 We're very excited to have Oren Rescheff present his work. If you have any questions, please feel free to ask them. Take it away. Thank you, Andre. Thank you for having me. I'm actually really excited for this talk. I just realized when I was kind of looking for the seminars that many of the people affiliated in this seminar are cited in this paper. I'm kind of very excited to hear your comments and feedback about this project. So today I'm going to be talking about the effective entry on incumbent firms and platform markets. And when I talk about, I'll call it platform markets. I'm referring more generally to platforms or two-sided markets. These markets have been spreading to numerous areas of the economy. So for example, the largest firms in the world today, Amazon, Alibaba, Microsoft, Apple, Google are all platforms. And also if you think about the largest firms of the future. So here, for example, I look at the most successful unicorns. So these are private firms with valuation of one billion or more. Again, 60 to 70 percent of those are platforms as well. So that means that platforms are very big and are growing very rapidly, meaning that more sellers and more firms are operating fully or at the very least partially within a platform setting. The goal of this paper is to try to understand how is this rapid entry of firms onto the platform is impacting the incumbent firms and how is that different or similar to more traditional markets. Now, what I'm going to be arguing in this paper is that the ultimate effect of entry on the incumbents is theoretically ambiguous. And basically the trade-off that I'm trying to pose is that similar to traditional markets, entry usually is associated with a direct negative effect. So there is a decrease in market share driving down profitability. On the other hand, because we're in a platform market, because we're in a two-sided market, these indirect externalities, indirect network effects are going to create positive spillovers between competitors on the same platform. Now, just to set the intuition, I want to start with a stylized example. And as you'll see, this is very similar to what I'll have in the empirical setting. So I do think it's worth spending a minute or two on that. So what we see here is some platform, and this is a geographically constrained platform. I marked here free incumbent firms in red. So you should think about it as free drivers operating on Uber or free listings on Airbnb, or in my setting, these are going to be free restaurants that use their platform for delivery. Now, what I'm going to be interested in understanding is what happens to those red incumbents when we start adding more of these blue competitors onto the platform. So are they going to perform better when the platform looks like this, or maybe when it looks like this, or maybe when we start adding even more competitors like we have on the graph on the right. Now, in traditional markets, again, all things equal, we would usually think that those red incumbents would be referred to operating a platform that looks like this, rather than one that looks like that just to alleviate competition and reduce this competitive effect. However, the main argument I'm going to be making is that in a platform market, not all things are equal. In particular, if you think about it from the consumer's point of view, if there are not a lot of firms on the platform, it's not going to be very appealing to consumers. There's not going to be a lot of things to find there. So we're not going to see a lot of, in my case, virtual traffic on the platform. If we start adding more and more businesses, the values of the platform as a whole increases. It attracts more consumers onto the market, and that might potentially benefit all of the firms, including those red incumbents. So really the trade-off here is between the increase in market size and the decrease in market share. Now, empirically, it might seem appealing to just compare platforms that look like this, or markets that look like this, to markets that look like that. However, of course, the main issue is endogeneity. So entry onto the platform is strategically determined. And of course, there are other unobservables that we as the researchers don't see. Here's one example. It could be the case that there are more businesses in this market because there are more consumers, and not the other way around. So that's a reverse causality issue, for example. So how do I solve this empirical issue? Well, in this research, I use a pretty cool data set. I use proprietary data from Yelp and YTP. YTP is the Yelp transaction platform. It is a subset of the Yelp review website, which allows for transactions between consumers and local services. I'm going to be focusing specifically on food order and food delivery. Now, in order to really identify the effect of entry, I'm going to be exploiting quasi-exogenous variation generated by YTP's agreement with Grubhub delivery service. So for those who don't know, Grubhub to date is the second largest food delivery platform in the U.S. At the time, I think it was the largest one. I'll talk more about exactly what this agreement did and how it changed the platform. But at a very high level, what it did is to sharply increase the number of restaurants on the platform. So very similar to the example that I just gave. And I'm going to use variation in that shock, in the intensity of that shock across different geographies in the main specification that's going to be cities in order to conduct a difference in differences analysis. So comparing treated areas to control areas. Then if I'll have time to talk about it, I'll mention it towards the end. I'll estimate a simple structural model, which will allow me to extrapolate to additional market conditions. In particular, I'm going to be asking what happens when the platform becomes very, very large, so much larger than what I see in my data. So just to formally state the research question. So I'm going to be asking, as I said, what is the net effect of entry on the incumbent firms operating on the platform? And what I think are the most interesting effects are actually not just the average effect, but the heterogeneities. So I'll ask, how does this effect depend on firm attributes? How does it depend on platform attributes? And then I'll also talk about how do incumbents respond to increased entry on the platform? I think there is a question. Yeah, so I'm thinking a little bit about the conference that many of us were at last week, where John Vickers gave an overview of this paper on patterns of competitive interaction. And one of the takeaways from that stream of work is that prices can go up with more competition, even in the absence of this additional consumers channel that you're describing. So the idea is, in a spatial model, let's say, without somebody in the middle, we're competing hard because we're also competing for the guys in the middle. If somebody comes in between us, we give up on those customers, and that softens the price competition between us. So you can get potentially higher prices in particular with entry, even in the absence of the kind of channel that you're describing. So I'm a little bit curious about what your outcome measures are and if this is an effect that's potentially one you could identify. Thanks for this comment. This is actually very interesting. Thank you for bringing this up. I will say that in this paper, the main outcomes are going to be looking at are sales and revenue and not prices. In fact, one thing that I'll say from the get-go, and it's actually not in a presentation, is that I don't see a lot of action on the prices margin. And I think that this might be for several reasons. One of them being the specific settings. So these are restaurants. They rarely change their prices. So it seems that this might be the thing that's moving in. Also, because it's only changed on the platform and prices off-platform and on-platform might be, are usually the same. That might be the case. They don't shift their generally profit-maximizing price. So you should think about this setting as prices remaining constant, and this is actually what I empirically get. And only what I'll be looking at is their sales and their revenue. All right. Yes, David. Thanks, yep. It's helpful. We have another question. David? Yeah, thanks. So a bit more related to the channel that you want to discuss, namely more customers coming online, and specifically to restaurant delivery apps, when they tend to expand very quickly or get flooded, I shouldn't say tend. I have one anecdote in mind, which is the city that I live in. And there are at least two platforms, and one of them has many fewer restaurants, but the perception of consumers is much higher quality. And it's not just about the number of these restaurants. And I don't really think that the ratings within the app give a full story there. And obviously that's intertwined with the channel that you want us to think about. I wonder if you have thoughts on that. Yeah, so I will be looking only... So again, I think this is one of the main issues. We're just kind of comparing platforms that look like this to platform that looks like this, right? Because you said, look, there are other things that are different between these two platforms. I'm going to be looking at a particular platform and looking at an exogenous, exogenous change within that platform. So you can think about these things kind of being held constant. And in the next slide, actually, when I'll talk about the results, you'll see that your intuition is exactly correct. And one of the main driving forces is really the quality of these firms. I am going to be using ratings as quality and have some of a better measure, but you will see that there's a lot of action based on the ratings of the firms. And that's really what's kind of driving most of these effects. Interesting things. All right. So let me kind of push forward with the research question. So I frame all of my research questions in terms of the firms on the platform. Of course, when I have implications to all of those individual firms operating on the platform, that's going to have an aggregate effect on the platform as a whole. And I'll talk about that towards the end of the talk. So let me start with a preview of the results. So I asked the question. Let me answer it. I'm going to spend the rest of the time just talking about how I got from the question to the answer. So what is the effect of entry and the performance of incumbents? We'll hear really the average effect that I find is actually positive. I find an increase in weekly revenue of about four and a half percent. And this is really seems to be driven by the fact that once we add more firms to the platform, that increases the consumer base. Using the simulation, what I find is actually an inverse U-shape relation between entry and sales. When the platform is relatively small, which is what I see in my data, entry is going to benefit the firms on the platform. When the platform becomes large enough, this trend reverses and additional entry is going to be harming the firms on the platform. Looking at firms heterogeneity, then I look at heterogeneity by rating on that platform. And what I find is that actually the average effect is masking considerable heterogeneities. In particular, all of the positive effect is coming from these high rated or high quality firms. They get an increase of about 15% in weekly revenue. In contrast, low quality firms suffer from a decrease of about 10% in weekly revenue. So what we actually get is this divergence on the platform story with high quality firms performing much better and low quality firms performing much worse. Going back to those inverse U shapes, if you look at the bliss points, the points at which profits are maximized, you find that for high quality firms these happen in a much higher saturation rate. Meaning that high quality firms prefer to operate in much larger markets or alternatively prefer to have more firms entering their platforms. And of course I'll talk about the implications of that. Finally, look at an accomplished response to firm entry. We already said I'm not going to find effect on the price margin. That might be just a result of the specific setting that I'm looking at. I do find suggestive evidence of increasing investment in quality which I'll talk about and changes in advertising behavior which I won't have time to talk about today. So before I get to the actual project let me just super quickly yes Aaron, another channel that you might think about is the nature of the products offered. On these platforms doing better do they offer more stuff for takeout rather than eat in? Do they change the kinds of offerings that they have to try and cater for this business or do they try and differentiate more let's say. That's a good point. There's probably other directions. Thanks for that. Let me just start with super quickly reviewing the literature. Hopefully this contributes to the work on network externalities. The first thing I'll show you is pretty what I hope is a clean estimate of the magnitude of network externalities at least in this setting. I hope to contribute to literature on two-sided markets. I won't start summarizing the literature to the people who wrote the literature. I just want to maybe mention the two papers that I think are most closely related to what I do here. So the first one is Coward Al. I think the latest version that I found is a working paper from 2018. They look at bike sharing apps in China and they have a relatively similar design when they have an existing platform and they have a staggered entry platform and they kind of look at the spillovers and the competitiveness between them. Their idea is fundamentally different because the platform is the one supplying the bike so it's not really two-sided and the network effects are only happening between individuals that are moving the bikes around so their kind of forces are somewhat different. What I think is more similar is more recent paper in the JPE by Carl et al. They ask what seems to be a different question. They ask why some industries we have that the industry is kind of converging to one big platform and then in other industries we have segregation into different platforms. It seems like a different question but actually just the flip side of the same thing because their model it's mostly theoretical paper is again thinking about this tradeoff between network effect and competitiveness on the platform. Here I provide I think much cleaner evidence that is consistent with their finding as well and adding these heterogeneities and more granular questions. More generally of course this relates to other markets with positive spillovers to the effect of entry on firms and of course the importance of rating and reputation mechanisms in online markets. So with that said let me actually start talking about what I do. So as I said I'm going to be focusing on YTP it was launched in 2013 and I'll be focusing on the food order and food delivery part of the platform. Now importantly for my setting Yelp or YTP is not actually a delivery service. You should think about it as an information aggregator so think about it like Kayak does for travel services. It aggregates multiple travel services and that allows you to choose a flight or a hotel or a rent car. The same way works here. And these are specialized firms like Grubhub, Eat Street, Chownow Delivery.com etc. And then these firms are the ones that contract with the individual restaurants. So in a way the network of restaurants that you see on the platform depends on the network of partners that Yelp is affiliated with. So this is what it looks like on the consumer's end. If you look on Yelp and you look on delivery you're directed to YTP Here you can see the offering of restaurants around Warshu if you click on a specific restaurant you can get here the menu and you can finalize the transaction all within the Yelp website. Now in the background what's going on is that this order is actually fulfilled by Grubhub. So it's not fulfilled by Yelp but it's fulfilled by one of Yelp's what they call it is partners. And the consumers cannot control that but it's actually going to be a Grubhub key that picks up the food and brings it to your house. How do I use this platform in order to figure out the effect of entry? What I'm going to be using is I'm going to be using Yelp's platform level partnership with Grubhub delivery service which came into effect in February 2018. So basically what happened is they Grubhub and Yelp signed the agreement. Grubhub app and website were still operating you can still order directly for Grubhub if you had the Grubhub app nothing changed for you. However, for YTP users basically overnight the network of restaurants that they could order from substantially increased because now they could order from all of these Grubhub restaurants. And this is what it looks like empirically. So here we have the week with the week of partnership this is the 8th week of 2018 is normalized to zero we see here the total number of businesses on YTP there's a steady growth of course over time there is a very sharp discontinuous jump in the number of restaurants on the platform generated by this addition of Grubhub. Now luckily for me or importantly for my setting food order is fundamentally geographically constrained. So you cannot order food from San Francisco to New York and for the most part you can't even order food from San Francisco to Oakley. So basically you can think about it as geographically constrained. Now what I find is that there's actually a lot of variation in the intensity of that job across different markets. So some areas were affected a lot of some areas were not very affected by that partnership. Now what does that depend on? It depends on the Yelp network before the partnership. So how big were their existing partners? It depends on the Grubhub network and it depends on the overlap between the two. What I show you here is the variation in intensity. This is the change in share. It's the share of restaurants on the delivery platform out of the total number of restaurants in that city. The reason I share is that adding 100 restaurants in New York City is different than adding 100 restaurants in Berkeley. So I kind of want to standardize it by market size. I will say that in the paper I do robustness test with different definitions the results hold. What we see here is that the median is actually not or only marginally affected by that partnership but we do see a very long tail of places that got a lot of new businesses because of this partnership and other ways to see it is just to look at distribution in the pre-period, distribution in the post-period and we do see it shifting to the right but of course not homogeneously. So some places are more affected than others. So I'll talk about how I use this for the design but let me just mention what the data is. As I said, this proprietary data from both Yelp and YTP, I have data on all food orders from 2017 and 2018 so the shock happens almost in the middle of the data. I can't tell you because of the NDA a lot of details about the data. I will say that it's in the millions of millions of orders and users. I see item level data including the description and price. I aggregated for the main results to the business week level. The main outcomes that I'm going to be interested in are weekly number of orders and weekly revenue. For descriptive stats, again, I cannot show you too much because of the NDA. I will say that I have about 56,000 businesses in about 4,000 municipalities so think about cities and towns. One thing that I do want you just to pay attention to is these two lines. So here we have the fraction of businesses on YTP in the period before the partnership. On average it's about 5% and this is pretty consistent with what we know about food delivery especially in the pre-Covid area now it's probably somewhat higher and then the change in share is about 2%. Even after the partnership we're still 7% to 10% of the market are on the platform. I rarely see places with 20% and I never see places with 60% and this is one of the reasons why I actually need to extrapolate out a sample in order to figure out what happens when we get to these sizes. This is the data and let me talk for just one slide about the research design. In the paper I spent a lot of time talking to it and I spent like 30 pages of appendix talking about it. Today I'll just talk about it in one slide and maybe why it should be valid in the setting. I'm happy if you have any question to address any concerns you might have. The under observation here is the business week as I said, why is going to be the outcome of interest and I apply the inverse hyperbolic sign transformation so all of my results would be interpreted as percentage change. Post is a dummy variable for where we have pre-partnership or post-partnership and then treatment is going to be the treatment intensity which is either the continuous change in share so 1%, 2%, 3%, etc. A binary measure where I just compared those that were not affected at all to those that were somewhat affected and a short binary where I basically dropped the areas that were only marginally affected so if the change was .0001 I'm just going to drop you from the sample. The results are consistent across specifications. I allow for establishment level fixed effect and for weak state fixed effect to absorb time state shocks standard errors are clustered at the CT level which is the unit at which treatment is assigned. Two words about a notification so identifying the assumption here is the parallel trend assumption meaning that absent of treatment treated areas and control areas would have developed or would have trended the same way. This is fundamentally untestable in an assumption. In the paper I go for great lengths to kind of try to justify why I think this holds in my setting. The first thing is that really the fact that this happens at a platform level is extremely important. If we had different firms or different cities strategically choosing whether to partner up or not, that would have been a major issue of selection. Luckily this doesn't happen. Now you might still worry that maybe places that were ultimately treated or even acts anti different than places that were not. Again, I cannot prove this is not the case but what I'll show you and I'll also show it today is that at least in the year leading up to the partnership soon to be treated areas and soon to be control areas seem to be trending the same way. Really we only start seeing this effect following the shock. Now I also check whether there's other things going on at the same time at the same time as the shock. So I look at other outcomes that should not be related to the platform but should be related to the restaurant industry and the help in that market and it doesn't seem to have any effect and I do a myriad of robustness tests. I already mentioned some like how I define market, how I define treatment randomization inference, matching procedures, splitting the sample allowing for city differential city level time trends and the results seem robust to all of this. I'm happy to discuss this after I show the main result. So if there are no questions I'll go directly to the results. Great. Okay, good. So the first thing that I'll show you is the effect on the market size. So this whole premise of this idea is that we have increase in market size decrease in market share. These graphs are going to be aggregated to the city level and will include both the new businesses the one that were added through Grubhub and the old businesses the one that were already on the platform. So just to explain all of my graphs would look the same. This is the week with the week of partnership normalized to zero again and the y-axis should always be interpreted as percentage change. So 0.2 meaning 20% and minus 0.2 meaning a decrease of 20%. What we see here is the number of unique consumers in treated areas compared to control areas. We see a relatively parallel trends in the pre-appearance and then an increase following the partnership. The elasticity here is about 0.4 and of course when we have more sellers and more buyers we also see an increase in total transaction volume or total weekly revenue on that platform. Here the effect is about 0.5-6 if I remember correctly depending on the specification. But of course remember that the pie is getting larger there's more revenue to share on the platform but there are a lot more firms on this side than they are on that side. That exactly was the idea that we have an entry of more firms into the platform. So in order to figure out which one of those effects is stronger I'm going to be looking from now on only on the incumbent firms. These are the firms that were already on the platform. These are not the firms that were added through Grubhub. These are the firms that actually deliver with other partners and I'm going to be looking at firm level outcomes. This is weekly revenue per firm. What we see here is again a relatively parallel trends in the pre-period and then following the partnership we see a mild increase of about 4-5% depending on the specification in weekly revenue. Of course this is statistically significant once I aggregate it to pre and post. So it does seem like the increase in market size is dominating that increased competition on the platform. But as I said what I really want to get it is the firm heterogeneity. And the way that I do it is I define high status firm and low status firm based on the relative star rating, the Yelp star rating within city on the eve of integration. So just conceptually think about it as on the day before the partnership if you're above median in your city you're coded as high. If you're below median you're coded as low and I do again other specifications or definitions as robustness. I will say that I keep it constant across time. So even if ratings change over time you're still going to be always coded as either high or low and in the post period I'll explain why I'll do it in a couple of slides. Formally this amounts to running a diff and diff and diff design. So I have here again the pre-post and I also have here the triple interaction with firm quality. Here beta 1 captures the effect on high quality firms and beta 2 captures the effect, sorry and beta 1 plus beta 2 captures the effect on low quality firms. But before I show you the table let me show you the graphs. So these are the exact same graphs as the one I showed you two slides ago but now I separated it by high rated firms and low rated firms and what we see here is a divergence better. So high quality firms do unambiguously worse following the partnership an increase of 10 to 15 percent and in contrast low quality firms do substantially worse following the partnership and their weekly revenue decreases by about 10 percent. Of course you take plus 15 and minus 10 you get this about 5 percent that we had on aggregate. So really the answer to the question is not just what is the effect but on whom it depends who the firm is. It looks the same if I do it by sales instead of revenue very similar pattern and this is what it looks like in an admittedly ugly regression table. These are the effects on high quality firms I won't go through exactly each specification but you can see that they are positive and statistically significant and again the sum of these two coefficients is the effect on low quality firms which again is generally negative and statistically significant. These results are robust to whatever I throw at it and as I said I do throw at it quite a lot so most of the appendix is dedicated to these robustness tests but instead of talking about these I want to talk a little bit. Do we have any questions? Wonderful. Let me talk a little bit about what might be what else is going on here. The first thing I want to talk about is the role of cannibalization. One limitation of the data is that I only see what happens on the YTP platform so I don't know what's happening to toll for revenue and that is a limitation. What you can be thinking about is maybe this increased source of revenue on YTP is cannibalizing other sources of revenue. Let's say I always order from an orange pizza and I used to just call and order directly and instead of using that channel X I'm switching to channel YTP. In my data that's going to look like revenue is increasing but in fact real firm revenue is staying exactly the same. Again I don't have a direct way to address it because I don't see revenue off the platform. I will say that this is somewhat of an asymmetric argument because for local difference we see a decrease so that would be like some sort of reverse cannibalization. But nevertheless I did want to do something to see how big of a concern that is. The argument that I'm making is if people already know what they want to order and all they're changing is the channel for which they do it then we expect to see different search patterns. In particular we expect search intensity to decrease. I know what I want so I will go on YTP, I will click orange pizza I will click on orange pizza's business and then I will finalize the transaction with that business. What I do is I look at several measures of search intensity. So I look at the number of search queries you typed I look at the number of businesses you clicked on. I looked at how long you spent on the platform and I look at how generic was the search query. So did you type orange pizza or did you just type pizza in my area? What I find is that search intensity does not seem to be decreasing and if anything it seems to be increasing which I think is more consistent with my model of people are first choosing the platform and then choosing what to buy on the platform rather than people are first choosing the product or the service and then choosing which channel they want to order it from. Other things that I look at that I will just mention here super briefly. One is the importance of differentiation and I think this came up so I look at whether you are a pizza place if another pizza place comes in or if another sushi place comes in how does that affect your performance and I can actually give you reasons why it would go either way it turns out that on average it mitigates the negative effect so the effect is more positive. Results admittedly are not as strong and then the importance of ordering of search results so where you are in the ranking of the search results as it turns out I for that would be a more important mechanism as it turns out this is not the main mechanism that's driving the results. One more thing that I want to talk about before I'll discuss the structural model is investment in quality so what we see here is the correlation between firms ratings and their orders this is in the pre-period and of course we see a positive correlation, higher rated firms sell more I don't think this is on its face very surprising what we see here is what happens post treatment here in the dashed lines we see that their relation is becoming steeper and this is consistent with what I just showed you high quality firms are doing better, low quality firms are doing worse so that means that now firms have more of an incentive to be high quality so if we think that firms can at least to some extent influence their ratings on the platform that now they have additional incentives to do so and what I test is exactly this so I look I run the exact same model but now the outcome being incumbents a subsequent Yelp rating on the platform and I find that there's a positive effect on average about 1% increase in ratings on the platform just to get an order of magnitude a 1% increase in rating is correlated with a 3% increase in weekly revenue so I do think it is economically meaningful now of course I call it suggestive evidence because I don't actually see investment in quality I just see the rating which is at best a proxy for quality I will say that I test for other stories such as rating inflation maybe different people are selecting into living reviews on the platform I look at what happens off of this delivery platform and the data doesn't seem to support any of these alternative explanations one thing that I at least for now I can check is whether they are actually investing in quality or maybe just needing more fake reviews so that's another way to effect your rating that's not directly investment in quality but it does seem like they are at least investing in the rating whether ethically or unethically and with that said let me briefly talk about the structural model and again the one goal of this model is just to see what happens when the platform becomes larger so what happens when we get to Amazon levels or what happens when we get multiple restaurants on the platform now as I said before the existing data is insufficient I mostly have a lot of firms up to the 10% of the firms in the market so what I'm going to do I'm going to estimate a simple model and then simulate firm performance under different levels of platform participation just to give the intuition for the model admittedly it's a very simple model so the way that the model works is consumers are going to have a two-step decision they first have to decide whether to join the platform and that's going to be I assume this is costly so you can think about it as a fee they have to join or in my case this is going to be like a hassle cost so you have to interact with the platform spend time figure out how it works maybe set up an account for the first time all of that is costly to the individual and they have to do it based on partial information so they have correct beliefs about what's on the platform but they don't know exactly which businesses and what items they have on the platform but that they actually have to engage with the platform which I think somewhat makes sense once they join the platform I assume searches is frictionless and I assume that products or in my case services are both vertically and horizontally differentiated so there is this rating component there like good restaurants and bad restaurants but also there's this match value so they have my favorite pizza topping or they have a vegan option what these two together create is basically a preference for differentiation and I know Andrew Hage has a paper about that so here I kind of get that from the model instead of assuming differentiation what that gives me is that I have I don't have to estimate differential differentiation preference parameter but instead I have to estimate the utility function which is relatively straightforward and the cost distribution the cost distribution or the parameters of the cost distribution it what pins down the relation between the number of restaurants and the size of the plant that's kind of what drives people into the platform when it's becoming larger so let me show you the main results so here this is what happens to the network effects so here we have the fraction of firms on the platform 100 being everybody zero being nobody you should think about this as a number of consumers in the market what I see in my data is basically this strip so I only see around 10 percent I will say that the model does very well in predicting the data in this area I have no way to test how it's doing in this area because I don't have good data here if I had to guess I will say it's probably doing pretty well here so so here I guess and probably not very well at the high end this is less of an issue for me because I don't think even in COVID times we don't get to 90 percent or 80 percent of the market so I think that that's fine and I do care more about this area that's the best I can what we see here is that the relation is concave so when the platform is small adding more businesses substantially increases the number of people on the platform but these network effects seem to be attenuating as the platform becomes larger and larger and what this does is create this inverse U shape relation right when the platform is small the network effect dominates the increase in competitiveness and all in all the effect is positive when the platform becomes larger the network effect become much weaker and this trend reverses if I do the same thing separately for low quality firms and high quality firms I get this pattern and of course high quality firms always do better but what I think is more interesting to see here is that the bliss points these maximas you can see that for low quality firms are for a lower number of firms compared to high quality firms so again that means that high quality firms should select into higher platforms would ideally want their platforms to become larger compared to low quality firms this is by the way what it looks like for the full model I can't really explain exactly what's going on here in the time that I have but we do see here that like for high quality firms the saturation rate the bliss point is about 40% 40 something for the median firm it's about 10% and we have this kind of monotonic increasing bliss points of course this I will admit that these precise numbers change when I change the specification but this general qualitative results are fairly robust so let me just sum up everything in the last 2 minutes and I think I'm right on time so in this paper I ask what is the effect of entry on incoming firms and I do think that this has implications to all of the stakeholders in this setting so for the businesses operating on the platform then we do see that there are maybe different implications than some markets in particular investment investing in barriers to entry or choosing to enter uncontested market may not be optimal especially if you're a high quality firms for platform and consumers I showed that in the paper and I don't think it's very surprising that profits of the platform well for the consumers is increasing with the supplier base what I think is interesting here is that there's an additional effect so adding more businesses doesn't just make your platform bigger it also on average makes it better because we now have consumers choosing better businesses we have businesses investing more in quality or in rating in this sense and we also should have a selection of better businesses onto the platform so all of this leads to an increase in the average quality on the platform finally for policymakers there's now talk about breaking up big tech and of course this would increase competitiveness between platforms this paper maybe addresses the downside of that which is the loss of these positive network effects to reduce competitiveness on the individual platform we don't get this increase in quality which might benefit everybody so basically that might have negative effects on all of the stakeholders in this setting so if there's like one take away here is that the effect of entry is ambiguous and it depends on the quality of the firm and on a platform size and I think I'm right on time so thank you very much thanks Aaron so I think next John will give his discussion and then there are a few questions in the chat maybe we can post both of those until John is done great thanks can everyone hear me okay? great I love my slides so let me just say first things first I really like this paper a great deal and one of the things I like about it specifically is that it's very connected to questions that I think actual platforms have to think about which is what kind of partnership should we be doing how much should we invest in trying to acquire more supply and so I think it really is kind of getting at the heart of a question that these sort of platforms have to think about pretty much all the time and that's kind of already a great starting point for some work and so now that I've said that I've praised the paper I can get into things that I think are not criticisms but just sort of thoughts that came up that if we were kind of looking at another paper this is what I would like to talk about next I'm not pointing out anything here that I think is not already in the paper but I think that is there a more micro level of competition where we can analyze and explore and so I think just a simple example it's easy for me to appreciate my pizza place might easily benefit from a new sushi place being available it brings more people to the platform it just sort of gets people are using it more intensively but it's just harder for my pizza place to benefit from a new pizza place being available and I think the results on how better firms benefit I think kind of the idea here is that we have two pizza places now the better pizza place is going to win out my question is not really a theory question is do we get enough kind of micro shocks like this so if you look at the distribution of how restaurants changes do we get enough kind of examples in the data where really a very direct editor for your kind of cuisine came on versus not something that we think it's really going to be that the network effects are going to dominate so this does require getting you into the kind of the micro structure but I wonder with the kind of amazing data we have here can we actually look and see what firms tend to co-occur and search results and kind of get this like okay this person had this choice set and therefore we're pretty sure that we can label these two types of firms as close competitors um you know the I think the kind of the story here is that demand for the platform is increasing in product variety and so if we went to that finer level of detail can we kind of explicitly characterize the effective variety on customer demand so you know you can imagine in a place that goes from four pizza places to five pizza places you know it just might not matter that much but going from you know Tibetan to one Tibetan might matter you know much more for the marginal new customer even the intensive margin from existing customers and I think Orrin might talk about this in paper so if I if I'm missing this my apologies but you know it would be nice at the level of the market to kind of show how much of an increase we got in variety however you can you know try to characterize that you know one thing I think is is specific to this context but I think it shares this context with a lot of platform goods is that um you know we have this bundles like no one wants to eat pizza every night no one wants to eat sushi every night and so there's there's a lot of kind of returns to having uh a lot of different variety and I think that in a nutshell is what kind of causes the cross-side externality effects to really dominate here but you know this isn't the case with something like you know ride-sharing where there's not that much differentiation and really the benefit is either going to come in terms of price or you know more probably in something like wait time um and then you know even you could go even finer than just you know whether how much variety matters and just kind of get down to is this like a within customer or between customer thing and so you know I think of like the Brinolson Bakos kind of world like you have you know is it you're each time you add a new restaurant you're picking up the customer that loves that kind of restaurant well probably not it's probably something more between customer preference for variety which I think would end up having some implications for disintermediation so you know if I only like one kind of restaurant I can sort of just go to them directly if I you know within my myself kind of vary in my taste for different cuisines over time the benefits to using a platform would just naturally be a lot a lot greater you know one effect in the paper that I thought was interesting you know Yelp is obviously an advertising platform and Orrin finds that the on platform advertising actually goes down and which kind of surprised me at first I was kind of thinking well you know you think you'd have more customers to get at and you have fiercer competition and so you want to grab more he proposes an interesting hypothesis though is that you know these better firms might I should say that the better firms are the ones that kind of have a decline and they're advertising their relative decline so the better firms are they writing up their marginal cost curve causing the returns to advertising to fall for them you know this kind of raises a natural question why not raise prices instead he doesn't find much on prices but he's got another paper I think that might explain this and there's other things he already talked about today that might explain this but then if it is a marginal cost thing you know does this effect sort of fade out over time you know do we see these restaurants kind of invest in more capital I mean I know he has this kind of reputation measure of improving quality you know maybe it's hard to get this kind of data but one idea I had is could you possibly explore this with transactional data on the firm component of wait time so with restaurants you know presumably if they kind of get really slammed with a big demand shock they kind of slow down and delivery times slow down perhaps if you see these firms that get this influx of new demand caused by the platform merger we see them actually you know have a higher throughput once they've had time to adjust so kind of this short run long run thing so I'm probably right at about five minutes so I you know mostly had a lot of questions but just to summarize I think this is a really fantastic paper I think it's going to be a really important contribution to the literature should I answer or should I wait for questions I don't know do you want to how should I do it maybe answer and then we'll collect questions so John thank you very much for these comments I really appreciate them I will answer some of these I'll answer all of them for some I'll have a bad answer for some so in terms of the more micro levels so there's kind of two things that I do in the paper to try to address it and again maybe there's more to do in this regard is first of all I in one specification I actually try to break up the market by food category and geography so that you know the St. Louis Pizza Market is a different market than the St. Louis Sushi market and then different market and so on the results are the same if I do it like this that the network that the positive networks are weaker and that makes sense because as you said people might go and look for pizza and sushi so there's positive also between treatment and control in this setting I will say one thing that I kind of glanced over today is that I look at the horizontal differentiation as well so if you're a pizza place as you said and the sushi place comes in and we actually gave the same example then it might be better for you because you're not competing for the same consumers but another argument would be well the people who are coming in because of the new sushi place are not pizza lovers so maybe you're not getting these people so theoretically it could be ambiguous I find again and maybe it's not in the version of the paper that's up I find it's actually on average if you get a different type of firms joining the effect is more positive on the incumbents one thing that you mentioned that I don't do is look at how like the increased variety attracts more people so if I add more pizza places versus if I add more sushi place that I didn't have how does that affect the network effect I didn't look at that and I think I can and I should in terms of the importance of variety so actually you gave the example of Uber and usually what I present I actually say that you need variety on Uber because you need the driver to be close to you so variety is not going to be whether you know which car to drive variety is going to be where they are if I'm in a city that doesn't have a lot of Uber drivers I'm going to use Lyft because I want I don't want to wait that long so here the value is going to be not waiting that much rather than so the match value is not if the food if it's food that I like but it's wherever they're close to me geographically I didn't look at within consumer between consumer I should do that I will say consumer data is not great it's very noisy but I can definitely do more to look at that I do have that data other measures of quality then again it's on my to-do list I should do it it's just you know it's a big project to do so I keep procrastinating but maybe I should prioritize that I think one of the downsides of knowing everything or having all the access is that you kind of unshackle your referees to ask you for everything under the sun because they know how much you have and I think that's the main the main point to your address but thanks there is a question from Eva I don't know if she wants to ask it sure hi I was wondering whether you can think about what's kind of behind these differential effects on low and high quality firms that could that be also driven by the fact that the entering firms are somehow low quality I mean this seems to be a little bit the case at least in Europe in some places that the firms that sell only via this delivery services are actually only take away places which have for example bad hygiene ratings and so on so I wonder if that could be behind that as well and that would also explain maybe this investment in quality result because then you would see that the firms that were before already above mean quality they actually gain in their ratings just because the incoming firms are very low quality so I don't know if you can check that somehow as well so that's a good question I will say a couple of things first of all just on average the restaurants that are entering the platform look very similar to the firms that are already on the platform in terms of rating I mean they think that like marginally lower it's like a 0.0 something less in average rating but on average they look kind of the same one way that I try to control for it because maybe some markets get more high quality enterance in some places do not is that when I code wherever you are high or low I also include the new businesses so this is where you are in the new or in the future distribution of ratings so that's kind of a way to try to control for difference between markets I did in trying to measure the network effects and I actually don't think it ended up in the paper because I just I thought it was pretty straightforward I tried to test whether if there are better businesses joining the platform that does that draw in more people compared to when bad firms joining the platform conditional the number of firms that joined there is of course a weak positive relation there so good firms on average weekly bring me weekly bring in more consumers okay thanks thanks Oslem I think you have a question yes thank you so I was wondering maybe Yelp also has some effect on by giving prominence differently to different type of firms after this partnership contract so maybe it's not about the entry exogenously happening but it's just the Yelp ranking of those firms are changing after the partnership and I was I also can expect Yelp to not to put so much down in the list those high quality ones that are already high ranked so maybe the ones that are really suffering from this lack of prominence are the incumbents that were not high status or ranked before the partnership have you thought about this or is there any way that you can see anything in the data yep so that's a good question I think the intuition you described goes in the other way if you say that they are they want to bring up the lower ones so that wouldn't make the I was thinking about you know if I'm Yelp and I want to change the prominence of course I don't want to put the really high ranked or high quality ones incumbents down in the list but those ones that are already low quality might be more suffering from this so it's the ordering of the search yes so I do look at that in the paper so admittedly actually when I you know when I got this result I thought this would be like one of the main things that's going on it turns out it doesn't seem to be really the case so first of all if you the correlation between the your star rating and where you are on the search result is actually it's negative right because high quality business as you mentioned are higher on the list or lower in terms of rank but it's only minus point one so it's a pretty weak correlation and because there's a lot of other things going on when you search for results when you search for a restaurant I did what I did try to do is I see your search rating so I try to control for that and see if that makes if that substantially changes the results right so it's all coming from just where you are rated and intuition is that when you add more firms the good ones they stay up but the bad ones they're getting pushed down so you just don't see them then that would mean that if I if I control for the research results this should go away it doesn't seem to be the case so I agree this was my intuition as well but that does not seem to be at least in my setting the main thing the driving this results does this answer your question yeah I mean one thing is thinking about how people are doing the search and you know thinking adding firms changing the way that they are ranking the search the other one is like I was really thinking about this objective ranking versus subjective ranking you know Yelp giving prominence to those firms that are more like having partnership with Grubhub now because of the fact that maybe there's some commission coming additional from this relationship so let me clarify so first of all I worked at Yelp I never had access to their search algorithm because this is like super proprietary but I will say that to my knowledge it hasn't changed significantly over time in particular it didn't change between treated areas and control areas right so everything that I do is different so I always have treatment versus control so if there are any changes to how they rank businesses they should be the same in both of these markets the last thing that I want to say just it's a correction I think it's important the businesses that I look at in the main specification are not the Grubhub businesses right so the Grubhub is not going to be they are not included in the main analysis I only look at the businesses that were already on the platform so just to kind of I think this is important okay thanks thanks so we're exactly at one hour so at this point I will stop the recording