Changing Their Tune: How Consumers' Adoption of Online Streaming Affects Music Consumption and Discovery

Marketing Science Institute Working Paper Series 2016 Report No. 16-136

Changing Their Tune: How Consumers’ Adoption of Online Streaming Affects Music Consumption and Discovery

Hannes Datta, George Knox, and Bart J. Bronnenberg

Revised May 2017

“Changing Their Tune: How Consumers’ Adoption of Online Streaming Affects Music Consumption and Discovery” © 2016 Hannes Datta, George Knox, and Bart J. Bronnenberg; Report Summary © 2016 Marketing Science Institute

MSI working papers are distributed for the benefit of MSI corporate and academic members and the general public. Reports are not to be reproduced or published in any form or by any means, electronic or mechanical, without written permission. Report Summary

Streaming has rapidly become a predominant technology for the distribution and consumption of digital goods like music and movies. Unlike in the ownership model, when consumers purchase content, streaming services provide consumers access to vast libraries where content is free at the margin. Hence, the price of variety decreases for adopters of streaming.

Constructing a unique panel data set of individual consumers’ listening behavior on digital music platforms, the authors estimate the short-, medium-, and long-term effects of adopting a particular streaming service on quantity, variety and discovery of new digital content.

Among key findings, they find that consumer adoption of streaming leads to:

Increases in quantity of consumption. A half-year after users adopt Spotify, consumption, measured in play counts, is up by 43%.

Increases in variety of consumption. For example, the number of unique artists listened to increases by 36% six months after adopting streaming.

Increases in discovery of new music. Relative to music ownership where experimentation is expensive, repeat listening increases for consumers’ best new discoveries.

Implications The results have implications for consumers and producers of music, and are also of interest to entertainment researchers, trade organizations (e.g., RIAA) and governmental agencies (e.g., FCC).

• Streaming revenues are climbing not only because more consumers are adopting streaming, but because consumers’ overall consumption of music is growing as well.

• Streaming creates a more level playing field for smaller artists; however, while it is easier to enter the consumption set, it is harder to stay there.

• Streaming expands consumers’ attention to a wider set of artists, potentially increasing demand for complementary goods, like live performances.

• Streaming alleviates a deadweight loss problem for varieties where valuation is positive but below the price of ownership.

• Streaming increases consumer welfare by reducing search frictions (e.g., enhancing discovery) and helping users discover new high-value content.

Hannes Datta is Assistant Professor of Marketing and George Knox is Associate Professor of Marketing, both at Tilburg University, The Netherlands. Bart J. Bronnenberg is Professor of Marketing, Tilburg University, and Research Fellow in Industrial Organization, CEPR, London.

Marketing Science Institute Working Paper Series 1 Acknowledgements The authors thank Wes Hartmann and Puneet Manchanda for comments on an earlier draft. They are also grateful for comments from seminar participants at Amsterdam Business School, INSEAD, Tilburg University, the University of Cologne, the Catholic University of Leuven and Vlerick Business School, the University of Oxford, the University of Michigan, Yale University, Eurosonic Noorderslag, participants at the 2016 Marketing Science and Marketing Dynamics Conferences, and executives at Buma/Stemra, Sony Music and Spotify. The authors acknowledge financial support from the Netherlands Foundation for Science (NWO 453-09-004) and from the Marketing Science Institute (MSI 4-1854).

Marketing Science Institute Working Paper Series 2 Introduction Traditionally, copyright-related industries have suffered as new digital technologies disrupted their revenue models. One such disruptive technology that is taking over the music industry is streaming, which allows consumers to rent unlimited access to a vast library of content. Digital music revenues, previously driven by purchases, now mostly derive from subscription fees; in 2015, streaming became the single largest source of music industry revenues in the U.S. (Friedlander 2016). A similar shift from ownership-based to streaming-based business models is taking place in other copyright-related industries (e.g., movies, games, books). Similar to research on file sharing, the rise of streaming has triggered a discussion among researchers about its effects on producers’ profits. Using song-level digital sales, Aguiar and Waldfogel (2015) find that streaming displaces ownership-based downloads. In a survey panel, Wlömert and Papies (2016) show that free, ad-supported streaming services cannibalize demand from other channels; since revenues from paid subscriptions more than offset this effect, streaming positively affects sales. Aguiar (2015) documents that ad-supported streaming of music increases visits to legal and illegal downloading websites among heavy users. Missing from the literature is an account of how streaming affects consumption behavior at the individual level. We study how streaming technology affects music consumption in three ways. We first ask to what extent streaming generates additional music consumption rather than displacing consumption from other platforms. Access to a wide variety of content on a streaming platform may entice consumers to consume more, potentially turning deadweight loss —music that is valued above zero but below its purchase price, and hence is not consumed— into surplus (e.g., see also Waldfogel 2012). Another possibility is that consumers merely shift consumption to other platforms: the songs may be different, but the time spent listening remains unchanged. Second, music is a consumption good for which most consumers have a love of variety. We study the effects of streaming on the nature and size of the subset of variety chosen. In the ownership model, when consumers purchase and download specific music titles, variety is costly at the margin; in the streaming model, variety is free at the margin. We measure the breadth of variety in terms of the number of distinct artists, songs, and genres consumed. Next, we measure how users reallocate their time listening, whether they concentrate on a few artists (e.g., superstars, Elberse 2008; Rosen 1981), songs or genres, or spread their time listening across a

Marketing Science Institute Working Paper Series 3 wider set of artists. If consumers concentrate their consumption on a few major artists, they perpetuate a winner-take-all market. If, on the other hand, consumers now broaden their consumption, streaming may level the playing field to the benefit of smaller music producers, e.g., indie artists or labels. Third, an important determinant of welfare are consumers’ search frictions when discovering new high-value content. We investigate how music discovery changes after adopting a streaming technology. By reducing the costs of exploring the variety of music, streaming allows—according to one estimate—the average user to discover 27 new artists per month (Kissel 2015). To what extent do these new discoveries actually yield highly played songs? We investigate this issue empirically by considering how streaming affects repeat consumption for new titles in general, and for selected popularity deciles of new discoveries. We examine the effects of streaming by focusing on the moment of adoption of Spotify, currently the largest streaming provider serving 100 million customers in 60 countries (Spotify 2016). We construct a unique panel data set capturing individual-level music consumption by using a third-party service that tracks consumers’ platform choices and listening behavior across a wide set of platforms. We identify (self) “treated” consumers who adopt streaming in our observation window, though they may continue to use other providers. Next, we match adopters to “control” users who do not adopt streaming. We also measure variety using a secondary data set with meta-level characteristics for more than 200,000 artists. We identify treatment effects of adopting a streaming service on total music consumption on all platforms in the short- (within two weeks), medium- (up to six months), and long-term (six to twelve months after adoption). We use a differences-in-differences (DID) approach that controls for unobserved user-level and time-varying characteristics. Two forms of selection complicate our identification strategy. First, our data lack a randomized assignment of consumers into treatment and control conditions; we use a quasi-experimental procedure to distinguish causal treatment effects from simple differences in the characteristics between adopters and non- adopters (e.g., Bronnenberg, Dubé, and Mela 2010). Second, the demographic profile of sampled consumers may not be representative of the larger population of potential adopters. Hence, we study local average treatment effects (LATE) among those consumer segments who adopt streaming.

Marketing Science Institute Working Paper Series 4 We find that adoption of streaming services leads to a long-run growth of 43% in overall music consumption across all platforms. A sizeable amount of consumption on streaming services comes at the expense of ownership platforms such as iTunes, Winamp and Windows Media Player. Breadth of variety increases and concentration drops as consumers expand their listening over a larger assortment of artists, songs, and genres. Streaming increases the rate at which consumers discover new music. After adopting streaming, consumers replay each new variety less on average; however, the best discoveries post-adoption generate increased listening. This suggests that adoption of streaming services leads consumers to discover higher valued music, which in turn may be caused by low-cost trial of a wider selection. We examine several sources of heterogeneity in the local average treatment effect. For example, if streaming increases music discovery, we would expect the effect to be larger for consumers with limited variety prior to adoption. In line with this intuition, we find that the effect of streaming on discovery is larger for users who consumed more “Top 100” artists before adoption and for older users. We verify the robustness of our results by using different variable operationalizations, functional forms, long-term effects, and definitions of our sample. In one particular robustness check, we address selection on unobservables by applying DID on adopters only, in which later adopters act as a control for those who have adopted several months earlier (similar to, e.g., Manchanda, Packard, and Pattabhiramaiah 2015). The results from this and other robustness checks indicate broad agreement with our reported results. Taken together, our results demonstrate a significant, long-term diversification of listening behavior, as consumers choose a wider set of variety and engage in less repeat listening after adopting the streaming model. Combined with the increase in new music discovery, we conclude that streaming services level the playing field for artists relative to music distribution based on ownership by consumers.

Variety in the music entertainment industry The music industry has been studied in economics (Adler 1985; Cameron and Collins 1997; Chung and Cox 1994; Rosen 1981), marketing (Chung, Rust, and Wedel 2009; Holbrook and Hirschman 1982; Lacher and Mizerksi 1994), law (Zentner 2006), and sociology (Lopes 1992). Because music variety is at the heart of consumer welfare, unsurprisingly, a central issue in the literature are the limits on variety consumption in demand and supply.

Marketing Science Institute Working Paper Series 5 On the demand side, variety in music can serve two purposes. First, it can cater to consumers with idiosyncratic tastes (Crain and Tollison 2002). In this setting, more variety meets the tastes of more consumers and enhances welfare along the extensive consumer margin. Alternatively, a broad selection in music can satisfy the demand for variety at the individual level (Adler 1985; Chung and Cox 1994; Kim, Allenby and Rossi 2002; Ratner, Kahn and Kahneman 1999), creating welfare along the intensive consumer margin. Another literature on the demand for variety does not consider its search cost (see, e.g., Elberse 2008), but the acquisition cost of purchasing quantity versus variety (see, e.g., Bronnenberg 2015). This literature suggests that the costs of a marginal variety lead to limited demand for variety. On the supply side, one of the earliest empirical observations concerning variety in the music industry is that a relatively small number of artists commands a large share of the revenue, i.e., that the music industry is characterized by limits to the supply of variety. Rosen (1981) presents a theory which proposes that “superstars” emerge from two conditions. One is that artists’ rewards are convex in talent. This occurs when consumers view variety and quality as substitutes and are willing to pay more for a single top performance than for several good but intermediate ones. Convexity turns small talent differences into large pay-off differences. The second condition is that the marginal cost of producing a consumer experience, e.g., the cost associated with producing an additional Mp3 file, is low. This creates a scale economy allowing few sellers to serve many consumers. An alternative view is that concentration arises —even among equally talented artists— from differences in search cost or complementary “consumption capital.” For instance, Adler (1985) and Chung and Cox (1994) view music as an experience good. Rather than requiring a distribution of talent (and a magnified distribution of rewards), superstardom can emerge from imitation behavior by fans who have incomplete information and choose popular artists to minimize search costs. Adler discusses this mechanism in the context of accumulating consumption capital that is produced by listening to music and “discussing it with other persons who know about it” (p. 208). This produces a reinforcing spillover among consumers and selects a handful of lucky performers to become stars. Empirically, the consumption capital explanation has gained some support. Lacher and Mizerski (1994) find that consumers are likely to purchase music more by its ability to create an

Marketing Science Institute Working Paper Series 6 absorbing experience than by liking alone. Using a measure of voice quality, Hamlen (1991) documents that small differences in talent do not lead to excessive differences in rewards. Entry of streaming providers in a market dominated by ownership models has affected the three different mechanisms of variety reduction mentioned above. First, and central to our paper, the ownership model charges a fee per variety, i.e., per song, whereas a streaming provider charges a subscription fee for the entire catalog, thus dramatically lowering the acquisition cost of variety. Indeed, variety on a streaming provider is free of charge (although some search cost may remain) and this paper seeks to exploit this shift in costs to measure its impact on consumption and demand. Second, there are additional effects on the supply of variety that may have secondary effects on variety consumption. Viewed through the lens of limited entry from convex rewards (e.g., Rosen 1981), a streaming provider changes an artist’s reward structure to a fee per play that does not depend on the quality or overall popularity of an artist. Thus, relative to a world where artists that are more popular command higher prices, the streaming reward schedule is less convex in popularity. Additionally, streaming provides low-cost information about artists and measures of consumption capital to consumers (e.g., by providing consumption information in the form of playlists). We now describe the data that we have collected to shed light on how streaming changes quantity and variety of music consumption, and the discovery of new music.

Data

Institutional background Recently, the music industry has witnessed a marked increase in the number of interactive streaming providers, which are the focus of this study.1 At present there are over 20 providers offering comparable services in terms of variety and price.2 The largest of these, Spotify, has 40 million paying subscribers in 60 countries (Spotify 2016); subscribers have unlimited access to a library of over 30 million songs in exchange for $9.99 per month. More than 60 million users are on Spotify’s free membership plan. They have less control over what they can listen to, and

1 Interactive providers give consumers free choice in which songs they want to consume. Non-interactive providers like Pandora, instead, are similar to radio broadcasting and offer a pre-selected set of songs only (Aguiar 2015). 2 For a comparison of streaming music providers, see https://en.wikipedia.org/w/index.php? title=Comparison_of_on-demand_streaming_music_services&oldid=728943470 [accessed August 3, 2016].

Marketing Science Institute Working Paper Series 7 advertisers pay Spotify to expose free users to commercials. Streaming providers pay royalties to copyright holders—labels, publishers, distributors, and artists—based on the number of times a song was streamed.

Sample We use a third-party music recommendation service (henceforth “service”) which wishes to remain anonymous to construct our data. The service builds detailed user profiles by recording users’ listening histories across multiple platforms. Consumers join the service to receive music recommendations based on consumption across all their music sources rather than those based on only one platform. The service supports more than 100 devices and clients, giving a comprehensive picture of music consumption, whether offline or online, mobile or desktop.3 We sampled from the service’s user base by repeatedly visiting its website in short intervals between April 22 and 29, 2014 (Oestreicher-Singer and Zalmanson 2013). Our web scrapers collected the usernames of recently active users that used the service during this period. From May 1, 2014 to August 1, 2015, we recorded the music consumption and platform choices of 5,003 randomly selected users from their profile pages. We use the first four weeks (until May 29, 2014) as an initialization period, and the remaining 62 weeks for estimation. Due to interruptions in our data collection for technical reasons, we record consumption on 431 out of 458 days. Missing days are not systematic. For each individual listener, we collect a unique user name, a timestamp, artist and song names, and platforms used to listen to each song. We observe 123 million plays for 4,033 active users. Data on the remaining 970 users is not available, either because these users changed their privacy settings or cancelled their accounts during our observation period. Using the service’s Application Protocol Interface (API), we collected additional information on music consumption, from January 6, 2013 onwards, i.e., predating our data collection through scraping which starts on May 1, 2014. These additional consumption data contain songs and artists, but not platform choices, and will be used later in matching adopters of streaming with non-adopters. We identify unique artists and songs using an algorithm described in Online Appendix A, so that “The Beatles” and “beatles” are counted as the same artist. Some songs have unrealistically

3 The service’s technology monitors song-level consumption on all platforms a user has activated (on average 4.17 platforms per user in our data). Music consumed offline is also monitored, and submitted whenever a connection becomes available. Traditional FM radio is not included.

Marketing Science Institute Working Paper Series 8 short song lengths; also, users may skip songs after listening only for a few seconds. To ensure that our findings are not driven by sampling, we remove all songs shorter than 30 seconds and songs that were skipped before half of the song has finished (in total, 7.2% of all plays). We retain 114 million plays for 4,033 users. In our estimation data set, the most played artists are Lana Del Rey, Taylor Swift, and Madonna; the main genres are rock, pop, and metal; average daily consumption of music is approximately 3 hours. The top platforms are Spotify and iTunes, with a market share of 22.8% and 18.3% in terms of the total number of plays. Other platforms are mainly used to play locally stored Mp3 files or listen to CDs on the computer, such as Winamp (12.5%), Foobar2000 (10.1%), and Windows Media Player (9.7%). We group all remaining platforms in an “other” category4 because they cannot be classified as streaming-based or ownership-based platforms (e.g., capturing consumption on a variety of websites), because they are non-interactive (see footnote 1), or because their market share in our sample is negligible.5 In Figure 1, we plot the market shares for the major platforms in terms of playcounts in our estimation sample. The figure shows that usage of Spotify grew steadily while that of iTunes and the remaining platforms declined. Hence, our aggregate data seem to confirm reports that Spotify is encroaching on the market shares of ownership-based platforms.

Identifying adoption of music streaming We consider Spotify and its adoption as the only streaming platform, because other streaming providers in our sample have negligible market shares. Because of potential left- truncation problems, we classify users as Spotify adopters only if we observe them not use Spotify for at least 45 days since the beginning of their recorded usage. Further, we require potential adopters to be active on at least one of the major ownership-based platforms (i.e., iTunes, Winamp, Windows Media Player, or Foobar2000). This procedure gives us 507 users

4 The “other” category consists of the following players (listed if market share > 1%): The service plugin (7.99%), Android service (4.33%), Simple service plugin (2.28%), service plugin for web (1.33%), Plugin for iOS (1.23%), and Clementine (1.08%). The remaining 211 niche players account for 8.4% of total playcounts. YouTube is part of our data, but largely tracked in combination with other web-based players. Hence, we cannot reliably distinguish YouTube from other services such as SoundCloud. 5 In fact, our sample includes other streaming platforms. However, due to their small market shares in our data, we would still be effectively reporting on Spotify if we were to consider their adoption. Their market shares are as follows: Grooveshark (.54%), Rdio (.42%), Google Play (.17%), simfy (.07%), WiMP (.03%) Deezer (.01%), Tidal (.01%), and 8tracks (.01%).

Marketing Science Institute Working Paper Series 9 that adopt Spotify in our 62-week estimation sample.6 Obviously, such users may continue to consume music on other platforms, too. A total of 1,471 users never adopt Spotify. We disregard 1,135 users who use Spotify before our 45-day cutoff, 714 users who never use a major ownership-based platform, and 206 light users who never listen in the initialization period or listen for less than 10 weeks in the estimation period. Estimating stable treatment effects requires that no major changes occur to Spotify during our observation period. We provide an overview of all significant changes to Spotify in Online Appendix B. In the first year of our estimation sample (between 29 May 2014 and 27 April 2015), major developments were confined to launching in new markets (Canada and Brazil), client updates (e.g., for Windows Phone), partnerships with hardware manufacturers (e.g., Sony’s PlayStation), and the introduction of a small number of new playlists (Dinner, Sleep, Folk Americana). Further, the supply of music remained stable (except for the removal of Taylor Swift, and the introduction of John Lennon and Rammstein). Important changes to Spotify’s recommendation algorithm were introduced in the last 16 weeks of our estimation period, such as “perfect playlists at your fingertips” (28 April 2015), playlists for workouts (“Spotify Running”), an improved starting page (both 20 May 2015), and personalized playlists with previously unheard content (“Discover weekly”, 20 July 2015). In our analysis section, we explain how we verify the robustness of our findings with regard to these innovations. Spotify’s main competitor, Apple Music, was launched in the last 4 weeks of our data (on 30 June 2015). We record each user’s demographic profile (i.e., age, gender, and country) from the service’s API. We also track artist-level meta-characteristics that allow us to characterize the type of variety consumers seek: we retrieve data on genre from Echonest.com (the industry’s leading music intelligence provider, using the service’s API), and Musicbrainz.org (an open- source music encyclopedia, using a virtual replication of their SQL server). This information is available for 53% of all artists.

6 See Figure C1 in Online Appendix C for the distribution of Spotify adoption times.

Marketing Science Institute Working Paper Series 10 Measuring Quantity, Variety and Discovery Quantity. The unit of analysis is the user-week: we measure the number of songs each user listens to in a week (i.e., the weekly playcount) on any platform. In Table C1 in Online Appendix C, we compare treatment and control groups on these measures. These statistics are only suggestive of an adoption effect, because they ignore heterogeneity and any differences in adoption effects in the short- and long-term. Variety. Measuring product variety is complex and difficult (Alexander 1997). Hence, we try to partially characterize variety and derive a set of metrics in two categories: breadth (e.g., Hoch, Bradlow and Wansink 1999), and concentration (e.g., Elberse 2008). We specify these metrics in a multi-attribute space (i.e., songs, artists, and an artist’s genre), facilitating interpretation in favor of collapsed and hence more abstract metrics (e.g., entropy, van Herpen and Pieters 2002). The first set of variety metrics relates to the breadth of variety consumed by users. We measure breadth by counting the number of distinct artists, songs, and genres consumed. The second set of metrics relates to the concentration of variety. We measure the popularity of consumed content as a share of the Top 20, 100, and 500 artists in a user’s region in the initialization period. We also calculate the concentration ratio based on each user’s own favorite top artists, songs and genres, as a share of total plays. Discovery. Crucial to our investigation is how the discovery of new content changes when consumers adopt streaming services. Because we observe listening history for more than a year prior to the beginning of our estimation sample, we can measure the consumption share of new music. As a proxy for value, we measure repeat consumption share for both new and known artists. To quantify the value of the “best” new discoveries, we calculate the ratio of top new variety plays to top overall plays (either new or known) over a rolling eight-week window. We explain the operationalization of all metrics in Table 1. Summary statistics for variety and discovery are in Table C2 in the Online Appendix. Do our measures reflect demand or supply? We assume that consumers are free to make their own consumption choices. However, certain features may restrict consumer choice or forbid it altogether, as is the case with the internet radio provider Pandora. If this is true, then the results of our analysis can still be interpreted as consumption effects, but they will not be informative about whether the effects are due to demand or supply. However, because Spotify is

Marketing Science Institute Working Paper Series 11 an interactive streaming provider, users can request content suggestions as a deliberate choice. Even if users were to receive recommendations (e.g., by listening to curated playlists), they could simply skip less preferred songs.7 Hence, we attribute changes in music consumption to shifts in demand, because it is unnecessary and unlikely that consumers listen to content on Spotify against their will. Second, a related concern is that differences in assortment (e.g., size, or indie vs. major) between ownership-based versus streaming platforms may drive our effects. However, both the music catalogue of iTunes and Spotify feature around 30 million songs and are thus comparable (Mitroff and Blanco 2015). Further, windowing strategies to release popular content on Spotify later than on iTunes were only introduced after our observation period. We acknowledge that, typically, there are more songs available for purchasing than for streaming (e.g., CDs bought at a local concert). However, an assortment of millions of songs on streaming platforms is not likely to constrain choice in a meaningful way.

Method

Identification strategy Our objective is to identify LATEs of adopting Spotify on quantity, variety and discovery. We face two major challenges. First, our data generation process lacks a randomized assignment of consumers into treatment and control conditions. A simple approach is to estimate a DID model, which removes any persistent consumer-specific effects that may, if ignored, introduce endogeneity due to self-selection into adopting Spotify. In addition, for our main results, we assume that we can control for the unobserved need for variety in absence of streaming by conditioning on a rich set of observed characteristics (for a similar approach, see Bronnenberg, Dubé, and Mela 2010). That is, we use a quasi-experimental matching procedure, in which we match adopters with similar non-adopters based on a propensity score, constructed from variables meant to capture users’ demand for variety. Second, treated and control consumers may exit our sample at different moments (e.g., because they stop using the service). To achieve comparability in terms of market trends we seek pairs of users with similar beginning- and end-points in our observation period. We combine all criteria by selecting pairs that minimize the Mahalanobis distance between treated and potential

7 As explained in the data section, we have excluded skipped songs from our data.

Marketing Science Institute Working Paper Series 12 control users based on the propensity score, and beginning- and end-points in our observation window. We now explain in detail each step of our identification strategy.

Comparison of treated and control groups As a first step, we compare the listening behavior in the initialization period (i.e., strictly prior to adoption), for consumers that eventually adopt Spotify with those that do not adopt. Table C3 in the Online Appendix (columns 1 and 2) shows the differences in key behavioral variables as well as demographic variables for unmatched users from the control and the treatment group. For the behavioral variables, we use an initialization period of 12 weeks before the start of our estimation sample. Adopters listen to more songs (49.95 plays vs. 42.27 plays) on more platforms (3.17 vs. 2.87) than non-adopters. Further, adopters are younger (22.41 years vs. 24.19 years), and more likely female (26% vs. 23%).8 We note that these differences are statistically small and insignificant. Nonetheless, to obtain conservative treatment effects, we use quasi-experimental methods.

Propensity Score estimation and matching Self-selection may arise due to differences in taste for variety. For example, consumers adopting streaming may do so because they value variety more. To induce quasi-randomization, we pair each consumer that adopted Spotify with a comparable consumer that is very likely to adopt, but “randomly” did not do so in our 62-weeks observation period.9 We achieve this by implementing a hybrid matching procedure (e.g., Gensler, Leeflang, and Skiera 2012, Bronnenberg, Dubé, and Mela 2010), which extends propensity score matching by enhancing the comparability of treated and control users beyond quasi-randomization. This is achieved by adding additional covariates to the matching procedure (e.g., to account for different activity windows, a point to which we will return shortly). We execute this procedure in three stages: First, we estimate each household’s adoption propensity as a function of observed variables (e.g., Rosenbaum and Rubin 1983):

8 In the cases for which age and gender are missing (22.4% and 12.4% of all users, respectively), we first estimate age and then gender based on an auxiliary regression model for age (and a logit model for gender) using covariates Zi, defined immediately following equation (1). 9 Note that Spotify launched in 2008. By the time of our observation window (2014-2015), Spotify already had 75 million users. Users who adopted likely belong to the early/late majority adopter segments; we expect them to have weaker tastes for variety than the innovators and early adopter segments in 2008-2014.

Marketing Science Institute Working Paper Series 13 Pr(adopti = 1) = Pr(β0 + Ziγ + ηi > 0) (1)

where 푍푖 is a vector of observed household-specific characteristics. The covariates entering

푍푖 describe a users’ listening behavior (e.g., quantity, variety, concentration, discovery, and repeat consumption) during 12 weeks prior to our sampling period starting at 29 May 2014. The relevant descriptive statistics are in Table C3 in the Online Appendix, columns 1 and 2. Our matching estimator is static: it uses the same pre-adoption observation window to predict adoption for early versus late adopters. This assumption greatly eases matching vis-à-vis more complicated, dynamic matching models. At the same time, we believe our assumption of more or less stable predictors is fair, as adoption occurs in a relatively narrow window of about a year.

We assume that the 휂푖 are IID random variables with a Type I Extreme Value distribution. This makes the probability in Equation (1) a binary logit model. Second, for further comparability of treatment and control, we ensure that each matched treated-control pair has the same observation window. We observe most users during our complete observation period, but about 10% of users stops using the service and exits the sample. If treated users were matched with control users in a different observation window, or vice versa, any estimated treatment effect would not be clearly attributable to adoption but be contaminated by timing effects. Thus, to ensure we are using the same observation window for matched treated-control pairs, we match users upon their first and last observed period of listening in our sample. In the third stage, we construct a three-dimensional distance metric of the composite of propensity score, and the first and last period of consumption. We compute the Mahalanobis distance for each treated and control user pair, and use the one-nearest neighbor algorithm to match users that are closest together.10 Some of the matches that we obtain are potentially non- unique, i.e., there are instances where the same non-adopter is the closest match for more than one adopter. We select unique matches sequentially, in order of closeness of their Mahalanobis distances. To ensure a sufficient matching quality, while avoiding the need to match with replacement, we match an adopter to his/her k-th best match. This strikes a balance between the number of adopters in our sample (increasing in k), and matching quality (decreasing in k). In

10 We drop control users outside the region of common support, defined as the overlap in propensity scores between treated and control users (e.g., Gensler, Leeflang, and Skiera 2012).

Marketing Science Institute Working Paper Series 14 preparing our final data set, we set k to 3, and retain those user-periods in which treatment- control observations overlap. The matching procedure yields 447 treated and 447 matched control users. In Table C4 in the Online Appendix, we report the results of the logit propensity score model. The directions of the effects have face validity: ceteris paribus, consumers with higher average plays, more superstar consumption and more platforms used are more likely to adopt Spotify. In turn, older users with more concentrated listening and users who already have access to and listen to new music are less likely to adopt Spotify. Notably, users from the regions “South America” and the “US/Canada” have higher adoption probabilities than users from other geographic regions, as Spotify launched in Brazil and Canada during our observation period. The hit rate is 66%, and McFadden Pseudo R2 is .093. After matching, both consumer groups are indistinguishable in terms of their adoption propensities (see Figure 2), observation windows (see Figure D2), as well as observed demographics and pre-adoption behavior. In Table C3 in the Online Appendix, column 3, we report pre-sample summary statistics for the matched control sample, which are now very close to those of the matched treatment group (column 4). Note that users, on average, are relatively young and predominantly male. Because these users subscribed to the service to track and publicly display their music consumption, they may also not be representative of the larger population of potential streaming adopters.

Differences-in-differences We use a DID approach to estimate the effect of adopting streaming on our outcome measures. We compare the outcome measures of adopters before and after their adoption with those of the matched non-adopters. We also investigate how long changes in consumption, variety and discovery last. We estimate models of the following type: ST Yit = αi + γt + β ∙ I(0 ≤ weeks_since_adoptionit ≤ 1) MT + β ∙ I(2 ≤ weeks_since_adoptionit ≤ 24) (2) LT + β ∙ I(weeks_since_adoptionit ≥ 25) + εit

where 푌푖푡 is the dependent variable, and the indicator variables 퐼(weeks_since_adoption푖푡) are 1 if the number of weeks since adoption for consumer i is within the indicated range, in week

Marketing Science Institute Working Paper Series 15 t. Further, 훼푖 is a consumer-level fixed effect, 훾푡 is a week-level fixed effect and 휀푖푡 is the error. This two-way fixed effects specification controls for time-invariant consumer characteristics, such as overall liking of music, as well as common time trends and week-to-week fluctuations.

An important identifying assumption is that the 휀푖푡 are orthogonal to the indicator variables

퐼(weeks_since_adoption푖푡). We distinguish between treatment effects in the short term (within the first 2 weeks of adoption, 훽푆푇), medium term (between 2 and 24 weeks after adoption, 훽푀푇), and the long term (25 weeks and after, 훽퐿푇). We use robust standard errors clustered at the user level to account for any serial correlation (Bertrand, Duflo and Mullainathan 2004). The DID approach relies on the assumption of parallel pre-treatment trends. To test for this, we carry out so-called “placebo” treatment tests using pre-adoption data (and matching the observation window in the control sample). In particular, we define a placebo “treatment” at the mid-point of a user’s pre-treatment data. Next, we estimate a DID model for all dependent variables in the study (see Table 1), across the dimensions songs, artists, and genres. For each but 2 of the 37 combinations of variables and dimensions we fail to reject the null-hypothesis of no treatment effect for placebo treatments (see Online Appendix, Table C5). This supports the idea that the pre-treatment trends are statistically equivalent across both user groups. Taken together, we combine a propensity score matching with a DID approach. Our propensity score model selects non-adopting consumers who are like adopters, except for not subscribing to music streaming. Because we include a rich set of behavioral measures of pre- existing consumer tastes for variety and discovery in our propensity score model, we assume that the unobserved components in the propensity equation (1), i.e., the 휂푖, are independent of the unobserved components of our regression model (2), i.e., the 휀푖푡. Given our method, we interpret the reported effects from the DID regressions as the average treatment effect on the adopters (compliers), i.e., as local average treatment effects (LATE). In this context, we highlight one additional type of analysis that our data permit us to run. Because adoption time differs across adopters (see Figure D1), we can estimate a treatment effect using only within-adopter variation, with late adopters acting as a control for early adopters. This approach also accounts for possible selection on unobservables shared by late and early adopters. We report on this analysis as a robustness check, after discussing the main results.

Marketing Science Institute Working Paper Series 16 Heterogeneity in Treatment Effects The impact of streaming may depend on individual characteristics and preferences, such as taste for variety. We investigate three potential moderators of our treatment effects: top 100 listening share prior to adoption, age and whether users use the premium or ad-based (free) plan of Spotify. We explain our predictions regarding these effects next. If streaming increases the variety and discovery of users, we would expect the effect to be larger for people with limited variety prior to adoption (i.e., users with a high share of listening to the Top 100 artists, or older users).11 Users on Spotify’s free plan may have less control than paying subscribers over what they consume;12 hence, free users may see less discoveries and varieties, compared to premium subscribers. Our strategy is to incorporate each of these terms as moderators of an adoption effect. To avoid cluttering, we do not distinguish between short-, medium-, and long-term effects but use a single treatment effect (adoption푖푡), which equals one on and after the week of adoption and zero otherwise. This is specified as follows:

Yit = αi + γt + δ ∙ adoptionit + ϕ1(adoptionit ∙ Top100sharei) (3) +ϕ2 (adoptionit ∙ Agei) + ϕ3 (adoptionit ∙ Freei) +εit The parameters of interest are the coefficients for the two-way interaction terms.

Results

Consumption growth and displacement The first questions we address are: does adopting Spotify lead to extra music consumption and, how long do these effects last? In Table 2, column (1), our dependent variable is the log total playcount consumed across all platforms by a given user on a given week. In the week of adoption and the week after, the number of plays grows by 120% (=exp(0.79) – 1). Total consumption is 58% higher (=exp(0.46) – 1) in the medium-term, from two weeks until 24 weeks after Spotify adoption. Even more than 25 weeks (nearly 6 months) after adoption, overall consumption is still 43% higher (=exp(0.36) – 1) than before adopting Spotify.

11 We thank a reviewer for this suggestion. 12 We do not directly observe whether listeners use the free or the premium version of Spotify. However, the free version features 30-second advertisements and the premium version does not. We know the length for each song, and the time elapsed between starting two adjacent songs. This allows us to do a median split of users on the frequency of 25-35 second gaps relative to the number of adjacent song pairs. We use this metric as a proxy for which consumers have the free version with advertisements.

Marketing Science Institute Working Paper Series 17 To what extent does Spotify adoption displace consumption on iTunes and other ownership-based platforms? In column (2) of Table 2, we see that iTunes consumption drops 21% in the first two weeks after Spotify adoption, 30% in the following 22 weeks, and is 33% lower about six months after Spotify adoption. Column (3) shows that consumption on Winamp, Windows Media Player, and Foobar2000 (henceforth, WWF) falls 23% in the first two weeks after Spotify adoption, 34% in the following 22 weeks, and 37% after 6 months. Both displacement effects grow over time. The impact on consumption on all remaining platforms is not significant in the first 24 weeks; after six months there is a 24% drop in column (4). Hence, we conclude that Spotify adoption leads to strong consumption growth. Moreover, this effect persists beyond 24 weeks after users adopt Spotify.13 Secondly, and in accordance with Figure 1, we see that Spotify increasingly displaces consumption from iTunes and WWF.14

Variety consumption Breadth of variety. We next investigate the effect of Spotify adoption on the variety of music consumption. Recall that adopting Spotify lowers the monetary cost of the marginal variety to zero. Hence, to the extent that this cost limits demand for variety, we expect users to consume more variety after adoption. Table 3 presents results for the log total number of unique artists, songs, and genres. We find a significant increase in all measures immediately following Spotify adoption. In the first two weeks after adoption, the number of unique artists heard increases by 63% (= exp(0.49) – 1), the number of unique songs increases by 49%, and the number of unique genres increases by 43%. The effect attenuates over time but is still larger than pre-adoption levels both statistically and substantively. For instance, 2-24 weeks after adoption the number of unique artists consumed weekly is 34% higher than pre-adoption levels. Similarly, after 25 weeks and up to approximately a year after adoption, we measure a 36% increase in the number of unique

13 Our data allows us to estimate these long-term effects reliably. For example, we observe 298 adopters for more than 24 weeks, 220 for more than 36 weeks, 125 for more than 48 weeks, and 79 for more than 52 weeks. 14 Spotify and iTunes/WWF may be compliments even if the estimated iTunes/WWF consumption effects are negative: users may sample on Spotify before buying on iTunes, decreasing purchases on iTunes but increasing preference fit due to pre-purchase sampling. We use repeat consumption not-on-Spotify as a proxy for consumer- preference fit and find that it decreases rather than increases (see new content metrics in Table D7 in the Online Appendix D). Hence, there is little evidence in our data that suggests that consumers purchase better content on iTunes/WWF once they adopt Spotify. We thank a reviewer for this suggestion.

Marketing Science Institute Working Paper Series 18 artists consumed. We see similar patterns for the other measures. These results strongly point to a permanent increase in the breadth of music consumption. Superstar consumption. One avenue through which superstars can arise according to our literature review —even in the absence of talent differences— is that consumers are uncertain about quality and “economize” on learning and search costs (Adler 1985). We argue that Spotify lowers search and learning costs by, among others, letting users costlessly sample their catalog of over 30 million songs and recommending its users playlists. Hence, we expect superstar consumption to decline post-adoption. Because consumption quantity and variety increase after Spotify adoption (cf. Tables 2 and 3), we measure superstar consumption as a share of unique varieties (columns 1–3, Table 4) and of total listening (columns 4–6). In Table 4, we empirically investigate the impact of Spotify adoption on the consumption share of artists in the top 20, 100, and 500. Inclusion in the set of top artists is determined by ranking artists in terms of total plays in an initialization period (cf. Table 1, footnote 1) and geographic region. Users shift their consumption out of the top artists, at least in the short and medium term. Column (2) reports that in the first two weeks following Spotify adoption the consumption share of top 100 artists drops 0.028 in terms of unique varieties, from a pre- adoption baseline of 0.17 (see Online Appendix, Table C6); in terms of plays, column (5) shows the consumption share of top 100 drops 0.020 from a pre-adoption baseline of 0.20. These drops represent a substantial 10% loss in superstar consumption. In the medium term, the superstar consumption share is still 0.016 lower (9% less) than pre-adoption levels in terms of unique artists; in terms of plays, the coefficient is also 0.016 lower (8% less). In the long term, the effects are negative but insignificant, albeit in several cases still substantial in magnitude. A similar pattern holds for the top 500 artists. The pattern is weaker if we define superstars as the top 20.15 Top Two (C2) and Top Ten (C10). In addition to the consumption share of common favorites (superstars) diminishing, consumers may allocate less of their listening to their own personal favorites, i.e., their own (weekly) top artists. Table 5 investigates the share of listening allocated to each user’s weekly top 2 artists, songs, and genres in columns (1) – (3). The results

15 To what extent does the content removal of a very popular artist (Taylor Swift) during our observation period affect our results? Confining our analysis to the period before her content removal (see Online Appendix D, Table D4), we find that the effects are even stronger. Hence, a decrease in superstar consumption is not likely driven by Taylor Swift. We thank a reviewer for making this suggestion.

Marketing Science Institute Working Paper Series 19 show a clear decrease in top listening share, the C2 concentration ratio, following Spotify adoption. For example, the share of listening of the top two artists (column 1) decreases from its pre-adoption baseline 0.46 by 0.075 in the short-term, 0.035 in the medium-term, and 0.044 in the long-term relative to pre-Spotify levels, which amounts to reductions of 16%, 8% and 10% respectively. This pattern holds for the share of listening to the top two artists and top two genres; it also holds if we use the top 10 instead of the top 2 in columns (4) – (6) of Table 5. These results show that music consumption fragments across a wider set of varieties as consumers allocate a smaller fraction of their total time to the top varieties each week. This shift in consumer behavior is persistent: the effect still holds in the long term.

Discovery Consumption of new content. If Spotify lowers the search cost for new music, we would expect that users discover more varieties, i.e., songs they have not consumed previously. We measure the share of listening of new varieties as a function of the number of plays and the number of unique varieties consumed each week. We define newness as a flow metric: an artist, song, or genre can only be new to the user in the first week of consumption. To operationalize newness, we use a long pre-sample history of listening behavior starting on January 6, 2013 (see Table 1, footnote 2). Newness is therefore specific to an individual’s listening history, irrespective of the release date of music. Table 6 shows the results, again as a share of unique varieties in columns (1) – (3) and plays in columns (4) – (6). The adoption of Spotify results in a marked increase in the share of new content consumed at the artist, song and genre level. The variety-share of new artists (column 1) increases substantially from its pre-adoption level of 0.16 by 0.14 in the short-term; it is 0.043 higher in the medium-term, and still 0.025 higher in the long-term. The pattern holds for songs and genres as well as if we measure the shares using play counts instead of distinct varieties. Thus, Spotify adoption triggers an increase in the rate of new variety consumption and we conclude that adopting Spotify makes consumers listen to more new varieties. Repeat consumption, welfare and the value of discovery. A key question about the value of the expansion of variety is whether the new variety has high match value. Empirically the effect of adopting Spotify could go either way, depending on whether consumers have good information about their match value with artists and songs on the market a priori. If they do, then variety expansion from lower cost likely is limited to downward selection into new but less

Marketing Science Institute Working Paper Series 20 preferred content. However, if we believe that music is an experience good (Adler 1985), consumers’ pre-consumption valuations need not be the same as their ex-post valuation (Rob and Waldfogel 2004). In short, consumers may not know their match value a priori. In this case, offering variety at a low cost leads to more experimentation and learning about new artists. It may therefore lead to the discovery of new favorites or upward selection. While a complete welfare analysis is beyond the scope of the paper, we perform two types of analyses to study the direction in which consumers expand variety. First, we use repeat listening as a proxy for value, assuming that consumers will repeatedly consume content they like; we next study repeat-consumption of new versus known music. In Table 7, columns (1) – (3), we measure the amount of known (i.e., to the consumer) artists, songs and genres played more than once as a share of total unique known artists, songs and genres consumed. Adoption of Spotify appears to have a small negative impact on this quantity in the long-term. In the medium- term, we see the share of repeat-consumed known songs increase. One explanation may be that adopters substitute out of known songs into new songs; the known songs abandoned for new ones are likely to be those that are not repeat-consumed, effectively trimming the pool of known songs towards repeat-consumption. To understand the welfare implications of new music discovery, we now repeat this exercise for new music in columns (4) – (6) in Table 7. The share of new artists consumed repeatedly, drops 0.096 directly following adoption. It is 0.046 lower in the medium-term, and it is 0.047 lower in the long-term compared to pre-adoption levels of .15. We conclude that repeat consumption of new artists is lower after the adoption of streaming. Similarly, column (5) shows that the share of repeatedly played new songs drops 0.035 in the short-term and .018 in the medium-term; it is insignificant thereafter. Finally, column (6) reveals that the Spotify effect on share of new genres played repeatedly is negative and substantial in size. It is also significant in the short and long-term. Collectively, these results support that new content found on Spotify is subject to downward selection and that newly discovered music after Spotify adoption is, on average, of lower value. Importantly, though, downward selection on average might simply be masking that consumers listen to larger quantities of new music to find new favorites. Consumers may be better off if their “best” discoveries are of high value. Hence, we investigate the effect of Spotify adoption on the match value of user’ best new artists, songs, and genres. In particular, we first

Marketing Science Institute Working Paper Series 21 construct the share of new music as the consumption that belongs to the user’s top 1 and 5 artists, songs, and genres that are new to the consumer in that week, where consumption is computed over the next eight weeks. Next, we take this share relative to the average consumption of the (not necessarily new) top 1 and 5 artist, song, and genre in that period.16 Hence, our measure is a ratio, similar to a top 1 or top 5 lift for a given user, of her top new varieties to the best overall varieties. This measure is between 0 and 1, depending on whether the user’s best discovery gets hardly any plays, or whether it is the new leader even among past favorites. Column (1) in Table 8 shows that the top new artist lift, i.e., the share of plays to the top new artist, immediately increases by 0.088 for adopters of Spotify. It remains higher than the pre-adoption period in the medium-term and is insignificant in the long-term. Column (2) shows that the same is true for top 5 artists. In Column (3), the share of plays to the top new song increases by 0.13 in the short-term and is 0.049 higher in the medium-term. Even 25+ weeks after adoption, top new songs are consumed more than the top songs discovered before Spotify adoption, suggesting that these new songs have higher match value. Genres reveal similar patterns, albeit weaker in the long-term. To conclude, Table 8 provides considerable evidence that the top new discoveries post-Spotify are consumed more (and hence provide more consumer welfare) after than before adoption. However, our estimate of the effects of adoption on the value on top discoveries becomes noisy in the long-term.

Heterogeneous Treatment Effects There are sizeable effects of adopting Spotify on consumption, variety and discovery. We would expect that effects differ in the population based on individual tastes such as love of variety. As mentioned earlier, we focus on three particular moderators: (1) Taste for popular artists, (2) Old (vs. young) users, and (3) Free (vs. premium) subscribers. We use a median split of users on each of these three moderators and estimate Equation (3). In Table 9, we investigate heterogeneous treatment effects for the share of new content; in Online Appendix E, we provide a full set of heterogeneous treatment effects for all outcome measures. Table 9 presents support across multiple measures that the Spotify adoption effect on

16 To make the columns for top 1 and 5 comparable, we divide the cumulative share of the 5 best (new and overall) by 5, scaling it so that it matches the scale of the top 1 shares. For a few consumers, less than 5 new artists, songs or genres were discovered on a weekly basis. In this case, we use as many new artists, songs, or genres as we can. In even fewer cases, no new content was discovered in a given week. For these cases, we report a 0 as the share of the most popular new artists, songs, or genres.

Marketing Science Institute Working Paper Series 22 variety and discovery is stronger for users who (in the pre-sample period) listen to more Top 100 artists; additionally older users in our sample discover more new content than younger users. These results suggest that the adoption effect is relatively larger for users with limited diversity prior to adoption. For instance, the effect of adoption on the share of plays from new artists (column 4) is close to 0 for the baseline group consisting of those with a below median play share of Top 100 artists, of below median age and who subscribe to the premium plan. But it is large for above median age users whose plays were previously concentrated the Top 100 artists. Last, the adoption effect appears smaller for users on an ad-based (free) subscription; these users typically have less control over what they can listen to, which limits their ability to discover new music.

Selection on Unobservables and Robustness If users who adopt Spotify are systematically different from non-adopters in some way that is not accounted for by the observable characteristics approach in the matching procedure but that affects our dependent measures, our estimated treatment effects will be biased. As a check against this, we re-do the entire analysis using only adopters. Because of the variation of adoption time, we can use late adopters as a control for early adopters. In particular, consider an early adopter who adopts streaming at t and a late adopter who adopts streaming at t +T. Next, we construct a two-way fixed effects DID estimator from comparing the difference in their behavior in periods [0,…,t-1] with the difference in their behavior at [t,…,t +T-1] (Manchanda, Packard, and Pattabhiramaiah 2015). Using this procedure, the short-term and medium-term effects of adoption on the log number of artists are 0.48 and 0.28 (we cannot estimate long-term effects, since the time horizon between early and late adoption is shorter). We can compare this to the short-term and medium- term estimates reported in Table 3, which are 0.49 and 0.29. The estimates are remarkably similar; the same is true for unique songs and genres. Online Appendix D shows this holds also for the other outcome measures. We conclude that our main results that are based on a comparison across adopters and non-adopters are similar to results from a within adopter analysis. This provides further support that our matching procedure is free from (static) selection on unobservables. We also report on a comprehensive set of additional robustness checks. Specifically, in addition to (1) controlling for unobservables using variation in adoption timing, we replicated

Marketing Science Institute Working Paper Series 23 our analysis (2) for different definitions of long term effects (e.g., >36 weeks instead of >24), (3) excluding periods after which Taylor Swift removed content, (4) excluding the last weeks of our data before Spotify improved its recommendations and Apple Music was launched, (5) dropping countries in which Spotify was launched recently, (6) focusing on consumption excluding Spotify, (7) using different functional forms (dependent variables in levels instead of logs, modeling shares), (8) estimating a single treatment effect instead of short-, medium-, and long- term effects, and (9) using the Herfindahl-Hirschman index as an alternative concentration metric (reported as additional variables in Table D5). Online Appendix D provides a full overview of the results. Small numerical differences notwithstanding, all of our reported results are substantively robust.

Implications The emergence of streaming services has provoked a wide-ranging debate about the benefits and drawbacks of ownership- versus streaming-based consumption. We document long-term effects of streaming on music consumption: a half year after users adopt Spotify, consumption, measured in play counts, is up by 43%. By setting the marginal price of variety to zero, Spotify alleviates a deadweight loss problem for varieties where valuation is positive but below the price of ownership. We also provide evidence that Spotify increases consumer welfare by reducing search frictions (e.g., enhancing discovery) and helping users discover new high-value content. Generally, while our results do not reveal a strong negative long-term effect of Spotify on superstar consumption, our remaining results —consumers spreading their consumption across a wider set of artists, discovering more artists, and engaging in less repeat listening— point to Spotify creating a more level playing field for smaller artists. Spotify expands consumers’ attention to a wider set of artists, potentially increasing demand for complementary goods, like live performances (Mortimer, Nosko and Sorensen 2010). The other side of the discovery coin is a drop in the staying power of songs and artists in the consumption set of consumers. With new discoveries and a constraint on time, the share of new artists played more than once drops in the long run. Thus, while it is easier to enter the consumption set, it is harder to stay there. Artists may thus need to exert more effort than before to stay top-of-mind.

Marketing Science Institute Working Paper Series 24 Conclusion Constructing a unique panel data set of music consumption on streaming-, and ownership- based platforms, we demonstrate the short-, medium-, and long-term effects of adoption of online streaming on quantity, variety in consumption, and new music discovery. We find that streaming increases total consumption, leads to more variety, and facilitates discovery of more highly valued music. While our analysis sheds light on the effects of adopting Spotify on music consumption, our data does not allow us to address the underlying mechanism that leads to these effects. We postulate that our effects are largely driven by the price shock on the marginal variety of adopting streaming, yet platform-specific features (e.g., personalized recommendations) may also be important. In particular, we believe that a fruitful area of future research is to examine the role of playlists and sharing of consumption capital on music consumption.

Marketing Science Institute Working Paper Series 25 References Adler, M. (1985). “Stardom and talent,” American Economic Review, 75(1) 208–212.

Aguiar, L. (2015), “Let the Music Play? Free Streaming, Product Discovery, and Digital Music Consumption,” JRC Working Paper.

Aguiar, L., and J. Waldfogel (2015), “Streaming reaches flood stage: Does Spotify stimulate or depress music sales?,” National Bureau of Economic Research.

Alexander, P. (1997), “Product variety and market structure: A new measure and a simple test,” Journal of Economic Behavior & Organization, 32(2) 207–214.

Bertrand, M., E. Duflo, and S. Mullainathan (2004), “How much should we trust differences-in- differences estimates?,” The Quarterly Journal of Economics, 119(1) 249–275.

Bronnenberg, B. (2015), “The provision of convenience and variety by the market,” Rand Journal of Economics, 46(3) 480–498.

Bronnenberg, B., J.-P. Dubé, and C. Mela (2010), “Do digital video recorders influence sales?,” Journal of Marketing Research, 47(6) 998–1010.

Cameron, S., and A. Collins (1997), “Transaction costs and partnerships: The case of rock bands,” Journal of Economic Behavior & Organization, 32(2) 171–183.

Chung, K., and A. Cox (1994), “A stochastic model of superstardom: An application of the Yule distribution,” The Review of Economics and Statistics, 76(4) 771–775.

Chung, T., R. Rust, and M. Wedel (2009), “My mobile music: An adaptive personalization system for digital audio players,” Marketing Science, 28(1) 52–68.

Crain, W., and R. Tollison (2002), “Consumer choice and the popular music industry: A test of the superstar theory,” Empirica 29 1–9.

Elberse, A. (2008), “Should you invest in the long tail?, “ Harvard Business Review, 86(7/8) 88– 96.

Friedlander, Joshua P. 2016, “News and Notes on 2015 RIAA Shipment and Revenue Statistics,” Available at https://www.riaa.com/wp-content/uploads/2016/03/RIAA-2015-Year-End- shipments-memo.pdf [accessed: 2 August 2016].

Gensler, S., P. Leeflang, and B. Skiera (2012), “Impact of online channel use on customer revenues and costs to serve: Considering product portfolios and self-selection,” International Journal of Research in Marketing, 29(2) 192–201.

Hamlen, W. (1991), “Superstardom in popular music: Empirical evidence,” The Review of Economics and Statistics, 73(4) 729–733.

Marketing Science Institute Working Paper Series 26 van Herpen, E., and R. Pieters (2002), “The variety of an assortment: An extension to the attribute-based approach,” Marketing Science, 21(3) 331–341.

Hoch, S. J., E. T. Bradlow, and B. Wansink (1999), “The variety of an assortment,” Marketing Science, 18(4) 527–546.

Holbrook, M., and E. Hirschman (1982), “The experiential aspects of consumption: Consumer fantasies, feelings, and fun,” Journal of Consumer Research, 9(2) 132–140.

Kim, J., G. M. Allenby, and P. E. Rossi (2002), “Modeling consumer demand for variety,” Marketing Science, 21(3) 229–250.

Kissel, C. (2015), “Spotify listeners discover roughly 27 new artists a month,“ Available at: http://diffuser.fm/spotify-listeners-discover-roughly-27-new-artists-a-month/ [accessed: 16 November 2015].

Lacher, K. T., and R. Mizerski (1994), “An exploratory study of the response and relationships involved in the evaluation of, and in the intention to purchase new rock music,” Journal of Consumer Research, 21(2) 366–380.

Lopes, P. (1992), “Innovation and diversity in the popular music industry, 1969 to 1990,” American Sociological Review, 57(1) 56–71.

Manchanda, P., G. Packard, and A. Pattabhiramaiah (2015), “Social Dollars: The Economic Impact of Customer Participation in a Firm-Sponsored Online Customer Community,” Marketing Science, 34(3) 367–387.

Mitroff, S., and X. Blanco (2015), “Apple Music vs Spotify: What’s the difference?,” https://www.cnet.com/news/apple-music-vs-spotify-whats-the-difference/ [accessed 7 October 2016].

Mortimer, J., C. Nosko, and A. Sorensen (2012), “Supply responses to digital distribution: Recorded music and live performances,” Information Economics and Policy, 24(1) 3–14.

Oestreicher-Singer, G., and L. Zalmanson (2013), “Content or community? A digital business strategy for content providers in the social Age,” MIS Quarterly. 37(2) 591–616.

Papke, L.E., and J.M. Wooldridge (1996), “Econometric methods for fractional response variables with an application to 401(K) plan participation rates,” Journal of Applied Econometrics, 11(6) 619–632.

Ratner, R.K., B.E. Kahn, and D. Kahneman (1999), “Choosing Less-Preferred Experiences for the Sake of Variety,” Journal of Consumer Research, 26(1) 1–15.

Rob, R., and J. Waldfogel (2004), “Piracy on the high C's: Music downloading, sales displacement, and social welfare in a sample of college students,” National Bureau of Economic Research.

Marketing Science Institute Working Paper Series 27 Rosen, S. (1981), “The economics of superstars,” American Economic Review, 71(5) 845–858.

Rosenbaum, P., and D. B. Rubin (1983), “The central role of the propensity score in observational studies for causal effects,” Biometrika 70(1) 41–55.

Spotify (2016), “Fast Facts,” Available at: https://press.spotify.com/us/about/ [accessed 13 October 2016].

Waldfogel, J. (2012), “Music Piracy and Its Effects on Demand, Supply, and Welfare,” National Bureau of Economic Research.

Wlömert, N., and D. Papies (2016), “On-demand streaming services and music industry revenues—Insights from Spotify’s market entry,” International Journal of Research in Marketing, 33(2) 314–327.

Zentner, A. (2006), “Measuring the effect of file sharing on music purchases,” Journal of Law & Economics, 49(1) 63–90.

Marketing Science Institute Working Paper Series 28 Table 1. Variable Operationalization Dimension Operationalization (1) Quantity • Log number of song plays (2) Variety Breadth • Log number of unique artists, songs, and genres listened to

Concentration • Number of unique / amount of plays to popular artists in a user’s geographic region (defined as the top 20, 100, and 500 artists listened to by users in a specific geographic region in an initialization period1), divided by the number of unique artists / total number of plays • Share of playcounts to a user’s weekly top two and top ten artists, songs, and genres (C2 and C10) (3) Discovery New content • Number of distinct new artists, songs, and genres listened to by a user consumption for the first time,2 divided by the total number of distinct artists, songs, and genres listened to • Amount of plays to new artists, songs, and genres listened to by a user for the first time,2 divided by the total number of plays Repeat • Number of unique new and known (i.e., not new) artists, songs, and consumption genres played more than once, divided by the total number of unique new and known artists, songs, and genres listened to • Amount of plays to the top 1 and 5 new artists, songs, and genres in the week of discovery and subsequent seven weeks (t, t+1, ..., t+7) ranked in order of playcounts, divided by the amount of plays to the overall (not necessarily new) top 1 and 5 artists, songs, and genres over the same time period. Notes: All variables are computed at the user-week level. 1 Initialization period is users’ listening history from 6 January 2013 until the start of our sample on 29 May 2014. Variables are computed at the user-week level, accounting for a user’s geographic region (European Union, South America, USA and Canada, and others). 2 A user’s first week of consumption, based on the users’ listening history on the service up to 6 January 2013.

Marketing Science Institute Working Paper Series 29 Table 2. Consumption Growth and Displacement across Platforms (1) (2) (3) (4) Log playcounts Log playcounts Log playcounts on Log playcounts on all platforms on iTunes Winamp, on other Windows Media platforms Player and Foobar2000 short-term (0-1) 0.79*** -0.24*** -0.26*** 0.041 (0.066) (0.069) (0.073) (0.069) medium-term (2-24) 0.46*** -0.35*** -0.42*** -0.10 (0.070) (0.078) (0.084) (0.078) long-term (25+) 0.36*** -0.40*** -0.47*** -0.27** (0.095) (0.11) (0.12) (0.10) R-squared 0.50 0.75 0.71 0.63 F 83.0 6.96 13.7 12.8 p-value 0.000 0.000 0.000 0.000 users 894 894 894 894 observations 52228 52228 52228 52228 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week. The dependent variable is the log number of songs heard by a panelist on a week (playcount). The independent variables are indicators for a user's time since adoption of Spotify, defined as short-term (within weeks 0 and 1), medium-term (within weeks 2 and 24), and long-term (weeks 25 and after). + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 30 Table 3. Breadth of variety: Unique Number of Consumed Artists, Songs, and Genres (1) (2) (3) Log number of Log number of Log number of unique artists unique songs unique genres short-term (0-1) 0.49*** 0.40*** 0.36*** (0.041) (0.043) (0.030) medium-term (2-24) 0.29*** 0.27*** 0.22*** (0.039) (0.041) (0.029) long-term (25+) 0.31*** 0.28*** 0.22*** (0.054) (0.057) (0.040) R-squared 0.56 0.54 0.56 F 53.3 69.6 45.6 p-value 0.000 0.000 0.000 users 894 894 894 observations 46585 46585 46585 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the log number of distinct artists, songs, and genres heard by a panelist on a week. The independent variables are indicators for a user's time since adoption of Spotify, defined as short-term (within weeks 0 and 1), medium-term (within weeks 2 and 24), and long-term (weeks 25 and after). + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 31 Table 4. Concentration of Variety (1): Share of Superstar Consumption (1) (2) (3) (4) (5) (6) Top 20 artists Top 100 artists Top 500 artists Top 20 artists Top 100 artists Top 500 artists (share of unique (share of unique (share of unique (share of all (share of all (share of all artists) artists) artists) plays) plays) plays) short-term (0-1) -0.0085** -0.028*** -0.036*** -0.0066 -0.020** -0.023** (0.0027) (0.0052) (0.0070) (0.0043) (0.0071) (0.0088) medium-term (2-24) -0.0053* -0.016*** -0.022*** -0.0045 -0.016* -0.017* (0.0023) (0.0048) (0.0061) (0.0035) (0.0064) (0.0074) long-term (25+) -0.0024 -0.0072 -0.022* 0.0016 -0.0050 -0.016 (0.0028) (0.0063) (0.0086) (0.0048) (0.0085) (0.010) R-squared 0.42 0.54 0.61 0.42 0.52 0.55 F 3.52 6.68 15.2 3.44 6.59 13.2 p-value 0.000 0.000 0.000 0.000 0.000 0.000 users 894 894 894 894 894 894 observations 46585 46585 46585 46585 46585 46585 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the number of unique / amount of plays to popular artists (measured in an initialization period) in a user's geographic region, divided by the number of unique artists / total number of plays. The independent variables are indicators for a user's time since adoption of Spotify, defined as short-term (within weeks 0 and 1), medium-term (within weeks 2 and 24), and long-term (weeks 25 and after). + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 32 Table 5. Concentration of Variety (2): Share of Top Two (C2) and Top Ten (C10) (1) (2) (3) (4) (5) (6) Artist Song Genre Artist Song Genre concentration concentration concentration concentration concentration concentration (C2) (C2) (C2) (C10) (C10) (C10) short-term (0-1) -0.075*** -0.032*** -0.048*** -0.067*** -0.071*** -0.018*** (0.0084) (0.0051) (0.0068) (0.0076) (0.0088) (0.0025) medium-term (2-24) -0.035*** -0.024*** -0.028*** -0.035*** -0.044*** -0.012*** (0.0077) (0.0041) (0.0055) (0.0071) (0.0079) (0.0024) long-term (25+) -0.044*** -0.024*** -0.029*** -0.044*** -0.047*** -0.015*** (0.011) (0.0061) (0.0079) (0.0097) (0.011) (0.0033) R-squared 0.45 0.36 0.53 0.58 0.45 0.60 F 21.4 8.23 13.5 18.6 20.9 7.94 p-value 0.000 0.000 0.000 0.000 0.000 0.000 users 894 894 894 894 894 894 observations 46585 46585 46585 46585 46585 46585 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the share of playcounts to a user's weekly top two and top ten artists, songs, and genres (C2 and C10). The independent variables are indicators for a user's time since adoption of Spotify, defined as short-term (within weeks 0 and 1), medium-term (within weeks 2 and 24), and long-term (weeks 25 and after). + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 33 Table 6. Discovery of New Content (1): New Artists, Songs, and Genres Consumed (1) (2) (3) (4) (5) (6) New artists New songs New genres New artists New songs New genres (share of unique (share of unique (share of unique (share of all (share of all (share of all artists) songs) genres) plays) plays) plays) short-term (0-1) 0.14*** 0.15*** 0.065*** 0.096*** 0.14*** 0.021*** (0.0097) (0.010) (0.0057) (0.0093) (0.010) (0.0037) medium-term (2-24) 0.043*** 0.050*** 0.019*** 0.020*** 0.040*** 0.0046** (0.0060) (0.0077) (0.0023) (0.0058) (0.0078) (0.0017) long-term (25+) 0.025** 0.022* 0.011*** 0.0030 0.016 0.0013 (0.0079) (0.010) (0.0031) (0.0077) (0.010) (0.0022) R-squared 0.35 0.43 0.11 0.29 0.38 0.073 F 24.5 17.2 4.29 15.9 14.7 1.21 p-value 0.000 0.000 0.000 0.000 0.000 0.128 users 894 894 894 894 894 894 observations 46585 46585 46585 46585 46585 46585 Notes: Regression model with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the number of distinct / amount of plays to distinct new artists, songs, and genres listened to by a user for the first time, divided by the total number of distinct artists, songs, and genres / total number of plays. New artists, songs, and genres are defined by a user's first week of consumption on the service up to 6 January 2013. The independent variables are indicators for a user's time since adoption of Spotify, defined as short-term (within weeks 0 and 1), medium-term (within weeks 2 and 24), and long-term (weeks 25 and after), and the log of unique artists, songs, and genres. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 34 Table 7. Discovery of New Content (2): Repeat Consumption (Known versus New) (1) (2) (3) (4) (5) (6) Known artists Known songs Known genres New artists New songs New genres played more played more played more played more played more played more than once than once than once than once than once than once (share of unique (share of unique (share of unique (share of unique (share of unique (share of unique known artists) known songs) known genres) new artists) new songs) new genres) short-term (0-1) -0.033*** 0.0028 -0.019** -0.096*** -0.035*** -0.089*** (0.0079) (0.0064) (0.0063) (0.013) (0.0075) (0.022) medium-term (2-24) -0.011 0.015** -0.0096+ -0.046*** -0.018** -0.051** (0.0069) (0.0054) (0.0051) (0.010) (0.0064) (0.017) long-term (25+) -0.016+ 0.0055 -0.012+ -0.047*** -0.015 -0.056* (0.0093) (0.0077) (0.0069) (0.014) (0.0091) (0.022) R-squared 0.45 0.45 0.32 0.33 0.33 0.27 F 1.95 4.99 1.74 2.49 4.05 1.55 p-value 0.000 0.000 0.000 0.000 0.000 0.004 users 894 894 894 894 894 887 observations 46114 45483 46527 35439 43720 15732 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the number of unique known (versus new) artists, songs, and genres played more than once, divided by the total number of unique known (versus new) artists, songs, and genres listened to. New artists, songs, and genres are defined by a user's first week of consumption on the service up to 6 January 2013. The independent variables are indicators for a user's time since adoption of Spotify, defined as short-term (within weeks 0 and 1), medium-term (within weeks 2 and 24), and long-term (weeks 25 and after). + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 35 Table 8. Discovery of New Content (3): Value of Top Discoveries (1) (2) (3) (4) (5) (6) Share of top 1 Share of top 5 Share of top 1 Share of top 5 Share of top 1 Share of top 5 new artist to new artist to new song to new song to new genre to new genre to overall top 1 overall top 5 overall top 1 overall top 5 overall top 1 overall top 5 artist artist song song genre genre short-term (0-1) 0.088*** 0.066*** 0.13*** 0.13*** 0.017*** 0.0089*** (0.017) (0.012) (0.019) (0.018) (0.0050) (0.0025) medium-term (2-24) 0.026* 0.013+ 0.049** 0.039** 0.0061* 0.0025+ (0.012) (0.0078) (0.016) (0.014) (0.0025) (0.0014) long-term (25+) 0.0097 0.0015 0.041+ 0.027 0.0050 0.0020 (0.017) (0.011) (0.022) (0.019) (0.0036) (0.0020) R-squared 0.22 0.25 0.22 0.26 0.077 0.079 F 16.0 17.5 9.75 11.6 1.40 1.50 p-value 0.000 0.000 0.000 0.000 0.033 0.012 users 883 883 882 882 883 883 observations 32448 32448 31702 31702 32448 32448 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week with at least one new song played. The dependent variable is the amount of plays to the top 1 and 5 new artists, songs, and genres in the eight week subsequent to discovery (t+1, ..., t+8) ranked in order of plays, divided by the amount of plays to the overall (not necessarily new) top 1 and 5 artists, songs, and genres over the same time period. Observations are excluded when the rolling 8-week window includes both pre-adoption and post-adoption periods, and when there are fewer than 8 weeks remaining at the end of each user's observation period. New artists, songs, and genres are defined by a user's first week of consumption on the service up to 6 January 2013. The independent variables are indicators for a user's time since adoption of Spotify, defined as short-term (within weeks 0 and 1), medium-term (within weeks 2 and 24), and long-term (weeks 25 and after). + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 36 Table 9. Discovery of New Content (1): New Artists, Songs, and Genres Consumed (1) (2) (3) (4) (5) (6)

New artists New songs New genres New artists New songs New genres (share of unique (share of unique (share of unique (share of all (share of all (share of all artists) songs) genres) plays) plays) plays) Adoption 0.037** 0.039* 0.017*** 0.0049 0.031* 0.0031 (0.012) (0.015) (0.0045) (0.012) (0.015) (0.0032) x Top 100 artists 0.020+ 0.022 0.014*** 0.019+ 0.018 0.0059* (play share) (0.011) (0.014) (0.0041) (0.010) (0.014) (0.0028) x Age 0.026* 0.036** 0.0061 0.030** 0.037** 0.0028 (0.010) (0.014) (0.0042) (0.0099) (0.014) (0.0030) x Free (vs -0.013 -0.017 -0.0070+ -0.0059 -0.020 -0.0022 premium) (0.011) (0.014) (0.0042) (0.010) (0.014) (0.0031) R-squared 0.34 0.42 0.10 0.29 0.38 0.072 F 22.3 14.0 3.77 14.6 11.8 1.08 p-value 0.000 0.000 0.000 0.000 0.000 0.313 users 894 894 894 894 894 894 observations 46585 46585 46585 46585 46585 46585 Notes: Regression model with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the number of distinct / amount of plays to distinct new artists, songs, and genres listened to by a user for the first time, divided by the total number of distinct artists, songs, and genres / total number of plays. New artists, songs, and genres are defined by a user's first week of consumption on the service up to 6 January 2013. The independent variables are indicators for a user's adoption of Spotify, and interaction effects with pre- sample measures to capture heterogeneous treatment effects. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 37

Figure 1. Market share (playcounts) in the raw data

Marketing Science Institute Working Paper Series 38

Figure 2. Distribution of Propensity Scores Before and After Matching

Marketing Science Institute Working Paper Series 39 Online Appendix A. Algorithm to identify unique artists and songs

The aim of this algorithm is twofold: First, we need to identify unique artists and songs (e.g., to make sure that “The Beatles” and “beatles” are counted as one artist, and not as two). Second, we need to link each unique artist in our data set to additional databases the open source music encyclopedia Musicbrainz.org and the music intelligence company Echonest.com which powers Spotify’s music catalogue to construct our variety measures (e.g., based on meta- characteristics such as an artist’s genre). Identifying and Matching Artists The raw data from the service contain information on each artist’s name as used on the service platform, and if available an artist’s Musicbrainz ID (MBID). We use a combination of clear-text names and MBIDs to establish the linkage with Echonest.com. We perform our matching procedure in the following steps, whereby each subsequent step is performed on the unmatched cases resulting from the previous step: 1) For each artist in the service data set, obtain the corresponding Echonest ID by querying Echonest for an artist’s MBID 2) For each unmatched artist, perform a fuzzy match to obtain an artist’s Echonest ID, using Echonest’s Developer API querying for an artist’s clear-text name. 3) For each unmatched artist (i.e., for artists without an Echonest ID), consider those artists identified whose MBID is non-missing in the service data set. 4) For each unmatched artist, link to already identified artists based on variants of their clear-text names in the following order: a. Put artist names to lower case and remove leading and trailing spaces (e.g., to match “Beatles” and “ beatles”) b. Remove & and “and” (e.g., to match “Mumford & Sons” with “Mumford and Sons”) c. Remove articles (“the” and “a”) (e.g., to match “The Flying Burrito Brothers” to “Flying Burrito Brothers”) d. Remove non-alphanumeric characters (e.g., to combine “Chef’special” with “Chef special”) e. Account for collaborations (e.g., “Patrice, Keziah Jones”) by retaining only the foremost artist (“Patrice”). We implement this by removing text after the collaboration qualifier (feat, featuring, vs, versus, with, dash (-), slash (/), semicolon (;), plus (+), and (&, and), comma (,)). We also replace numbers by letters (e.g., to match “30 seconds to Mars” to “Thirty Seconds to Mars”), and remove double spaces.

Identifying and Matching Songs The raw service data set contain information on each songs’ name, but no universal identifier that can be readily used to cross-reference the data with Musicbrainz.org or Echonest.com. Therefore, we identify unique songs exclusively by their clear-text name. We perform the

Marketing Science Institute Working Paper Series 40 algorithm for each artist identified in the previous step after making changes to the recorded song names as described below: 1) Put song names to lower case and remove leading and trailing spaces (e.g., to match Racoon’s “Good To See You” to “good to see you”) 2) Replace “&” by “and” (e.g., to match Capital Cities’ “Safe and Sound” to “Safe & Sound) 3) Remove articles “a” and “the” (e.g., to match Avicii’s “The Nights” to “Nights”) 4) Retain only alphanumeric characters (e.g., to match Patrice’s “Have You Seen It” to “Have You Seen It?”) Note that we do not remove text in brackets or after a dash (“ – “) to account for a song’s remixes (e.g., Curtis Mayfield’s “Move on up – Extended Version” or Charley Winston’s “Too Long (Radio Edit)”)

Marketing Science Institute Working Paper Series 41 Online Appendix B. Innovations at Spotify

Table B1. Innovations at Spotify during our observation period Date Event Source Launch in new markets 2014/05/27 Launch in Brazil https://news.spotify.com/us/2014/05/28/brazil/ 2014/09/30 Launch in Canada https://news.spotify.com/us/2014/09/30/hello-canada- spotify-here/ Hardware cooperations 2014/06/16 Cooperation with https://news.spotify.com/us/2014/06/16/new-samsung- Samsung wireless-audio-multi-room-speakers-the-first-multi-room- speakers-with-spotify-connect/ 2014/07/21 Cooperation with https://news.spotify.com/us/2014/07/21/connect-to-some- Libratone good-vibrations 2014/07/24 Cooperation with US https://news.spotify.com/us/2014/07/24/our-tv-app-is- Smart TV manufacturer now-on-vizio-internet-apps-plus-smart-tvs 2014/08/26 Promotion when https://news.spotify.com/us/2014/08/26/spotify-premium- purchasing Mini Jambox free-for-3-months-with-a-mini-jambox-or-big-jambox/ or Big Jambox 2014/08/27 Cooperations with Bose, https://press.spotify.com/us/2014/08/27/spotify-connect- Panasonic, and turns-one-new-partners-new-devices-and-now-on-your- Gramofon smart-tv/ 2014/09/02 Cooperation with Denon https://news.spotify.com/us/2014/09/02/get-amazing- multi-room-sound-with-the-new-heos-by-denon-and- spotify-connect/ 2014/09/15 Cooperation with https://news.spotify.com/us/2014/09/15/enjoy-millions- Amazon Fire of-songs-on-your-amazon-fire-tv-with-spotify-connect-2/ 2014/10/08 Cooperation with Bose https://news.spotify.com/us/2014/10/08/spotify-connect- is-now-available-on-soundtouch-the-new-multi-room- sound-systems-from-bose/ 2014/11/07 Spotify connect for https://news.spotify.com/us/2014/11/07/connect-for- computers computers/ 2014/11/17 Cooperation with Uber https://news.spotify.com/us/2014/11/17/uber/ 2014/11/18 Cooperation with BMW https://news.spotify.com/us/2014/11/18/bmw-and-mini- and Mini bring-spotify-into-the-car/ 2015/01/28 Cooperation with https://news.spotify.com/us/2015/01/28/hello-playstation- Playstation 4 spotify-here 2015/03/30 Cooperation with https://news.spotify.com/us/2015/03/30/spotify-on- Playstation Music playstation-music-is-now-available/ 2015/06/23 Cooperation with Ford’s https://news.spotify.com/us/2015/06/23/hit-the-road- SYNC3 entertainment with-spotify-in-ford-vehicles/ system 2015/07/20 Nike+ Running App https://news.spotify.com/us/2015/07/20/nike-running- integrates Spotify delivers-new-ways-to-motivate-runners-through-music/

Marketing Science Institute Working Paper Series 42 Content availability 2014/10/07 John Lennon content https://press.spotify.com/us/2014/10/07/the-complete- now available john-lennon-catalogue-now-available-on-spotify/ 2014/11/03 Taylor Swift removes https://news.spotify.com/us/2014/11/03/taylor-swifts- content decision/ 2014/11/26 Rammstein content now https://news.spotify.com/us/2014/11/26/rammstein-now- available on-spotify/ Recommendations and music discovery 2014/07/02 New dinner and sleep https://news.spotify.com/us/2014/07/02/fill-your- browse categories evenings-with-music-introducing-dinner-and-sleep- categories-to-browse/ 2014/09/19 New browse category https://news.spotify.com/us/2014/09/19/new-browse- “Folk Americana” category-folk-americana 2014/12/11 Introducing “Top tracks https://news.spotify.com/us/2014/12/11/top-tracks-in- in your own network” your-network/ 2015/04/28 Perfect Playlists at Your https://news.spotify.com/us/2015/04/28/perfect-playlists- Fingertips at-your-fingertips/ 2015/05/20 Recommendations based https://news.spotify.com/us/2015/05/20/say-hello-to-the- on time of day and most-entertaining-spotify-ever/ running tempo, and release of video clips and exclusive audio content 2015/06/15 Facilitating exploration https://news.spotify.com/us/2015/06/15/taste-rewind/ of music from earlier decades 2015/07/20 Launch of Discovery https://press.spotify.com/it/2015/07/20/introducing- Weekly (personalized discover-weekly-your-ultimate-personalised-playlist/ playlists) Client updates 2014/05/13 Update for Windows https://news.spotify.com/us/2014/05/13/introducing- Phone discover-and-browse-to-windows-phone/ 2014/07/26 Removing fee for https://news.spotify.com/us/2014/08/26/windows-phone- Windows Phone client free/ 2014/10/30 New look for iPad App https://news.spotify.com/us/2014/10/30/new-look- spotify-for-ipad/ 2015/01/19 New look for Windows https://news.spotify.com/us/2015/01/19/new-look-for- Phone App windows-phone 2015/01/22 Better song preview (by https://news.spotify.com/us/2015/01/22/touch-preview/ touch) 2015/02/26 Introducing Lyrics in https://news.spotify.com/us/2015/02/26/desktop-with- Desktop client lyrics-musixmatch/ 2015/05/28 Rollout to Android Ware https://news.spotify.com/us/2015/05/28/spotify-arrives- to-android-wear/ Other 2014/05/27 Data of one user was https://news.spotify.com/us/2014/05/27/important-notice- accessed illegally to-our-users/

Marketing Science Institute Working Paper Series 43 2014/08/13 Selling band https://news.spotify.com/us/2014/08/13/bandpage-merch/ merchandise 2014/09/08 Video ads and sponsored https://news.spotify.com/us/2014/09/08/introducing- sessions spotify-for-brands-video-ads/ 2014/09/10 Introducing Spotify Fan https://news.spotify.com/us/2014/09/10/introducing- insights spotify-insights/ 2014/10/20 Introducing Spotify https://news.spotify.com/us/2014/10/20/introducing- family subscriptions at spotify-family-one-account-for-the-whole-band/ reduced rates 2015/04/16 Playlist targeting for https://press.spotify.com/us/2015/04/16/spotify-launches- brands playlist-targeting-for-brands/ 2015/05/18 Starbucks and Spotify https://news.spotify.com/us/2015/05/18/starbucks-and- redefine retail experience spotify-redefine-retail-experience-by-connecting-spotify- streaming-music-service-with-world-class-store-and- digital-platform/ 2015/06/30 Launch of Apple Music https://www.apple.com/pr/library/2015/06/08Introducing- Apple-Music-All-The-Ways-You-Love-Music-All-in- One-Place-.html

Notes: Retrieved from https://press.spotify.com and https://blog.spotify.com [accessed 14 December 2015], if not indicated otherwise.

Marketing Science Institute Working Paper Series 44 Online Appendix C. Descriptive statistics, matching procedure, and placebo tests

Table C1. Summary Statistics: Quantity Spotify All iTunes Winamp, Other Windows Media Player and Foobar2000 Adopters (pre-adoption) mean 0 218.5 71.5 107.1 39.8 sd 0 293.2 184.4 238.2 97.3 min 0 0 0 0 0 max 0 2721 2651 2721 1755 Adopters (post-adoption) mean 79.8 240.7 56.1 63.1 41.7 sd 170.8 280.2 158.2 162.1 116.3 min 0 0 0 0 0 max 2439 2450 2399 2287 2000 Non-adopters (pre-adoption) mean 0 219.1 72.7 110.6 35.8 sd 0 255.5 188.2 198.9 101.8 min 0 0 0 0 0 max 0 2462 2462 1775 1829 Non-adopters (post- adoption) mean 0 196.1 59.7 94.6 41.8 sd 0 264.0 178.8 193.6 130.9 min 0 0 0 0 0 max 0 2493 2493 2043 2432 Observations 52228 Notes: Summary statistics are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; the unit of analysis is the user-week.

Marketing Science Institute Working Paper Series 45 Table C2. Summary Statistics: Variety mean sd min max Breadth of variety - Number of unique artists 36.59 49.67 1.00 882.00 - Number of unique songs 152.85 163.35 1.00 2192.00 - Number of unique genres 13.65 12.61 1.00 209.00 Concentration of variety - Top 20 artists (share of unique artists) .05 .09 .00 1.00 - Top 100 artists (share of unique artists) .14 .17 .00 1.00 - Top 500 artists (share of unique artists) .32 .25 .00 1.00 - Top 20 artists (share of all plays) .06 .13 .00 1.00 - Top 100 artists (share of all plays) .17 .23 .00 1.00 - Top 500 artists (share of all plays) .37 .29 .00 1.00 - Artist concentration (C2) .45 .26 .02 1.00 - Song concentration (C2) .13 .16 .00 1.00 - Genre concentration (C2) .67 .22 .14 1.00 - Artist concentration (C10) .79 .22 .06 1.00 - Song concentration (C10) .33 .26 .01 1.00 - Genre concentration (C10) .95 .08 .48 1.00 Discovery of new content - New artists (share of unique artists) .18 .20 .00 1.00 - New songs (share of unique songs) .35 .26 .00 1.00 - New genres (share of unique genres) .04 .09 .00 1.00 - New artists (share of all plays) .16 .21 .00 1.00 - New songs (share of all plays) .36 .27 .00 1.00 - New genres (share of all plays) .02 .07 .00 1.00 Repeat consumption - Known artists played more than once (share of unique known .67 .25 .00 1.00 artists) - Known songs played more than once (share of unique known .21 .21 .00 1.00 songs) - Known genres played more than once (share of unique known .77 .20 .00 1.00 genres) - New artists played more than once (share of unique new artists) .59 .37 .00 1.00 - New songs played more than once (share of unique new songs) .22 .25 .00 1.00 - New genres played more than once (share of unique new .56 .45 .00 1.00 genres) - Share of top 1 new artist to overall top 1 artist .20 .32 .00 1.00 - Share of top 5 new artist to overall top 5 artist .12 .21 .00 1.00 - Share of top 1 new song to overall top 1 song .55 .40 .00 1.00 - Share of top 5 new song to overall top 5 song .45 .36 .00 1.00 - Share of top 1 new genre to overall top 1 genre .01 .07 .00 1.00 - Share of top 5 new genre to overall top 5 genre .01 .04 .00 1.00 Notes: Summary statistics are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; the unit of analysis is the user-week when there is at least one song played. The relevant statistics are computed on the basis of non-missing artist characteristics.

Marketing Science Institute Working Paper Series 46 Table C3. Comparison of Adopters and Non-Adopters Before and After Matching (1) (2) (3) (4) All control All treated Matched Matched users users control treated users users Mean/sd Mean/sd Mean/sd Mean/sd Average daily playcount 42.27 49.95 47.03 48.24 (37.57) (43.82) (38.75) (42.77) Number of unique artists 4.00 3.32 3.22 3.40 (4.51) (3.75) (2.59) (3.94) Top 100 artists (share of all plays) 0.15 0.24 0.21 0.22 (0.18) (0.22) (0.22) (0.21) Artist concentration (C2) 0.20 0.21 0.20 0.22 (0.15) (0.14) (0.14) (0.15) Artists played more than once (share of 0.73 0.74 0.74 0.74 unique artists) (0.17) (0.16) (0.15) (0.16) Number of platforms used 2.87 3.17 3.06 3.13 (1.01) (1.19) (1.13) (1.21) Gender (female = 1) 0.23 0.26 0.25 0.26 (0.42) (0.44) (0.43) (0.44) Age 24.19 22.41 22.63 22.78 (6.30) (5.90) (4.29) (6.11) European Union (dummy) 0.33 0.29 0.32 0.32 (0.47) (0.45) (0.47) (0.47) South America (dummy) 0.15 0.38 0.32 0.32 (0.36) (0.49) (0.47) (0.47) Canada/US (dummy) 0.10 0.09 0.09 0.10 (0.30) (0.28) (0.28) (0.30) Observations 1471 507 447 447 Notes: Means with standard deviations parentheses, calculated in the initialization period before May 29, 2014; the unit of analysis is an individual user.

Marketing Science Institute Working Paper Series 47 Table C4. Propensity Score Model

Mean effect Std. error

Average daily playcount 0.004* 0.002 Number of unique artists -0.027 0.021 Top 100 artists (share of all plays) 1.191*** 0.324 Artist concentration (C2) -0.749+ 0.451 New artists (share of all plays) -0.961** 0.370 Artists played more than once (share of unique -0.251 0.446 artists) Number of platforms used 0.266*** 0.052 Gender (female = 1) 0.107 0.129 Age -0.025* 0.011 Canada/US (dummy) 0.566** 0.204 European Union (dummy) 0.558*** 0.143 South America (dummy) 1.275*** 0.151 Constant -1.644** 0.539 Observations 1978 Notes: Logit model with standard errors in parentheses. Estimates are calculated on an unmatched sample of 507 adopters and 1471 non-adopters in the initialization period before May 29, 2014; the unit of analysis is an individual user. The dependent variable is whether a user adopted Spotify (adopt=1), or not (adopt=0). + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 48 Table C5. Placebo tests Placebo effect N R-squared Log playcounts on all platforms -0.042 (0.065) 22192 0.564 Log playcounts on iTunes -0.042 (0.061) 22192 0.826 Log playcounts on Winamp, Windows Media Player and -0.036 (0.072) 22192 0.762 Foobar2000 Log playcounts on other platforms 0.102 (0.065) 22192 0.705 Log number of unique artists 0.068* (0.034) 19804 0.603 Log number of unique songs 0.066+ (0.037) 19804 0.595 Log number of unique genres 0.034 (0.026) 19804 0.609 Top 20 artists (share of unique artists) 0.002 (0.002) 19804 0.467 Top 100 artists (share of unique artists) -0.001 (0.004) 19804 0.609 Top 500 artists (share of unique artists) 0.002 (0.006) 19804 0.660 Top 20 artists (share of all plays) -0.001 (0.004) 19804 0.493 Top 100 artists (share of all plays) -0.002 (0.006) 19804 0.591 Top 500 artists (share of all plays) 0.001 (0.007) 19804 0.604 Artist concentration (C2) -0.011 (0.007) 19804 0.511 Song concentration (C2) -0.002 (0.005) 19804 0.400 Genre concentration (C2) -0.002 (0.005) 19804 0.588 Artist concentration (C10) -0.006 (0.006) 19804 0.631 Song concentration (C10) -0.010 (0.007) 19804 0.511 Genre concentration (C10) -0.000 (0.002) 19804 0.650 New artists (share of unique artists) -0.000 (0.006) 19804 0.378 New songs (share of unique songs) 0.004 (0.007) 19804 0.445 New genres (share of unique genres) -0.001 (0.003) 19804 0.140 New artists (share of all plays) 0.002 (0.006) 19804 0.314 New songs (share of all plays) 0.004 (0.008) 19804 0.398 New genres (share of all plays) -0.001 (0.002) 19804 0.113 Known artists played more than once (share of unique known -0.003 (0.007) 19637 0.496 artists) Known songs played more than once (share of unique known -0.004 (0.005) 19441 0.485 songs) Known genres played more than once (share of unique known -0.003 (0.005) 19778 0.362 genres) New artists played more than once (share of unique new artists) -0.008 (0.011) 14414 0.378 New songs played more than once (share of unique new songs) -0.000 (0.007) 18405 0.382 New genres played more than once (share of unique new genres) 0.012 (0.021) 6303 0.360 Share of top 1 new artist to overall top 1 artist 0.007 (0.012) 12985 0.247 Share of top 5 new artist to overall top 5 artist 0.006 (0.008) 12985 0.276 Share of top 1 new song to overall top 1 song 0.014 (0.017) 12717 0.261 Share of top 5 new song to overall top 5 song 0.018 (0.016) 12717 0.304 Share of top 1 new genre to overall top 1 genre 0.002 (0.002) 12985 0.155 Share of top 5 new genre to overall top 5 genre 0.002 (0.001) 12985 0.170 Notes: Placebo adoption regressions with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters in a sample prior to adoption; user- and week-specific fixed effects are used and the unit of analysis is the user-week. The independent variable is an indicator for users' placebo adoption. For each user, this indicator is 0 until half the pre-adoption time period elapses; then the value takes on one until the end of the pre-adoption time period. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 49 Table C6. Summary Statistics in Levels for Adopters and Non-adopters Before and After Adoption (1) (2) (3) (4) Adopters Adopters Non-adopters Non-adopters (pre-adoption) (post-adoption) (pre-adoption) (post-adoption) mean mean mean mean Playcounts on all platforms 239.5 222.8 230.1 180.6 Playcounts on iTunes 81.9 48.7 78.1 52.6 Playcounts on Winamp, Windows Media 110.0 58.5 112.6 89.8 Player and Foobar2000 Playcounts on other platforms 47.5 40.0 39.5 38.2 Number of unique artists 33.1 37.9 35.0 29.8 Number of unique songs 152.0 149.0 149.9 120.6 Number of unique genres 12.6 14.1 12.9 11.5 Top 20 artists (share of unique artists) 0.056 0.041 0.052 0.044 Top 100 artists (share of unique artists) 0.17 0.13 0.16 0.13 Top 500 artists (share of unique artists) 0.38 0.30 0.35 0.29 Top 20 artists (share of all plays) 0.074 0.049 0.071 0.055 Top 100 artists (share of all plays) 0.20 0.15 0.20 0.16 Top 500 artists (share of all plays) 0.43 0.35 0.41 0.33 Artist concentration (C2) 0.46 0.45 0.46 0.50 Song concentration (C2) 0.14 0.13 0.14 0.16 Genre concentration (C2) 0.68 0.66 0.68 0.70 Artist concentration (C10) 0.80 0.79 0.80 0.82 Song concentration (C10) 0.34 0.34 0.35 0.40 Genre concentration (C10) 0.96 0.95 0.96 0.96 New artists (share of unique artists) 0.16 0.24 0.16 0.19 New songs (share of unique songs) 0.33 0.42 0.33 0.37 New genres (share of unique genres) 0.043 0.061 0.043 0.039 New artists (share of all plays) 0.15 0.20 0.15 0.18 New songs (share of all plays) 0.35 0.43 0.35 0.39 New genres (share of all plays) 0.023 0.027 0.023 0.023 Known artists played more than once (share of 0.68 0.65 0.68 0.67 unique known artists) Known songs played more than once (share of 0.21 0.21 0.22 0.21 unique known songs) Known genres played more than once (share of 0.78 0.76 0.78 0.77 unique known genres) New artists played more than once (share of 0.61 0.54 0.63 0.62 unique new artists) New songs played more than once (share of 0.24 0.19 0.25 0.23 unique new songs) New genres played more than once (share of 0.59 0.51 0.60 0.61 unique new genres) Share of top 1 new artist to overall top 1 artist 0.18 0.24 0.18 0.22 Share of top 5 new artist to overall top 5 artist 0.11 0.15 0.11 0.14 Share of top 1 new song to overall top 1 song 0.54 0.60 0.52 0.55 Share of top 5 new song to overall top 5 song 0.44 0.50 0.43 0.46 Share of top 1 new genre to overall top 1 genre 0.013 0.016 0.015 0.013 Share of top 5 new genre to overall top 5 genre 0.0069 0.0090 0.0087 0.0075

Marketing Science Institute Working Paper Series 50

Figure D1. Adoption Timing for Spotify Adopters

Marketing Science Institute Working Paper Series 51

Figure D2. Observation Overlap Before and After Matching

Note: The decreases in %-weekly active users in weeks 2, 12, and 33 are caused by interruptions in our data collection, such that a week consists of less than 7 days. Week 62 is the last week in our sample, covering 3 out of 7 days (30 July – 1 August 2015).

Marketing Science Institute Working Paper Series 52 Online Appendix D. Robustness Checks

We expose our empirical findings to an expansive set of robustness checks. First, the data set allows us to conduct several natural experiments to verify the stability of our findings. Some of these natural experiments are the removal of Taylor Swift content from Spotify, the sudden improvement of recommendation algorithms, or the launch of Apple Music. Second, we verify the robustness of our model specification with regard to its underlying functional forms (e.g., log versus levels, modeling shares), the assessment of the overall adoption effect, and the definition of the long-term impact. Third, we use the data to compute alternative measures of concentration

(Herfindahl, instead of C2, and C10).

Table D1 summarizes all the concerns that may potentially affect our findings, along with details on how we assess their robustness. In what follows, Tables D2-D8 replicate the findings of Tables 2-8, and addition list the results of each robustness checks on the short-, medium-, and long-term adoption effects. For some robustness checks, the definition of the long-term is not available as the sample coverage is less than 24 weeks. For example, Taylor Swift’s content was removed from Spotify on 14 November 2014, 22 weeks into our sample.

Marketing Science Institute Working Paper Series 53 Table D1. Overview of Robustness Checks Concern Proposed Robustness Check 1) Consumers that choose Spotify may be Estimate models with adopters only. Later adopters systematically different in some (median split; 23 weeks into our main sample) act unobserved way from those that remain on as a control for those who have adopted earlier ownership-based platforms like iTunes. (Manchanda, Packard, and Pattabhiramaiah 2015). 2) The long-term effect (>24 weeks, ~6 Allow for a longer long-term period (>36 weeks, months) captures only transitory effects of ~ 9 months). Spotify adoption. 3) The altered supply of music (e.g., the Confine sample to the period before the removal of removal of Taylor Swift’s content from Taylor Swift’s music from Spotify (3 November Spotify may have affected the results. 2014; 22 weeks retained). 4) To what extent does the recommendation Confine sample to the period before 28 April 2015, system of Spotify drive users’ changes in excluding the introduction of Spotify Running, and music consumption? the launch of new (personalized) playlists (46 weeks retained). Further, this period excludes the launch of Apple Music on 30 June 2015. For details on these events, see Online Appendix B. 5) What is the role of innovators versus late Estimate models without matched treated-control adopters? pairs where at least one user comes from Canada or Brazil, two countries in which Spotify was launched during the observation period (514 users retained). We also provide these estimates only for matched treated-control pairs where both users are from Canada or Brazil (160 users retained). 6) Do Spotify adopters also change music Estimate all variety models with the dependent consumption on other platforms, or does variable defined over consumption on all platforms the change in music consumption mostly except Spotify. come from Spotify? 7) The results may not be robust to the Estimate (a) those models for which logged functional forms (e.g., logs, shares) variables have been used in levels, and (b) those models for which the dependent variables represent shares with fractional logit models (Papke and Wooldridge 1996). 8) It is hard to assess the overall effect of Estimate models with only one treatment dummy adoption from the three dummy variables variablea. short-, medium-, and long-term adoption. 9) Verify the robustness with regard to Estimate concentration models with Herfindahl alternative measures of concentration. index, instead of C2 and C10 (reported as additional variables in Table D5) a For expositional purposes, we do not create a separate column for this estimate, but list it under medium-term adoption effects.

Marketing Science Institute Working Paper Series 54 Table D2. Robustness Checks for Consumption Growth and Displacement across Platforms Short-term Medium-term Long-term b b b Log playcounts (all) Main Analysis 0.79*** 0.46*** 0.36*** Control for Unobservables 0.66*** 0.32** Different Long-Term Effect (36 weeks) 0.80*** 0.45*** 0.40*** Before Removal of Taylor Swift 0.65*** 0.33*** Before Improving Recommendations 0.81*** 0.47*** 0.31* Without Launch Countries 0.80*** 0.39*** 0.25* Only With Launch Countries 0.66*** 0.64*** 0.69** Dependent Variable in Levels 63.4*** 35.4*** 19.6 With One Treatment Dummy 0.49*** Log playcounts (iTunes) Main Analysis -0.24*** -0.35*** -0.40*** Control for Unobservables -0.16+ -0.28** Different Long-Term Effect (36 weeks) -0.24*** -0.35*** -0.47*** Before Removal of Taylor Swift -0.16+ -0.27** Before Improving Recommendations -0.22** -0.30*** -0.39** Without Launch Countries -0.26** -0.31** -0.38** Only With Launch Countries -0.054 -0.26+ -0.40 Dependent Variable in Levels -13.7* -14.0* -9.89 With One Treatment Dummy -0.34*** Log playcounts (WWF) Main Analysis -0.26*** -0.42*** -0.47*** Control for Unobservables -0.18* -0.35** Different Long-Term Effect (36 weeks) -0.25*** -0.42*** -0.38** Before Removal of Taylor Swift -0.18* -0.34*** Before Improving Recommendations -0.22** -0.41*** -0.42** Without Launch Countries -0.23* -0.44*** -0.49*** Only With Launch Countries -0.60** -0.56** -0.53+ Dependent Variable in Levels -20.6** -34.3*** -41.1*** With One Treatment Dummy -0.40*** Log playcounts (other) Main Analysis 0.041 -0.10 -0.27** Control for Unobservables -0.022 -0.16 Different Long-Term Effect (36 weeks) 0.048 -0.13 -0.30* Before Removal of Taylor Swift -0.011 -0.16+ Before Improving Recommendations 0.020 -0.14+ -0.29* Without Launch Countries -0.0059 -0.16 -0.32* Only With Launch Countries -0.055 -0.0089 -0.32 Dependent Variable in Levels 1.14 -0.66 -7.95 With One Treatment Dummy -0.100 Notes: Regressions with robust standard errors. Estimates are calculated for the main analysis, and several robustness checks (details are given in Table D1). User- and week-specific fixed effects are used and the unit of analysis is the user-week. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 55 Table D3. Robustness Checks for Breadth of Variety: Unique Number of Consumed Artists, Songs, and Genres Short-term Medium-term Long-term b b b Log number of unique artists Main Analysis 0.49*** 0.29*** 0.31*** Control for Unobservables 0.48*** 0.28*** Different Long-Term Effect (36 weeks) 0.49*** 0.29*** 0.33*** Before Removal of Taylor Swift 0.48*** 0.29*** Before Improving Recommendations 0.50*** 0.28*** 0.29*** Without Launch Countries 0.50*** 0.24*** 0.27*** Only With Launch Countries 0.53*** 0.43*** 0.46*** Consumption Not on Spotify -0.12** -0.24*** -0.25*** Dependent Variable in Levels 14.0*** 9.42*** 11.9*** With One Treatment Dummy 0.32*** Log number of unique songs Main Analysis 0.40*** 0.27*** 0.28*** Control for Unobservables 0.36*** 0.23*** Different Long-Term Effect (36 weeks) 0.39*** 0.27*** 0.27*** Before Removal of Taylor Swift 0.37*** 0.26*** Before Improving Recommendations 0.40*** 0.26*** 0.25*** Without Launch Countries 0.41*** 0.22*** 0.22** Only With Launch Countries 0.40*** 0.37*** 0.42** Consumption Not on Spotify -0.21*** -0.33*** -0.35*** Dependent Variable in Levels 38.6*** 25.5*** 26.2** With One Treatment Dummy 0.28*** Log number of unique genres Main Analysis 0.36*** 0.22*** 0.22*** Control for Unobservables 0.36*** 0.21*** Different Long-Term Effect (36 weeks) 0.36*** 0.22*** 0.24*** Before Removal of Taylor Swift 0.37*** 0.21*** Before Improving Recommendations 0.37*** 0.21*** 0.20*** Without Launch Countries 0.38*** 0.19*** 0.20*** Only With Launch Countries 0.39*** 0.31*** 0.29** Consumption Not on Spotify -0.085* -0.17*** -0.20*** Dependent Variable in Levels 4.07*** 2.81*** 3.32*** With One Treatment Dummy 0.24*** Notes: Regressions with robust standard errors. Estimates are calculated for the main analysis, and several robustness checks (details are given in Table D1). User- and week-specific fixed effects are used and the unit of analysis is the user-week. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 56 Table D4. Robustness Checks for Concentration of Variety (1): Share of Superstar Consumption Short-term Medium-term Long-term b b b Top 20 (unique share) Main Analysis -0.0085** -0.0053* -0.0024 Control for Unobservables -0.0032 -0.0049 Different Long-Term Effect (36 weeks) -0.0085** -0.0049* -0.00088 Before Removal of Taylor Swift -0.0046 -0.0068* Before Improving Recommendations -0.0076** -0.0044+ -0.0022 Without Launch Countries -0.0076** -0.0040 -0.0032 Only With Launch Countries -0.020* -0.0066 -0.00091 Consumption Not on Spotify 0.0011 0.00084 0.0056+ Fractional Response Model -0.18** -0.11* -0.060 With One Treatment Dummy -0.0055* Top 100 (unique share) Main Analysis -0.028*** -0.016*** -0.0072 Control for Unobservables -0.022** -0.017* Different Long-Term Effect (36 weeks) -0.029*** -0.015** -0.0085 Before Removal of Taylor Swift -0.026*** -0.022*** Before Improving Recommendations -0.027*** -0.017** -0.0027 Without Launch Countries -0.024*** -0.012* -0.0089 Only With Launch Countries -0.052** -0.020 -0.0035 Consumption Not on Spotify -0.0071 0.0013 0.011+ Fractional Response Model -0.22*** -0.12*** -0.057 With One Treatment Dummy -0.017*** Top 500 (unique share) Main Analysis -0.036*** -0.022*** -0.022* Control for Unobservables -0.036*** -0.024** Different Long-Term Effect (36 weeks) -0.037*** -0.022*** -0.032** Before Removal of Taylor Swift -0.040*** -0.029*** Before Improving Recommendations -0.036*** -0.027*** -0.019* Without Launch Countries -0.037*** -0.022** -0.024* Only With Launch Countries -0.045* -0.024 -0.026 Consumption Not on Spotify -0.0068 -0.0055 0.0019 Fractional Response Model -0.16*** -0.088*** -0.090* With One Treatment Dummy -0.024***

Marketing Science Institute Working Paper Series 57 Top 20 (play share) Main Analysis -0.0066 -0.0045 0.0016 Control for Unobservables -0.0059 -0.0049 Different Long-Term Effect (36 weeks) -0.0069 -0.0038 0.0019 Before Removal of Taylor Swift -0.0068 -0.0046 Before Improving Recommendations -0.0061 -0.0038 0.00057 Without Launch Countries -0.0091* -0.0056 -0.0026 Only With Launch Countries -0.021 -0.00071 0.018 Consumption Not on Spotify 0.00093 0.00066 0.010+ Fractional Response Model -0.12+ -0.088 0.0096 With One Treatment Dummy -0.0044 Top 100 (play share) Main Analysis -0.020** -0.016* -0.0050 Control for Unobservables -0.015 -0.012 Different Long-Term Effect (36 weeks) -0.020** -0.014* -0.0063 Before Removal of Taylor Swift -0.019* -0.018+ Before Improving Recommendations -0.017* -0.017* 0.00029 Without Launch Countries -0.014+ -0.0078 -0.0023 Only With Launch Countries -0.047* -0.022 0.0011 Consumption Not on Spotify -0.0016 -0.0016 0.015+ Fractional Response Model -0.13** -0.11** -0.040 With One Treatment Dummy -0.015* Top 500 (play share) Main Analysis -0.023** -0.017* -0.016 Control for Unobservables -0.024* -0.015 Different Long-Term Effect (36 weeks) -0.024** -0.017* -0.025* Before Removal of Taylor Swift -0.027* -0.020+ Before Improving Recommendations -0.018* -0.018* -0.0071 Without Launch Countries -0.023* -0.017* -0.016 Only With Launch Countries -0.040+ -0.025 -0.019 Consumption Not on Spotify -0.0074 -0.0086 0.0038 Fractional Response Model -0.091* -0.063* -0.058 With One Treatment Dummy -0.018* Notes: Regressions with robust standard errors. Estimates are calculated for the main analysis, and several robustness checks (details are given in Table D1 and in the paper). User- and week-specific fixed effects are used and the unit of analysis is the user-week. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 58 Table D5. Robustness Checks for Concentration of Variety (2): Share of Top Two (C2), Top Ten (C10), and Herfindahl Index Short-term Medium-term Long-term b b b Artist concentr. (C2) Main Analysis -0.075*** -0.035*** -0.044*** Control for Unobservables -0.075*** -0.033** Different Long-Term Effect (36 weeks) -0.074*** -0.036*** -0.046*** Before Removal of Taylor Swift -0.074*** -0.033** Before Improving Recommendations -0.074*** -0.034*** -0.045*** Without Launch Countries -0.074*** -0.026* -0.035* Only With Launch Countries -0.097*** -0.069*** -0.092*** Consumption Not on Spotify 0.026** 0.054*** 0.049*** Fractional Response Model -0.30*** -0.14*** -0.18*** With One Treatment Dummy -0.041*** Song concentr. (C2) Main Analysis -0.032*** -0.024*** -0.024*** Control for Unobservables -0.032*** -0.020** Different Long-Term Effect (36 weeks) -0.032*** -0.024*** -0.021** Before Removal of Taylor Swift -0.035*** -0.027*** Before Improving Recommendations -0.034*** -0.024*** -0.026*** Without Launch Countries -0.029*** -0.017** -0.016* Only With Launch Countries -0.033* -0.032** -0.040** Consumption Not on Spotify 0.026*** 0.031*** 0.039*** Fractional Response Model -0.29*** -0.21*** -0.19*** With One Treatment Dummy -0.025*** Genre concentr. (C2) Main Analysis -0.048*** -0.028*** -0.029*** Control for Unobservables -0.056*** -0.028** Different Long-Term Effect (36 weeks) -0.049*** -0.028*** -0.032*** Before Removal of Taylor Swift -0.052*** -0.025** Before Improving Recommendations -0.049*** -0.028*** -0.029** Without Launch Countries -0.048*** -0.022** -0.028** Only With Launch Countries -0.068*** -0.045*** -0.040* Consumption Not on Spotify 0.020* 0.031*** 0.035*** Fractional Response Model -0.21*** -0.13*** -0.13*** With One Treatment Dummy -0.031*** Artist concentr. (C10) Main Analysis -0.067*** -0.035*** -0.044*** Control for Unobservables -0.066*** -0.030** Different Long-Term Effect (36 weeks) -0.067*** -0.036*** -0.050*** Before Removal of Taylor Swift -0.065*** -0.031*** Before Improving Recommendations -0.071*** -0.036*** -0.044*** Without Launch Countries -0.065*** -0.029** -0.043** Only With Launch Countries -0.082*** -0.059*** -0.076** Consumption Not on Spotify 0.013+ 0.030*** 0.024* Fractional Response Model -0.38*** -0.21*** -0.27*** With One Treatment Dummy -0.040*** Song concentr. (C10) Main Analysis -0.071*** -0.044*** -0.047*** Control for Unobservables -0.069*** -0.044*** Different Long-Term Effect (36 weeks) -0.071*** -0.044*** -0.044*** Before Removal of Taylor Swift -0.069*** -0.047*** Before Improving Recommendations -0.077*** -0.046*** -0.051*** Without Launch Countries -0.071*** -0.033** -0.033*

Marketing Science Institute Working Paper Series 59 Only With Launch Countries -0.077*** -0.066** -0.077** Consumption Not on Spotify 0.044*** 0.068*** 0.076*** Fractional Response Model -0.32*** -0.19*** -0.20*** With One Treatment Dummy -0.048*** Genre concentr. (C10) Main Analysis -0.018*** -0.012*** -0.015*** Control for Unobservables -0.019*** -0.0088* Different Long-Term Effect (36 weeks) -0.018*** -0.012*** -0.013*** Before Removal of Taylor Swift -0.019*** -0.0081** Before Improving Recommendations -0.019*** -0.011*** -0.013*** Without Launch Countries -0.021*** -0.010** -0.015** Only With Launch Countries -0.020*** -0.015** -0.014* Consumption Not on Spotify 0.0025 0.0055* 0.0039 Fractional Response Model -0.39*** -0.26*** -0.32*** With One Treatment Dummy -0.013*** Artist concentr. (Herfindahl) Main Analysis -0.065*** -0.033*** -0.037*** Control for Unobservables -0.064*** -0.031** Different Long-Term Effect (36 weeks) -0.065*** -0.034*** -0.038*** Before Removal of Taylor Swift -0.065*** -0.034*** Before Improving Recommendations -0.062*** -0.031*** -0.037*** Without Launch Countries -0.066*** -0.024** -0.025* Only With Launch Countries -0.082*** -0.064*** -0.080*** Consumption Not on Spotify 0.020* 0.046*** 0.047*** Fractional Response Model -0.42*** -0.20*** -0.21*** With One Treatment Dummy -0.038*** Song concentr. (Herfindahl) Main Analysis -0.021*** -0.016*** -0.015*** Control for Unobservables -0.022*** -0.013** Different Long-Term Effect (36 weeks) -0.021*** -0.016*** -0.014** Before Removal of Taylor Swift -0.024*** -0.018*** Before Improving Recommendations -0.021*** -0.015*** -0.016*** Without Launch Countries -0.019*** -0.012*** -0.0094* Only With Launch Countries -0.022** -0.019** -0.024** Consumption Not on Spotify 0.017** 0.019*** 0.026*** Fractional Response Model -0.51*** -0.34*** -0.28*** With One Treatment Dummy -0.016*** Genre concentr. (Herfindahl) Main Analysis -0.061*** -0.035*** -0.030*** Control for Unobservables -0.064*** -0.033** Different Long-Term Effect (36 weeks) -0.062*** -0.035*** -0.038*** Before Removal of Taylor Swift -0.064*** -0.034*** Before Improving Recommendations -0.058*** -0.033*** -0.029** Without Launch Countries -0.063*** -0.032*** -0.028* Only With Launch Countries -0.080*** -0.055*** -0.053* Consumption Not on Spotify 0.015+ 0.035*** 0.045*** Fractional Response Model -0.27*** -0.15*** -0.13*** With One Treatment Dummy -0.038*** Notes: Regressions with robust standard errors. Estimates are calculated for the main analysis, and several robustness checks (details are given in Table D1). User- and week-specific fixed effects are used and the unit of analysis is the user-week. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 60 Table D6. Robustness Checks for Discovery of New Content (1): New Artists, Songs, and Genres Consumed Short-term Medium-term Long-term b b b New artists (unique share) Main Analysis 0.14*** 0.043*** 0.025** Control for Unobservables 0.13*** 0.047*** Different Long-Term Effect (36 weeks) 0.14*** 0.041*** 0.028** Before Removal of Taylor Swift 0.13*** 0.045*** Before Improving Recommendations 0.14*** 0.049*** 0.033*** Without Launch Countries 0.14*** 0.042*** 0.029** Only With Launch Countries 0.16*** 0.058*** 0.035* Consumption Not on Spotify 0.023** -0.0070 -0.016* Fractional Response Model 0.81*** 0.27*** 0.13** With One Treatment Dummy 0.055*** New song (unique share) Main Analysis 0.15*** 0.050*** 0.022* Control for Unobservables 0.14*** 0.064*** Different Long-Term Effect (36 weeks) 0.16*** 0.047*** 0.034** Before Removal of Taylor Swift 0.14*** 0.067*** Before Improving Recommendations 0.15*** 0.057*** 0.032** Without Launch Countries 0.14*** 0.054*** 0.030* Only With Launch Countries 0.21*** 0.068*** 0.027 Consumption Not on Spotify 0.019+ -0.014* -0.021* Fractional Response Model 0.64*** 0.21*** 0.089* With One Treatment Dummy 0.062*** New genre (unique share) Main Analysis 0.065*** 0.019*** 0.011*** Control for Unobservables 0.065*** 0.025*** Different Long-Term Effect (36 weeks) 0.066*** 0.018*** 0.014*** Before Removal of Taylor Swift 0.064*** 0.023*** Before Improving Recommendations 0.068*** 0.021*** 0.016*** Without Launch Countries 0.062*** 0.019*** 0.014*** Only With Launch Countries 0.077*** 0.022*** 0.014 Consumption Not on Spotify 0.0077+ -0.0023 -0.0043 Fractional Response Model 1.05*** 0.42*** 0.26*** With One Treatment Dummy 0.025***

Marketing Science Institute Working Paper Series 61 New artists (play share) Main Analysis 0.096*** 0.020*** 0.0030 Control for Unobservables 0.084*** 0.020* Different Long-Term Effect (36 weeks) 0.097*** 0.018** 0.0073 Before Removal of Taylor Swift 0.084*** 0.017+ Before Improving Recommendations 0.094*** 0.024*** 0.0076 Without Launch Countries 0.096*** 0.022** 0.0096 Only With Launch Countries 0.12*** 0.018 -0.0012 Consumption Not on Spotify 0.0094 -0.017** -0.023** Fractional Response Model 0.61*** 0.14*** 0.012 With One Treatment Dummy 0.029*** New song (play share) Main Analysis 0.14*** 0.040*** 0.016 Control for Unobservables 0.12*** 0.052*** Different Long-Term Effect (36 weeks) 0.14*** 0.037*** 0.026* Before Removal of Taylor Swift 0.12*** 0.052*** Before Improving Recommendations 0.14*** 0.046*** 0.023+ Without Launch Countries 0.13*** 0.046*** 0.024+ Only With Launch Countries 0.18*** 0.052** 0.022 Consumption Not on Spotify 0.011 -0.018* -0.025* Fractional Response Model 0.57*** 0.17*** 0.062 With One Treatment Dummy 0.051*** New genre (play share) Main Analysis 0.021*** 0.0046** 0.0013 Control for Unobservables 0.019*** 0.0061+ Different Long-Term Effect (36 weeks) 0.022*** 0.0042* 0.0036 Before Removal of Taylor Swift 0.017** 0.0042 Before Improving Recommendations 0.022*** 0.0045* 0.00095 Without Launch Countries 0.022*** 0.0070** 0.0051+ Only With Launch Countries 0.021** -0.0012 -0.0051 Consumption Not on Spotify 0.0036 -0.00048 -0.00039 Fractional Response Model 0.72*** 0.17** -0.015 With One Treatment Dummy 0.0066*** Notes: Regressions with robust standard errors. Estimates are calculated for the main analysis, and several robustness checks (details are given in Table D1). User- and week-specific fixed effects are used and the unit of analysis is the user-week. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 62 Table D7. Robustness Checks for Discovery of New Content (2): Repeat Consumption (Known versus New) Short-term Medium-term Long-term b b b Known artists (unique share) Main Analysis -0.033*** -0.011 -0.016+ Control for Unobservables -0.041*** -0.028* Different Long-Term Effect (36 weeks) -0.034*** -0.011+ -0.020+ Before Removal of Taylor Swift -0.046*** -0.031** Before Improving Recommendations -0.037*** -0.010 -0.017 Without Launch Countries -0.029** -0.013 -0.031* Only With Launch Countries -0.052** -0.017 0.0048 Consumption Not on Spotify -0.016+ -0.0087 -0.015 Fractional Response Model -0.15*** -0.049+ -0.071+ With One Treatment Dummy -0.014* Known songs (unique share) Main Analysis 0.0028 0.015** 0.0055 Control for Unobservables -0.0040 0.0088 Different Long-Term Effect (36 weeks) 0.0032 0.014** 0.0042 Before Removal of Taylor Swift -0.0031 0.0099 Before Improving Recommendations 0.0022 0.017** 0.0057 Without Launch Countries 0.014 0.012+ -0.0014 Only With Launch Countries -0.033* 0.019 0.023 Consumption Not on Spotify 0.00083 0.0038 -0.0067 Fractional Response Model 0.017 0.094** 0.041 With One Treatment Dummy 0.013* Known genres (unique share) Main Analysis -0.019** -0.0096+ -0.012+ Control for Unobservables -0.023* -0.012 Different Long-Term Effect (36 weeks) -0.019** -0.010+ -0.018* Before Removal of Taylor Swift -0.031** -0.019* Before Improving Recommendations -0.021** -0.010+ -0.017* Without Launch Countries -0.016* -0.0054 -0.020* Only With Launch Countries -0.029+ -0.022+ -0.0014 Consumption Not on Spotify -0.020** -0.017** -0.018* Fractional Response Model -0.11** -0.050+ -0.067+ With One Treatment Dummy -0.011*

Marketing Science Institute Working Paper Series 63 New artists (unique share) Main Analysis -0.096*** -0.046*** -0.047*** Control for Unobservables -0.11*** -0.048* Different Long-Term Effect (36 weeks) -0.097*** -0.046*** -0.053*** Before Removal of Taylor Swift -0.11*** -0.045* Before Improving Recommendations -0.11*** -0.049*** -0.067*** Without Launch Countries -0.077*** -0.031* -0.045** Only With Launch Countries -0.11*** -0.091*** -0.042 Consumption Not on Spotify -0.058*** -0.038*** -0.039** Fractional Response Model -0.40*** -0.19*** -0.20*** With One Treatment Dummy -0.053*** New songs (unique share) Main Analysis -0.035*** -0.018** -0.015 Control for Unobservables -0.049*** -0.034** Different Long-Term Effect (36 weeks) -0.036*** -0.018** -0.018+ Before Removal of Taylor Swift -0.047*** -0.032** Before Improving Recommendations -0.036*** -0.018* -0.021* Without Launch Countries -0.024* -0.015+ -0.019 Only With Launch Countries -0.063*** -0.024 0.014 Consumption Not on Spotify -0.020* -0.0097 -0.015 Fractional Response Model -0.22*** -0.12*** -0.093+ With One Treatment Dummy -0.020** New genres (unique share) Main Analysis -0.089*** -0.051** -0.056* Control for Unobservables -0.11** -0.050 Different Long-Term Effect (36 weeks) -0.090*** -0.052** -0.065* Before Removal of Taylor Swift -0.11** -0.057+ Before Improving Recommendations -0.095*** -0.037+ -0.058* Without Launch Countries -0.091** -0.059** -0.062* Only With Launch Countries -0.10+ -0.076+ -0.073 Consumption Not on Spotify -0.021 -0.013 -0.0083 Fractional Response Model -0.38*** -0.22*** -0.25*** With One Treatment Dummy -0.058*** Notes: Regressions with robust standard errors. Estimates are calculated for the main analysis, and several robustness checks (details are given in Table D1). User- and week-specific fixed effects are used and the unit of analysis is the user-week. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 64 Table D8. Robustness Checks for Discovery of New Content (3): Value of Top Discoveries Short-term Medium-term Long-term b b b Top 1 new artist (share) Main Analysis 0.088*** 0.026* 0.0097 Control for Unobservables 0.040 0.00035 Different Long-Term Effect (36 weeks) 0.089*** 0.025* -0.0056 Before Removal of Taylor Swift 0.036 -0.0061 Before Improving Recommendations 0.091*** 0.031* 0.019 Without Launch Countries 0.10*** 0.033* 0.023 Only With Launch Countries 0.090** 0.012 -0.030 Consumption Not on Spotify -0.013 -0.026* -0.042* Fractional Response Model 0.49*** 0.12* 0.000038 With One Treatment Dummy 0.039*** Top 5 new artists (share) Main Analysis 0.066*** 0.013+ 0.0015 Control for Unobservables 0.031+ -0.0062 Different Long-Term Effect (36 weeks) 0.067*** 0.013 -0.0029 Before Removal of Taylor Swift 0.030+ -0.0079 Before Improving Recommendations 0.068*** 0.018* 0.011 Without Launch Countries 0.069*** 0.016 0.0072 Only With Launch Countries 0.081*** 0.015 -0.0067 Consumption Not on Spotify -0.0093 -0.017* -0.032** Fractional Response Model 0.52*** 0.076 -0.043 With One Treatment Dummy 0.024** Top 1 new song (share) Main Analysis 0.13*** 0.049** 0.041+ Control for Unobservables 0.14*** 0.061+ Different Long-Term Effect (36 weeks) 0.14*** 0.049** 0.042 Before Removal of Taylor Swift 0.13*** 0.054 Before Improving Recommendations 0.14*** 0.059*** 0.062* Without Launch Countries 0.14*** 0.047* 0.014 Only With Launch Countries 0.13** 0.099* 0.12* Consumption Not on Spotify -0.027 -0.037* -0.059** Fractional Response Model 0.57*** 0.20*** 0.17* With One Treatment Dummy 0.065***

Marketing Science Institute Working Paper Series 65 Top 5 new songs (share) Main Analysis 0.13*** 0.039** 0.027 Control for Unobservables 0.15*** 0.068* Different Long-Term Effect (36 weeks) 0.13*** 0.040** 0.040+ Before Removal of Taylor Swift 0.15*** 0.071* Before Improving Recommendations 0.14*** 0.049** 0.040+ Without Launch Countries 0.13*** 0.038* 0.014 Only With Launch Countries 0.14** 0.076* 0.074 Consumption Not on Spotify -0.035+ -0.042** -0.056** Fractional Response Model 0.53*** 0.15*** 0.097 With One Treatment Dummy 0.057*** Top 1 new genre (share) Main Analysis 0.017*** 0.0061* 0.0050 Control for Unobservables 0.010 -0.0014 Different Long-Term Effect (36 weeks) 0.017*** 0.0061* 0.0048 Before Removal of Taylor Swift 0.0064 -0.0037 Before Improving Recommendations 0.018*** 0.0063* 0.0042 Without Launch Countries 0.016** 0.0095** 0.0094* Only With Launch Countries 0.024 -0.0047 -0.011 Consumption Not on Spotify 0.0045 0.0022 0.00089 Fractional Response Model 0.79*** 0.17 -0.13 With One Treatment Dummy 0.0082*** Top 5 new genres (share) Main Analysis 0.0089*** 0.0025+ 0.0020 Control for Unobservables 0.0045 -0.0019 Different Long-Term Effect (36 weeks) 0.0090*** 0.0025+ 0.0027 Before Removal of Taylor Swift 0.0023 -0.0025 Before Improving Recommendations 0.0096*** 0.0026 0.0016 Without Launch Countries 0.0091** 0.0041* 0.0040+ Only With Launch Countries 0.011 -0.0016 -0.0036 Consumption Not on Spotify 0.0039 0.0010 0.00045 Fractional Response Model 0.79*** 0.074 -0.24 With One Treatment Dummy 0.0037** Notes: Regressions with robust standard errors. Estimates are calculated for the main analysis, and several robustness checks (details are given in Table D1). User- and week-specific fixed effects are used and the unit of analysis is the user-week. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 66 Online Appendix E. Heterogenous Treatment Effects

Table E1. Consumption Growth and Displacement across Platforms (1) (2) (3) (4)

Log Log playcounts Log playcounts on Log playcounts on playcounts on on iTunes Winamp, other platforms all platforms Windows Media Player and Foobar2000 Adoption 0.48*** -0.20 -0.35* 0.0076 (0.14) (0.16) (0.18) (0.18) x Top 100 artists 0.034 -0.51*** 0.15 -0.0092 (play share) (0.13) (0.15) (0.16) (0.16) x Age 0.11 0.019 -0.070 -0.25 (0.13) (0.15) (0.16) (0.16) x Free (vs -0.13 0.25+ -0.18 0.066 premium) (0.12) (0.14) (0.15) (0.14) R-squared 0.50 0.75 0.71 0.63 F 77.7 6.94 13.5 12.6 p-value 0.000 0.000 0.000 0.000 users 894 894 894 894 observations 52228 52228 52228 52228 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week. The dependent variable is the log number of songs heard by a panelist on a week (playcount). The independent variables are indicators for a user's adoption of Spotify, and interaction effects with pre-sample measures to capture heterogeneous treatment effects. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 67 Table E2. Breadth of variety: Unique Number of Consumed Artists, Songs, and Genres (1) (2) (3)

Log number of Log number of Log number of unique artists unique songs unique genres Adoption 0.29*** 0.29*** 0.20** (0.082) (0.083) (0.062) x Top 100 artists 0.11 0.027 0.11* (play share) (0.074) (0.076) (0.055) x Age 0.036 0.087 0.026 (0.074) (0.076) (0.055) x Free (vs -0.12 -0.14+ -0.079 premium) (0.071) (0.073) (0.052) R-squared 0.56 0.54 0.56 F 52.0 68.3 44.5 p-value 0.000 0.000 0.000 users 894 894 894 observations 46585 46585 46585 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the log number of distinct artists, songs, and genres heard by a panelist on a week. The independent variables are indicators for a user's adoption of Spotify, and interaction effects with pre-sample measures to capture heterogeneous treatment effects. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 68 Table E3. Concentration of Variety (1): Share of Superstar Consumption (1) (2) (3) (4) (5) (6)

Top 20 artists Top 100 artists Top 500 artists Top 20 artists Top 100 artists Top 500 artists (share of unique (share of unique (share of unique (share of all (share of all (share of all artists) artists) artists) plays) plays) plays) Adoption -0.00064 0.012 0.023* 0.0034 0.026* 0.029* (0.0038) (0.0085) (0.011) (0.0055) (0.011) (0.013) x Top 100 artists -0.020*** -0.061*** -0.088*** -0.030*** -0.078*** -0.091*** (play share) (0.0039) (0.0088) (0.011) (0.0060) (0.012) (0.013) x Age 0.0073+ 0.0072 0.0064 0.010+ 0.0019 0.012 (0.0042) (0.0090) (0.011) (0.0061) (0.012) (0.013) x Free (vs 0.0036 -0.0024 -0.0090 0.0051 -0.0033 -0.013 premium) (0.0040) (0.0082) (0.0100) (0.0059) (0.011) (0.012) R-squared 0.42 0.54 0.61 0.42 0.52 0.55 F 3.53 7.35 16.9 3.56 7.17 14.4 p-value 0.000 0.000 0.000 0.000 0.000 0.000 users 894 894 894 894 894 894 observations 46585 46585 46585 46585 46585 46585 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the number of unique / amount of plays to popular artists (measured in an initialization period) in a user's geographic region, divided by the number of unique artists / total number of plays. The independent variables are indicators for a user's adoption of Spotify, and interaction effects with pre- sample measures to capture heterogeneous treatment effects. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 69 Table E4. Concentration of Variety (2): Share of Top Two (C2) and Top Ten (C10) (1) (2) (3) (4) (5) (6)

Artist Song Genre Artist Song Genre concentration concentration concentration concentration concentration concentration (C2) (C2) (C2) (C10) (C10) (C10) Adoption -0.039* -0.026** -0.026* -0.048** -0.044** -0.014* (0.015) (0.0078) (0.012) (0.015) (0.016) (0.0061) x Top 100 artists -0.026+ -0.0071 -0.018+ -0.016 -0.012 -0.0037 (play share) (0.014) (0.0073) (0.010) (0.014) (0.014) (0.0052) x Age 0.0016 -0.0044 0.0052 0.0063 -0.018 0.0025 (0.014) (0.0075) (0.010) (0.014) (0.015) (0.0050) x Free (vs 0.023+ 0.014* 0.0049 0.026* 0.026+ 0.0047 premium) (0.014) (0.0073) (0.0097) (0.013) (0.014) (0.0044) R-squared 0.45 0.36 0.53 0.58 0.45 0.60 F 20.8 8.13 13.3 17.9 20.6 7.72 p-value 0.000 0.000 0.000 0.000 0.000 0.000 users 894 894 894 894 894 894 observations 46585 46585 46585 46585 46585 46585 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the share of playcounts to a user's weekly top two and top ten artists, songs, and genres (C2 and C10). The independent variables are indicators for a user's adoption of Spotify, and interaction effects with pre-sample measures to capture heterogeneous treatment effects. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 70 Table E5. Discovery of New Content (2): Repeat Consumption (Known versus New) (1) (2) (3) (4) (5) (6)

Known artists Known songs Known genres New artists New songs New genres played more played more played more played more played more played more than once than once than once than once than once than once (share of unique (share of unique (share of unique (share of unique (share of unique (share of unique known artists) known songs) known genres) new artists) new songs) new genres) Adoption 0.0051 0.0042 -0.0025 -0.071*** -0.017 -0.065* (0.013) (0.0096) (0.0092) (0.020) (0.013) (0.031) x Top 100 artists -0.031** 0.0078 -0.016+ -0.0084 0.0030 0.022 (play share) (0.012) (0.0096) (0.0089) (0.020) (0.011) (0.032) x Age -0.0047 0.011 0.0000052 0.036+ 0.0040 -0.0074 (0.012) (0.0095) (0.0088) (0.019) (0.011) (0.031) x Free (vs -0.00055 -0.0023 0.00026 0.0046 -0.015 0.0012 premium) (0.012) (0.0096) (0.0085) (0.018) (0.011) (0.029) R-squared 0.45 0.45 0.32 0.33 0.33 0.27 F 1.79 4.81 1.75 2.14 3.91 1.47 p-value 0.000 0.000 0.000 0.000 0.000 0.011 users 894 894 894 894 894 887 observations 46114 45483 46527 35439 43720 15732 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week when there is at least one song played. The dependent variable is the number of unique known (versus new) artists, songs, and genres played more than once, divided by the total number of unique known (versus new) artists, songs, and genres listened to. New artists, songs, and genres are defined by a user's first week of consumption on the service up to 6 January 2013. The independent variables are indicators for a user's adoption of Spotify, and interaction effects with pre-sample measures to capture heterogeneous treatment effects. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 71 Table E6. Discovery of New Content (3): Value of Top Discoveries (1) (2) (3) (4) (5) (6)

Share of top 1 Share of top 5 Share of top 1 Share of top 5 Share of top 1 Share of top 5 new artist to new artist to new song to new song to new genre to new genre to overall top 1 overall top 5 overall top 1 overall top 5 overall top 1 overall top 5 artist artist song song genre genre Adoption 0.028 0.016 0.052+ 0.043 0.0011 -0.00076 (0.022) (0.015) (0.030) (0.027) (0.0039) (0.0025) x Top 100 artists 0.052** 0.031* 0.024 0.024 0.0069+ 0.0042* (play share) (0.019) (0.013) (0.028) (0.025) (0.0036) (0.0021) x Age 0.015 0.012 0.034 0.036 0.0034 0.0020 (0.019) (0.013) (0.027) (0.024) (0.0035) (0.0020) x Free (vs -0.055** -0.032* -0.038 -0.040 0.0036 0.0026 premium) (0.020) (0.013) (0.027) (0.025) (0.0037) (0.0021) R-squared 0.22 0.25 0.22 0.26 0.076 0.078 F 15.4 16.8 9.20 10.9 1.34 1.27 p-value 0.000 0.000 0.000 0.000 0.051 0.096 users 883 883 882 882 883 883 observations 32448 32448 31702 31702 32448 32448 Notes: Regression with robust standard errors in parentheses. Estimates are calculated on a matched sample of 447 adopters and 447 non-adopters observed over 62 weeks starting May 29, 2014; user- and week-specific fixed effects are used and the unit of analysis is the user-week with at least one new song played. The dependent variable is the amount of plays to the top 1 and 5 new artists, songs, and genres in the eight week subsequent to discovery (t+1, ..., t+8) ranked in order of plays, divided by the amount of plays to the overall (not necessarily new) top 1 and 5 artists, songs, and genres over the same time period. Observations are excluded when the rolling 8-week window includes both pre-adoption and post-adoption periods, and when there are fewer than 8 weeks remaining at the end of each user's observation period. New artists, songs, and genres are defined by a user's first week of consumption on the service up to 6 January 2013. The independent variables are indicators for a user's adoption of Spotify, and interaction effects with pre-sample measures to capture heterogeneous treatment effects. + p < 0.10, * p < 0.05, ** p < .01, *** p < .001.

Marketing Science Institute Working Paper Series 72