Network effects on the iPhone platform: An empirical examination

Network effects on the iPhone platform: An empirical examination

Telecommunications Policy 39 (2015) 877–895 Contents lists available at ScienceDirect Telecommunications Policy URL: www.elsevier.com/locate/telpol ...

388KB Sizes 3 Downloads 70 Views

Telecommunications Policy 39 (2015) 877–895

Contents lists available at ScienceDirect

Telecommunications Policy URL: www.elsevier.com/locate/telpol

Network effects on the iPhone platform: An empirical examination Daniel D. Garcia-Swartz a,n, Florencia Garcia-Vicente b a b

Charles River Associates, 1 South Wacker Drive, # 3400, Chicago, IL 60606, USA Belly Chicago, 600 West Chicago Avenue, Chicago, IL 60654, USA

a r t i c l e i n f o

abstract

Available online 21 August 2015

Smartphones are “systems,” that is, collections of (hardware and software) components that work well with one another. They can also be viewed as multi-sided platforms that connect two or more distinct types of customers (users, application developers, handset manufacturers, network operators, and advertisers) and are characterized by the existence of indirect network effects—an extra customer on one side generates benefits for customers on the other side. In this study we focus on one specific aspect of the growth dynamics of Apple's iPhone: the feedback effects between users and applications (or application developers). After briefly tracing the history of the iPhone, we construct time series of users and apps. We then carry out cointegration analysis and estimate dynamic OLS (DOLS) and vector error correction models (VECM). We find evidence of indirect network effects of a substantial magnitude. & 2015 Elsevier Ltd. All rights reserved.

Keywords: iPhone Multisided platforms Indirect network effects Cointegration Dynamic OLS Vector error correction models

1. Introduction Smartphones are “systems,” that is, collections of components connected through an interface that work well with one another (Farrell & Klemperer, 2007; Katz & Shapiro, 1994). They are also multi-sided platforms, which are platforms with three characteristics: first, they serve two or more distinct types of customers; second, there are indirect externalities arising from the interconnection among different customer types; and third, an intermediary is required to internalize the externalities flowing from one side of the platform to the other. The early papers on multisided platforms include Schmalensee (2002), Caillaud and Jullien (2003), and Rochet and Tirole (2003). Later contributions include Parker and Van Alstyne (2005) and Armstrong (2006), among others. Evans (2003) explores the connections between multisided platforms and antitrust. Rochet and Tirole (2006) and Rysman (2009), among others, provide summaries of the key issues raised in the literature. For a recent discussion see Hagiu and Wright (2015).1 The smartphone's “customers” are users, application developers, handset makers, network operators, and advertisers. The indirect externalities are particularly obvious for some customer pairs, including users and application developers: additional users enhance the platform's monetization

n

Corresponding author. E-mail address: [email protected] (F. Garcia-Vicente). 1 An early discussion of the economics of networks more generally (and network externalities more specifically) can be found in Economides (1996). Farrell and Klemperer (2007) analyze network effects within the more general context of coordination and lock-in. http://dx.doi.org/10.1016/j.telpol.2015.07.011 0308-5961/& 2015 Elsevier Ltd. All rights reserved.

878

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

opportunities and thus make it more attractive for developers, and at the same time additional developers (and applications) make the platform more useful and thus enhance its attractiveness for users.2 Recently a few studies have attempted to interpret the evolution of the smartphone market in light of economic and management theory, and some of those studies have relied on a version of platform theory. Among recent studies of the industry, West and Mace (2010) examine browsing as the “killer app” that explains the rise of Apple's iPhone; Kenney and Pon (2011) analyze the strategies of several smartphone vendors in light of the technological platform theory; CampbellKelly, Garcia-Swartz, Lam, and Yang (in press) present a comprehensive examination of the smartphone industry against the backdrop of multi-sided platform theory; Pon, Seppälä and Kenney (2014) focus on the strategies of Google, Amazon, and Xiaomi within the context of the Android platform; and Cecere, Corrocher, and Battaglia (2015) raise the question whether a dominant design has arisen in the smartphone industry. None of these studies, however, attempt to quantify the role of indirect network effects on a specific smartphone platform. There is also a recent and highly innovative empirical literature on mobile applications. Bresnahan, Davis, and Yin (2013) use data on smartphone users and apps to study the industry from the perspective of application developers. Yin, Davis, and Muzyrya (2013) rely on data from the iPhone app ecosystem to examine how the development of “killer apps” depends on the nature of the market and on the characteristics of specific applications. Davis, Muzyrya, and Yin (2013) examine data on firms producing iPhone applications to analyze the links between experimentation strategies and entrepreneurial innovation. Although this literature is unique in its creative use of highly detailed data on apps and developers, to our knowledge it has not made an attempt at quantifying indirect network effects. There is also a recent literature that focuses on estimating demand for mobile applications. A key study in this literature is that by Ghose and Han (2014), in which a structural econometric model of demand for mobile apps is built and estimated with the goal of uncovering consumer preferences for various app characteristics. The authors find that app demand increases with the in-app purchase option and decreases with the in-app advertisement option. Again, although this study is highly innovative, it makes no attempt to assess the strength of indirect network effects. Our study views smartphones as both dynamic systems and multi-sided platforms. More specifically, we focus on one such platform—Apple's iPhone—and make an attempt at providing a quantitative assessment of the role of the indirect feedback effects between users, on the one hand, and applications (or application developers), on the other. We collect timeseries information on the various sides of the iPhone platform and analyze the interactions between users and applications relying on the toolkit of time-series econometrics. More precisely, we carry out cointegration analysis and estimate dynamic OLS (DOLS) and vector error correction models (VECM). Our study is organized as follows. Section 2 presents background information on the origins and growth of Apple's iPhone. Section 3 summarizes the literature on network effects with a specific focus on empirical studies. This literature review lays the foundation for the rest of our study, since in this section we highlight the two functions—hardware demand and software supply —that are at the core of most studies of indirect network effects. In Section 4 we present the key results from the econometric analysis: we carry out tests of cointegration for the iPhone's users and apps and estimate dynamic OLS (DOLS) and vector error correction models (VECM) to determine the magnitude of the indirect network effects. Section 5 concludes. In three appendices we provide additional information on network-effects studies, hedonic price indices for smartphones, and robustness checks. Appendix A summarizes the key features of various empirical studies on network effects in platform industries. Appendix B presents details on the estimation of hedonic models for smartphones—we use the results derived from these models to construct hedonic price indices for smartphones, and then use the iPhone's hedonic price index in the estimation of some of the econometric models presented in the core of this paper. Appendix C, finally, presents one additional robustness check for one of the models estimated in this study: we compare the magnitude of indirect network effects obtained from OLS, DOLS, and instrumental-variable models. We believe our study makes a number of contributions. First, we collect and analyze a set of time series that provides a comprehensive overview of the growth of the iPhone as an evolving system. Second, in the process of studying the growth dynamics of smartphones, we construct hedonic price indices for Apple (and Android) smartphones through 2011. Third, we present what to our knowledge is the first quantitative investigation of indirect network effects on the iPhone platform. Fourth, we show that indirect network effects do play an important role in the growth of the system and provide an econometrically-based assessment of that role.

2. Background on the origins and growth of the iPhone The smartphone market changed radically with the arrival of Apple's iPhone in June 2007 (Campbell-Kelly et al., in press; Campbell-Kelly & Garcia-Swartz, 2015, chap 11; Vogelstein, 2008; Vogelstein, 2013). Apple's management started exploring the development of an Apple phone in 2002, soon after the introduction of the iPod MP3 player. One of the key problems consumers were facing in the early 2000s, and especially after the iPod was introduced, was mobile device proliferation— people usually carried a PDA, a phone, an MP3 player, and sometimes a digital camera as well. Some attractive products had 2 This description is appropriate for a platform that has proprietary software at its core. Open-source software platforms have many of the same features, but sometimes may have a not-for-profit institution playing the role of the intermediary. Further, the “monetization” arrangements vary considerably in an open-source software platform.

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

879

started appearing in the market offering many functions in one device—the Palm Treo 600, for example, became extremely popular by combining a phone, a PDA, and email capabilities. By 2004 the iPod business, which accounted for about 16 percent of Apple's annual revenues, looked promising but also vulnerable, since mobile phones were becoming increasingly sophisticated and the price of storage was plunging. Further, music stores rivaling Apple's iTunes, which had been introduced in April 2003, were proliferating. Apple responded with a first “music phone,” the ROKR, jointly developed with Motorola and introduced in 2005. The ROKR—which had an unfriendly interface, could hold no more than 100 songs, and had to be synced with a personal computer to complete a purchase from the iTunes store—failed to take off. As the ROKR episode was unfolding, Apple established a business relationship with Cingular, the wireless carrier. Apple and Cingular—which was acquired by AT&T in December 2006—held prolonged negotiations that led to a business arrangement centered on the introduction of an Apple phone, internally known as Purple 2. Apple's iPhone, developed in an atmosphere of complete secrecy in 2005–2006, was announced by Steve Jobs in January 2007. It became commercially available in June of that year. Jobs presented the iPhone as a “revolutionary mobile phone” that combined the capabilities of a phone, an iPod, and an “Internet communicator” (Thomas, 2007). The entire front surface of the iPhone was a touchscreen and all of its functions were activated by touch. It ran the Mac OS X operating system (later rebranded iOS when used on the iPhone), offered a radically new and improved Internet access capability, and was able to wirelessly download music and movies from the iTunes store. In the second quarter of 2007, when the iPhone became commercially available, Symbian was the undisputed leader in the world smartphone market with a share of almost 66 percent of all units shipped (Gartner Research, 2008; Markoff, 2007). The iPhone's share of world smartphone shipments grew quickly, from less than one percent at the time of introduction to about 11 percent in late 2008. The rise of the iPhone was accompanied by a steep decline in the share of both Symbian and Windows CE, while the share of the Blackberry OS kept on growing through the end of 2009. The App Store opened for business in July 2008. A new version of the iPhone, the iPhone 3G, was released in that month as well (with App Store support). The number of third-party apps available in the App Store grew quickly, as did app downloads, to the point that by early 2009 about 500 million downloads had already taken place. By late 2008 the iPhone was well on its way to becoming an important player in the world smartphone market. Its ascent, however, did not remain uncontested for long, since in September 2008T-Mobile released the first smartphone running on Android, the operating system developed by Google (Lendino, 2012). This first Android smartphone was reviewed poorly and did not sell well (Vogelstein, 2011). The first successful Android smartphone was Motorola's DROID, which reached the market in October 2009 (McCracken, 2009). The battle between Apple and Google soon centered on attracting app developers to each platform. In March 2008, before the opening of the App Store, Apple unveiled the software development kit for the iPhone, and the venture-capital firm Kleiner Perkins Caufield & Byers set up the iFund, a $100-million fund to invest in “game-changing applications” for Apple's smartphone (Olsen, 2008a). Google responded with the $10-million Android Developer Challenge (Olsen, 2008b). The share of Android in world smartphone shipments grew quickly from late 2009 on. By early 2011 Android had become the leading operating system in the world market with a share of about 36 percent of all shipments, followed by Symbian, Apple's iOS, the BlackBerry OS, and Windows Mobile.

3. Empirical studies on network effects Over the last couple of decades, several studies have attempted to quantify the strength of (direct and indirect) network effects in various platform industries. We provide a reasonably comprehensive, although by no means exhaustive, survey of these studies in Appendix A. The industries and markets covered in these studies include personal-computer hardware and software, personal digital assistants, home video game systems, and DVD systems. A subset of these studies has focused on network effects in telecommunications markets, although none of them have focused on the iPhone specifically. A review of the empirical network effects studies that excludes the most recent articles can be found in Gandal (2005). The published empirical network-effect studies are remarkably heterogeneous. Some of them rely on structural models to derive a set of estimating equations, but many of them estimate reduced-form models. Several of them use either timeseries data or a single cross section of data to estimate the magnitude of the network effects, and a few of them rely on panel data sets, which not only are much richer in terms of the information included but also allow for the implementation of identification strategies that are not feasible in a single cross section. Leaving aside the heterogeneity in modeling approach (structural versus reduced-form) and in the econometric techniques used, several studies end up estimating a version of what could be described as a hardware-demand function and a software-supply function. Strictly speaking, the core product in the system is not always hardware—it could well be operating-system software. Thus, the hardware-demand function can be better expressed as a core-product (CP) function of the following general form: COMP CP Q CP Þ t ¼ f ðP t ; Q t

ð1Þ

880

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

Here the quantity demanded of the core product at a given point in time, Q CP t , is a function of the price of the core product in that time period, P CP t , and the quantity of complements for the core product available in the marketplace in that . From theory, the expectation is, of course, that time period, Q COMP t

∂Q CP t o0 ∂P CP t

∂Q CP

t and ∂Q COMP 40. That is, the quantity demanded of t

the core product increases as the (potentially quality-adjusted) price of the core product declines and as the quantity of complementary products increases. The theory is not entirely precise as to whether the relevant quantities are stocks or flows, and different studies rely on different interpretations of the quantities that matter. Furthermore, several studies point out that the key right-hand-side variables may be endogenous, which usually leads to the search for an econometric technique—frequently instrumental variables—that will address the problem. Since the quality of the core product may be changing dramatically over time, some studies construct a quality-adjusted price via a hedonic approach and use it in the estimation of the model. One study that implements such an approach is that by Gandal, Kende and Rob (2000), which focuses on compact disc players. The so-called software-supply function is more precisely expressed as a complement-supply function of the following general form: Q COMP ¼ gðQ CP t t Þ

ð2Þ

That is, the supply of complements for the core product at a given point in time, Q COMP , is a function of the quantity of the t core product (installed base) that is available in the marketplace in that time period, Q CP t . From theory, the expectation is that

∂Q COMP t 4 0. ∂Q CP t

That is, a larger installed base of the core product provides stronger incentives for developers to supply

complementary products for the core product. Some measure of the cost of creating the complementary products is sometimes included—as in Gandal et al. (2000)—and sometimes not. In time-series models, inclusion only makes a difference if there is evidence that the relevant costs have changed dramatically over time. Some models do not include the cost of creating complementary products but rather the price of those products. Here again, there may be endogeneity issues affecting the estimation of the coefficient on the Q CP t variable, which sometimes leads to a search for instrumental variables. In their study of video games, Clements and Ohashi (2005) include a vintage effect on the right hand side of the software-supply equation but other studies do not. Another set of studies takes a slightly different approach and estimates a set of equations in which the emphasis is mostly on indirect feedback effects (and changes in costs or prices are basically disregarded). For example, Gandal, Greenstein, and Salant (1999) estimate a vector autoregressive model (VAR) involving operating system software for microcomputers and complementary software products, whereas Stremersch, Tellis, Franses, and Binken (2007) estimate a similar model for a variety of industries including black-and-white and color television. The Stremersch et al. (2007) study does include a core-product price in the relevant equation, but the authors acknowledge that their constructed price series has serious limitations for several industries. Further, they do not include a price variable in the complementary-product equation and they point out that this price is usually of much lesser importance in the context of multisided platform models. We estimate a variety of models to gauge the strength of indirect network effects on the iPhone platform. Our starting point is a model in the spirit of that presented in Gandal et al. (1999). The estimation of a VAR model assumes that the relevant time series are stationary. Thus, rather than proceeding to estimate a VAR directly, we start off by testing whether there is evidence that the relevant series are indeed stationary or not. We find that, for the most part, they are not. Having rejected the stationarity assumption for most of the relevant time series, there are two paths that economists take. The traditional wisdom used to be that non-stationary time series can be used in first differences, since differencing makes them stationary. It is now well established that it is better to first test for cointegration of the original series, since it is also well established that differencing time series that are cointegrated leads to a misspecification error (see, for example, Enders, 2010, p. 356 and more generally chapter 6). Testing for cointegration is equivalent to testing for the existence of a long-run, equilibrium relationship among the time series at stake. We thus proceed to test for cointegration and find that there is a cointegrating relationship among several of the time series we focus on. Having found cointegration, we know we can estimate the relevant relationships via OLS but we also know that more efficient estimators exist, including DOLS and VECM. Thus, we rely on these approaches to estimate the relevant cointegrating relationships. But we also go one step further. Since we suspect that the decline in the price of the iPhone may have been partially responsible for encouraging adoption of the device among consumers, we construct a quality-adjusted price for the iPhone via a hedonic approach. Then we incorporate this quality-adjusted price in the core-product demand equation and we reestimate the strength of the indirect network effects with the price variable in the model. In the presence of a cointegrating relationship both OLS and DOLS are consistent estimators of the long-term equilibrium relationship between iPhone users and applications. But in order to provide an additional test of robustness we also estimate the core-product equation via instrumental variables. Thus, our study provides a whole battery of estimates of the strength of the indirect network effects that are at work on the iPhone platform.

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

881

4. Time-series analysis: iPhone users and applications The focus of our study is providing a quantitative assessment of the role of indirect network effects on the iPhone platform. We would like to determine whether there is a long-term relationship between iPhone users and applications, or application developers. If that relationship exists, we would like to provide a range of estimates for the magnitude of the indirect network effects. Network-effects studies based on panel data usually rely on a model of the demand for the core product of the following form: uj;t ¼ β 0 þxj βx þ βp pj;t þ γ Nj;t þ ξj;t þ εj;t

ð3Þ

In this model it is assumed that a representative individual (or household) maximizes a utility function uj; t , which is the utility derived from consuming the services of core product j at time t. The individual (or household) maximizes such a utility function by choosing core product j at time t among J t þ 1 alternatives, one of which is the option not to consume these products at all. If we had repeated observations on smartphone models over time, j would index phone model, and we would have to consider the possibility that individuals might choose not to use any smartphone at all. In Eq. (3), xj is a vector of observed characteristics of product j, pj; t is the price of product j at time t, and Nj; t is the quantity of complementary products available for core product j at time t. It is often assumed that xj β x captures the average benefit from the product's technology, and ξj; t is an error term that captures the deviation from the average (with Eðξj; t Þ ¼ 0). The error term also accounts for other factors that lead consumers to purchase a specific core product and that are not observable in the data. This model usually serves as the foundation for the estimation of a logit or nested-logit model along the lines of Berry (1994) and Berry, Levinsohn, and Pakes (1995). As far as the supply of complementary products is concerned, researchers often assume the existence of single-product firms that supply their complementary products for core-product j. Other assumptions sometimes include increasing returns to scale in, and free entry into, the creation of complementary products, as well as aCES demand for complementary products on the part of individuals (or households) who own core product j (see, for example, Clements and Ohashi (2005)). The symmetric Bertrand equilibrium in this model determines the available variety of complementary products as follows: Nj; t ¼ Aj ðIBj; t Þσ

ð4Þ

In this model, the quantity of complementary products available for core product j at time t is, in essence, a function of the installed base ðIBÞ of the core product at that time. Our data have limitations that do not allow us to estimate a full-blown version of this model. We do not a have complete panel dataset available on all smartphone models over time. Although the logit or nested-logit models of core-product demand can be estimated with a single cross section of data, the estimation of such models requires the availability of coreproduct quantity data for all relevant products on the market, which are not available to us. Thus, in essence we use time-series data to estimate reduced-form versions of a core-product demand equation and of a complementary-products supply equation—with the iPhone as the core product and the iPhone apps as the complementary products. More precisely, we estimate the following core-product demand equation for the iPhone: COMP CP lnðQ CP Þ þ ν1; t t Þ ¼ α1 þ β 1 lnðP t Þ þ δ1 lnðQ t

ð5Þ

And we estimate the following complementary-products supply equation: lnðQ COMP Þ ¼ α2 þ δ2 lnðQ CP t t Þ þ ν2; t

ð6Þ

About four decades ago Granger and Newbold (1974) alerted us of the possibility of estimating what they called spurious regressions, that is, of finding statistically significant relationships among variables that are, in reality, unrelated. Spurious regressions are likely to arise when estimating relationships among variables that contain stochastic trends (that is, movements generated by the cumulative effects of stochastic changes). In light of the stochastic-trend literature launched by the Granger-Newbold (1974) contribution, it is standard practice to follow a sequence of steps when estimating time–series relationships among variables (see, for example, Heij, de Boer, Franses, Kloek, & van Dijk, 2004, pp. 673–674). First, we must check whether the variables at stake are stationary (or the alternative, whether they contain stochastic trends). Second, if they are stationary, the relationships among the variables can be estimated via OLS. Third, if alternatively they contain stochastic trends, we must check whether the variables are cointegrated, that is, whether they share the same stochastic trend. (That non-stationary variables are cointegrated means that there is at least one linear combination of them that is stationary.) Fourth, if they are not cointegrated, we need to check whether the first differences of the original variables are stationary and, if they are, estimate the relationship in first differences (keeping in mind that the meaning of the estimated relationship has changed). Fifth, if alternatively the variables are cointegrated, we can estimate the relationships among them either via the Dynamic OLS (DOLS) estimator developed by Stock and Watson (1993) or the vector error correction model (VECM) developed by Johansen (1988) and Johansen and Juselius (1990). We start off with an approach inspired by the Gandal et al. (1999) study, which examines the evolution of the CP/M operating system. That is, we first analyze purely the relationships between iPhone shipments (as a proxy for iPhone users) and iPhone app submissions (as a proxy for the developer side of the business). However, we go one step further and take

882

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

the evolution of iPhone prices into account. More specifically, we construct a hedonic price index for the iPhone and incorporate the iPhone's hedonic price index in some of our estimated models. This is another contribution of our study, since we know of no previous attempt at constructing hedonic price indices for the iPhone, or for smartphones more generally. In order to have a benchmark with which to compare the evolution of the iPhone's hedonic price, we also construct a hedonic price index for Android smartphones through 2011. The data and methodology for the elaboration of the hedonic price indices are presented in Appendix B.

4.1. Data The core data for our investigation come from a variety of sources. As far as smartphone users are concerned, we have Gartner-Research data on world quarterly shipments of smartphones (in units) for the key operating systems starting in early 2007. We use these data on the quarterly flow of smartphones to construct various measures of the stock of smartphones by operating system at various points in time. Our focus is on the iPhone, which did not exist before 2007. This has two implications for the construction of a measure of the stock of iPhones. First, we can safely start with a stock of zero in early 2007. Second, a measure of the world stock of iPhones at every point in time only requires an assumption about what proportion of the flow of new iPhones available on the market every quarter replaces a portion of the pre-existing stock. Since we do not want our results to depend on an assumption such as this one, we work with three different measures of the stock—one assumes that the replacement rate is 50 percent, another one that it is 40 percent, and a third one that it is 30 percent. (We assume that replacement of old iPhones with new ones does not begin before early 2009.) Table 1 presents the data on world shipments of smartphones for all operating systems. Table 2 focuses on the iPhone: it shows the quarterly shipment figures as well as the three measures of the stock based on the replacement-rate assumptions just mentioned. The “Stock 50” column assumes a replacement rate of 50 percent, the “Stock 40” column a replacement rate of 40 percent, and the “Stock 30” column a replacement rate of 30 percent. The other side of the market involves developers and applications. We do not have time–series data on the number of developers creating apps for the iPhone. We know that, as of late 2013, there were almost one million active applications on Apple's App Store and more than 262,000 active publishers (developers) associated with the App Store. We do have monthly information from the 148apps.biz site on both the flow of application submissions to the App Store and the total stock of active applications available on the App Store. We also have analogous information from the pocketgamer.biz site and we can compare the two sources for consistency. Further, on the basis of information contained in a variety of trade-press articles (available from the authors upon request) we have constructed an alternative time series for the total stock of applications available on the App Store at various points in time (and the total number of app downloads up to each point in time). Table 3 presents the data on the flow of application submissions and Table 4 the data on the stock Table 1 Worldwide smartphone sales to end-users (operating systems), volume in 000. Source: Gartner press releases.

Q1-2007 Q2-2007 Q3-2007 Q4-2007 Q1-2008 Q2-2008 Q3-2008 Q4-2008 Q1-2009 Q2-2009 Q3-2009 Q4-2009 Q1-2010 Q2-2010 Q3-2010 Q4-2010 Q1-2011 Q2-2011 Q3-2011 Q4-2011 Q1-2012 Q2-2012 Q3-2012 Q4-2012 Q1-2013

Symbian

Android

RIM

Windows

iPhone

Other

Total

15,844 18,273 20,664 22,903 18,400 18,405 18,179 17,949 17,825 20,881 18,314 23,858 24,068 25,387 29,480 32,642 27,599 23,852 19,500 17,458 12,467 9072 4405 2569 1349

0 0 0 0 0 0 0 0 575 756 1425 4043 5227 10,653 20,500 30,845 36,268 46,776 60,490 75,906 81,067 98,529 122,480 144,720 156,186

2080 2471 3192 4025 4312 5594 5800 7443 7534 7782 8523 10,508 10,753 11,629 11,908 13,162 13,004 12,652 12,701 13,185 9939 7991 8947 7333 6219

2931 3212 4180 4374 3858 3874 4053 4714 3739 3830 3260 4203 3696 3059 2248 3375 3659 1724 1702 2759 2713 4087 4058 6186 5989

0 270 1104 1928 1726 893 4720 4079 3848 5325 7040 8676 8360 8743 13,484 16,011 16,883 19,629 17,295 35,456 33,121 28,935 23,550 43,457 38,332

4087 3628 3612 3536 4113 3456 3763 3958 2986 2398 2531 2517 2403 2588 2912 5212 3357 3106 3497 4278 5085 5072 5738 3397 1971

24,943 27,855 32,752 36,766 32,408 32,221 36,515 38,143 36,507 40,972 41,093 53,804 54,506 62,058 80,533 101,248 100,769 107,739 115,185 149,042 144,392 153,686 169,179 207,662 210,046

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

883

of active applications on the App Store between mid-2008 and late 2013. Both tables are based on the information available on the 148apps.biz web site. For comparison purposes, Table 5 presents information (collected from a variety of sources) on apps available on the App Store at various points in time and the total number of apps downloaded from that store up to each point in time; it also includes analogous information for Android Market (more recently renamed Google Play). A perusal of the data reveals one interesting fact: the iPhone did not really take off before the opening of the App Store in mid-2008 (see also Vogelstein, 2013, p. 185). Leaving aside the buzz that the introduction of the first iPhone generated in mid-2007, shipments of the first-generation iPhone peaked at less than two million in late 2007 and declined thereafter.

Table 2 iPhone shipments and stock, volume in 000. Source: Shipment figures come from the Gartner Research press releases.

Q1-2007 Q2-2007 Q3-2007 Q4-2007 Q1-2008 Q2-2008 Q3-2008 Q4-2008 Q1-2009 Q2-2009 Q3-2009 Q4-2009 Q1-2010 Q2-2010 Q3-2010 Q4-2010 Q1-2011 Q2-2011 Q3-2011 Q4-2011 Q1-2012 Q2-2012 Q3-2012 Q4-2012 Q1-2013

iPhone shipments

Stock 50

Stock 40

Stock 30

0 270 1104 1928 1726 893 4720 4079 3848 5325 7040 8676 8360 8743 13,484 16,011 16,883 19,629 17,295 35,456 33,121 28,935 23,550 43,457 38,332

0 270 1374 3302 5028 5921 10,641 14,720 16,644 19,307 22,827 27,165 31,345 35,716 42,458 50,464 58,905 68,720 77,367 95,095 111,656 126,123 137,898 159,627 178,793

0 270 1374 3302 5028 5921 10,641 14,720 17,029 20,224 24,448 29,653 34,669 39,915 48,006 57,612 67,742 79,519 89,896 111,170 131,043 148,404 162,534 188,608 211,607

0 270 1374 3302 5028 5921 10,641 14,720 17,414 21,141 26,069 32,142 37,994 44,114 53,553 64,761 76,579 90,319 102,426 127,245 150,430 170,684 187,169 217,589 244,422

Table 3 App submissions to the App Store, by month, selected dates. Source: 148Apps.biz web site. Month

# Non-game apps

# Games

# Total

Average/day

2008-06 2008-09 2008-12 2009-03 2009-06 2009-09 2009-12 2010-03 2010-06 2010-09 2010-12 2011-03 2011-06 2011-09 2011-12 2012-03 2012-06 2012-09 2012-12 2013-03 2013-06 2013-09

11 1483 2973 5462 9121 10,857 22,619 17,482 17,287 15,816 22,783 18,212 19,266 18,741 20,294 22,333 19,564 23,641 23,548 24,729 22,624 25,782

3 465 817 1211 1569 2083 3384 2512 2700 2321 4393 3689 3473 3488 4752 4087 3566 4180 4892 4468 3916 4856

14 1948 3790 6673 10,690 12,940 26,003 19,994 19,987 18,137 27,176 21,901 22,739 22,229 25,046 26,420 23,130 27,821 28,440 29,197 26,540 30,638

0 63 122 215 345 417 839 645 645 585 877 706 734 717 808 852 746 897 917 942 856 988

884

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

There is clearly a qualitative jump in the number of shipments between Q2 2008 (less than one million) and Q3 2008 (almost five million). From then on, iPhone shipments and app submissions to Apple's App Store grow together, suggesting that the dynamics of a multi-sided platform is at work. 4.2. Results: stationarity We start off by testing whether the relevant variables—iPhone users and applications—are stationary, for which we rely on the Augmented Dickey–Fuller (ADF) test (Dickey & Fuller, 1979). There are other test statistics for this purpose, but Stock (1994) has shown that the ADF performs reasonably well relative to the alternatives. In addition, we also relied on the soTable 4 Active apps on the App Store, by month, selected dates. Source: 148Apps.biz web site. Month

# Non-game apps

# Games

# Total

2008-09 2008-12 2009-03 2009-06 2009-09 2009-12 2010-03 2010-06 2010-09 2010-12 2011-03 2011-06 2011-09 2011-12 2012-03 2012-06 2012-09 2012-12 2013-03 2013-06 2013-09

42,371 46,094 52,822 60,940 72,679 90,707 113,456 139,637 166,691 199,994 231,276 265,602 302,427 342,962 387,756 439,560 490,260 552,727 620,202 691,975 763,802

37,394 38,515 40,105 42,225 44,965 48,809 52,519 57,443 62,143 68,480 74,836 81,300 88,277 95,720 103,634 112,654 121,725 132,300 143,544 155,882 170,061

79,765 84,609 92,927 103,165 117,644 139,516 165,975 197,080 228,834 268,474 306,112 346,902 390,704 438,682 491,390 552,214 611,985 685,027 763,746 847,857 933,863

Table 5 Apps available on, and apps downloaded from, the App Store and Google Play. Sources: Multiple articles in the general and trade press. App Store Apps available Jul-08 Jan-09 Mar-09 Jun-09 Jul-09 Nov-09 Dec-09 Mar-10 Apr-10 Aug-10 Sep-10 Oct-10 Jan-11 Jun-11 Jul-11 Oct-11 Dec-11 Feb-12 Mar-12 Jun-12 Sep-12 Oct-12 Jan-13

Google Play Downloads (M)

800 15,000 25,000 50,000 65,000 100,000

10 500 800 1000 1500 2000

150,000 200,000

3000 4500

250,000 300,000 350,000 425,000 425,000 500,000

6500 7000 10,000 14,000 15,000 18,000

500,000 550,000 650,000 700,000

24,000 25,000 30,000 35,000

800,000

40,000

Apps available

Downloads (M)

2,300

16,000 30,000 38,000 80,000

1000

100,000

250,000 319,000 380,297 450,000 600,000 675,000 700,000

6000 10,000

20,000 25,000

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

885

called DF-GLS test developed by Elliott, Rothenberg, and Stock (1996), and found that the results were not radically different from those obtained with the more conventional ADF test. The null hypothesis in the ADF test is that the relevant variable contains a unit root (that is, it is not stationary), and the test takes the following form:

Δyt ¼ α þ βt þ γ yt  1 þ

s X i¼1

δi Δyt  i þ εt

ð7Þ

In this model, y is the variable under investigation, Δyt is the first difference operator (that is, Δyt ¼ yt  yt  1 ), t is a linear time trend that captures a deterministic trend in the process, s is the number of lags of Δy that we rely on to correct for possible serial correlation, and ε is a random error term. We run the ADF on the following variables: the flow of iPhone shipments, the flow of iPhone app submissions, the iPhone's hedonic price index, the log of iPhone shipments, the log of iPhone app submissions, the log of the iPhone's hedonic price index, the various versions of the stock of iPhones, the stock of apps available for the iPhone, the log of the various versions of the iPhone stock, and the log of the stock of apps available for the iPhone. To check the robustness of our results, we estimate the ADF with 1, 2, 3, and 4 lags, and both with and without a deterministic trend. Table 6 reports the results: for each test, the table reports the p-value associated with the null hypothesis that the time series has a unit root (that is, it is not stationary). For the most part, the tests overwhelmingly fail to reject the null hypothesis. 4.3. Results: cointegration Since for the most part the tests fail to reject the null hypothesis of a unit root, the next step is to test whether the time series are cointegrated, that is, whether they share a stochastic trend. In order to test for cointegration we first determine the optimal number of lags for the test and then compute a trace statistic proposed by Johansen (1995). We rely on a variety of tests to choose the number of lags to include—the likelihood-ratio (LR) criterion, Hannan and Quinn information criterion (HQIC), the final prediction error (FPE) criterion, and the Akaike information criterion (AIC). The Johansen tests find a cointegrating relationship for the following variables: the log of iPhone shipments and the log of app submissions with three lags, the log of iPhone shipments and the log of the iPhone's hedonic price with two and four lags, all three variables combined (log shipments, log app submissions, and log price) with three lags, and (various versions of) the stock of iPhones and the stock of apps available with three lags. Table 7 summarizes this information. 4.4. Results: flows on flows On the basis of these findings we estimate the long-run relationship between these variables. First, we apply the StockWatson DOLS estimator: yt ¼ α þ βxt þ

p X

θj Δxt  j þ μt

ð8Þ

j ¼ p

The DOLS estimator is like an augmented OLS regression of users on applications (or applications on users) in which the β coefficient is an estimate of the long-run relationship among the relevant variables. The regression is “augmented” with lags and leads of the first difference of the right-hand-side variables. The DOLS estimator is efficient in large samples, and it is also consistent even if the xt variables are endogenous (assuming, of course, that the relevant variables are cointegrated) (Stock & Watson, 2003, pp. 556–558). Table 8A reports the DOLS estimates for the regression of log iPhone shipments on log app submissions and the regression of log app submissions on log shipments (with 3 lags and 3 leads of the first differences). There are five models reported Table 6 P-values from Augmented Dickey–Fuller tests. 1 lag

2 lags

3 lags

4 lags

Deterministic trend?

No

Yes

No

Yes

No

Yes

No

Yes

iPhone shipments iPhone app submissions iPhone hedonic price log iPhone shipments log iPhone app submissions log iPhone hedonic price iPhone stock iPhone app stock log of iPhone stock log of iPhone app stock

0.79 0.46 0.39 0.19 0.00 0.80 1.00 1.00 0.05 0.79

0.00 0.32 0.81 0.00 0.00 0.08 1.00 1.00 0.33 0.84

0.99 0.46 0.34 0.11 0.00 0.79 1.00 1.00 0.44 0.48

0.62 0.45 0.82 0.00 0.00 0.07 1.00 1.00 0.09 0.70

0.99 0.41 0.29 0.28 0.00 0.78 0.99 1.00 0.05 0.22

0.76 0.62 0.83 0.00 0.00 0.08 0.98 1.00 0.14 0.40

0.97 0.37 0.23 0.39 0.00 0.77 0.66 1.00 0.13 0.05

0.12 0.55 0.85 0.00 0.01 0.08 0.55 1.00 0.75 0.11

Note: Cases in which the null hypothesis of a unit root is rejected are highlighted in bold.

886

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

in this table—the first three are models of log shipments on log app submissions, and the last two of log app submissions on log shipments. Column (1) is the basic DOLS model of log shipments on log app submissions, column (2) incorporates the log hedonic price on the right hand side, and column (3) further adds quarterly binary variables to account for seasonal effects. Column (4) is the basic DOLS model of log app submissions on log shipments, and column (5) adds quarterly binary variables. App submissions have a positive and statistically significant effect on iPhone shipments. In the basic model the elasticity of shipments relative to app submissions is about 1.2. The incorporation of the log hedonic price index on the right hand side, however, considerably changes the magnitude of the estimated indirect network effects: the coefficient on log app submissions declines from about 1.2 (in model 1) to about 0.19 once log price is included in the estimating equation (in models 2 and 3). The own price elasticity of iPhone shipments is about  0.93. It is also the case that iPhone shipments have a positive and significant effect on app submissions (see models 4 and 5). The elasticity of app submissions vis-à-vis shipments is about 0.79. Thus, the DOLS models provide evidence of statistically significant indirect network effects in both directions. We also estimate a VECM of the following form:

Δyt ¼ αðβ1 yt  1 þ β2 xt  1 Þ þ

s X i¼1

λi Δyt  i þ

s X

ϕi Δxt  i þ μt

ð9Þ

i¼0

This model specifies the first difference of the left-hand-side variable, which is stationary, as a function of linear lagged values of the first difference of xt , which are stationary as well, and stationary combinations of the non-stationary variables yt and xt , which represent the long-run relationships among these variables. More specifically, the long-term relation between the non-stationary variables is given by the elements of the vector β, and is known as the cointegrating vector. The rate at which yt responds to disequilibrium in the long-term relationship is given by α. Table 8B reports the results from estimating a VECM for the log of shipments on the log of app submissions (without and with the log of the hedonic price), and for the log of app submissions on the log of shipments. Model 1 is the basic model of log shipments on log app Table 7 Cointegrating relationships between different dimensions of the iPhone platform: flow of shipments, flow of app submissions, hedonic price, stock of phones, and stock of active apps (number of lags between parentheses). iPhone shipments App submissions Hedonic price Log app submissions Log hedonic price Active app stock Log active app stock

Log iPhone shipments

No No

iPhone stock

Log iPhone stock

No Yes (3) Yes (2 and 4)

No Yes (3) No

Table 8A DOLS of log iPhone shipments, log iPhone app submissions, and log iPhone hedonic price (with 3 lags and 3 leads of the first differences and Newey-West standard errors).

log app submissions

(1) Log shipments

(2) Log shipments

(3) Log shipments

1.223n (0.02)

0.197 (0.07)  0.929nnn (0.00)

0.186nn (0.00)  0.927nnn (0.00)  0.030 (0.75)  0.220n (0.01) 0.070 (0.24)

log hedonic price Q2 Q3 Q4 log shipments Constant Observations p-Values in parentheses. n p o 0.05. nn po 0.01. nnn p o0.001.

3.350 (0.52) 48

10.880nnn (0.00) 48

11.031nnn (0.00) 48

(4) Log app submissions

(5) Log app submissions

0.777nnn (0.00)  2.173 (0.41) 51

0.074 (0.67)  0.170 (0.50)  0.026 (0.91) 0.794nnn (0.00)  2.436 (0.37) 51

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

887

submissions, model 2 adds the log hedonic price on the right hand side, and model 3 incorporates seasonal binary variables. Model 4 is the basic model of log submissions on log shipments and model 5 adds seasonal binary variables. The same pattern that we observed in Table 8A is evident here. App submissions have a positive and significant effect on shipments. In model 1 the effect is not only significant but also substantial: the relevant elasticity is 4.8. The incorporation of log hedonic price on the right hand side (in models 2 and 3) generates a considerable decline in the magnitude of the relevant elasticity. In model 2 the elasticity is 0.95 and in model 3, which further adds seasonal controls, it is 0.58. The priceelasticity of shipments is estimated at about 0.90, not much different from the estimated elasticity in the DOLS models. It is also the case that shipments have a positive and significant impact on app submissions. The elasticity of app submissions relative to iPhone shipments is about 0.44 in model 5, which incorporates quarterly binary variables. Thus, these models also provide evidence consistent with the existence of statistically significant indirect network effects in both directions. We know that the DOLS estimators are consistent estimators of the long-run equilibrium relationships even if the righthand-side variables are endogenous. As a further robustness check we estimate the model of log shipments on log app submissions with instrumental variables for log app submissions and log hedonic price. These results, which are reported in Appendix C, again provide evidence of a positive and statistically significant relationship between the flow of iPhone shipments and the flow of iPhone app submissions.

4.5. Results: stocks on stocks As explained above, the cointegration tests also found a cointegrating relationship between the stock of iPhones and the stocks of active apps for the iPhone. Table 9A reports the DOLS estimates from the regression of the stock of iPhones on the stock of active apps. Table 9B reports the DOLS estimates from the regression of the stock of active apps on the stock of iPhones. (For space reasons we do not report the first-difference coefficients in either table.) It should be noted that we did not find a cointegrating relationship between the stock of iPhones and the iPhone's hedonic price. Further, the incorporation of hedonic price in the stock models does not change the reported results in any substantive way. The results from Table 9A and B suggest that there is a positive and statistically significant relationship between the stock of iPhone users and the stock of active apps available for the iPhone. More specifically, the results in Table 9A show that an extra app in the stock of active apps is associated with between 271 and 386 extra individuals in the stock of iPhone users (depending on the definition of the stock). The results in Table 9B show that 1000 extra users in the stock of iPhone users is associated with between 2.5 and 3.6 extra apps in the stock of active apps. Table 9C reports the results from estimating the stock-on-stock models via a VECM. We report the results for the stock 50 definition only for space reasons. These models show, once again, that there is a positive and statistically significant relationship between the stock of iPhones and the stock of active iPhone apps. An extra app in the stock of active apps is associated with an additional 381 units in the stock of iPhones (model 1), and an extra 1000 units in the stock of iPhones is associated with 2.6 extra apps in the stock of active apps (model 2). Both stock-on-stock indirect network effects relationships are positive and statistically significant, and are reasonably close in magnitude to those estimated via DOLS. An important caveat should be noted in connection with the impact of additional devices on the stock of active applications, and especially on the impact of additional applications on the stock of iPhones. We are estimating relationships that abstract from heterogeneity both in the stock of devices and in the stock of applications. Recent studies on mobile applications—such as the ones listed in the introduction to this paper—point out that there are indeed “killer apps” that have a disproportionate impact on the evolution of the marketplace. These are, by definition, applications that are in high demand and that are likely to have a high impact on the demand of the core product for which they were designed.

Table 8B VECM of log iPhone shipments, log iPhone app submissions, and log iPhone hedonic price (with 3 lags).

log app submissions log hedonic price log shipments

p-Values in parentheses. po 0.05; nnnp o 0.001. nn p o 0.01.

n

(1) Log shipments

(2) Log shipments

(3) Log shipments

4.800nn (0.00)

0.954nn (0.00) -0.849nn (0.00)

0.575nn (0.00) -0.900nn (0.00)

(4) Log app submissions

(5) Log app submissions

0.208 (0.14)

0.435nn (0.00)

888

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

5. Conclusions In this study we have analyzed the role of indirect network effects in explaining the dynamic behavior of the iPhone platform. Here we summarize our main findings and suggest directions for further research.

5.1. Key findings Most of the time series that reflect the evolution of the iPhone system are not stationary but it is possible to find cointegrating relationships among them. More specifically, we found cointegrating relationships among flows (log of iPhone shipments, log of app submissions, and log of hedonic price) and among stocks as well (stock of iPhone users and stock of apps available). This suggests that the relationships between the user and developer sides of the multi-sided platform are not spurious.

Table 9A DOLS models for the stock of iPhones on the stock of active apps (with 3 lags and 3 leads of the first differences and NeweyWest standard errors).

Stock of active apps Constant Observations

(1) Stock 50

(2) Stock 40

(3) Stock 30

271.00nnn (0.00)  3.4eþ 06nnn (0.00) 48

328.37nnn (0.00)  7.8e þ06nnn (0.00) 48

385.73nnn (0.00)  1.2e þ07nnn (0.00) 48

p-Values in parentheses. p o 0.05; nnpo 0.01. nnn p o 0.001.

n

Table 9B DOLS models for the stock of active iPhone apps on the stock of iPhones (with 3 lags and 3 leads of the first differences and Newey-West standard errors). (1) Stock of active apps iPhone stock 50

(2) Stock of active apps

0.0036nnn (0.00) 0.0029nnn (0.00)

iPhone stock 40

0.0025nnn (0.00) 3.1e þ 04nnn (0.00) 52

iPhone stock 30 1.2e þ04nnn (0.00) 52

Constant Observations

(3) Stock of active apps

2.4eþ 04nnn (0.00) 52

p-Values in parentheses. n po 0.05; nnp o0.01. nnn po 0.001.

Table 9C VECM for the stock 50 of iPhones and the stock of active iPhone apps. (1) iPhone stock 50 iPhone stock of active apps iPhone stock 50

p-Values between parentheses. n p o 0.05; nnnpo 0.001. nn p o0.01.

(2) iPhone stock of active apps

381.28nn (0.00) 0.0026nn (0.00)

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

889

As far as the relationships between flows are concerned, the point estimates from the basic model estimated in a DOLS framework (model 1 in Table 8A) suggest that the elasticity of shipments vis-à-vis apps is somewhat larger than the elasticity of apps vis-à-vis shipments. They are both close to one, however. In fact, we cannot reject the hypothesis that the elasticity of shipments relative to app submissions is one, and we can barely reject the hypothesis that the elasticity of app submissions relative to shipments is one. The incorporation of the hedonic price index in the models does change the magnitude of the elasticity of shipments visà-vis app submissions—a portion of the indirect network effects is now captured by the decline in the hedonic price index. For example, in the DOLS models, the point estimate of the elasticity of shipments relative to app submissions declines from 1.22 to about 0.19 when we incorporate the hedonic price index in the regression. The same pattern appears when we estimate the indirect network effects via a VECM. Overall, our conclusions regarding the flow relationships can be summarized as follows. First, the growth of app submissions contributed in a statistically significant manner to the growth of iPhone shipments. The elasticity of shipments vis-à-vis app submissions is roughly between 0.19 and 0.58 in models that also include the iPhone's hedonic price index. Second, the elasticity of shipments vis-à-vis the iPhone's hedonic price index is around  0.90 and statistically significant. Third, the growth of shipments contributed in a statistically significant manner to the growth of app submissions. The elasticity of app submissions relative to shipments is roughly between 0.44 and 0.79. We also found economically substantial and statistically significant relationships between the stock of iPhones and the stock of active apps on the App Store. An extra app on the stock of active apps is associated with roughly between 271 and 386 additional users, whereas 1000 additional users in the user stock are associated with roughly 2.5–3.6 additional apps in the stock of active apps. The estimated relationships tend to be statistically significant, which suggests that indirect network effects have indeed played an important role in the growth of the iPhone. Put differently, the multi-sided platform view of the iPhone finds support in the data. Note, in passing, the complexities of pricing on a multisided platform such as the iPhone. In pricing its services, the platform has to consider, first, the traditional elasticity measures—that is, the elasticity of shipments relative to the price of the core product, which we call ηC; PC ; and the elasticity of app submissions relative to the cost of developing applications for the platform (or relative to the “price” developers face when they “sell” their applications on the platform), which we call ηD; PD . In addition, the platform also has to take into account the elasticity of one side's participation relative to the participation of the other side—that is, the elasticity of core-product shipments relative to the availability of complementary products, which we call ηC; D , and the elasticity of submitted applications relative to shipments, which we call ηD; C . Particularly in the early stages of the life of a multi-sided platform, when the platform is seeking to achieve take-off toward self-sustained growth, intuition would suggest that the platform is better off encouraging (subsidizing) the participation of that customer type that has the strongest potential to attract customers of the other type (or types). This intuition is consistent with a more formal model of multisided pricing developed by Armstrong (2006), where optimal pricing to each side of the platform depends not only on that side's elasticity of demand vis-à-vis the price it faces for participation but also on the extent to which the participation of each side encourages the other side to join the platform.

5.2. Directions for further research A natural next step is the estimation of a structural model of hardware demand and software supply that incorporates the network effects that we have studied here. The literature has tended to model demand in the context of a nested-logit framework inspired by the contributions of Berry (1994) and Berry et al. (1995). Although this model can be estimated with a single cross section of data, the ideal data set for such estimation would be a panel data set containing information on prices, quantities, and characteristics for each smartphone device. We have taken a first step in the construction of such a dataset by collecting information on smartphone prices and characteristics for all iPhone and Android devices between 2007 and 2011, and we have used these data in the estimation of the hedonic models reported in Appendix B. But the estimation of nested logit models requires information on market quantities for each device, which we have not been able to obtain. A panel dataset would also allow us (or others) to use dynamic-panel-data (DPD) approaches in the construction of instrumental variables, an approach that has been used in recent years in panel-data estimation of a variety of econometric models. DPD “system” models, which combine an equation in levels and an equation in first differences, rely on lagged first differences of the endogenous variables as instruments for the current levels and on lagged levels of the endogenous variables as instruments for the current-period first differences. Estimation of such a structural model in a panel-data setting would provide an additional robustness check on our measures of indirect network effects on the iPhone platform. Ideally, structural models of indirect network effects on smartphone platforms should take into account not only heterogeneity in devices (smartphones) but in applications as well. It is well established that there are “killer apps” that are likely to have a large impact on the demand for the core product (device) for which they were designed. A complete structural study of the market should take this app heterogeneity into consideration in the analysis of indirect network effects.

890

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

Contributions

 First, we collect and analyze a set of time series that provides a comprehensive overview of the growth of the iPhone as an evolving system.

 Second, in the process of studying the growth dynamics of smartphones, we construct hedonic price indices for Apple (and Android) smartphones through 2011.

 Third, we present what to our knowledge is the first quantitative investigation of indirect network effects on the iPhone platform.

 Fourth, we show that indirect network effects do play an important role in the growth of the system and provide an econometrically-based assessment of that role.

Appendix A. Empirical studies on network effects In Table A1 we have summarized a variety of empirical studies on (direct and indirect) network effects. In the table we include the year when the study was published, the authors, the industry or market that the study focuses on, and a short description of the methodological approach. Appendix B. Hedonic prices for smartphones The literature on hedonic prices has a long history and we will not develop all the nuances here. Among the early studies on the subject see especially Dulberger (1989), Gordon (1989), and Triplett (1989). Among the more recent studies, see Pakes (2002), Chwelos, Berndt, and Cockburn (2004), and White, Abel, Berndt, and Monroe (2004). Suffice it to say that, in essence, hedonic models usually take the following form: X X lnðpi Þ ¼ α þ βj X j; i þ γ t Di; t þ εi ðB1Þ In this equation pi is the price of the product in question, in this case a given smartphone model, and X j; i is a vector of characteristics of each smartphone model, where i ¼ 1; :::; I indexes models and j ¼ 1; …; J indexes characteristics. The Di; t variables are binary variables for the specific time period t in which smartphone model i was introduced (a month, a quarter, or a year). The estimated γ t coefficients are then used in the construction of a hedonic price index for smartphones. More specifically, the level of the index for time period t is simply expð _ γ t Þ, where _ γ t is the coefficient for binary variable Di; t estimated from model (B1). We collected information on prices and characteristics for all iPhone models introduced between 2007 and 2012, as well as for all Android smartphones introduced between 2008 and 2011. We were particularly careful to retrieve information on prices at the time when a specific device was introduced—as opposed to the price that some of those devices command today, since some of them are still available in the marketplace at a discount. Further, in the process of gathering price information we were especially careful in identifying whether an advertised price for a device was with or without a contract, since this makes a substantial difference for the posted price. For prices we followed fundamentally PC MAGAZINE (www.pcmag.com), and when in doubt we also checked CNET (www.cnet.com). If after checking these two sources we were still uncertain about the nature of the price information, we checked a variety of additional sources. We also collected information on a wide variety of characteristics for each device, including when the device was introduced, who the manufacturer was, on which network it operated, the operating system it used, whether it included a GPS, whether it included a digital camera, whether it included a digital music player, the physical dimensions of the phone (width, depth, height, and weight), the technology it relied on (say, GSM), whether it had a touchscreen, whether it had a keyboard, the clock speed (in MHz), the maximum amount of memory that could be included on a memory card (in MB), whether a card was included and if so how much memory it contained (in GB), the amount of internal memory the phone had (in MB), RAM memory (in MB), ROM memory (in MB), the camera pixels, the touchscreen's display technology (say, TFT), display size (in inch.), display resolution (in PPI), maximum talk time (in min), and battery capacity (in mAh). All of these characteristics were taken from GSMARENA (www.gsmarena.com). We estimated a hedonic econometric model for iPhone devices from 2007 through 2012. More precisely, we estimated four different hedonic models for iPhone devices: a model with annual dummy variables and the characteristics in levels, another model with annual dummy variables and the characteristics in logs, a third model with quarterly dummy variables and the characteristics in levels, and a fourth model with quarterly dummy variables and the characteristics in logs. We estimated various versions of these four models with various subsets of characteristics included on the right hand side. Leaving aside the year-specific binary variables, the models presented in Table B1 include the following control variables on the right hand side: a binary variable for whether the price corresponds to a device on a contract or not, the digital camera's pixels, the display resolution (in pixels per in.), and the internal memory of the device (in MB). For space reasons, we only report the models with annual dummy variables. The results look quite reasonable. Devices on a contract are considerably cheaper than those sold without a contract. Further, extra camera pixels, better display resolution and additional units of internal memory are associated with higher

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

891

Table A1 Selected literature on the empirical estimation of network effects. Reference

Product or market

Approach

Gandal (1994)

Spreadsheets

The author estimates hedonic price equations for spreadsheets to test whether Lotus compatibility commands a premium

Saloner and Shepard (1995)

ATMs

The authors use standard duration models to test whether the speed of adoption of ATM technology by a bank is a function of the number of branches it has (interpreted as a proxy for expected network size)

Majumdar and Telecommunications The authors focus on the US telecommunications industry and rely on OLS cross-sectional Venkataraman (1998) regressions to study the impact on adoption of three different effects: (a) a conversion effect driven by operations-related increasing returns to scale, (b) a consumption effect driven by demand-side increasing returns to scale, and (c) an imitation effect Gandal et al. (1999)

Microcomputers

The authors estimate a vector autoregressive (VAR) model to explore the magnitude of the feedback effects between operating-system and application software in the early microcomputer industry (with a focus on the DOS and CP/M operating systems)

Gandal et al. (2000)

CD players

The authors rely on a two-equation, time-series structural model to estimate the direct elasticty of CD-technology adoption relative to CD player prices and the cross elasticity with respect to the variety of CD titles available in the marketplace

Goolsbee and Klenow (2002)

Home computers

The authors estimate cross-sectional regression models on a micro-data set to determine whether households are more likely to acquire their first computer in areas where a high fraction of households already own computers

Dranove and Gandal (2003)

DVD standards

The authors rely on time-series regressions to estimate the impact on the adoption of DVD players of (a) the availability of movies in DVD format (network effects), and (b) the preannouncement of DIVX technology

Kim and Kwon (2003)

Mobile phones

On the basis of a consumer survey for the Korean mobile telephony market, the authors estimate a conditional logit model to determine whether users prefer carriers with larger numbers of subscribers

Shankar and Bayus (2003)

Home video games

The authors use monthly time-series data on sales of Sega and Nintendo home video game systems and a structural model of demand to quantify the strength of direct and indirect network effects

Gowrisankaran and Stavins (2004)

Electronic payments The authors use a panel dataset on bank adoption and usage of ACH services (and a static model of technology adoption) to determine the extent to which network externalities matter for the ACH electronic payments system

Park (2004)

VCRs

The author relies on a firm-level panel data set for the US home VCR market (for 1981–1988), and estimates a standard nested logit model with the goal of measuring the extent to which network externalities contributed to the standardization of the VHS format in the US

Nair, Chintagunta and Dube (2004)

Personal Digital Assistants

The authors use panel data on sales of Palm and Microsoft personal digital assistants to estimate a structural model that includes a nested-logit hardware demand equation and a software-provision equation

Rysman (2004)

Yellow pages

The author focuses on the market for Yellow Pages, and estimates three models: (a) consumer demand for directory usage, (b) advertiser demand for advertising, and (c) the publisher's first-order condition; the ultimate goals are to determine whether advertisers value consumer usage and whether consumers value advertising

Clements and Ohashi (2005)

Video games

The authors use a panel dataset on videogame consoles and titles to estimate (a) a standard nested logit model on the console-demand side, and (b) a reduced-form model on the software-entry side; the objective is to determine the extent to which software variety matters for console adoption

Doganoglu and Grzybowski (2007)

Mobile phones

The authors focus on the market for mobile telecommunications services in Germany; they use panel data to estimate a nested logit model with the goal of measuring the extent to which network effects played a role in encouraging the growth of mobile subscriptions

Birke and Swann (2007) Mobile phones

The authors estimate an OLS time-series regression to determine whether mobile users in the UK between 1999 and 2003 made more on-net calls than would be expected based on price differentials only

Corts and Lederman (2007)

Home video games

The authors estimate two equations (hardware demand and software supply) with the goal of measuring the extent to which there are indirect network effects affecting users of competing and incompatible hardware platforms

Stremersch et al. (2007) Multiple industries

The authors estimate various time-series models to gauge the impact of indirect network effects in a variety of industries (black and white television, CD, CD-ROM, color television, DVD, Game Boy, imode, Internet, and laser disc)

Lee (2013)

The author uses a panel dataset to estimate a dynamic structural model of consumer demand for both videogame hardware and software for the period 2000–2005; the model explicity incorporates heterogeneity in consumer preferences and videogame quality

Video games

892

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

Table B1 Hedonic regressions for iPhone models only, 2007–2012.

Contract Log camera pixels Log display resolution Log internal memory Year 2008 Year 2009 Year 2010 Year 2011 Year 2012

(1) Log–log

(2) Log-level

 0.95184nnn (0.00) 0.94375nnn (0.00)  0.00138 (0.98) 0.44883nnn (0.00)  0.93144nnn (0.00)  1.85798nnn (0.00)  2.29307nnn (0.00)  2.72816nnn (0.00)  2.72818nnn (0.00)

 0.95184nnn (0.00)

Camera pixels Display resolution Internal memory Constant Observations

2.73044nnn (0.00) 35

 0.65880n (0.01)  1.29882nnn (0.00)  2.00350nnn (0.00)  2.70819nnn (0.00)  2.70164nnn (0.00) 0.23483nnn (0.00) 0.00164nn (0.01) 0.00001nnn (0.00) 6.44196nnn (0.00) 35

p-Values in parentheses. n p o0.05. nn p o 0.01. nnn p o 0.001

prices. Most of the estimated coefficients have the expected sign and are statistically significant. The annual binary variables suggest that the quality of the iPhone improved dramatically between 2007 and 2012. For comparison purposes we also estimated a model for Android phones between 2008 and 2011. Table B2 presents the key results for econometric models with annual binary variables. The Android hedonic model is slightly different from the iPhone hedonic model we estimated. For example, in the Android hedonic model we were able to include manufacturer and network fixed effects. Further, we defined the “memory” variable in the Android hedonic model as a binary variable that takes on the value 1 if a memory card with at least 16 GB of memory is included with the device (since the internal memory variable often had missing values for Android smartphones). Leaving these caveats aside, most of the variables have the expected signs and are statistically significant. The annual binary variables suggest that a steady process of quality improvement led to a decline in quality-adjusted prices for Android devices. They also suggest that quality improvement on the Android platform happened at a slower pace than on the iPhone platform, at least through 2011. Appendix C. Instrumental variables as an additional robustness check on the log shipments equation We know that when time series are cointegrated, both the OLS estimator and the DOLS estimator are consistent for the relevant parameter, that is, for the parameter that captures the strength of the long-term, equilibrium relationship between the series. It could be the case, however, that the estimates we present in the core of this study are still tainted by some degree of small-sample bias. Thus, here we estimate the relevant relationships via an instrumental-variable approach. More specifically we rely on instrumental variables to re-estimate the following equation: COMP CP lnðQ CP Þ þ ν1; t t Þ ¼ α1 þ β 1 lnðP t Þ þ δ1 lnðQ t

This is the equation of log iPhone shipments on log hedonic price and on the log of the flow of app submissions. We use a number of variables that capture the evolution of costs as instruments for the price of the iPhone. More specifically, since components for the iPhone are manufactured in Korea, Taiwan, and China, we use the log of the real exchange rate between the currency of each one of these countries and the US dollar as instruments for price. Further, we use the log of the real producer price index for semiconductors as an instrument for app submissions, under the assumption that (exogenous)

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

893

Table B2 Hedonic regressions for Android smartphones only, 2008–2011 (with handset manufacturer and network fixed effects).

Contract Keyboard Log camera pixels Log display resolution Log display size Memory Year 2009 Year 2010 Year 2011

(1) Log–log

(2) Log-level

 0.83566nnn (0.00) 0.16262 (0.07) 0.61332nnn (0.00) 0.29853 (0.31) 1.46757nn (0.00) 0.28430nn (0.00)  0.24603 (0.22)  0.62354nnn (0.00)  0.87030nnn (0.00)

 0.83219nnn (0.00) 0.16732 (0.07)

0.26780nn (0.01)  0.16699 (0.37)  0.56527nnn (0.00)  0.82994nnn (0.00) 0.10857nn (0.00) 0.00154 (0.25) 0.44697nnn (0.00) 4.23641nnn (0.00) 138

Camera pixels Display resolution Display size Constant

2.41696 (0.06) 138

Observations p-Values in parentheses.np o0.05. nn po 0.01. nnn p o0.001.

Table C1 A comparison of the OLS, DOLS, and IV estimators for the model of log iPhone shipments on log app submissions and log hedonic price.

log app submissions log hedonic price Q2 Q3 Q4 Constant Observations

(1) OLS

(2) DOLS

(3) IV

0.172nn (0.00)  0.852nnn (0.00)  0.032 (0.64)  0.143n (0.02) 0.059 (0.10) 11.418nnn (0.00) 48

0.186nn (0.00)  0.927nnn (0.00)  0.030 (0.75)  0.220n (0.01) 0.070 (0.24) 11.031nnn (0.00) 48

0.146nn (0.01)  0.883nnn (0.00)  0.033 (0.63)  0.149nn (0.00) 0.053 (0.07) 11.594nnn (0.00) 48

p-Values in parentheses. n p o 0.05. nn po 0.01. nnn p o0.001.

technological change in semiconductor components makes app developers more productive and thus increases the supply of iPhone applications. Of course, the producer price index for semiconductors can be viewed as an instrumental variable for the price of the iPhone as well. Relying on cost-related variables as instruments for price is a well-established approach and has been applied in some indirect network effect studies, for example Clements and Ohashi (2005). For comparison purposes, in Table C1 we present three different sets of estimates of the parameters that matter in the log-shipment model: one set obtained via OLS, another one obtained via DOLS, and a third one obtained via an IV approach.

894

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

The two key parameters are the elasticity of iPhone shipments relative to the iPhone's hedonic price and the elasticity of iPhone shipments vis-à-vis app submissions. The elasticity of shipments relative to price is between  0.85 and  0.93, and it is always statistically significant. The elasticity of iPhone shipments vis-à-vis app submissions is between 0.15 and 0.19 and it is also statistically significant in all models. If we compare the IV model with the OLS model, we find that the elasticity of shipments relative to price is larger (in absolute value) in the IV model and that the elasticity of shipments relative to apps is smaller in the IV model. Thus, OLS appears to be slightly overestimating responsiveness to apps and slightly underestimating responsiveness to price. References Armstrong, M. (2006). Competition in two-sided markets. Rand Journal of Economics, 37(3), 669–691. Berry, S. (1994). Estimating discrete-choice models of product differentiation. Rand Journal of Economics, 25, 242–262. Berry, S., Levinsohn, J., & Pakes, A. (1995). Automobile prices in market equilibrium. Econometrica, 63, 841–890. Birke, D., & Swann, G. (2007). Network effects and the choice of mobile phone operator. In U. Cantner, & F. Malerba (Eds.), Berlin and Heidelberg: Springer. Bresnahan, T., Davis, J., & Yin, P. (2013). Economic value creation in mobile applications. SIEPR discussion paper 13-007. Caillaud, B., & Jullien, B. (2003). Chicken and egg: Competition among intermediary service providers. Rand Journal of Economics, 34(2), 309–328. Campbell-Kelly, M., & Garcia-Swartz, D. (2015). From mainframes to smartphones: A history of the international computer industry. Cambridge: Harvard University Press. Campbell-Kelly, M., Garcia-Swartz, D., Lam, R., & Yang, Y. (2015). Economic and business perspectives on smartphones as multi-sided platforms. Telecommunications Policy in press, http://dxdoi.org/10.1016/j.telpol.2014.11.001. Cecere, G., Corrocher, N., & Battaglia, R. (2015). Innovation and competition in the smartphone industry: Is there a dominant design?. Telecommunications Policy, 39(3–4), 162–175. Chwelos, P., Berndt, E., & Cockburn, I. (2004). Faster, smaller, cheaper: A hedonic price analysis of PDAs. National bureau of economic research working paper 10746. Clements, M., & Ohashi, H. (2005). Indirect network effects and the product cycle: Video games in the US, 1994–2002. Journal of Industrial Economics, 53(4), 515–542. Corts, K., & Lederman, M. (2007). Software exclusivity and the scope of indirect network effects in the U.S. home video game market. Available at 〈http://papers. ssrn.com/sol3/papers.cfm?abstract_id=1029856〉. Davis, J. P., Muzyrya, Y., & Yin, P., (2013). Experimentation strategies and entrepreneurial innovation: Inherited market differences in the iPhone ecosystem. SIEPR discussion paper 13-029. Dickey, D., & Fuller, W. (1979). Distribution of the estimators for auto regressive time series with a unit root. Journal of the American Statistical Association, 7, 427–431. Doganoglu, T., & Grzybowski, L. (2007). Estimating network effects in mobile telephony in Germany. Information Economics and Policy, 19(1), 65–79. Dranove, D., & Gandal, N. (2003). The DVD vs. DIVX standard war: Empirical evidence on vaporware. Journal of Economics and Management Strategy, 12(3), 363–386. Dulberger, E. (1989). The application of a Hedonic model to a quality-adjusted price index for computer processors. In D. Jorgenson, & R. Landau (Eds.), Cambridge and London: The MIT Press. Economides, N. (1996). The economics of networks. International Journal of Industrial Organization, 14, 673–699. Elliott, G., Rothenberg, T., & Stock, J. (1996). Efficient tests for an autoregressive unit root. Econometrica, 64, 813–8136. Enders, W. (2010). Applied econometric time series 3rd edition). Hoboken, NJ: John Wiley & Sons. Evans, D. (2003). The antitrust economics of multi-sided platform markets. Yale Journal of Regulation, 20(2), 325–381. Farrell, J., & Klemperer, P. (2007). Coordination and lock-in: Competition with switching costs and network effects. In M. Armstrong, & R. Porter (Eds.), Handbook of industrial organization, Vol 3 (pp. 1967–2072). New York: Elsevier. Gandal, N. (1994). Hedonic price indexes for spreadsheets and an empirical test for network externalities. Rand Journal of Economics, 25, 160–170. Gandal (2005). Network effects—Empirical studies. Available at 〈http://www.tau.ac.il/  gandal/empirical%20studies.pdf〉. Gandal, N., Greenstein, S., & Salant, D. (1999). Adoptions and orphans in the early microcomputer market. Journal of Industrial Economics, 47, 87–105. Gandal, N., Kende, M., & Rob, R. (2000). The dynamics of technological adoption in hardware/software systems: The case of compact disc players. Rand Journal of Economics, 31, 43–61. Gartner Research (2008). Gartner says worldwide smartphone sales grew 16 per cent in second quarter of 2008 (September 8). Available at 〈http://www. gartner.com/newsroom/id/754112〉 last visited on 08.02.13. Ghose, A., & Han, S. (2014). Estimating demand for mobile applications in the new economy. Available at 〈http://papers.ssrn.com/sol3/papers.cfm? abstract_id=2378007〉. Gordon, R. (1989). The postwar evolution of computer prices. In D. Jorgenson, & R. Landau (Eds.), Cambridge and London: The MIT Press. Goolsbee, A., & Klenow, P. (2002). Evidence on learning and network externalities in the diffusion of home computers. Journal of Law and Economics, 45(2), 317–343. Gowrisankaran, G., & Stavins, J. (2004). Network externalities and technology adoption: Lessons from electronic payments. Rand Journal of Economics, 35, 260–276. Granger, C., & Newbold, P. (1974). Spurious regressions in econometrics. Journal of Econometrics, 2, 111–140. Hagiu, A., & Wright, J. (2015). Multi-sided platforms. Harvard business school working paper 15-037. Heij, C., de Boer, P., Franses, P., Kloek, T, & van Dijk, H. (2004). Econometric methods with applications in business and economics. Oxford: Oxford University Press. Johansen, S. (1988). Statistical analysis of cointegrating vectors. Journal of Economic Dynamics and Control, 12, 231–245. Johansen, S. (1995). Likelihood-based inference in cointegrated vector autoregressive models. Oxford: Oxford University Press. Johansen, S., & Juselius, K. (1990). Maximum likelihood estimation and inference on cointegration with application to the demand for money. Oxford Bulletin of Economics and Statistics, 52, 169–209. Katz, M., & Shapiro, C. (1994). Systems competition and network effects. Journal of Economic Perspectives, 8(2), 93–115. Kenney, M., & Pon, B. (2011). Structuring the smartphone industry: Is the mobile internet OS platform the key?. Journal of Industry, Competition, and Trade, 11(3), 239–261. Kim, H., & Kwon, N. (2003). The advantage of network size in acquiring new subscribers: A conditional logit analysis of the Korean mobile telephony market. Information Economics and Policy, 15, 17–33. Lendino, J. (2012). Happy 4th Birthday, Android. PCMag.com (September 20). Available at 〈http://www.pcmag.com/article2/0,2817,2409863,00.asp〉 last visited on 08.02.13. Lee, R. (2013). Vertical integration and exclusivity in platform and two-sided markets. Available at 〈http://pages.stern.nyu.edu/ rslee/papers/VIExclusivity.pdf〉. Majumdar, S., & Venkataraman, S. (1998). Network effects and the adoption of new technology: Evidence from the U.S. telecommunications industry. Strategic Management Journal, 19, 1045–1062.

D.D. Garcia-Swartz, F. Garcia-Vicente / Telecommunications Policy 39 (2015) 877–895

895

Markoff, J. (2007). I, Robot: The man behind the Google phone. New York Times (November 4), Available at 〈http://www.nytimes.com/2007/11/04/technology/ 04google.html?pagewanted=all〉 last visited on 08.02.13. McCracken, H. (2009). Verizon's Droid targets iPhone. PCWorld (October 18). Available at 〈http://www.pcworld.com/article/173865/verizons_droid_target s_iphone.html〉 last visited on 08.02.13. Nair, H., Chintagunta, P., & Dube, J. (2004). Empirical analysis of indirect network effects in the market for personal digital assistants. Quantitative Marketing and Economics, 2(1), 23–58. Olsen, S. (2008a). Advice for Apple iPhone start-ups. CNET News (March 6). Available at 〈http://news.cnet.com/8301-10784_3-9888320-7.html〉 last visited on 08.02.13. Olsen, S. (2008b). Apple, Google vie for hearts (and wallets) of developers. CNET News (March 31). Available at 〈http://news.cnet.com/ 8301-10784_3-9905858-7.html〉 last visited on 08.02.13. Pakes, A. (2002). A reconsideration of Hedonic price indices with an application to PC's. National Bureau of Economic Research working paper 8715. Park, S. (2004). Quantitative analysis of network externalities in competing technologies: The VCR case. Review of Economics and Statistics, 86, 937–945. Parker, G., & Van Alstyne, M. (2005). Two-sided network effects: A theory of information product design. Management Science, 51, 1494–1504. Pon, B., Seppälä, T., & Kenney, M. (2014). Android and the demise of operating system-based power: Firm strategy and platform control in the post-PC world. Telecommunications Policy, 38(11), 979–991. Rochet, J., & Tirole, J. (2003). Platform competition in two-sided markets. Journal of the European Economic Association, 1, 990–1029. Rochet, J., & Tirole, J. (2006). Two-sided markets: Where we stand. Rand Journal of Economics, 37(3), 645–666. Rysman, M. (2004). Competition between networks: A study of the market for yellow pages. Review of Economic Studies, 71, 483–512. Rysman, M. (2009). The economics of two-sided markets. Journal of Economic Perspectives, 23, 125–143. Saloner, G., & Shepard, A. (1995). Adoption of technologies with network externalities: An empirical examination of the adoption of automated teller machines. Rand Journal of Economics, 26, 479–501. Schmalensee, R. (2002). Payment systems and interchange fees. Journal of Industrial Economics, 50(2), 103–122. Shankar, B., & Bayus, B. (2003). Network effects and competition: An empirical analysis of the home video game industry. Strategic Management Journal, 24, 375–384. Stock, J. (1994). Units roots, structural breaks, and trend. In R. Engle, & D. McFadden (Eds.), New York: Elsevier. Stock, J., & Watson, M. (1993). A simple estimator of cointegrating vectors in higher order integrated systems. Econometrica, 61, 783–820. Stock, J., & Watson, M. (2003). Introduction to econometrics. Boston: Addison-Wesley. Stremersch, S., Tellis, G., Franses, P., & Binken, J. (2007). Indirect network effects in new product growth. Journal of Marketing, 71, 52–74. Thomas, O. (2007). Apple: Hello, iPhone. CNNMoney (January 9). Available at 〈http://money.cnn.com/2007/01/09/technology/apple_jobs/index.htm〉 last visited on 08.02.13. Triplett, J. (1989). Price and technological change in a capital good: A survey of research on computers. In D. Jorgenson, & R. Landau (Eds.), Cambridge and London: The MIT Press. Vogelstein, F. (2008). The untold story: How the iPhone blew up the wireless industry. Wired (January 9). Available at 〈http://www.wired.com/gadgets/ wireless/magazine/16-02/ff_iphone?currentPage=all〉 last visited on 08.02.13. Vogelstein, F. (2011). How the Android ecosystem threatens the iPhone. Wired (April 14). Available at 〈http://www.wired.com/magazine/2011/04/mf_android/ 〉 last visited on 08.02.13. Vogelstein, F. (2013). Dogfight: How Apple and Google went to war and started a revolution. New York: Sarah Chrichton Books. West, J., & Mace, M. (2010). Browsing as the killer app: Explaining the rapid success of Apple's iPhone. Telecommunications Policy, 34, 270–286. White, A., Abel, J., Berndt, E., & Monroe, C. (2004). Hedonic price indexes for personal computer operating systems and productivity suites. National bureau of economic research working paper 10427. Yin, P., Davis, J., & Muzyrya, Y. (2013). Entrepreneurial innovation: Killer apps and the iPhone ecosystem. SIEPR discussion paper 13-028.