How Users Drive Value in Two-sided Markets - Platform Designs That Matter

How Users Drive Value in Two-sided Markets: Platform Designs That Matter Abstract One of the deepest platform challenges is understanding how users create network value for each other and which investments provide leverage. Is more value created by advertising to attract users, discounting to subsidize users, or investing in architecture to connect and retain users? Having grown a user network, which promotes “winner-take-all” dominance, why do platforms with large user bases fail? To address these challenges, we build a theoretical and empirical model to simultaneously measure within and across period network e↵ects for two-sided markets. This yields three main results. First, we extend the customer lifetime value (CLV) literature to network markets, allowing us to measure how di↵erent interventions drive CLV on both sides of two-sided markets. Second, we apply our model to the case of Groupon and empirically estimate the strength of within period and across period attraction. We find significant within period attraction between merchants and consumers that does not persist through time. This o↵ers one explanation for the puzzle that even strong network e↵ects might not yield market dominance. Third, we find that users attract users more strongly in experience than search goods markets, as user input drives CLV more than price does. Together these results provide numerous points of leverage for understanding and boosting CLV in order to increase platform value. 1 1 Introduction If the world’s first trillion-dollar businesses are platforms based on network e↵ects (Cusumano et al., 2020), why might firms with strong networks falter? Stronger network e↵ects drive platforms toward greater market dominance (Katz and Shapiro, 1985, 1986; Shapiro and Varian, 1998; Schilling, 1999; Park, 2004; Dubé et al., 2010). This encourages platforms to invest in user growth, hoping that jump-starting a user base leads to self-propagating growth. Empirically, however, platforms with large user bases have both succeeded and failed (McIntyre and Srinivasan, 2017; Hagiu and Rothman, 2016). Firms such as Apple, Amazon, Google, and Facebook have grown large user bases, and with high margins and market shares, their valuations have followed. On the other hand, platforms such as Uber, DiDi, Lending Club, and Groupon have grown large user bases yet their valuations have fallen. Venture capitalists have subsidized such firms, despite an absence of profits, under the belief that investing in user growth until network e↵ects achieve critical mass can lead to market dominance. Success simply takes patient capital. Central questions for both platform investors and managers are: When is a large user base insufficient for a “winner-take-all” market, and which user investments create the most value? Answering these questions informs both investment and strategy. Focusing on platforms, this research explores the role network e↵ects play in creating long term customer value. Current models of customer lifetime value (CLV), however, omit network e↵ects (Gupta et al., 2004; Pfeifer and Farris, 2004; Zhang et al., 2012; McCarthy and Fader, 2018), missing the positive externalities they create for each other. Standard models of network e↵ects are static (Rochet and 2 Tirole, 2003; Parker and Van Alstyne, 2005; Armstrong, 2006), missing the chain of value creation over time. The main contribution of this research is to build a dynamic within and across period model of two-sided network e↵ects, from which we develop a two-sided model of customer lifetime value, CLV2. We validate this empirically, then show how it applies across di↵erent users, goods, and locations. Our theoretical question is how to properly measure network e↵ects over time in order to understand CLV2. Using this model, our empirical question is to measure the strength of attraction in a two-sided market, accounting for time and transacting across various goods and services. Our point of leverage is to use time, within and across periods, to see if we can observe attraction, spillovers, and stickiness. Observing these factors, we also identify precursors that moderate CLV2. Answers thus provide not only investment advice but also insight into odd market behaviors such as low market values despite strong network e↵ects. In the case of Groupon, for example, we find significant cross-side attraction but not stickiness suggesting that better platform design and investment strategies are warranted. We take “network e↵ect” to mean that a product or service has the property of rising in value as more parties consume it (Katz and Shapiro, 1985; Liebowitz and Margolis, 1994; Shapiro and Varian, 1998). This rise in value naturally attracts consumers whose usage then naturally creates value. Interaction of these properties creates the flywheel investors hope lead to market dominance. This conventional definition of a network e↵ect, however, has several nuances worth clarifying. One is that how users create value can be intentional or unintentional (Parker et al., 2016). The explicit case is represented by reviews, user generated content, and users actively 3 linking to other users. The implicit case arises when machine learning algorithms observe users’ behavior and improve the experience of other users or o↵er value adding recommendations. The latter requires no intentional act to create value for other users. Network e↵ects can also exhibit strength or weakness based on whether they are local or non-local (Sundararajan, 2008; Parker et al., 2016). Rideshare, home services, and food delivery, for example, give rise to stronger local than nonlocal attraction as it makes no sense to match a rider in New York to a driver in Paris. By contrast, e-commerce, operating systems, and social networks have stronger nonlocal network e↵ects because a match need not be location specific. Further, network e↵ects can be classified based on whether they are “same-side” versus “cross-side” (Eisenmann et al., 2006; Rochet and Tirole, 2006) meaning that the e↵ect operates within group or across groups respectively.1 In our context, a same-side network e↵ect is represented by the value that consumers create for other consumers or the value that merchants create for other merchants. A cross-side e↵ect is represented by the value that consumers create for merchants and vice versa. Thus a two-sided market can exhibit up to four network e↵ects – two same-side and two cross-side – with no requirement that these necessarily exist or have equivalent strengths. For example, competition among online merchants or among rideshare drivers could eliminate positive spillovers and attraction among them. Adding the time dimension, representing within and across periods, doubles this count to eight possible factors. We will find it convenient to simplify this complexity in the analysis that follows. 1 Certain literature also uses the terms “direct” and “indirect” for same-side and cross-side network e↵ects respectively (Clements, 2004; Church et al., 2008). 4 To give precision to these eight factors, we introduce the following conventions. Let “IJt ” designate network attraction from the class of user i to the class of user j with the subscript representing time from t to the present. Thus, CM0 represents a cross-side network attraction from consumers to merchants within the current period and CC 1 represents a same-side network attraction from consumers to consumers from the prior period to the present. This also generalizes. We can use SS0 to represent a Same-Side within period network e↵ect for both consumers CC0 and merchants M M0 or use CS 1 to represent a Cross-Side network e↵ect from prior period for both consumers to merchants CM 1 and vice versa M C 1 collectively. We summarize nomenclature in Table 1, which also illustrates how attraction occurs. Table 1: Di↵erent Dimensions of Network E↵ects Within-period Across Period Same-side Within-period same-side network e↵ect (SS0 : CC0 , M M0 ) E.g., Gamers interacting with gamers Inter-temporal same-side e↵ect (SS 1 : CC 1 , M M 1 ) E.g., Customers learning from prior customer reviews Cross-side Within-period cross-side network e↵ect (CS0 : CM0 , M C0 ) E.g., Customers transacting with current period merchants Inter-temporal cross-side e↵ect (CS 1 : CM 1 , M C 1 ) E.g., Merchants learning from prior customer reviews Measuring network e↵ects faces special challenges due to the well-known “reflection problem” (Manski, 1993, 1999; Rysman, 2019) and represents one reason empirical research is underdeveloped (Jullien et al., 2021). The adoption choice of an individual depends on the adoption choice by the group yet the group’s choice depends on the collective choices of the individuals that comprise it. This simultaneity of decisions has led much of the theoretical literature to embrace fulfilled expectations equilibria where group members correctly anticipate the behavior of other members (Katz and Shapiro, 1985; Cabral, 1990; Krugman, 1991; Economides, 1996). The fact that internal expectations within period are not externally observable confounds 5 empirical measurement because even use of instruments treats all members of that group the same. Much of this work, however, predates development of two-sided network models (Parker and Van Alstyne, 2000; Rochet and Tirole, 2003; Caillaud and Jullien, 2003) that allow choices of one group to depend on those of a matched but di↵erent group. We therefore follow Rysman (2019) and identify instrumental variables that treat only one side of our two-sided network. Further, by introducing the time dimension, choices within group may depend on group choice in the prior period, which is externally verifiable (Zhang et al., 2012; Chu and Manchanda, 2016).2 Careful measurement thus motivates our focus on same-side prior-period network e↵ects and on cross-side within-period network e↵ects. Using our notation, these are SS 1 and CS0 respectively. This focus not only reduces complexity but also avoids the reflection problem. Di↵erent factors modulate product attraction. User generated content (UGC), for example, is more valuable for products with higher uncertainty (Li et al., 2011). Prior sales information also has a greater impact on a customer’s choice of an experience good than of a search good (Li and Wu, 2018). In the same vein, we hypothesize that same-side attraction for experience goods is stronger than that for search goods based on peer to peer learning e↵ects (Hagiu and Wright, 2020). Greater information asymmetry renders accumulated information more valuable for goods that must be tried in order to know their value than for goods whose price is the deciding factor. Distance might also matter (Sundararajan, 2008). Since Groupon customers trans2 Another way to resolve the reflection problem is to shrink the time interval of the observation window. This can shift observation of a specific behavior from the current to a prior period, avoiding a simultaneous choice. We thank an anonymous referee for this insight, which also motivates a focus on SS 1 rather than SS0 . 6 act with Groupon merchants in their local markets, we expect local network e↵ects to prove stronger than non-local network e↵ects. Using our model of inter-temporal attraction, we explore both factors and find that both matter. Among customers, same side attraction CC 1 is estimated to be positive and significant for experience goods but insignificant for search goods. For both experience and search goods, however, the cross side e↵ect of merchants on customers M C0 is estimated to be positive and significant. Among merchants, same side M M 1 and cross side CM0 e↵ects are estimated to be positive and significant for both experience goods and search goods, but M M 1 for experience goods is approximately 2.5 times as large as that for search goods while CM0 is qualitatively similar across these categories. Using model coefficients, we can also account for two-sided market spillovers and explore marketing and information systems interventions to boost customer and merchant lifetime value CLV2 and MLV2. For example, does CLV2 rise more based on adding another customer, another merchant, increasing consumption, or increasing retention? Marketing interventions might add same or cross side market participants. Information systems interventions might capture and present user data or use data to boost stickiness. Interestingly, in the extreme where same side e↵ects SS 1 are absent on the platform, strong cross-side e↵ects CS0 only create value in one period, and CLV2 is equal to the value created in the current period. However, when SS 1 is significant and positve, users continue creating value in future periods. In practical terms, these tools allow investment analysis of di↵erent options such as methods to enhance user participation (Huang et al., 2019; Chen et al., 2019) or to alter user behaviors that drive di↵erences in platform value (McIntyre and Srinivasan, 7 2017). Amazon was among the first platforms to invest in recommender systems and UGC to make their undi↵erentiated books more valuable (Dellarocas, 2003). Using information systems data to present consumers with their own “best value” ranking can drive adoption (Ghose et al., 2012). In SAP’s enterprise software ecosystem, independent software vendors (ISVs) compete for contracts and initially hid their expertise from one another. The launch of SAP’s developer network, however, rewarded partners for sharing ideas and answering each other’s questions in a public forum, which also boosted adoption and retention. ISVs have increased each other’s sales, and thus customer value, at statistically significant levels (Ceccagnoli et al., 2012). Airbnb’s host forum, likewise, helps hosts solve problems, deal with unruly guests, and share tips for improved ratings (Hardy and Dolnicar, 2017). Platforms can also design information systems to simplify user adoption and avoid disintermediation (Zhu and Iansiti, 2012; Halaburda et al., 2016). For example, platforms can schedule bookings, process payments, and handle logistics to drive platform adoption and retention. As platform users and merchants become stickier due to these value-added services, customer and merchant lifetime value increase, such that user (and/or merchant) acquisition strategies become cost-e↵ective. Parameter estimates for strengths of these relationships help target investments. This research o↵ers theoretical and practical insights. First, we develop a model of inter-temporal network e↵ects and within-period network e↵ects to complement the static models of two-sided markets in extant literature. Strong within-period attraction does not necessarily imply strong inter-temporal attraction across periods. When the inter-temporal network e↵ect is weak, the within-period network e↵ect 8 may not persist, and a user growth strategy is not e↵ective. Given this model, we further propose a framework to estimate the customer and merchant lifetime values, CLV2 and MLV2, in two-sided markets. Existing CLV models do not capture the value created through network e↵ects and underestimate that of platforms such as Amazon and eBay (Gupta and Mela, 2008; Zhang et al., 2012). Our innovation extends this literature to a pair of joint calculations accounting for spillovers across time and across sides of two-sided markets. To the best of our knowledge, this is the first study that proposes a theory-driven customer and merchant value calculation that can be adopted easily by practice. Second, this paper empirically distinguishes and estimates di↵erent network effects based on time, location, and type of good. In our context, the inter-temporal network e↵ect is weak overall, yielding poor user stickiness. A user growth strategy is less e↵ective for such platforms. All else equal, CLV 2 and M LV 2 are higher in markets with higher same side inter-temporal attraction SS 1 . For markets with weak SS 1 , we show that platform design interventions that enhance stickiness and spillovers can increase lifetime value, whereby acquisition strategies through marketing interventions become cost-e↵ective. Overall, these findings remind platform managers not to overemphasize user growth when user or merchant stickiness and spillovers are poor and, instead, to focus on platform designs that enhance stickiness and spillovers as well as attraction. Our paper proceeds as follows. Section 2 reviews related literature. Section 3 introduces our theoretical and empirical models. Section 4 then describes our data set and identification strategies. We present empirical findings in Section 5. Follow- 9 ing coefficient estimation, Section 6 analyses strategic responses when a platform’s user stickiness is poor. Section 7 o↵ers commentary on limitations and directions for future research. Finally, Section 8 concludes with managerial and investor implications, and highlights research contributions. 2 Related Literature Our research builds on contributions from four research streams including two-sided markets, market dynamics, customer lifetime value, and platform design. Research extensions and contrasts for each stream appear below. 2.1 Static Models of Two-sided Markets Static models have developed to study the price structure of two-sided markets, how they lead to free goods markets, and which side merits a subsidy (Rochet and Tirole, 2003; Caillaud and Jullien, 2003; Parker and Van Alstyne, 2005; Armstrong, 2006). These theoretical models highlight how one side of a market attracts the other side, requiring use of a coordination mechanism, such as free pricing or a price stimulus to one side, in order to get both sides on board. Empirical works in this context have focused on estimating network e↵ect size and discussing implications for platform competition (Rysman, 2004). Due to the intrinsicly dynamic nature of platform growth, however, static models are insufficient for understanding time dependent aspects of user adoption on a platform’s net present value. Our paper extends the static two-sided market framework by adding time. Although the per- 10 sistence of network e↵ects plays a critical role in platforms’ user growth strategies, little empirical work has investigated their inter-temporal durability (McIntyre and Srinivasan, 2017). Considering market entry, Biglaiser et al. (2019) discuss how incumbents maintain a profit advantage via network e↵ects even as entrants match or exceed price and quality. In this research, we control for factors such as platform quality, the presence of a competitor, and how users themselves contribute to persistent network e↵ects. 2.2 Dynamic Models of Network E↵ect Markets Network e↵ects have attracted a long-standing interest in market coordination and evolution. The adoption decision is often modeled as a function of users’ expectations of network size (Katz and Shapiro, 1994; Shapiro and Varian, 1998; Cabral, 2011; Krugman, 1991; Farrell and Klemperer, 2007). Empirical contributions in this domain have estimated the magnitude of di↵erent cross-side network e↵ects. For example, Chu and Manchanda (2016) quantified the extent to which current consumers attract future suppliers (and vice versa) in online retail. Zhang et al. (2012) examined how UGC contributors created value for UGC consumers. Importantly, these models use only lagged prior period terms, and not within period terms, omitting measures of current period attraction. We include current period attraction and use an instrumental variables approach to address the reflection problem. Li et al. (2020) further studied user growth across multiple categories of goods and across di↵erent time periods in a two-sided market. Insights from these studies show how to optimize investments in user growth conditional on the e↵ectiveness of 11 di↵erent cross-side network e↵ects. The assumption that cross-side network e↵ects depend on user beliefs of network size is reasonable when user switching costs are high. Platform research has shown that switching cost is an important determinant of user value (Farrell and Klemperer, 2007; Metcalfe, 2013; Katz and Shapiro, 1985). At present, however, user switching costs are low on many digital platforms. In search, rideshare, group-buying, and e-commerce, for example, users can easily observe competing platform o↵ers and switch without lock-in. The possibility of platform switching can dissipate rents and intensify competition (Rysman, 2009). A large supplier base in one period does not necessarily guarantee consumer adoption in the next period. Cross-side attraction in existing papers, often cast as an inter-temporal e↵ect (Chu and Manchanda, 2016), corresponds to the mixture of SS 1 and CS0 in our model. Since di↵erent types of user e↵ects correspond to di↵erent mechanisms and imply a need for di↵erent platform strategies, we distinguish them and estimate them separately. Our model shows that platform user stickiness plays a critical role: when user stickiness is poor, no user growth strategies can generate high returns on investment (ROI) because the network is just a “leaky bucket.” In contrast to conventional wisdom, investing in either side or in di↵erent categories of goods might not be cost e↵ective. Rather, we suggest that a platform should improve its design so as to enhance user stickiness. Our user-value strategy is particularly relevant when switching costs are low, as is the case for many digital platforms. 12 2.3 Network E↵ect Design The importance of network e↵ects is such that they imply changes in the structure of the firm. Parker et al. (2017) show that product firms transition to platform firms as the strength of network e↵ects increase. Because network e↵ects scale faster among external users than among internal employees, the firm “inverts” in the sense that value creation shifts from inside to outside. Network e↵ects, however, are not simply exogenous environmental factors but might instead be the consequence of conscious design (Schrage, 2012). Platforms can enhance network e↵ects through investments in connecting users to one another, enabling file sharing, creating complements, and hacking compatibility (McIntyre and Subramaniam, 2009). Piggybacking on social media to promote word-of-mouth (WOM) can generate social influence. Capturing and presenting user-generated-content (UGC) builds engagement. Extant research has shown that reviews, ratings, and other UGC a↵ect the attractiveness of a platform over time (Dellarocas, 2003; Forman et al., 2008; Li et al., 2011). Several recent studies also model network e↵ects as a learning mechanism because UGC sheds light on product/service quality, which reduces uncertainty as customers make their purchase decisions (Hagiu and Wright, 2018; Wu et al., 2015). Firms also give away reference designs and system developer toolkits (Boudreau, 2010), or open application programming interfaces (Benzell et al., 2017). These design choices attract developers who improve platform quality, which attracts users, creating platform value (Tellis et al., 2009; Zhu and Iansiti, 2012). Of particular importance for our paper, network size alone might not be the best predictor of value, which might instead depend more 13 on network density or topology (Suarez, 2005). Complementing these works, our paper proposes a framework to manage user stickiness modeled as di↵erent categories of same-side and cross-side network ties . We show that investments in network designs can make current users more valuable for future users, magnifying each user’s lifetime value, and raising the value of the platform. This is consistent with investing for the purpose of “making customers better to make better customers” (Schrage, 2012). As a further illustration, Airbnb trained hosts to take more attractive photos of their homes, which increased guest bookings, which inn turn increased guest listings. Network e↵ects need not be assumed, they may be designed. 2.4 Customer Lifetime Value An important stream of research that is closely related to firm value is the concept of CLV in marketing (Gupta et al., 2004; McCarthy and Fader, 2018). This line of research emphasizes user stickiness and explicitly incorporates user retention in CLV calculation (Gupta et al., 2004; Pfeifer and Farris, 2004; Gupta, 2005). However, extant literature is largely restricted to businesses without network externalities. Despite its popularity, the extant CLV calculation is insufficient to estimate the true value of users on digital platforms. One primary objective of our research is to explicitly model how network externalities create value across groups and across time. We thus extend the CLV framework to two-sided markets. Furthermore, extant CLV research does not distinguish how di↵erent factors contribute to user value. Following Zhang et al. (2012) and Li et al. (2020), we control for factors such 14 as platform quality and focus on analyzing how di↵erent types of network e↵ects a↵ect value creation. Our study is distinct from extant studies that also addressed network e↵ects (Tucker and Zhang, 2011; Chu and Manchanda, 2016; Li et al., 2020). Whereas those studies focus on choosing a preferred investment side or within versus cross-category investment, our research explicitly provides a framework that not only quantifies CLV2 for both sides in a two-sided markets but also sheds light on the relative contributions of network e↵ects and net margin in boosting user value. This makes a comparison of advertising, discounting, and architectural design investments feasible. 3 Modeling Approach This section presents our model specification. First, we explain how di↵erent types of network e↵ects influence the dynamics of two-sided markets and specify an aggregatedata discrete-choice model to study users’ participation choices (Berry et al., 1995). Second, based on the estimates of network e↵ect sizes, we develop a model to calculate consumer and merchant lifetime values, CLV 2 and M LV 2 , in two-sided markets. We discuss platform heterogeneity and demonstrate how the strength of SS 1 a↵ects the e↵ectiveness of the user growth strategy. 3.1 Demand Dynamics in Two-sided Markets Built upon the static model of Parker and Van Alstyne (2005) and the dynamic model of Chu and Manchanda (2016), we incorporate the time dimension and study 15 demand dynamics in two-sided markets. Figure 1 describes the demand evolution on the two sides of a online market, which allows consumers (merchants) in the pervious period to influence consumers (merchants) in the current period. To illustrate market evolution, consider a shock to consumer base at t 1 and its dynamic e↵ect, as shown in Figure 1. First, the shock to consumer base at t 1 a↵ects the consumer base at t via SS 1 . Second, an increase in the current consumer base would lead to growth in the current merchant base via the CS0 . The prices and sizes of each user group combinedly determine the profit during each period in the dynamic system. Here nt , mt , pnt , pm t correspond to the consumer base, merchant base, price for consumers, and price for merchants, respectively, in period t. The same logic applies to a shock to the merchant base. Next, we present how these di↵erent network e↵ects are modeled empirically. For exposition, we refer to the two sides as the “consumer side” and the “merchant side”, although our modeling framework can be applied to a two-sided market beyond the retailing setting. 3.1.1 Consumer Side On the consumer side, several factors jointly determine the utility that a representative buyer i derives from using the platform in market j week t. First, the utility is related to the platform’s attractiveness, which is further decomposed into four components: (1) consumer i0 s intrinsic preference for the focal platform ↵jn (in subsequent analysis, superscript ‘n’ denotes the consumer side), which is assumed to be varying by market but constant over time; (2) the number of consumers in the previous period nj,t 1 , to capture CC 1 that is attributable to word-of-mouth and 16 Figure 1: Each period shows dynamic participation of consumers (nt ) and merchants (mt ). Link nt 1 ! nt represents CC 1 while link nt ! mt represents CM0 . Links anchored at mt 1 and mt are symmetric in network e↵ects M M 1 and M C0 respectively. UGC; (3) the number of merchants mjt , to capture M C0 that is due to the amount of supply;3 and (4) average price pjt and other product characteristics Xjt . Furthermore, the utility depends on (i) the attractiveness of the rival platform Rjt (Rochet and Tirole, 2003; Armstrong, 2006), which includes rivals’ user base in market j at t, ii) the macro-economic trends of the industry and seasonalities that are common to all markets and unobserved to researchers Ttn , (iii) the unobserved market and time n specific shocks to the demand ⇠jt , and (iv) the idiosyncratic error "nijt . Formally, the net indirect utility is specified as follows: n n n Uijt = f (↵jn , nj,t 1 , mjt , pjt , Xjt )f (Rjt )f (Ttn , ⇠jt , ✏ijt ). 3 (1) The lags of merchant base are excluded from Equation 2 because we model a setting with trivial switching cost. In such a context, consumers make decisions after observing current supply on competing platforms, and thus, they do not need to form expectations about current supply based on past supply. 17 We assume that the utility takes the Cobb-Douglas form (Chu and Manchanda, 2016) and rewrite the indirect net utility as: unijt = ↵jn + 1 ln nj,t 1 + 1 ln mjt + ⇢1 ln pjt + ⌘1 Xjt n 1 Tt + ✓1 ln Rjt + + n ⇠jt + (2) "nijt . The inter-temporal and within-period network e↵ect parameters are our prin mary interest in Equation 2. There are also market-specific demand shocks ⇠jt that make the platform more or less attractive across di↵erent markets over time. These shocks are observed by platforms and users but unobserved to researchers, causing an endogeneity concern. All other factors are absorbed in the idiosyncratic errors "nijt . We now denote the utility as unijt = n jt + "nijt and normalize the utility of the outside option as zero, which corresponds to consumers choosing the rivals or not participating in the market. Assuming that "nijt follows independent and identically distributed (i.i.d.) Type-I extreme value distribution, we derive the market share of the focal platform as: n snjt e jt = 1+e n jt . (3) Thus, the platform’s relative market share is: snjt njt /Njt ln n = ln s0jt Njt njt /Njt = ↵jn + 1 ln nj,t + ✓1 ln Rjt + 1 n 1 Tt + 1 ln mjt + ⇢1 ln pjt + ⌘1 Xjt n + ⇠jt , 18 (4) where sn0jt is the market share of the outside option and Njt is the market size for buyers in market j time t. Estimating the parameters of interest based on Equation 4 poses challenges due to endogeneity and market heterogeneity. We discuss our identification strategy in detail in Section 6. 3.1.2 Merchant Side The utility specification for the merchants is similar to that of the consumer side. For a representative merchant (seller) l in market j at time t, its utility of working with the platform is related to the following factors: (1) merchants l0 s preference for the focal platform ↵jm (in subsequent analysis, superscript ‘m’ denotes the merchant side), which is assumed to be varying by market but constant over time; (2) the number of existing merchants mj,t 1 to capture M M 1 ; (3) the number of current consumers njt to capture CM0 ; (4) product price Pljt ; and (5) whether the merchant has experiences with the platform Xljt , which captures the switching cost on the platform. Note that the merchant base in the past, the M M 1 , can a↵ect the participation decision for a current merchant for several reasons. Existing merchants could create a word-of-mouth e↵ect, which helps the platform to attract new merchants. Too many merchants, however, could increase competition and deter future entry, causing a negative same-side network e↵ect. Thus, the M M 1 on the merchant side could be positive or negative, depending on which e↵ect dominates. The CM0 from consumers to merchants is expected to be positive, because more consumers would bring higher 19 sales and profits, increasing the attractiveness of the platform to merchants 4 . For specific platforms such as Groupon, the consumer base helps increase the consumer reach and awareness for the merchant’s brand, creating a marketing e↵ect (Dholakia, 2011; Edelman et al., 2016). Nevertheless, the CM0 term in our model captures the cross-side e↵ect of consumers on both the current profits and future profits. The merchant’s utility further depends on (i) the attractiveness of the rival platform Rjt , (ii) the macro-economic trends of the industry that are common to all markets and unobserved to researchers Ttm , (iii) the unobserved market and time m specific shocks to the demand ⇠jt , and (iv) the idiosyncratic error "m ljt . Formally, the net indirect utility for a merchant is specified as follows: m um ljt = ↵j + 2 ln mj,t 1 + 2 m 2 Tt + ✓2 ln Rjt + ln njt + ⇢2 ln pljt + ⌘2 Xljt + m ⇠jt + (5) "m ljt . Following a similar step as for the consumer side, we denote the utility as um ljt = m m jt +"ljt and normalize the utility of outside option as zero. Assuming that "m ijt is i.i.d. Type-I extreme value distributed, we derive the relative market share of merchants choosing the platform as follows: sm mjt /Mjt jt ln m = ln s0jt Mjt mjt /Mjt = ↵jm + 2 ln mj,t + ✓2 ln Rjt + 1 m 2 Tt + 2 ln njt + ⇢2 ln pjt + ⌘2 Xjt (6) m + ⇠jt , 4 Following Li et al. (2017), the lags of the consumer base are excluded from Equation 5 because merchants have correct expectations about future demand. 20 where sm 0jt is the market share for merchants choosing the outside option and Mjt is the market size for potential merchants. Variables pjt and Xjt are the average deal price and characteristics in market j week t. Our estimation and identification strategy is discussed in Section 6. Table 2 provides a summary of all variable and parameter definitions used in our model. Table 2: Variable and Parameter Definitions Notation n m N M R p X T 3.2 Definition consumer base merchant base consumer market size merchant market size rival user base and price price product features industry-level common trend Notation 1, 2 1, 2 n ↵j , ↵jm snjt , sm jt ✓1 , ✓2 ⇢1 , ⇢2 ⌘1 , ⌘ 2 1, 2 n m ⇠jt , ⇠jt Definition CC 1 , M M 1 M C0 , CM0 user preference market share competitive e↵ect price elasticity product attractiveness time e↵ects unobserved shocks Customer Lifetime Value in Two-sided Markets In their classic work, Gupta et al. (2004) define the CLV as the expected sum of discounted future earnings generated by a customer. Following their definition, we develop a model to estimate the CLV 2. Take the consumer’s CLV 2 as an example. We assume a one-unit increase in the consumer base and quantify the increase of user bases on both sides of the market. To compute the CLV 2, we transform the estimated elasticity of the market share into a marginal e↵ect of the user base using 21 Y / X = Ȳ /X̄ ⇥ the standard relationship: ln Y / ⇥ (1 s̄Y ), where sY is the ln X average market share of Y (Berry et al., 1995; Trusov et al., 2009). Thus, the marginal e↵ects of inter-temporal user attraction are ˜1 = s̄n ) and ˜2 = 1 ⇥ (1 and the marginal e↵ects of within-period user attraction are ˜1 = and ˜2 = m̄ n̄ ⇥ 2 ⇥ (1 n̄ m̄ ⇥ 2 ⇥ (1 1 n̄ p̄ ⇥ ⇢1 ⇥ (1 s̄n ) and ⇢˜2 = m̄ p̄ ⇥ ⇢2 ⇥ (1 s̄m ), respectively. In the long term, a one-unit increase in the consumer base leads to tional consumers via CC 1 , and on merchants. Furthermore, 1 s̄n ) s̄m ). The marginal e↵ects of deal price on the consumer and merchant bases are ⇢˜1 = the M M ⇥ (1 s̄m ), ˜2 1 ˜1 ˜2 1 ˜1 1 1 ˜1 addi- additional merchants via the CM0 of consumers merchants lead to ˜2 1 ˜1 ⇥ 1 1˜ total merchants via 2 of merchants. For each consumer and merchant, we use ⇡ n and ⇡ m , re- spectively, to denote the value they generate in one period. We use the discounted cash flow approach to compute value perpetuity (Brealey et al., 2012). Thus, the consumer value is represented by the total increased profits associated with a oneunit increase in the consumer base.5 Formally, the consumer’s CLV 2 is computed as follows: CLV 2 = 1 1 | ⇥ ⇡n + ˜1 1 {z } | consumer side value ˜2 ˜1 ⇥ 1 1 ˜2 {z 1 1 | ⇥ ⇡m + ˜2 1 {z } | merchant side value (7) ⇥ ⇡n . } (8) merchant side value Based on the same logic, the merchant’s M LV 2 is: M LV 2 = ⇥ ⇡m . } ˜1 ˜2 ⇥ 1 1 ˜ {z 1 consumer side value We can then examine the e↵ectiveness of a user growth strategy by comparing 5 To simplify the calculation, we assume there is no discount of profits across periods. 22 the CLV 2 and the consumer acquisition cost, and the M LV 2 and the merchant acquisition cost, respectively. We denote the consumer acquisition cost and merchant acquisition cost as AC n and AC m , respectively. A user growth strategy is e↵ective only when CLV 2 > AC n or M LV 2 > AC m . Importantly, Equations 7 and 8 can also be used to answer the question: what is a free customer worth (Gupta and Mela, 2008). The direct e↵ect of ⇡ n in the first equation or of ⇡ m in the second equation could be zero and yet both CLV 2 and M LV 2 could be positive due to non-zero cross-side value. 3.3 Platform Heterogeneity in User Growth Strategy An interesting question to consider is why the e↵ectiveness of a user growth strategy varies by platforms with strong CS0 . According to Equation 7 and 8, the CLV 2 and M LV 2 are jointly determined by the strength of both SS of SS 1 1 and CS0 . The existence makes the influence of CS0 persist over time and thus amplifies the impact of CS0 on CLV 2 and M LV 2. Table 3: How SS ASn 1 1 0 2 0.5 3 0.667 1 Amplifies The Impact of CS0 on CLV 2. 4 0.75 To illustrate the impact of SS 5 0.8 1 6 0.833 7 0.857 8 0.875 9 0.889 on CLV 2, we vary the value of 10 0.9 1 while keep other parameters unchanged, then compute the changes in CLV 2. We use ASn to denote the amplification size of the consumer’s SS 23 1 on CLV 2. According to our Figure 2: Cash Flow in Each Period When SS ASn = 1, ASn = 4, model, ASn = 1: 1 . 1 ˜1 1 1 =0 1 Changes. ASn = 2, = 0.75 ASn = 8, 1 1 = 0.5 = 0.875 Table 3 shows the correspoding relationship between ASn and not only does ASn increase with when 1 1, but also the increase in ASn accelerates becomes larger. Figure 2 depicts the impact of various SS in future periods. In an extreme case when the SS 1 1 on cash flow is absent on the platform, the CS0 only creates value in one period, and the CLV 2 is equal to the value created in the current period. However, when the SS 1 exists, users continue creating value in future periods, leading to amplified CLV 2. Therefore, a weak SS 1 could explain why platforms with strong CS0 may fail in using a user growth strategy. We theorize that one reason for the heterogeneous SS 1 is because of the “product learning” mechanism: product with higher uncertainty would expect a higher SS 1 . Using Groupon data, Li and Wu (2018) show that 24 previous sales information has a greater impact on consumer choices in the category of experience goods than search goods because experienced goods are associated with higher uncertainty. In the same vein, we hypothesize that the SS 1 of experienced goods is stronger than that of search goods. In the following sections, we discuss our identification strategies and test this hypothesis. 4 Data and Estimation This section presents the empirical setting and our identification strategy. 4.1 Empirical Setting and Data We estimate our model and discuss the analysis of the user growth strategy using data from the daily deal market. Daily deal platforms such as Groupon emerged around 2008 as two-sided markets connecting merchants and consumers with discounted deals. Figure 3 shows the growth of Groupon’s consumer base, profit, revenue, and market value over the past ten years. Groupon experienced rapid convex growth in its first several years, before reaching 2013 revenues of $2.6 billion, but then it almost stopped growing. As a forecast of future growth, its market value experienced a sharp decline in 2012. Analysts observed that the consumer base and revenues had a highly correlated trend. For years, a central question for Groupon senior management was: Should Groupon further promote the growth of its user base? This question makes Groupon a valuable setting for our research. The daily deal market is ideal to answer our research questions for several rea- 25 sons. The market was largely dominated by two leading platforms: Groupon and LivingSocial, making it easier to control for the competition e↵ects. Around our data collection period, Groupon and LivingSocial made up roughly 59% and 17% of the total revenue in the U.S. daily deal market, respectively.6 In our estimation, we use Groupon as the focal platform to estimate the parameters of interest and use LivingSocial to control for competition. Perhaps more importantly, it is critical for us to examine how SS 1 and CS0 vary by product categories (especially between experience goods and search goods) after the overall quality and brand name of the platform are controlled for. Comparing di↵erent platforms would face tremendous identification challenges, because the nature of products o↵ered by di↵erent platforms would be indistinguishable from the intrinsic quality of the platforms. Fortunately, Groupon o↵ers products from both experience and search categories. By leveraging the panel data structure across categories and markets, we can estimate the e↵ects of users after eliminating the e↵ects attributable to platform quality or brand name. Our sample includes Groupon’s largest 108 markets from January 2012 to December 2012. Table 4 provides summary statistics by market and week. Note that the platform’s consumer base is not directly observed in this study. In the estimation, it is represented using the transaction volume. Using transaction volume has an advantage of capturing the number of active users rather than dormant users who did not make a purchase. Compared with the number of registered accounts, transaction volume is better aligned with the decision-making process underlying Equations 2 and 5, because consumers and merchants can observe the transaction 6 Statistica 2013, http://www.statista.com/statistics/322293/groupon-market-share-us/. 26 Figure 3: Groupon Growth from 2010-2019. Bars Show Observation Period volume but not the number of inactive consumers. Table 4: Descriptive Statistics by Market and Week 4.2 Variable Obs Mean SD Consumer base Merchant base Average deal price ($) Rival’s consumer base Rival’s merchant base Returning merchant indicator Deal duration (days) 5,707 4,685.12 6,085.98 5,707 22.4 17.19 5,707 78.06 53.94 5,707 2,828.47 8,935.2 5,707 10.9 9.16 5,707 0.59 0.2 5,707 4.15 0.97 Min Max 4 1 11 0 0 0 0 111,820 130 770 425,030 57 1 13.54 Estimation and Identification In this section, we present the estimation and identification strategy for the intertemporal and within-period network e↵ects. 27 4.2.1 Identification for the Inter-temporal Network E↵ect We begin by discussing the estimation for the SS 1 . As shown in Equations 2 and 5, the SS 1 is captured by the lag(s) of the consumer base and merchant base. As the dependent variable, i.e., the logarithm of relative market share, is also a function of the user base, our specification is a variant of the dynamic panel linear model. This specification has the advantage of controlling for heterogeneity and allowing the distinction between the short-run and long-run dynamics, but calls for special attention on estimation. It is well known that the ordinary least square (OLS) estimator for the lagged user base (i.e., ln nj,t 1 1 and 2) is biased in dynamic panel linear models. By construction, in Equation 2 and ln mj,t 1 in Equation 5 are correlated with the market fixed e↵ects. To eliminate the unobserved market e↵ects, we apply the first-di↵erence approach proposed by Anderson and Hsiao (1981) and Arellano and Bond (1991). Take Equation 2 as an example. After the first-di↵erence transformation, it becomes n ln Yj,t = 1 ln nj,t n where Yj,t = snjt /sn0jt , 1+ 1 ln mjt +⇢1 ln pjt +✓1 n n ln Yj,t = ln Yj,t n ln Yj,t 1, ln Rjt +⌘1 Xjt + ln nj,t 1 = ln nj,t n Ttn + ⇠jt , 1 1 ln nj,t 2 and so on. The error terms are assumed to have zero serial correlations for the same n n cross-sectional unit, i.e., E(⇠jt ⇠jt0 ) = 0 for all t 6= t0 7 . Furthermore, the levels of the dependent variable lagged two periods or more are valid instruments in the equation of first-di↵erences (Arellano and Bond, 1991). 7 The test results for the auto-correlation assumption is presented in the results section. 28 The identification restriction is specified as E(ln nj,t p n ⇠jt ) = 0, where p 2 {2, 3, .., t 1}; t 2 {3, 4, ..., T }. The instruments are used to form the objective function for the generalized method of moments (GMM) (Hansen, 1982; Wooldridge, 2010), forming the Di↵erence GMM (DGMM) estimator (Arellano and Bover, 1995). 4.2.2 Identification for the Within-period Network E↵ect The identification challenge underlying the parameters for the classic network effect is that the consumer base and merchant base are simultaneously determined in Equations 2 and 5, causing an endogeneity concern. We address this problem by providing instrumental variables for ln njt and ln mjt , respectively. We first present the instrumental variables for the merchant base ln mjt in Equation 2. A valid instrument should be correlated with the number of merchants but n orthogonal to the demand shock ⇠jt . We used two sets of instruments here. First, following the pre-determined variable in Arellano and Bond (1991), we use the lag of merchant size (after logarithm transformation) as the instrument. The argument is that the demand shock to consumer base in period t should not be correlated with the merchant size in period t 2, after controlling for all the other variables in the model. Second, we also leverage political advertising as an exogenous variation for identification. Our data collection year, 2012, happened to be a presidential election year in the U.S.. During an election year, political candidates and interest groups (including party committees and outside political action groups known as PACs and 29 superPACs) invest heavily in television advertising, leading to an increase in ad prices (Moshary et al., 2021). The increased cost of advertising leads merchants to look for alternative marketing avenues such as participating in daily deals (Dholakia, 2011; Edelman et al., 2016). Thus, the number of merchants on deal platforms should be correlated with the amount of political election advertising in the market.8 We collect data on political advertisements across four types of elections that year (presidential, gubernatorial, House, and Senate elections) (Fowler and Ridout, 2015), and compute the total amount of air time and ad spending across all candidates and parties in each market j during week t.9 In this empirical setting, it often takes some time for the merchants to negotiate with the platform to determine the terms (Li et al., 2017). Thus, we use the political advertising in period t 1 as the instrument because such a specification captures the time lag associated with the merchant base in t. The lagged advertising variable also helps the exogenous assumption in that past political advertising should not directly a↵ect the current n deal demand ⇠jt . We follow a similar identification strategy for the consumer base ln njt in Equation 5. We use two sets of instruments: (1) the lag of consumer base with a degree 8 In our data, we find positive correlations between merchant base and political advertising (the Pearson correlation coefficient was 0.19 (p<0.001) with ad spending and 0.05 (p<0.001) with air time, indicating that there was a positive association between the number of merchants on Groupon and the amount of TV election ads in the market. We provide the detailed first-stage analysis in the results section. 9 The data on political advertising are at the designated market area (DMA) level. We map the advertising data onto the Groupon divisions (i.e., markets). The markets that fall into the same DMA have the same value for the advertising variables. Including ads from gubernatorial, House, and Senate elections helped address the unequal allocation of campaign resources between battleground and non-battleground states for presidential elections. In our data, only 5 (4.7%) DMAs received zero political advertising in 2012 and the remaining 95.3% all had some levels of election advertising on TV. 30 of 2, and (2) the precipitation and temperature in each market. The identification argument for precipitation and temperature is based on the finding that consumers’ online shopping behaviors are shown to be influenced by weather conditions. Using data from large-scale field experiments, Li et al. (2017) provided robust evidence that consumer online shopping behaviors deviate significantly on rainy days from sunny days, because of the impacts of weather on the mood and psychological states of consumers. Based on these, we argue that the local weekly temperature and precipitation should correlate with consumers’ intention to shop Groupon deals, providing the first-stage variation for identification. However, precipitation and temperature at m time t should not be correlated with the supply-side error ⇠j,t , because the merchant base has been determined prior to t due to the processing and negotiation time gap required for merchants to show up on the platform. Thus, the exogenous requirement for using precipitation and temperature as the instrumental variable is met. 4.2.3 Other Variables Market size. Our specification requires the “size” for each market. On the consumer side, the market size is defined as the total number of users who could participate in the daily deal market. Because anyone with Internet access could use a deal site, we use the number of Internet users as the measure of market size. The data are retrieved from the “October 2012 School Enrollment and Internet Use Survey,” a supplement to the Current Population Survey (CPS) by the U.S. Census Bureau. On the merchant size, we measure the market size using the number of businesses in each local market in 2012, which is extracted from the County Business Patterns 31 (CBP) database from the the U.S. Census Bureau 10 . This data source tracks the number of businesses in each market along with the NAICS industry code and a short description. We matched the NAICS with the merchant types in Groupon and LivingSocial and selected the business categories that have participated in the daily deal industry. Experience versus Search Goods. We code product categories on Groupon into experience goods and search goods based on the nature of the product and services. Following Li and Wu (2018), we administered an online survey to classify whether a good or service possesses more experience attributes or search attributes. A total of 818 respondents were asked to imagine that “user reviews are not available” on a platform and then asked to assess for each category, “to what extent it is easy (hard) to evaluate the quality without seeing or trying it.”11 The score closer to “easy” (coded as 1) corresponds to more of a search good, and vice versa for experience goods if the score is closer to “hard” (coded as 7) (Nelson, 1974). To ensure that respondents can finish the survey within 5 minutes, each respondent answered a random selection of 15 categories (on average, each category was rated by 136.3 respondents). Table 5 presents the summary statistics from the survey. Across the total of 78 categories, the mean score is 4.72 (Standard Deviation = 1.02) and the median is 4.97. We used the median split and code each category into two groups: experience 10 https://www.census.gov/programs-surveys/cbp/data/datasets.html. Accessed Feb 2021. A total of 1,112 respondents were recruited on Amazon MTurk to participate the survey. 294 (26%) respondents failed the attention filler questions and hence were removed from the analysis. The detailed survey design and summary statistics are included in the Appendix. 11 32 goods if the average score of that category is greater than the median and vice versa for search goods. Among our categories, the highest i.e. experience goods scores were “hair salons,” “facials,” and “makeup services,” while those with the lowest, i.e. search goods, scores were “holiday decor,” “gifts: candles, phone cases, stationary,” and “office supplies.” Note that the mean split yielded very similar results since only two categories in the middle would switch groups. In subsequent analysis, we used the MTurk responses to classify the categories. Table 5: Summary Description of Amazon MTurk Online Survey N Mean All categories 78 4.72 Experience Goods 39 Search Goods 39 SD Median Min Max 1.02 4.97 1.98 6.39 5.59 0.36 5.51 5.01 6.39 3.85 0.65 3.97 1.98 4.94 To cross-check our experience-versus-search goods classification, we conducted an o✏ine survey using a di↵erent design. Twenty-eight business-school graduate students at a major U.S. university participated in the survey. Each student was asked to code the product/service categories on Groupon using a binary response: the “search” type (denoted as Ei = 0) or the “experience” type (denoted as Ei = 1). Across all respondents, the Fleiss’ Kappa is 0.603, indicating a moderate level of average inter-coder reliability (Fleiss, 1971). For each category, we used the majority voting: a category was classified as the experience type if more than half of the respondents coded Ei = 1. Out of the categories, 8 (10.3%) were coded as search goods by MTurk respondents but switched to experience type in the validation survey; 5 (6.4%) switched from experience to search type; and the rest 65 (83.3%) were consis33 tent between the two surveys. We estimated the main models based on the validation survey and obtained qualitatively similar results. The detailed model estimates are presented in the appendix. Price. Lastly, deal price in Equations 2 and 5 could also be endogenous. A rational platform would strategically adjust the deal price in response to an expected shock in demand and supply. For example, if the demand is expected to be low, a platform would be incentivized to reduce the price. In the same vein, if the supply is expected to be strong, the platform would also be incentivized to adjust prices to increase profits. To address this issue, we use the lagged deal prices as the instrument for pjt , as proposed and tested in Villas-Boas and Winer (1999). The identification assumption is that, the prices from the previous periods should be correlated with the current price due to the common cost factors. But after controlling for the market fixed e↵ects and other variables, they should be exogenous to the demand and supply shock in the current period. We consider this assumption reasonable in the Groupon setting, because many merchants on the platform provide services to local customers and thus the cost of service within the market should be correlated over time. However, Groupon typically feature di↵erent merchants in consecutive weeks and thus the lagged prices should be uncorrelated with the current demand or supply shock. 34