 Okay, welcome. Welcome everyone to economics of platforms online seminar. Today we have Seth Benzell presenting how to govern Facebook. He's joined work with Avinash Collies and Alex White discussing. Seth will have 40 minutes for presentation and Alex will have five minutes for discussion then we will have general Q&A and discussion and I will stop recording after an hour but I will stay here for as long as there is a need and a discussion. So with that Seth, the screen is yours. Take it away. Hey everyone, thank you so so much for having me here at Telys. I have to say that this paper has some pretty deep Telysa ties. First off, obviously whenever you're writing about multi-sided platforms you're going to be heavily indebted to Professor Ty Roll. But more specifically I can say that the Telys digital economy workshop was one of the first digital economy really the first digital economy conference I ever attended really got me excited about the field and my co-author Avinash Collies who I think is on the line did his master's here at TSE. So there's some pretty deep TSE roots here in this paper and which is titled how to govern Facebook a structural model for taxing and regulating big tech. So I'm going to launch right into it. Again this is joint work with Avinash Collies. Let's take it away. So digital platform businesses create value by enabling connections. So we can think about lots of different platforms. You know I'm going to focus here on Facebook because they're relatively easier to collect data on. They are very clearly making their value added in enabling connections between pairs of paid people. But if you think about multi-sided platform like Uber where you're connecting riders to drivers or eBay where you're connecting auctioneers to buyers and the fundamental thing going on in these digital platforms is the enabling of a connection which then the platform is going to sort of monetize in some way. These digital platforms are able to scale really rapidly because they have pretty low marginal costs and pretty fixed capital needs right. If Airbnb wants to roll out for a new city they don't have to buy or build a new hotel in the way that foreseasons if they enter a new city needs to. Rather they flip a switch on a server and people show up with all of the fixed capital that's needed in order to provide their share of the service. And then finally these companies all collect lots of data. These features have made digital platforms a really fundamental part of our COVID-19 world. I know that it's really kind of hard to think about trying to go down the slot through this lockdown without our social media and our digital communication tools. But all of this power that these digital economy platforms have also create large supply and demand economies of scale and therefore raise monopoly concerns right. So on the supply side we've got this economy of scale that just comes from being an information good right which is you write the code once and then you can sell it a million times. But then these platform goods also have these demand side economies of scale which is as more people use the platform the quality of the platform itself improves at least in general especially because most of these platforms have tools to make sure that you're only connecting to people that you have a positive network value from. In general you can imagine a negative network effect from some people. Okay so that's kind of the big concern is that these platforms have lots of power and if anything COVID-19 has made them more powerful but what should we do about that? Well there are a bunch of ideas that we'd like to try to evaluate and put side by side. So for example we know in France you've got the digital services tax that you've been talking about which would be incident on these giant companies and for a company like Facebook this tax is going to be basically incident on its advertisements because advertisements are 95 percent of its revenue. Another and that's so that's an idea for trying to take some of the profit and sort of redistribute it towards the users you know French users say I'm giving my eyeballs towards Facebook and Amazon and Google I want some share of the monopoly profits that's being created by that. A sort of more direct way of attacking the same issue is this idea of data as labor which maybe Alex will have some insights on. So there the thinking is kind of promote collective bargaining on behalf of the people who are using the platforms and then they can go to the platform and say hey as a group we're only going to use the platform if you give us x percentage of the revenue or you can imagine the negotiating being more complex than that. Another approach is to try to promote competition on these digital platforms when the idea that that'll lower that'll sort of raise quality and drive out these monopoly profits and Fiona Scott Morton has a lot of ideas about how to do that particularly this idea of encouraging interoperability so people could enter with a platform and then use essential data that was originally created on another platform such as I could enter with a social network and then just use Facebook social graph that would promote entry and then finally people have talked sort of more broadly about just sort of breaking up these companies and those breakups could either be horizontal so you could end up with two sort of baby Facebooks that do the same thing or if they could be vertical right you could split up Facebook from Instagram for example and maybe those things those two platforms are more compliments than substitutes although obviously that's an open question so let me give you sort of a very very high level result preview first we're going to find that Facebook generates seven times as much social surplus as it creates from ad revenues in the US right so it's the important place to start is when we measure it we're finding a lot of social surplus from Facebook in particular and again we're taking a more general model we're going to calibrate for Facebook we find that Facebook's market power lowers welfare by nine percentage points versus the first best which is a nationalization of Facebook and then a subsidy for use and then four percent versus perfect competition but those those numbers are going to vary depending on exactly what counts in social welfare kind of most obviously in our calculations we have no negative externalities from Facebook or social media use although that's a big sort of social concern right now is is there something about the social media that's poisoning our politics or etc etc but also it's going to be important for thinking about Facebook's motivations because we're going to find that Facebook wants to have a larger user base that's implied by pure profit maximization and so we want to try to understand why does Facebook want to be big because that's going to inform our decision about do we want to reward that behavior of being big taxes are mostly incident on Facebook and property to properly targeted taxes can actually rise consumer surplus so the intuition there is if Facebook both has the goals of making ads through revenues and maintaining a large user base you can actually get them to substitute into their large user base motivation by taxing advertising and so when they shift into that that's better for society because we think that there's an inframarginal network effect that Facebook doesn't fully internalize that really we would want more people to be on Facebook because of these positive network effects that again Facebook doesn't 100% internalize and Alex can tell us all about the expense distortion and then finally a botch break up would be disastrous again this is very much in what Alex and Wilde talked about right which is letting the right one win sort of the worst case scenario is you break up one of these big platforms but don't any get any gains in competition whereas data and labor potentially even better than what I call the first best above because then you could you know show people advertisements and sort of arbitrage that because people get less disutility from viewing advertisements than they create in revenues from watching those advertisements so somehow data as labor could preserve that arbitrage there's a surplus to be made okay briefly literature review I want to just highlight this series of papers here obviously I'm indebted to professor Ty roll as well as professors Parker and Van Alstine and then another set of papers I'm indebted to is while and while in white and before them comes Ross 1974 which is this really interesting paper about Bell Tele in the Bell Journal of Economics about telephone networks but a lot of those ideas really only became manifest today with today's social networks in today's digital platforms okay so what does our model look like so what I'm going to do is I'm going to lay out my model I'm going to tell you some sort of theoretical results for it then we're going to calibrate from Facebook and then I'm going to be able to in this structural model do some counterfactual experiments of how will welfare and profits change under different universes so what's in our paper what's in our model so we've got a platform firm which is a monopolist and it's going to maximize profits but maybe also cares about having a shadow value for a large user base and in order to maximize their objective function which is the sum of their profits and wanting to have a large user base they're going to set prices subsidies or advertising levels and then obviously then we could adapt this to different settings with different sort of pricing schemes consumers decide whether or not to use the platform and importantly you're because this is a network good your desire to use the platform is going to be a function of others usage as well as how heavily the platform is being monetized so when I say monetized I could either mean you know how much I'm being charged in terms of a monthly subscription or how many ads am I being required to watch when I use it and now this sort of network dependence really important you know because if something happens to the platform and I leave the platform that's going to have reverberations on everyone on the platform who likes me so like let's say that Alex likes being connected to me on Twitter if I stop using Twitter then that's going to have a negative effect on Alex's use of Twitter and then if he leaves there's some potentially cascading effect of everyone he likes leaving we're going to say that consumers are potentially heterogeneous in what they value and that's what makes this a multi-sided platform instead of what's typically thought of instead of like typical two or one-sided networks right so when I say they're heterogeneous what they value for example some some demographic groups might have a larger opportunity cost for using Facebook so you might have just something better to do than using whatever platform we're talking about you have a better opportunity or alternatively there can be differences in network effects right so I'm going to show you some interesting heterogeneity that we find in what sort of connections are value on Facebook across different demographic groups and then finally we can allow for the different groups to have different disutility from view from monetization in this case viewing ads okay so this is what the platform's problem looks like we're going to maximize our profit which is fi times pi minus f so fi is going to be the net revenue and you can also think that this includes the shadow value in there is the net benefit of having person i use your platform and then pi is the probability that i use as your platform and then from the perspective of the platform your pi so the probability that any individual uses your platform is going to be a function of this vector of monetizations that you do on every right and so you can imagine for every individual the platform is choosing the monetization strategy that gets them the most revenue for the least disutility right and you're going to choose a vector of these points for every individual and that's going to lead to a certain equilibrium consumers on the other hand are going to participate if their utility from participating is higher than their opportunity cost which is going to be have follow a distribution in our calibration we're going to be district we're going to assume it's distributed such that people's desire to use the platform is library in observables their utility from participating as a function of all others decision participate so that's the equilibrium the network equilibrium so i's probability use the platform p is a function of others use and how much i'm being monetized so i only in so if we're on twitter i only indirectly care if alex is being forced to watch lots of ads as long as alex is on the platform i get the benefit of alex being on the platform i don't care if he's being slightly unhappy or when he's on twitter because he has to watch ads based on only indirectly care if those advertisements get him to stop using twitter and then i might stop using twitter okay so uh there's going to be a distribution of opportunity cost that's going to give you a logic demand function for each group you're going to be in equilibrium where you've got this symmetry right the probability of me using is optimal given the probability of everybody using and then the way kind of computationally we're going to solve for moving to new equilibria is that after there's a disturbance so for example after a new tax shows up and facebook pages at the level of advertising we're going to move to a new equilibrium after a series of cascades of the network right so again imagine advertisements go up on alex he uses the platform 50 percent less that convinces me to use the platform 25 percent less everyone connected to me is now using the platform 10 percent less etc etc what's important is that for the equal current equilibrium to be stable the this infinite sum of cascades has to have a finite sum right um it doesn't have to have a finite sum and that means that you were in an unstable equilibrium start the way i'm thinking about it in this time of corona viruses with r not right so we all know that if r not is the odds that any individual travel transmit the virus to another person if you have r not equal to or greater than one eventually the entire you know one infection is going to infect the entire society but if you have r not slightly less than one one infection is going to infect point nine people is going to affect point eight people blah blah blah and that sum is going to be not infinite um and here's just an illustration of stable versus unstable equilibria and so we're going to assume that facebook starts in a stable equilibria and actually actually that's not actually an assumption we can um in alter the model that we write down can certainly computationally say hey you're starting in an unstable equilibrium i can say empirically it looks like facebook starting in a stable equilibrium which makes this analysis possible okay sorry sir yeah please can i ask a question what's the distinction between omega and mu you've got omega of mu and why do you need both of them why do i need omega of mu oh yeah um the reason is that mu is the utility that i get from the platform given this setting and then omega is converts the utility of the platform into a probability of use right so basically what the omega has in it is the random opportunity cost and how are you going to be able to measure distinguish them because the so what we're actually going to do and you'll get more details later is we're actually going to kind of directly solicit utility so we're going to ask people what you know um how much would you be willing to pay to give up facebook that's going to give us kind of a demand curve and we're going to say what's generating that demand curve is these random opportunity costs and i can i'll go into more detail if that's not clear after i get to that part okay so again theory side what we're going to do is just ask what is um what's the first order condition for the platform again that's going to be an infinite series so just to keep things simple here i'm just showing the first two terms of the infinite series right so should i increase the level of advertising on person number one well if i do i get more money from them if they participate p i but i lose the odds that they stop using the platform because i raise the fee on them times the amount by which i raise the fee so d mu i d fi that's the disutility from the fee and then d omega i d mu i that's how much less do you use the platform because your utility went down so that's the distinguishment and then now we have this network effect which is if person i leaves that's going to discourage all of the jays that i likes or encourage all of the jays that i doesn't like to leave or stay and of course if the jays participate then you get to collect net revenue from them fee subject okay and of course you can some additional terms into this so what's kind of the takeaway here well this guy over here is just sort of classic third degree price discrimination and we just get this extra series of terms which is sort of the net the right way to price discriminate if you have these network effects so you should increase fees on those who inelastically demand the platform and decrease fees for those who provide lots of value to other people especially when those other people pay a lot of money to the platform again so this is very much in the tradition for genre you know do i subsidize side a or side b well you should you shouldn't subsidize the side that inelastically demands the platform you should subsidize the elastic side and the side that makes network effects so okay now we're going to calibrate the model so to do that we're going to survey facebook users we're going to use facebook because it's ubiquitous social network so it's easy to find people who have opinions about facebook on one polling we're going to get 73,000 responses on google surveys of those we're going to use about 50 in the final analysis we're going to use google surveys and google surveys has a registration of subjects age range and gender so that's how we're going to divide up the universe into 12 demographic categories by age and gender the way that we're going to solicit these demand functions is we're going to ask questions of the form would you accept $10 a month to give a facebook would you accept $15 a month to give a facebook would you accept $20 a month to give a facebook so that's following the alcott at all a ER paper as well as brunelsson collis and eggers so that strategy but we're not going to come around the end and actually do the enforcing of the the way that you actually would hypothetically guarantee that someone's revealing their true preference because then you do this as a sealed bid auction right we don't actually do the sealed bid auction part we just ask people to you know yes no do you accept this offer in the alcott at all paper the sort of the non double blind sealed bid responses aren't very different than the ones you get if you just ask so we feel on sort of solid ground and you'll see that our demand curve that we estimate overall is very similar to what these other guys found so I wrote down a very very general utility function but we're going to have to be more specific to actually bring it to the data so this mu i so what's the value from the platform if you decide to use it it's going to be the sum of the utility that person i gets from j if j uses the platform times the number of people who are type j in america times the probability that someone in group j uses the platform times z i of j which is the probability that we're friends given that both of us are using the platform and then minus the disutility from advertising so what I want to point out here is all of these guys are linear so and that's just has to do with data constraints of course we think in the real world that people's utility from digital platforms might be highly nonlinear in lots of ways you can think of this as an approximation that's going to be better when we're closer to the start and maybe a little bit weaker as we think about more extreme scenarios we're going to think about opportunity costs being distributed such that omega i is logistically distributed so that when we fit a logic curve to the demand function that's going to make sense and then we're going to think about that demand curve and the initial probability of use being separately identified for each demographic group and then of course we're also bringing data from facebook on revenues for each demographic group which are based on facebook's quarterly reports and api again there's some approximation there because they don't tell you everything you would want to know ideally like i said the questions are generally of the form would you give up facebook for one month in exchange for x dollars choose yes if you don't use facebook and then we can ask the same question for different demographic groups would you give up all connections to people on facebook of demographic group x for one month in exchange for x dollars we rebalance those so that those add up to 100 of the values so we're just going to assume like we do here that the value of facebook is the sum of the value of connections to all individual friends on facebook which again isn't necessarily the case of course facebook can have other benefits beyond that okay so this is what our overall demand curve for all of the different groups look like so again we're asking these questions would you give up facebook for five dollars a month and so this dashed line is going to be our demand curve overall but of course we're separately doing it for each group and as you can see it's a little it's a lot more precise than bring y'all sin at all because we're asking way way more questions about a much narrower range of prices but you can see it's sort of overall consistent with what they found for this demand and as you can see there's this sort of really long tail here right you can get up to a hundred couple hundred dollars and there will be a significant action of a population who would not give up facebook and so when you see these sort of average values of facebook a lot of that is kind of being driven by a long tail of people with high valuations i will say these previous papers they they have this sort of demand curve in them but they mainly focus on reporting median values so one sort of innovation in this paper is taking that entire demand curve seriously and seeing if we can do something with the calibrated model rather than just focusing on median values um so the next so one of the things we calculate is sort of this network of network effects that each demographic group gives another demographic group so here in this figure it's a little bit of a hairball um shows the 124 values of all connections to from a member of one demographic group to another demographic group or to their own demographic group um and it's uh the size of the nodes has to do with the initial size of the user base so older women are the largest user base on facebook right now so these are the most valuable connections each of these connections are on average worth more than 50 cents a month and so the arrow is pointing towards who's getting the value so female 65 plus very much value connections to young men male 65 plus really value connections to middle aged men and middle aged women really value connections to older men um so this is these are the most it's so and that's kind of a general pattern that kind of the value flows tends to go on facebook from younger male people to older and more female people here's the top 10 most valuable connections you can kind of see that flow that i talked about this is just showing just um how how kind of flipped it can be so female 65 plus getting a lot of value from everybody especially young men um whereas providing value out not so much to young men at all more than middle aged people um we can flip it around to young men get very little value for most people but provide a lot of value out so um and then finally so um we try to talk to facebook to try to get as much internal data as they could uh as we could to calibrate this model as well as we can they were not super helpful as you might imagine but they did give us one data set which allows us to confirm one of the things we did so this is from a data set that they have on connections between people of different age groups the frequency of connections amongst people of different age groups and so on the x axis is what we find as the percentage of friends who are between those parent age groups and on the y axis you have what they find and you can see that there's a pretty strong correlation um we don't perfectly hit what they think the ground truth is there's no reason to think that we would be exact because we're kind of asking very loosely about friends and they're taking they have a sort of very technical definition of what an active friend is um but I think there's some signal here okay um so now we get to actually give you the results so we estimate that facebook in late 2019 for americans 18 and older generated monthly 1.8 billion in revenue 12 billion in consumer surplus and has 154 million users um so the first question we need to ask is is facebook currently profit maximizing because we suspect they also have this large user base motivation that a lot of platforms has so we're going to sort of infer what that value is and then use it in all the simulation so that we can justify the current equilibrium through that part of their utility function and then we're going to evaluate all these different policies that I promised that I would so what's right here is we just ask the question what would happen if facebook just tried to maximize current revenues and didn't care about having a large user base at all so what they would do is they would raise advertisements a lot so they would basically more than double the level of advertising that's going to initially lead to a decrease in consumer surplus for lots of groups by a sizable fraction but it's going to get worse as that network effect starts to percolate so first we increase the level of advertising on these groups then what's going to happen is those groups are going to use the platform a little bit less that's going to make everyone slightly more unhappy because now not only is there a higher level of advertisers there's advertisements there's less users on the platform and that's going to just keep on going until you reach a new equilibrium one sort of interesting result we have is that you don't need to calculate a whole lot of cascades as you can see here at least in this current calibration after three cascades the we basically achieve the new equilibrium to three significant digits so we say profit maximization would increase facebook revenues by 2.3 a month 2.3 billion a month so why not do that well because it decreases the user base by 49 right and presumably facebook doesn't want that there's lots of different reasons why facebook might not want a shrinking user base and you can kind of divide them into two categories as is there a motivation pro social or sort of neutrally social on the pro social side you might think that this is for future monetization right so i need a big user base so later i'm going to be able to sell people oculus and libra and get good network effects from those you might think that i'm collecting data because i'm going to make this really great product that i'm not selling right now but eventually i'm gonna um and then what a final possibility is in this model we're only modeling america but if americans had positive network externalities on users of the platform not in america that would be a benefit to having a large american user base that's not captured in our monetary value because we're only modeling america on the other hand you might think that the platform just wants to be big to deter entry or that it's based on sort of stolen data that the users haven't really internalized the disutility of that and it's in some way sort of dishonest this big user base motivation so it's a little bit ambiguous which of this is the right way of thinking about it right now i'm going to focus on the positive interpretation that facebook having a big user base is for these sort of positive reasons of their sort of developing new products in the future um as well as you know trying to prevent their their system from unraveling okay so this is what i'm going to call the first best so this is asking how would a benign social planner run facebook of course it's a little bit unrealistic but for just to give us a baseline and then the other thing assumption here which is why it's going to be possible for data is labor to actually look better than the first s is here i'm assuming that ads must be eliminated rather than arbitrage so here i'm saying facebook um the nationalized facebook is going to be making big investments in raising platform quality in addition by part by cutting ads which is different than um the data is labor scenario where like platform quality doesn't change but people are receiving checks in the mail for using facebook so what do we get from this well we see that obviously not obvious but we see that the nationalization first best entails running a subsidy for using facebook which is not surprising given that we think that there are some inframarginal network effects that a platform doesn't internalize um and we find that we can boost social welfare by 9.6 percent by running this big subsidy um now let's think about six different tax redistributive and regulatory policies and think about how close can we get to the first best with stuff that's a little bit less onerous than bernie sanders running facebook first i'm going to think about an ad revenue tax so very similar to what uh france has proposed or is has implemented i don't know if they've actually raised any of that revenue yet we can think about a per user tax which is sort of logically sort of the opposite approach what if you wanted to sort of discourage usage of facebook and get a smaller user base that would be an approach and then finally data is labor which is model is facebook rebating some percentage of ad revenues to users here just to make the example sort of really extreme where to assume that 100 percent of ad revenues are just rebated to users of facebook um again sort of the sort of the high level theoretical result i want to focus on here is that if you just multiply a one one minus tau if you multiply this big tax on the front of that first order condition that i showed you earlier you'll see it doesn't really if it hits every part of facebook's revenues at tax it's actually not going to move the the optimization condition at all right so that's the sense in which facebook is sort of inelastic to taxes what makes them elastic to taxes what makes them keeps them from being perfectly inelastic to taxes in this model is that a tax on digital ad revenues isn't going to be incident on this shadow value of maintaining a large user base again so if you tax ad revenues substitute into the large user base motivation and then the opposite so if you tax the number of users facebook is going to have the optimal optimal excuse me the exact opposite incentive they're going to want a smaller number of users who are each getting monetized more intensely and so here are the results from those simulations the things i want to highlight is a 3 tax on facebook can potentially raise consumer surplus by 1.3 percentage points and social welfare by 1.1 percentage points so the difference between those two coming from less ads being shown and so facebook losing out on those revenues alternatively if you tax the number of users and in such a way that you collect the same amount of revenue you're going to have basically no effect on welfare this is actually it's like negative point zero this is like negative point zero one so let's talk a little bit more about taxes so digital service taxes and ad taxes what they're doing is they're changing firms aside versus revenue trade-off and then theory suggests that platform quality and participation is going to be too small due to the expense distortion so i've been talking about that let me say exactly what i mean by that when you're a platform what you care about is not maximizing total utility but what you do care about is creating these network effects for certain j's right but who are the j's that you care j being a certain altar right but who are the altars that you care about creating utility for you only really care about creating utility for people who are going to pay you a lot of money in terms of advertising revenues if they use the platform and people who are marginal right you don't want to create utility for people who are always going to use your platform no matter what right from the perspective of the platform that's wasted however from a social perspective creating utility for people who are always going to use your platform no matter what is actually a positive and so that's where the expense distortion arises it's the inframarginal network effect as opposed to the marginal network okay um so again just to summarize our tax results and then of course there might be more nebulous externalities right now right now those aren't in this current model but could be easily added as just one more thing you keeping track of like we say we think that the tax would have this effect um and now if you're a foreign country now from the perspective of France taxing a foreign company using this strategy is going to be even more attractive presumably because more Americans own stock in facebook than french people own stock in facebook so but why not so the first thing is if you have these tax taxes that end at a border that's going to potentially create an additional distortion if there are international network effects so that's one thing to keep track of that i'm not modeling here and then a second thing that's not modeling here so what keeps this from going to infinity what keeps france from setting a 50 digital service tax if it's good for them rather than a three percent digital service tax well i think proposition 22 in california gives us a hand does anybody remember was anybody following proposition 22 in california okay then this is a fun little story this will be a good way to close out so in california um there was a movement uh to try to get uber and lef and all of the big sort of uh gig economy platforms to offer more generous contracts to the people who uh did the work for them so the contractors for the people who drive for uber and lef etc the way that this would have basically cashed out is as sort of a compensation floor for those people right so it basically regulations raising the amount of sort of fringe benefits that the contractors would get treating them more analogous to employees than contractors in the context of our model you can view this as sort of a forced increase in or forced change in the price to what's being charged one side to the platform so in other words facebook is going to have to subsidize the driver side of the platform more or charge them less right this is kind of similar to the logic of the tax that we just talked about so what keeps but what happened so when this ballot initiative was proposed uber and lef said we will just stop operating in california if you pass this law right and what happened the voters in california totally backed down they totally said you know what uber and lef we'd love it if you were nicer to your drivers but we'd be really really sad if there was no uber and lef and california 100 back down right and so what that suggests is that the ratio of platform profit to user surplus is going to be really important for understanding who's going to win these fights moving forward in terms of taxes on the digital economy because our sort of old principles of you know taxes are always bad might not super apply here in terms of consumer welfare finally i have a couple of results about regulation like i said before this is going to capture alex's intuition that if you break up the platform without getting any additional competition it's kind of a lose lose but we also imagine what if we were able to achieve perfect competition with interoperability so there's no destruction of network effects um and as you can see uh social welfare goes up by four and a half percent in this perfect competition scenario so it kind of gets you halfway to the first best um whereas a vertical breakup or a horizontal breakup that doesn't boost competition at all are going to be daring amounts of bad um yeah and so those are the headline results and finally i just kind of want to close with call to action um so our calibration is going to be in many ways limited we're we're based on this imperfect survey data um in order to calibrate the model we're going to have to assume that all of these aspects of the utility function are linear but um as far as i know the and having talked to facebook i'm pretty confident that this is sort of the most robust model structural model of a social network out there for trying to ask these sort of policy questions for policymakers to make wise choices they need to be able to compel platforms to share data and identify that we need to identify demand curves and network effects and platforms and regulators should develop and contrast quantitative models so that we can kind of start moving past more quantitative qualitative debates about the role of taxes and regulations and actually start putting numbers next to each other and saying well i think x i think x will go up five percent no i think y will go down two percent so thank you guys so much for your time thank you very much very interesting now let me invite alex white to provide a short discussion okay thank you interesting set of topics in this paper and i enjoyed i enjoyed reading it oh sorry can everybody hear me i i'm getting a beep in my bluetooth you got it okay now we can thank you okay thank you sets um very interesting set of issues and and discussion in the paper i just want to make a couple of quick comments um so the um it's definitely from a theoretical perspective kind of a puzzle why facebook doesn't seem to be profit maximizing according to the specification of the model i think that intuitively we we all sort of agree that that's what we would expect to happen but it doesn't really happen in the model and so i think along your in your list of um potential reasons it's good to talk it would be good to think a bit more about competition it's after all what you have is a monopoly model and so due to the fact that it's a monopoly model you're probably getting a prediction that their profit maximizing behavior involves higher prices than what you would get in a model with competition among different social networks now of course you could respond to that immediately by saying that well facebook basically is a monopoly um and so why do we need to take competition seriously um but this gets to a broader issue in this debate and just if you look for example at the um recent reports wonderful reports by people like jack primair and fiona scott morten and then the firman report in the uk a big theme there is the distinction between competition in the market right so it would be like two different facebook a and facebook b who compete head to head versus competition for the market and i think everybody kind of agrees that to the extent that market forces can discipline platforms it has to be driven by some version of competition for the market rather than sort of plain vanilla competition in the market and what exactly does that mean does that mean that there are going to be startups or um you know less less popular competitors who are there constantly acting whose shadow is constantly disciplining the dominant firm from behaving in a way that's anti competitive or harmful to consumers or not and um i mean i think that at the moment people feel as though this threat of competition for the market is not a sufficiently strong disciplining factor but it's also an issue that needs to be studied more and so i'm a bit hesitant to just draw model draw um conclusions from a monopoly model when i think that um what we're really interested in in in these markets is understanding under what circumstances can competition can or cannot competition for the market act as a disciplining device to induce good behavior by dominant platform so that's sort of a broad um thematic issue that i think would be interesting to add to your list um also i should mention that um there's a paper about by am about amazon by rimers and walled focal called the throwing the books at them where um they also address a similar puzzle of why amazon doesn't seem to be behaving in a profit maximizing way with respect to its uh i think it's with the with the book pricing so um you you might check that out and while i'm on the topic of of um quantitative models of platform competition there's a forthcoming paper by min j song looking at advertising in um uh i think new uh magazines um it's coming out in a j micro that involves a structural model structural estimation exercise um you know effectively in two-sided markets and i think that that hasn't caught your attention so take a take a look at that um the other um point that i just want to briefly mention regarding you didn't really talk about it in the in the talk so i'd be happy to discuss it more offline but with respect to well i guess it'd be it'd be online but not with everybody else here the um uh the the the question i have is about the need for approximation in your solution to the model so your model is um very similar to the wild 2010 model the a r paper which uh you know you you you certainly mention and acknowledge um i think there's some confusion about the technical aspects of that paper and the insulating tariffs so in that monopoly model one question is how do you write what what are the dimensions of maximization in solving that model and then another question is how does the platform implement implement the solution and um the the way it works is that you can maximize a multi-sided platform model with respect to the number of users served on each side and come up with an exact first order condition for the um representing the the solution and that pins down the the optimal price for the platform the optimal actual final price that the platform wants to charge and then the problem is that if the platform just announces those prices there can be multiple equilibria in the next stage of the game and it's in response to that problem that um glenn uses the idea of insulating tariffs which is you know it comes from this paper by um uh in the in the 70s in journal of public economics similar idea um where they where they talk about it in implementing a public good but the point is that contingent pricing in in the context of um monopoly two-sided markets is a separate issue from how to identify the um the optimal prices and so i think you might get you might you might be able to rewrite your maximization problem just with respect to the number of users served on each side and not have to deal with the approximation so i can explain that more um um you know without everybody having to worry about the details but i think that that could make things simpler for you so um and i and i like very much your your idea of using surveys to elicit data about people's willingness to pay for facebook that's not my area of expertise but it certainly is nice to see that because it's something that i'm very interested in and i've never done myself so um very good um i think i think that's that those are the points i wanted to to make so uh uh thanks very much thank you alex so let me yes so so said um before we move on to general q&a uh feel free to to respond to alex so yeah alex thank you so much for those really thoughtful comments um thing number one obviously we we writing this paper we thought really hard about is facebook really a monopolist um we leaned away from modeling explicit com market competition in the model because we were worried about uh and creating too many multiple equilibria so it's kind of a modeling simplification to think about a monopoly um but i think the model i think if we could estimate a cost of entry of entering with a facebook replacement i think that could be incorporated to the model it's just how how am i supposed to calibrate that and if you have any papers or ideas about how to think about um calibrating the threat of entry i'd be that's something i'd love to try to incorporate um and but yes clearly our shadow value of facebook wanting a large user base is a combination of both what what you might think of this anti-competitive uh motivation right building a big moat so that no one else can enter against you as well as maybe these more pro-social motivations you're a hundred percent right which is in order to understand the welfare impacts of our different simulations you would really want to understand which of those two those are um as well as think more seriously about monopolies just generally yeah remind me just sorry to interrupt i just remind there's this there's a new paper by zingalis and co-authors about the kill zone that you might take a take a look at that's quite a that's quite an interesting paper on that topic yeah yeah that's a good idea too um so along those lines i guess what i would say is even though we do have this sort of simulation of perfect competition in there that is competition in the market right that is imagining that all of facebook's kind of secret information becomes available and then people can enter with just basically portals to read that news feed and read that social graph so that is a scenario that has been painted that you could kind of maybe get in the market again most of this paper is more thinking about managing a natural monopoly than about how would you go about encouraging or harnessing competition um the min the min the model that you talked about magazine advertisements and a multi-sided platform reminds me a lot of the riceman paper on multi-sided platforms in the in the yellow pages with advertisements so that's a paper i'm familiar with and certainly have learned from but i'm looking forward to reading about this paper um and then this sort of insulating character discussion yeah we can talk about whether computationally it would be better or not to try to use that approach i think that as a practical matter we wrote down utility functions such that there's going to be a unique equilibrium with these nice logic demand curves um but uh so i'm not super worried about like insulating am i at a local equilibrium it's the local maximum instead of the local maximum i'm not so worried about that um and i think that there's some sort of just sort of realism and just sort of the intuitive clarity about talking about real world uh pricing well no but this is my this is this is my whole point which is that to solve the maximization problem with respect to number of users rather than with respect to price you're not using insulating tariffs you're just representing and what one can represent the problem as a as a problem with respect to quantity and get exact solutions without the need for approximation that imputes the prices that infers prices and then you just say those are the prices it doesn't have anything to do with it's it's it's simply a question of how you represent the so the problem not a question of the platform's behavior in a monopoly model so that's that's my point fair fairly play all right yeah let's take some questions so uh if uh whoever has a question i think you just feel free to unmute and and ask it jack ever are you need to you're on mute jack thank you very much the first thank you very much so this is uh really a good work it's kind of nice to have a paper which looks at bigger policy issues we don't do enough of those in economics um the my question is about how do you model your model of advertising are you assuming that if you decrease advertising by half the revenue will decrease by half so that there's no price elasticity of advertising yeah so we're so we're another big simplification in this model is we're going to view is exactly that so there's going to be a linear relationship between advertising revenues and user disutility so if you cut advertising revenues by half we're not literally keeping track of number of ads we're keeping track of disutility caused by ads so if you decrease the disutility caused by ads by half you're going to lower your revenues by half okay do you have i mean this seems to be a very strong assumption and seems to be rather important because uh yes uh incentive i mean all the revenue part is is is crucial so do you have any idea whether you could test this or uh you know so the way that we calibrated here is we ask people about how much that they would value and add free facebook so how much would you we do a willingness to pay experiment so how much would you be willing to pay to eliminate ads on facebook which i guess which totally is not perfect um i know i'm vaguely aware of some uh papers that in particular platforms look at you know how how elastic is users to number of advertisements we could try to grab some parameters from them um but as far i am not aware and i'm really eager if anybody you guys know that looks at a non-linear relationship between advertisements and disutility um so i think that's a really important open question right both like what is the social i mean that's it i wish i had marketing phd friends that could answer this question like is is their social value to advertising and how bad are they for us privately in terms of unpleasantness because certainly it's probably not non-linear and it's probably highly non-linear in your life i think the only way to do this research is to roam the bars around the offices of facebook and get engineers drunk and get them to to tell you yeah exactly and and exactly and that led me to my final point which is a lot of the information you would want to know to really model facebook properly is private information that facebook has and i guess that's a common problem in antitrust but it's really exacerbated in this setting so said uh uh haski has a question so haski can just unmute yourself and ask the question it's uh it's less a question more just a comment following up from from jack's observation so you know some of the consumers are a bigger deal in terms of their network externalities and for advertisers as well some some consumers are more valuable to to advertise to so um this is just kind of exacerbating the the treating advertisements uniformly is um uh a little counter to the application so we uh we allow for price discrimination so yeah in the model the model does allow for facebook to charge it's third-degree price discrimination so different groups user groups are allowed to be charged uh different levels of advertising uh at least at least in in the model i yeah i um so when you run your advertising counterfactual you do the same uh i think it's possible i have to look at it um i know that in some of these simulations we allow for full price discrimination across the groups and i know that in some of them we do not um i have the top of my head i can't remember which simulations are which but that's certainly my code allows for it can i ask a question about the assumption or maybe i missed one point about the impact of advertising because you seem to assume that people dislike ads so i believe that this is a good assumption in yellow pages world but uh for facebook it's funny because he has the opposite assumption in yellow pages world but keep going yeah yeah but uh but you know at least people believe that you know people dislike ads in general uh in standard you know when i'm interested in content but facebook does more than that right they are matching also your what you want to see in the news feed and this is more about matching my taste to the content rather than showing me you know just ads uh random ads so how should we then take this point into account in the in the analysis that's a really good point and the way i would think of it is as just another way of saying shocks point right which is in the real world there's probably lots of non-linearities but what i the way i would put it is it's got to be that at the margin that these advertisements are disutilists right because at the margin they've got to be crowding out better content or else you just show a news feed of advertisements right so again what we're trying to figure out is sort of at the margin what is the disutility from raising another dollar of advertisements in terms of disutility of using the platform that's the number we would want to know in principle and that would be sort of locally correct even if sort of inframarginally sort of like the very first ad that you show me is really valuable and helps me a lot and actually raises my utility there's got to be at some point at which the 100th ad is having a negative marginal effect and that's kind of the way i would interpret it is the more you're worried that this is a problem the more you should think well i'm only going to trust sets model locally but i also have a question about this margins because uh at the beginning i thought that you are measuring in the model with the probability of usage the intensity of usage but then when you set up the survey it was more like participation margin so i think these things might not be necessarily the same right so people who value facebook user who value facebook membership might be marrying it just to see others content rather than using it themselves so then there might be some mismatch between the usage intensity and the membership so i don't know so the second part of that i think we have i we do have heterogeneous network effects so it can be the case that some people like being connected to another person but don't provide value back to them so that that we have what we don't have that your right to point out is um i'm sorry i totally lost it what what what margin you are talking about i think is i was also confused when you presented the model i thought that you were measuring oh yeah we don't we don't have intent sorry yes i when i was speaking at the early on i was speaking very loosely about intensity what i should have say is people like seth benzel use it less on average right instead of saying seth any individual ex ante is going to 100 percent participate or 100 percent not participate but because we're thinking about a bunch of people within a demographic group i was loosely saying um yeah so people like seth are participating 30 less right um you're totally right that intensity is another margin we can think about here and that actually might be the next thing that we add to the model that is important thank you so um i will stop the recording now let me thank seth and alex for the presentation discussion for the purpose of recording but feel free to stay and discuss and i'm staying staying here as well