 The title of the paper I'm going to present is a larger analysis in the upper economy, an empirical model of other sponsored media. So in the last few years, we have seen many important antitrust and regulation issues in the upper economy. And one of the biggest would be the recent case between Epic Games and Apple. And there are many reasons why the analysis in the upper economy is not straightforward. But I think one of the major factor is that multiple business model coexist in this economy. Some apps are paid, but some apps are for free, but they are other sponsored. And sometimes an app combines these two monetization policies. And this poses some difficulty in traditional empirical antitrust analysis that are primarily based on prices and price variations. And this is creating some room for misguided policies. To this end, in this paper, we would like to discuss how to deal with this situation. Specifically, we would like to ask, how we can still estimate relevant demand and supply parameters such an impact competition at the sponsored media, where multiple monetization policies can coexist. And based on an estimated model, how we can conduct the usual measure analysis, including relevant market definition and margin simulation. And also we would like to discuss how we can analyze relevant issues such as the effect of changing transaction fee imposed by the marketplace on welfare. In particular, we are going to develop an empirical model at the sponsored media. In this model, we consider consumers who face both budget and time constraint. And we consider a specific framework in which output developers are regarded as competing in utility through price and advertisement setting. In this setting, we can introduce well-defined notion of consumer's cost for using an app and this will facilitate following margin analysis. After establishing this empirical model, we introduce an estimator based on available data and then apply this model to the data about Google Play in Japan. But potentially this can be applied to the data outside Japan. Finally, using this estimated model, we would like to conduct an SSNIC test, an extended version of SNIC test, run margin simulation and study the effect of transaction fee. We find from our estimate that the disability of watching an app for a consumer is approximately five to 6% of the app's advertising revenue. And I would say this number is slightly lower than the number that is found in the literature. Also, we find that game apps are more segmented by categories than categories other than games. And this has some implication for the market definition and model simulation. We conduct an extended version of SNIC test to define a relevant market and in particular we try whether categories in Google Play can constitute an antitrust relevant market. And we find that some game categories such as action game or puzzle game and role playing games can be considered to form an antitrust relevant market. But none of non-game categories will form an antitrust relevant market. And we conduct corresponding model simulation and find that the implication of model simulation is mostly aligned with the simpler analysis based on SNIC test. In particular, only the modules within a category that is judged as relevant market by SNIC test has large impact on consumer welfare. Validating the use of SNIC test as a convenient screening device for, but modules. And finally, we conduct a policy experiment of changing transaction fees and see how this affects price and advertisement and welfare. And we find different from my prior belief reduction in transaction fee can increase price rather than decrease the price and reduce up. And this is especially true for non-game apps. I would like to discuss this mechanism well when we introduce this result. This is a brief introduction. So if you have any clarification question at this point, I can stop and hear your question. Okay, so far no clarification question. Yeah, okay, then I move on. So first I would like to describe our model as a sponsored media competition. We consider a weekly static market indexed by D. And in each market, there is a set of apps indexed by J. And there is a set of app developers indexed by D. An app developer can own multiple apps. And there is a massive consumers. A developer of app J sets its download price FJ. You can consider that this includes both subscription fee and in-app purchase. We summarize all of them as a download price. And it also sets the advertising intensity, A.J. Advertising intensity is the number of advertisement that a consumer has to watch to use the app for a unit of time. Consumer I observe this menu proposed by app developers and downloads at most one app and choose the usage time QJ of download it up. In this setting, we consider a static pure strategy in Azure Kiribati. In this setting, we have unique Kiribati as long as the random coefficients standard deviation is small enough. And if the standard deviation is too high, we cannot say anything about the uniqueness of the Kiribati. Okay, first I would like to describe a consumer's problem. Consumers problem is formulated by a discrete continuous choice model. The indirect utility from downloading app J consists of the usage-related surplus is J and download-related surplus. And this download-related surplus is modeled as usual as usual in the discrete choice model. Xj is an observed characteristics of app J and we have random coefficient on these characteristics. Fj is the price and alpha y is the disability of losing a unit of money. And Qcdj is the app-specific download-related and observed heterogeneity. That can be endogenous with Fj, that can be correlated with Fj. And if your Ij is an idiosyncratic preference shock. Here, one assumption is that we do not have a random coefficient on the price for the reason I explained earlier. The usage surplus is J is evaluated at the optimal usage time Qj. And this function, Vj is written like this and this first term is the marginal utility of usage. And the marginal utility of usage is determined by the observed characteristics and this is the important part. This aj is the number of us that a consumer have to watch. When the consumer uses this app for a unit of time. And this alpha a is the disability of watching a unit of advertisement. So if you divide alpha a by alpha y, you get the disability in terms of Japanese year. And w is a wage and captures the opportunity cost and this coefficient is the same with this coefficient on the price. And the last term minus yator times Qj square captures effect of satiation. And the couple are multiplied this surplus to roughly capture the expected number of weeks that a consumer will use the app when the consumer decided to download. Because we can have a free dynamic model, we approximate dynamic nature by this parameter. So this is a compromise. In this specification, we can derive analytically the usage time Qj and derive the download choose probability for each consumer and by simulating the random coefficient we can evaluate this share of an app. Oh, I missed to mention another important assumption. On top of this alpha y, we do not allow for random coefficient for the usage related utility. Okay, I will discuss this later. Next, I would like to describe an app developer's problems. An app developer can have multiple apps, but the profit per app Qj is written in this way. This is the download share and this is a profit per download. The profit per download consists of the revenue from consumers and revenue from advertisers. Fj is the price that the app charges to the consumer and low is the transaction fee imposed by Google. During this period, this low was 0.3. aj is, as I said, the advertising intensity and Qj is the usage time. So Qj times aj is a total number of advertisement that the app showed on this platform. And R is an advertisement price. This is a revenue per unit of advertisement for the app developer. Lambda is the marginal cost of changing the usage level. And epsilon j is an idiosyncratic download-related marginal cost shock. The total profit of app developer D is the sum of per app profit. Each developer chooses aj and fj simultaneously of the owned apps to maximize the total profit. In this framework, oh, sorry, this is subject to non-negativity constraints on aj and fj. In this framework, the free apps and other free apps can be simply captured by a corner solution with respect to these non-negativity constraints. So here in this framework, we can transform this model about price and advertisement into an equivalent model of competition in utility. First, in our framework, the mean utility is sufficient statistics of price and advertisement for consumers. This is due to the assumption of no random coefficient in price and the usage-related utility. Because of this, the interaction with consumers and interaction competitors are only through the choice of data. Therefore, we can calculate the optimal combination of price and advertisement for a given data. Then we can consider the payoff as the profit as a function of the data where price and advertisement are evaluated at this optimal combination for this given data. So we can transform this model into gain where the payoff is profit evaluated at the optimal combination of price and advertisement for a given data and the action is to choose data. If we accept this framework, the nice aspect is that we can define the cost for a consumer to use an app. The cost for a consumer to use an app is defined by the difference in the mean utility achieved by their price and us. And the actual mean utility under the optimal combination of price and advertisement that achieves this mean utility data chain. If there is no advertisement and if there is only a price competition, this cost reduces to the price times the disability of our supply. Therefore, the notion of cost generalizes the notion of price. And we can use this notion of cost to define a market, even if there is a free product in the market. So this is a basic framework that we are going to use. At this point, I would like to discuss a few limitations and assumptions in our model. First, there was a marginal cost of changing the usage time for an app developer. However, we didn't have any direct marginal cost of changing advertising intensity. This is one important assumption that would be necessary for identifying some key parameters. Second, we assumed competitive advertisement market. So no publisher, no developer has a market power in the advertisement market. This is due to the limitation in available data set. So we can discuss a large app developers, such as Facebook, who should have some market power in the advertisement market in the current paper. To overcome this issue, we have to get richer data set, including the app specific advertisement price and advertisement quantity. At this point, we don't have this data set. So we had to assume this nature. Third, as I already discussed, no consumer heterogeneity is allowed for the coefficient of the price and in the usage surplus. And finally, we stick to a stick framework. So we cannot discuss entry, innovation or potential customer base accumulation or something like this. So this introduces the model. And again, I would like to ask if there is any question for qualification or for other types of questions. Looks so far so good. Okay. Now I would like to introduce the data we use to estimate this model. So we focus on Google Play and we do not look at the data from Apple Store because we couldn't get all the relevant data set for the Apple Store. We selected measure apps to represent the share of business models, which includes free and unsponsored app and paid and unfree apps and paid and also unsponsored apps. We check this share in the original data set. And then by keeping this share, we focused on the top apps who regularly showed up in the ranking of download and usage. The period of the data covers from March 2015 to January 2017. But estimation is based on some random sampling from this period. Here is a summary statistics at the week up level. One thing that we can see is that apps that belongs to non-game category tend to have higher number of download. And an app that belongs to a game category tend to have a higher usage time. Here, I'm showing the shares of business model for each product category in non-game categories. So one fact that we can see here is that there are the majority of non-game apps are for free and unsponsored. However, if you include download price and subscription fee and in-app purchases, there is a non-negligible share of apps that are charging price to consumers. This is the second fact. The third fact is that half of paid apps both paid apps and unsponsored apps. Okay. And this share has implication for the effect of changing transaction fee. So I want you to remember this number. In case of game apps, which I didn't include in this slide, there are more paid apps. And when they are paid apps, there is few paid apps that actually are sponsored as well. If they are paid apps, they are usually out of free. This is the difference between game apps and non-game apps. To construct app-specific observed characteristics, we scraped the description of apps from Google Play. This is a description of Dropbox in Japanese. And then based on that, we constructed some characteristics such as the number of words and the something like this that are used in the literature. In particular, used by a paper, by ghosts and harm. In addition to these traditional variables, we also constructed a word vector. In particular, we converted each word in an app description into a 300-dimensional word vector using the national language web corpus of Japanese. Then for each app, we took the average of the world vectors weighted by their uniqueness and then regarded it as a numerical vector that somehow captures the nature of the description of the app. And then we use this as part of the observed characteristics of app, XJ, and we augmented the data with the market level data, which includes the advertisement price and hourly wage and the market size. In this analysis, we define the market size as the number of active devices. Time is some constant to endure that no total market share exceeds them. Today, I skip the technical detail of the estimation. Essentially, we invert the equilibrium condition for an observed fixed effect XJ and then use this moment condition for those unobserved fixed effect for the demand side. And then for the supply side, we use pricing optimality condition and so on to construct a moment. Rather here, I would like to focus on two fundamental identification issues. The first issue is this. Even though the advertising intensity, AJ, plays an important role, we do not observe this quantity in the data. What we can observe in the data is just whether advertisement is shown or not. So in the literature, like a paper by Goose and Ham, they included this indicator variable into the demand function. However, we would like to estimate, we would like to use this AJ. So what we do? Usually, we identify marginal cost from observed price and price optimality condition. However, in this paper, we elicit this equilibrium advertisement intensity AJ from the advertising optimality condition, assuming that the direct marginal cost of changing advertisement is zero. We believe this assumption is justifiable in the up economy because of app technology. Once app developers introduce app technology and pages fixed cost, after that, adjusting advertisement intensity is almost automatic. This is different from traditional media such as newspaper and TVs. That uses a lot of resource to acquire you and retain sponsors. So in summary, by assuming that the direct marginal cost of changing advertisement is zero, we elicit implied equilibrium advertising intensity, which is valued by advertisement price in the data, and then this is used to identify the disutility of advertisement. This is the first fundamental identification issue. Second identification issue is this. When there is a free app, the first-order condition is only inequality. So the price optimality condition cannot point identify the idiosyncratic marginal cost shock free apps. In other words, we don't know how marginally or deeply free free apps are. So for making prediction, some extrapolation is necessary. In the current version of the paper, we try to identify the distribution of the cost of free apps by assuming that free paid version of a pair of sibling apps has the same marginal cost. This may introduce some bias, but I think this is the most possible way to overcome this issue. So I would like to clarify these two identification issues. So this is a summation result, part of the summation result. And one takeaway is that the disutility of watching a unit of advertisement is roughly 27 yen. This is like a 27 cent US dollar in US dollar. And this is slightly smaller than the number known in the literature. Based on this estimated model, we would like to conduct market definition and margin simulation and then move on to the analysis of transaction fee. Here we extend the standard SNP test to SNP test by changing the price to the consumer's cost. SNP test considers how the profit are hypothetical monopolist. That owns a set of apps changes if the price is forced to increase by 5,000. The set of apps are regarded to form the market if the profit increases as a result of this increase in the price. Because we cannot use a SNP test for free apps, we use SNP test for this purpose. I skip this part because the time is running out. So here what I did was like this. We consider a hypothetical monopolist who owns all the apps in our sample belonging to particular category like comics. Then we force them to increase the consumer's cost by 5%, and then we check whether this result in the increase of profit. And what we found here is that if you focus on non-game categories, none of the profit will increase as a result of forced increase in the price by 5%. That means none of these categories will form a relevant market. In other words, they are competing for consumers' time across categories. On the other hand, if you focus on game categories, we find some categories where the profit can increase as a result of forced increase in the price. This means that consumers are segmented by these categories and they do not go away even if the price increases. So we can judge that categories such as action, casino, casual, puzzle, and so on will constitute an antitrust relevant market. We can check this result based on full merger simulation as well. Similarly, we consider a hypothetical merger who owns all the apps in our sample that belongs to a particular category. And then we run a merger simulation. This time, we can incorporate the strategic response of the competitors and optimal pricing by the hypothetical monopolist. In this case, we check whether the equilibrium average consumers' cost increases more than 5% or not. And what we found was the result is aligned with sneak test. And the damage to consumer surplus is only pronounced in the categories where sneak test found that these categories are relevant market. So this will validate the use of sneak test as a convenient screening device for potentially anti-competitive modules. Finally, we would like to discuss this issue. What happens if Google Play reduces the transaction fee? If we only consider a standard market, decrease in the proportional transaction fee will decrease the price. However, in the current market, this can increase the price. First, in case of free other sponsored medium, if the transaction fee on the price is reduced, they may find it more profitable to change the business model from free other sponsored medium to paid up. Then the price will increase. Second, in case of up, that charges price and show advertisement to consumer. Special feature of proportional fee can increase the price. Specifically speaking, and proportional fee inflates the marginal cost relative to the price. Therefore, if the effective cost of download is negative due to the advertisement revenue, reduction in the transaction fee can increase the price. And this happens for non-game apps that charges prices and also show advertisement to consumers. So theoretically, the impact of transaction fee on the prices is ambiguous. And according to our estimation estimated model, this actually increases the price of apps for non-game apps. If the transaction fee is reduced from 30% to a smaller number, the transaction fee of non-game apps can increase the amount. And we have similar result for games, but the impact is much more smaller because the aforementioned channels work more for non-game apps where there are more free apps and there are more paid and other sponsored apps. And this result in welfare implication. In case of non-game apps, it is possible that the reduction in the transaction fee can only benefit app developer but can damage the consumer welfare and total surplus because the price can increase. This depends on the relative disunity of watching advertisement compared to paying a money or a consumer. Finally, in case of game apps, impact on consumer surplus and total surplus is almost negligible and they look changing but the magnitude is very small. And this is within a numerical error. So this can be almost dismissed. So in case of game apps where the aforementioned channel do not work, the reduction in the transaction fee is regarded as a pure transfer between app developers and Google Play. So we observe different implications of changing the transaction fee for games and non-game apps. Okay. So this is all I have today. In summary, our model allows for coexistence business models and enables the usual margin analysis including market definition and margin simulation even if there is a free app. Even if multiple monetization policies coexist. Then we showed that some categories of game apps form a relevant market whereas non-of-none game categories form relevant market. And we showed that margin simulation result is mostly aligned with sneak test. So we may be able to use sneak test as a convenient screen device. And finally, we pointed out a theoretical possibility that a reduction in transaction fee can increase the price. And demonstrated that this may be a real threat especially for non-game apps in the Google Play in Japan.