 align:middle line:84% and I get n draws from the price distribution 00:58:06.480 --> 00:58:09.700 align:middle line:84% has this Pareto distribution here. 00:58:09.700 --> 00:58:16.020 align:middle line:84% And so with that, consumers compute their reservation value, 00:58:16.020 --> 00:58:21.660 align:middle line:84% which is the minimizer of the search cost 00:58:21.660 --> 00:58:25.560 align:middle line:84% plus the expected price that I get. 00:58:25.560 --> 00:58:28.590 align:middle line:84% So again, this here is a common cost component. 00:58:28.590 --> 00:58:31.500 align:middle line:84% If I'm not successful, I'm getting the second order 00:58:31.500 --> 00:58:34.500 align:middle line:84% statistics out of taking two draws. 00:58:34.500 --> 00:58:37.050 align:middle line:84% And when I'm successful, I'm replacing this 00:58:37.050 --> 00:58:40.890 align:middle line:84% with the second-order statistic out of n draws. 00:58:40.890 --> 00:58:46.450 align:middle line:84% This is this delta n that we defined beforehand. 00:58:46.450 --> 00:58:49.110 align:middle line:90% The difference in expectations. 00:58:49.110 --> 00:58:54.150 align:middle line:84% So it will be useful to keep track of this reservation value 00:58:54.150 --> 00:58:57.730 align:middle line:84% as a part that is common to all consumers 00:58:57.730 --> 00:58:59.460 align:middle line:84% and this idiosyncratic component here, 00:58:59.460 --> 00:59:02.100 align:middle line:84% which depends on the consumer's idiosyncratic search 00:59:02.100 --> 00:59:05.120 align:middle line:84% cost, as well as the number of lenders in the market. 00:59:05.120 --> 00:59:07.630 align:middle line:90% 00:59:07.630 --> 00:59:12.840 align:middle line:84% So in other words, consumers are different in terms of how costly 00:59:12.840 --> 00:59:14.290 align:middle line:90% it is to search for them. 00:59:14.290 --> 00:59:16.650 align:middle line:84% They exert in this continuous effort. 00:59:16.650 --> 00:59:19.500 align:middle line:84% This gives them some reservation value 00:59:19.500 --> 00:59:21.125 align:middle line:84% comes out of this optimization problem. 00:59:21.125 --> 00:59:24.830 align:middle line:90% 00:59:24.830 --> 00:59:28.260 align:middle line:84% And there are some consumers who will not 00:59:28.260 --> 00:59:32.345 align:middle line:84% search at all if their search cost is too high. 00:59:32.345 --> 00:59:36.100 align:middle line:90% 00:59:36.100 --> 00:59:39.550 align:middle line:84% So we are now moving to stage one, 00:59:39.550 --> 00:59:45.567 align:middle line:84% which is the lender who gives the initial quote decides 00:59:45.567 --> 00:59:46.650 align:middle line:90% what that quote should be. 00:59:46.650 --> 00:59:52.395 align:middle line:90% 00:59:52.395 --> 00:59:56.080 align:middle line:84% So for that, we define the distribution 00:59:56.080 --> 00:59:59.440 align:middle line:84% of eta, which was this fixed cost of searching. 00:59:59.440 --> 01:00:04.600 align:middle line:84% So the consumer who decides whether or not to take more 01:00:04.600 --> 01:00:08.355 align:middle line:84% draws privately observes this fixed cost. 01:00:08.355 --> 01:00:11.260 align:middle line:90% 01:00:11.260 --> 01:00:14.710 align:middle line:84% And the probability that the consumer goes out 01:00:14.710 --> 01:00:17.260 align:middle line:84% to take more draws is the probability 01:00:17.260 --> 01:00:21.730 align:middle line:84% that eta i plus the reservation value 01:00:21.730 --> 01:00:26.050 align:middle line:84% is lower than the initial quote that I get. 01:00:26.050 --> 01:00:28.210 align:middle line:84% The lender observes ui, so the lender 01:00:28.210 --> 01:00:30.790 align:middle line:84% knows what your reservation value is. 01:00:30.790 --> 01:00:33.110 align:middle line:84% But the lender does not know eta i. 01:00:33.110 --> 01:00:35.385 align:middle line:90% So this is privately observed. 01:00:35.385 --> 01:00:40.360 align:middle line:90% 01:00:40.360 --> 01:00:43.720 align:middle line:84% So the lender who sets the initial quote 01:00:43.720 --> 01:00:47.200 align:middle line:84% maximizes, again, the margin times the probability 01:00:47.200 --> 01:00:49.415 align:middle line:84% that the consumer does not walk away. 01:00:49.415 --> 01:00:51.040 align:middle line:84% And there's an implicit assumption here 01:00:51.040 --> 01:00:53.290 align:middle line:84% that when the consumer walks away, then 01:00:53.290 --> 01:00:55.900 align:middle line:84% I'm losing the business of the consumer for sure. 01:00:55.900 --> 01:00:59.680 align:middle line:84% So the consumer then takes one of the offers 01:00:59.680 --> 01:01:01.690 align:middle line:84% from the other lenders and doesn't compare this 01:01:01.690 --> 01:01:03.780 align:middle line:90% to my initial m0 quote. 01:01:03.780 --> 01:01:09.290 align:middle line:90% 01:01:09.290 --> 01:01:15.020 align:middle line:84% So this is basically a monopoly pricing problem, 01:01:15.020 --> 01:01:19.160 align:middle line:84% where I'm pricing against this residual demand curve here, 01:01:19.160 --> 01:01:23.360 align:middle line:84% which is whether or not the consumer walks away or not. 01:01:23.360 --> 01:01:25.700 align:middle line:84% And you get this solution here where 01:01:25.700 --> 01:01:29.620 align:middle line:84% the margin is equal to the inverse hazard rate. 01:01:29.620 --> 01:01:33.200 align:middle line:84% So a pretty standard solution for these problems 01:01:33.200 --> 01:01:35.480 align:middle line:84% where consumers have private information. 01:01:35.480 --> 01:01:37.910 align:middle line:84% So the consumer earns some information rent, 01:01:37.910 --> 01:01:42.110 align:middle line:84% and that leads in principle to somewhat lower prices 01:01:42.110 --> 01:01:44.490 align:middle line:84% than the reservation value would imply. 01:01:44.490 --> 01:01:48.710 align:middle line:84% In the empirical application, however, they're 01:01:48.710 --> 01:01:51.943 align:middle line:90% going to just set eta i to 0. 01:01:51.943 --> 01:01:53.360 align:middle line:84% So the empirical application, they 01:01:53.360 --> 01:01:56.360 align:middle line:84% assume that consumers do not earn any information 01:01:56.360 --> 01:01:58.460 align:middle line:84% rent and lenders are therefore just paying them 01:01:58.460 --> 01:02:00.950 align:middle line:84% their reservation value, which is because they 01:02:00.950 --> 01:02:02.485 align:middle line:90% observe ui publicly observed. 01:02:02.485 --> 01:02:11.890 align:middle line:90% 01:02:11.890 --> 01:02:17.770 align:middle line:84% OK, so with this solution, where there is basically now, 01:02:17.770 --> 01:02:21.070 align:middle line:84% according to this very stark restriction, no more 01:02:21.070 --> 01:02:23.710 align:middle line:84% any private information of consumers, 01:02:23.710 --> 01:02:27.010 align:middle line:84% you can directly map what comes out 01:02:27.010 --> 01:02:32.020 align:middle line:84% of this search problem of the consumer to the observed 01:02:32.020 --> 01:02:35.220 align:middle line:84% distribution of negotiated margins in the after period. 01:02:35.220 --> 01:02:39.500 align:middle line:84% So remember, this here is just for the treated markets 01:02:39.500 --> 01:02:40.840 align:middle line:90% in the after period. 01:02:40.840 --> 01:02:45.220 align:middle line:84% The fraction of consumers who get a negotiated margin less 01:02:45.220 --> 01:02:46.930 align:middle line:90% than m. 01:02:46.930 --> 01:02:49.270 align:middle line:84% And we can write this as the probability 01:02:49.270 --> 01:02:52.240 align:middle line:84% that this reservation value, which 01:02:52.240 --> 01:02:54.880 align:middle line:84% had the common component and the consumer idiosyncratic 01:02:54.880 --> 01:02:57.610 align:middle line:90% component, is less than m. 01:02:57.610 --> 01:03:01.720 align:middle line:84% And if we define hu to be the distribution 01:03:01.720 --> 01:03:04.060 align:middle line:84% of these idiosyncratic search costs, 01:03:04.060 --> 01:03:06.730 align:middle line:84% we can instead rewrite this as follows. 01:03:06.730 --> 01:03:14.218 align:middle line:84% So we can write the quantiles of the consumer types, which is 01:03:14.218 --> 01:03:15.510 align:middle line:90% how costly it is to search for. 01:03:15.510 --> 01:03:17.885 align:middle line:84% You have to correspond to the quantiles of the negotiated 01:03:17.885 --> 01:03:19.550 align:middle line:90% margin. 01:03:19.550 --> 01:03:22.610 align:middle line:84% So this is a strategy that you will also 01:03:22.610 --> 01:03:25.880 align:middle line:84% see in auctions where there's something kind of idiosyncratic 01:03:25.880 --> 01:03:27.180 align:middle line:90% that's unobserved. 01:03:27.180 --> 01:03:30.110 align:middle line:84% And consumers are ordered in some way. 01:03:30.110 --> 01:03:32.450 align:middle line:84% And that's going to map to something 01:03:32.450 --> 01:03:33.890 align:middle line:90% observed that's kind of ordered. 01:03:33.890 --> 01:03:37.050 align:middle line:84% In this case, the price that consumers get. 01:03:37.050 --> 01:03:42.930 align:middle line:84% And again, the way this works here is I have my search costs. 01:03:42.930 --> 01:03:44.900 align:middle line:84% The lender observes that and just pays me 01:03:44.900 --> 01:03:46.410 align:middle line:90% off my reservation value. 01:03:46.410 --> 01:03:49.190 align:middle line:84% And so you get this kind of perfect ordering 01:03:49.190 --> 01:03:51.080 align:middle line:84% of the negotiated margin that corresponds 01:03:51.080 --> 01:03:52.945 align:middle line:84% to the ordering of these unobserved types. 01:03:52.945 --> 01:03:59.730 align:middle line:90% 01:03:59.730 --> 01:04:04.770 align:middle line:84% Of course, this here only works for a given set of parameters 01:04:04.770 --> 01:04:09.870 align:middle line:84% of the remaining problem here, which still includes the firm's 01:04:09.870 --> 01:04:13.770 align:middle line:84% cost variation as well as lambda i, 01:04:13.770 --> 01:04:16.980 align:middle line:84% which was determining how costly it is for consumers to search. 01:04:16.980 --> 01:04:23.040 align:middle line:84% So we can't quite directly back out this distribution hu, 01:04:23.040 --> 01:04:25.740 align:middle line:84% because we don't know yet the parameters, the remaining 01:04:25.740 --> 01:04:30.090 align:middle line:84% parameters of this problem, which 01:04:30.090 --> 01:04:35.045 align:middle line:84% are the search effort parameters as well as the cost of lenders. 01:04:35.045 --> 01:04:39.410 align:middle line:90% 01:04:39.410 --> 01:04:44.480 align:middle line:84% So one thing that you get out of this 01:04:44.480 --> 01:04:50.060 align:middle line:84% is that consumers that have higher search costs 01:04:50.060 --> 01:04:53.190 align:middle line:84% benefit less from the merger than consumers with lower search 01:04:53.190 --> 01:04:53.690 align:middle line:90% costs. 01:04:53.690 --> 01:04:58.460 align:middle line:84% So for u prime larger u, you get a difference 01:04:58.460 --> 01:05:03.650 align:middle line:84% in reservation values for less sophisticated consumers 01:05:03.650 --> 01:05:06.740 align:middle line:84% that's smaller than for consumers 01:05:06.740 --> 01:05:09.140 align:middle line:90% that have a lower search cost. 01:05:09.140 --> 01:05:11.810 align:middle line:84% So in other words, what this model here generates 01:05:11.810 --> 01:05:15.168 align:middle line:84% is this difference that we already saw in the data 01:05:15.168 --> 01:05:17.210 align:middle line:84% that we see larger treatment effects at the lower 01:05:17.210 --> 01:05:19.160 align:middle line:84% end of the price distribution, where 01:05:19.160 --> 01:05:21.980 align:middle line:84% we would expect to be the consumers that have lower search 01:05:21.980 --> 01:05:25.500 align:middle line:84% costs relative to at the high of the price distribution. 01:05:25.500 --> 01:05:28.564 align:middle line:90% 01:05:28.564 --> 01:05:34.260 align:middle line:84% So in other words, this pricing function 01:05:34.260 --> 01:05:38.640 align:middle line:84% under symmetric information basically is monotonically 01:05:38.640 --> 01:05:41.820 align:middle line:90% increasing in ui. 01:05:41.820 --> 01:05:43.260 align:middle line:90% It's weakly decreasing in n. 01:05:43.260 --> 01:05:45.690 align:middle line:90% So competition helps. 01:05:45.690 --> 01:05:50.400 align:middle line:84% And there are these increasing marginal effects of the merger 01:05:50.400 --> 01:05:53.345 align:middle line:84% when you get to consumers of lower search costs. 01:05:53.345 --> 01:05:58.540 align:middle line:90% 01:05:58.540 --> 01:06:05.140 align:middle line:84% So with all of this put together, what they're doing is 01:06:05.140 --> 01:06:10.360 align:middle line:84% they now want to explain the actual observed differences 01:06:10.360 --> 01:06:12.520 align:middle line:90% in the treatment effects. 01:06:12.520 --> 01:06:16.600 align:middle line:84% So we had basically large treatment effects 01:06:16.600 --> 01:06:19.420 align:middle line:84% at the bottom of the price distribution, small treatment 01:06:19.420 --> 01:06:20.660 align:middle line:90% effects at the top. 01:06:20.660 --> 01:06:25.210 align:middle line:84% And they're going to find parameters, 01:06:25.210 --> 01:06:29.260 align:middle line:84% which is the common cost of lenders, the consumer search 01:06:29.260 --> 01:06:32.350 align:middle line:84% strategy, as well as the variation of the idiosyncratic 01:06:32.350 --> 01:06:36.700 align:middle line:84% cost of lenders, that best explain 01:06:36.700 --> 01:06:40.900 align:middle line:84% the differences in negotiated margins going from before to 01:06:40.900 --> 01:06:42.690 align:middle line:90% after the merger. 01:06:42.690 --> 01:06:48.190 align:middle line:84% So this is, I think, quite nice because they basically 01:06:48.190 --> 01:06:51.460 align:middle line:84% let the model be informed by this treatment effect 01:06:51.460 --> 01:06:54.070 align:middle line:90% that they observe in the data. 01:06:54.070 --> 01:07:01.240 align:middle line:84% And so because of admittedly a fairly stark assumption, 01:07:01.240 --> 01:07:05.860 align:middle line:84% they get this direct mapping between the observed negotiated 01:07:05.860 --> 01:07:09.580 align:middle line:84% margin and what is unobserved to the researcher, which 01:07:09.580 --> 01:07:11.870 align:middle line:84% is the type distribution of consumers. 01:07:11.870 --> 01:07:19.850 align:middle line:84% They can impose basically that quantiles of consumer types 01:07:19.850 --> 01:07:22.318 align:middle line:84% have to correspond to this object 01:07:22.318 --> 01:07:23.610 align:middle line:90% that they get out of the model. 01:07:23.610 --> 01:07:25.550 align:middle line:84% So they're basically directly imposing 01:07:25.550 --> 01:07:27.230 align:middle line:84% the fact that these observed margins 01:07:27.230 --> 01:07:31.510 align:middle line:84% have to be equal to these unobserved consumer types. 01:07:31.510 --> 01:07:37.640 align:middle line:90% 01:07:37.640 --> 01:07:41.630 align:middle line:84% So again, the way this works is they only 01:07:41.630 --> 01:07:44.300 align:middle line:84% have these three parameters here. 01:07:44.300 --> 01:07:46.320 align:middle line:84% This is a method of moments problem. 01:07:46.320 --> 01:07:49.430 align:middle line:84% And for each time they evaluate the objective function, 01:07:49.430 --> 01:07:52.310 align:middle line:84% they directly impose this correspondence 01:07:52.310 --> 01:07:56.170 align:middle line:84% between observed margins and unobserved consumer types. 01:07:56.170 --> 01:08:01.240 align:middle line:90% 01:08:01.240 --> 01:08:03.680 align:middle line:84% So what do they get out of this in the end? 01:08:03.680 --> 01:08:07.150 align:middle line:84% So one thing that's typically, of course, 01:08:07.150 --> 01:08:09.100 align:middle line:84% when people compute structural models, 01:08:09.100 --> 01:08:11.780 align:middle line:84% they use them to compute counterfactuals. 01:08:11.780 --> 01:08:14.630 align:middle line:84% Here, the counterfactual is already observed in the data. 01:08:14.630 --> 01:08:17.770 align:middle line:84% So the additional thing that you get out of computing this model 01:08:17.770 --> 01:08:20.859 align:middle line:84% is that you can learn something about the unobserved search 01:08:20.859 --> 01:08:22.130 align:middle line:90% costs of consumers. 01:08:22.130 --> 01:08:27.850 align:middle line:84% They estimate that this is about $1,700 for an average loan size 01:08:27.850 --> 01:08:32.140 align:middle line:90% of about $152,000. 01:08:32.140 --> 01:08:35.140 align:middle line:84% And because they also get lenders' costs from this, 01:08:35.140 --> 01:08:37.479 align:middle line:84% they can compute their markups, which 01:08:37.479 --> 01:08:41.859 align:middle line:84% are at the order of 5% to 10% And so, 01:08:41.859 --> 01:08:45.550 align:middle line:84% again, this model, I mean, it's constructed in this way, 01:08:45.550 --> 01:08:48.460 align:middle line:84% but it directly maps to this empirical observation. 01:08:48.460 --> 01:08:52.390 align:middle line:84% The consumers at the bottom end of the price distribution 01:08:52.390 --> 01:08:54.819 align:middle line:84% are hurt more by the merger than consumers 01:08:54.819 --> 01:08:56.969 align:middle line:90% who are searching much less. 01:08:56.969 --> 01:09:03.600 align:middle line:90% 01:09:03.600 --> 01:09:06.720 align:middle line:84% I think this is a very nice paper, because again, it 01:09:06.720 --> 01:09:12.540 align:middle line:84% uses an actual change in market structure 01:09:12.540 --> 01:09:13.870 align:middle line:90% that's observed in the data. 01:09:13.870 --> 01:09:16.050 align:middle line:90% So an actual merger. 01:09:16.050 --> 01:09:19.590 align:middle line:84% It highlights what I think is a new angle on the effects 01:09:19.590 --> 01:09:22.870 align:middle line:84% of mergers, is that different consumers fare very differently. 01:09:22.870 --> 01:09:26.010 align:middle line:84% And that could potentially mean that we underestimate the market 01:09:26.010 --> 01:09:27.310 align:middle line:90% power effect of a merger. 01:09:27.310 --> 01:09:31.710 align:middle line:84% Because if, let's say, consumers don't search very much 01:09:31.710 --> 01:09:35.100 align:middle line:84% or we have markets in which search costs are very high, 01:09:35.100 --> 01:09:38.220 align:middle line:84% and we do, let's say, a simple before-after merger analysis, 01:09:38.220 --> 01:09:41.010 align:middle line:84% we might observe that prices don't change all that much. 01:09:41.010 --> 01:09:44.010 align:middle line:84% But that could be because consumers are just not searching 01:09:44.010 --> 01:09:44.645 align:middle line:90% very much. 01:09:44.645 --> 01:09:46.020 align:middle line:84% And so to the extent that we want 01:09:46.020 --> 01:09:48.460 align:middle line:84% to learn what's the market power effect of a merger, 01:09:48.460 --> 01:09:51.720 align:middle line:84% it's important to account for that. 01:09:51.720 --> 01:09:54.149 align:middle line:84% And it provides this what I think 01:09:54.149 --> 01:09:57.168 align:middle line:84% is very tight link between the treatment effect analysis 01:09:57.168 --> 01:09:58.210 align:middle line:90% and the structural model. 01:09:58.210 --> 01:10:00.700 align:middle line:84% And that, of course, required some, 01:10:00.700 --> 01:10:06.334 align:middle line:84% as I think I've pointed out, fairly strong assumptions. 01:10:06.334 --> 01:10:10.360 align:middle line:84% In particular, to come back to those, the fact 01:10:10.360 --> 01:10:12.760 align:middle line:84% that there are no information asymmetries 01:10:12.760 --> 01:10:16.090 align:middle line:84% between the borrowers and the lenders 01:10:16.090 --> 01:10:21.100 align:middle line:84% means that search is actually not happening in equilibrium. 01:10:21.100 --> 01:10:26.020 align:middle line:84% So the lenders-- think about the quote of the initial lender. 01:10:26.020 --> 01:10:28.000 align:middle line:84% The lender is always paying consumers such 01:10:28.000 --> 01:10:31.570 align:middle line:84% that they do not find it worth their while to go out and get 01:10:31.570 --> 01:10:33.640 align:middle line:90% additional quotes. 01:10:33.640 --> 01:10:39.160 align:middle line:84% So that means if you want to, for instance, quantify the, 01:10:39.160 --> 01:10:42.340 align:middle line:84% let's say, lost economic surplus due to search costs, 01:10:42.340 --> 01:10:44.710 align:middle line:84% this wouldn't be the right model for that. 01:10:44.710 --> 01:10:51.070 align:middle line:84% And we might think that in some markets, the fact 01:10:51.070 --> 01:10:53.770 align:middle line:84% that some consumers have lower search 01:10:53.770 --> 01:10:57.040 align:middle line:84% costs than others can have disciplined prices in the market 01:10:57.040 --> 01:10:58.638 align:middle line:90% overall. 01:10:58.638 --> 01:11:01.180 align:middle line:84% So you had seen this effect, for instance, in the style model 01:11:01.180 --> 01:11:03.850 align:middle line:84% that if you get more sophisticated consumers 01:11:03.850 --> 01:11:07.910 align:middle line:84% in the market, we move towards more competitive pricing. 01:11:07.910 --> 01:11:10.760 align:middle line:84% And this kind of effect you don't get here 01:11:10.760 --> 01:11:14.000 align:middle line:84% because of this perfect price discrimination where every type 01:11:14.000 --> 01:11:15.055 align:middle line:90% gets kind of his price. 01:11:15.055 --> 01:11:17.734 align:middle line:90% 01:11:17.734 --> 01:11:21.190 align:middle line:84% So I think that's a potential shortcoming 01:11:21.190 --> 01:11:27.860 align:middle line:84% of this model that is kind of interesting to explore. 01:11:27.860 --> 01:11:30.070 align:middle line:84% But I think they had good reasons 01:11:30.070 --> 01:11:31.570 align:middle line:84% to construct it this way, because it 01:11:31.570 --> 01:11:37.640 align:middle line:84% provided this tight link to the empirical specification. 01:11:37.640 --> 01:11:40.750 align:middle line:90% 01:11:40.750 --> 01:11:43.450 align:middle line:84% Any questions or comments about this paper? 01:11:43.450 --> 01:11:44.200 align:middle line:90% Yes. 01:11:44.200 --> 01:11:47.440 align:middle line:84% STUDENT: I guess here the search costs are just 01:11:47.440 --> 01:11:50.110 align:middle line:90% based on each consumer. 01:11:50.110 --> 01:11:54.290 align:middle line:84% There's nothing that the firm does to change the search costs. 01:11:54.290 --> 01:11:56.680 align:middle line:84% But it seems like in a mortgage market, 01:11:56.680 --> 01:11:59.890 align:middle line:84% I could imagine that maybe the firm does less advertising 01:11:59.890 --> 01:12:02.660 align:middle line:84% or I make you walk into my office rather than call me. 01:12:02.660 --> 01:12:06.100 align:middle line:84% And that would be, like, if [INAUDIBLE] strategy 01:12:06.100 --> 01:12:08.410 align:middle line:84% for changing the search cost depending on how 01:12:08.410 --> 01:12:10.210 align:middle line:90% much competition there is. 01:12:10.210 --> 01:12:12.970 align:middle line:84% And so what is the reasoning for trying 01:12:12.970 --> 01:12:16.720 align:middle line:84% to think of firm strategies just in the prices 01:12:16.720 --> 01:12:19.600 align:middle line:84% that they give consumers rather than any way in which they could 01:12:19.600 --> 01:12:22.870 align:middle line:90% be changing the search costs? 01:12:22.870 --> 01:12:25.420 align:middle line:84% TOBIAS SALZ: So first of all, there 01:12:25.420 --> 01:12:28.090 align:middle line:84% are papers mostly in the marketing literature 01:12:28.090 --> 01:12:31.510 align:middle line:84% that actually do what you suggest. 01:12:31.510 --> 01:12:34.750 align:middle line:84% They think about advertising as basically firms 01:12:34.750 --> 01:12:41.030 align:middle line:84% providing a search, a free quote, to consumers. 01:12:41.030 --> 01:12:43.240 align:middle line:84% There's actually a fairly recent paper by Andrea, 01:12:43.240 --> 01:12:46.120 align:middle line:84% Siminoff, and Raluca Ursu, and some others 01:12:46.120 --> 01:12:50.320 align:middle line:84% that do exactly that, where they try to see 01:12:50.320 --> 01:12:56.950 align:middle line:84% how does a free quote or free presentation of a product 01:12:56.950 --> 01:12:59.030 align:middle line:84% influence a consumer's search strategy? 01:12:59.030 --> 01:13:00.730 align:middle line:84% And they do this with online data. 01:13:00.730 --> 01:13:02.505 align:middle line:84% So people have thought about this. 01:13:02.505 --> 01:13:06.190 align:middle line:90% 01:13:06.190 --> 01:13:09.610 align:middle line:90% And I think it's interesting. 01:13:09.610 --> 01:13:14.590 align:middle line:84% I think in part, it's tractability. 01:13:14.590 --> 01:13:16.050 align:middle line:90% But yeah, that's a good thought. 01:13:16.050 --> 01:13:19.550 align:middle line:90% 01:13:19.550 --> 01:13:21.480 align:middle line:90% Any other questions about this? 01:13:21.480 --> 01:13:28.190 align:middle line:90% 01:13:28.190 --> 01:13:31.910 align:middle line:84% So I'll probably not get through all of this, 01:13:31.910 --> 01:13:35.000 align:middle line:84% but I just want to very briefly talk about my job market 01:13:35.000 --> 01:13:41.690 align:middle line:84% paper, which ties in quite nicely with this previous paper 01:13:41.690 --> 01:13:44.210 align:middle line:84% that we discussed, because in some sense, one 01:13:44.210 --> 01:13:46.010 align:middle line:84% of the kind of economic differences 01:13:46.010 --> 01:13:53.090 align:middle line:84% that I'm after here is precisely getting at these spillover 01:13:53.090 --> 01:13:55.910 align:middle line:84% effects of having more sophisticated consumers 01:13:55.910 --> 01:13:56.830 align:middle line:90% in the market. 01:13:56.830 --> 01:14:00.410 align:middle line:84% So again, the idea, which is, for instance, 01:14:00.410 --> 01:14:04.520 align:middle line:84% in the Salop-Stiglitz model also, 01:14:04.520 --> 01:14:08.330 align:middle line:84% that when there are sophisticated consumers 01:14:08.330 --> 01:14:10.440 align:middle line:84% in the market, both sophisticated and 01:14:10.440 --> 01:14:12.440 align:middle line:84% unsophisticated ones, and firms can actually not 01:14:12.440 --> 01:14:14.120 align:middle line:90% tell them apart. 01:14:14.120 --> 01:14:17.690 align:middle line:84% Then the fact that some consumers are searching 01:14:17.690 --> 01:14:20.120 align:middle line:84% disciplines the firm's pricing, because they 01:14:20.120 --> 01:14:22.587 align:middle line:84% don't know which kinds of consumers they're dealing with. 01:14:22.587 --> 01:14:24.170 align:middle line:84% And this effect was kind of completely 01:14:24.170 --> 01:14:27.160 align:middle line:90% shut down in the previous paper. 01:14:27.160 --> 01:14:30.390 align:middle line:84% And so in particular, I want to look 01:14:30.390 --> 01:14:36.150 align:middle line:84% at this effect with an application to intermediaries. 01:14:36.150 --> 01:14:38.940 align:middle line:84% So intermediaries are in this setting firms 01:14:38.940 --> 01:14:41.640 align:middle line:90% that help buyers search. 01:14:41.640 --> 01:14:45.000 align:middle line:84% And if you think about who are the buyers that 01:14:45.000 --> 01:14:48.757 align:middle line:84% would hire somebody else to conduct their search for them, 01:14:48.757 --> 01:14:51.090 align:middle line:84% it's going to be buyers that have very, very high search 01:14:51.090 --> 01:14:52.410 align:middle line:90% costs. 01:14:52.410 --> 01:14:54.180 align:middle line:84% So you could construct a model where 01:14:54.180 --> 01:14:58.350 align:middle line:84% there's some sorting very closely related to the Hong 01:14:58.350 --> 01:15:01.020 align:middle line:84% and Shum model that I showed you, where you have basically 01:15:01.020 --> 01:15:03.640 align:middle line:90% these different consumer bins. 01:15:03.640 --> 01:15:05.130 align:middle line:84% Consumers with higher search costs 01:15:05.130 --> 01:15:07.800 align:middle line:84% take more draws from the price distribution. 01:15:07.800 --> 01:15:10.030 align:middle line:84% Sorry, take fewer draws from the price distribution. 01:15:10.030 --> 01:15:11.655 align:middle line:84% Once your search cost becomes too high, 01:15:11.655 --> 01:15:14.130 align:middle line:90% you instead use an intermediary. 01:15:14.130 --> 01:15:18.450 align:middle line:84% And so what we get out of this is 01:15:18.450 --> 01:15:21.210 align:middle line:84% that the highest search cost types 01:15:21.210 --> 01:15:23.430 align:middle line:90% are kind of leaving the market. 01:15:23.430 --> 01:15:25.710 align:middle line:84% They're going instead through the intermediary. 01:15:25.710 --> 01:15:33.260 align:middle line:84% And firms in their pricing will kind of condition on the fact 01:15:33.260 --> 01:15:36.470 align:middle line:84% that in equilibrium, they're not encountering the very highest 01:15:36.470 --> 01:15:39.310 align:middle line:90% search cost types. 01:15:39.310 --> 01:15:43.540 align:middle line:84% So again, think about the Hong and Shum search-cost 01:15:43.540 --> 01:15:44.290 align:middle line:90% distribution. 01:15:44.290 --> 01:15:47.840 align:middle line:84% You have at the very top the people that leave the market. 01:15:47.840 --> 01:15:50.410 align:middle line:84% And that could lead to lower prices for all consumers 01:15:50.410 --> 01:15:52.390 align:middle line:90% in the market. 01:15:52.390 --> 01:15:54.760 align:middle line:84% And so this, of course, only works 01:15:54.760 --> 01:15:57.520 align:middle line:84% if you now do have this private information, if firms do not 01:15:57.520 --> 01:15:59.380 align:middle line:90% know whom they're dealing with. 01:15:59.380 --> 01:16:03.190 align:middle line:84% And that introduces some technical challenges 01:16:03.190 --> 01:16:05.980 align:middle line:90% in how to deal with that. 01:16:05.980 --> 01:16:09.070 align:middle line:84% And so basically what I do in this paper 01:16:09.070 --> 01:16:12.790 align:middle line:84% is construct a model that essentially puts together 01:16:12.790 --> 01:16:19.390 align:middle line:84% the Hong and Shum search strategy of consumers, where 01:16:19.390 --> 01:16:21.190 align:middle line:84% consumers have heterogeneous search costs 01:16:21.190 --> 01:16:23.190 align:middle line:84% and decide how many draws to take from the price 01:16:23.190 --> 01:16:26.080 align:middle line:84% distribution, with a model of suppliers 01:16:26.080 --> 01:16:30.320 align:middle line:84% that also have heterogeneous costs for providing services. 01:16:30.320 --> 01:16:32.980 align:middle line:84% And in these kinds of settings, you now 01:16:32.980 --> 01:16:36.760 align:middle line:84% have unobserved source of variation 01:16:36.760 --> 01:16:39.640 align:middle line:90% on both sides of the market. 01:16:39.640 --> 01:16:44.830 align:middle line:84% So again, one contribution is here 01:16:44.830 --> 01:16:49.540 align:middle line:84% to construct a model for these kinds of decentralized markets 01:16:49.540 --> 01:16:52.720 align:middle line:84% where firms don't know what type of consumer 01:16:52.720 --> 01:16:53.890 align:middle line:90% they're dealing with. 01:16:53.890 --> 01:16:56.920 align:middle line:84% And one data contribution is to find a setting 01:16:56.920 --> 01:17:01.250 align:middle line:84% where you actually find prices for decentralized markets. 01:17:01.250 --> 01:17:04.610 align:middle line:84% So typically good settings are those that are regulated. 01:17:04.610 --> 01:17:08.883 align:middle line:84% So privacy, in the case of the mortgage paper. 01:17:08.883 --> 01:17:10.300 align:middle line:84% The reason that they had this data 01:17:10.300 --> 01:17:12.870 align:middle line:84% is because people needed to get insurance. 01:17:12.870 --> 01:17:15.550 align:middle line:84% So they collected what is essentially administrative data, 01:17:15.550 --> 01:17:21.860 align:middle line:84% because the lenders had to submit this to a regulator. 01:17:21.860 --> 01:17:24.430 align:middle line:84% And so a similar thing in my setting 01:17:24.430 --> 01:17:27.910 align:middle line:84% here where I'm going to use data from the New York City trade 01:17:27.910 --> 01:17:31.090 align:middle line:84% waste market, which were very idiosyncratic reasons, which 01:17:31.090 --> 01:17:34.870 align:middle line:84% is basically organized crime in the city of new York 01:17:34.870 --> 01:17:37.398 align:middle line:90% was starting to be regulated. 01:17:37.398 --> 01:17:39.190 align:middle line:84% And all the firms have to submit their data 01:17:39.190 --> 01:17:41.610 align:middle line:90% to a central regulator. 01:17:41.610 --> 01:17:47.360 align:middle line:84% And so you can actually see individual level contract data 01:17:47.360 --> 01:17:49.460 align:middle line:84% that allow you to make some progress 01:17:49.460 --> 01:17:51.070 align:middle line:90% on these kinds of markets. 01:17:51.070 --> 01:17:54.510 align:middle line:90% 01:17:54.510 --> 01:17:59.060 align:middle line:84% OK, so this is the non-residential waste disposal 01:17:59.060 --> 01:17:59.560 align:middle line:90% market. 01:17:59.560 --> 01:18:02.340 align:middle line:84% So trade waste market is a fancy name for that. 01:18:02.340 --> 01:18:05.280 align:middle line:84% And so what is really striking about this market 01:18:05.280 --> 01:18:08.700 align:middle line:84% is that there are a lot of sellers in New York City 01:18:08.700 --> 01:18:11.620 align:middle line:84% relative to most other cities in the US. 01:18:11.620 --> 01:18:15.870 align:middle line:84% So in most other cities, you have large companies 01:18:15.870 --> 01:18:17.670 align:middle line:84% like waste management and others that 01:18:17.670 --> 01:18:19.830 align:middle line:90% are supplying the entire city. 01:18:19.830 --> 01:18:25.450 align:middle line:84% Here you have about 100 sellers in some zip codes of Manhattan. 01:18:25.450 --> 01:18:29.670 align:middle line:90% You can call up to 35 carters. 01:18:29.670 --> 01:18:31.780 align:middle line:84% So there seems to be this very fragmented 01:18:31.780 --> 01:18:36.103 align:middle line:84% and, from an outside observers point of view, redundant supply. 01:18:36.103 --> 01:18:42.220 align:middle line:90% 01:18:42.220 --> 01:18:44.740 align:middle line:84% The market is regulated by the Business Integrity 01:18:44.740 --> 01:18:49.090 align:middle line:84% Commission, which was established by the city in 1995. 01:18:49.090 --> 01:18:53.530 align:middle line:84% And again, was a response to the property and racketeering 01:18:53.530 --> 01:18:55.750 align:middle line:84% system that has been in place in New York. 01:18:55.750 --> 01:19:00.250 align:middle line:84% Famously, many crime shows and movies, that the waste market 01:19:00.250 --> 01:19:04.240 align:middle line:84% was one of the predominant industries where 01:19:04.240 --> 01:19:05.750 align:middle line:90% organized crime was active. 01:19:05.750 --> 01:19:10.510 align:middle line:84% And that led the city to actually collect all this data. 01:19:10.510 --> 01:19:12.370 align:middle line:84% So I want to use my last few minutes 01:19:12.370 --> 01:19:15.490 align:middle line:84% to just show you quickly the model 01:19:15.490 --> 01:19:18.640 align:middle line:84% and how it works just so you get a sense. 01:19:18.640 --> 01:19:21.850 align:middle line:84% It's a bit fast, but just to get you the idea of what 01:19:21.850 --> 01:19:23.860 align:middle line:90% is different about this model. 01:19:23.860 --> 01:19:27.100 align:middle line:90% So here again, we have buyers. 01:19:27.100 --> 01:19:28.750 align:middle line:84% These are businesses in New York City 01:19:28.750 --> 01:19:31.790 align:middle line:84% that need to contract for their waste disposal services. 01:19:31.790 --> 01:19:32.810 align:middle line:90% We have sellers. 01:19:32.810 --> 01:19:35.420 align:middle line:90% These are carting companies. 01:19:35.420 --> 01:19:37.900 align:middle line:84% And then we have brokers in the middle. 01:19:37.900 --> 01:19:39.610 align:middle line:84% And some buyers could decide to delegate 01:19:39.610 --> 01:19:42.760 align:middle line:84% to a broker and the broker instead searches for them, 01:19:42.760 --> 01:19:45.730 align:middle line:84% but by searching here means it's an auction. 01:19:45.730 --> 01:19:48.760 align:middle line:84% And so what is challenging about this model 01:19:48.760 --> 01:19:53.770 align:middle line:84% is that we have basically two sources of unobserved variation 01:19:53.770 --> 01:19:56.740 align:middle line:84% that gives rise to prices, which are unobserved 01:19:56.740 --> 01:20:00.760 align:middle line:84% costs of businesses to provide service and the unobserved 01:20:00.760 --> 01:20:05.110 align:middle line:84% search costs of buyers to search, 01:20:05.110 --> 01:20:08.380 align:middle line:90% which are here called kappa. 01:20:08.380 --> 01:20:11.290 align:middle line:84% And so if you, for instance, contrast this 01:20:11.290 --> 01:20:14.050 align:middle line:84% with the Hong and Shum model, in Hong and Shum, 01:20:14.050 --> 01:20:17.410 align:middle line:84% the cost of sellers was assumed to be constant. 01:20:17.410 --> 01:20:22.240 align:middle line:84% And the heterogeneity that you observe in prices basically 01:20:22.240 --> 01:20:26.000 align:middle line:84% maps to an unobserved distribution of search costs. 01:20:26.000 --> 01:20:31.110 align:middle line:84% So here you have to disentangle the two. 01:20:31.110 --> 01:20:33.210 align:middle line:90% So let me skip some of this. 01:20:33.210 --> 01:20:34.540 align:middle line:90% I told you about this. 01:20:34.540 --> 01:20:37.890 align:middle line:84% I just want to briefly show you how the model works. 01:20:37.890 --> 01:20:45.000 align:middle line:84% So we have buyers that have heterogeneous search expenses, 01:20:45.000 --> 01:20:46.620 align:middle line:84% and they can, like in Hong and Shum, 01:20:46.620 --> 01:20:49.710 align:middle line:84% decide how many draws to take from the price distribution 01:20:49.710 --> 01:20:54.010 align:middle line:84% or whether or not they're using a broker instead. 01:20:54.010 --> 01:20:56.010 align:middle line:84% And so because you know the Hong and Shum model, 01:20:56.010 --> 01:20:57.570 align:middle line:90% I can skip ahead a bit here. 01:20:57.570 --> 01:21:00.930 align:middle line:84% So you get, again, a partition of the search cost distribution, 01:21:00.930 --> 01:21:03.405 align:middle line:84% but those with the very highest cost, they're 01:21:03.405 --> 01:21:04.655 align:middle line:90% going to use a broker instead. 01:21:04.655 --> 01:21:11.330 align:middle line:90% 01:21:11.330 --> 01:21:13.640 align:middle line:84% So as before, we want to keep track 01:21:13.640 --> 01:21:16.550 align:middle line:84% of buyers that are the mass of buyers 01:21:16.550 --> 01:21:17.820 align:middle line:90% between these marginal types. 01:21:17.820 --> 01:21:19.620 align:middle line:84% This gives rise to these search weights. 01:21:19.620 --> 01:21:22.037 align:middle line:84% So these are the buyers that search m times from the price 01:21:22.037 --> 01:21:25.304 align:middle line:90% distribution. 01:21:25.304 --> 01:21:30.020 align:middle line:84% And so because the seller's cost is also heterogeneous, 01:21:30.020 --> 01:21:32.930 align:middle line:84% we don't get a mixed strategy pricing equilibrium, 01:21:32.930 --> 01:21:35.180 align:middle line:84% but we instead get for any different cost 01:21:35.180 --> 01:21:38.570 align:middle line:84% that a seller could draw, we get a bidding 01:21:38.570 --> 01:21:43.795 align:middle line:84% function that maps the seller's unobserved cost to a price. 01:21:43.795 --> 01:21:47.840 align:middle line:90% 01:21:47.840 --> 01:21:50.540 align:middle line:84% And so just to show you this briefly, 01:21:50.540 --> 01:21:55.070 align:middle line:84% I know I'm going fast here, but we now get a profit maximization 01:21:55.070 --> 01:21:56.850 align:middle line:90% problem that looks like this. 01:21:56.850 --> 01:21:59.720 align:middle line:84% The sellers are maximizing for any given consumer 01:21:59.720 --> 01:22:03.860 align:middle line:84% that contacts them their margin times the weighted average 01:22:03.860 --> 01:22:06.050 align:middle line:84% across different consumers that they could encounter 01:22:06.050 --> 01:22:07.425 align:middle line:84% would take n draws from the price 01:22:07.425 --> 01:22:10.460 align:middle line:84% distribution times the probability 01:22:10.460 --> 01:22:12.830 align:middle line:84% that they're selling to that consumer. 01:22:12.830 --> 01:22:15.530 align:middle line:84% Now, what is different here is that these marginal costs are 01:22:15.530 --> 01:22:18.710 align:middle line:84% not uniform as before as in Hong and Shum, 01:22:18.710 --> 01:22:21.140 align:middle line:84% but they're drawn from a distribution. 01:22:21.140 --> 01:22:24.650 align:middle line:84% And so what's nice about thinking about this problem 01:22:24.650 --> 01:22:30.770 align:middle line:84% is that this actually leads to a closed form 01:22:30.770 --> 01:22:33.830 align:middle line:84% solution for the bidding function that 01:22:33.830 --> 01:22:38.630 align:middle line:84% is like in a first price procurement auction. 01:22:38.630 --> 01:22:40.910 align:middle line:84% I know we'll do auctions later in this semester, 01:22:40.910 --> 01:22:44.780 align:middle line:84% but this is a first price procurement auction 01:22:44.780 --> 01:22:49.030 align:middle line:90% solution to an IPV problem. 01:22:49.030 --> 01:22:51.480 align:middle line:84% And this is essentially just a weighted average 01:22:51.480 --> 01:22:56.100 align:middle line:84% across what you would bid if the number of competitors 01:22:56.100 --> 01:22:56.718 align:middle line:90% were known. 01:22:56.718 --> 01:22:58.260 align:middle line:84% And so here the number of competitors 01:22:58.260 --> 01:23:00.230 align:middle line:84% depends on the consumer search strategy. 01:23:00.230 --> 01:23:03.300 align:middle line:84% So just to wrap up and put the pieces together, 01:23:03.300 --> 01:23:06.360 align:middle line:84% so basically this kind of takes the Hong and Shum model 01:23:06.360 --> 01:23:11.193 align:middle line:84% and then introduces heterogeneity on the supply side 01:23:11.193 --> 01:23:13.110 align:middle line:84% and kind of maps this essentially back to what 01:23:13.110 --> 01:23:14.940 align:middle line:90% is an auction problem. 01:23:14.940 --> 01:23:17.140 align:middle line:90% And I'll leave it here. 01:23:17.140 --> 01:23:21.380 align:middle line:84% I'll probably tell those who come to [14.]273 some more about 01:23:21.380 --> 01:23:21.880 align:middle line:90% this. 01:23:21.880 --> 01:23:24.290 align:middle line:90% But yeah, that's it for today. 01:23:24.290 --> 01:23:25.970 align:middle line:90% Thank you. 01:23:25.970 --> 01:23:38.000 align:middle line:90%