Download Impact of Platform Mergers on User Behavior: A Case Study of Rover and DogVacay and more Schemes and Mind Maps Literature in PDF only on Docsity!
Submitted to Management Science
manuscript
Authors are encouraged to submit new papers to INFORMS journals by means of a style file template, which includes the journal title. However, use of a template does not certify that the paper has been accepted for publication in the named journal. INFORMS journal templates are for the exclusive purpose of submitting to an IN- FORMS journal and should not be used to distribute the papers in print or online or to submit the papers to another publication.
Dog Eat Dog:
Balancing Network Effects and Differentiation in a
Digital Platform Merger
(Authors’ names blinded for peer review)
Network effects are often used to justify platform strategies such as acquisitions and subsidies that aggregate users to a single dominant platform. However, when users have heterogeneous preferences, a single platform may be worse than multiple platforms, both from a strategic and antitrust perspective. We study the role of network effects and platform differentiation in the context of the merger between the two largest platforms for pet-sitting services. To obtain causal estimates of network effects, we leverage geographic variation in pre- merger market shares and a difference-in-differences approach. We find that users of the acquiring platform benefit from the merger because of network effects, but users of the acquired platform are hurt because their preferred option is removed. Network effects and differentiation offset each other such that at the market level, users are not substantially better off with a combined platform rather than two separate platforms. Our results have strategic and regulatory implications, and highlight the importance of platform differentiation even in the presence of network effects. Key words : mergers and acquisitions, two-sided platforms, peer-to-peer markets, network effects, platform growth, antitrust
- Introduction
Companies face many strategic choices when pursuing growth, including how to innovate and at- tract new customers, whether to acquire competitors, and if so, how to integrate their processes into the merged company. These strategic choices become more complex with network effects, which occur when the value per user from a product or service increases with the number of other users (Katz and Shapiro (1985)). Network effects are often considered a defining characteristic of plat- forms (Rochet and Tirole (2003)) and a main driver of their growth (Dub´e et al. (2010)). Network effects are also used to justify first-mover advantages or equilibria where a single winner eventu- ally dominates the market (Lieberman and Montgomery (1988)). In practice however, the mere
1
2 Article submitted toAuthors’ names blinded for peer review Management Science; manuscript no.
presence of network effects is not enough to draw these conclusions, because other countervailing forces may push in opposite directions. We consider the role of platform differentiation in countervailing network effects in the context of digital transaction platforms (Cusumano et al. (2019), platforms henceforth). These platforms help buyers and sellers find each other and safely transact. Examples include Airbnb, Amazon Marketplace, and Uber. Sometimes, platforms are designed to cater to subsets of users with specific preferences, for example by emphasizing original and unique items (e.g., Etsy) versus delivery speed and convenience (e.g., Amazon). Even platforms that offer very similar services can attract different types of users due to subtle differences in design (Jia et al. (2021)). When platforms are differentiated and network effects are not too large, strategies designed to drive all users on a single platform may not be justified, for platform managers and regulators alike (Farrell and Shapiro (2000)). We study the relative importance of network effects and platform differentiation in a merger between two platforms competing in the local services industry, in which the largest platform acquired and then shut down its largest competitor. We find that buyers on the acquiring platform engaged in more transactions after users of the acquired platform joined. This first result confirms that some buyers benefited from the merger because of network effects. However, many buyers on the acquired platform left by choosing not to switch to the acquiring platform. This second result suggests that some buyers were hurt by the removal of their preferred platform. Network effect benefits and the loss of platform differentiation offset each other so that on average buyers are not significantly better off with a single platform compared to two competitors. Even though our focus is on buyer outcomes, we find similar results for platform and seller revenues, which remain constant after the merger compared to the sum of revenues from the two competing platforms before the merger. Our findings on the role of network effects and platform differentiation provide insights on two key decisions that managers face when considering acquisitions. First, the net result that users are not better off with one compared to two platforms makes it more difficult for regulators and managers to justify platform acquisitions solely on the basis of network effects. Second, even if an acquisition is approved, it may be beneficial for a company to operate multiple platforms rather than merge them. More generally, our findings call into question growth strategies based on first- mover advantage and winner-take-all equilibria. Measuring network effects on platforms is generally difficult since changes in the number of users are typically endogenous. For our identification strategy, we use the sudden increase in the number of buyers and sellers induced by a platform merger. In March 2017, Rover, the largest US platform for pet-sitting services, acquired DogVacay, their closest and largest competitor. A single platform
4 Article submitted toAuthors’ names blinded for peer review Management Science; manuscript no.
To evaluate whether network effects are large enough to justify a single platform, we study the effects of the merger on the market, aggregating data from both platforms.^1 If network effects were large enough, combining the two platforms would lead to larger user benefits in geographies where both platforms were equally large before the merger compared to geographies where one platform was already dominant. This is because in split geographies, the merger effectively doubles the number of users who can interact. We use a difference-in-differences strategy to measure the effects of the merger, comparing out- comes before and after the acquisition, and across zip codes with different market shares. We explicitly address selection into market shares with matching. We find that after the merger, ex- isting Rover buyers increased platform usage more in geographies where Rover received a bigger influx of users from DogVacay. Existing DogVacay buyers similarly benefited from network effects, but, relative to existing Rover buyers, they decreased their platform usage after the merger. Many of these buyers chose not to switch to Rover, and those who switched transacted less frequently and matched at lower rates than comparable Rover buyers. We find support for two related mech- anisms that partially explain these effects: a coordination failure and disintermediation, whereby DogVacay buyers have a harder time finding their previous providers on Rover and may be led to transact with those same providers off the platform. Attrition by Dogvacay buyers almost perfectly offsets the increased usage of Rover buyers so that at the market level, we find no evidence that the combined platform substantially improves market outcomes compared to the sum of the two separate platforms: not on the extensive margins such as user adoption, retention or total transactions, nor on the intensive margins, such as match rates or ratings. Although we predominantly focus on buyer outcomes, we confirm that our results are not simply due to a redistribution of value across buyers, sellers, and the platform. Our results imply that even if network effects are strong in online platforms, preference het- erogeneity can offset the benefits of a single platform compared to multiple competing platforms, even when competitors appear to be close substitutes. This result is true across different types of geographies: geographies with a small versus large baseline number of users, and geographies where users have lower versus higher propensity to multi-home. The rest of the paper is structured as follows. In Section 2, we present the relevant literature. Section 3 presents a stylized model motivating our empirical analysis. Section 4 describes the con- text and relevant data while Section 5 presents our empirical specification. Results are in Section 6. In Section 7, we conclude by discussing implications for platform strategy and antitrust regulation.
(^1) When we say “we study the effects of the merger on the market,” we can actually only measure the effects on buyers who used one of the two platforms for which we have data, which represent the vast majority of online pet-sitting. Our assumptions on the value of the outside option (Section 3) imply that the value enjoyed by consumers who choose the outside option after the merger is either constant or lower compared to before.
Authors’ names blinded for peer review Article submitted to Management Science; manuscript no. 5
- Literature Review In this section, we present the mostly theoretical literature on platforms and network effects, and describe how the setting in this paper is ideal for studying network effects empirically. Early theoretical work focuses on competition and product compatibility in the presence of net- work externalities (Katz and Shapiro (1985) and Farrell and Saloner (1985)), but the pioneering models of multi-sided platforms came with Rochet and Tirole (2003), Caillaud and Jullien (2003), Parker and Van Alstyne (2005), and Armstrong (2006). In their models, platform businesses are characterized by multiple user groups and the presence of positive cross-side network effects, where each user benefits from having more users in other groups. The early papers focused on platform pricing strategies (Weyl (2010)). Other strategic choices, such as entry, vertical integration, and degree of openness are the focus of Hagiu and Wright (2014), Suarez et al. (2015), Zhu and Ian- siti (2012), Eisenmann et al. (2011), and Boudreau (2010), among others. More recently, Bakos and Halaburda (2019), Jeitschko and Tremblay (2020), and Park et al. (2021) explore how plat- form strategies change as a function of multi-homing, i.e., the propensity of users to join multiple platforms. In the theoretical literature on platforms, the presence of network effects has led to several strate- gic implications. Platforms entering first have an advantage (Lieberman and Montgomery (1988)), markets with multiple competitors tend to tip towards a single platform (Dub´e et al. (2010)), and that single platform will eventually control the entire market (Cennamo and Santalo (2013)). In such cases when a dominant platform emerges, Tan and Zhou (2020) and Nikzad (2020) predict that the interaction of network effects, product variety, and pricing power lead to theoretically am- biguous effects of platform dominance on consumer surplus. Argenziano (2008) even theorizes that the competitive outcome is inefficient when platforms are differentiated. Our work adds empirical evidence to this literature by emphasizing the importance of platform variety in counterbalanc- ing network effect benefits. Our insights challenge unconditional tipping by estimating network effects that are too weak to naturally lead to winner-take-all equilibria. We also provide some unique empirical evidence on the extent of multi-homing, finding that albeit limited, multi-homing is predominantly concentrated on the supply side among the largest sellers. The empirical literature on network effects dates back to Greenstein (1993), Gandal (1994), Saloner and Shepard (1995), and more recently Gowrisankaran and Stavins (2004) and Tucker (2008), who show early evidence that network effects are present in the adoption of a broad range of technologies, from banks’ ATMs to video-messaging software. One of the first to empirically study and find evidence of positive cross-side network externalities is Rysman (2004) in the market for Yellow Pages, while Chu and Manchanda (2016) find similar evidence on e-commerce platforms.
Authors’ names blinded for peer review Article submitted to Management Science; manuscript no. 7
- Theoretical Framework This section presents a model that highlights the key trade-off between network effects and platform differentiation. The model gives us expressions for buyers’ utilities before and after a merger of two competitors to guide our empirical analysis. Our model, like our later analysis, focuses on buyers, implicitly assuming away any redistribution of merger gains between buyers, sellers, and the platform. In practice, this means that if buyers captured, say, 20%, of value before the merger, they still capture 20% after the merger.^2 This simplification, which is supported by the data, makes the model more tractable and intuitive. Our model also does not capture the separate effect of increasing the number of buyers versus sellers. In a two-sided platform, doubling buyers hurts each individual buyer due to a crowding out effect while doubling sellers benefits them because of cross-side network effects. However, the combination of cross-side network effects from each user group to the other implies that doubling both buyers and sellers should benefit each individual buyer. It is this combination of cross-side network effects that we focus on, so we assume that the number of buyers relative to sellers is fixed, equal to 1 for simplicity, so that doubling the number of users means increasing the number of buyers and sellers at the same rate. In our model, there are two platforms — platform α, the acquiring platform, and platform β, the acquired platform — and a unit-mass of buyers that are located on a Hotelling line. Platform α is located at 0 while platform β is located at 1. Each buyer also has a value for the outside option. Buyer types are identified by their location on the Hotelling line, di ∼ U (0, 1), and their value for the outside option, i ∼ U (− 1 , 1). A buyer i located at point di on the Hotelling line has utility for platform α equal to uiα(nα) = v(nα) − di, where nα is the mass of buyers using platform α. Horizontal preferences are given by the parameter di. Network effects exist whenever v() is increasing in its argument. We assume that v() is not too small nor too large so that the share of buyers located at di who choose the outside option is strictly between 0 and 1 along the entire Hotelling line. We have two periods, the pre-merger period in which both platforms α and β are available but each user is only aware of one of them, and the post-merger period in which only platform α is available and everyone is aware of platform α. Buyers do not expect the merger to occur. Pre- merger, when both platforms are available, we assume that advertising and customer acquisition efforts effectively split buyers in two groups, each of which is only aware of a single platform. We posit that there is an exogenous cutoff, k 1 , such that buyers to the left of the cutoff (di ≤ k 1 ) consider only platform α and the outside option, while to the right of k 1 buyers only consider
(^2) This assumption ignores cost efficiencies that the platform may enjoy as a result of the merger or changes to sellers’ service costs.
8 Article submitted toAuthors’ names blinded for peer review Management Science; manuscript no.
platform β and the outside option. Buyers have rational expectations over the equilibrium number of buyers choosing the various options. They select the option of which they are aware that gives them the highest utility given their type (di, i). In particular, buyer i for whom di ≤ k 1 joins platform α if and only if uiα(nα) ≥ i. Similarly, buyer i for whom di > k 1 joins platform β if and only if uiβ (nβ ) ≥ i.
Figure 1 Buyer Types
A B
C D
k
k
k
k k
−1.
−0.
0.00 0.25 0.50 0.75 1. Horizontal Preference (d)
Value of Outside Option (
ε)
This figure divides the space of buyer types according to an exogenous cutoff, k 1 , and their optimal choices conditional on that cutoff. A denotes buyers who choose platform α both before and after the merger. B denotes buyers who switch from platform β to α. C denotes buyers who switch from the outside option to platform α. D denotes buyers who switch from platform β to the outside option.
Buyer choices result in two indifference conditions, depicted in Figure 1. The first condition is the point along the vertical axis, k 2 , where buyers are indifferent between the outside option and platform α: v(nα) = k 2. Similarly, the second condition determines the point of indifference, k 3 , between platform β and the outside option: v(nβ ) = k 3. The two indifference conditions and the exogenous cutoff, k 1 , allow us to find an equilibrium in (k 1 , k 2 , k 3 ). The two market shares nα and nβ can be derived from k 1 , k 2 , and k 3. They are graphically depicted as the A area (for nα) and the B+D area (for nβ ) in Figure 1. Note that this model could in principle have multiple equilibria, although for our purposes equilibrium selection is not important. At the realized equilibrium, the average per-buyer utility on platform α is equal to:
u¯α = v(nα) −
∫ (^) k 1 0
dig(di)∂di, (1)
10 Article submitted toAuthors’ names blinded for peer review Management Science; manuscript no.
Hypothesis 1: The benefits of the merger to existing buyers on platform α is decreasing in nα (or, equivalently, increasing in nβ ).
The hypothesis states that with network effects, the increase in average value for existing buyers on platform α is bigger in geographies where platform α was smaller before the merger. To evaluate the role of horizontal preferences, we compare the post- and pre-merger utility of buyers who switch from platform β to platform α (switchers). The change in utility is equal to
[v(n∗) − v(nβ )] −
[∫ 1
k 1
(di + k 1 − 1)f (di)∂di
]
where f (di) is the distribution of switchers along the Hotelling line (area B in Figure 1). Switchers benefit from network effects because n∗^ > nβ , but are also on average farther from their platform of choice. Our model does not yield a sharp prediction about how utility changes after the merger for switchers. Depending on whether network effects dominate over horizontal preferences, switchers may be better or worse off after the merger. We note, however, that there is a close relationship between the gains of buyers from platform α and platform β. In particular, suppose we compare the benefits to platform α’s buyers from a merger in a geography where platform α had ¯n buyers pre-merger and the benefits to platform β’s buyers in a geography where platform β had ¯n buyers. The benefits to platform α’s buyers from the merger are greater than the benefits to platform β’s buyers from the symmetric merger and the difference is solely due to the role of platform differentiation. That is because in both Equation (3) and Equation (4), the network effect benefits are v(n∗) − v(¯n), while the reduction in platform differentiation only hurts switchers (the integral in Equation (4)). If users have horizontal preferences over different platforms, the difference between Equation (3) and Equation (4) when nα = nβ = ¯n is negative. This is another testable implication.
Hypothesis 2: Consider two geographies, one where platform α has ¯n number of buyers and the other where platform β has ¯n buyers. If buyers have horizontal preferences over platforms, switchers in the second geography benefit less from the merger than stayers in the first geography.
Are network effects large enough for a single platform to create more value for buyers than two separate platforms? For this to be true, network effects need to dominate over horizontal preferences. We have already argued that stayers should benefit and switchers may or may not benefit. Joiners (area C) are definitely better off by switching to the now larger platform α from
Authors’ names blinded for peer review Article submitted to Management Science; manuscript no. 11
Figure 2 Change in Buyer Utility Platform Gains from Network Effects
Platform Losses from Differentiation
Platform Gains from Merger
Market Surplus from Merger
−0.
0.00 0.25 (^) Pre−Merger Platform0.50 (^) α Share0.75 1.
Merger Effects on Combined Platform Utility
The figure plots the change in aggregate utility experienced by platform buyers after the merger as a function of platform α’s pre-merger market share. Market share is computed as (^) n 1 n+^1 n 2. The solid line represents the total gains by the platform from the merger. The top line represents the platform’s benefits of the merger due to network effects. The bottom line represents the platform’s costs from the loss of platform differentiation. The dot-dash line represents the total change in utility for all buyers, which includes the outside option.
the outside option. Leavers (area D in Figure 1) are definitely worse off by switching to the outside option, which was already available pre-merger. Instead of providing the algebraically-complicated equations determining the change in buyer values, we provide graphical intuitions from Figure 1. Platform managers care about how the merger affects their users, regardless of the alternative choices those users have at their disposal. This implies that platform managers care about the change in utility of buyers in areas A and B, to which they add the post-merger utility for buyers in C and subtract the pre-merger utility for buyers in D. This comparison is displayed in Figure 2, which plots the change in aggregate utility created by the platform as a solid line. The figure also separates the net change into its two components: the gains from network effects (dashed) and the losses from the removal of platform β (dotted). To more closely map the model to our empirical strategy, we plot the change in buyer utility as a function of market shares that we can compute in our data, (^) nαn+αnβ. Network effect gains are maximized in geographies where platform α’s pre-merger market share is 0.5. Similarly, the losses from platform differentiation are largest at the same point. If network effects dominate, as in Figure 2, the benefits from the merger are maximized in more competitive geographies, where the two platforms have similar market shares. This is our last testable hypothesis.
Authors’ names blinded for peer review Article submitted to Management Science; manuscript no. 13
Figure 3 Rover’s and DogVacay’s Landing Pages
(a) Rover.com, March 2017. (b) Dogvacay.com, March 2017.
The figures show the landing page of Rover and DogVacay before the acquisition. The screenshots are accessible on Wayback Machine (https: // web. archive. org/ web/ 20170307101746/ https: // www. rover. com/ and https: // web. archive. org/ web/ 20170228165616/ https: // dogvacay. com/ )
The pet industry market is large and growing. According to the American Pet Products Asso- ciation,^5 in 2019 pet owners in the US spent $95.7 billion on their pets, including $10.7 billion in services like boarding, grooming, training, pet sitting, and walking. That constitutes a 5.5% increase over the previous year. In the US, 84.9 million households, or 68% of all households, own a pet. Of them, 75% own a dog. Dog owners (buyers) use Rover – and DogVacay before the acquisition – to find pet care services from sitters (sellers).^6 The services range from dog walking to in-home pet grooming, but their largest category is dog boarding. Before the acquisition, Rover and DogVacay were the largest players in the online dog boarding market. At the time, the next largest competitor was Wag Labs (Wag). Wag, which mainly offered dog-walking services, started offering overnight boarding only in 2016,^7 although it never grew to become their largest service category. In 2017, Rover earned five times higher revenues than Wag.^8 Offline competitors include more traditional businesses like kennels and dog hotels, and more informal alternatives such as friends and family. Although we do not have data on these alternatives, our theoretical model rules out that kennels change the prices or quality of their offerings in response to the acquisition. On the surface, Rover and DogVacay appear to be close substitutes, especially in comparison to competing platforms in other industries. In particular, Rover and DogVacay had similar interfaces (^5) https://www.americanpetproducts.org/pr (accessed April 2020). (^6) It is fairly easy to join the platform as a pet sitter. One of us signed up on Rover by creating a sitter profile. Platform approval was quickly granted after a general background check. Additional background checks can be performed at the sitter’s will (https://www.rover.com/background-checks/, accessed July 2020). (^7) https://www.vox.com/the-goods/2018/9/12/17831948/rover-wag-dog-walking-app, accessed December 2020. (^8) https://secondmeasure.com/datapoints/wag-rover-dog-walking-sales/, accessed December 2020. Note that this figure includes total sales, not just from dog boarding.
14 Article submitted toAuthors’ names blinded for peer review Management Science; manuscript no.
(Figure 3) and transaction flows, which remained constant at least until the end of our study. When buyers need pet care services, they initiate a search for sellers available in the preferred category,^9 for a given location, and for the dates needed. As is typical in online platforms for local services, buyers then see a list of search results for available sellers ranked by the companies’ proprietary algorithms. For each seller, buyers see their name, picture, location, online ratings, and nightly price. Buyers can then choose to contact sellers to discuss their needs and confirm availability. An exchange is not finalized until both users accept the transaction. Transactions come with reservation protection, trust and safety support, and a secure payment system provided by the platform. A deeper comparison uncovers a number of differences between Rover and DogVacay. Platforms use proprietary algorithms to rank sitters in search results, weighing sitter characteristics differ- ently.^10 DogVacay used to offer a ‘meet and greet’ option before finalizing a match whereas Rover did not. Lastly, user sorting across the platforms could create differences in the user experience, either due to path-dependence or due to strategic decisions by the platforms regarding which types of users to attract (Halaburda et al. (2018)). Just before the acquisition, both Rover and DogVacay took about 20% of gross transaction volume in commission fees, up from 15% when they first started. Sellers would set the prices for their services.^11 As of 2018, fees are divided into a provider (seller) fee and a owner (buyer) fee. The provider fee is 15% for providers who joined before March 2016, and 20% for providers who joined after March 2016. The owner fee is zero if the owner joined before September 2015, while it varies but is never more than $50 per booking for owners who joined after September 2015.^12 DogVacay had a very similar fee structure and its commissions closely tracked those of Rover throughout the period between 2012 and 2017 (Figure 4).
4.1. The Acquisition On March 29, 2017, Rover announced it would buy DogVacay.^13 DogVacay was reportedly strug- gling to keep up with the recent cash injections that Rover had received from venture capitalists,^14 (^9) The service categories include pet overnight boarding, sitting, drop-ins, daycare, and walking. (^10) Details on how the current search algorithm works on Rover can be found at https://www.rover.com/blog/ sitter-resources/how-rover-search-works/ (accessed October 2020). (^11) At the time of our study, the only price suggestion available was Rover’s “holiday rate” feature, which suggested sellers to increase their prices during holidays. (^12) Before July 2019, the maximum owner fee was $25 per booking, according to screenshots on Wayback Ma- chine. These screenshots can be accessed at https://web.archive.org/web/20190705174452/https:/support. rover.com/hc/en-us/articles/205385304-What-are-the-service-fees-. Information on current policies is avail- able at https://support.rover.com/hc/en-us/articles/205385304-What-are-the-service-fees- (accessed De- cember 2020). (^13) https://techcrunch.com/2017/03/29/rover-dogvacay-merge/ (accessed July 2019). (^14) https://www.latimes.com/business/technology/la-fi-tn-dogvacay-rover-20170329-story.html (accessed June 2020).
16 Article submitted toAuthors’ names blinded for peer review Management Science; manuscript no.
4.2. Data We observe all service requests, buyer-seller booking inquiries, matches, and reviews from both platforms before and after the acquisition. A request refers to a buyer’s need for a sitter (e.g. dog boarding in Seattle from August 16th^ until August 18th) and is created when a buyer initiates a search or contacts a sitter directly. Contacts for the same request with different sellers are recorded as booking inquiries. A search leads to a recorded request only if a buyer sends at least one booking inquiry to a sitter. If a booking inquiry leads to a transaction, it is matched to a stay. Both DogVacay and Rover have multiple service categories, but we restrict attention to dog overnight boarding, which constitutes 70% of gross transaction volume on Rover and 91% on DogVacay before the acquisition. We consider all buyer-seller booking inquiries initiated between June 2011 and January 2018 for requests between January 2012 and January 2018 included. Out of all booking inquiries, we remove those whose duration – i.e., number of nights requested – is recorded as negative or greater than 1 month (0.6% of requests), those with lead times – i.e., time between start date and request date – recorded as negative or greater than one year (1.1%), price outliers in terms of total price or commission fee percentage (2.3%). In particular, we remove prices lower than $1 or higher than $200 per night, and commission fees greater than 30%. In total, we exclude 4.2% of total requests, and 3.8% of transactions. We now describe the nature of competition between Rover and DogVacay before the acquisition, which suggests that a merger is likely to generate network effects if those exist in digital platforms like ours. First, the two platforms were of similar size in the dog overnight boarding category before the acquisition, with Rover transacting at a 25% higher volume compared to DogVacay in the quarter before the acquisition.^20 Second, the local nature of the services exchanged implies that buyers are typically interested in transacting with sellers within the same city. Indeed, 79% of booking inquiries and 81% of stays occur within a buyer’s Core-Based Statistical Area (CBSA).^21 This means that we can measure competition between Rover and DogVacay at the local rather than aggregate level. Third, we investigate multi-homing. Few users, and fewer buyers than sellers, use both platforms. We define a user as multi-homing if they transact at least once on both platforms over the 5 years before the acquisition. Only 3.3% of buyers and 7.6% of sellers multi-home. Not surprisingly, multi-homing users tend to transact more frequently than single-homing users. 27% of transactions are made by multi-homing sellers and 8% are made by multi-homing buyers.^22 (^20) Across all service categories, Rover was 62% larger than DogVacay. Appendix Figure C.2 plots the number of monthly stays on DogVacay since January 2012, in log scale. Despite being founded after Rover, DogVacay immedi- ately outgrew Rover in overnight boarding services, before being surpassed again around March 2015. (^21) CBSAs roughly coincide with metropolitan and micropolitan areas. (^22) Appendix Figure C.3 plots the share of a user’s transactions occurring on DogVacay prior to the acquisition, separately for buyers and sellers. On average, only 4.2% of users are both buyers and sellers of services on any given
Authors’ names blinded for peer review Article submitted to Management Science; manuscript no. 17
Figure 4 Average Fees
The figure plots the average commission fee, as a percentage of the price that buyers pay. The vertical line identi- fies March 2017, when the acquisition was publicly announced. Levels on the y-axis are hidden to protect company information.
During the period before the acquisition, DogVacay sellers were expected to receive about $3. more per night (13% more) than sellers on Rover.^23 After controlling for geographic and time ob- servables, the price difference decreases to about 6% but it completely disappears once we compare prices of multi-homing sellers transacting on both Rover and DogVacay within the same month (Appendix Table C.1). This suggests that although sellers may have different qualities across plat- forms, which also may induce demand sorting, multi-homing sellers consider the two platforms as close substitutes. Figure 4 plots the average commission fee on the two platforms, computed as the ratio of plat- form total fees over the price paid by buyers. The figure shows that commission fees were very similar across platforms, and they continued their pre-acquisition upward trend after Rover ac- quired DogVacay. The upward trend is due to the higher fee schedule for buyers and sellers who joined after September 2015 and March 2016, respectively, whose shares increased steadily over time. As is clear from the figure, commission fees did not increase discontinuously after the ac- quisition, suggesting that Rover did not take advantage of its increased market power to increase prices.
year. Buyers rarely act as service providers on the platforms. In the years before the acquisition, on average 4.8% of buyers also transacted as sellers on any given year. Sellers are more often buying pet-sitting services on the platforms. Indeed, 25.8% of sellers also transacted as buyers on any given year. (^23) The payment that a seller receives is equal to what the buyer pays minus the platform commission fees. Tipping is not required, and is not recorded on the platform. However dog owners are not prevented from tipping sitters outside of the platform (https://support.rover.com/hc/en-us/articles/206199686-Should-I-tip-my-sitter-, accessed July 2019).
Authors’ names blinded for peer review Article submitted to Management Science; manuscript no. 19
To test our hypotheses, we cannot simply compare zip codes before and after the merger because aggregate shocks, e.g., due to seasonality or changes in business operations following the acquisition, may confound the results. Instead, we need a control group, which we expect to be relatively unaffected by network effects and platform differentiation. We create such a control group using the zip codes where Rover was already dominant pre-merger (i.e., Rover had more than 80% of the market share). We divide the remaining markets into four treatment groups, corresponding to the other market share groups displayed in Figure 5. It is important to allow for the treatment effects to vary across markets with differing market shares since our theory predicts non-monotonic effects across these markets. Zip codes where either Rover or DogVacay were dominant before the acquisition tend to be more rural, have fewer residents, lower population densities, and lower shares of college graduates. Areas where Rover is particularly successful also tend to have higher pet ownership rates.^25 Given these differences, we may be concerned that the main assumption behind a difference-in-differences approach, that zip codes with different market shares have the same latent trends in platform performance, does not hold. To ensure that zip codes in treated market share groups are as similar as possible to zip codes in the control group, we employ a matching estimator that accounts for covariate imbalance across groups (Imai et al. 2018). We match one zip code from the control group to each “treated” zip code using covariate balancing propensity score matching (CBPS), introduced by Imai and Ratkovic (2014). Distances are calculated on the total number of active sellers in each month up to a year before the acquisition, where an active seller is defined as a seller who was involved in at least one booking inquiry in the given month. We hold the matched control group constant as we measure the effects of combining the two platforms across different outcomes of interest. Matching on number of sellers ensures that treated and control groups have similar number of participants across the two platforms combined, but our results do not depend on whether we match on the number of buyers, the number of sellers, or a combination of both (Section A.2). Appendix Table C.4, which provides descriptive statistics for the matched samples, shows that we are able to improve matching on a number of covariates that we do not explicitly use in the matching procedure.^26 However, platform performance metrics that are not explicitly considered in matching (e.g. prices, match rates, and share of repeat transactions) fail to balance across treatment and control group. Some of this imbalance is expected — for example we know that prices are higher on DogVacay and average prices will therefore be higher in markets with a higher DogVacay share.
(^25) Appendix Figures C.5 and C.6, together with Appendix Table C.3, provide comparisons for a large set of observable demographics and platform performance metrics. (^26) Appendix Table C.3 presents descriptives for the unmatched zip codes.
20 Article submitted toAuthors’ names blinded for peer review Management Science; manuscript no.
Other differences reflect the fact that platform performance metrics tend to positively correlate with a platform’s market share. We should note however, that our empirical strategy, described below, does not require identical levels of pre-treatment outcomes, but rather parallel trends. The figures in Section 6 provide support for this assumption. Given matched zip codes, let yzt be the outcome in treated zip code z and year-month t. Sepa- rately for each treated market share group [0 − 20%), [20% − 40%), [40% − 60%), and [60% − 80%), we estimate the following regression:
yzt − yz′t = αt + z,z′,t, (5)
where z is the treated zip code, and z′^ is the matched control zip code. The coefficients αt should be interpreted as changes in the outcome variable relative to the control group, and relative to February 2017, the month before the acquisition announcement. Cluster-robust standard errors are calculated using the method from Aronow et al. (2015).^27 Equation (5) allows us to test Hypothesis 1 and Hypothesis 3. Hypothesis 1 posits that, due to network effects, the coefficients αt after the merger should be positive and increasing as Rover market share decreases. For Hypothesis 3, if network effects are large enough to justify a single combined platform, we would expect the largest benefits from network effects to arise in the zip codes with intermediate market shares. To test Hypothesis 2 we need a different approach. Recall that in order to evaluate the role of platform differentiation, we need to estimate the extent to which DogVacay buyers are worse off relative to Rover buyers who experienced the same change in platform size. Rover buyers in markets with Rover’s pre-merger market share of ¯n experience a change in platform size similar to DogVacay buyers in markets with Rover’s pre-merger market share of 1 − ¯n. We attribute any difference in outcomes between Rover and DogVacay buyers in these symmetric markets to a reduction in platform differentiation. Let s ∈ { 0 , 20%, 40%, 60%, 80%} denote the lowest Rover’s market share in each of our market share groups. For each of the five s, we consider the outcomes of Rover buyers in zip codes with market shares within [s, s + 20%) and the outcomes of DogVacay buyers in zip codes with market shares within [80% − s, 100% − s). With these outcomes we estimate the following regression:
yzt = βt + γt 1 {z has market share in [80% − s, 100% − s)} + νz + zt, (6) where yzt is the outcome of Rover buyers in zip code z and year-month t if z ∈ [s, s + 20%), or the outcome of DogVacay buyers in zip code z and year-month t if z ∈ [80% − s, 100% − s). The
(^27) Each matched pair, or dyad, is no longer independently informative, as a single control market can impact the estimates of multiple dyads. The method proposed in Aronow et al. (2015) accounts for the correlation in error terms between each matched pair.
