WEBVTT 00:00:00.000 --> 00:00:01.968 align:middle line:90% [SQUEAKING] 00:00:01.968 --> 00:00:03.444 align:middle line:90% [RUSTLING] 00:00:03.444 --> 00:00:04.920 align:middle line:90% [CLICKING] 00:00:04.920 --> 00:00:16.250 align:middle line:90% 00:00:16.250 --> 00:00:18.397 align:middle line:84% GLENN ELLISON: On bounded rationality, 00:00:18.397 --> 00:00:20.480 align:middle line:84% I told you there are several different ways people 00:00:20.480 --> 00:00:22.770 align:middle line:84% do boundedly rational models last class. 00:00:22.770 --> 00:00:24.920 align:middle line:84% And the one I emphasized was this sort 00:00:24.920 --> 00:00:26.690 align:middle line:84% of psychology and economics style, 00:00:26.690 --> 00:00:29.960 align:middle line:84% where you take one of these rationality-like models 00:00:29.960 --> 00:00:32.990 align:middle line:84% you see in behavioral economics classes, 00:00:32.990 --> 00:00:35.720 align:middle line:84% and you think about what happens when consumers behave like that 00:00:35.720 --> 00:00:38.600 align:middle line:90% and firms exploit them. 00:00:38.600 --> 00:00:41.120 align:middle line:84% But we say, to motivate the rule of thumb approach, 00:00:41.120 --> 00:00:43.760 align:middle line:84% I guess one philosophical question of like, why is it 00:00:43.760 --> 00:00:47.660 align:middle line:84% we ever even do economics with rational models to begin with? 00:00:47.660 --> 00:00:51.020 align:middle line:84% When we have-- you write down some rational model, 00:00:51.020 --> 00:00:53.000 align:middle line:84% and then you have some very complicated-- 00:00:53.000 --> 00:00:55.167 align:middle line:84% if you're going to publish a paper in a top journal, 00:00:55.167 --> 00:00:57.470 align:middle line:84% you've got very complicated mathematical argument 00:00:57.470 --> 00:00:59.960 align:middle line:84% of eight pages of calculations proving what happens 00:00:59.960 --> 00:01:01.560 align:middle line:90% under this set of assumptions. 00:01:01.560 --> 00:01:04.709 align:middle line:84% And then you say, therefore this happens in this industry. 00:01:04.709 --> 00:01:07.100 align:middle line:84% And your question is like, why do we believe 00:01:07.100 --> 00:01:08.390 align:middle line:90% that in the first place? 00:01:08.390 --> 00:01:09.770 align:middle line:84% There's some parts of rationality 00:01:09.770 --> 00:01:11.478 align:middle line:84% that we really do believe, that if people 00:01:11.478 --> 00:01:14.020 align:middle line:84% can do something that's better for themselves, they'll do it. 00:01:14.020 --> 00:01:15.390 align:middle line:90% And they have some preferences. 00:01:15.390 --> 00:01:17.980 align:middle line:84% They have some goals they're trying to achieve. 00:01:17.980 --> 00:01:20.110 align:middle line:84% You can't keep fooling them over and over again. 00:01:20.110 --> 00:01:21.700 align:middle line:90% They will catch on to things. 00:01:21.700 --> 00:01:24.450 align:middle line:84% But the full rationality, I think, 00:01:24.450 --> 00:01:26.020 align:middle line:90% we don't really believe it. 00:01:26.020 --> 00:01:28.260 align:middle line:84% And the same with these behavioral models. 00:01:28.260 --> 00:01:30.720 align:middle line:84% Like, OK, we believe people are short sighted, 00:01:30.720 --> 00:01:33.420 align:middle line:84% but do we really believe the beta, beta hat, 00:01:33.420 --> 00:01:35.320 align:middle line:90% delta, quasi-hyperbolic model? 00:01:35.320 --> 00:01:37.180 align:middle line:84% Why not any other functional form? 00:01:37.180 --> 00:01:40.230 align:middle line:84% And if there are people are being fooled, 00:01:40.230 --> 00:01:42.715 align:middle line:84% is that going to keep working on them in that model, 00:01:42.715 --> 00:01:45.090 align:middle line:84% or are they going to change the way they behave when they 00:01:45.090 --> 00:01:47.100 align:middle line:84% realize that their life is going badly because 00:01:47.100 --> 00:01:48.220 align:middle line:90% of something that they do? 00:01:48.220 --> 00:01:51.120 align:middle line:84% And are they going to introspect and change it? 00:01:51.120 --> 00:01:52.810 align:middle line:84% And so, I mean, I think one answer 00:01:52.810 --> 00:01:55.840 align:middle line:84% to why we do that is it does prevent you from cheating. 00:01:55.840 --> 00:01:58.440 align:middle line:84% It prevents you from just assuming people do this 00:01:58.440 --> 00:02:00.880 align:middle line:84% and then getting your conclusion right away. 00:02:00.880 --> 00:02:03.120 align:middle line:84% But you can still cheat with rational modeling. 00:02:03.120 --> 00:02:05.070 align:middle line:84% And you just, when you're writing down the preferences 00:02:05.070 --> 00:02:06.300 align:middle line:84% and you're writing down the cost functions 00:02:06.300 --> 00:02:08.310 align:middle line:84% and you're writing down the other things 00:02:08.310 --> 00:02:10.800 align:middle line:84% and the information, you slip in a lot of extra assumptions 00:02:10.800 --> 00:02:14.760 align:middle line:84% that, in some sense, let you get to your conclusion anyway. 00:02:14.760 --> 00:02:19.110 align:middle line:84% And so then maybe, sociology of economics, 00:02:19.110 --> 00:02:22.070 align:middle line:84% one of the reasons we do that is because to publish a paper, 00:02:22.070 --> 00:02:24.350 align:middle line:84% you need to have eight pages of complicated-seeming 00:02:24.350 --> 00:02:26.810 align:middle line:84% complications to show that you've accomplished something 00:02:26.810 --> 00:02:28.850 align:middle line:84% hard in your paper, and then that 00:02:28.850 --> 00:02:31.500 align:middle line:90% rationality lets you do that. 00:02:31.500 --> 00:02:37.250 align:middle line:84% But then the question is, is it really a good model 00:02:37.250 --> 00:02:39.680 align:middle line:84% if it takes a lot of pages of complicated calculations 00:02:39.680 --> 00:02:40.770 align:middle line:90% to find out what happens? 00:02:40.770 --> 00:02:42.312 align:middle line:84% Because that's when we don't really-- 00:02:42.312 --> 00:02:44.630 align:middle line:84% all the minor things that depart from rationality 00:02:44.630 --> 00:02:47.367 align:middle line:84% don't matter or could have an effect. 00:02:47.367 --> 00:02:48.950 align:middle line:84% And so what the rule-of-thumb approach 00:02:48.950 --> 00:02:53.627 align:middle line:84% tends to do is say that, rather than writing down 00:02:53.627 --> 00:02:56.210 align:middle line:84% a rational model and sticking with it, even though it's really 00:02:56.210 --> 00:02:58.160 align:middle line:84% hard to solve and we may not believe it, 00:02:58.160 --> 00:03:01.310 align:middle line:84% why don't we just write down a model that we think 00:03:01.310 --> 00:03:04.250 align:middle line:84% is more like what-- we think is kind of like what consumers do, 00:03:04.250 --> 00:03:07.310 align:middle line:84% and we're happy to buy that as an approximation that's 00:03:07.310 --> 00:03:08.540 align:middle line:90% easy to solve. 00:03:08.540 --> 00:03:11.630 align:middle line:84% And so instead of thinking of the assumptions on information 00:03:11.630 --> 00:03:14.150 align:middle line:84% and costs and whatever and epsilon 00:03:14.150 --> 00:03:17.630 align:middle line:84% ijk as being logit distributed, let's 00:03:17.630 --> 00:03:19.052 align:middle line:84% just go straight to the behavior, 00:03:19.052 --> 00:03:21.260 align:middle line:84% write down a behavior that makes the model tractable, 00:03:21.260 --> 00:03:23.610 align:middle line:84% and let us describe what happens. 00:03:23.610 --> 00:03:26.690 align:middle line:90% So I'll start with an example. 00:03:26.690 --> 00:03:34.050 align:middle line:84% Smallwood and Conlisk is one of my favorite classic economics 00:03:34.050 --> 00:03:38.820 align:middle line:84% papers that didn't get taught when I took IO. 00:03:38.820 --> 00:03:41.220 align:middle line:84% But Smallwood and Conlisk start with-- 00:03:41.220 --> 00:03:46.590 align:middle line:84% they have a question they're starting from of, 00:03:46.590 --> 00:03:48.570 align:middle line:84% are superior products going to win out 00:03:48.570 --> 00:03:49.830 align:middle line:90% in the long run in the market? 00:03:49.830 --> 00:03:53.790 align:middle line:84% And do superior products take over markets and outcompete 00:03:53.790 --> 00:03:55.683 align:middle line:84% inferior products in a world where 00:03:55.683 --> 00:03:57.600 align:middle line:84% people don't know the quality of every product 00:03:57.600 --> 00:03:59.670 align:middle line:90% when it's first invented? 00:03:59.670 --> 00:04:03.720 align:middle line:84% And so they use the example of, suppose you have someone buying 00:04:03.720 --> 00:04:04.523 align:middle line:90% a box of cereal. 00:04:04.523 --> 00:04:06.690 align:middle line:84% And they're standing here looking at the supermarket 00:04:06.690 --> 00:04:09.870 align:middle line:84% shelf, and they're trying to decide what to buy. 00:04:09.870 --> 00:04:12.060 align:middle line:84% Obviously, they've only tasted some small fraction 00:04:12.060 --> 00:04:12.870 align:middle line:90% of those cereals. 00:04:12.870 --> 00:04:14.412 align:middle line:84% And so they have to go there and look 00:04:14.412 --> 00:04:16.290 align:middle line:84% at the prices of some subset of them 00:04:16.290 --> 00:04:20.519 align:middle line:84% and think about how much they're going to like them and will 00:04:20.519 --> 00:04:21.670 align:middle line:90% they like it or not. 00:04:21.670 --> 00:04:24.040 align:middle line:84% If I buy it, what's the probability I'll like it? 00:04:24.040 --> 00:04:26.677 align:middle line:84% Will I like it enough that I'll want to buy it again? 00:04:26.677 --> 00:04:28.260 align:middle line:84% And then how much extra utility will I 00:04:28.260 --> 00:04:30.052 align:middle line:84% get from the future purchase of that cereal 00:04:30.052 --> 00:04:31.990 align:middle line:84% now that I have all the extra information? 00:04:31.990 --> 00:04:33.782 align:middle line:84% And they said, well, what's even much worse 00:04:33.782 --> 00:04:36.510 align:middle line:84% is imagining that, when you're going to buy the box of cereal, 00:04:36.510 --> 00:04:39.150 align:middle line:84% you're just walking up to the shelf, 00:04:39.150 --> 00:04:44.058 align:middle line:84% and you see that person in front of you buying a box of cereal. 00:04:44.058 --> 00:04:45.600 align:middle line:84% And then what are you supposed to do? 00:04:45.600 --> 00:04:47.475 align:middle line:84% You're supposed to say, OK, well, that person 00:04:47.475 --> 00:04:48.730 align:middle line:90% is buying this box of cereal. 00:04:48.730 --> 00:04:50.790 align:middle line:84% If they bought that box of cereal, 00:04:50.790 --> 00:04:52.570 align:middle line:90% they must like that cereal. 00:04:52.570 --> 00:04:54.870 align:middle line:84% And so if they like that cereal and then our epsilon-- 00:04:54.870 --> 00:04:59.215 align:middle line:84% these things are-- there's a xi j and an epsilon ijk. 00:04:59.215 --> 00:05:01.590 align:middle line:84% The xi j must be high for that cereal conditional on that 00:05:01.590 --> 00:05:02.830 align:middle line:90% person buying it. 00:05:02.830 --> 00:05:04.560 align:middle line:84% But then you think, well, wait, no, 00:05:04.560 --> 00:05:07.260 align:middle line:84% this person might be buying it because he's tasted before, 00:05:07.260 --> 00:05:09.000 align:middle line:84% or that person might be buying it 00:05:09.000 --> 00:05:14.168 align:middle line:84% because, two days earlier, that person was sitting 00:05:14.168 --> 00:05:15.960 align:middle line:84% on their couch at home, and there was a TV, 00:05:15.960 --> 00:05:18.480 align:middle line:84% and they were watching a TV, and there was a commercial. 00:05:18.480 --> 00:05:21.960 align:middle line:84% And they saw the TV commercial for that cereal. 00:05:21.960 --> 00:05:25.440 align:middle line:84% But if they did see the TV commercial, 00:05:25.440 --> 00:05:28.300 align:middle line:84% what's the signaling value of the TV commercial that they saw? 00:05:28.300 --> 00:05:30.540 align:middle line:84% Because the firm wouldn't have advertised the cereal 00:05:30.540 --> 00:05:32.832 align:middle line:84% unless it knew people were going to repeat purchase it. 00:05:32.832 --> 00:05:35.680 align:middle line:84% So maybe I have to have a prior over the cost 00:05:35.680 --> 00:05:38.290 align:middle line:84% of that cereal and the reasons why the TV commercial would 00:05:38.290 --> 00:05:39.490 align:middle line:90% have been on. 00:05:39.490 --> 00:05:41.080 align:middle line:84% But then again, maybe this person 00:05:41.080 --> 00:05:43.810 align:middle line:84% is buying it because the person who was two people before them 00:05:43.810 --> 00:05:46.600 align:middle line:84% in the aisle bought it and they saw the person before take it 00:05:46.600 --> 00:05:48.790 align:middle line:84% off that aisle, or maybe they're just buying it 00:05:48.790 --> 00:05:50.805 align:middle line:90% because it's at eye level. 00:05:50.805 --> 00:05:52.180 align:middle line:84% But if it's at eye level, there's 00:05:52.180 --> 00:05:56.287 align:middle line:84% this signaling problem of, why is it at eye level? 00:05:56.287 --> 00:05:58.120 align:middle line:84% Is it eye level because it's really popular, 00:05:58.120 --> 00:06:01.720 align:middle line:84% or is it at eye level because the firm paid the store 00:06:01.720 --> 00:06:02.750 align:middle line:90% to put it at eye level? 00:06:02.750 --> 00:06:04.420 align:middle line:84% And would they have paid more or less to put it at eye level 00:06:04.420 --> 00:06:05.890 align:middle line:90% if it was good or bad? 00:06:05.890 --> 00:06:08.110 align:middle line:84% But especially if this person saw the person before, 00:06:08.110 --> 00:06:11.950 align:middle line:84% it's almost like I need this prior over all possible 00:06:11.950 --> 00:06:15.070 align:middle line:84% purchases by every person since the invention of cereal 00:06:15.070 --> 00:06:18.700 align:middle line:84% in the 19th century and who saw what and what they did and what 00:06:18.700 --> 00:06:20.210 align:middle line:90% that did to the popularity. 00:06:20.210 --> 00:06:23.860 align:middle line:84% And then I've got to have that prior over that enormous network 00:06:23.860 --> 00:06:25.750 align:middle line:84% of a century's worth of cereal purchases, 00:06:25.750 --> 00:06:28.060 align:middle line:84% and then invert it and figure out, now 00:06:28.060 --> 00:06:29.870 align:middle line:84% what's my posterior on that cereal box? 00:06:29.870 --> 00:06:33.340 align:middle line:84% And Smallwood and Conlisk's view is that this is insane. 00:06:33.340 --> 00:06:34.740 align:middle line:90% This is not what people do. 00:06:34.740 --> 00:06:35.540 align:middle line:90% People walk up. 00:06:35.540 --> 00:06:37.080 align:middle line:90% They look at the cereal box. 00:06:37.080 --> 00:06:38.940 align:middle line:84% They're buying 50 different things at the supermarket. 00:06:38.940 --> 00:06:40.010 align:middle line:84% Most of the time, they just take off 00:06:40.010 --> 00:06:42.030 align:middle line:84% the shelf whatever they normally take off the shelf. 00:06:42.030 --> 00:06:44.697 align:middle line:84% And then every once in a while, they've got a little extra time, 00:06:44.697 --> 00:06:46.640 align:middle line:84% and they put a little bit of time 00:06:46.640 --> 00:06:48.650 align:middle line:84% into thinking, which one do I buy? 00:06:48.650 --> 00:06:53.660 align:middle line:84% And so why don't we think about consumers more like that 00:06:53.660 --> 00:06:58.290 align:middle line:84% and ask about, then what would happen? 00:06:58.290 --> 00:07:00.090 align:middle line:84% Would good cereal brands take over, 00:07:00.090 --> 00:07:03.320 align:middle line:84% or do we think that Cheerios is really popular even 00:07:03.320 --> 00:07:08.090 align:middle line:84% though Cheerios isn't very good or whatever it is? 00:07:08.090 --> 00:07:11.330 align:middle line:84% So here's what Smallwood and Conlisk did 00:07:11.330 --> 00:07:14.000 align:middle line:84% and this idea of just trying to write down-- 00:07:14.000 --> 00:07:17.150 align:middle line:84% let's write down a dynamic model that seems reasonable to us 00:07:17.150 --> 00:07:19.970 align:middle line:84% and talk about, what's the long-run implications 00:07:19.970 --> 00:07:22.430 align:middle line:84% in terms of, do superior brands take over 00:07:22.430 --> 00:07:26.810 align:middle line:84% when consumers are fairly naive in their shopping? 00:07:26.810 --> 00:07:29.630 align:middle line:84% And I think they very much do believe that consumers 00:07:29.630 --> 00:07:35.360 align:middle line:84% are naive when they shop and that this is a better 00:07:35.360 --> 00:07:37.200 align:middle line:90% model than the rational one. 00:07:37.200 --> 00:07:38.750 align:middle line:90% So anyway, here's their model. 00:07:38.750 --> 00:07:42.240 align:middle line:90% Products differ in quality. 00:07:42.240 --> 00:07:45.170 align:middle line:84% A consumer of product k has a bad experience, 00:07:45.170 --> 00:07:47.552 align:middle line:84% which they'll call a breakdown with probability bk. 00:07:47.552 --> 00:07:49.760 align:middle line:84% So if you're talking about people buying light bulbs, 00:07:49.760 --> 00:07:51.385 align:middle line:84% that could be literally the probability 00:07:51.385 --> 00:07:54.200 align:middle line:84% that the light bulb burns out in any given-- 00:07:54.200 --> 00:07:55.910 align:middle line:90% between any given shopping trip. 00:07:55.910 --> 00:07:58.497 align:middle line:84% If it's a cereal, it just could be that every once in a while, 00:07:58.497 --> 00:08:00.080 align:middle line:84% you're eating your cereal, and it just 00:08:00.080 --> 00:08:01.370 align:middle line:84% occurs to you in the morning, you know, 00:08:01.370 --> 00:08:02.580 align:middle line:90% this really isn't all that good. 00:08:02.580 --> 00:08:04.163 align:middle line:84% Maybe I should buy a different cereal. 00:08:04.163 --> 00:08:06.560 align:middle line:84% But products differ in the probability 00:08:06.560 --> 00:08:08.310 align:middle line:90% of giving a bad experience. 00:08:08.310 --> 00:08:09.957 align:middle line:84% And so the highest quality products 00:08:09.957 --> 00:08:11.290 align:middle line:90% are the ones with the lowest bk. 00:08:11.290 --> 00:08:14.220 align:middle line:90% 00:08:14.220 --> 00:08:18.690 align:middle line:84% What consumers do is they always continue to buy the same product 00:08:18.690 --> 00:08:22.980 align:middle line:84% they bought previously until one of these breakdowns occur. 00:08:22.980 --> 00:08:26.430 align:middle line:84% When the breakdown occurs, they decide 00:08:26.430 --> 00:08:29.670 align:middle line:84% to buy product k with some probability 00:08:29.670 --> 00:08:33.360 align:middle line:84% mk of t raised to the sigma power, where mk of t 00:08:33.360 --> 00:08:35.549 align:middle line:84% is the market share of the product. 00:08:35.549 --> 00:08:38.159 align:middle line:84% So they just buy it with some probability 00:08:38.159 --> 00:08:41.320 align:middle line:84% that's related to the product's current market share. 00:08:41.320 --> 00:08:41.820 align:middle line:90% OK. 00:08:41.820 --> 00:08:45.000 align:middle line:90% So one example, sigma equals 0. 00:08:45.000 --> 00:08:48.570 align:middle line:84% Sigma equals 0 would just be when a breakdown occurs, 00:08:48.570 --> 00:08:52.620 align:middle line:84% I just buy a product in proportion to its popularity. 00:08:52.620 --> 00:08:55.185 align:middle line:84% So it could be that the supermarket stock shelves 00:08:55.185 --> 00:08:57.180 align:middle line:84% so that the number of boxes of each cereal 00:08:57.180 --> 00:09:00.455 align:middle line:84% is related to how many people buy it, 00:09:00.455 --> 00:09:01.830 align:middle line:84% and therefore I just pick one off 00:09:01.830 --> 00:09:03.860 align:middle line:90% the shelf entirely at random. 00:09:03.860 --> 00:09:05.550 align:middle line:90% OK. 00:09:05.550 --> 00:09:06.270 align:middle line:90% Sorry. 00:09:06.270 --> 00:09:09.300 align:middle line:84% Sigma 0 would be that there's one box of each cereal. 00:09:09.300 --> 00:09:11.430 align:middle line:84% I pick the box of the cereal off at random 00:09:11.430 --> 00:09:13.920 align:middle line:90% when there's one box of each. 00:09:13.920 --> 00:09:21.910 align:middle line:84% Sigma equals 1 is, I buy it in proportion to the market share. 00:09:21.910 --> 00:09:26.070 align:middle line:84% So that could be they stock more boxes of-- 00:09:26.070 --> 00:09:29.070 align:middle line:84% the more popular cereals get more shelf space, 00:09:29.070 --> 00:09:30.730 align:middle line:84% and you pick according to shelf space. 00:09:30.730 --> 00:09:33.810 align:middle line:84% Or it could be that I watch the person before me, 00:09:33.810 --> 00:09:36.990 align:middle line:84% I see them pick off a cereal, and I say, OK, 00:09:36.990 --> 00:09:39.450 align:middle line:84% that person likes it, maybe I will try it. 00:09:39.450 --> 00:09:43.290 align:middle line:84% So that would be a sigma equals 1 model. 00:09:43.290 --> 00:09:47.820 align:middle line:84% A larger sigma could be some explicit reasoning 00:09:47.820 --> 00:09:49.530 align:middle line:84% about popularity, saying, this one 00:09:49.530 --> 00:09:51.330 align:middle line:84% seems to be the most popular, maybe I 00:09:51.330 --> 00:09:54.150 align:middle line:84% will try that, that I think popularity is probably 00:09:54.150 --> 00:09:57.180 align:middle line:84% correlated with fewer breakdowns, therefore, 00:09:57.180 --> 00:09:59.330 align:middle line:90% I will buy a more popular brand. 00:09:59.330 --> 00:10:02.110 align:middle line:84% And so, sigma going to infinity would 00:10:02.110 --> 00:10:06.900 align:middle line:84% be, you buy the most popular brand with probability 1. 00:10:06.900 --> 00:10:09.840 align:middle line:84% And then the main result of Smallwood and Conlisk 00:10:09.840 --> 00:10:14.010 align:middle line:84% is that, with fairly naive consumers, 00:10:14.010 --> 00:10:16.133 align:middle line:84% social learning can work very well. 00:10:16.133 --> 00:10:17.550 align:middle line:84% And what it has to be is they have 00:10:17.550 --> 00:10:20.630 align:middle line:84% to use the right amount of weight on popularity. 00:10:20.630 --> 00:10:23.940 align:middle line:84% So the theorem is, if sigma equals 00:10:23.940 --> 00:10:28.650 align:middle line:84% 1, the most popular product dominates in the t goes 00:10:28.650 --> 00:10:29.770 align:middle line:90% to infinity limit. 00:10:29.770 --> 00:10:33.570 align:middle line:84% That is, the product that has the smallest bk-- 00:10:33.570 --> 00:10:39.390 align:middle line:90% so it's that mk of t goes to 1. 00:10:39.390 --> 00:10:42.840 align:middle line:84% So limit as t goes to infinity of mk of t 00:10:42.840 --> 00:10:54.000 align:middle line:84% goes to 1 for k equals the argmin of bk. 00:10:54.000 --> 00:10:58.260 align:middle line:84% The product that has the lowest breakdown probability 00:10:58.260 --> 00:11:00.940 align:middle line:84% has its market share converge to 1 in the long run. 00:11:00.940 --> 00:11:03.130 align:middle line:90% So the best product takes over. 00:11:03.130 --> 00:11:05.350 align:middle line:90% All other products die out. 00:11:05.350 --> 00:11:08.620 align:middle line:84% And that happens when people put the sigma equals 00:11:08.620 --> 00:11:10.880 align:middle line:90% 1 weight on probability. 00:11:10.880 --> 00:11:13.000 align:middle line:90% So they're not really explicit. 00:11:13.000 --> 00:11:16.150 align:middle line:84% It's as if you just do the "watch one person in front 00:11:16.150 --> 00:11:18.710 align:middle line:90% of you" model. 00:11:18.710 --> 00:11:21.220 align:middle line:84% So it's using very minimal information about probability. 00:11:21.220 --> 00:11:24.560 align:middle line:90% That converges to 1. 00:11:24.560 --> 00:11:28.640 align:middle line:84% When sigma is less than 1, we converge to a state 00:11:28.640 --> 00:11:30.770 align:middle line:90% with all k products active. 00:11:30.770 --> 00:11:34.230 align:middle line:84% So each product has then a market share between 0 and 1 00:11:34.230 --> 00:11:35.270 align:middle line:90% in the long run. 00:11:35.270 --> 00:11:38.010 align:middle line:84% Inferior products stay alive in the market. 00:11:38.010 --> 00:11:39.990 align:middle line:84% Superior products have bigger market shares. 00:11:39.990 --> 00:11:42.020 align:middle line:90% They don't go to 1. 00:11:42.020 --> 00:11:46.220 align:middle line:84% And then, anything like the sigma equals infinity, 00:11:46.220 --> 00:11:47.870 align:middle line:84% but any sigma greater than 1 where 00:11:47.870 --> 00:11:49.700 align:middle line:84% you pick the most popular product, 00:11:49.700 --> 00:11:52.890 align:middle line:84% then inferior products can come to dominate the market, 00:11:52.890 --> 00:11:55.580 align:middle line:84% where if you just get some initial popularity, 00:11:55.580 --> 00:11:57.920 align:middle line:84% you're initially popular, people buy you, 00:11:57.920 --> 00:12:00.800 align:middle line:84% your popularity just goes to 1, and superior products 00:12:00.800 --> 00:12:03.332 align:middle line:84% can't outcompete you because you're more popular. 00:12:03.332 --> 00:12:04.790 align:middle line:84% So Smallwood and Conlisk's argument 00:12:04.790 --> 00:12:08.180 align:middle line:84% was that markets can select things, 00:12:08.180 --> 00:12:12.330 align:middle line:84% but there's this awkward part of it, 00:12:12.330 --> 00:12:15.000 align:middle line:84% that consumers have to be fairly naive to get the market 00:12:15.000 --> 00:12:16.180 align:middle line:90% to select the best products. 00:12:16.180 --> 00:12:19.062 align:middle line:84% They have to put fairly little weight on popularity. 00:12:19.062 --> 00:12:20.520 align:middle line:84% They put some weight on popularity, 00:12:20.520 --> 00:12:22.860 align:middle line:84% but it's got to be the equivalent to the sigma equals 00:12:22.860 --> 00:12:27.220 align:middle line:84% 1 weight on popularity to get social learning to work out. 00:12:27.220 --> 00:12:27.720 align:middle line:90% OK. 00:12:27.720 --> 00:12:30.137 align:middle line:84% Anyway, so I'm just going to then just talk about a couple 00:12:30.137 --> 00:12:36.280 align:middle line:84% of other papers that have built on this and done other things. 00:12:36.280 --> 00:12:43.590 align:middle line:84% So I have a paper with Drew, 1995, 00:12:43.590 --> 00:12:47.310 align:middle line:84% where we try to take the Smallwood and Conlisk idea 00:12:47.310 --> 00:12:51.900 align:middle line:84% and make it a bit more like a classic economics utility 00:12:51.900 --> 00:12:55.140 align:middle line:90% framework. 00:12:55.140 --> 00:12:59.430 align:middle line:84% So here now, when consumer i's consume products, 00:12:59.430 --> 00:13:02.190 align:middle line:84% they receive utility, and the utility 00:13:02.190 --> 00:13:06.330 align:middle line:84% you get from consuming a product is a product specific mean. 00:13:06.330 --> 00:13:09.040 align:middle line:84% This is like the xi j or whatever. 00:13:09.040 --> 00:13:11.890 align:middle line:84% So you get a product specific mean utility. 00:13:11.890 --> 00:13:15.070 align:middle line:84% And you get an epsilon ikt, which 00:13:15.070 --> 00:13:17.616 align:middle line:84% is your own idiosyncratic preference. 00:13:17.616 --> 00:13:20.450 align:middle line:90% 00:13:20.450 --> 00:13:25.070 align:middle line:84% And we're going to assume that the utility 00:13:25.070 --> 00:13:28.220 align:middle line:84% difference between product 1 and probability 2 00:13:28.220 --> 00:13:32.270 align:middle line:84% is a random variable, where sometimes product 1 is better, 00:13:32.270 --> 00:13:34.230 align:middle line:84% or sometimes product 2 is better. 00:13:34.230 --> 00:13:36.020 align:middle line:84% So this is something about these products 00:13:36.020 --> 00:13:40.870 align:middle line:84% that there is a period-specific choice to them that makes one 00:13:40.870 --> 00:13:42.120 align:middle line:90% of them better than the other. 00:13:42.120 --> 00:13:44.620 align:middle line:84% So for instance, it could be that these two products are not 00:13:44.620 --> 00:13:46.220 align:middle line:90% actually cereal boxes. 00:13:46.220 --> 00:13:49.080 align:middle line:84% It could be those two products are ways to drive to work. 00:13:49.080 --> 00:13:50.760 align:middle line:84% Do I drive to work on the Mass Pike? 00:13:50.760 --> 00:13:52.740 align:middle line:84% Do I drive to work on Storrow Drive? 00:13:52.740 --> 00:13:57.140 align:middle line:84% And every day, one of those two ways to drive to work is better, 00:13:57.140 --> 00:14:00.710 align:middle line:90% but it changes from day to day. 00:14:00.710 --> 00:14:02.180 align:middle line:84% It could be that these are products 00:14:02.180 --> 00:14:05.450 align:middle line:84% that you're buying, where there are two restaurants, 00:14:05.450 --> 00:14:08.360 align:middle line:84% and because the food quality varies from day to day 00:14:08.360 --> 00:14:10.070 align:middle line:84% due to the quality of the ingredients 00:14:10.070 --> 00:14:14.313 align:middle line:84% or the chef that's working, on some days shop A, 00:14:14.313 --> 00:14:15.980 align:middle line:84% the food is better than shop B, and then 00:14:15.980 --> 00:14:19.380 align:middle line:90% some days shop B is better. 00:14:19.380 --> 00:14:21.388 align:middle line:84% Or it could be that these are different-- 00:14:21.388 --> 00:14:23.430 align:middle line:84% you're buying from different insurance companies. 00:14:23.430 --> 00:14:26.010 align:middle line:84% And if there's a bad event, then one of them 00:14:26.010 --> 00:14:27.302 align:middle line:90% was better to have bought from. 00:14:27.302 --> 00:14:29.343 align:middle line:84% And if there was no bad event, then the other one 00:14:29.343 --> 00:14:30.610 align:middle line:90% is better to have bought from. 00:14:30.610 --> 00:14:33.690 align:middle line:84% So we have something that makes the quality of the two products 00:14:33.690 --> 00:14:36.540 align:middle line:90% differ. 00:14:36.540 --> 00:14:39.330 align:middle line:84% Product 1 is better with probability p. 00:14:39.330 --> 00:14:43.470 align:middle line:84% So you would like to buy the product 00:14:43.470 --> 00:14:45.750 align:middle line:84% 1 if p is bigger than a half and buy product 2 if p 00:14:45.750 --> 00:14:47.660 align:middle line:90% is less than a half. 00:14:47.660 --> 00:14:53.270 align:middle line:84% And then we have Smallwood and Conlisk-like consumers. 00:14:53.270 --> 00:14:54.560 align:middle line:90% We don't have breakdowns. 00:14:54.560 --> 00:14:58.010 align:middle line:84% So the inertia part is, consumers buy the same product 00:14:58.010 --> 00:15:01.790 align:middle line:84% they bought previously with probability 1 minus alpha. 00:15:01.790 --> 00:15:04.490 align:middle line:84% And with probability alpha, they try 00:15:04.490 --> 00:15:06.680 align:middle line:84% to decide what-- they consider changing 00:15:06.680 --> 00:15:08.870 align:middle line:84% and think about what they should buy. 00:15:08.870 --> 00:15:11.450 align:middle line:84% And when they consider changing, what they do 00:15:11.450 --> 00:15:18.470 align:middle line:84% is ask N of their friends about their most recent experience. 00:15:18.470 --> 00:15:26.180 align:middle line:90% I form u1t bar to be the sum-- 00:15:26.180 --> 00:15:30.800 align:middle line:90% or I form u1t bar and u2t bar. 00:15:30.800 --> 00:15:33.590 align:middle line:84% This is the average of the utilities received by all 00:15:33.590 --> 00:15:35.330 align:middle line:90% my friends who use product 1. 00:15:35.330 --> 00:15:37.580 align:middle line:84% This is the average of the utilities of all my friends 00:15:37.580 --> 00:15:38.840 align:middle line:90% who use product 2. 00:15:38.840 --> 00:15:42.050 align:middle line:84% And then I pick whichever one has the better average when 00:15:42.050 --> 00:15:44.030 align:middle line:90% I've surveyed them. 00:15:44.030 --> 00:15:46.820 align:middle line:90% Again, I asked the friends who-- 00:15:46.820 --> 00:15:50.660 align:middle line:84% I asked the friends who took Storrow Drive which is better 00:15:50.660 --> 00:15:51.835 align:middle line:90% for them, and vice versa. 00:15:51.835 --> 00:15:55.100 align:middle line:90% 00:15:55.100 --> 00:15:57.530 align:middle line:84% And then if all friends-- if everyone I ask 00:15:57.530 --> 00:15:59.910 align:middle line:84% bought the same product, then I just buy that one. 00:15:59.910 --> 00:16:03.590 align:middle line:84% So I take an average of 0 over 0 is negative infinity, 00:16:03.590 --> 00:16:05.240 align:middle line:90% so I don't buy that one. 00:16:05.240 --> 00:16:09.140 align:middle line:84% And so the observation is, in this model, 00:16:09.140 --> 00:16:12.050 align:middle line:84% again, we can sometimes get optimal social learning 00:16:12.050 --> 00:16:15.260 align:middle line:84% from just asking friends about products. 00:16:15.260 --> 00:16:19.340 align:middle line:84% Typically, this occurs when N is fairly small. 00:16:19.340 --> 00:16:24.710 align:middle line:84% So this is a graph showing, as a function of alpha, what's 00:16:24.710 --> 00:16:28.850 align:middle line:90% the range of Ns for which-- 00:16:28.850 --> 00:16:31.970 align:middle line:84% what's the range of Ns for which this model gives 00:16:31.970 --> 00:16:33.690 align:middle line:90% efficient social learning? 00:16:33.690 --> 00:16:35.900 align:middle line:84% And this graph seems to be showing us 00:16:35.900 --> 00:16:39.530 align:middle line:84% that if alpha is close to 1, almost everybody is asking 00:16:39.530 --> 00:16:40.770 align:middle line:90% their friends every period. 00:16:40.770 --> 00:16:42.690 align:middle line:84% Then you need to be asking three, four, five, 00:16:42.690 --> 00:16:44.070 align:middle line:90% or six friends. 00:16:44.070 --> 00:16:45.800 align:middle line:84% Most of the time, when alpha is small, 00:16:45.800 --> 00:16:48.710 align:middle line:84% this model with the "ask one friend what they bought" 00:16:48.710 --> 00:16:49.692 align:middle line:90% works well. 00:16:49.692 --> 00:16:51.900 align:middle line:84% So in some sense, asking one friend what they bought, 00:16:51.900 --> 00:16:53.540 align:middle line:84% that's a lot like Smallwood and Conlisk 00:16:53.540 --> 00:16:56.580 align:middle line:84% because that's like choosing in proportion to market share. 00:16:56.580 --> 00:16:59.820 align:middle line:84% But then when alpha is bigger, what you want to do 00:16:59.820 --> 00:17:01.680 align:middle line:90% is gather more information. 00:17:01.680 --> 00:17:05.670 align:middle line:84% And the more information helps you find the better products. 00:17:05.670 --> 00:17:12.740 align:middle line:84% And again, what's going on here is that there's this more-- 00:17:12.740 --> 00:17:16.000 align:middle line:84% better products do have higher mean utilities. 00:17:16.000 --> 00:17:20.210 align:middle line:84% And then there's also some effect of popularity weighting 00:17:20.210 --> 00:17:22.532 align:middle line:84% in that if something is more popular, 00:17:22.532 --> 00:17:23.990 align:middle line:84% you're more likely to hear about it 00:17:23.990 --> 00:17:26.130 align:middle line:84% and more likely to not hear about the popularity. 00:17:26.130 --> 00:17:29.300 align:middle line:84% And the combination of those two things 00:17:29.300 --> 00:17:32.570 align:middle line:84% can give us social learning but doesn't give us social 00:17:32.570 --> 00:17:35.970 align:middle line:90% learning if N is too big. 00:17:35.970 --> 00:17:39.673 align:middle line:84% And then one other paper along these lines. 00:17:39.673 --> 00:17:42.360 align:middle line:90% 00:17:42.360 --> 00:17:46.530 align:middle line:84% Ran Spiegler has a paper, another follow-on paper, 00:17:46.530 --> 00:17:51.420 align:middle line:84% he calls "The Market for Quacks," which is the-- 00:17:51.420 --> 00:17:54.840 align:middle line:84% my paper with Drew, we were following Smallwood and Conlisk 00:17:54.840 --> 00:17:57.540 align:middle line:84% in not having any prices at all in the model, 00:17:57.540 --> 00:17:59.827 align:middle line:84% which works well for the people learning 00:17:59.827 --> 00:18:02.160 align:middle line:84% which way to drive to work or learning things like that. 00:18:02.160 --> 00:18:06.480 align:middle line:84% Doesn't work as well necessarily for buying products. 00:18:06.480 --> 00:18:10.230 align:middle line:84% Spiegler thinks about what happens in word-of-mouth models 00:18:10.230 --> 00:18:12.720 align:middle line:90% if you let the firms set prices. 00:18:12.720 --> 00:18:14.580 align:middle line:84% So he calls it "The Market for Quacks" 00:18:14.580 --> 00:18:16.570 align:middle line:84% because he's thinking about the application-- 00:18:16.570 --> 00:18:21.750 align:middle line:84% his application he's thinking about is people selling the many 00:18:21.750 --> 00:18:25.830 align:middle line:84% medications and health things or whatever that you see advertised 00:18:25.830 --> 00:18:29.902 align:middle line:84% online that are probably no better than not buying some kind 00:18:29.902 --> 00:18:31.860 align:middle line:84% of health supplement that's supposed to improve 00:18:31.860 --> 00:18:33.780 align:middle line:90% your whatever. 00:18:33.780 --> 00:18:37.420 align:middle line:84% And so what he imagines is that he has this model where you have 00:18:37.420 --> 00:18:41.320 align:middle line:84% this large number of products, and whenever you consume 00:18:41.320 --> 00:18:43.840 align:middle line:84% one of these products, with probability alpha, 00:18:43.840 --> 00:18:46.870 align:middle line:84% you do get a good outcome after having consumed it. 00:18:46.870 --> 00:18:49.330 align:middle line:84% But of course, there's an outside option that gives you 00:18:49.330 --> 00:18:53.560 align:middle line:84% utility that costs 0, and you also get a positive experience 00:18:53.560 --> 00:18:56.080 align:middle line:84% sometimes after not consuming any of these products. 00:18:56.080 --> 00:18:59.920 align:middle line:84% And in particular, he's got that the outside option-- 00:18:59.920 --> 00:19:05.800 align:middle line:84% there's an outside option which also 00:19:05.800 --> 00:19:09.700 align:middle line:84% gives you a good experience with probability alpha and costs 0. 00:19:09.700 --> 00:19:11.950 align:middle line:84% So the outside option is actually just better 00:19:11.950 --> 00:19:13.125 align:middle line:90% than all of these products. 00:19:13.125 --> 00:19:16.790 align:middle line:90% 00:19:16.790 --> 00:19:20.090 align:middle line:84% So what happens is, the utility that a consumer i gets in period 00:19:20.090 --> 00:19:25.460 align:middle line:84% t if they buy product k is 1 minus pkt with probability alpha 00:19:25.460 --> 00:19:30.260 align:middle line:84% and minus pkt with probability 1 minus alpha. 00:19:30.260 --> 00:19:36.157 align:middle line:84% And what consumers do in each period is, again, 00:19:36.157 --> 00:19:38.240 align:middle line:84% like Smallwood and Conlisk, we have this continuum 00:19:38.240 --> 00:19:39.090 align:middle line:90% of consumers. 00:19:39.090 --> 00:19:41.250 align:middle line:84% Some consumers are using every product. 00:19:41.250 --> 00:19:43.400 align:middle line:84% What consumers do in each period is, 00:19:43.400 --> 00:19:47.420 align:middle line:84% ask one user of each product about their experience 00:19:47.420 --> 00:19:49.790 align:middle line:84% and buy the product that would give the highest 00:19:49.790 --> 00:19:53.090 align:middle line:84% utility assuming the same experience if it 00:19:53.090 --> 00:19:54.680 align:middle line:90% gives positive utility. 00:19:54.680 --> 00:19:59.130 align:middle line:84% So I ask-- there are k different products. 00:19:59.130 --> 00:20:00.680 align:middle line:84% I find one person using each product. 00:20:00.680 --> 00:20:02.370 align:middle line:90% Well, how did you like it? 00:20:02.370 --> 00:20:03.243 align:middle line:90% Did it work for you? 00:20:03.243 --> 00:20:05.910 align:middle line:84% And three people tell you that, yeah, the product worked for me. 00:20:05.910 --> 00:20:09.630 align:middle line:84% It gave me utility 1, but I paid this for it. 00:20:09.630 --> 00:20:11.370 align:middle line:84% Seven people tell you I didn't like it. 00:20:11.370 --> 00:20:12.410 align:middle line:90% It gave me utility 0. 00:20:12.410 --> 00:20:14.970 align:middle line:84% And I paid money for it, so I got negative utility. 00:20:14.970 --> 00:20:20.487 align:middle line:84% And so then what consumers do is they pick the-- 00:20:20.487 --> 00:20:23.070 align:middle line:84% in some sense, they're going to choose among the products that 00:20:23.070 --> 00:20:25.560 align:middle line:84% gave utility 1 to their friends the one that's 00:20:25.560 --> 00:20:27.690 align:middle line:84% offered at the lowest price this period. 00:20:27.690 --> 00:20:29.430 align:middle line:84% And if every one of their friends 00:20:29.430 --> 00:20:31.950 align:middle line:84% got 0, got negative utility, then they just 00:20:31.950 --> 00:20:32.975 align:middle line:90% buy the outside option. 00:20:32.975 --> 00:20:36.820 align:middle line:90% 00:20:36.820 --> 00:20:37.320 align:middle line:90% OK. 00:20:37.320 --> 00:20:38.970 align:middle line:90% What happens here? 00:20:38.970 --> 00:20:42.510 align:middle line:84% This model is-- if you think back 00:20:42.510 --> 00:20:45.390 align:middle line:84% to the Stahl model of price dispersion, 00:20:45.390 --> 00:20:48.120 align:middle line:84% this model is going to be kind of like that Stahl 00:20:48.120 --> 00:20:53.610 align:middle line:84% model of price dispersion, because imagine you're a firm. 00:20:53.610 --> 00:20:55.290 align:middle line:84% You're a firm, and you know that there's 00:20:55.290 --> 00:20:56.957 align:middle line:84% going to be some consumers out there who 00:20:56.957 --> 00:20:59.880 align:middle line:84% heard positive reviews of your product 00:20:59.880 --> 00:21:02.580 align:middle line:84% and negative reviews of every other product. 00:21:02.580 --> 00:21:05.190 align:middle line:84% And so if they heard a positive news about your product 00:21:05.190 --> 00:21:06.690 align:middle line:84% and negative about everybody else's, 00:21:06.690 --> 00:21:08.868 align:middle line:90% you'd want to set p equals 1. 00:21:08.868 --> 00:21:10.410 align:middle line:84% But then there can be other consumers 00:21:10.410 --> 00:21:12.660 align:middle line:84% who heard positive things about your product and about 00:21:12.660 --> 00:21:14.077 align:middle line:90% exactly one other product. 00:21:14.077 --> 00:21:16.410 align:middle line:84% And so if they heard positive things about two products, 00:21:16.410 --> 00:21:19.513 align:middle line:84% including yours, you want to be epsilon less than the product 00:21:19.513 --> 00:21:21.180 align:middle line:84% that they heard-- the other product they 00:21:21.180 --> 00:21:22.740 align:middle line:90% heard good things about. 00:21:22.740 --> 00:21:25.190 align:middle line:84% And so it's kind of like that kind of Stahl model, 00:21:25.190 --> 00:21:26.940 align:middle line:84% where there's multiple types of consumers, 00:21:26.940 --> 00:21:29.070 align:middle line:84% the consumers who heard good things only 00:21:29.070 --> 00:21:32.398 align:middle line:84% about yours, where they're like the non-shoppers, 00:21:32.398 --> 00:21:34.440 align:middle line:84% and the consumers who heard good things about you 00:21:34.440 --> 00:21:35.700 align:middle line:90% and several others. 00:21:35.700 --> 00:21:38.830 align:middle line:84% And there, you want to be undercutting the other prices. 00:21:38.830 --> 00:21:42.540 align:middle line:84% So what you're going to get is some kind of mixed strategy 00:21:42.540 --> 00:21:46.020 align:middle line:84% equilibrium, where you're mixing over some interval, 00:21:46.020 --> 00:21:48.630 align:middle line:84% and you're sometimes charging a high price 00:21:48.630 --> 00:21:50.700 align:middle line:84% and sometimes charging a low price 00:21:50.700 --> 00:21:53.970 align:middle line:84% to try to get the people who've heard about multiple products. 00:21:53.970 --> 00:21:57.058 align:middle line:84% It's not as extreme as that Stahl model in that there's-- 00:21:57.058 --> 00:21:59.100 align:middle line:84% it's going to be very rare for consumers who have 00:21:59.100 --> 00:22:00.600 align:middle line:90% heard about all N products. 00:22:00.600 --> 00:22:03.570 align:middle line:84% So you don't need to be the single lowest price product 00:22:03.570 --> 00:22:04.742 align:middle line:90% in the market to get people. 00:22:04.742 --> 00:22:07.200 align:middle line:84% They've probably only heard about two or three or something 00:22:07.200 --> 00:22:08.170 align:middle line:90% like that. 00:22:08.170 --> 00:22:11.640 align:middle line:84% And so anyway, you'll have some distribution-- 00:22:11.640 --> 00:22:17.126 align:middle line:84% I imagine it probably looks like this-- of what prices look like. 00:22:17.126 --> 00:22:22.210 align:middle line:84% An interesting observation is that in this model, 00:22:22.210 --> 00:22:26.140 align:middle line:84% price is inversely related to product quality. 00:22:26.140 --> 00:22:31.660 align:middle line:84% So if you imagine, if the products are all-- 00:22:31.660 --> 00:22:33.310 align:middle line:90% this is common product quality. 00:22:33.310 --> 00:22:36.670 align:middle line:84% If the products are all good and people actually 00:22:36.670 --> 00:22:39.130 align:middle line:84% have positive experiences with them, 00:22:39.130 --> 00:22:42.250 align:middle line:84% then you as a firm selling are going to be 00:22:42.250 --> 00:22:44.620 align:middle line:84% like, OK, if someone heard about my-- 00:22:44.620 --> 00:22:47.080 align:middle line:84% their friend had a good experience with my product, 00:22:47.080 --> 00:22:48.580 align:middle line:84% they probably also talked to friends 00:22:48.580 --> 00:22:50.622 align:middle line:84% who had good experiences with the other products. 00:22:50.622 --> 00:22:53.590 align:middle line:84% So I'm probably competing with three or four or five or six 00:22:53.590 --> 00:22:55.280 align:middle line:84% other products for their business. 00:22:55.280 --> 00:22:56.980 align:middle line:84% Therefore, I need to price aggressively 00:22:56.980 --> 00:22:59.830 align:middle line:84% to be-- beat those four or five or six products. 00:22:59.830 --> 00:23:04.330 align:middle line:84% But if the products are terrible and there's a 1 in 100 chance 00:23:04.330 --> 00:23:06.490 align:middle line:84% that they actually do anything, then you 00:23:06.490 --> 00:23:08.370 align:middle line:84% would think, OK, well, if someone 00:23:08.370 --> 00:23:10.120 align:middle line:84% had a positive experience with my product, 00:23:10.120 --> 00:23:12.042 align:middle line:90% then when they're asking-- 00:23:12.042 --> 00:23:13.750 align:middle line:84% they found a friend who liked my product. 00:23:13.750 --> 00:23:14.830 align:middle line:84% They probably didn't find a friend 00:23:14.830 --> 00:23:16.330 align:middle line:84% who liked any of the other products. 00:23:16.330 --> 00:23:18.430 align:middle line:84% So I'm probably a monopolist, so I'm probably 00:23:18.430 --> 00:23:20.480 align:middle line:90% safe pricing close to 1. 00:23:20.480 --> 00:23:25.880 align:middle line:84% So in equilibrium, competition between low quality products, 00:23:25.880 --> 00:23:27.050 align:middle line:90% we get high prices. 00:23:27.050 --> 00:23:30.470 align:middle line:84% Competition between high quality products, we get low prices. 00:23:30.470 --> 00:23:33.200 align:middle line:84% And the reason is that the competition between high quality 00:23:33.200 --> 00:23:36.200 align:middle line:84% products is actually more real competition, 00:23:36.200 --> 00:23:38.540 align:middle line:84% whereas the competition for the low quality products 00:23:38.540 --> 00:23:40.010 align:middle line:84% is just getting the people who have 00:23:40.010 --> 00:23:44.160 align:middle line:90% the rare mistaken impression. 00:23:44.160 --> 00:23:47.920 align:middle line:84% So this is not a fully rational model. 00:23:47.920 --> 00:23:54.200 align:middle line:84% The consumers are not somehow inverting the equilibrium price 00:23:54.200 --> 00:23:56.150 align:middle line:84% distribution or inverting the market shares 00:23:56.150 --> 00:23:59.420 align:middle line:84% and realizing, OK, if prices are high, 00:23:59.420 --> 00:24:01.230 align:middle line:84% it must be that the quality is low. 00:24:01.230 --> 00:24:04.407 align:middle line:84% Therefore, I'm learning something about-- 00:24:04.407 --> 00:24:05.990 align:middle line:84% I had a prior about alpha, and now I'm 00:24:05.990 --> 00:24:07.970 align:middle line:84% learning something about alpha, or I'm not 00:24:07.970 --> 00:24:11.540 align:middle line:84% updating alpha from realizing that I'm just getting i.i.d. 00:24:11.540 --> 00:24:13.910 align:middle line:84% These are all-- these seem to be i.i.d draws 00:24:13.910 --> 00:24:15.410 align:middle line:90% as I draw them every period. 00:24:15.410 --> 00:24:18.510 align:middle line:84% They're just every period taking this person liked it, 00:24:18.510 --> 00:24:19.965 align:middle line:90% therefore this is quality 1. 00:24:19.965 --> 00:24:21.090 align:middle line:90% This person didn't like it. 00:24:21.090 --> 00:24:23.690 align:middle line:90% These are all quality 0. 00:24:23.690 --> 00:24:27.380 align:middle line:84% But I think Spiegler's view is that this 00:24:27.380 --> 00:24:33.860 align:middle line:84% seems like a reasonable force and that the naive model may 00:24:33.860 --> 00:24:36.830 align:middle line:84% be a better way to think about, that this force of competition 00:24:36.830 --> 00:24:41.800 align:middle line:84% or quality drives down prices could be a real thing. 00:24:41.800 --> 00:24:42.680 align:middle line:90% OK. 00:24:42.680 --> 00:24:46.080 align:middle line:84% Second thing I want to do is-- and I 00:24:46.080 --> 00:24:49.650 align:middle line:84% figured I would do something fun today 00:24:49.650 --> 00:24:51.790 align:middle line:84% in honor of being here before the holidays. 00:24:51.790 --> 00:24:54.210 align:middle line:84% So anyway, I'm going to talk about-- this 00:24:54.210 --> 00:24:58.270 align:middle line:84% is another rule-of-thumb paper I've written. 00:24:58.270 --> 00:25:00.240 align:middle line:84% It's not actually really an IO paper, 00:25:00.240 --> 00:25:03.180 align:middle line:84% but I feel it's useful for graduate students who have 00:25:03.180 --> 00:25:04.860 align:middle line:90% thought about these issues. 00:25:04.860 --> 00:25:09.900 align:middle line:84% So this is a paper thinking about the journal refereeing 00:25:09.900 --> 00:25:10.480 align:middle line:90% process. 00:25:10.480 --> 00:25:14.588 align:middle line:84% So 6 or 7 years from now, many of you will be professors, 00:25:14.588 --> 00:25:16.380 align:middle line:84% and you'll be sitting in offices somewhere, 00:25:16.380 --> 00:25:18.450 align:middle line:84% and this is what you'll be complaining about, 00:25:18.450 --> 00:25:21.120 align:middle line:84% is the journal refereeing process. 00:25:21.120 --> 00:25:26.550 align:middle line:84% So in particular, what's very difficult, as people find out, 00:25:26.550 --> 00:25:30.300 align:middle line:84% is somehow it's very, very hard to publish an economics paper, 00:25:30.300 --> 00:25:32.430 align:middle line:84% and it takes years and years and years and years. 00:25:32.430 --> 00:25:34.380 align:middle line:84% And you try to write about something topical, 00:25:34.380 --> 00:25:36.510 align:middle line:84% and then it takes five years to get your paper 00:25:36.510 --> 00:25:38.612 align:middle line:84% through the journal refereeing process. 00:25:38.612 --> 00:25:41.070 align:middle line:84% And after five years getting through the journal refereeing 00:25:41.070 --> 00:25:43.952 align:middle line:84% process, your wonderfully topical thing 00:25:43.952 --> 00:25:45.910 align:middle line:84% is now completely irrelevant because the things 00:25:45.910 --> 00:25:49.720 align:middle line:84% you're writing about are no longer topical. 00:25:49.720 --> 00:25:52.610 align:middle line:84% Actually, the experience-- I'm a department head. 00:25:52.610 --> 00:25:55.115 align:middle line:84% I'm also on Promotions and Tenure Committee for people 00:25:55.115 --> 00:25:55.990 align:middle line:90% in other departments. 00:25:55.990 --> 00:25:58.870 align:middle line:84% And in other social science fields where people write books, 00:25:58.870 --> 00:26:00.860 align:middle line:84% this problem also just appears in spades. 00:26:00.860 --> 00:26:03.170 align:middle line:84% We're constantly reading cases of, here's this person. 00:26:03.170 --> 00:26:05.080 align:middle line:84% They've been a faculty member for five years. 00:26:05.080 --> 00:26:07.692 align:middle line:84% Their book, which was based on their graduate student thesis, 00:26:07.692 --> 00:26:09.400 align:middle line:84% is still not out yet because they've just 00:26:09.400 --> 00:26:12.110 align:middle line:84% been revising their book for the past five years. 00:26:12.110 --> 00:26:13.960 align:middle line:84% And the book is five years less timely 00:26:13.960 --> 00:26:17.080 align:middle line:84% than it was when they wrote it originally. 00:26:17.080 --> 00:26:20.440 align:middle line:84% So anyway, how do I think about why publication works this way? 00:26:20.440 --> 00:26:21.440 align:middle line:90% Here's a theory. 00:26:21.440 --> 00:26:28.330 align:middle line:84% So in the model, papers have a two-dimensional quality q and r. 00:26:28.330 --> 00:26:33.310 align:middle line:84% q I think of as the fundamental idea in the paper 00:26:33.310 --> 00:26:35.830 align:middle line:84% or the quality of the unchangeable insight 00:26:35.830 --> 00:26:37.563 align:middle line:90% that this paper has. 00:26:37.563 --> 00:26:38.980 align:middle line:84% The q is what you would teach when 00:26:38.980 --> 00:26:40.855 align:middle line:84% you teach that paper in your graduate course. 00:26:40.855 --> 00:26:42.320 align:middle line:90% This is the paper's idea. 00:26:42.320 --> 00:26:44.008 align:middle line:84% This is how that paper changes the way 00:26:44.008 --> 00:26:45.050 align:middle line:90% we think about something. 00:26:45.050 --> 00:26:47.970 align:middle line:84% This was their interesting idea they came up with. 00:26:47.970 --> 00:26:52.860 align:middle line:84% And then the r is the other dimensions of quality 00:26:52.860 --> 00:26:55.270 align:middle line:84% that you can improve in a revision. 00:26:55.270 --> 00:26:58.050 align:middle line:84% So rather than having linear functional forms, 00:26:58.050 --> 00:27:01.090 align:middle line:84% you can do everything with arbitrary functional forms. 00:27:01.090 --> 00:27:03.120 align:middle line:84% You can go through the seven most 00:27:03.120 --> 00:27:06.120 align:middle line:84% recent econometric techniques published in Econometrica 00:27:06.120 --> 00:27:09.040 align:middle line:84% and re-estimate it using these seven different estimators 00:27:09.040 --> 00:27:11.413 align:middle line:84% and talk about how all those estimators do things. 00:27:11.413 --> 00:27:13.830 align:middle line:84% Obviously, some parts of the things we improve in revision 00:27:13.830 --> 00:27:16.750 align:middle line:84% are making us more confident in the results. 00:27:16.750 --> 00:27:18.040 align:middle line:90% They are valuable. 00:27:18.040 --> 00:27:21.032 align:middle line:84% Many of them are just doing everything-- 00:27:21.032 --> 00:27:23.490 align:middle line:84% addressing every potential criticism that anyone could have 00:27:23.490 --> 00:27:27.540 align:middle line:84% come up with and just continuing to just make your paper harder 00:27:27.540 --> 00:27:31.650 align:middle line:90% and harder to read. 00:27:31.650 --> 00:27:36.390 align:middle line:84% So we value both aspects of quality, the idea part 00:27:36.390 --> 00:27:39.240 align:middle line:90% and the craftsmanship part. 00:27:39.240 --> 00:27:41.880 align:middle line:84% And referees know that they're supposed-- 00:27:41.880 --> 00:27:44.250 align:middle line:84% that there's a social norm in the profession 00:27:44.250 --> 00:27:46.510 align:middle line:84% that we put alpha weight on the q 00:27:46.510 --> 00:27:49.000 align:middle line:84% and we put 1 minus alpha weight on the r. 00:27:49.000 --> 00:27:52.040 align:middle line:84% And they know that they're supposed to recommend acceptance 00:27:52.040 --> 00:27:56.100 align:middle line:84% if this weighted average of q and r is at least z. 00:27:56.100 --> 00:28:01.840 align:middle line:84% But referees being people who are, in general, 32 00:28:01.840 --> 00:28:04.300 align:middle line:84% years old or whatever, not in the profession for all 00:28:04.300 --> 00:28:07.783 align:middle line:84% that long, don't necessarily know what alpha and z are. 00:28:07.783 --> 00:28:08.950 align:middle line:90% And they're trying to learn. 00:28:08.950 --> 00:28:10.750 align:middle line:84% They're just trying to be good referees 00:28:10.750 --> 00:28:12.760 align:middle line:84% and evaluate papers appropriately 00:28:12.760 --> 00:28:14.420 align:middle line:90% when asked to evaluate a paper. 00:28:14.420 --> 00:28:16.960 align:middle line:84% And so at time t, the profession ends up 00:28:16.960 --> 00:28:19.600 align:middle line:84% using these weights, alpha t and zt. 00:28:19.600 --> 00:28:21.820 align:middle line:84% And all the referees in the profession 00:28:21.820 --> 00:28:24.070 align:middle line:84% are just trying to update all the time 00:28:24.070 --> 00:28:26.860 align:middle line:84% and learn, what is alpha t, what's the weight on these two, 00:28:26.860 --> 00:28:28.780 align:middle line:84% and what's the quality threshold where I'm 00:28:28.780 --> 00:28:30.090 align:middle line:90% supposed to say yes versus no? 00:28:30.090 --> 00:28:32.830 align:middle line:90% 00:28:32.830 --> 00:28:35.530 align:middle line:84% I assume that authors rationally divide their time 00:28:35.530 --> 00:28:38.260 align:middle line:84% between producing the two types of quality. 00:28:38.260 --> 00:28:41.320 align:middle line:84% They start by picking some amount of time tq. 00:28:41.320 --> 00:28:44.410 align:middle line:84% So 1 is the total amount of time you have available. 00:28:44.410 --> 00:28:50.230 align:middle line:84% You start by spending some time tq in 0,1 to get a random draw 00:28:50.230 --> 00:28:51.760 align:middle line:90% on an idea. 00:28:51.760 --> 00:28:54.010 align:middle line:84% And at the end of that time tq that you 00:28:54.010 --> 00:28:57.880 align:middle line:84% spent developing random draws on the ideas, 00:28:57.880 --> 00:29:01.250 align:middle line:84% then you get an idea of quality q. 00:29:01.250 --> 00:29:05.780 align:middle line:84% And there's some CDF F of q given tq that's increasing. 00:29:05.780 --> 00:29:09.050 align:middle line:84% So for instance, it could be that just q is uniform on 0 00:29:09.050 --> 00:29:10.450 align:middle line:90% to tq, is one. 00:29:10.450 --> 00:29:14.480 align:middle line:90% 00:29:14.480 --> 00:29:20.410 align:middle line:84% But you could also make it normal mean tq or whatever. 00:29:20.410 --> 00:29:24.257 align:middle line:84% And then after you get your idea, 00:29:24.257 --> 00:29:25.840 align:middle line:84% you send your paper in to the journal, 00:29:25.840 --> 00:29:28.690 align:middle line:90% and they evaluate your paper. 00:29:28.690 --> 00:29:31.490 align:middle line:84% And then they tell you how good-- 00:29:31.490 --> 00:29:34.360 align:middle line:84% they give you some answer, r of q, 00:29:34.360 --> 00:29:36.640 align:middle line:84% which is like, if you can make your paper-- 00:29:36.640 --> 00:29:39.610 align:middle line:84% get your craftsmanship up to quality r of q, 00:29:39.610 --> 00:29:40.860 align:middle line:90% we would take your paper. 00:29:40.860 --> 00:29:42.830 align:middle line:84% Otherwise, no, it's not good enough. 00:29:42.830 --> 00:29:44.650 align:middle line:90% So here's your paper. 00:29:44.650 --> 00:29:50.200 align:middle line:84% In general, actually, you send a paper to a lot of top economics 00:29:50.200 --> 00:29:53.440 align:middle line:84% journals today, you're going to get five referee reports that 00:29:53.440 --> 00:29:55.780 align:middle line:84% are going to be a total of like 18 pages long 00:29:55.780 --> 00:30:00.200 align:middle line:84% and have 67 comments of things you could improve in your paper. 00:30:00.200 --> 00:30:04.210 align:middle line:84% And you then decide-- they're telling you 00:30:04.210 --> 00:30:07.040 align:middle line:84% how much you have to do to get the paper into the journal. 00:30:07.040 --> 00:30:10.330 align:middle line:84% And then authors decide how much time of their remaining 00:30:10.330 --> 00:30:14.590 align:middle line:84% time between 0 and 1 minus tq to spend improving the paper. 00:30:14.590 --> 00:30:18.010 align:middle line:84% And time spent improving the craftsmanship increases r 00:30:18.010 --> 00:30:23.420 align:middle line:84% to h of tr plus eta, where eta is again some random variable. 00:30:23.420 --> 00:30:26.120 align:middle line:84% So sometimes you're able to do the things well, 00:30:26.120 --> 00:30:28.200 align:middle line:84% sometimes you're not able to do the things well. 00:30:28.200 --> 00:30:33.560 align:middle line:84% And then I make eta, again, a uniform distribution. 00:30:33.560 --> 00:30:36.530 align:middle line:84% And then what editors do is editors 00:30:36.530 --> 00:30:40.430 align:middle line:84% accept the highest quality papers that are resubmitted. 00:30:40.430 --> 00:30:42.680 align:middle line:84% The editors all also know alpha t. 00:30:42.680 --> 00:30:46.560 align:middle line:84% They all know the current market weights on these things. 00:30:46.560 --> 00:30:48.740 align:middle line:84% The editors, though, don't use zt. 00:30:48.740 --> 00:30:50.900 align:middle line:84% The editors, they know how many papers 00:30:50.900 --> 00:30:53.570 align:middle line:84% they're supposed to accept, so they just accept the best. 00:30:53.570 --> 00:30:55.190 align:middle line:84% Even if the referee has told people 00:30:55.190 --> 00:30:56.925 align:middle line:84% to do more than was feasible, the editors 00:30:56.925 --> 00:30:58.550 align:middle line:84% have to go against the referee and say, 00:30:58.550 --> 00:31:00.030 align:middle line:90% I have to accept something. 00:31:00.030 --> 00:31:03.260 align:middle line:84% And so the editors just accept some fraction tau 00:31:03.260 --> 00:31:05.690 align:middle line:84% of papers that are highest on this metric alpha q 00:31:05.690 --> 00:31:06.835 align:middle line:90% plus 1 minus alpha r. 00:31:06.835 --> 00:31:09.370 align:middle line:90% 00:31:09.370 --> 00:31:13.630 align:middle line:84% So anyway, that's my model of academics writing and revising 00:31:13.630 --> 00:31:14.130 align:middle line:90% papers. 00:31:14.130 --> 00:31:17.100 align:middle line:90% 00:31:17.100 --> 00:31:19.710 align:middle line:90% First observation. 00:31:19.710 --> 00:31:24.383 align:middle line:84% This model has a continuum of consistent social norms. 00:31:24.383 --> 00:31:27.420 align:middle line:90% 00:31:27.420 --> 00:31:29.580 align:middle line:84% If you think about it, if I pick any alpha, 00:31:29.580 --> 00:31:33.030 align:middle line:84% if I pick alpha equals 1 so only q counts, 00:31:33.030 --> 00:31:35.010 align:middle line:84% if only q counts, what people are going to do 00:31:35.010 --> 00:31:37.410 align:middle line:84% is just spend 100% of their time producing ideas, 00:31:37.410 --> 00:31:39.780 align:middle line:84% have zero craftsmanship, and there's 00:31:39.780 --> 00:31:43.810 align:middle line:84% going to be some level z that clears the market. 00:31:43.810 --> 00:31:48.780 align:middle line:84% So if you spend 100% of your time on q quality, 00:31:48.780 --> 00:31:51.510 align:middle line:84% a fraction tau of people are going to get qualities 00:31:51.510 --> 00:31:54.210 align:middle line:84% above this level, and 1 minus tau 00:31:54.210 --> 00:31:57.270 align:middle line:84% are going to get qualities below that level. 00:31:57.270 --> 00:31:58.685 align:middle line:90% And that clears the market. 00:31:58.685 --> 00:32:01.470 align:middle line:90% 00:32:01.470 --> 00:32:04.650 align:middle line:84% If you choose alpha equals 0, then people 00:32:04.650 --> 00:32:06.720 align:middle line:84% are going to spend zero time thinking up ideas, 00:32:06.720 --> 00:32:10.530 align:middle line:84% and they're just going to have very, very 00:32:10.530 --> 00:32:14.490 align:middle line:84% complicated estimations or proofs of their non-ideas. 00:32:14.490 --> 00:32:16.980 align:middle line:84% They're going to have incredibly high levels of r quality, 00:32:16.980 --> 00:32:18.980 align:middle line:84% but they're going to have no ideas in the paper. 00:32:18.980 --> 00:32:21.640 align:middle line:84% And again, the people who get the best draws-- 00:32:21.640 --> 00:32:23.290 align:middle line:90% everyone spends time 1. 00:32:23.290 --> 00:32:26.350 align:middle line:84% The people who get the best draws here, the tau fraction, 00:32:26.350 --> 00:32:28.315 align:middle line:84% get the best draws, get their papers published. 00:32:28.315 --> 00:32:29.940 align:middle line:84% The other people get the bad draws on r 00:32:29.940 --> 00:32:31.482 align:middle line:84% and don't get their papers published. 00:32:31.482 --> 00:32:33.910 align:middle line:84% And then, for every weight in between, 00:32:33.910 --> 00:32:36.820 align:middle line:84% people are going to spend some time in q and some time in r, 00:32:36.820 --> 00:32:38.710 align:middle line:84% and there's some quality threshold 00:32:38.710 --> 00:32:40.160 align:middle line:90% that clears the market. 00:32:40.160 --> 00:32:43.900 align:middle line:84% So in some sense, there's nothing 00:32:43.900 --> 00:32:47.890 align:middle line:84% in this model that pins down whether we care about main ideas 00:32:47.890 --> 00:32:49.820 align:middle line:84% or whether we care about craftsmanship. 00:32:49.820 --> 00:32:51.740 align:middle line:84% Profession could have any set of norms. 00:32:51.740 --> 00:32:52.845 align:middle line:90% They're all stable. 00:32:52.845 --> 00:32:55.980 align:middle line:90% 00:32:55.980 --> 00:32:58.920 align:middle line:84% And this would be an example of what this equilibrium would 00:32:58.920 --> 00:33:00.275 align:middle line:90% look like. 00:33:00.275 --> 00:33:02.650 align:middle line:84% This is a weight where people are putting weight on both. 00:33:02.650 --> 00:33:04.698 align:middle line:90% So people choose some time q. 00:33:04.698 --> 00:33:05.740 align:middle line:90% It must be a pretty high. 00:33:05.740 --> 00:33:08.130 align:middle line:90% People put some weight on q. 00:33:08.130 --> 00:33:10.530 align:middle line:84% This is equal weights, looks like. 00:33:10.530 --> 00:33:13.660 align:middle line:84% People put some weight on q that's between 0 and 1. 00:33:13.660 --> 00:33:17.010 align:middle line:90% They come up with ideas. 00:33:17.010 --> 00:33:19.590 align:middle line:84% If at the end of your time q you end up with a bad idea, 00:33:19.590 --> 00:33:21.382 align:middle line:84% you send it to the journal, and the journal 00:33:21.382 --> 00:33:24.510 align:middle line:84% is like, in order to get your paper accepted, 00:33:24.510 --> 00:33:26.820 align:middle line:84% you're going to have to reach this super high threshold 00:33:26.820 --> 00:33:30.070 align:middle line:84% on craftsmanship because you don't have much of an idea, 00:33:30.070 --> 00:33:31.980 align:middle line:84% you look at it and you're like, given 00:33:31.980 --> 00:33:35.460 align:middle line:84% that I made the distribution that r 00:33:35.460 --> 00:33:42.120 align:middle line:84% is h of tr plus eta, where eta is uniform 0 sigma, 00:33:42.120 --> 00:33:43.890 align:middle line:84% there's an upper bound to how good 00:33:43.890 --> 00:33:47.290 align:middle line:84% my craftsmanship can be even if I spend all the time on it. 00:33:47.290 --> 00:33:49.218 align:middle line:90% So you've got a bad idea. 00:33:49.218 --> 00:33:50.260 align:middle line:90% They tell you what to do. 00:33:50.260 --> 00:33:51.910 align:middle line:84% You're like, there's no way I can do that. 00:33:51.910 --> 00:33:52.618 align:middle line:90% You just give up. 00:33:52.618 --> 00:33:55.780 align:middle line:84% You don't revise your paper, and you go home. 00:33:55.780 --> 00:34:01.120 align:middle line:84% If your paper is better, you revise it to death. 00:34:01.120 --> 00:34:04.390 align:middle line:90% You set tr to be 1 minus tq. 00:34:04.390 --> 00:34:06.220 align:middle line:84% You do all the revisions you can. 00:34:06.220 --> 00:34:08.760 align:middle line:90% You get this draw on r quality. 00:34:08.760 --> 00:34:10.800 align:middle line:84% And if you get a good draw, you get accepted. 00:34:10.800 --> 00:34:13.409 align:middle line:84% You get a bad draw, you get rejected. 00:34:13.409 --> 00:34:16.920 align:middle line:84% And then, if your quality is super high, 00:34:16.920 --> 00:34:19.740 align:middle line:84% then they recognize your quality is super high and they tell you, 00:34:19.740 --> 00:34:21.810 align:middle line:90% you only need to do this-- 00:34:21.810 --> 00:34:24.130 align:middle line:84% reach this level of r quality to get accepted. 00:34:24.130 --> 00:34:27.750 align:middle line:84% So what you do is there, you're able to get it accepted 00:34:27.750 --> 00:34:29.648 align:middle line:90% with probability 1 here. 00:34:29.648 --> 00:34:31.940 align:middle line:84% And up here, you can get it accepted with probability 1 00:34:31.940 --> 00:34:34.560 align:middle line:84% and even take a vacation, some vacation time. 00:34:34.560 --> 00:34:36.820 align:middle line:90% So here, you take some time off. 00:34:36.820 --> 00:34:39.659 align:middle line:84% But any of these papers that are super high quality, they get 00:34:39.659 --> 00:34:42.960 align:middle line:84% accepted with probability 1, and you just do the-- 00:34:42.960 --> 00:34:50.155 align:middle line:84% you set-- you choose h of tr star is just equal to the thing 00:34:50.155 --> 00:34:52.780 align:middle line:84% that you're being asked to do so that even if you get the worst 00:34:52.780 --> 00:34:55.389 align:middle line:84% data draw possible, the paper gets accepted with probability 00:34:55.389 --> 00:34:56.139 align:middle line:90% 1. 00:34:56.139 --> 00:34:59.710 align:middle line:84% So you ensure that your paper gets in. 00:34:59.710 --> 00:35:03.400 align:middle line:84% In my model, people have these lexicographic preferences 00:35:03.400 --> 00:35:06.500 align:middle line:84% over paper acceptance and leisure. 00:35:06.500 --> 00:35:08.042 align:middle line:84% So anything that adds any probability 00:35:08.042 --> 00:35:09.792 align:middle line:84% to your paper getting accepted, you do it, 00:35:09.792 --> 00:35:12.460 align:middle line:84% and you only take leisure when your probability of getting in 00:35:12.460 --> 00:35:16.930 align:middle line:84% is 0 or when it's 0 marginal impact. 00:35:16.930 --> 00:35:17.540 align:middle line:90% Anyway. 00:35:17.540 --> 00:35:22.420 align:middle line:84% So this is what we think would happen. 00:35:22.420 --> 00:35:26.170 align:middle line:84% And so that's what an equilibrium would look like. 00:35:26.170 --> 00:35:28.750 align:middle line:84% And then the question about, where 00:35:28.750 --> 00:35:31.450 align:middle line:84% do these dynamics come from, because it's 00:35:31.450 --> 00:35:33.980 align:middle line:84% gotten harder and harder to publish papers over the years. 00:35:33.980 --> 00:35:36.310 align:middle line:90% Why do I think that happens? 00:35:36.310 --> 00:35:39.350 align:middle line:84% What I then add is an overconfidence bias. 00:35:39.350 --> 00:35:41.980 align:middle line:84% So everybody thinks their papers are 00:35:41.980 --> 00:35:43.630 align:middle line:90% better than they actually are. 00:35:43.630 --> 00:35:45.638 align:middle line:84% So either you think your q is better 00:35:45.638 --> 00:35:47.680 align:middle line:84% than it really is, your idea is more interesting, 00:35:47.680 --> 00:35:49.690 align:middle line:84% or you think your craftsmanship is better 00:35:49.690 --> 00:35:52.338 align:middle line:84% than your craftsmanship really is because you're using the best 00:35:52.338 --> 00:35:54.880 align:middle line:84% techniques you know about and you think, therefore, I'm state 00:35:54.880 --> 00:35:56.810 align:middle line:90% of the art, when you're not. 00:35:56.810 --> 00:35:59.410 align:middle line:84% But everybody just thinks their paper 00:35:59.410 --> 00:36:04.570 align:middle line:84% is epsilon better than other people think their paper is. 00:36:04.570 --> 00:36:06.160 align:middle line:90% OK? 00:36:06.160 --> 00:36:11.920 align:middle line:84% And so what's going to happen when people are learning is, 00:36:11.920 --> 00:36:16.080 align:middle line:84% this norm with an overconfidence bias 00:36:16.080 --> 00:36:19.435 align:middle line:84% is going to confuse people because they're going to be-- 00:36:19.435 --> 00:36:21.310 align:middle line:84% two things are going to be happening to them. 00:36:21.310 --> 00:36:23.160 align:middle line:84% One is, they're going to be submitting papers, 00:36:23.160 --> 00:36:24.827 align:middle line:84% and they're going to be told that you've 00:36:24.827 --> 00:36:27.690 align:middle line:84% got to reach this really, really high-- well, what 00:36:27.690 --> 00:36:30.290 align:middle line:84% seems to you to be a really, really high line because you 00:36:30.290 --> 00:36:32.790 align:middle line:84% think your paper is better than it is, and you're like, wow, 00:36:32.790 --> 00:36:34.620 align:middle line:84% it's really, really hard to get a paper in. 00:36:34.620 --> 00:36:36.570 align:middle line:84% For my paper to get accepted, I have 00:36:36.570 --> 00:36:38.610 align:middle line:84% to do this laundry list of 67 things 00:36:38.610 --> 00:36:40.410 align:middle line:90% better than I'm doing today. 00:36:40.410 --> 00:36:44.370 align:middle line:84% And yet you also referee papers and you say, 00:36:44.370 --> 00:36:45.850 align:middle line:90% this is not as good as mine. 00:36:45.850 --> 00:36:47.080 align:middle line:90% You should reject it. 00:36:47.080 --> 00:36:49.000 align:middle line:84% And then you see the journal taking the paper. 00:36:49.000 --> 00:36:51.250 align:middle line:84% And so you have this sort of cognitive dissonance. 00:36:51.250 --> 00:36:53.680 align:middle line:84% It's like, why is it that I can't get my papers in 00:36:53.680 --> 00:36:56.010 align:middle line:84% and these mediocre things either that I refereed 00:36:56.010 --> 00:36:58.950 align:middle line:84% or I see in the journal are getting published? 00:36:58.950 --> 00:37:01.380 align:middle line:84% And so you have to try to reconcile those two 00:37:01.380 --> 00:37:03.240 align:middle line:84% things in your mind and think, maybe I 00:37:03.240 --> 00:37:05.245 align:middle line:84% misunderstood the alpha z process. 00:37:05.245 --> 00:37:08.790 align:middle line:90% 00:37:08.790 --> 00:37:13.260 align:middle line:84% Again, here, I could do something rational where-- 00:37:13.260 --> 00:37:15.480 align:middle line:84% something fully rational where people 00:37:15.480 --> 00:37:19.170 align:middle line:84% have this prior overall alphas and z's 00:37:19.170 --> 00:37:23.310 align:middle line:84% that the world might have and a prior over the whole process 00:37:23.310 --> 00:37:28.390 align:middle line:84% that is being used to generate acceptance and rejections. 00:37:28.390 --> 00:37:30.600 align:middle line:84% And think about the rational Bayesian updating. 00:37:30.600 --> 00:37:33.780 align:middle line:84% When you're getting these two conflicting pieces of evidence, 00:37:33.780 --> 00:37:34.980 align:middle line:90% how do you best-- 00:37:34.980 --> 00:37:37.380 align:middle line:84% in some sense, it's a Bayesian updating 00:37:37.380 --> 00:37:39.480 align:middle line:84% where the true state of the world 00:37:39.480 --> 00:37:44.790 align:middle line:84% is outside the support of your prior 00:37:44.790 --> 00:37:46.950 align:middle line:84% because you don't think that you're biased. 00:37:46.950 --> 00:37:51.060 align:middle line:84% But anyway, what I did instead was think about just 00:37:51.060 --> 00:37:52.510 align:middle line:90% a different objective function. 00:37:52.510 --> 00:37:57.040 align:middle line:84% That is, I'm basically getting two types of evidence. 00:37:57.040 --> 00:38:04.020 align:middle line:84% One is, I'm being asked, these are the r of q's on my paper. 00:38:04.020 --> 00:38:08.365 align:middle line:84% And I'm trying to figure out, what is the correct-- 00:38:08.365 --> 00:38:10.490 align:middle line:84% I'm trying to figure out, what is the correct line? 00:38:10.490 --> 00:38:12.130 align:middle line:84% And this is what I think r of q is 00:38:12.130 --> 00:38:14.980 align:middle line:90% on the observations on my paper. 00:38:14.980 --> 00:38:17.770 align:middle line:84% And I'm going to say that people are in some-- one thing they're 00:38:17.770 --> 00:38:21.040 align:middle line:84% trying to do is fit a line to the data 00:38:21.040 --> 00:38:22.720 align:middle line:84% that they're getting on their own papers 00:38:22.720 --> 00:38:25.307 align:middle line:84% with some least squares procedure. 00:38:25.307 --> 00:38:26.890 align:middle line:84% But then the other thing they're doing 00:38:26.890 --> 00:38:30.950 align:middle line:84% is they're seeing other papers are accepted or rejected. 00:38:30.950 --> 00:38:33.580 align:middle line:84% And every time a paper-- if this paper 00:38:33.580 --> 00:38:37.630 align:middle line:84% is accepted but this paper is also accepted, 00:38:37.630 --> 00:38:40.390 align:middle line:84% and this one is rejected and this one is rejected 00:38:40.390 --> 00:38:44.657 align:middle line:84% and this one is rejected, and this one's accepted, 00:38:44.657 --> 00:38:46.240 align:middle line:84% they're trying to, in some sense, also 00:38:46.240 --> 00:38:49.570 align:middle line:84% fit a line to this data that separates the accepted papers 00:38:49.570 --> 00:38:52.060 align:middle line:90% from the rejected papers. 00:38:52.060 --> 00:38:54.730 align:middle line:84% And what I do, entirely to make the model tractable, 00:38:54.730 --> 00:38:57.430 align:middle line:84% is say that you're using a least squares loss 00:38:57.430 --> 00:39:00.490 align:middle line:84% function for the things you're told about your own papers, 00:39:00.490 --> 00:39:03.040 align:middle line:84% and you're using a sum of absolute deviations loss 00:39:03.040 --> 00:39:05.930 align:middle line:84% function for the acceptance and rejections 00:39:05.930 --> 00:39:08.703 align:middle line:84% and trying to fit the two data together with the same line. 00:39:08.703 --> 00:39:11.120 align:middle line:84% And what's going on is that they're-- actually, basically, 00:39:11.120 --> 00:39:12.200 align:middle line:90% there are-- 00:39:12.200 --> 00:39:14.550 align:middle line:84% what's happening is in reality is there are two lines. 00:39:14.550 --> 00:39:18.170 align:middle line:84% This is the r of q line, what you perceive 00:39:18.170 --> 00:39:19.970 align:middle line:84% you're being asked to do, and this 00:39:19.970 --> 00:39:22.950 align:middle line:84% is the acceptance-rejection line, which is lower. 00:39:22.950 --> 00:39:24.620 align:middle line:84% And you think you're being asked to do 00:39:24.620 --> 00:39:26.995 align:middle line:84% more because you think your quality is better than it is. 00:39:26.995 --> 00:39:29.863 align:middle line:84% So the data is really coming in on these two lines, 00:39:29.863 --> 00:39:32.030 align:middle line:84% and you're trying to fit a single line to this data, 00:39:32.030 --> 00:39:34.460 align:middle line:84% not recognizing that the observations from your papers 00:39:34.460 --> 00:39:35.918 align:middle line:84% are different from the observations 00:39:35.918 --> 00:39:37.160 align:middle line:90% on the other papers. 00:39:37.160 --> 00:39:39.800 align:middle line:84% And the observation of the paper is 00:39:39.800 --> 00:39:43.650 align:middle line:84% if people have views on this, what happens? 00:39:43.650 --> 00:39:50.010 align:middle line:84% What happens is, people perceive this line to be the line. 00:39:50.010 --> 00:39:52.810 align:middle line:84% And everyone thinks it's harder to publish than it is. 00:39:52.810 --> 00:39:55.900 align:middle line:84% They always tell the authors when they're refereeing, 00:39:55.900 --> 00:39:57.990 align:middle line:84% you need to get really, really good to get in, 00:39:57.990 --> 00:40:01.910 align:middle line:84% and you need to reach this high r of q line. 00:40:01.910 --> 00:40:04.370 align:middle line:84% But then the editors can't fill the journal 00:40:04.370 --> 00:40:06.990 align:middle line:84% with papers who reach the high r of q line. 00:40:06.990 --> 00:40:08.510 align:middle line:84% So there are these people down here 00:40:08.510 --> 00:40:12.320 align:middle line:84% who had the medium quality ideas that revised them 00:40:12.320 --> 00:40:15.320 align:middle line:84% to death, who fell just short of what the referees told 00:40:15.320 --> 00:40:15.850 align:middle line:90% them to do. 00:40:15.850 --> 00:40:17.600 align:middle line:84% And then the editors take the paper anyway 00:40:17.600 --> 00:40:19.980 align:middle line:84% because they need to fill the journal. 00:40:19.980 --> 00:40:28.120 align:middle line:84% And so they take the papers that were slightly lower quality 00:40:28.120 --> 00:40:30.360 align:middle line:84% on the r dimension than what they were told to do. 00:40:30.360 --> 00:40:33.040 align:middle line:84% So that's what the data actually looks like. 00:40:33.040 --> 00:40:36.340 align:middle line:84% And then when people are trying to fit these two data 00:40:36.340 --> 00:40:39.220 align:middle line:84% sources together, the r of q data 00:40:39.220 --> 00:40:42.130 align:middle line:84% and the acceptance-rejection data, the best way 00:40:42.130 --> 00:40:47.800 align:middle line:84% to do that is by tilting the line down because the surprise 00:40:47.800 --> 00:40:52.810 align:middle line:84% acceptances are papers that are low q, high r, and get in. 00:40:52.810 --> 00:40:54.670 align:middle line:84% And so the best way that you can explain 00:40:54.670 --> 00:40:56.470 align:middle line:84% that on these really great papers, 00:40:56.470 --> 00:40:59.650 align:middle line:84% I'm being asked to do a lot, and yet down here I 00:40:59.650 --> 00:41:03.550 align:middle line:84% see these low q, high r papers getting in, is to say, 00:41:03.550 --> 00:41:07.900 align:middle line:84% oh, I think it must be that the best way to fit this data 00:41:07.900 --> 00:41:11.500 align:middle line:84% is to tilt the line down and think that r quality is 00:41:11.500 --> 00:41:13.250 align:middle line:90% more important than I thought. 00:41:13.250 --> 00:41:16.570 align:middle line:84% And these high r papers are getting in 00:41:16.570 --> 00:41:18.970 align:middle line:90% because r quality matters more. 00:41:18.970 --> 00:41:21.490 align:middle line:84% And so the argument is that what's going to happen is, 00:41:21.490 --> 00:41:24.430 align:middle line:84% people are just going to keep being surprised that papers get 00:41:24.430 --> 00:41:26.390 align:middle line:90% accepted that they refereed. 00:41:26.390 --> 00:41:29.330 align:middle line:84% The surprise acceptances are overwhelmingly low q, 00:41:29.330 --> 00:41:30.155 align:middle line:90% high r papers. 00:41:30.155 --> 00:41:32.870 align:middle line:90% 00:41:32.870 --> 00:41:35.960 align:middle line:84% And so the best way to explain that those surprise acceptances 00:41:35.960 --> 00:41:39.560 align:middle line:84% and yet the high demands being made for the high q papers 00:41:39.560 --> 00:41:42.200 align:middle line:84% are tilting the line and thinking r quality 00:41:42.200 --> 00:41:44.070 align:middle line:84% is a little bit more important than I am. 00:41:44.070 --> 00:41:45.570 align:middle line:84% And so we end up with this situation 00:41:45.570 --> 00:41:49.760 align:middle line:84% where just perpetually, everyone always is holding people 00:41:49.760 --> 00:41:53.617 align:middle line:84% to standards that are slightly too high, and then slightly 00:41:53.617 --> 00:41:55.700 align:middle line:84% twisting their views to think about the importance 00:41:55.700 --> 00:41:58.610 align:middle line:84% of craftsmanship versus the importance of ideas. 00:41:58.610 --> 00:42:03.540 align:middle line:84% And this graph shows you-- if you 00:42:03.540 --> 00:42:06.570 align:middle line:84% think of this as the set of social norms 00:42:06.570 --> 00:42:08.910 align:middle line:84% that would all be consistent with alpha on this axis 00:42:08.910 --> 00:42:11.880 align:middle line:84% and z on this axis, we're always staying slightly 00:42:11.880 --> 00:42:16.300 align:middle line:90% above that set of social norms. 00:42:16.300 --> 00:42:18.000 align:middle line:84% And then this dynamics of this model 00:42:18.000 --> 00:42:19.860 align:middle line:84% would just slowly drift and drift and drift 00:42:19.860 --> 00:42:20.980 align:middle line:90% and drift and drift. 00:42:20.980 --> 00:42:23.400 align:middle line:90% And the argument is that-- 00:42:23.400 --> 00:42:26.610 align:middle line:84% I'm going to argue that this was a 30-year disequilibrium 00:42:26.610 --> 00:42:27.220 align:middle line:90% process. 00:42:27.220 --> 00:42:29.290 align:middle line:84% For 30 years, we've been out of equilibrium. 00:42:29.290 --> 00:42:32.160 align:middle line:84% Referees are always slightly too high. 00:42:32.160 --> 00:42:34.410 align:middle line:84% I'm just trying to reconcile the difference 00:42:34.410 --> 00:42:37.230 align:middle line:84% in their mind between the slightly too high standards they 00:42:37.230 --> 00:42:39.777 align:middle line:84% think are being applied with the acceptances they see. 00:42:39.777 --> 00:42:42.360 align:middle line:84% They just keep thinking that it must be the craftsmanship that 00:42:42.360 --> 00:42:43.173 align:middle line:90% got that one in. 00:42:43.173 --> 00:42:45.090 align:middle line:84% And then they keep just changing their beliefs 00:42:45.090 --> 00:42:46.900 align:middle line:84% about the importance of craftsmanship. 00:42:46.900 --> 00:42:51.420 align:middle line:90% 00:42:51.420 --> 00:42:53.240 align:middle line:90% Any questions there? 00:42:53.240 --> 00:42:55.250 align:middle line:84% So again, the rule-of-thumb approach 00:42:55.250 --> 00:42:58.970 align:middle line:84% here is just I'm trying to say what the referees do. 00:42:58.970 --> 00:43:03.320 align:middle line:84% And I've actually-- some chance I will rewrite this paper doing 00:43:03.320 --> 00:43:04.820 align:middle line:84% the fully Bayesian version, which 00:43:04.820 --> 00:43:07.520 align:middle line:84% I think there's been big advances in Bayesian learning 00:43:07.520 --> 00:43:11.120 align:middle line:84% with misspecified priors in the last 20 years, 00:43:11.120 --> 00:43:13.460 align:middle line:84% where I think maybe now I could solve this Bayesian 00:43:13.460 --> 00:43:16.220 align:middle line:84% model with misspecified priors and show that kind of what 00:43:16.220 --> 00:43:19.100 align:middle line:84% I say happens here, happens in a fully Bayesian model 00:43:19.100 --> 00:43:20.258 align:middle line:90% with misspecified priors. 00:43:20.258 --> 00:43:22.550 align:middle line:84% But I think that the idea of the rule-of-thumb approach 00:43:22.550 --> 00:43:26.132 align:middle line:84% is just to say, think about what you think people would do. 00:43:26.132 --> 00:43:28.590 align:middle line:84% Think about what you think they would do that's reasonable. 00:43:28.590 --> 00:43:29.882 align:middle line:90% I don't think it's reasonable-- 00:43:29.882 --> 00:43:32.372 align:middle line:84% I think it seems like referees are fairly naive, 00:43:32.372 --> 00:43:34.830 align:middle line:84% and this is how they're trying to reconcile the two things. 00:43:34.830 --> 00:43:35.580 align:middle line:90% This could happen. 00:43:35.580 --> 00:43:38.420 align:middle line:84% And then write down a model that you can actually solve. 00:43:38.420 --> 00:43:41.205 align:middle line:84% And like, why do I have a quadratic loss 00:43:41.205 --> 00:43:43.580 align:middle line:84% function in one part and a least absolute deviations loss 00:43:43.580 --> 00:43:44.455 align:middle line:90% function in one part? 00:43:44.455 --> 00:43:47.370 align:middle line:84% And it's exactly because that was the model I could solve. 00:43:47.370 --> 00:43:48.860 align:middle line:84% We might as well solve-- write down 00:43:48.860 --> 00:43:54.100 align:middle line:84% models we can solve and say what we can about them. 00:43:54.100 --> 00:43:56.680 align:middle line:84% The, let me say, empirics on this, 00:43:56.680 --> 00:44:00.820 align:middle line:84% I have another kind of depressing paper in the JPE, 00:44:00.820 --> 00:44:03.203 align:middle line:84% looking at how long it takes a paper 00:44:03.203 --> 00:44:04.870 align:middle line:84% to get accepted from the time it's first 00:44:04.870 --> 00:44:07.450 align:middle line:90% submitted historically. 00:44:07.450 --> 00:44:13.300 align:middle line:84% And in the 1970s, papers were typically getting accepted 00:44:13.300 --> 00:44:16.270 align:middle line:84% at all the top journals 6 to 8 months after they were first 00:44:16.270 --> 00:44:17.550 align:middle line:90% accepted. 00:44:17.550 --> 00:44:20.200 align:middle line:84% And it's kind of remarkable when you realize, at the time, 00:44:20.200 --> 00:44:22.780 align:middle line:84% the Xerox machine had not been invented. 00:44:22.780 --> 00:44:25.640 align:middle line:84% And so how did refereeing a paper work? 00:44:25.640 --> 00:44:30.850 align:middle line:84% You would type up two or three copies of your paper. 00:44:30.850 --> 00:44:32.758 align:middle line:84% You would send-- or actually, in the 1960s, 00:44:32.758 --> 00:44:35.050 align:middle line:84% you would send in a single draft of the manuscript that 00:44:35.050 --> 00:44:36.340 align:middle line:90% your secretary typed. 00:44:36.340 --> 00:44:38.470 align:middle line:84% But you would send the paper in-- 00:44:38.470 --> 00:44:41.890 align:middle line:84% so '60s, certainly, your secretary typed one copy 00:44:41.890 --> 00:44:43.810 align:middle line:84% of the paper, you sent it to the journal, 00:44:43.810 --> 00:44:45.620 align:middle line:90% they sent it to the referee. 00:44:45.620 --> 00:44:48.513 align:middle line:84% The referee read the paper, made comments, 00:44:48.513 --> 00:44:49.930 align:middle line:84% sent it back to the journal office 00:44:49.930 --> 00:44:51.800 align:middle line:90% because that was the only copy. 00:44:51.800 --> 00:44:54.590 align:middle line:84% The journal office then sent the paper to a second referee. 00:44:54.590 --> 00:44:56.900 align:middle line:84% That second referee made comments, sent it back 00:44:56.900 --> 00:44:58.110 align:middle line:90% to the journal office. 00:44:58.110 --> 00:45:02.575 align:middle line:84% The journal office made comments and mailed you a decision. 00:45:02.575 --> 00:45:03.950 align:middle line:84% Somehow, at the time, papers were 00:45:03.950 --> 00:45:07.850 align:middle line:84% getting published 5, 6 months after they were submitted. 00:45:07.850 --> 00:45:10.910 align:middle line:84% Starting in the '70s, when the revise and resubmit was 00:45:10.910 --> 00:45:14.180 align:middle line:84% invented, things just got longer and longer and longer 00:45:14.180 --> 00:45:14.910 align:middle line:90% and longer. 00:45:14.910 --> 00:45:18.772 align:middle line:84% And by the year 2000, some journals were averaging 2 00:45:18.772 --> 00:45:21.230 align:middle line:84% and 1/2 years from the time a paper was submitted to get it 00:45:21.230 --> 00:45:24.990 align:middle line:84% published, some were averaging two years, and so on. 00:45:24.990 --> 00:45:30.253 align:middle line:84% And this is the time that's conditional on a good outcome. 00:45:30.253 --> 00:45:31.670 align:middle line:84% The way the process tends to work, 00:45:31.670 --> 00:45:36.120 align:middle line:84% though, is the top journals have a 96% rejection rate. 00:45:36.120 --> 00:45:38.902 align:middle line:84% So you send it to journal one, they screw around with it 00:45:38.902 --> 00:45:40.860 align:middle line:84% for a year and a half, and then they reject it. 00:45:40.860 --> 00:45:42.110 align:middle line:84% And then the second one, and then they 00:45:42.110 --> 00:45:44.100 align:middle line:84% mess around with it for six months, and they reject it. 00:45:44.100 --> 00:45:46.170 align:middle line:84% And the third one, it's a year, and then they reject it. 00:45:46.170 --> 00:45:47.545 align:middle line:84% And then finally, you get someone 00:45:47.545 --> 00:45:50.360 align:middle line:84% to spend three years revising it before they then-- 00:45:50.360 --> 00:45:52.310 align:middle line:84% two or three years, and then they publish it. 00:45:52.310 --> 00:45:55.670 align:middle line:84% So actually, I have this very depressing statistic 00:45:55.670 --> 00:45:59.450 align:middle line:84% that if you look at the absolute stars on the IO market 00:45:59.450 --> 00:46:03.380 align:middle line:84% who do empirical IO and how long it takes them to get their job 00:46:03.380 --> 00:46:06.650 align:middle line:84% market papers published, the median time is now 00:46:06.650 --> 00:46:10.430 align:middle line:84% 7 years, which, you know, it's a long time, 00:46:10.430 --> 00:46:12.920 align:middle line:84% because then you spent also three years writing that paper 00:46:12.920 --> 00:46:15.540 align:middle line:84% before you first graduated from school. 00:46:15.540 --> 00:46:16.760 align:middle line:90% So. 00:46:16.760 --> 00:46:19.915 align:middle line:84% Anyway, I hope future generations 00:46:19.915 --> 00:46:22.290 align:middle line:84% will recognize what they're doing and solve this problem. 00:46:22.290 --> 00:46:26.750 align:middle line:84% And I've made some attempts with limited success. 00:46:26.750 --> 00:46:29.210 align:middle line:90% Any questions? 00:46:29.210 --> 00:46:29.770 align:middle line:90% Yeah. 00:46:29.770 --> 00:46:31.520 align:middle line:84% AUDIENCE: How important is it in the model 00:46:31.520 --> 00:46:33.380 align:middle line:84% that the authors perfectly observe 00:46:33.380 --> 00:46:36.320 align:middle line:90% the quality of their ideas? 00:46:36.320 --> 00:46:38.798 align:middle line:84% GLENN ELLISON: So I don't think it's at all important. 00:46:38.798 --> 00:46:40.340 align:middle line:84% I think what's going on is I'm having 00:46:40.340 --> 00:46:49.200 align:middle line:84% the authors do this sort of least squares fit of this. 00:46:49.200 --> 00:46:51.260 align:middle line:84% So they're being asked to do-- every time 00:46:51.260 --> 00:46:54.590 align:middle line:84% you submit a paper with quality q, you're being asked to r of q. 00:46:54.590 --> 00:46:58.890 align:middle line:84% And you just fit a least squares line to that. 00:46:58.890 --> 00:47:01.166 align:middle line:90% So if there was sort of-- 00:47:01.166 --> 00:47:06.230 align:middle line:84% if your paper qualities were like this, fitting a line to it, 00:47:06.230 --> 00:47:08.720 align:middle line:84% I think, would do roughly the same thing. 00:47:08.720 --> 00:47:12.740 align:middle line:84% What's important is the argument that the very high quality 00:47:12.740 --> 00:47:14.450 align:middle line:84% papers, you're asked to do a lot, 00:47:14.450 --> 00:47:16.520 align:middle line:84% and they do all get accepted because you 00:47:16.520 --> 00:47:19.010 align:middle line:90% do what you're being asked. 00:47:19.010 --> 00:47:21.170 align:middle line:84% And the low quality ones are the ones 00:47:21.170 --> 00:47:24.480 align:middle line:84% that get in not doing what they were asked. 00:47:24.480 --> 00:47:26.852 align:middle line:84% So the high quality ones don't get 00:47:26.852 --> 00:47:29.060 align:middle line:84% in not doing what they were asked because they always 00:47:29.060 --> 00:47:30.780 align:middle line:90% do what they were asked enough. 00:47:30.780 --> 00:47:33.050 align:middle line:90% That's, I think, the main thing. 00:47:33.050 --> 00:47:35.750 align:middle line:84% But the randomness I don't think would affect-- 00:47:35.750 --> 00:47:38.300 align:middle line:84% on either one shouldn't affect it. 00:47:38.300 --> 00:47:41.150 align:middle line:84% Let me say, advertising is not something 00:47:41.150 --> 00:47:44.810 align:middle line:84% I can do any-- do justice in this paper. 00:47:44.810 --> 00:47:46.448 align:middle line:84% There is a field called marketing. 00:47:46.448 --> 00:47:48.740 align:middle line:84% I don't know if any of you are from the Sloan Marketing 00:47:48.740 --> 00:47:51.140 align:middle line:84% Department, but most business schools 00:47:51.140 --> 00:47:52.500 align:middle line:90% have marketing departments. 00:47:52.500 --> 00:47:55.153 align:middle line:84% Marketing departments have lots and lots of professors. 00:47:55.153 --> 00:47:57.070 align:middle line:84% There are like a thousand marketing professors 00:47:57.070 --> 00:47:58.130 align:middle line:90% in the United States. 00:47:58.130 --> 00:47:59.830 align:middle line:84% There are many, many marketing journals, 00:47:59.830 --> 00:48:01.610 align:middle line:84% just like there are economics journals. 00:48:01.610 --> 00:48:04.370 align:middle line:84% Those marketing journals are full of papers. 00:48:04.370 --> 00:48:07.840 align:middle line:84% And obviously, people in marketing do advertising. 00:48:07.840 --> 00:48:09.910 align:middle line:84% They teach entire-- all of their courses 00:48:09.910 --> 00:48:12.810 align:middle line:84% are entirely marketing, not one week of my class or something 00:48:12.810 --> 00:48:13.310 align:middle line:90% like that. 00:48:13.310 --> 00:48:16.930 align:middle line:84% So there's an enormous literature out there. 00:48:16.930 --> 00:48:19.480 align:middle line:90% Marketing is a field where-- 00:48:19.480 --> 00:48:22.670 align:middle line:84% I put here influences on marketing. 00:48:22.670 --> 00:48:25.000 align:middle line:84% So if you look at what's being cited by marketing 00:48:25.000 --> 00:48:28.305 align:middle line:90% papers, the largest-- 00:48:28.305 --> 00:48:30.680 align:middle line:84% well, obviously, mostly they cite other marketing papers. 00:48:30.680 --> 00:48:33.970 align:middle line:84% So this is, what non-business things are they citing? 00:48:33.970 --> 00:48:36.040 align:middle line:84% They do cite a lot of psychology papers, 00:48:36.040 --> 00:48:39.370 align:middle line:84% and there's been of a rise of psychology in marketing over 00:48:39.370 --> 00:48:43.210 align:middle line:84% the last 40-year time span, where people in marketing are 00:48:43.210 --> 00:48:46.330 align:middle line:84% paying more and more attention to what psychologists are saying 00:48:46.330 --> 00:48:48.200 align:middle line:84% and citing what psychologists are saying. 00:48:48.200 --> 00:48:50.350 align:middle line:84% There's also been an increasing trend 00:48:50.350 --> 00:48:52.880 align:middle line:84% in citing what economists are doing. 00:48:52.880 --> 00:48:57.040 align:middle line:84% And so now it's 6% of the citations in marketing papers 00:48:57.040 --> 00:48:59.780 align:middle line:84% are to economics papers, rather than to marketing papers. 00:48:59.780 --> 00:49:02.540 align:middle line:84% So there's a substantial influence 00:49:02.540 --> 00:49:04.170 align:middle line:90% of economics on marketing. 00:49:04.170 --> 00:49:06.980 align:middle line:84% And if you look at what within economics are marketing people 00:49:06.980 --> 00:49:10.400 align:middle line:84% citing, IO, which is the yellow line here, 00:49:10.400 --> 00:49:12.630 align:middle line:84% is a big part of what's going on in marketing. 00:49:12.630 --> 00:49:13.522 align:middle line:90% And so there's just-- 00:49:13.522 --> 00:49:15.980 align:middle line:84% there are marketing papers that look like psychology papers 00:49:15.980 --> 00:49:17.563 align:middle line:84% or look very different from economics. 00:49:17.563 --> 00:49:19.340 align:middle line:84% But there are a lot of marketing papers 00:49:19.340 --> 00:49:21.980 align:middle line:84% that look like IO theory papers and that look 00:49:21.980 --> 00:49:23.550 align:middle line:90% like empirical IO papers. 00:49:23.550 --> 00:49:27.050 align:middle line:84% So they're just-- marketing people write all about the way 00:49:27.050 --> 00:49:31.130 align:middle line:84% that firms can price or make their products appealing 00:49:31.130 --> 00:49:32.220 align:middle line:90% to customers. 00:49:32.220 --> 00:49:35.720 align:middle line:84% And they write a lot about empirical demand estimation 00:49:35.720 --> 00:49:38.820 align:middle line:84% and how advertising campaigns affect demand and so on. 00:49:38.820 --> 00:49:42.025 align:middle line:84% And both of those parts look a lot like IO. 00:49:42.025 --> 00:49:43.400 align:middle line:84% So what I'm going to do today is, 00:49:43.400 --> 00:49:45.980 align:middle line:84% today, I'm just going to do a few classic models 00:49:45.980 --> 00:49:47.450 align:middle line:90% of advertising. 00:49:47.450 --> 00:49:49.800 align:middle line:84% And then it's, again, a week from Monday, 00:49:49.800 --> 00:49:54.160 align:middle line:84% I'm going to do some empirics on advertising. 00:49:54.160 --> 00:49:56.410 align:middle line:90% So models of advertising. 00:49:56.410 --> 00:49:58.210 align:middle line:84% Generally, I think you can see the way 00:49:58.210 --> 00:50:03.880 align:middle line:84% people in the economics literature break down three 00:50:03.880 --> 00:50:04.780 align:middle line:90% different-- 00:50:04.780 --> 00:50:06.625 align:middle line:84% obviously, to explain advertising, 00:50:06.625 --> 00:50:08.500 align:middle line:84% you've got to say, why is it that advertising 00:50:08.500 --> 00:50:10.070 align:middle line:90% changes what people purchase? 00:50:10.070 --> 00:50:12.370 align:middle line:84% If you start with our basic model of consumers 00:50:12.370 --> 00:50:15.160 align:middle line:90% have preferences ui jt. 00:50:15.160 --> 00:50:18.250 align:middle line:84% If they buy product j at time t and it's 00:50:18.250 --> 00:50:22.450 align:middle line:84% xjt beta minus alpha pjt and they buy the best thing, 00:50:22.450 --> 00:50:26.195 align:middle line:84% then what does an advertisement do if it doesn't change your-- 00:50:26.195 --> 00:50:27.820 align:middle line:84% if that's your utility function, you're 00:50:27.820 --> 00:50:29.158 align:middle line:90% always maximizing your utility. 00:50:29.158 --> 00:50:30.700 align:middle line:84% Advertising shouldn't affect anything 00:50:30.700 --> 00:50:32.260 align:middle line:90% that goes on in the world. 00:50:32.260 --> 00:50:34.510 align:middle line:84% And yet we see firms spend lots and lots 00:50:34.510 --> 00:50:35.690 align:middle line:90% of money on advertising. 00:50:35.690 --> 00:50:39.640 align:middle line:84% It must be advertising changes behavior or changes something 00:50:39.640 --> 00:50:40.970 align:middle line:90% that increases profits. 00:50:40.970 --> 00:50:42.490 align:middle line:90% What is it? 00:50:42.490 --> 00:50:46.000 align:middle line:84% One thing, one class of models is advertising 00:50:46.000 --> 00:50:48.010 align:middle line:90% can have informational content. 00:50:48.010 --> 00:50:51.550 align:middle line:84% It can tell you that a product exists, 00:50:51.550 --> 00:50:53.200 align:middle line:84% what the product's price is, what 00:50:53.200 --> 00:50:55.050 align:middle line:90% the product's attributes are. 00:50:55.050 --> 00:50:57.055 align:middle line:84% But in some sense, it's like a search model. 00:50:57.055 --> 00:50:59.180 align:middle line:84% You don't know everything, but then the advertising 00:50:59.180 --> 00:51:01.160 align:middle line:84% tells you things that you would otherwise 00:51:01.160 --> 00:51:05.090 align:middle line:84% have to invest search effort in figuring out, 00:51:05.090 --> 00:51:07.730 align:middle line:84% and that increases the number of people who buy the product. 00:51:07.730 --> 00:51:11.630 align:middle line:90% And classic examples like this. 00:51:11.630 --> 00:51:12.870 align:middle line:90% Informative advertising. 00:51:12.870 --> 00:51:16.680 align:middle line:84% You just get an ad in the mail or in your newspaper for Target. 00:51:16.680 --> 00:51:18.650 align:middle line:84% And the advertisement for Target tells you 00:51:18.650 --> 00:51:23.300 align:middle line:84% you can get these boots for $10 and those jackets for $15 or $20 00:51:23.300 --> 00:51:25.430 align:middle line:90% and these clothes for $4 or $5. 00:51:25.430 --> 00:51:26.930 align:middle line:84% They show you the pictures of things 00:51:26.930 --> 00:51:29.760 align:middle line:84% to tell you what Target is selling and what it costs. 00:51:29.760 --> 00:51:33.680 align:middle line:84% And this would be informative advertising. 00:51:33.680 --> 00:51:35.990 align:middle line:84% A second thing that advertising can do 00:51:35.990 --> 00:51:39.410 align:middle line:84% is not necessarily tell you anything about the product 00:51:39.410 --> 00:51:44.360 align:middle line:84% itself, but just signal to you something about the quality. 00:51:44.360 --> 00:51:46.840 align:middle line:84% So for instance, let me look at these two ads. 00:51:46.840 --> 00:51:51.440 align:middle line:84% You guys are probably all too young to remember pets.com. 00:51:51.440 --> 00:51:53.300 align:middle line:84% Pets.com in the first internet boom-- 00:51:53.300 --> 00:51:55.640 align:middle line:84% just like when there was amazon.com, 00:51:55.640 --> 00:51:58.700 align:middle line:84% pets.com started up around the same time. 00:51:58.700 --> 00:52:02.780 align:middle line:84% Pets.com, just a miraculous rise and fall 00:52:02.780 --> 00:52:05.450 align:middle line:84% of a company in the space of a year or two, 00:52:05.450 --> 00:52:08.612 align:middle line:84% where pets.com just burned through hundreds 00:52:08.612 --> 00:52:09.570 align:middle line:90% of millions of dollars. 00:52:09.570 --> 00:52:10.790 align:middle line:90% They had this sock puppet. 00:52:10.790 --> 00:52:13.160 align:middle line:84% And they put the sock puppet on the Super Bowl 00:52:13.160 --> 00:52:16.160 align:middle line:84% and spent millions of dollars just having the sock puppet 00:52:16.160 --> 00:52:17.930 align:middle line:90% just appear on the Super Bowl. 00:52:17.930 --> 00:52:20.660 align:middle line:84% The sock puppet did not tell you anything about-- 00:52:20.660 --> 00:52:21.870 align:middle line:90% obviously, what is pets.com? 00:52:21.870 --> 00:52:23.550 align:middle line:90% They sell stuff for pets. 00:52:23.550 --> 00:52:25.620 align:middle line:84% You know what's in a Petco store. 00:52:25.620 --> 00:52:28.460 align:middle line:84% You probably know what's on pets.com. 00:52:28.460 --> 00:52:30.170 align:middle line:84% But the sock puppet, in some sense, 00:52:30.170 --> 00:52:33.830 align:middle line:84% maybe was sort of saying that, wow, they just spent 00:52:33.830 --> 00:52:38.000 align:middle line:84% $2.4 million to have the sock puppet just interviewing 00:52:38.000 --> 00:52:40.160 align:middle line:90% somebody during the Super Bowl. 00:52:40.160 --> 00:52:42.660 align:middle line:84% It must be they think they have a good product. 00:52:42.660 --> 00:52:44.180 align:middle line:84% And if I actually go to pets.com, 00:52:44.180 --> 00:52:45.950 align:middle line:84% I'll buy stuff because otherwise, 00:52:45.950 --> 00:52:48.110 align:middle line:84% how could they have burned through $2 million 00:52:48.110 --> 00:52:49.920 align:middle line:90% in 30 seconds like that? 00:52:49.920 --> 00:52:52.287 align:middle line:84% And so you get this sort of signaling idea. 00:52:52.287 --> 00:52:53.870 align:middle line:84% I would say the same thing about-- you 00:52:53.870 --> 00:52:56.960 align:middle line:90% look at the Patagonia ad here. 00:52:56.960 --> 00:52:59.210 align:middle line:90% Again, I don't know. 00:52:59.210 --> 00:53:01.670 align:middle line:84% I mean, maybe it's all Photoshopped, 00:53:01.670 --> 00:53:04.940 align:middle line:84% but it must have been expensive to send some guy with a camera 00:53:04.940 --> 00:53:08.240 align:middle line:84% up to the top of some mountain wearing those Patagonia stuff 00:53:08.240 --> 00:53:09.660 align:middle line:90% and taking the pictures. 00:53:09.660 --> 00:53:13.798 align:middle line:84% And it must be that their clothing-- either this person 00:53:13.798 --> 00:53:16.340 align:middle line:84% is really not uncomfortable up there on that mountain wearing 00:53:16.340 --> 00:53:18.410 align:middle line:84% the Patagonia stuff and so the clothing is good, 00:53:18.410 --> 00:53:20.240 align:middle line:84% or they're just spending a lot of money 00:53:20.240 --> 00:53:23.300 align:middle line:84% on the ads because they know that if I bought the clothing, 00:53:23.300 --> 00:53:25.380 align:middle line:84% I would like it or something like that. 00:53:25.380 --> 00:53:28.550 align:middle line:84% And then I guess the third thing we think about advertising is 00:53:28.550 --> 00:53:30.920 align:middle line:84% advertising can be changing your preferences 00:53:30.920 --> 00:53:33.230 align:middle line:84% or directly affecting purchase decisions 00:53:33.230 --> 00:53:34.530 align:middle line:90% of non-rational things. 00:53:34.530 --> 00:53:37.740 align:middle line:84% So if I look at this commercial, "Let's Grab a Beer." 00:53:37.740 --> 00:53:41.510 align:middle line:84% And those people are having just a great time having those 00:53:41.510 --> 00:53:45.470 align:middle line:84% beers outdoors at some thing, and they're all 00:53:45.470 --> 00:53:48.230 align:middle line:84% relatively attractive or whatever, it must be I 00:53:48.230 --> 00:53:52.250 align:middle line:84% think that somehow, if only I was having that beer, 00:53:52.250 --> 00:53:53.930 align:middle line:84% my life would be like that instead 00:53:53.930 --> 00:53:55.680 align:middle line:90% of what my life is like today. 00:53:55.680 --> 00:53:59.880 align:middle line:84% And I would like to be those people. 00:53:59.880 --> 00:54:05.420 align:middle line:84% And somehow, if I just had that beer, this is me, or something. 00:54:05.420 --> 00:54:09.290 align:middle line:84% It changes your preferences from being 00:54:09.290 --> 00:54:11.320 align:middle line:84% a person who wants to work on a problem set 00:54:11.320 --> 00:54:13.820 align:middle line:84% to being a person who wants to be out on the street drinking 00:54:13.820 --> 00:54:14.340 align:middle line:90% beer. 00:54:14.340 --> 00:54:19.410 align:middle line:84% And your tastes change, and therefore your purchases change. 00:54:19.410 --> 00:54:19.910 align:middle line:90% OK. 00:54:19.910 --> 00:54:24.440 align:middle line:84% So how do-- let me see-- how do I model these, 00:54:24.440 --> 00:54:27.270 align:middle line:84% and what do we get out of the model? 00:54:27.270 --> 00:54:31.370 align:middle line:84% So first classic model of informative advertising 00:54:31.370 --> 00:54:34.250 align:middle line:90% is Butters, 1977. 00:54:34.250 --> 00:54:40.110 align:middle line:84% Butters was thinking about N firms selling identical goods. 00:54:40.110 --> 00:54:42.810 align:middle line:84% There's this unit mass of identical consumers, 00:54:42.810 --> 00:54:45.060 align:middle line:84% and they're all going to get utility v minus p 00:54:45.060 --> 00:54:47.970 align:middle line:84% if they buy this good at price p. 00:54:47.970 --> 00:54:52.350 align:middle line:84% All the firms have some cost c of production. 00:54:52.350 --> 00:54:58.290 align:middle line:84% And this is a problem where the consumers don't know either 00:54:58.290 --> 00:55:01.320 align:middle line:84% that the product exists or where they would buy it. 00:55:01.320 --> 00:55:05.670 align:middle line:84% So what the firms can do is they can spend amount of money-- 00:55:05.670 --> 00:55:09.030 align:middle line:84% so if they spend A of x on advertising, 00:55:09.030 --> 00:55:13.560 align:middle line:84% their ads will reach a fraction x of consumers. 00:55:13.560 --> 00:55:17.520 align:middle line:84% So it's, I guess, good to think about this as putting ads out 00:55:17.520 --> 00:55:20.470 align:middle line:84% there on the radio or print airwaves. 00:55:20.470 --> 00:55:23.310 align:middle line:84% And just every ad you put out there, 00:55:23.310 --> 00:55:24.940 align:middle line:84% you're going to hit more consumers. 00:55:24.940 --> 00:55:28.500 align:middle line:84% There's decreasing returns because the first ads 00:55:28.500 --> 00:55:30.340 align:middle line:90% I send out all hit somebody. 00:55:30.340 --> 00:55:35.160 align:middle line:84% But then the later ads that I send out 00:55:35.160 --> 00:55:37.810 align:middle line:84% are mostly hitting people who already heard about my product. 00:55:37.810 --> 00:55:39.840 align:middle line:84% And so by having them hear about my product, 00:55:39.840 --> 00:55:44.320 align:middle line:84% I'm just wasting duplicated messages on them. 00:55:44.320 --> 00:55:46.000 align:middle line:84% And so I'm sending out these messages. 00:55:46.000 --> 00:55:47.910 align:middle line:84% And the more people I want to reach, 00:55:47.910 --> 00:55:50.425 align:middle line:84% the higher is the A of x, the money 00:55:50.425 --> 00:55:51.550 align:middle line:90% I'm going to have to spend. 00:55:51.550 --> 00:55:53.653 align:middle line:84% And it could be that even getting to x equals 1. 00:55:53.653 --> 00:55:55.570 align:middle line:84% This thing could almost asymptote to infinity. 00:55:55.570 --> 00:55:59.370 align:middle line:84% If I'm trying to reach every person in the world, it's hard. 00:55:59.370 --> 00:56:02.280 align:middle line:84% And so they consider a game where firms simultaneously 00:56:02.280 --> 00:56:04.350 align:middle line:84% choose two things, the amount of advertising 00:56:04.350 --> 00:56:07.930 align:middle line:84% x they're going to do, and the price is p. 00:56:07.930 --> 00:56:11.290 align:middle line:84% And then consumers who buy at least one ad 00:56:11.290 --> 00:56:14.980 align:middle line:84% buy from the cheapest firm that they heard about. 00:56:14.980 --> 00:56:16.550 align:middle line:84% And again, the goods are identical. 00:56:16.550 --> 00:56:19.900 align:middle line:84% So it's just, if v minus pj is positive, 00:56:19.900 --> 00:56:24.270 align:middle line:84% then I'll buy it from the cheapest firm I found out there. 00:56:24.270 --> 00:56:27.800 align:middle line:90% 00:56:27.800 --> 00:56:33.470 align:middle line:84% And if I don't hear about a firm, I can't buy from them. 00:56:33.470 --> 00:56:35.030 align:middle line:84% This could be a good model for things 00:56:35.030 --> 00:56:36.962 align:middle line:84% advertised through infomercials on TV also, 00:56:36.962 --> 00:56:37.920 align:middle line:90% or something like that. 00:56:37.920 --> 00:56:39.560 align:middle line:84% It's like, if you see the infomercial, 00:56:39.560 --> 00:56:43.420 align:middle line:84% you can then buy the product and write down the phone number. 00:56:43.420 --> 00:56:43.920 align:middle line:90% OK. 00:56:43.920 --> 00:56:45.290 align:middle line:90% Observations. 00:56:45.290 --> 00:56:47.720 align:middle line:84% Again, this was one of the early papers 00:56:47.720 --> 00:56:50.060 align:middle line:84% that had an equilibrium with price dispersion. 00:56:50.060 --> 00:56:52.770 align:middle line:84% This is like the Spiegler paper I talked about earlier. 00:56:52.770 --> 00:56:55.490 align:middle line:84% This is like Stahl from a few weeks ago. 00:56:55.490 --> 00:56:58.930 align:middle line:84% Again, some consumers are going to only get one ad. 00:56:58.930 --> 00:57:01.430 align:middle line:84% If the consumer only got your ad and nobody else's, then you 00:57:01.430 --> 00:57:05.840 align:middle line:84% want to set p equals v, but if they've got multiple ads, 00:57:05.840 --> 00:57:08.840 align:middle line:84% we're again in this situation where, in equilibrium, we're 00:57:08.840 --> 00:57:10.670 align:middle line:84% going to get some distribution on prices 00:57:10.670 --> 00:57:14.000 align:middle line:84% that has an upper bound of v and then goes down from there. 00:57:14.000 --> 00:57:17.600 align:middle line:84% And it's just determined by, I want to undercut, 00:57:17.600 --> 00:57:22.070 align:middle line:84% but if they only see one ad, I'd like to be the highest-- 00:57:22.070 --> 00:57:24.590 align:middle line:84% I'd like to set at price equals v. But then they 00:57:24.590 --> 00:57:28.130 align:middle line:84% might see two or three or four or five ads, and so I undercut. 00:57:28.130 --> 00:57:30.660 align:middle line:84% We get this mixed strategy pricing equilibrium. 00:57:30.660 --> 00:57:33.950 align:middle line:90% 00:57:33.950 --> 00:57:36.260 align:middle line:90% Advertising levels. 00:57:36.260 --> 00:57:41.310 align:middle line:84% A prime of 0 is small and A prime of 1 is really big. 00:57:41.310 --> 00:57:44.930 align:middle line:84% So if I make it even more stark, if this thing starts out 00:57:44.930 --> 00:57:47.660 align:middle line:84% like in advertising, the first people is really cheap, 00:57:47.660 --> 00:57:50.160 align:middle line:84% and then reaching the last people is really expensive, 00:57:50.160 --> 00:57:54.320 align:middle line:84% there's going to be some internal level x star 00:57:54.320 --> 00:57:56.360 align:middle line:90% of advertising that I pick. 00:57:56.360 --> 00:57:58.340 align:middle line:84% And I'm going to pick it x star so 00:57:58.340 --> 00:58:00.740 align:middle line:84% that the marginal cost of reaching one more consumer 00:58:00.740 --> 00:58:05.600 align:middle line:84% equals the marginal profit I get off the last consumer I reach. 00:58:05.600 --> 00:58:07.940 align:middle line:84% So the first order condition is going 00:58:07.940 --> 00:58:12.860 align:middle line:84% to be the profit on the marginal consumer reached 00:58:12.860 --> 00:58:16.080 align:middle line:84% equals the cost of the marginal consumer reached. 00:58:16.080 --> 00:58:18.700 align:middle line:90% OK? 00:58:18.700 --> 00:58:20.620 align:middle line:84% Interesting observation that Butters 00:58:20.620 --> 00:58:25.660 align:middle line:84% has is that in this model, advertising levels 00:58:25.660 --> 00:58:27.418 align:middle line:90% are socially efficient. 00:58:27.418 --> 00:58:29.710 align:middle line:84% You might have thought this is like entry where there's 00:58:29.710 --> 00:58:31.900 align:middle line:84% business stealing and because of business stealing, 00:58:31.900 --> 00:58:33.880 align:middle line:90% you could have too much entry. 00:58:33.880 --> 00:58:37.540 align:middle line:84% Or it could be because you don't internalize consumer surplus 00:58:37.540 --> 00:58:40.330 align:middle line:84% gains, there could be too little entry or entry being 00:58:40.330 --> 00:58:41.470 align:middle line:90% advertising. 00:58:41.470 --> 00:58:44.170 align:middle line:84% But in this model, with the technologies he's got, 00:58:44.170 --> 00:58:46.610 align:middle line:84% advertising is socially efficient. 00:58:46.610 --> 00:58:48.160 align:middle line:90% And we think of that like-- 00:58:48.160 --> 00:58:51.190 align:middle line:84% it's an argument that here's a special case where there's not 00:58:51.190 --> 00:58:53.350 align:middle line:84% a distortion, and the informative advertising 00:58:53.350 --> 00:58:56.050 align:middle line:84% in the world should exactly match what a social planner 00:58:56.050 --> 00:58:57.820 align:middle line:90% would want it to be. 00:58:57.820 --> 00:58:59.410 align:middle line:90% And why is that? 00:58:59.410 --> 00:59:04.150 align:middle line:84% So first order condition for x star 00:59:04.150 --> 00:59:06.515 align:middle line:84% gives the marginal cost of reaching the consumer 00:59:06.515 --> 00:59:08.140 align:middle line:84% equals the marginal benefit to the firm 00:59:08.140 --> 00:59:10.150 align:middle line:90% of reaching the consumer. 00:59:10.150 --> 00:59:14.290 align:middle line:84% Marginal cost of reaching the consumer is A prime of x star. 00:59:14.290 --> 00:59:17.500 align:middle line:90% What's the marginal benefit? 00:59:17.500 --> 00:59:20.510 align:middle line:84% The marginal benefit, I'm indifferent. 00:59:20.510 --> 00:59:21.980 align:middle line:90% I'm mixing over these prices. 00:59:21.980 --> 00:59:24.190 align:middle line:84% So I'm both choosing a pure strategy x star, 00:59:24.190 --> 00:59:26.260 align:middle line:90% and I'm mixing over prices. 00:59:26.260 --> 00:59:28.580 align:middle line:90% So it must be I'm indifferent. 00:59:28.580 --> 00:59:31.750 align:middle line:84% My profits are the same regardless of which price 00:59:31.750 --> 00:59:33.830 align:middle line:84% I choose from my price distribution. 00:59:33.830 --> 00:59:35.740 align:middle line:84% And so my profit on the marginal consumer 00:59:35.740 --> 00:59:37.840 align:middle line:84% is the same, regardless of which profit price 00:59:37.840 --> 00:59:39.730 align:middle line:84% I choose from the price distribution. 00:59:39.730 --> 00:59:43.180 align:middle line:84% So I'm going to evaluate it, looking at the case 00:59:43.180 --> 00:59:45.520 align:middle line:84% where I set a price at the upper bound of the price 00:59:45.520 --> 00:59:50.530 align:middle line:84% distribution v. So what's the marginal value of reaching 00:59:50.530 --> 00:59:53.140 align:middle line:84% one more consumer when I set a price of v? 00:59:53.140 --> 00:59:55.030 align:middle line:84% Well, I get v minus c is the profit 00:59:55.030 --> 00:59:56.920 align:middle line:90% from selling to that consumer. 00:59:56.920 --> 01:00:01.840 align:middle line:84% And then I get that multiplied by the probability 01:00:01.840 --> 01:00:05.830 align:middle line:84% that the consumer saw no other ad, because when I set price v, 01:00:05.830 --> 01:00:09.010 align:middle line:84% I sell to them only if they got no other ad, because if they got 01:00:09.010 --> 01:00:10.930 align:middle line:84% any other ad, they'll find a lower price, 01:00:10.930 --> 01:00:13.120 align:middle line:90% and they won't buy from me. 01:00:13.120 --> 01:00:18.800 align:middle line:84% But this is the social benefit of an ad, 01:00:18.800 --> 01:00:21.240 align:middle line:84% of reaching a consumer who's never been reached before. 01:00:21.240 --> 01:00:23.407 align:middle line:84% Because if a consumer has never been reached before, 01:00:23.407 --> 01:00:24.800 align:middle line:84% the social surplus from the firms 01:00:24.800 --> 01:00:30.950 align:middle line:84% interacting with that consumer goes from 0 to v minus c. 01:00:30.950 --> 01:00:37.610 align:middle line:84% So what we're getting is that the private return 01:00:37.610 --> 01:00:40.550 align:middle line:84% to reach one more consumer is the social return 01:00:40.550 --> 01:00:43.670 align:middle line:90% to reaching one more consumer. 01:00:43.670 --> 01:00:46.220 align:middle line:84% And it's not obvious if you think about any other price. 01:00:46.220 --> 01:00:48.290 align:middle line:84% When you set any other price, you 01:00:48.290 --> 01:00:50.060 align:middle line:84% are doing some business stealing, 01:00:50.060 --> 01:00:53.612 align:middle line:84% and you are stealing consumers away from other firms, 01:00:53.612 --> 01:00:55.070 align:middle line:84% and you are increasing the consumer 01:00:55.070 --> 01:00:56.670 align:middle line:90% surplus of the consumers. 01:00:56.670 --> 01:01:00.500 align:middle line:84% But somehow, this model is such that those two things do exactly 01:01:00.500 --> 01:01:01.790 align:middle line:90% offset. 01:01:01.790 --> 01:01:05.605 align:middle line:84% And so there's no social inefficiency here. 01:01:05.605 --> 01:01:08.430 align:middle line:90% 01:01:08.430 --> 01:01:11.940 align:middle line:84% So it's at least one case where we don't worry about advertising 01:01:11.940 --> 01:01:14.640 align:middle line:90% being distortionary. 01:01:14.640 --> 01:01:18.600 align:middle line:84% And obviously, this is an extreme. 01:01:18.600 --> 01:01:22.590 align:middle line:84% Butters has come up with a special case where this is true. 01:01:22.590 --> 01:01:28.680 align:middle line:84% If you had the firm selling imperfect substitutes rather 01:01:28.680 --> 01:01:32.970 align:middle line:84% than perfect substitutes, we wouldn't get this-- 01:01:32.970 --> 01:01:36.430 align:middle line:84% we could get this kind of mixed strategy equilibrium. 01:01:36.430 --> 01:01:38.640 align:middle line:84% We get the kind of pure strategy equilibrium pricing 01:01:38.640 --> 01:01:41.070 align:middle line:84% like we got with search, with differentiated goods. 01:01:41.070 --> 01:01:42.810 align:middle line:84% You'd have the business stealing effects, 01:01:42.810 --> 01:01:45.130 align:middle line:84% you'd have the consumer surplus effects. 01:01:45.130 --> 01:01:47.125 align:middle line:84% But I think, in the entry literature, 01:01:47.125 --> 01:01:48.750 align:middle line:84% we have this intuition that one of them 01:01:48.750 --> 01:01:50.740 align:middle line:84% was often going to be bigger than the others. 01:01:50.740 --> 01:01:52.930 align:middle line:84% Here, I think we don't have that intuition. 01:01:52.930 --> 01:01:56.515 align:middle line:90% 01:01:56.515 --> 01:01:57.015 align:middle line:90% OK. 01:01:57.015 --> 01:02:00.170 align:middle line:90% 01:02:00.170 --> 01:02:02.600 align:middle line:84% Second class of models I want to talk about then 01:02:02.600 --> 01:02:06.080 align:middle line:84% is, how do we think about signaling models of advertising, 01:02:06.080 --> 01:02:09.410 align:middle line:90% and when would they work? 01:02:09.410 --> 01:02:12.860 align:middle line:84% So Milgrom and Roberts is your classic paper doing this 01:02:12.860 --> 01:02:14.900 align:middle line:90% in the 1980s. 01:02:14.900 --> 01:02:19.767 align:middle line:84% Milgrom and Roberts' model, all firms 01:02:19.767 --> 01:02:21.350 align:middle line:84% that are entering the market, they try 01:02:21.350 --> 01:02:22.790 align:middle line:90% to come up with some product. 01:02:22.790 --> 01:02:25.540 align:middle line:90% They invent some quality q. 01:02:25.540 --> 01:02:27.540 align:middle line:84% We're just going to make it a two-point support. 01:02:27.540 --> 01:02:31.430 align:middle line:84% Your quality is either low or high. 01:02:31.430 --> 01:02:33.800 align:middle line:84% Obviously, all the firms are trying to invent quality q, 01:02:33.800 --> 01:02:35.830 align:middle line:84% but they sometimes come up with quality H 01:02:35.830 --> 01:02:37.205 align:middle line:84% or sometimes come up with quality 01:02:37.205 --> 01:02:44.115 align:middle line:84% L. There's some common prior that the consumers and the firm 01:02:44.115 --> 01:02:45.990 align:middle line:84% have over the probability that the quality is 01:02:45.990 --> 01:02:48.090 align:middle line:90% high of a new product. 01:02:48.090 --> 01:02:50.730 align:middle line:84% The firm observes its own quality 01:02:50.730 --> 01:02:54.390 align:middle line:84% and then chooses a price p for the product 01:02:54.390 --> 01:02:58.410 align:middle line:90% and a level of advertising A. 01:02:58.410 --> 01:03:00.930 align:middle line:84% Important in this model is consumers 01:03:00.930 --> 01:03:06.480 align:middle line:84% observe A. The consumers observe that they just put that sock 01:03:06.480 --> 01:03:07.630 align:middle line:90% puppet on the Super Bowl. 01:03:07.630 --> 01:03:09.720 align:middle line:90% That was $2.4 million. 01:03:09.720 --> 01:03:11.630 align:middle line:84% In some sense, this is what we tend 01:03:11.630 --> 01:03:14.040 align:middle line:84% to refer to as a money burning theory of advertising. 01:03:14.040 --> 01:03:17.970 align:middle line:84% It almost doesn't matter what they did with the money. 01:03:17.970 --> 01:03:21.600 align:middle line:84% It's like, yeah, those guys just put $2.4 million in dollar bills 01:03:21.600 --> 01:03:23.920 align:middle line:84% in a big pile and lit it on fire. 01:03:23.920 --> 01:03:27.450 align:middle line:84% That proves to me that they must have 01:03:27.450 --> 01:03:31.680 align:middle line:84% thought it was worth it for them to burn $2.4 million. 01:03:31.680 --> 01:03:33.690 align:middle line:84% So they must be the type of firm that wants 01:03:33.690 --> 01:03:36.720 align:middle line:90% to make big bonfires of money. 01:03:36.720 --> 01:03:43.500 align:middle line:84% So anyway, consumers observe A. They Bayesian update 01:03:43.500 --> 01:03:47.040 align:middle line:84% on what q is, knowing what different firms would 01:03:47.040 --> 01:03:50.010 align:middle line:84% do in terms of burning money or not. 01:03:50.010 --> 01:03:53.190 align:middle line:84% And so then they form a posterior on what q is, 01:03:53.190 --> 01:03:55.440 align:middle line:90% and they buy the product. 01:03:55.440 --> 01:03:57.023 align:middle line:84% And then the thought here is that this 01:03:57.023 --> 01:03:59.148 align:middle line:84% is like a dynamic model, although they don't really 01:03:59.148 --> 01:04:00.340 align:middle line:90% model the dynamics fully. 01:04:00.340 --> 01:04:02.670 align:middle line:84% But then after you buy the product, 01:04:02.670 --> 01:04:05.370 align:middle line:84% you learn something more about the quality, 01:04:05.370 --> 01:04:07.590 align:middle line:84% and then you may buy the product again, 01:04:07.590 --> 01:04:09.300 align:middle line:84% or you may buy it a third time, you 01:04:09.300 --> 01:04:10.950 align:middle line:84% may continue to buy it for many years, 01:04:10.950 --> 01:04:13.680 align:middle line:84% you may tell your friends about it, whatever. 01:04:13.680 --> 01:04:17.350 align:middle line:84% So you observe the A. You decide whether to buy. 01:04:17.350 --> 01:04:19.950 align:middle line:84% If you buy, you learn something about the quality. 01:04:19.950 --> 01:04:24.040 align:middle line:84% And then there's some repeat purchasing going on later on. 01:04:24.040 --> 01:04:25.330 align:middle line:90% OK. 01:04:25.330 --> 01:04:28.780 align:middle line:84% They say, whatever's going on in all those later stages, 01:04:28.780 --> 01:04:31.760 align:middle line:84% we're just going to model this in a reduced form way. 01:04:31.760 --> 01:04:34.360 align:middle line:84% We're going to say that the profits the firm gets 01:04:34.360 --> 01:04:38.770 align:middle line:84% are pi of p given q and q hat that 01:04:38.770 --> 01:04:43.120 align:middle line:84% depend both on the true quality q and on the quality q hat 01:04:43.120 --> 01:04:46.310 align:middle line:84% that consumers expect to receive when making purchase decisions. 01:04:46.310 --> 01:04:49.450 align:middle line:84% And generally, profit is going to be increasing 01:04:49.450 --> 01:04:51.760 align:middle line:90% in both of these arguments. 01:04:51.760 --> 01:04:53.823 align:middle line:84% If consumers think your product is good, 01:04:53.823 --> 01:04:55.240 align:middle line:84% even though it's not good, they're 01:04:55.240 --> 01:04:56.920 align:middle line:84% still going to buy it once, and you're 01:04:56.920 --> 01:04:59.170 align:middle line:84% going to make money from them thinking it's good. 01:04:59.170 --> 01:05:01.120 align:middle line:84% But then conditional on how good they 01:05:01.120 --> 01:05:03.320 align:middle line:84% think it is, what makes them buy it or not, 01:05:03.320 --> 01:05:05.327 align:middle line:84% if it's actually good, then when they buy it, 01:05:05.327 --> 01:05:06.410 align:middle line:90% they're going to enjoy it. 01:05:06.410 --> 01:05:07.730 align:middle line:90% They're going to buy it again. 01:05:07.730 --> 01:05:11.080 align:middle line:84% And so your profit's increasing in both the true quality q 01:05:11.080 --> 01:05:14.950 align:middle line:84% and in q hat, which is their posterior on q 01:05:14.950 --> 01:05:17.313 align:middle line:84% after you've had your advertising campaign 01:05:17.313 --> 01:05:18.105 align:middle line:90% and set your price. 01:05:18.105 --> 01:05:24.990 align:middle line:90% 01:05:24.990 --> 01:05:26.320 align:middle line:90% Observations. 01:05:26.320 --> 01:05:29.040 align:middle line:84% This model can have a separating equilibrium, 01:05:29.040 --> 01:05:33.570 align:middle line:84% where only firms with high quality advertise. 01:05:33.570 --> 01:05:36.240 align:middle line:84% So we get this equilibrium where you 01:05:36.240 --> 01:05:40.560 align:middle line:84% have these two things, a low price and zero advertising 01:05:40.560 --> 01:05:44.220 align:middle line:84% and a high price and some level of advertising. 01:05:44.220 --> 01:05:46.690 align:middle line:84% And all the low quality firms do this. 01:05:46.690 --> 01:05:48.650 align:middle line:84% All the high quality firms do this. 01:05:48.650 --> 01:05:53.790 align:middle line:90% 01:05:53.790 --> 01:05:56.790 align:middle line:90% When is it that this can happen? 01:05:56.790 --> 01:05:59.380 align:middle line:84% I think two things you're going to want to have. 01:05:59.380 --> 01:06:03.900 align:middle line:84% First is that repeat purchases have to be important. 01:06:03.900 --> 01:06:06.200 align:middle line:84% If there were no repeat purchases in the model, 01:06:06.200 --> 01:06:11.240 align:middle line:84% this wouldn't happen because if there were no repeat purchases, 01:06:11.240 --> 01:06:14.900 align:middle line:84% low and high quality firms would earn exactly the same profits. 01:06:14.900 --> 01:06:16.910 align:middle line:84% So if the high quality firms found 01:06:16.910 --> 01:06:19.430 align:middle line:84% it better to set a high price and burn money, 01:06:19.430 --> 01:06:22.580 align:middle line:84% the low quality firms would also set a high price, 01:06:22.580 --> 01:06:23.743 align:middle line:90% profit, and burn money. 01:06:23.743 --> 01:06:25.160 align:middle line:84% So what you have to have happen is 01:06:25.160 --> 01:06:26.577 align:middle line:84% there's got to be repeat purchases 01:06:26.577 --> 01:06:28.635 align:middle line:84% so that the high quality firms are 01:06:28.635 --> 01:06:30.260 align:middle line:84% willing to burn this money because they 01:06:30.260 --> 01:06:32.610 align:middle line:84% know the consumers will keep coming back. 01:06:32.610 --> 01:06:35.270 align:middle line:84% The low quality firms are only going to sell to consumers once. 01:06:35.270 --> 01:06:39.770 align:middle line:84% So you have to burn an amount of money such that the-- 01:06:39.770 --> 01:06:45.680 align:middle line:84% it's got to be that pH minus pL or pH 01:06:45.680 --> 01:06:54.740 align:middle line:84% minus c D of pH or D1 of pH minus pH minus c-- 01:06:54.740 --> 01:06:55.900 align:middle line:90% pL minus c. 01:06:55.900 --> 01:06:58.760 align:middle line:90% 01:06:58.760 --> 01:07:01.650 align:middle line:84% So imagine that this is the extra profits 01:07:01.650 --> 01:07:05.650 align:middle line:84% I get in the first period from burning the money. 01:07:05.650 --> 01:07:09.120 align:middle line:84% It's got to be that this is less than A 01:07:09.120 --> 01:07:12.558 align:middle line:84% because if this is less-- if this was bigger than A, then 01:07:12.558 --> 01:07:14.100 align:middle line:84% even the low quality firms would just 01:07:14.100 --> 01:07:15.600 align:middle line:84% go ahead and burn the money in order 01:07:15.600 --> 01:07:18.120 align:middle line:84% to get the extra demand in the first period. 01:07:18.120 --> 01:07:20.820 align:middle line:84% So it's got to be that this is less than A, 01:07:20.820 --> 01:07:23.100 align:middle line:84% but this plus the long-run benefit that you get 01:07:23.100 --> 01:07:26.640 align:middle line:84% from the repeat purchases at t equals 2 or t equals 3 01:07:26.640 --> 01:07:29.500 align:middle line:84% or whatever does make it worth advertising. 01:07:29.500 --> 01:07:33.370 align:middle line:84% And so we need a repeat purchasing effect. 01:07:33.370 --> 01:07:37.260 align:middle line:84% And then we also need some kind of small-- 01:07:37.260 --> 01:07:39.420 align:middle line:84% we need word of mouth learning effects 01:07:39.420 --> 01:07:41.947 align:middle line:84% to be small because if word of mouth learning effects 01:07:41.947 --> 01:07:43.530 align:middle line:84% were going to be small, you could just 01:07:43.530 --> 01:07:50.910 align:middle line:84% come in, set a price of pH, do zero advertising, 01:07:50.910 --> 01:07:52.800 align:middle line:84% have only a few people buy your product. 01:07:52.800 --> 01:07:54.300 align:middle line:84% Then those few people tell everybody 01:07:54.300 --> 01:07:56.300 align:middle line:84% and they tell everybody and they tell everybody, 01:07:56.300 --> 01:07:59.520 align:middle line:84% and the thing takes off, and the long-run benefits are there, 01:07:59.520 --> 01:08:01.770 align:middle line:84% even without having spent the money on the advertising 01:08:01.770 --> 01:08:03.100 align:middle line:90% in the first place. 01:08:03.100 --> 01:08:09.150 align:middle line:84% So what we need is something about repeat purchases help-- 01:08:09.150 --> 01:08:11.910 align:middle line:84% getting people to buy once, the repeat purchases 01:08:11.910 --> 01:08:14.110 align:middle line:84% make it worth spending the advertising. 01:08:14.110 --> 01:08:16.470 align:middle line:84% And you can't do it without spending the advertising 01:08:16.470 --> 01:08:19.000 align:middle line:84% because you can't do it by just a small release 01:08:19.000 --> 01:08:21.725 align:middle line:84% and then let people tell their friends instead. 01:08:21.725 --> 01:08:26.510 align:middle line:90% 01:08:26.510 --> 01:08:29.630 align:middle line:84% Technologically, this advertising is money burning. 01:08:29.630 --> 01:08:32.630 align:middle line:84% It's not changing the utility functions at all, 01:08:32.630 --> 01:08:36.979 align:middle line:84% but it can increase-- it can be socially valuable if it 01:08:36.979 --> 01:08:38.790 align:middle line:90% increases the match quality. 01:08:38.790 --> 01:08:41.600 align:middle line:84% So if you imagine a model where you have-- 01:08:41.600 --> 01:08:43.370 align:middle line:84% there are some people who only really 01:08:43.370 --> 01:08:45.319 align:middle line:90% have value for a high quality-- 01:08:45.319 --> 01:08:46.979 align:middle line:84% some people who are serious hikers. 01:08:46.979 --> 01:08:49.340 align:middle line:84% And so if it's not a really good Patagonia coat that's 01:08:49.340 --> 01:08:51.260 align:middle line:84% actually going to keep you warm on a mountain, 01:08:51.260 --> 01:08:52.196 align:middle line:90% you don't want it. 01:08:52.196 --> 01:08:54.529 align:middle line:84% And then there's some low quality consumers who are just 01:08:54.529 --> 01:08:55.830 align:middle line:90% wearing it around town. 01:08:55.830 --> 01:08:58.502 align:middle line:84% And if it doesn't actually prevent the wind 01:08:58.502 --> 01:08:59.960 align:middle line:84% from getting you, it doesn't really 01:08:59.960 --> 01:09:02.210 align:middle line:84% matter because you're just wearing it around the city. 01:09:02.210 --> 01:09:05.870 align:middle line:84% If the advertising helps us with the match quality 01:09:05.870 --> 01:09:08.930 align:middle line:84% and lets the people who want high quality buy 01:09:08.930 --> 01:09:11.750 align:middle line:84% high quality, people who want low quality, buy low quality, 01:09:11.750 --> 01:09:14.810 align:middle line:84% then this sort of money burning can raise utility. 01:09:14.810 --> 01:09:17.660 align:middle line:84% And we might not want to ban it, even though it's not 01:09:17.660 --> 01:09:19.800 align:middle line:90% technologically valuable. 01:09:19.800 --> 01:09:24.060 align:middle line:84% But it can be the only way to signal information here. 01:09:24.060 --> 01:09:27.920 align:middle line:90% 01:09:27.920 --> 01:09:32.460 align:middle line:84% And in some models like this, you 01:09:32.460 --> 01:09:36.609 align:middle line:84% can actually even do signaling without any advertising at all. 01:09:36.609 --> 01:09:39.330 align:middle line:84% So it's possible that there can be-- 01:09:39.330 --> 01:09:40.529 align:middle line:90% don't go in advertising. 01:09:40.529 --> 01:09:44.010 align:middle line:84% There can be equilibria like this, where some firms set 01:09:44.010 --> 01:09:47.340 align:middle line:84% a really high price, and then people 01:09:47.340 --> 01:09:50.550 align:middle line:84% know that they're only going to sell 01:09:50.550 --> 01:09:51.760 align:middle line:90% to a small number of people. 01:09:51.760 --> 01:09:53.040 align:middle line:84% So it must be they're expecting to sell 01:09:53.040 --> 01:09:55.613 align:middle line:84% to that small number of people over and over and over again. 01:09:55.613 --> 01:09:57.280 align:middle line:84% Whereas you can set a low price and sell 01:09:57.280 --> 01:10:01.210 align:middle line:84% to a lot of people who just buy once or something like that. 01:10:01.210 --> 01:10:05.580 align:middle line:84% So you can get you can get price as a signal. 01:10:05.580 --> 01:10:08.170 align:middle line:84% You can get price as a signal even without advertising. 01:10:08.170 --> 01:10:11.820 align:middle line:84% But advertising makes it easier to get signaling models 01:10:11.820 --> 01:10:14.370 align:middle line:90% to work out. 01:10:14.370 --> 01:10:14.870 align:middle line:90% OK. 01:10:14.870 --> 01:10:20.270 align:middle line:90% 01:10:20.270 --> 01:10:20.870 align:middle line:90% OK. 01:10:20.870 --> 01:10:21.680 align:middle line:90% Questions? 01:10:21.680 --> 01:10:23.960 align:middle line:84% So then I thought the final one, I 01:10:23.960 --> 01:10:27.440 align:middle line:84% will go to things based on my beer ad. 01:10:27.440 --> 01:10:31.220 align:middle line:84% And here, I'm covering an unpublished paper that's 01:10:31.220 --> 01:10:34.110 align:middle line:84% unpublished for 15 years or whatever. 01:10:34.110 --> 01:10:37.010 align:middle line:84% I'm sure Jesse has long since given up on this paper, 01:10:37.010 --> 01:10:39.560 align:middle line:84% but I always really liked it, so I'm going 01:10:39.560 --> 01:10:41.010 align:middle line:90% to keep covering it anyway. 01:10:41.010 --> 01:10:44.960 align:middle line:84% So this is a behavioral economics paper 01:10:44.960 --> 01:10:47.940 align:middle line:84% about advertising changing preferences. 01:10:47.940 --> 01:10:49.850 align:middle line:84% And it starts out with noting, there's 01:10:49.850 --> 01:10:55.010 align:middle line:84% a lot of very, very good evidence about fallibility 01:10:55.010 --> 01:10:57.890 align:middle line:84% of memory and manipulability of memory 01:10:57.890 --> 01:11:00.210 align:middle line:84% that makes you think that these things can work. 01:11:00.210 --> 01:11:04.370 align:middle line:84% So, for instance, one of these very classic experiments, 01:11:04.370 --> 01:11:09.590 align:middle line:84% the famous person experiment is, I give you a list of people. 01:11:09.590 --> 01:11:14.330 align:middle line:84% And I write down, I give you a list of kind of famous people, 01:11:14.330 --> 01:11:17.880 align:middle line:84% and I ask you, which of these people are famous? 01:11:17.880 --> 01:11:21.510 align:middle line:84% And I put in a bunch of names on there like, Ban Ki-moon 01:11:21.510 --> 01:11:23.500 align:middle line:84% and Sebastian Weisdorf and whatever. 01:11:23.500 --> 01:11:26.485 align:middle line:84% And these people are people who you've kind of heard of. 01:11:26.485 --> 01:11:28.110 align:middle line:84% And you're like, OK, yeah, Ban Ki-moon, 01:11:28.110 --> 01:11:31.710 align:middle line:84% that's some guy who did something UN 01:11:31.710 --> 01:11:32.715 align:middle line:90% official or something. 01:11:32.715 --> 01:11:34.090 align:middle line:84% I don't remember quite who he is. 01:11:34.090 --> 01:11:37.240 align:middle line:84% Sebastian Weisdorf, that just doesn't ring a bell. 01:11:37.240 --> 01:11:38.430 align:middle line:90% I'm going to put not famous. 01:11:38.430 --> 01:11:41.220 align:middle line:84% And you put names of actresses and whatever 01:11:41.220 --> 01:11:44.370 align:middle line:84% that, again, you know, I have no idea what that person's ever 01:11:44.370 --> 01:11:47.430 align:middle line:84% been in, but I know that's the name of some actress 01:11:47.430 --> 01:11:48.257 align:middle line:90% and whatever. 01:11:48.257 --> 01:11:50.340 align:middle line:84% And people are pretty good at this task of saying, 01:11:50.340 --> 01:11:51.155 align:middle line:90% is someone famous? 01:11:51.155 --> 01:11:52.530 align:middle line:84% And one of the interesting things 01:11:52.530 --> 01:11:55.210 align:middle line:84% is you give people that task, you give them, here's 20 people, 01:11:55.210 --> 01:11:56.980 align:middle line:90% is this person famous or not? 01:11:56.980 --> 01:11:58.740 align:middle line:84% And some of them are just made up names. 01:11:58.740 --> 01:12:00.430 align:middle line:84% And people are very, very good at that task. 01:12:00.430 --> 01:12:02.730 align:middle line:84% They can write down famous, famous, not famous, famous, 01:12:02.730 --> 01:12:04.605 align:middle line:84% not famous, even though they have no idea who 01:12:04.605 --> 01:12:06.630 align:middle line:90% the people actually are. 01:12:06.630 --> 01:12:08.970 align:middle line:84% And then you bring the same students back in the list-- 01:12:08.970 --> 01:12:10.240 align:middle line:90% into the lab. 01:12:10.240 --> 01:12:13.680 align:middle line:84% One week later, you hand them exactly the same list 01:12:13.680 --> 01:12:15.600 align:middle line:90% that they said yes or no on. 01:12:15.600 --> 01:12:18.180 align:middle line:84% And they're much worse at the task. 01:12:18.180 --> 01:12:20.820 align:middle line:84% And it's because, OK, Sebastian Weisdorf, 01:12:20.820 --> 01:12:22.500 align:middle line:90% I've kind of heard that. 01:12:22.500 --> 01:12:24.120 align:middle line:84% But is that because Sebastian Weisdorf 01:12:24.120 --> 01:12:25.578 align:middle line:84% is some guy I've heard of, or is it 01:12:25.578 --> 01:12:28.960 align:middle line:84% because he was a guy who was on the list that I did last week? 01:12:28.960 --> 01:12:31.680 align:middle line:84% And people no longer can remember the things 01:12:31.680 --> 01:12:34.350 align:middle line:84% as well as they did the week before or whatever. 01:12:34.350 --> 01:12:37.350 align:middle line:84% He also gives these examples of these marketing 01:12:37.350 --> 01:12:39.808 align:middle line:84% experiments with orange juice, where you 01:12:39.808 --> 01:12:41.100 align:middle line:90% have people taste orange juice. 01:12:41.100 --> 01:12:43.500 align:middle line:84% And you prime them before with the correct answers, 01:12:43.500 --> 01:12:49.650 align:middle line:84% like here's an orange juice that's got a really tangy taste 01:12:49.650 --> 01:12:50.415 align:middle line:90% or whatever. 01:12:50.415 --> 01:12:53.040 align:middle line:84% And then here's an orange juice that's got a really sweet taste 01:12:53.040 --> 01:12:53.520 align:middle line:90% or whatever. 01:12:53.520 --> 01:12:54.437 align:middle line:90% Which one do you like? 01:12:54.437 --> 01:12:58.920 align:middle line:84% And then the tangy taste one is just like, 01:12:58.920 --> 01:13:00.420 align:middle line:84% you put vinegar in the orange juice, 01:13:00.420 --> 01:13:01.808 align:middle line:90% and you just made it worse. 01:13:01.808 --> 01:13:03.600 align:middle line:84% But if you've described it like, oh, here's 01:13:03.600 --> 01:13:04.950 align:middle line:84% a really sour orange juice taste, 01:13:04.950 --> 01:13:06.570 align:middle line:84% or here's a really tangy one taste, somehow, 01:13:06.570 --> 01:13:08.130 align:middle line:84% if you've told them it's a really tangy one, 01:13:08.130 --> 01:13:10.140 align:middle line:84% they're like, oh, yeah, I like that one, that's tangy, 01:13:10.140 --> 01:13:11.260 align:middle line:90% or something like that. 01:13:11.260 --> 01:13:15.390 align:middle line:84% And so his thought on the beer example is, 01:13:15.390 --> 01:13:18.150 align:middle line:84% people have had beer many times in their lives, 01:13:18.150 --> 01:13:20.980 align:middle line:84% and they've had parties with their friends many times. 01:13:20.980 --> 01:13:24.720 align:middle line:84% And the thing is, you like to go to events where 01:13:24.720 --> 01:13:26.400 align:middle line:90% you're having a good time. 01:13:26.400 --> 01:13:28.980 align:middle line:84% And you don't remember whether it was me 01:13:28.980 --> 01:13:30.740 align:middle line:84% who had a good time drinking the Budweiser 01:13:30.740 --> 01:13:34.740 align:middle line:84% or if it was those people I saw on TV having a good time 01:13:34.740 --> 01:13:36.010 align:middle line:90% drinking the Budweiser. 01:13:36.010 --> 01:13:37.290 align:middle line:84% And so then when I'm there in the store, 01:13:37.290 --> 01:13:38.707 align:middle line:84% I'm trying to remember, oh, should 01:13:38.707 --> 01:13:39.858 align:middle line:90% I buy the Budweiser or not? 01:13:39.858 --> 01:13:42.150 align:middle line:84% And I'm just trying to remember some previous time when 01:13:42.150 --> 01:13:43.720 align:middle line:84% I had a party and I served Budweiser, 01:13:43.720 --> 01:13:45.250 align:middle line:90% did people have a lot of fun? 01:13:45.250 --> 01:13:47.140 align:middle line:84% And I can't keep track of whether it 01:13:47.140 --> 01:13:49.810 align:middle line:84% was my party with the Budweiser where people had the fun 01:13:49.810 --> 01:13:52.840 align:middle line:84% or it was that one on TV I saw with people 01:13:52.840 --> 01:13:54.980 align:middle line:84% having Budweiser, where everyone had a great time. 01:13:54.980 --> 01:13:57.760 align:middle line:84% And in some sense, they've changed my memory 01:13:57.760 --> 01:13:59.593 align:middle line:84% and put all these memories in there 01:13:59.593 --> 01:14:01.510 align:middle line:84% that I can't distinguish from my own memories. 01:14:01.510 --> 01:14:03.370 align:middle line:90% And therefore I can't-- 01:14:03.370 --> 01:14:05.860 align:middle line:84% and so you're changing my preferences 01:14:05.860 --> 01:14:10.330 align:middle line:84% by changing what I remember about my prior experiences 01:14:10.330 --> 01:14:12.820 align:middle line:84% with the product, confusing my experiences 01:14:12.820 --> 01:14:15.640 align:middle line:84% with the people on TV's experiences with the product. 01:14:15.640 --> 01:14:19.990 align:middle line:84% So the way he does this is he has this continuum of consumers 01:14:19.990 --> 01:14:22.030 align:middle line:90% who have types theta i. 01:14:22.030 --> 01:14:26.140 align:middle line:84% And the type theta consumers get benefit v with probability theta 01:14:26.140 --> 01:14:27.620 align:middle line:90% when they consume the good. 01:14:27.620 --> 01:14:30.100 align:middle line:84% So could this just be-- depending on your friend group, 01:14:30.100 --> 01:14:31.600 align:middle line:84% you could be in a friend group where 01:14:31.600 --> 01:14:35.260 align:middle line:84% serving Budweiser at your party, 100% chance the party goes well. 01:14:35.260 --> 01:14:37.510 align:middle line:84% You could be in a friend group where serving Budweiser 01:14:37.510 --> 01:14:40.210 align:middle line:84% at your party, just 0% chance the party goes well 01:14:40.210 --> 01:14:42.940 align:middle line:84% because your friends all hate Budweiser. 01:14:42.940 --> 01:14:43.690 align:middle line:90% Anyway. 01:14:43.690 --> 01:14:47.290 align:middle line:84% So each consumer is different in their experiences 01:14:47.290 --> 01:14:50.020 align:middle line:90% when they've consumed the good. 01:14:50.020 --> 01:14:54.700 align:middle line:84% Consumers have consumed the good on N prior occasions 01:14:54.700 --> 01:14:56.290 align:middle line:90% and gotten these draws. 01:14:56.290 --> 01:14:59.120 align:middle line:90% But they don't know theta. 01:14:59.120 --> 01:15:01.750 align:middle line:84% I can't go, every time in my life 01:15:01.750 --> 01:15:04.610 align:middle line:84% that I've bought Budweiser for a party, how did that party go? 01:15:04.610 --> 01:15:06.830 align:middle line:84% Let me count them all up and add them all up. 01:15:06.830 --> 01:15:09.820 align:middle line:84% What I do is I just remember often just one. 01:15:09.820 --> 01:15:11.270 align:middle line:90% I just call to mind. 01:15:11.270 --> 01:15:14.090 align:middle line:84% Let me try to remember some time when I served Budweiser. 01:15:14.090 --> 01:15:15.830 align:middle line:90% I remember that experience. 01:15:15.830 --> 01:15:17.320 align:middle line:90% I remember how it went. 01:15:17.320 --> 01:15:20.770 align:middle line:90% And then I Bayesian update. 01:15:20.770 --> 01:15:24.910 align:middle line:84% So I have some-- there's some common prior over the thetas. 01:15:24.910 --> 01:15:26.530 align:middle line:84% And then I Bayesian update to form, 01:15:26.530 --> 01:15:30.820 align:middle line:84% what's my expected view of my preference for the product, 01:15:30.820 --> 01:15:34.630 align:middle line:84% conditional on the events that I'm remembering? 01:15:34.630 --> 01:15:37.510 align:middle line:84% And then what Jesse does, he models advertising 01:15:37.510 --> 01:15:42.700 align:middle line:84% as potentially altering memory in one of two ways. 01:15:42.700 --> 01:15:47.560 align:middle line:84% One is an ad could just take a memory out of your head 01:15:47.560 --> 01:15:50.110 align:middle line:84% and stick a fake memory in instead. 01:15:50.110 --> 01:15:52.420 align:middle line:84% So it's just taking one of your N observations 01:15:52.420 --> 01:15:54.390 align:middle line:84% out, putting another observation in 01:15:54.390 --> 01:15:56.140 align:middle line:84% so that when you do the Bayesian updating, 01:15:56.140 --> 01:15:59.830 align:middle line:90% you might recall that memory. 01:15:59.830 --> 01:16:04.550 align:middle line:84% The other thing is a notion of priming memories. 01:16:04.550 --> 01:16:08.770 align:middle line:84% That is, we all have all these memories in our heads. 01:16:08.770 --> 01:16:12.430 align:middle line:84% If I get you to make one memory, call that memory 01:16:12.430 --> 01:16:14.305 align:middle line:84% to the front of your mind, then the next time 01:16:14.305 --> 01:16:16.680 align:middle line:84% you're trying to call a memory to the front of your mind, 01:16:16.680 --> 01:16:18.190 align:middle line:84% that one will come forward again. 01:16:18.190 --> 01:16:21.310 align:middle line:84% It's been more recent and more primed to come forward. 01:16:21.310 --> 01:16:25.660 align:middle line:84% And so here, this is, you show people a positive experience 01:16:25.660 --> 01:16:27.010 align:middle line:90% while they're watching the ad. 01:16:27.010 --> 01:16:29.920 align:middle line:84% They remember the most similar experience that they've had. 01:16:29.920 --> 01:16:31.420 align:middle line:84% And then because that experience has 01:16:31.420 --> 01:16:34.490 align:middle line:84% been brought through to the front of their mind, 01:16:34.490 --> 01:16:37.500 align:middle line:84% it's more likely to be drawn the next time it comes out. 01:16:37.500 --> 01:16:40.640 align:middle line:84% So this is like you're manipulating memories 01:16:40.640 --> 01:16:44.397 align:middle line:84% by making people's memories more-- 01:16:44.397 --> 01:16:46.730 align:middle line:84% changing the probability of recall of different memories 01:16:46.730 --> 01:16:48.500 align:middle line:90% that they have. 01:16:48.500 --> 01:16:51.890 align:middle line:84% And then the paper itself, it mostly focuses on this. 01:16:51.890 --> 01:16:57.590 align:middle line:84% It gets into this highly rational case where consumers 01:16:57.590 --> 01:17:00.620 align:middle line:84% are aware that their memories are being manipulated, 01:17:00.620 --> 01:17:02.840 align:middle line:84% and they are aware of the equilibrium 01:17:02.840 --> 01:17:06.510 align:middle line:84% level of memory manipulation that's going on by the firms. 01:17:06.510 --> 01:17:09.000 align:middle line:90% And then they try to actually-- 01:17:09.000 --> 01:17:11.412 align:middle line:90% 01:17:11.412 --> 01:17:12.870 align:middle line:84% if I were him, I might have started 01:17:12.870 --> 01:17:15.515 align:middle line:84% with more of the non-fully rational thing, where 01:17:15.515 --> 01:17:16.890 align:middle line:84% people are just being manipulated 01:17:16.890 --> 01:17:17.765 align:middle line:90% and don't realize it. 01:17:17.765 --> 01:17:19.710 align:middle line:84% But he's mostly doing the, they are 01:17:19.710 --> 01:17:22.570 align:middle line:84% trying to offset the manipulation of their memories. 01:17:22.570 --> 01:17:25.060 align:middle line:84% And so, in some sense, they downweight it. 01:17:25.060 --> 01:17:27.710 align:middle line:84% They're like, I know this firm in equilibrium is advertising. 01:17:27.710 --> 01:17:30.210 align:middle line:84% Therefore, I know that this memory might be mine or might be 01:17:30.210 --> 01:17:31.350 align:middle line:90% someone else's. 01:17:31.350 --> 01:17:33.690 align:middle line:84% Therefore, when I'm doing the Bayesian updating, 01:17:33.690 --> 01:17:36.750 align:middle line:84% I'm doing the Bayesian updating of expectation of theta 01:17:36.750 --> 01:17:39.318 align:middle line:84% given ri and given A star, where I 01:17:39.318 --> 01:17:40.860 align:middle line:84% know what the equilibrium advertising 01:17:40.860 --> 01:17:42.190 align:middle line:90% level is in the world. 01:17:42.190 --> 01:17:46.230 align:middle line:84% And so I know, and I'm trying to undo this. 01:17:46.230 --> 01:17:49.980 align:middle line:84% But nonetheless, the point is, even in this model, 01:17:49.980 --> 01:17:52.860 align:middle line:84% where memories are being artificially manipulated, 01:17:52.860 --> 01:17:56.010 align:middle line:84% firms might still want to manipulate the memories. 01:17:56.010 --> 01:18:00.720 align:middle line:84% And this is kind of like for information design reasons. 01:18:00.720 --> 01:18:04.360 align:middle line:84% So you're not changing the mean-- 01:18:04.360 --> 01:18:07.360 align:middle line:84% you're not changing the mean level of people's posterior 01:18:07.360 --> 01:18:11.710 align:middle line:84% on theta, but you are changing the distributions on the theta. 01:18:11.710 --> 01:18:13.870 align:middle line:84% And so if you can change the distribution of thetas 01:18:13.870 --> 01:18:17.200 align:middle line:84% in a way that lets you extract more of the social surplus, 01:18:17.200 --> 01:18:21.070 align:middle line:84% it may make sense to advertise even if the advertising is 01:18:21.070 --> 01:18:22.190 align:middle line:90% being undone. 01:18:22.190 --> 01:18:24.040 align:middle line:84% So if the advertising is not being undone, 01:18:24.040 --> 01:18:25.820 align:middle line:84% it's clear you want to advertise. 01:18:25.820 --> 01:18:27.490 align:middle line:84% But even when it's being undone, you 01:18:27.490 --> 01:18:30.610 align:middle line:84% might want to do it for information design 01:18:30.610 --> 01:18:32.350 align:middle line:84% reasons, where if I can get everyone 01:18:32.350 --> 01:18:36.040 align:middle line:84% to have the same posterior, then I can charge that posterior 01:18:36.040 --> 01:18:37.940 align:middle line:84% and extract all the social surplus. 01:18:37.940 --> 01:18:39.640 align:middle line:84% Or if there's more heterogeneity, 01:18:39.640 --> 01:18:42.010 align:middle line:84% I only extract a fraction of the social surplus. 01:18:42.010 --> 01:18:44.530 align:middle line:90% 01:18:44.530 --> 01:18:45.140 align:middle line:90% Anyway. 01:18:45.140 --> 01:18:47.780 align:middle line:84% So it's just got some interesting observations. 01:18:47.780 --> 01:18:49.360 align:middle line:84% And it's also interesting about-- 01:18:49.360 --> 01:18:51.760 align:middle line:84% you can think about when advertising 01:18:51.760 --> 01:18:53.530 align:middle line:84% is going to be more and less likely 01:18:53.530 --> 01:18:58.143 align:middle line:84% because it's going to depend on product quality 01:18:58.143 --> 01:19:00.310 align:middle line:84% because if you actually have a high-quality product, 01:19:00.310 --> 01:19:02.602 align:middle line:84% you don't need-- you don't get as much marginal benefit 01:19:02.602 --> 01:19:04.630 align:middle line:84% from replacing memories in people's minds 01:19:04.630 --> 01:19:07.190 align:middle line:84% with fake memories because the true memories are good. 01:19:07.190 --> 01:19:09.610 align:middle line:84% And so you can get this sort of advertising 01:19:09.610 --> 01:19:12.690 align:middle line:84% as a signal of bad quality and various interesting things. 01:19:12.690 --> 01:19:16.460 align:middle line:90% 01:19:16.460 --> 01:19:17.180 align:middle line:90% OK. 01:19:17.180 --> 01:19:18.600 align:middle line:90% Anyway, I am out of time. 01:19:18.600 --> 01:19:22.070 align:middle line:84% So anyway, Monday, I'm going to-- again, 01:19:22.070 --> 01:19:23.820 align:middle line:90% I'm going to do auctions. 01:19:23.820 --> 01:19:25.650 align:middle line:84% It's going to be mostly auction theory. 01:19:25.650 --> 01:19:28.080 align:middle line:84% I am going to do one empirical paper that I really like. 01:19:28.080 --> 01:19:33.260 align:middle line:84% This is the Hendricks and Porter paper about offshore oil 01:19:33.260 --> 01:19:35.180 align:middle line:90% drilling auctions. 01:19:35.180 --> 01:19:37.160 align:middle line:84% And then, yeah, that's the plan for next week. 01:19:37.160 --> 01:19:38.580 align:middle line:90% I'll do auctions. 01:19:38.580 --> 01:19:41.480 align:middle line:84% Tobias comes in Wednesday and does 01:19:41.480 --> 01:19:43.040 align:middle line:90% structural empirical auctions. 01:19:43.040 --> 01:19:46.330 align:middle line:84% Then we go back to advertising empirics the week after. 01:19:46.330 --> 01:19:58.000 align:middle line:90%