Three Essays on Network Development, Social Influence, And

Home , Anime Network

THREE ESSAYS ON NETWORK DEVELOPMENT,

SOCIAL INFLUENCE, AND USER ENGAGEMENT IN

ONLINE COMMUNITIES

Mina Ameri

APPROVED BY SUPERVISORY COMMITTEE:

Ying Xie, Chair

Elisabeth Honka, Co-Chair

Dmitri Kuksov

Xiaolin Li

Brian T. Ratchford Copyright c 2018 Mina Ameri

SOCIAL INFLUENCE, AND USER ENGAGEMENT IN

ONLINE COMMUNITIES

MINA AMERI, BS, MS

DISSERTATION

Presented to the Faculty of

The University of Texas at Dallas

in Partial Fulﬁllment

of the Requirements

for the Degree of

DOCTOR OF PHILOSOPHY IN

MANAGEMENT SCIENCE

THE UNIVERSITY OF TEXAS AT DALLAS

May 2018 ACKNOWLEDGMENTS

Writing this dissertation was only possible because of the help and support of my amazing advisors, Drs. Ying Xie and Elisabeth Honka. I would like to also thank my other committee members, Drs. Dmitri Kuksov, Xiaolin Li, and Brian T. Ratchford for their consistent encouragement and support.

March 2018

v THREE ESSAYS ON NETWORK DEVELOPMENT,

SOCIAL INFLUENCE, AND USER ENGAGEMENT IN

ONLINE COMMUNITIES

Mina Ameri, PhD The University of Texas at Dallas, 2018

Supervising Professors: Ying Xie, Chair Elisabeth Honka, Co-Chair

I study inﬂuential factors on users’ activities in online communities in three chapters. Using unique data coming from an online anime (Japanese cartoon) platform containing data on users’ friendship networks, anime adoptions, and generated content over time, I look into consumers’ decisions in the context of new media such as online streaming.

In the first chapter of this dissertation, I study the e↵ects of observational learning and word- of-mouth (volume and valence) on consumers’ product adoptions. Understanding whether these two social learning devices provide di↵erent and unique information or whether one is redundant in the presence of the other is crucial for companies’ information provision strategies. I di↵erentiate between the e↵ects of word-of-mouth and observational learning from friends (“personal network”) and the e↵ects of word-of-mouth and observational learning from the whole community (“community network”). The relative importance of word-of- mouth and observational learning at each network level provides guidance for companies’ platform design. My results reveal that both word-of-mouth and observational learning from both the community and personal networks have significant and positive e↵ects on individual users adoption decisions. I find that word-of-mouth valence from the community

vi network is the largest adoption driver among the social learning forces I study. Lastly, I test for asymmetric observational learning from positive and negative actions and observational learning creating product awareness versus transferring unobserved quality information.

In the second chapter, I quantify the e↵ects of binge-watching on consumers’ engagement with media franchises in two areas: interactive and personal engagement. The former involves auser’scontentgenerationrelatedtoafocalmediaproductandthelatterconcernsthe adoption of franchise extensions (sequels and other extensions). I ﬁnd that binge-watching has a negative e↵ect on the production of user-generated content. The e↵ect of binge- watching on the adoption of franchise extensions critically depends on both the availability of franchise extensions at the time of watching the focal anime and the extension type: if it is available, binge-watching increases (decreases) the probability that a user watches the sequel (other franchise extensions).

In the third chapter, I develop a structural model for the co-evolution of individuals’ friendship tie formations and their concurrent online activities (product adoptions and production of user-generated content) within a social network. Explicitly modeling the endogenous formation of the network and accounting for the interdependence between decisions in these two areas (friendship formations and online activities) provides a clean identiﬁcation of peer e↵ects and of important drivers of individuals’ friendship decisions. My results reveal that, compared to a potential friend’s product adoptions and content generation activities, total number of friends and number of common friends this potential friend has with the focal individual are the most important drivers of friendship formation. Further, while having more friends does not make a person more active, having more active friends increases a user’s activity levels in terms of both product adoptions and content generation through peer e↵ects. Via counterfactuals I assess the e↵ectiveness of various seeding strategies in increasing website trac while taking the endogenous network formation into account.

vii TABLE OF CONTENTS

ACKNOWLEDGMENTS ...... v ABSTRACT ...... vi LISTOFFIGURES ...... xi LISTOFTABLES...... xii CHAPTER 1 WORD-OF-MOUTH, OBSERVATIONAL LEARNING, AND PROD- UCT ADOPTION: EVIDENCE FROM AN ANIME PLATFORM ...... 1 1.1 Introduction ...... 1 1.2 Relevant Literature ...... 6 1.3 Data ...... 9 1.3.1 Data Collection, Cleaning, and (Re-)Construction ...... 13 1.3.2 Data Description ...... 17 1.4 Model and Estimation ...... 21 1.4.1 Challenges ...... 21 1.4.2 Model Description ...... 23 1.5 ResultsandDiscussion ...... 24 1.5.1 E↵ects of Word-Of-Mouth and Observational Learning ...... 24 1.5.2 PositiveandNegativeObservationalLearning ...... 28 1.5.3 Awareness versus Learning about Unobserved Quality ...... 29 1.6 Limitations and Future Research ...... 29 1.7 Conclusion ...... 30 CHAPTER 2 THE EFFECTS OF BINGE-WATCHING ON MEDIA FRANCHISE ENGAGEMENT ...... 33 2.1 Introduction ...... 33 2.2 Relevant Literature ...... 37 2.2.1 UserEngagement...... 37 2.2.2 Binge-Watching...... 39 2.2.3 Online Movie Streaming ...... 41 2.3 Data ...... 42

viii 2.3.1 Data Cleaning ...... 44 2.3.2 Engagement ...... 45 2.3.3 Binge-Watching...... 47 2.3.4 Data Description ...... 48 2.4 Model and Estimation ...... 54 2.5 Results ...... 57 2.6 Robustness Checks ...... 61 2.7 Limitations and Future Research ...... 63 2.8 Conclusion ...... 63 CHAPTER 3 A MODEL OF NETWORK DYNAMICS: TIE FORMATION, PROD- UCT ADOPTION, AND CONTENT GENERATION ...... 67 3.1 Introduction ...... 67 3.2 Relevant Literature ...... 73 3.2.1 NetworkFormation...... 74 3.2.2 Peer E↵ects ...... 76 3.2.3 Seeding ...... 77 3.3 Data ...... 79 3.3.1 Estimation Sample ...... 80 3.3.2 Data Description ...... 84 3.4 Model ...... 89 3.4.1 Tie Formation ...... 90 3.4.2 ProductAdoptionandContentGeneration...... 92 3.4.3 IntegratingAllActions...... 92 3.4.4 UtilitySpeciﬁcations ...... 93 3.5 Estimation ...... 96 3.6 Identiﬁcation ...... 100 3.7 Results ...... 101 3.8 Counterfactual ...... 107 3.8.1 How to Increase In-Site Activities through Platform-Wide Stimulation System? ...... 109

ix 3.8.2 How to Increase In-Site Activities through Seeding? ...... 109 3.9 Limitations and Future Research ...... 113 3.10 Conclusion ...... 114 APPENDIX A ORIGINAL AND FINAL DATA ...... 116 APPENDIX B VARIABLE (RE-)CONSTRUCTION ...... 119 APPENDIX C ROBUSTNESS CHECKS ...... 122 APPENDIXD PROBITRESULTS ...... 123 APPENDIX E COMPLETE RESULTS FOR OTHER CONTINENTS ...... 125 APPENDIX F ALTERNATIVE CLASSIFICATIONS OF BINGE-WATCHING . . . 129 APPENDIXG LIKELIHOODDERIVATION ...... 131 REFERENCES...... 133 BIOGRAPHICALSKETCH...... 141 CURRICULUM VITAE

x LIST OF FIGURES

1.1 ExampleofanAnimePage ...... 10 1.2 ExampleofaUserPage ...... 10 1.3 ExampleofaUserWatchList...... 12 1.4 DatesUsersJoinedMyAnimeList.Net...... 14 1.5 Percentage of Friends Added During First 2 Years After Joining MyAnimeList.net (grouped by length of membership) ...... 15 1.6 Average Number of News Articles (shaded area denotes 5th and 95th percentiles) 16 1.7 Histograms of the Number of Friends and of Descriptives Related to Adoption . 18 1.8 WOMandOLfromPersonalandCommunityNetworks ...... 19 2.1 DatesUsersJoinedMyAnimeList.Net...... 43 2.2 Number of Days After Release of First or Final Episode in a Season That Animes WereWatched...... 46 2.3 Watch Period Distribution (truncated at 200 days) ...... 48 2.4 NumberofHoursWatchedPerDay...... 49 2.5 PercentageofAUser’sWatchListThatIsBinge-Watched ...... 50 2.6 Binge-WatchingvsNon-Binge-WatchingAcrossTime ...... 51 2.7 DistributionofRatingsforBingedvs. Non-BingedCases ...... 54 2.8 ExamplesofReleaseDatesonNetﬂix...... 65 3.1 DatesUsersJoinedMyAnimeList.Net...... 80 3.2 NumberofDaysBetweenActivities...... 82 3.3 UserSamplingStrategy ...... 83 3.4 Network Co-Evolution Over Time Lines Between Nodes Indicate Friendship Ties. Node Size Increases with More Animes Watched. Node Color Darkens with More PostsWritten...... 85 3.5 Average Activity Levels Over Time Since Joining (New Users) ...... 87 3.6 Percentage of Observations with Certain Activities Conditional on Performing at leastoneActivity...... 88 3.7 NumberofActivitiesinEachAreaPerDay ...... 89 3.8 Number of Activities and Active Users Per Day Under Di↵erent Recommendation Strategies (Colors Represent Seeding in Indicated Areas of Activities.) . . . . . 110

xi LIST OF TABLES

1.1 Descriptive Statistics ...... 20 1.2 Results ...... 25 2.1 NumberofEpisodesinandDurationofaSeason ...... 45 2.2 ProbabilityofEngagementAction...... 51 2.3 TimingofEngagementActions ...... 53 2.4 Weekday Frequencies of Start Dates ...... 57 2.5 Results-NorthAmerica ...... 58 2.6 E↵ects of Binge-Watching Across Di↵erentRegions ...... 62 3.1 Descriptive Statistics ...... 86 3.2 Results ...... 103 3.3 Counterfactual Results ...... 111 A.1 DescriptiveStatistics ...... 118 C.1 Robustness Checks ...... 122 D.1 ProbitResults-NorthAmerica ...... 123 D.2 E↵ects of Binge-Watching Across Di↵erentRegions ...... 124 E.1 Results-SouthAmerica ...... 125 E.2 Results - Europe ...... 126 E.3 Results-Asia...... 127 E.4 Results - Oceania ...... 128 F.1 Probability of Engagement Action – Binge-Watching Deﬁnition: More than 2 Hours129 F.2 Probability of Engagement Action – Binge-Watching Deﬁnition: More than 4 Hours130

xii CHAPTER 1

WORD-OF-MOUTH, OBSERVATIONAL LEARNING, AND PRODUCT

ADOPTION: EVIDENCE FROM AN ANIME PLATFORM

1.1 Introduction

Social learning has been shown to play an important role in consumers’ product adoptions

(e.g., Aral and Walker 2011; Chen et al. 2011). Consumers can learn from and be inﬂuenced by their social interactions with others through two di↵erent mechanisms, namely, through word-of-mouth (WOM hereafter) and through observational learning (OL hereafter). In

WOM, consumers extract product information directly from others’ opinions, while in OL, consumers infer information about products from others’ previous actions indirectly. Nu- merous studies have shown that volume and valence of WOM can have a signiﬁcant impact on consumers’ purchase and adoption behaviors (to name a few, Godes and Mayzlin 2004;

Chevalier and Mayzlin 2006; Liu 2006; Moe and Trusov 2011; Lovett and Staelin 2016).

Although OL has not been studied to that extent in the marketing literature, a few recent empirical papers have shown that OL can a↵ect consumers’ decisions leading to information cascades and herding behavior (e.g., Cai et al. 2009; Zhang 2010; Herzenstein et al. 2011;

Zhang and Liu 2012).

Although both WOM and OL have been separately studied as elements of social learning, there remain important questions unanswered. First, almost all extant literature has studied either WOM or OL as the single social learning device that inﬂuences consumers’ product adoptions (e.g., Godes and Mayzlin 2004; Zhang 2010). Although information about product quality can be extracted or inferred from both mechanisms, consumers may still interpret

WOM and OL information di↵erently and therefore be inﬂuenced by these two forces to varying degrees. On the one hand, one can argue that, compared with OL, WOM conveys more diagnostic information about product quality; therefore it should play a more prominent

1 role. On the other hand, actions speak louder than words. In the presence of OL, the product information a consumer can obtain from consumer reviews may seem unreliable or redundant and therefore the role of WOM may be diminished. To the best of my knowledge, (Chen et al., 2011) is the only paper that studies the e↵ects of both WOM and OL at the aggregate product sales level. No study that I are aware of has simultaneously examined the di↵erential e↵ects of information from WOM versus OL on individual consumers’ adoption decisions.1 Second and more importantly, both WOM and OL can operate at di↵erent levels of a network. Many online platforms provide various tools and functions to facilitate socialization among their users. Users can become friends with each other and form their own personal social networks within the larger community. In this context, a user can be influenced by his friends’ actions and/or opinions, while, at the same time, he can also observe product adoptions, online reviews, and ratings by users beyond his personal network. Throughout this paper I refer to a user’s network of friends as the “personal” network and to the network as a whole (which includes his personal network) as the “community” network. Although extant empirical studies have led support for the significant e↵ects of WOM or OL from either the community or the personal network (e.g., Godes and Mayzlin 2004; Zhang 2010; Nair et al. 2010; Aral and Walker 2011), it remains an unanswered empirical question whether and to what extent WOM and OL influence product adoptions when both types of information are available from both network levels. The answer to this question will provide useful guidance for companies’ platform design. On the one hand, friends’ actions and opinions may be viewed as more informative and provide more relevant guidance (Zhang et al. 2015). This is because, when users make their product adoption decisions based on both personal preferences and product quality,

1Note that my definition of OL in this paper di↵ers from the classic definition of OL used by previous literature (e.g., Cai et al. 2009; Zhang 2010; Herzenstein et al. 2011; Zhang and Liu 2012): In the classic definition, OL is conceptualized as observations of adoptions without additional information. In my paper, OL is conceptualized as observations of adoptions after controlling for WOM information (see also Chen et al. 2011). I thank an anonymous reviewer for pointing this out.

2 the higher certainty in preferences of the personal network makes the extraction of quality information easier. On the other hand, when community networks are large, they provide more “accurate” information in terms of being less prone to cascades than personal networks (Zhang et al. 2015). The finding whether one network level is dominant or both network levels are equally important will provide useful information to guide companies’ platform design decisions on what socialization tools and functions should be made available to consumers. In this paper, I aim to answer these questions in the empirical context of anime (Japanese cartoon) watching. I choose this market as my empirical context for the following reasons: Movies and shows such as animes are cultural products. With the rapid expansion of online streaming services in recent years,2 consumers face an overwhelmingly large and constantly growing choice set when deciding which specific movies or shows to watch. In this scenario, consumers tend to rely on various informational cues to learn about product availability as well as to lower their ex-ante uncertainty about product utility. Moreover, in contrast to other online markets, the marginal product price is zero in online streaming.3 Therefore product popularity and rating information from social networks are likely to play significant roles in consumers’ product adoptions, making it an ideal context to study social learning. Iobtainmydatafromaspecialinterestonlinecommunitywebsiteforanimescalled MyAnimeList.net. This website provides a gathering place for anime fans to share their enthusiasm and exchange their opinions about animes. Aside from online ratings, forum posts, rankings, and news, the website provides a platform for users to interact with each other and to form friendships. Furthermore, users can not only create their personal watch

2Online streaming of movies and (TV) shows has grown rapidly over the last decade (see McKinsey&Company’s Global Media Report 2015). In 2015, over 40% of U.S. households sub- scribed to at least one video streaming service (http://www.nielsen.com/us/en/insights/reports/2015/the- total-audience-report-q4-2014.html). 70% of North American internet trac in 2015 consisted of streaming video and audio content and Netﬂix alone accounted for 37% of all internet trac in North America (https://www.sandvine.com/pr/2015/12/7/sandvine-over-70-of-north-american-trac-is- now-streaming-video-and-audio.html).

3Through legal channels, there are usually ﬁxed costs of online streaming through subscription fees.

3 lists, i.e., a list of animes that they have watched, and rate the movies on their watch list, but they can also check other users’ watch lists and the ratings these users have submitted. Users receive information about their friends’ anime adoptions and the ratings thereof through three means: through automatic updates about their friends’ recent activities, by looking at friends’ watch lists, and by checking the adopter list for an anime. Users can also check community-wide popularity (based on the number of adoptions) and average rating scores for all animes listed on the platform. This dual nature enables us to tease apart di↵erent sources of information and to study their separate inﬂuence on users’ product adoptions.

One of the major challenges of working with network data is distinguishing between correlation and causation. As (Hartmann et al., 2008) discuss correlation in behavior can be due to three di↵erent reasons: endogenous group formation, correlated unobservables, and simultaneity. I take two steps to solve the challenge of endogenous group formation: first, I only look at users who have been in the network for more than one year before the release of the first anime under study since my data indicate that users mostly form their friendships in the first six months after joining. And second, to address the issue of homophily which arises due to endogenous group formation, I exploit the rich panel structure of my data and include user-anime fixed e↵ects to control for each user’s preference for a specific anime and user-release week fixed e↵ects to control for each user’s propensity to adopt earlier as opposed to later in my model. To account for common shocks that lead to correlated unobservables, I include (calendar) week fixed e↵ects and the number of news pieces collected from MyAnimeList.net and other websites in my model. To address simultaneity, I use lagged versions of variables describing friends’ actions and opinions.

I model users’ adoption decisions for 103 animes using a linear probability model to be able to accommodate the large number of ﬁxed e↵ects in my model speciﬁcation (e.g.,

Bandiera and Rasul 2006; Nair et al. 2010; Bollinger and Gillingham 2012). My results reveal that both WOM and OL have signiﬁcant e↵ects on users’ anime adoptions. At the community

4 network level, while both WOM valence and WOM volume have significant positive e↵ects on users’ adoptions, the e↵ect of WOM valence is larger than the e↵ect of WOM volume. Furthermore, OL also has a significant, albeit smaller than WOM valence, positive e↵ect on product adoptions: as an anime gains more popularity in the community, users become more likely to watch the anime. Similarly, at the personal network level, both WOM (volume and valence) and OL from friends have significant positive e↵ects on users’ adoptions with WOM volume having the largest e↵ect among the three. Comparing the magnitudes WOM and OL e↵ects across both network levels based on my predictive exercise results, I find that WOM valence from the community network is the largest adoption driver related to social learning. Further, I find evidence that users di↵erentiate between positive OL (from their friends’ positive actions) and negative OL (from their friends’ negative actions). And finally, I find OL to both create awareness for an anime and to let users learn about the unobserved quality of an anime. The contribution of this paper is two-fold. First, I contribute to the social learning and product adoption literatures by disentangling the e↵ects of WOM and OL, the two prevalent social learning devices. My findings provide empirical support for the di↵erential and unique e↵ects that product information inferred from WOM versus OL has on consumers’ product adoption decisions. In particular, my result that the e↵ect of community WOM valence overshadows the e↵ect of community OL is consistent with the predominant business practice to display average product ratings. And second, I demonstrate the relative importance of social learning at di↵erent network levels: the community network versus the personal network. My finding that social learning from the community network has a larger impact on consumer product adoption than social learning from the personal network corroborates the theoretical prediction from (Zhang et al., 2015) that community networks provide more accurate information to consumers than personal networks when they are suciently large. The remainder of the paper is organized as follows: In the next section, I discuss the relevant literature. In Sections 3.3 and 3.4, I describe my data, introduce my model and

5 estimation approach. I present and discuss my results in Section 3.7. In the following section, I examine limitations of the current work and opportunities for future research. Finally, I conclude by summarizing my ﬁndings in Section 3.10.

1.2 Relevant Literature

In this section, I review relevant streams of literature on word-of-mouth and observational learning and delineate the positioning of my research vis-a-vis the findings from extant research. WOM has been largely studied in the context of reviews and online opinions. There is strong empirical support for the positive e↵ect of online opinions in di↵erent industries: TV shows (Godes and Mayzlin 2004; Lovett and Staelin 2016), movies (Liu 2006; Dellarocas et al. 2007; Duan et al. 2008; Chintagunta et al. 2010), books (Chevalier and Mayzlin 2006; Li and Hitt 2008), bath and beauty (Moe and Trusov 2011), and video games (Zhu and Zhang 2010). The consensus of these studies is that WOM created by community networks influences consumers’ product adoptions. At the same time, there are few papers that have studied the e↵ects of WOM within personal networks. (Aral and Walker, 2011) study consumers’ app adoptions. They find WOM in the form of active-personalized messaging to be more e↵ective than in the form of passive broadcasting viral messaging in encouraging adoption per message. (Brown and Reingen, 1987) trace referral WOM of music teachers in local neighborhoods and quantify the e↵ects of WOM in weak and strong ties. They find that strong ties are likely to be used as sources for product related information. While these studies show the significant e↵ect of WOM on adoption at both the community and the personal network level, how these two levels of WOM influence an individual’s decision simultaneously is not clear. (Zhang and Godes, 2013) study how an individual’s (purchase) decision quality improves based on information received from strong and weak ties in the network while controlling for WOM valence and variance at the community network

6 level. While (Zhang and Godes, 2013) have data from an online community similar to the one under study in this paper, they do not have information on product adoptions (neither from friends nor the whole community). Therefore (Zhang and Godes, 2013) are not able to study OL and, instead, focus on WOM as the main social learning device. In addition, they also do not observe either the valence or the content of information individuals receive from their personal networks and instead use the number of ties as a proxy for the quantity of information received. In the current study, I study the e↵ect of both WOM and OL on an individual’s product adoption decisions. Furthermore, I treat WOM extracted from the community network and WOM received from one’s own personal network as separate information sources and identify their relative importance in driving individuals’ product adoption behavior.

Next, I discuss the relevant literature on OL. With limited information available, people use others’ observed prior decisions in addition to their private information to shape their beliefs and to make decisions (Banerjee 1992; Bikhchandani et al. 1992). This can lead to information cascades (Bikhchandani et al. 1992) and herding behavior. This e↵ect is stronger when consumers are uncertain about the product, have imperfect information, and infer their own utility from observing others’ prior decisions (Cai et al. 2009; Duan et al.

2009). (Zhang, 2010) uses data from the kidney market to show that patients draw negative quality inferences from earlier refusals by unknown people in the queue even though they themselves do not have information about the quality of the kidney. (Cai et al., 2009) show that displaying popularity of dishes in a restaurant increases orders of those dishes. (Zhang and Liu, 2012) study lenders’ funding decisions using data from an online microloan platform and ﬁnd evidence for rational herding among lenders. Studying individual choices under the inﬂuence of personal networks, (Nair et al., 2010) and (Wang et al., 2013) show that the volume of usage, expertise, or popularity of friends are key factors that a↵ect adoptions in medicine, technology, and fashion goods, respectively.

7 However, the influences of the community and personal networks have not been recognized as two di↵erent sources of OL until recently. (Zhang et al., 2015) employ a game-theoretical approach to study OL in networks of friends vs. strangers. They define friends as groups of users with homogenous preferences and strangers as groups of users with heterogeneous preferences. They show that, when the network is small, friends’ actions provide more information, while the network of strangers becomes more e↵ective as it grows in size. (Sun et al., 2012) study herding behavior of consumers under the influence of friends’ and the community’s choices. In their specific context, users do not infer quality information about achoice,justthepopularityofachoice.Theyshowthatpeoplearemorelikelytodiverge from the popular choice among their friends as the adoption rate of a choice increases, but do not respond to the popular choice in the community. This is because the community does not form an opinion about the person whereas friends do. These two studies suggest that OL can happen at both the personal and the community network level. In this paper, I observe choices of individuals when they receive product popularity information from both their personal and the community network and study how each of these two sources influences consumers’ product adoptions simultaneously.

To the best of my knowledge, almost all extant marketing literature has either studied

WOM or OL as the single mechanism through which consumers extract product information to facilitate their adoption decisions. The only exception is (Chen et al., 2011) in which the authors examine the role of both WOM and OL on aggregate online product sales at Amazon.com. They ﬁnd that, while negative WOM is more inﬂuential than positive

WOM, positive OL information signiﬁcantly increases sales but negative OL information has no e↵ect. No study that I are aware of has investigated the e↵ects of WOM and OL simultaneously on individual consumers’ product adoptions. Although information about product quality can be extracted or inferred from both mechanisms, consumers may still interpret WOM and OL information di↵erently. In this study, I aim to ﬁll in the gap

8 by examining the di↵erential e↵ects of information from WOM versus OL on individual consumers’ adoption decisions.

1.3 Data

My data come from MyAnimeList.net. This website is a consumption-related online community (Kozinets 1999) where online interactions are based upon shared enthusiasm for a speciﬁc consumption activity. MyAnimeList.net was created to allow anime fans to gather and share their excitement and opinions about animes. In addition, the website has developed into one of the most comprehensive online sources of information about animes (Japanese cartoons) and mangas (Japanese comics). In this paper, I focus on animes. On MyAnimeList.net, both animes and users have their own pages. Figure 1.1 shows an example of an anime page. Each anime page contains detailed information about the anime including acontentsummary,anepisodeguide,productiondetails,userratings,andrankings. Figure 1.2 shows an example of a user page. Note that all information contained in a user’s page is available to the public.4 On a user page, one can see information about the animes and mangas the user has adopted and the adoption dates, his opinion about adopted animes and mangas, his website join date, his in-site activities, the identities of his friends and other information. Users can become friends with other users upon mutual acceptance of a friendship request. After becoming friends, users can see automatic updates about friends’ recent in-site activities on their own pages. Moreover, instant-messaging and communication tools are provided to enable in-site communication between two friends. Users can create a list of animes that they plan to watch or have watched (I refer to this list as “watch list” throughout this paper).5 Figure 1.3 shows an example of a user’s watch

4Users have the option to hide their proﬁle page from the public, but less then 5% of users use this option.

5I do not account for platform choice in this paper because, in general, users can watch animes either legally or illegally through a number of di↵erent channels such as netﬂix.com, hulu.com, funimation.com, crunchyroll.com, aniplexusa.com and others.

9 Figure 1.1: Example of an Anime Page

Figure 1.2: Example of a User Page

10 list. Note that that all animes on the watch list are correctly and uniquely identiﬁed because users are required to use a search function to add animes to the list. Users can assign di↵erent stati to the animes on their watch list: “watched,” “watching,” “on hold,” “dropped,” or

“plan to watch.” I define a user as having adopted an anime if the anime is assigned to any of the first four stati on his watchlist.6 While this definition of adoption might seem very broad, note that the stati “watched,” “watching,” “on hold,” and “dropped” all imply that the user has at least started to watch, i.e., adopted, the anime.7 Further, users can indicate their opinion about the animes on their watch list by rating them on a scale ranging from 1 to 10 (10 being the highest rating). Throughout this paper, I refer to ratings given to animes on watch lists as “user ratings.” Users can also discuss the animes they have watched in the forum section of the website. Lastly, users can indicate the date they started watching an anime and the website also automatically registers the date users last updated the entry for an anime. I use these two dates to infer the time of adoption.8

Iaimatquantifyingthee↵ectsofWOMandOLfromboththepersonalandthecom- munity network on product adoption. I use the number of friends who adopted the anime

6My adoption data are self-reported. Thus accuracy in the reporting of adoptions is a potential concern. Note that in contrast to incentivized surveys, there are no incentives for users on MyAnimeList.net to falsely report their true anime watching behavior. Furthermore, in the similar setting of TV shows, (Lovett and Staelin, 2016) compare survey panelists’ self-reported viewing data and the actual streaming data and ﬁnd that people tend to correctly report their actual watching behavior. Thus I are conﬁdent that the self-reported adoption data are reliable in my context.

7In an additional model, I di↵erentiate between OL coming from positive and negative product adoption experiences. To do so, I deﬁne positive OL as product adoptions under the stati “watched,” “watching,” and “on hold” and negative OL as product adoptions under the status “dropped.” I discuss the results from this additional model in Section 5.2.

8My data contain the start dates and the dates of the last updates for 95% and 5% of observations, respectively. As a robustness check, I move the adoption dates of the 5% of observations for whom I only have the dates of the last update one week, i.e., mark the adoption time as one week prior to the date of the last update. I then re-estimate my model using these modiﬁed adoption times for observations with last updates and ﬁnd my results to be robust (see column (iv) in Table C.1 in Appendix C).

11 Figure 1.3: Example of a User Watch List to measure OL from the personal network.9 Further, I use the average rating of an anime given by the user’s friends to measure WOM valence from the personal network. Following previous literature (e.g., Godes and Mayzlin 2004; Chintagunta et al. 2010), I measure WOM volume from the personal network using the number of ratings and forum posts submitted by the user’s friends. With regard to the community network, users see the community-wide total number of adoptions for an anime on the anime page (see “Members” in the bottom left corner in Figure 1.1) - this is my measure of community OL for the anime. Similarly, users also see the average rating for an anime based on ratings submitted by all users on the anime page (see “Score” in the bottom left corner in Figure 1.1) - this is my measure

9I test the robustness of my results by using the percentage of friends (instead of the number of friends) who adopted the anime as my measure of OL from the personal network. I ﬁnd my results to be qualitatively similar (see model (i) in Table C.1 in Appendix C).

12 of community WOM valence.10 And lastly, users can see the number of ratings from the community network and the number of forum posts on another tab of the anime page. I use the total number of ratings and forum posts to measure community WOM volume.11

1.3.1 Data Collection, Cleaning, and (Re-)Construction

MyAnimeList.net was established in November 2004, but its main activities did not begin until 2007 when the website moved to a public domain and its user base started to grow rapidly (see Figure 2.1). At the point in time when I started the data collection (March

2015), there were more than 2.6 million users on the website among which were about 2.2 million stand-alone users with no friends and little to no activity.12 Since I are interested in the e↵ects of social learning on product adoption, I collected data on a network of nearly

380,000 users.13

There are over 10,000 animes listed on the website. These animes range from short 20- minutes single-episode animes to anime series with more than 50 episodes. I use data on

103 anime series in my analysis. These animes were selected based on release dates, being the first season of an anime (if multiple seasons exist), and viewership. More specifically, I chose animes that were released between July 2012 and January 2014 and I focus on the first

10Note that there is also an alternative measure of OL from community network: popularity rank (based on the number of adoptions) on the anime page (see “Popularity” in the bottom left corner in Figure 1.1). Similarly, there is an alternative measure of WOM valence from the community network: users can see the rank of an anime based on its average rating from all users (see “Ranked” in the bottom left corner of Figure 1.1). I estimated my model using these two alternative measures of OL and WOM valence from the community network and my results are robust (see models (ii) and (iii) in Table C.1 in Appendix C).

11Note that all my WOM variables from both the personal and the community network are conditional on friends’ and all users’ adoptions, respectively.

12The 2.2 million inactive stand-alone users represent a characteristic of this social media platform that is consistent with the well-known 90-9-1 rule in social media (see e.g., https:www.nngroup.comarticlesparticipation-inequality).

13This is the largest and oldest network on MyAnimeList.net. It includes the website owner and users who were members of the website before 2007.

13 ie ogrta h td eid(ihamnmmo . n aiu f74tmslne) hsI Thus total longer). the times to 7.4 compared of when maximum period a study 2.7 the and average during 1.4 on higher of is, significantly minimum period is observation a period. rate the observation (with adoption anime that period the first Note study that the 2015. conclude the March of than until longer date release times release from i.e., the period, to observation prior users year all one dropped than I less network adoption, the product joined and had formation tie who of simultaneity the avoid to First, the across weeks 52 to 19 the from after varies year) period (one animes. study weeks 103 the 52 Thus or exist) shorter. seasons is multiple whichever (if release, season second the of sample. release account final the together my and in animes users) 103 of these 2% include least I at share. by market i.e., users viewership years, of million two 68% 2.6 least for the at (among of MyAnimeList.net period of a over users have and animes 50,000 103 than animes, 535 more these by Among viewed animes. been 535 to criteria down two animes these of on list Based the season. narrowed previous I a from e↵ects spillover potential avoid to season 14 Itookthefollowingstepstoarriveatthesetofuserstobeincludedinthefinaldataset: until anime an of release the from period time the as study under period the define I 4 faotosars h 0 nmshpe uigtesuypro scmae otetotal the to compared as period study the during happen animes 103 the across adoptions of 74% 14

iue14 ae sr ondMyAnimeList.Net Joined Users Dates 1.4: Figure Density 01jan2004 0 .001 .002 .003 01jan2006 01jan2008 01jan2010 14 Join Date 01jan2012 01jan2014 01jan2016 0.70 Days Member 30 days 0.60 60 days 90 days 0.50 180 days 1 yr 0.40 2 yr 3 yr 4 yr 0.30 5 yr Percentage 6 yr 0.20 7 yr 8 yr

0.10

0.00 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Months

Figure 1.5: Percentage of Friends Added During First 2 Years After Joining MyAnimeList.net (grouped by length of membership) under study.15 The choice of a one year cut-o↵was driven by the data. In Figure 1.5, I show the average percentage of friends added over the years for di↵erent groups of users based on their join date. Users grow their friendship network mostly during the ﬁrst six months after joining the website. I chose a conservative cut-o↵of one year. Second, I removed users who showed no activity after the release of the last anime. I deﬁne activity as any update to the watch list. For these users, I would not be able to di↵erentiate between them not adopting an anime under study because they did not want to or due to their inactivity. Therefore, I look at users who added at least one anime (not necessarily one of the selected animes in this study) to their watch list after the release of the last anime under study. Third, I dropped users who reported fewer than 10 adoptions of any anime (not only the ones selected for this study) over the entire observation period (i.e., at least 4 years). This is a very conservative criterion which ensures a minimal interest and activity level. And lastly, for some users I do not have data on all their friends’ adoptions e.g., because one of their friends’ watch lists is not public. Therefore I restrict my data to

15I refer the reader to Section 1.4.1 where I discuss in detail why this is necessary.

15 Figure 1.6: Average Number of News Articles (shaded area denotes 5th and 95th percentiles) users for whom I have adoption data on more than 95% of their friends. Note that this only a↵ects a small number of users. After applying these four criteria, the remaining data contain information on 39,652 users with nearly 170 million weekly observations. I use data on a random sample of 5,000 users with 21,853,295 weekly observations for the empirical analysis.

To account for the e↵ects of common shocks on adoption, I gathered data on the number of news published for each anime online and on MyAnimeList.net. To collect data on online news, I used google.com/news search results. One advantage of using Google news is that

Google also provides information on whether the same news article was published on several webpages or not. This allows us to not only follow the number of news for each anime over time, but also to capture the volume of news at each point in time. Figure 1.6 shows the average number of news articles online and on MyAnimeList.net for the animes under study over time.16

16The grey shaded area in Figure 1.6 shows the area between the 5th and 95th percentile.

16 Further, I also considered another type of common shock: the availability of an anime through legal online streaming channels. However, I found that more than 90% of the animes under study were available for online streaming within hours to up to three days after their original airing in Japan.17 Since my data are at the weekly level, I conclude that availability through legal channels is synonymous with original episode airings and do not include it as aseparatevariableinmyempiricalmodel.

1.3.2 Data Description

Table 3.1 summarizes key statistics of my data. On average, users have 18 friends, watch 76 animes per year, and adopt 17 of the 103 animes under study. Figures 1.7(a), 1.7(b), 1.7(c), and 1.7(d) show histograms of the number of friends, the average number of adopted animes per year, the number of adopted animes among those under study, and the adoption weeks

(for each anime relative to its own release date), respectively. Note that there is considerable variation in all four variables and that the distributions have very long right tails. Further,

35% of users indicate their gender as “Female,” 53% as “Male,” and the remainder did not specify their gender. On average, users adopt an anime in week 16 with a median adoption week of 13. Two spikes in adoptions around week 13 and week 26 are noticeable. Note that most animes have 13 or 26 episodes and are aired on a weekly basis. Thus these two spikes are likely due to a signiﬁcant number of users waiting for all episodes in a season to be available before they start to watch an anime.

In Figure 1.8, I show the average levels (across users and animes) of my six key variables capturing WOM and OL from the personal and community networks across time. The shaded area in each graph displays the 5th and 95th percentiles at any point in time. Figure 1.8a shows the average cumulative number of friends who adopted the anime (this is my measure

17Animes were mostly available for immediate online streaming on the international website crunchyroll.com.

17 (b) Average Number of Animes Adopted Per (a) Number of Friends Year (truncated at 200 friends) (truncated at an annual average of 1,000 animes) 4000 6000 3000 4000 2000 Frequency Frequency 2000 1000 0 0 0 50 100 150 200 0 200 400 600 800 1000 Number of Total Friends Number of Average Yearly Animes Adopted

(c) Number of Adopted Animes Among Animes (d) Adoption Week Under Study (relative to each anime’s release date) 6000 1.0e+05 8.0e+04 4000 6.0e+04 Frequency Frequency 4.0e+04 2000 2.0e+04 0 0 0 20 40 60 80 100 0 10 20 30 40 50 Number of Animes Adopted Weeks to Adopt

Figure 1.7: Histograms of the Number of Friends and of Descriptives Related to Adoption

18 (a) OL from Personal Network (b) OL from Community Network

(e) WOM Volume from Personal Network (f) WOM Volume from Community Network

Figure 1.8: WOM and OL from Personal and Community Networks

19 Table 1.1: Descriptive Statistics

Mean Std. Dev. Min Median Max N Age 23 6 11 23 84 24,584 Gender (% Females) 35 39,652 Gender (% Males) 53 39,652 Gender(%NotSpeciﬁed) 13 39,652 NumberofFriends 18 31 1 91,720 39,652 AverageNumberofAnimesAdoptedperYear 76 68 1 58 2,144 39,652 Number of Animes Adopted Among AnimesUnderStudy 17 18 0 8 103 39,652 AdoptionWeek(ConditionalonAdoption) 16 13 1 13 52 614,048 of OL from the personal network). Figures 1.8c and 1.8e show the average of ratings given by friends and the volume of ratings and forum posts by friends (these are my measures of

WOM valence and WOM volume from the personal network). Figures 1.8b, 1.8d and 1.8f show the number of community adoptions, the average community rating, and the number of community ratings and forum posts, respectively.18 These last three variables capture OL,

WOM valence, and WOM volume from the community network. The graphs show that my key variables vary considerably across time. More importantly, the shaded areas displaying the 5th and 95th percentiles at each point in time indicate that there is also considerable variation in my key variables across animes and users, especially for the personal network measures. For example, the cumulative number of friends who watched an anime ranges from 0 to 2 across users and animes by the end of week 1 and from 0 to 5 by the end of week

20. The average rating given by friends varies from 3.5 to 10 across users and animes by the end of week 1 and from 4 to 9.5 by the end of week 20. These patterns suggest that I have sucient variation in all my WOM and OL measures to identify their e↵ects on product adoptions.

18I refer the reader to Appendix B for details on these variables.

20 1.4 Model and Estimation

In this section, I start by discussing the three challenges I face in modeling choice interdependence in networks. Subsequently, I present the model and discuss my estimation approach.

1.4.1 Challenges

I face three main challenges in modeling and estimating the e↵ects of social learning on product adoptions with network data: endogenous group formation, correlated unobservables, and simultaneity (Hartmann et al. 2008). In this section, I explain how these issues pose challenges and how I address each of them. The challenge of endogenous group formation has two aspects. First, social ties can be formed to facilitate sharing common interests among people (Kozinets 1999). Observing others’ past actions can be used as a source of information to find individuals with similar interests. To study how people influence each other, I have to take into account that, while friends influence each others’ product adoptions, friendships themselves are formed under the influence of previous product adoptions. To solve this part of the endogeneity of tie formation, I focus on users who have been a member of the website for at least one year before the release of the first anime. As mentioned in Section 3.1, I observe users to form friendships mostly during the first six months (see Figure 1.5). Using data on users who have been members for at least one year enables us to assume that the networks are exogenous and fixed. Second, additional diculty arises due to the existence of homophily,19 i.e., friendship ties among users have been formed because users share the same interests. While two friends adopting the same product might be due to one influencing the other, it might as well be due to those similar interests. To tease homophily apart from influence, I need to control

19Homophily, which (Manski, 1993) referred to as “correlated e↵ects,” is the more prominent aspect of the endogenous network formation challenge.

21 for both each user’s intrinsic preference for a speciﬁc anime and each user’s propensity to adopt earlier as opposed to later. I do so by taking advantage of the rich panel structure of my data (Hartmann et al. 2008) and by incorporating user-anime and user-release week

ﬁxed e↵ects in my model.20

The correlated unobservables problem21 is caused by common shocks that inﬂuence both users and their friends’ product adoptions. In such a case, even if both the user and his friend adopt the product due to the shock, it can be mistaken as the user who adopted the product earlier inﬂuencing his friend. To account for common shocks, I use the following two approaches. First, I control for a variable that can a↵ect the adoption decisions of all users: the number of online news pieces collected from MyAnimeList.net and other websites.

Since animes are available through legal online streaming immediately after airing in Japan and the users of the website are located all over the world, I believe online news and in-site news posts on MyAnimeList.net are the main sources of common shocks.22 Therefore I use previous week’s number of online and in-site news to control for common shocks. And second, to account for other unobserved shocks that are common among all users (e.g., seasonality or platform malfunction), I incorporate (calendar) week ﬁxed e↵ects in my model.

The simultaneity problem, which is also known as the “reﬂection problem” (Manski

1993), arises due to a potentially simultaneous decision-making by a user and his friends.

20My empirical strategy of including user-anime and user-release week fixed e↵ects goes beyond previous literature’s approach of including group fixed e↵ects (Lee 2007; Lee et al. 2010; Ma et al. 2014). Note that the user-anime and user-release week fixed e↵ects subsume any group and/or any group-anime and/or any group-release week fixed e↵ects. Further, because the user-anime fixed e↵ects subsume any group (and any group-anime) fixed e↵ects, my identification approach for peer e↵ects does not rely on a specific definition/operationalization of “group” and even allows the “group“ to vary from one anime to another.

21(Manski, 1993) referred to this issue as “exogenous (contextual) e↵ects.”

22Note that the platform was created mainly to provide fans with an environment to connect to other fans. It was a non-proﬁt, commercial free, and completely user-driven platform until 2016 (i.e., until after the end of my observation period). No targeted actions or similar strategies (display ads, emails, or any other kind of targeting tools) were employed by the platform. Similarly, the platform did not provide users with any product or friend recommendations during the observation period.

22 In other words, a user might be inﬂuenced by his friends and, at the same time, inﬂuence those friends. I are able to address this challenge by using the lagged versions of variables capturing aggregated friends’ actions.

1.4.2 Model Description

The set-up of the model is as follows: Suppose there are i =1,...,N individuals and j =1,...J animes that an individual can adopt at time t =1,...,T¯j.Ideﬁneeachtimeperiodt as the tth week since release of anime j.Iobserveeachindividuali until his adoption of anime j in time period Tij or until the end of the study period for anime j, T¯j,ifindividuali does not adopt anime j.Iassumethattheendofthestudyperiodisindependentofanindividual’s adoption, i.e., there is no censoring of time. Given that I as researchers chose the length of the study period ex post, this assumption is reasonable. The adoption status of anime j by user i at time t is shown by yijt.Ifuseri adopts anime j at week t, yijt equals 1 and 0otherwise.Imodelusers’adoptiondecisionsusingalinearprobabilitymodel.Inother words, yijt is given by

cal yijt = ↵ij + it + t + Xijt1 + Zjt2 + Cijt3 + ✏ijt (1.1)

cal where ↵ij are user-anime fixed e↵ects, it are user-release week fixed e↵ects, t are calendar week fixed e↵ects, Xijt contains WOM and OL variables from the personal network, and

23 Zjt includes WOM and OL variables from the community network. Cijt contains other variables whose e↵ects I control for, namely, the number of animes adopted by individual i in week t,thenumberofnewspublishedaboutanimej in week t 1, a dummy variable indicating whether the season finale was aired in week t or t 1, and the interactions of the season finale dummy with each of the community OL and WOM variables. I include interaction e↵ects between the season finale dummy and the community WOM and OL

23Note I use one-week lagged versions of my WOM and OL variables to avoid the simultaneity problem.

23 variables to control for sudden jumps in the community WOM and OL variables due to increased adoptions around the season ﬁnale. 1 and 2 capture the e↵ects of WOM and

OL from the personal and the community networks, respectively. ✏ijt is the error term and

cal is assumed to follow a standard normal distribution. Finally, ✓ = ↵ij,it,t ,1,2,3 is the set of parameters to be estimated.

1.5 Results and Discussion

The estimation results for my main model are presented in column (ii) in Table 1.2. For comparison, I also show the results of a model without user-anime, user-release week, and calendar week fixed e↵ects in column (i) of Table 1.2. In interpreting my results, I focus on my main model shown in column (ii). I start by discussing the parameter estimates for the control variables. The parameters for the number of adopted animes in week t and the number of online news are, as expected, both positive and significant. I find that the season finale dummy has a significant negative main e↵ect and significant positive interaction e↵ects with the community WOM and OL variables. For a typical anime, the total e↵ect of the season finale is positive.

1.5.1 E↵ects of Word-Of-Mouth and Observational Learning

Next, I discuss the e↵ects of my key variables of WOM and OL from the personal and the community networks. I first start with the community network. As expected, community ratings have a significant positive e↵ect on a user’s anime adoption decisions. Recall that community ratings capture the valence of WOM since they are the average ratings given to animes by the whole community, while the community-wide number of ratings and forum posts for an anime captures the volume of community WOM. I find both WOM valence and volume to have positive and significant e↵ects. In other words, users are more likely to adopt animes that generate more buzz and more positive buzz in the community. The coecient

24 Table 1.2: Results

(i) (ii) (iii) (iv) Homogenous Model Main Model Asymmetric OL Model Awareness Model

Word-of-Mouth Friends’Av.RatingDummy -0.0010*** -0.0044*** -0.0042*** -0.0043*** (0.0000) (0.0002) (0.0002) (0.0002) Friends’Av.RatingInteraction 0.0002*** 0.0003*** 0.0003*** 0.0003*** (0.0000) (0.0000) (0.0000) (0.0000) Friends’NumberofRatings -0.0001 0.0038*** 0.0038*** 0.0037*** and Forum Postsa (0.0001) (0.0001) (0.0001) (0.0001) Community Rating 0.0009*** 0.0214*** 0.0214*** 0.0209*** (0.0000) (0.0002) (0.0002) (0.0002) CommunityNumberofRatings -0.0027*** 0.0002*** 0.0002*** 0.0003*** and Forum Postsa (0.0000) (0.0000) (0.0000) (0.0000)

Observational Learning Cum. Number of Friends Who Adopteda 0.0007*** 0.0029*** 0.0035*** (0.0000) (0.0001) (0.0001) Cum. Number of Friends Who Watcheda 0.0026*** (0.0001) Cum. Number of Friends Who Droppeda 0.0020*** (0.0002) Dummy for First Adoption by Friend 0.0032*** (0.0001) Cum.NumberofCommunityUsers 0.0012*** 0.0030*** 0.0030*** 0.0029*** Who Adopteda (0.0000) (0.0001) (0.0001) (0.0001)

Other Parameters NumberofAnimesWatchedDuring 0.0118*** 0.0078*** 0.0078*** 0.0078*** the Weeka (0.0000) (0.0000) (0.0000) (0.0000) Number of Online Newsa 0.0009*** 0.0001*** 0.0001*** 0.0001*** (0.0011) (0.0000) (0.0000) (0.0000) SeasonFinaleDummy -0.1585*** -0.1192*** -0.1192*** -0.1192*** (0.0000) (0.0010) (0.0010) (0.0010) Season Finale Dummy Community 0.0017*** 0.0019*** 0.0019*** 0.0019*** Rating⇥ (0.0000) (0.0001) (0.0001) (0.0001) Season Finale Dummy Community 0.0182*** 0.0137*** 0.0137*** 0.0137*** Number of Ratings and⇥ Forum Postsa (0.0000) (0.0002) (0.0002) (0.0002) Season Finale Dummy Cum.Number 0.0014*** 0.0007** 0.0007** 0.0007** of Community Users Who⇥ Adopteda (0.0003) (0.0002) (0.0002) (0.0002)

Constant 0.0042*** (0.0000) User-Anime Fixed E↵ects No Yes Yes Yes User-Release Week Fixed E↵ects No Yes Yes Yes Calendar Week Fixed E↵ects No Yes Yes Yes AdjustedR-Squared 0.0186 0.1238 0.1238 0.1239 NumberofObservations 21,853,295 21,853,295 21,853,295 21,853,295 Standard errors in parentheses * p<0.05, ** p<0.01, *** p<0.001 a Measured on logarithmic scale.

25 for the number of adoptions in the community, which captures the e↵ect of OL from the whole community, is positive and signiﬁcant. Therefore, as expected, my result suggests that the more popular an anime gets, the more likely it will be adopted by an individual. Note that, the signiﬁcant positive e↵ect of OL is after controlling for WOM valence and volume, suggesting that OL provides additional information to users.

To judge the relative magnitudes of the e↵ects, I use my parameter estimates to predict adoption probabilities for several scenarios. Speciﬁcally, I evaluate the changes in the likelihood of adopting an anime resulting from a 1% increase (calculated at the mean levels) in

WOM (valence and volume) and OL due to incoming adoptions, ratings, and posts. For OL, this is represented by 80 additional adoptions; for WOM volume, this is represented by 50 additional ratings and posts about an anime; and for WOM valence, this is represented by

49 additional ratings with an average of 7.55 for these additional ratings.24 For each of these three scenarios, I predict the change in average adoption likelihood (across all users and all animes) and ﬁnd the average adoption likelihoods to increase by 0.49%, 0.03%, and 0.6% due to a 1% increase in OL, WOM volume, and WOM valence, respectively. To put it in words, the 1% increase in WOM valence produces the largest increase in adoptions followed by the increase in OL and WOM volume.

Inowturntothee↵ectsofWOMandOLfromthepersonalnetwork.Iusethreevariables to capture the e↵ects of WOM from the personal network: a friends’ rating dummy which equals one if at least one friend in an individual’s personal network has submitted a rating for the anime and zero otherwise; friends’ average rating conditional on the friends’ rating dummy being one to capture WOM valence within the personal network; and the number of friends’ ratings and forum posts to capture WOM volume within the personal network. I include the friend rating dummy variable because, for some users, I observe a time period

2449 additional ratings represent 1% of the mean number of ratings (4,911) submitted for an anime. The mean rating in my sample is 7.48, i.e., the additional 49 ratings are 1% higher than the mean rating.

26 after the anime release when none of their friends have rated the anime yet. Given this data pattern, the friends’ rating dummy captures the e↵ect of the first rating submitted by a friend and friends’ average rating captures the valence of the ratings.25 Ifindthee↵ectofthefriends’ratingdummytobenegativeandsignificant,whileboth WOM valence and WOM volume from friends have significant positive e↵ects on a user’s anime adoption decisions. I also find the e↵ect of OL from one’s personal network to be positive and significant: as the number of friends who have watched an anime increases, an individual becomes more likely to adopt the anime. These results for the personal network indicate that OL and WOM provide users with di↵erent and unique information and have separate influences on individuals’ product adoption decisions. Thus the influence of social learning is not fully captured when only WOM or only OL is considered in an empirical study. Similar to the predictive exercise I conducted for WOM and OL from the community network, I evaluate the changes in anime adoption likelihood resulting from a 1% increase at the average WOM valence, WOM volume and OL levels from the personal network. However, in contrast to the community network, the 1% change in WOM volume from friends has a larger e↵ect on users’ adoption decisions (with an average increase in adoption likelihood of 0.27%) than those of OL or WOM valence (with average increases in adoption likelihoods of 0.21% and 0.01%, respectively). Lastly, using the results from the prediction exercises, I compare the e↵ects of WOM valence and OL across the two network levels. WOM valence from the community network has a much larger influence than WOM valence from the personal network. In fact, it is the largest adoption driver related to social learning overall both across network levels and learning mechanisms (i.e., OL versus WOM). Similar to WOM valence, OL from the

25A dummy of similar nature could be used for the average rating from the community network as well, but due to the large number of users in the network, all 103 animes under study have at least one rating in the ﬁrst week after release.

27 community network has a larger inﬂuence than OL from the personal network. This is probably due to the fact that the community of MyAnimeList.net is a large one consisting of more than 2.6 million users. As (Zhang et al., 2015) point out: When community networks are large, they provide more information to individuals than personal networks.

1.5.2 Positive and Negative Observational Learning

As discussed in Section 3.3, users can assign di↵erent stati to the animes on their watch list:

“watched,” “watching,” “on hold,” “dropped,” or “plan to watch.” I define a user as having adopted an anime if the anime is on his watch list under any of the first four stati in my main model. However, one can argue that the four stati contain di↵erent information about product adoptions. More specifically, while the adoption information can be viewed as either positive or neutral for the stati “watched,” “watching,” and “on hold,” it is clearly negative for the status “dropped” as this status suggests product abandonment after trial. I therefore use this more nuanced information on adoption stati to estimate an additional model in which

Idi↵erentiatebetweenOLcomingfrompositiveandnegativeproductadoptionexperiences within the personal network. To do so, I deﬁne positive OL as the act of adopting the product under the stati “watched,” “watching,” and “on hold” and negative OL as the act of adopting and abandoning the product under the status “dropped.”

The results are shown in column (iii) in Table 1.2. As expected, I find a significant positive coecient for positive OL from friends. I also find a significant positive, albeit significantly smaller coecient for negative OL. This finding is consistent with my expectation that the e↵ect of positive OL is indeed larger than that of negative OL. In other words, users do pay attention to the di↵erential informational content in their friends’ adoption stati when making their own watching decisions. However, to my surprise, the negative OL still enhances users’ adoption likelihood of an anime. One plausible explanation is that, compared to animes that no friend would even want to (watch and then) drop, animes with negative

28 OL from friends are perceived to be of better quality because they crossed some friends’ bar for an initial trial.

1.5.3 Awareness versus Learning about Unobserved Quality

OL from the personal network can influence a user in his decision whether to watch an anime in two ways: a user can become aware of an anime through his friends’ adoptions and/or he can learn about the unobserved quality of an anime from his friends’ adoptions (see also Fafchamps et al. 2016). To put it di↵erently, when a user first observes that a friend has watched an anime, this can both create awareness for the anime and let the user learn about the unobserved quality of the anime. However, friends’ subsequent adoptions only inform a user about the unobserved quality and do not create awareness for the anime since that has already been achieved through the first adoption by a friend. To turn this around, if I do not find a significant e↵ect of the first adoption by a friend, but a significant e↵ect for friends’ subsequent adoptions, it implies that quality information transfer and not awareness creation is the underlying mechanism for the e↵ect of OL from friends in my setting. IestimateanadditionalmodelinwhichIincorporateseparatecoecientsforthefirst adoption by a friend and for subsequent adoptions by friends. The results are shown in column (iv) in Table 1.2. My results reveal significant positive coecients for both the first and subsequent adoptions by friends. This finding implies that users both become aware of an anime and learn about its unobserved quality through OL from the personal network and is similar to the results found in (Fafchamps et al., 2016) in the context of an airtime transfer service.

1.6 Limitations and Future Research

There are several limitations to my research. First, I only have data on online WOM and OL. While in my empirical context of animes online information is likely to be the primary

29 source of information due to the special interest nature of animes, accounting for o✏ine WOM and OL might be important in other contexts. Second, while I observe five di↵erent stati (“watched,” “watching,” “on hold,” “dropped” and “plan to watch”) for each anime, I only model initial adoptions of an anime (episode) and do not investigate what drives individuals to watch multiple episodes, take a break in watching a series or drop it altogether. I leave this for future research to study. Third, the influence of WOM and OL may vary across users. I view not modeling the varying degree of susceptibility to peer e↵ects across users as a limitation of my study and leave this very interesting question for future research. And finally, I chose to focus on studying the adoption behavior of users who have at least one friend in the anime community in this paper. This decision is largely due to little to no activity and very sparse adoption behavior among stand-alone users in the anime community and implies that my paper does not address questions related to the adoption behavior of stand-alone users. I leave the topic of social learning of stand-alone users for future research.

1.7 Conclusion

Advances in technology have enabled ﬁrms to directly facilitate and manage social interactions and information sharing among consumers. A good understanding of the di↵erential and unique e↵ects of various social learning devices at di↵erent levels of a network is essential for ﬁrms to develop successful information provision strategies and eciently design their websites. In this paper, I study the role of social learning in individual consumers’ product adoptions. Drawn from the previous literature, I conceptualize that an individual can learn from and be a↵ected by peers in his personal network as well as all other users in the community network through two di↵erent mechanisms, namely, WOM and OL. Utilizing a unique data on individual users’ friendship networks and movie watching decisions from an anime website, I examine the e↵ects of both WOM and OL on users’ product adoptions

30 and quantify the relative importance of information obtained from one’s personal network as compared to the information obtained from the community network. My study thus complements the growing body of literature investigating the role of social learning in individuals’ online purchases and consumption decisions.

My empirical analysis reveals that both OL and WOM (both valence and volume) have signiﬁcant and positive e↵ects on individual user’s anime adoption decisions. Moreover, this

ﬁnding holds true for WOM and OL information coming from both the community network and the personal network. Thus my results highlight that WOM and OL provide unique and di↵erent information that individuals use in their product adoption decisions. I also

ﬁnd that social learning from the community network has a larger impact on individuals’ product adoptions than social learning from one’s immediate personal network. This result is consistent with the theoretical prediction in (Zhang et al., 2015) that community networks provide more accurate information to consumers when they are suciently large.

My results o↵er noteworthy policy implications for firms operating online streaming platforms. First, the predominant business practice in the online streaming industry has been to only display community-level movie ratings and popularity statistics. For a short time period in 2013, Netflix gave users the option to link their Netflix to their Facebook accounts and thus enabled direct information sharing about movies among friends. Currently, to the best of my knowledge, none of the major online streaming services in the U.S. provides users with the tools necessary to form personal n etworks. My results suggest that the less significant role the personal network plays vis-a-vis the community network in individuals’ movie watching decisions may explain online streaming platforms’ strategic decision not to provide information from personal networks within the platform.

Second, I ﬁnd that the e↵ect of community WOM valence is the largest adoption driver and overshadows the e↵ect of community OL (and also WOM volume). This result is consistent with the current product information provision practice found among leading online

31 streaming platforms. The top four U.S. online streaming platforms, i.e., Netflix, Hulu, Ama- zon and HBO Now, all provide average user ratings for movies and (TV) shows available at their websites. Netflix and Hulu also provide some information partially based on adoptions: Netflix shows “Top Picks” which are based on viewership and customization to an individual’s tastes and “Trending Now.” Hulu has a “Popular Shows/Episodes” and a “Popular Networks” category. However, it is unclear how and to what extent actual adoptions by individuals influence these featured categories. My results suggest that, from the perspective of enhancing movie watching, it might be worthwhile for Amazon and HBO Now to provide popularity information for their movies and (TV) shows alongside average ratings. Online streaming platforms can also consider displaying OL information directly in terms of adoptions, rankings, or similar metrics to encourage adoptions more eciently.

32 CHAPTER 2

THE EFFECTS OF BINGE-WATCHING ON MEDIA FRANCHISE

ENGAGEMENT

2.1 Introduction

The global entertainment and media industry reported revenues of $1.72 trillion in 2015

(Statistica 2016b) with $38.3 billion coming from the box oce and $286 billion from the TV and video industry (Statistica 2016a). A notable trend on both big and small screens is the rising success of media franchises. For example, the top-grossing movies of 2015 all belonged to franchises such as Star Wars, Jurassic World and Avengers. Franchise series also ruled the small screen as witnessed by the exploding online streaming trac at Netﬂix drawn by

Breaking Bad and House of Cards. I broadly deﬁne “media franchise” as a collection of media in which several derivative works have been developed in response to the popularization of an original creative work and the commercial exploitation of such through licensing agreements

(Aarseth 2006). For example, the media franchise of the American sitcom “Friends” consists of ten seasons of the TV series and a spin-o↵TV series named “Joey;” the media franchise of the Japanese anime “Pokemon” began as two video games and now spans into animated

TV shows and movies, trading card games, comic books, and toys.

Although industry observers have regarded media franchises as the overt success recipe for Hollywood because of the built-in awareness and interest with audiences (Garrahan 2014;

Gonzales 2014), little is known about the factors that contribute to consumers’ engagement with a media franchise. Marketing scholars and business practitioners have long been interested in consumer engagement or customer brand engagement which highlights customers’ interactive and co-creative experiences with ﬁrms and other customers (e.g., Bowden 2009;

Van Doorn et al. 2010; Mollen and Wilson 2010; Vivek et al. 2012). Empirical studies have shown that engaged customers play a key role in viral marketing activities by providing

33 product referrals and recommendations, in new product development, and in co-creating experiences and value (e.g., Nambisan and Nambisan 2008; Brakus et al. 2009; Hoyer et al. 2010) across various industries. However, to the best of my knowledge, no empirical study to date has systematically examined consumer engagement in the context of media franchises. Another prominent recent trend in the entertainment and media industry is the immense popularity of binge-watching, i.e., the practice of watching multiple episodes (of a series) in rapid succession. The percentage of consumers who indicate that they binge-watch increased from 62% in 2013 (Shannon-Missal 2013) to 92% in 2015 (TiVo 2015). Anecdotal evidence is abundant that binge-watching might increase viewer engagement with sequels and spin-o↵s. For example, Breaking Bad creator Vince Gilligan previously told Mashable that the show “may have met its demise after season two, had it not been for streaming video on demand. It ushered in new viewers and encouraged time-starved individuals to keep watching at their own pace resulting in enormous growth from season to season” that reached its climactic end in September 2013 with 10.3 million viewers (the show’s highest viewership ever) (Hernandez 2014). Similarly, for popular series such as “Supernatural,” Netﬂix starts streaming previous season(s) shortly before the release of a new season (on traditional TV). Despite what anecdotes and common practice suggest, there is little systematic empirical evidence to support the claim that binge-watching increases consumer engagement with a media franchise. Furthermore, how this engagement manifests itself is of interest to both academics and practitioners. (Calder et al., 2009) identify two types of engagement with media: personal engagement such as enjoyment and relaxation, and interactive engagement such as socialization and participation in a community. They associate the former with intrinsic motivation that leads an individual to getting caught up in the ﬂow of an activity and being absorbed by it (Csikszentmihalyi 1997), and the latter with extrinsic motivation that leads to an individual’s content generation and promotion of a focal media product. Therefore, increased engagement with a TV series might result in the viewer watching subsequent series, i.e., sequels, or other franchise extensions of the same series and/or in the

34 viewer promoting the TV series and producing user-generated content (UGC) about it. I explore these questions in this paper.

If binge-watching increases consumer engagement with a media franchise, this finding would have important implications for both online streaming services and traditional TV networks. For online streaming services, it would validate their practice of releasing a whole season of a series at once and thereby making it bingeable. For TV networks, it would provide support for their new strategy of promoting a new season shown on traditional TV by making older seasons available through online streaming services. This strategic tool could represent an especially important benefit for TV networks since it would not only increase immediate profits through higher advertising revenues (for the new season on traditional TV), but also extend the “life” of a series, making it more likely to reach five seasons at which point the series is a candidate for syndication, a very profitable path for networks.

If binge-watching does not increase media franchise engagement or if it does not do so for all series or consumers, it is important to understand why and when this is the case. For example, does the timing of the release through online streaming services matter? Or does the type of franchise extensions matter, for example, sequels might beneﬁt more from binge- watching than other types of franchise extensions such as spin-o↵s? Furthermore, given the varying popularity of online streaming and binge-watching across di↵erent countries, are consumers from some countries a↵ected more by binge-watching than consumers from other countries? In this paper, I provide systematic empirical evidence of this new mode of viewing and its e↵ects on consumers’ media franchise engagement.

My data come from MyAnimeList.net, an online forum that attracts anime (Japanese cartoons) fans from all over the world. I observe a user’s adoption of animes including information on the number of days it took a user to watch the whole season of an anime.

This allows us to classify user-anime combinations into “binged” and “not binged” cases.

Further, I observe a user’s self-generated content about an anime in the form of posts on

35 the discussion forum and recommendations. My data also contain information on a user’s decision to watch the next season (sequel) of an adopted anime and to watch other franchise extensions such as summaries, spin-o↵s, side stories, and remakes.1 And lastly, I have a large set of control variables including a user’s rating of watched animes and the user’s geographic location. Iusebivariatebinaryprobitmodelstoquantifythee↵ectsofbinge-watchingonauser’s actions related to media franchise engagement. The first binary probit equation describes the e↵ects of binge-watching and the second binary probit equation models the user’s decision to binge. I incorporate an instrumental variable that satisfies the exclusion criterion in the latter equation to account for the potential endogeneity of the decision to binge-watch. By simultaneously modeling the decision to binge and the decision to engage with a media franchise, I allow correlated unobservables to drive both decisions. My results for North American consumers show that binge-watching reduces a user’s probability of producing UGC. The e↵ect of binge-watching on a user’s franchise adoption, however, largely depends on both the availability of the franchise at the time of watching the focal series and the type of franchise extensions. If the franchise extension is available, bingeing the prior season significantly increases a user’s probability of watching the subsequent season (sequel), but decreases the adoption of other franchise extensions. If the franchise is not available at the time of watching the focal series, bingeing decreases the probability of watching other types of franchise extensions, but has no e↵ect on the adoption probability of the sequels. I extend my analysis using data on consumers from other continents (South America, Europe, Asia and Oceania). I find the e↵ects of binge-watching to be qualitatively consistent across the five regions if the coecient estimates are statistically significant. My paper makes the following contributions. First, I contribute to the vast consumer engagement literature by systematically examining the factors that drive consumer engagement

1I deﬁne each of these types of franchise extension in the Data Section.

36 in the context of a media franchise. By quantifying the e↵ect of binge-watching on consumer engagement with a media franchise in two broad areas, interactive and personal engagement, my paper provides empirical evidence that the modus of consumption, on top of product adoption, inﬂuences consumer brand engagement. And second, my paper adds to the small but rapidly growing literature on binge-watching and online streaming. To the best of my knowledge, I are the ﬁrst to establish the e↵ects of binge-watching on consumers’ subsequent media consumption and word-of-mouth behavior. My results have important managerial implications for both online streaming services and traditional TV networks regarding content provision and the timing thereof.

The remainder of the paper is organized as follows: In the next section, I discuss the relevant literature. In Sections 3.3 and 3.4, I describe my data, introduce my model and estimation approach. I present my results in Section 3.7. In Section 2.6, I conduct robustness checks and discuss limitations and future research in the following section. Finally, I conclude by summarizing my ﬁndings and discussing managerial implications in the last section.

2.2 Relevant Literature

In this section, I review relevant streams of literature on customer engagement, binge- watching, and online movie streaming.

2.2.1 User Engagement

Customer “engagement” has been extensively studied in the marketing literature (e.g., Bow- den 2009; Mollen and Wilson 2010; Van Doorn et al. 2010; Vivek et al. 2012).2 It di↵ers from similar relational concepts such as participation or involvement in that it highlights customers’ interactive and co-creative experiences in networked relationships with multiple

2I refer readers to (Brodie et al., 2011) for an extensive review of the marketing literature on engagement.

37 stakeholders including service personnel, ﬁrms, and/or other customers (Brodie et al. 2011).

Empirical studies have shown that engaged customers play a key role in viral marketing activities by generating referrals and recommendations for products and services, in new product development, and in co-creating experiences and value across various industries

(e.g., Nambisan and Nambisan 2008; Brakus et al. 2009; Hoyer et al. 2010). However, to the best of my knowledge, no empirical studies to date have systematically examined consumer engagement in the context of media franchises. (Wei, 2016) studies the movie industry’s decision of producing original vs. imitative work and rationalizes the industry’s exceeding reliance on sequels and franchises as a result of ﬁrms attempting to reduce demand uncertainty and securing ﬁnances for new movies. Despite the popularity and success enjoyed by media franchises on both big and small screens, little is known about the factors that contribute to consumers’ engagement with a media franchise.

In a separate stream of literature, (Calder et al., 2009) define engagement in terms of the di↵erent motivational experiences that consumers have with a media product. Using a confirmatory factor analysis, they identify two types of engagement with media, personal engagement and interactive engagement. The former includes individualistic experiences such as enjoyment and relaxation, while the latter is especially relevant to online media and includes experiences such as socialization and participation in a community. While personal engagement with a media product is associated with intrinsic motivation, i.e., the individual getting caught up in the flow of consuming the product and being absorbed by it (Csik- szentmihalyi 1997), interactive engagement is associated with extrinsic motivation, i.e., the individual creating content and voluntarily engaging in media promotion activities. Follow- ing (Calder et al., 2009), I study users’ interactive and personal engagement with a media franchise. In my empirical context of an online anime platform, the former includes a user’s content generation and promotion of a focal media product, e.g., a user’s recommendations, comments, and responses published in a community discussion forum regarding a TV se-

38 ries. The latter includes a user’s self-enjoyment of the focal product and the adoption of its franchised extensions including sequels, spin-o↵s, summaries, side stories, and remakes.

2.2.2 Binge-Watching

The Merriam-Webster dictionary deﬁnes binge-watching as “Watch(ing) many episodes (of a television program) in rapid succession, typically by means of DVDs or digital streaming (Merriam-Webster.com 2017).” This deﬁnition is consistent with (Schweidel and Moe,

2016) who consider “the consumption of multiple episodes of a television series in a short period of time” as binge watching. Many regard the element of control as an essential part of binge-watching, which distinguishes binge-watching from watching marathon releases of series episodes back to back on regular TV channels (Jenner 2015; Pittman and Sheehan

2015). In other words, binge-watching is not only about watching multiple episodes in one sitting, but it is also about a user’s control and decision on when and what to watch. In addition, the presence or absence of interruptions such as commercials separates marathon releases on TV channels from binge-watching by means of DVDs or digital streaming (Jenner

2015).

There is disagreement on how much watching is considered binge-watching. Many studies rely on respondents’ perception of what is considered binge-watching without deﬁning a speciﬁc amount (e.g., Devasagayam 2014; Pena 2015). Based on a survey of their users,

Netflix defines binge-watching as watching at least two episodes in one sitting (Netflix 2013).

This is in line with the idea that binge-watching is a violation of what is considered the norm, regular TV watching or “appointment watching” (Jenner 2015). The number of two episodes is not agreed upon by everyone though. For example, Amazon made the ﬁrst

3 episodes of its series “Alpha House” available to its viewers at once, implying that it considers 3 episodes as a bingeable amount. Some studies view binge-watching as watching with the purpose of ﬁnishing a whole season in a short period of time (Devasagayam 2014;

39 Pena 2015). However, this view is not necessarily supported by consumer surveys. In a

MarketCast study, 71% of respondents indicated that they do not plan on bingeing, but they end up doing so. Furthermore, these deﬁnitions focus on the number of episodes without di↵erentiating between one-hour dramas (about 40 minutes without commercials) and 30- minute sitcoms (about 20 minutes without commercials). It is debatable whether watching

8 episodes of a sitcom corresponding to about 2.5 hours is considered binge-watching. In this paper, I suggest a clear deﬁnition of binge-watching which is based on the time spent watching a whole season and test its robustness.

Many reasons have been found for binge-watching. People binge to catch up on a series they missed when it was aired on TV (MarketCast 2013; TiVo 2015) or to be able to participate in the word-of-mouth created by the series (Pittman and Sheehan 2015). According to TiVo’s annual binge behavior report, 32% of respondents indicated that they postpone watching a series until it has aired completely so that they can binge the whole season (TiVo

2015). This ﬁnding is consistent with (MarketCast, 2013)’s ﬁnding that one of the main reasons for binge-watching is that viewers cannot or do not want to wait for each next episode.

The (TiVo, 2015) study also ﬁnds that 39% of the respondents consider it more enjoyable to binge a series as opposed to appointment watch it (Pittman and Sheehan 2015; TiVo 2015).

Some people binge-watch TV to relax (Devasagayam 2014; Pittman and Sheehan 2015). For example, after a week of hard work, they binge-watch during the weekend to restore or as a reward to the point that they even plan for it beforehand. On the other hand, on weekends, holidays, or summer holidays for students, people might binge-watch because they are bored, have no better alternative, or feel lonely and want to compensate for their limited social life

(Devasagayam 2014; Pittman and Sheehan 2015; Sung et al. 2015).

The underlying mechanism that drives binge-watching is related to the concept of “ﬂow”

(e.g., Ho↵man and Novak 1996), which describes a state of focus concentration, intrinsic enjoyment, and time distortion. Previous research has found that users who experience the

40 ﬂow are more likely to repeat their behaviors or even become addicted (e.g., Kubey and

Csikszentmihalyi 2002; Chou and Ting 2003). This mechanism also provides a plausible explanation for the interplay between advertisements and binge-watching as documented in

(Schweidel and Moe, 2016): Advertisements in a viewing session discourage binge-watching and binge-watchers are less responsive to advertisements compared to non-binge-watchers.

While there has been a considerable amount of research on the reasons for binge-watching, few studies have focused on the consequences of binge-watching. In the (TiVo, 2015) study,

52% of respondents indicated that they felt sad when they finished bingeing a series; 31% reported that they have lost sleep due to binge-viewing. Binge-watching - due to the intensity of the experience and the flow it creates - has been suggested to create loyalty to a series, lead to fandom or, at the very least, behavior similar to fandom such as purchasing ancillary materials, creating fandom pages or posting or creating content (Jenner 2015). However, empirical evidence supporting these claims is very limited. To the best of my knowledge, this paper is the first to carry out a systematic empirical examination on the e↵ects of binge-watching on user engagement with a media franchise.

2.2.3 Online Movie Streaming

Despite its wide popularity, research on online movie streaming is scarce. (Cha and Chan-

Olmsted, 2012) study the plausible cannibalization e↵ect of online video platforms on traditional TV by examining the perceived substitutability between the former and the latter.

They ﬁnd that users of online video platforms believe that online video platforms have unique functionality and therefore are not substitutes for traditional TV. However, non-users of online video platforms perceive online video platforms as substitutes for traditional TV because of their perceived similar functionality. (Cha, 2013) ﬁnds that the more consumers perceive online video platforms to di↵er from traditional TV in satisfying their needs, the more likely they are to use online video platforms.

41 Studying consumer behavior within online streaming services, (Ameri et al., 2016) investigate the drivers of consumers’ anime adoption decisions. They ﬁnd average anime ratings and popularity rank from the community network to have larger e↵ects on consumers’ adoption decisions compared to the same type of information obtained from the personal network. (Zhang et al., 2013) develop a new class of “clumpiness” measures and, using data from Hulu.com, show that the “clumpiness phenomenon” is widely prevalent in digital content consumption. In a separate study, (Zhang et al., 2015) extend the traditional re- cency/frequency/monetary value (RFM) segmentation framework to include the clumpiness measure (RFMC). In particular, they show that the RFMC framework can help companies with bingeable content (such as online streaming platforms) uncover previously unseen customer segments. And lastly and most closely related to this paper, (Schweidel and Moe,

2016) simultaneously examine the drivers of users’ binge-watching behavior and their responses to advertisements using data provided by Hulu.com. They ﬁnd that binge-watchers are less responsive to advertising compared to non-binge-watchers.

2.3 Data

My data come from MyAnimeList.net. This website was established in November 2004, but its main activities did not begin until 2007 when the website moved to a public domain and its user base started to grow rapidly (see Figure 2.1). At the point in time when I started the data collection (March 2015), there were more than 2.5 million users on the website.

MyAnimeList.net is a consumption-related online community where online interactions are based upon shared enthusiasm for a speciﬁc consumption activity (Kozinets 1999).

MyAnimeList.net was created to allow anime fans to gather and share their excitement and opinions about animes. Over the years, the website has developed into one of the most comprehensive online sources of information about animes (Japanese cartoons). On

42 acigbhvo.Tu r ofietta h efrpre dpindt r eibei ycontext. my in reliable are data actual adoption their self-reported self- report the panelists’ correctly to that survey tend confident behavior. compare surveys, people watching are 2016) that incentivized find I anime Staelin, and to Thus true and data contrast their streaming (Lovett behavior. actual watching in shows, report the TV and falsely that data of to note viewing setting MyAnimeList.net Further reported similar on the users 2.4. in for Figure Furthermore, incentives discussing no when are concern there this funimation.com, address hulu.com, I netflix.com, as such channels erent ↵ di of others. number date and a the aniplexusa.com through crunchyroll.com, indicate illegally can or users Lastly, legally ratings.” “user as lists watch on animes on to them given rating ratings by list Further, watch their identified. on animes uniquely the and ascalerangingfrom1to10(10beingthehighestrating).Throughoutthispaper,Ireferto about correctly opinion are their indicate animes also all can that users so function search a using paper). this throughout list” “watch as list date. join website the or location geographic user’s addition the in as shown such characteristics is personal recommendations her and to posts adopted about forum opinion and her ratings), dates), numerical the informa- (via (including page, animes adopted user’s has a user the On animes pages. the own about tion their have users and animes both MyAnimeList.net, 4 3 sr a raeals faie htte lnt ac rhv ace Irfrt this to refer (I watched have or watch to plan they that animes of list a create can Users yaoto aaaesl-eotd hsacrc nterprigo dpin saptnilconcern. potential a is adoptions of reporting the in accuracy either Thus animes watch self-reported. are can data users adoption general, My in because, paper this in choice platform for account not do I

iue21 ae sr ondMyAnimeList.Net Joined Users Dates 2.1: Figure Density 01jan2004 0 .001 .002 .003 01jan2006 01jan2008 3,4 01jan2010 43 Join Date oeta sr d nmst hi ac lists watch their to animes add users that Note 01jan2012 01jan2014 01jan2016 they started watching an anime series, the date they ﬁnished watching an anime series and the website also automatically registers the date users last updated the entry for an anime. Iusethesedatestoinferthetimeofadoption.

2.3.1 Data Cleaning

I scraped data on 370,000 users from the website through snowball sampling. Not all users list start dates for (all or any) animes they have adopted on their watch list. After excluding all anime-user combinations for which I did not have start dates, I were left with 92,273 users. I then dropped (i) animes for which I did not have the release date or information on the number of episodes; (ii) anime-user combinations for which the watch period seemed unreasonably long, i.e., more than 3,000 days; (iii) observations for days on which users indicated to have watched more than 24 hours of animes; (iv) observations with start dates before 2008 since, although the website was launched in 2004, its main activities did not start until mid 2007 (see Figure 2.1); (v) observations with start dates after the end of 2014. Using the remaining 89,422 users and 4,896 animes (3,481,664 user-anime combinations), I took the following steps to get to my final data. First, I dropped animes that would take less than 3 hours to watch. Table 2.1 shows the distribution of animes with respect to their number of episodes and durations. Movies or short anime series generally take less than 3 hours to be watched and thus, according to my operationalization of binge-watching (see Section “Binge-Watching”), cannot be binged. Note that even if a user watches 3 movies back to back, since they are not part of a series, I do not consider this instance as binge-watching. Second, I dropped cases in which a user did not watch the whole season. Even if a user binge-watches the first half of a season, her behavior might be di↵erent compared to someone who finished the whole season. To be able to attribute the di↵erence in behavior to the viewing modus of binge-watching and not the completion of the whole season, I only consider cases in which the user finished watching the whole season.

44 Table 2.1: Number of Episodes in and Duration of a Season

Number of Episodes Freq. Percent Duration of Series (Hours) Freq. Percent 1781.59lessthan186717.71 274015.091-290618.50 3-7 892 18.19 2-3 307 6.27 8-11 192 3.92 3-4 142 2.90 12 691 14.09 4 - 5 715 14.60 13 627 12.79 5 - 6 440 8.99 14 - 27 956 19.50 6 - 10 417 8.52 28 - 56 566 11.54 10 - 15 495 10.11 57 and more 161 3.28 15 - 20 252 5.15 20 and more 355 7.25

Third, I only consider cases in which users have the option to binge the anime, but may choose not to do so, i.e., I only consider watching incidences after the season ﬁnale of an anime has been aired. It is noteworthy that most of my observations are for such cases (see Figure

2.2). After these three steps, my ﬁnal data sample for the empirical analysis related to UGC contains 73,346 users and 2,715 animes (1,298,786 user-anime combinations). Next, to study user engagement in the form of watching franchise extensions, I only consider animes that have a franchise extension, i.e., next season (sequel) or other franchise extensions (side story, spin-o↵, summary, remakes). Note that some animes have multiple franchise extensions such as a spin-o↵and a summary. I consider adoption of each type of franchise extensions as a separate adoption, but only consider the ﬁrst adoption out of multiple adoptions of the same type.

2.3.2 Engagement

Istudyauser’sengagementwithaseriesintwoareas:interactiveandpersonalengagement.

First, a user might engage with an anime by producing UGC in the form of recommendations and posts on the discussion forum. On the platform, users can communicate with other users about the anime on an anime’s discussion forum page. I operationalize communication

45 (a) Binge-watchers: Number of days (b) Binge-watchers: Number of days after first episode (truncated at 1,000 after season finale (truncated at 1,000 days) days) .008 .004 .006 .003 Density Density .004 .002 .002 .001 0 0 0 200 400 600 800 1000 0 200 400 600 800 1000 Days after release Days after finale (c) Non-binge-watchers: Number of (d) Non-binge-watchers: Number of days after first episode (truncated at days after season finale (truncated at 500 days) 500 days) .15 .06 .1 .04 Density Density .05 .02 0 0 0 100 200 300 400 500 −500 0 500 Days after release Days after finale

Figure 2.2: Number of Days After Release of First or Final Episode in a Season That Animes Were Watched engagement as two indicator variables: whether a use wrote (at least) one recommendation and whether a user wrote (at least) one post on the discussion forum.5

Second, a user who is engaged with an anime might get more caught up in the story line and be more likely to watch its franchise extensions. My data contain information on whether a user watched the next season (sequel) and/or other franchise extensions, including spin-o↵s, summaries, side stories, and remakes. A “side story” is a short story related to the

5Conditional on writing at least one recommendation or at least one forum post, the median number of recommendations and forum posts users write are 1 and 2, respectively.

46 main characters in the context of the focal series. For example, the movie “Sherlock: The Abominable Bride” is a side story for the “Sherlock” series. A “spin-o↵” is a story taken from the focal series, however, unrelated to the main story. It usually tells the story of a secondary character following a di↵erent storyline, almost like a new series. For example, the “Joey” series is a spin-o↵from the popular sitcom series “Friends.” A “summary” is a short series or amoviesummarizingtheeventsofthefocalseries.Forexample,the“PinkPanther”movie is a summary of the events in the identically titled TV series. A “remake” is a remake of the series, usually with small di↵erences in the plot or a di↵erent ending. For example, there are several “Batman” series that are remakes of the same story. I operationalize personal engagement as an indicator variable: whether a user watched the franchise extension (at any point in time in the future).

2.3.3 Binge-Watching

Idefineauserashavingbinge-watchedananimeiftheuserwatchestheseriesforover3hours on a single day, a more conservative measure than Netflix’s.6 In the Robustness Section, Itesttherobustnessofthisdefinitionwithrespecttoshorterandlongerwatchtimes.To di↵erentiate binge from non-binge incidences, I use the average daily time that a user spent watching a series (calculated by dividing the total length of the series by the number days that it took the user to finish watching the series). For example, if a user watches more than an average of 3 hours a day (corresponding to about 8 25-minutes long episodes, excluding the few minutes of openings and endings), I mark this incidence as binge-watching.7 Iuseinformationonthestartandenddatesfromusers’watchliststomeasure“watch period,” the number of days it took a user to watch an anime season. I have both pieces

6Netflix defines binge-watching as watching at least two episodes in one sitting (Netflix 2013).

7Note that a user might have watched more than 3 hours on a Sunday, but it took him Monday to Friday to gradually watch the remaining 3 episodes (about 1 hour) and ﬁnish the series. My data do not allow us to identify the watching behavior on Sunday as binge-watching.

47 nsc ae,Imgtlblabnewthn aea o-ig-acigcs ftepro ewe start between period the if case non-binge-watching long. a is the as date case that update binge-watching Note last a the label respectively. and might date binge, I not cases, such did In and binged they which during day a on animes are cases take user-anime not of does portion 2.3 significant non-binged binged. Figure potentially being (which a hours vs. that of shows binged terms 2.3 into in Figure season account), classification into a my of While length the on two. depends or combinations animes day user-anime of a 50% 18.62% within than with days, watched More 5 sample. being within my watched in are days) cases (in user-anime the periods watch of of distribution the shows 2.3 Figure the for Description date Data finish the date. 2.3.4 of finish a lieu provided in not update did last user the registered for which this in entry i.e., use cases the watchlist, user-anime in I her change last on anime. the anime specific made an user that the to for when change registers a website the made anime user automat- individual website a each the time However, user-anime last date. of the finish the 7% registers not remaining ically but the date, For start the have combinations. I user-anime combinations, of 93% for information of 8 nFgrs24a n .() ipa h oa ubro or niiul watched individuals hours of number total the display I 2.4(b), and 2.4(a) Figures In h nycag srcnmk fe niaigta h a opee h eisi digarating. a adding is series the completed has she that indicating after make can user a change only The iue23 ac eidDsrbto tuctda 0 days) 200 at (truncated Distribution Period Watch 2.3: Figure

Density 0 .1 .2 .3 .4 .5 0 50 Watch Period−Days 48 100 150 200 8 (a) Days with Binge-Watching (b) Days with No Binge-Watching 1.5 .6 1 .4 Density Density .5 .2 0 0 0 3 6 9 12 15 18 21 24 0 3 6 9 12 15 18 21 24 Binge−Hours Per Day Non Binge−Hours Per Day

Figure 2.4: Number of Hours Watched Per Day total number of hours includes everything the user watched, i.e., all animes the user binged on that day and any other animes the user might have watched on that day. On days during which users binge-watched, the vast majority of users watched between 3 and 6 hours with a second, smaller group of users watching between 9 and 11 hours. While the distribution has a long right tail, very few users report watching more than 16 hours a day. This gives us conﬁdence in the accuracy of the self-reported watching behavior (see also (Netﬂix, 2013)).

On days during which users do not binge-watch, almost all users watch less than 3 hours.

This is not a direct result of my deﬁnition of binge-watching since Figure 2.4 shows the total number of hours users spent on anime watching. For example, users who watch 7 episodes of one anime series and 7 episodes of another anime series would not be classiﬁed as bingeing on that day, but would have watched more than 3 hours.

Figure 2.5 shows the distribution of the fraction of animes on a user’s watch list that can be classiﬁed as binged vs. not binged using the 3-hour cut-o↵. About 41.8% of users do not binge-watch at all, while for 6.5% of users bingeing is how they watch all animes.

This implies that, although some users can be called binge-watchers and others non-binge- watchers, most of the users binge some and gradually watch other animes. This empirical observation is consistent with previous ﬁndings (e.g., MarketCast 2013; Schweidel and Moe

49 (a) Including 0% and 100% (b) Excluding 0% and 100% .1 .4 .08 .3 .06 .2 Density Density .04 .1 .02 0 0 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 percentage of animes percentage of animes

Figure 2.5: Percentage of A User’s Watch List That Is Binge-Watched

2016). On average, I classify 20.4% of animes on a user’s watch list as binged with a standard deviation of 28% and a median of 8.3%.

Figure 2.6 displays how the number of binged vs. non-binged cases has evolved over time.

Up until about 2013, both the number of binged and the number of non-binged cases is gradually increasing. Starting in 2013, the number of binged cases continues to increase, while the number of non-binged cases starts to decrease, implying that more users are bingeing animes instead of gradually watch them. This pattern of an increasing proportion of users who binge-watch is consistent with ﬁndings reported in several survey studies (e.g., TiVo

2015).

In Table 2.2, I report statistics related to users’ engagement measures between binged and not binged cases across the five continents. I first describe the data patterns in UGC behavior for North American users. 11% of users who binged made at least one forum post about the anime they watched, while this percentage is 14% for users who did not binge (di↵erence is statistically significant at p<0.001). Furthermore, 5% of users who binged wrote a recommendation for the anime, while 6% of users who did not binge wrote arecommendation.Thisdi↵erenceisstatisticallysignificantatp<0.005. Both results

50 ig-ac .5 .6 .9 .5 .1 0.036 0.044 0.049 0.119 0.035 0.135 0.044 0.353 0.085 0.047 0.343 0.105 0.291 0.040 0.298 0.081 0.288 0.566 0.037 0.319 0.095 0.304 0.050 0.317 0.082 0.558 0.757 0.306 0.522 0.058 0.328 0.113 0.329 0.341 0.111 0.683 Binge-Watch 0.551 0.749 Non-Binge-Watch 0.325 0.536 0.345 0.139 Oceania 0.345 0.299 0.702 Binge-Watch 0.545 0.755 Non-Binge-Watch 0.345 0.557 0.307 Asia 0.287 0.708 Binge-Watch 0.557 0.762 Non-Binge-Watch 0.280 0.537 Europe 0.722 Binge-Watch 0.543 0.724 Non-Binge-Watch America South 0.689 Binge-Watch Non-Binge-Watch America North srcmiain htbne esnwthdtenx esn(eul,wieti per- this while (sequel), season next the watched season a binged that combinations user the for patterns data The content. similar. produce very to are likely regions less other are binge who users that suggest et ics srbhvo eae ofacieetnin.O vrg,6%o anime- of 69% average, On extensions. franchise to related behavior user discuss I Next, iue26 ig-acigv o-ig-acigArs Time Across Non-Binge-Watching vs Binge-Watching 2.6: Figure vial o vial vial o vial ou ot Recommendation Posts Forum Available Not Available Available Not Available etSao te rnhssUGC Franchises Other Season Next

al .:Poaiiyo naeetAction Engagement of Probability 2.2: Table 01jan2008 Frequency 0 1.0e+04 2.0e+04 3.0e+04

01jan2009

01jan2010

01jan2011 igdNotBinged Binged 51

Adoption 01jan2012

01jan2013

01jan2014

01jan2015 centage is 67% for non-bingeing cases (not shown in Table 2.2). However, the data patterns related to the adoption of the sequel crucially depend on the availability of the next season at the time of the viewership of the prior season. In Table 2.2, I show the percentage of adoptions of the next season for binged and not binged cases depending on whether the next season was (not) available. If the sequel was available, a person who binged the prior season was significantly more likely to watch the next season compared to someone who did not binge (72% vs. 69%; di↵erence is statistically significant at p<0.001). However, if the sequel was not available yet, users who binged and did not binge were equally likely to adopt the next season (once it became available). For other franchise extensions, the proportions of users who watch these franchise extensions are similar across binge-watchers and non-binge-watchers. I find similar patterns across the five continents.9

In Table 2.3, I compare the timing of when users produce UGC and/or watch franchise extensions depending on whether they binged the focal anime. On average, users who binged posted in the forum section 4 days after they started to watch, while users who did not binge posted in the forum section 8 days after they started to watch (di↵erence is statistically significant at p<0.001). Users who binged wrote recommendations for a series 167 days after starting to watch, while users who did not binge wrote recommendations 194 days after starting to watch (di↵erence is statistically significant at p<0.001). If the sequel was not available at the time of watching the prior season, users who binged started watching the next season 77 days (median) after its release while users who did not binge the prior season started watching the next season 72 days (median) after its release. The di↵erence, however, is not statistically significant. If the sequel was available at the time of watching the prior season, binge-watchers started watching it earlier than non-binge-watchers (a median of 1

9In Tables F.1 and F.2 in Appendix E, I show the same statistics for alternative definitions of binge- watching (2 and 4 hours compared to my main specification of 3 hours). My empirical findings are mostly robust to these alternative definitions.

52 Table 2.3: Timing of Engagement Actions

Median (Days) Obs Next Season After Release of Focal Season If Next Season Is not Available Non-Binge-Watching 72 59,420 Binge-Watching 77 13,036 After Finishing Focal Season If Next Season Is Available Non-Binge-Watching 2 213,848 Binge-Watching 1 53,504

Other Franchises After Release of Focal Season If Franchise Is not Available Non-Binge-Watching 105 28,288 Binge-Watching 121 6,469 After Finishing Focal Season If Franchise Is Available Non-Binge-Watching 4 164,577 Binge-Watching 1 46,369

UGC Posting in Discussion Forum After Starting to Watch Non-Binge-Watching 8 13,035 Binge-Watching 4 2,510 Posting Recommendation After Starting to Watch Non-Binge-Watching 194 1,185 Binge-Watching 167 313 vs. 2 days after finishing the prior season; di↵erence statistically significant at p<0.001). I find a similar pattern for other franchise extensions. Figure 2.7 compares the distributions of ratings for binged and non-binged cases. Both distributions are very similar with slightly more ratings of 9 and 10 for the binged cases. The average ratings for binged and non-binged cases are 7.85 and 8.04, respectively. Even though the di↵erence in average ratings is small, the average ratings are significantly di↵erent at p<0.001. This implies that a person who binges an anime either thinks better of the anime because of bingeing it and/or has higher intrinsic interest in the anime and therefore binges it compared to someone who does not binge. This empirical observation is consistent with previous findings on reasons for binge-watching (e.g., Pittman and Sheehan 2015).

53 (a) Binged Cases (b) Non-Binged Cases .3 .3 .25 .25 .2 .2 .15 .15 Density Density .1 .1 .05 .05 0 0 0 2 4 6 8 10 0 2 4 6 8 10 Rating Rating

Figure 2.7: Distribution of Ratings for Binged vs. Non-Binged Cases

2.4 Model and Estimation

Iquantifythee↵ectsofbinge-watchingontwoareasofusers’mediafranchiseengagement: (1) production of user-generated content, i.e., comments in the discussion forum and recommendations, and (2) viewership of franchise series, i.e., watching of the next season (sequel) or other franchise extensions (spin-o↵s, side stories, summaries, and remakes). User i’s propensity to take an action related to anime j is a latent continuous variable and denoted by yij⇤ .Inthedata,Iobserveuseri to take an action yij for anime j (e.g., writing of a recommendation or watching of the next season) if her propensity supersedes a threshold. User i’s propensity to produce UGC is given by

yij⇤ = ↵i + Bingeij + Cij + Gj + T + ✏ij (2.1) 1 yij⇤ > 0 yij = 8 , <>0 otherwise where ↵ is a user-speciﬁc intercept which comes from a normal distribution with N(↵,2 ) i :> ↵ and Bingeij is a dummy variable indicating whether user i binge-watched anime j.Further,

Iincludeothervariableswhosee↵ectsIcontrolfor. Cij contains a dummy variable indicating whether user i submitted a rating for anime j and, if so, the rating user i gave to anime j,

54 the popularity rank of anime j based on the number of users who have adopted it, and the average user rating given to anime j by all adopters on the website by the time of adoption of user i. Note that all these variables are publicly visible on the platform. Gj consists of anime-speciﬁc variables, namely, anime j’s genre dummies, the number of episodes in a season, and the length of each episode in minutes. And lastly, T contains year dummies.

User i’s propensity to adopt a franchise extension of anime j is modeled similarly. In addition to controlling for Cij, Gj,andT , I also include a dummy variable for series type (sequel or not) and an interaction term between series type and binge watching, i.e.,

y⇤ = ↵ + Binge + ⌘Sequel + Binge Sequel + C + G + T + ✏ ij i ij j ij ⇥ j ij j ij (2.2) 1 yij⇤ > 0 yij = 8 , <>0 otherwise > where Sequel: j is a dummy variable indicating whether the franchise extension is a sequel. The e↵ects of binge-watching on users’ franchise adoptions (next season and other franchise extensions) depend on the availability of the franchise extension at the time of the viewership of the focal anime (see Table 2.2). Consequently, I estimate separate models for the e↵ects of binge watching on franchise extension adoptions when the extension is and is not available. Furthermore, in the case that the franchise extension is not available, I control for the number of days the user has to wait after ﬁnishing the focal anime until the franchise is released.

Potential endogeneity of the decision to binge-watch is a concern. To account for this concern, I simultaneously model user i’s decision to binge-watch the focal anime and to produce UGC or to adopt a franchise series, respectively, allowing for the error terms across the two equations to be correlated (see Heckman 1978, Maddala 1983, p. 123-159, Wilde

2000, Wooldridge 2010, p. 595-507). User i’s decision whether to binge-watch anime j is given by

55 Bingeij⇤ = ↵i0 + 0wij + 0Cij + 0Gj + 0T + ✏ij0 (2.3) 1 Bingeij⇤ >e0 Bingeij = 8 , <>0 otherwise where Binge is the underlying latent variable capturing user i’s propensity of bingeij⇤ :> watching anime j.ThevariableBingeij (whose realizations I observe in the data) equals

1ifBingeij⇤ is positive and 0 otherwise. Bingeij⇤ is a function of user random e↵ects ↵i0,a weekend dummy wij,andCij which contains the popularity rank of anime j based on the number of users who have adopted it at the time of adoption of user i,andtheaverageuser e rating given to anime j by all adopters by the time of adoption of user i. Additionally, I control for Gj and T which contain the same sets of variables as described for equation (2.1). And lastly, I allow the two error terms in equations (2.1) and (2.3) and equations (2.2) and 1 ⇢ (2.3) to be correlated, i.e., ✏ij,✏ij0 N(0, ⌃) with ⌃= 2 3. ⇠ ⇢ 1 6 7 Iincludeaweekenddummy,wij,asanexclusionvariableinequation(2.3).Iexpectusers4 5 to have more time on weekends and thus to be more likely to binge-watch, while this variable should have no e↵ect on users’ subsequent adoption of franchise extensions and production of UGC. Table 2.4 shows the distribution of weekdays when users start to watch animes. Both binge-watching and non-binge-watching cases occur more often on weekends compared to weekdays (average of 15.78% vs. 13.69%). However, binge-watching is 1.76% more likely to happen on weekends (32.44% vs. 30.68%). Apotentialconcernisthatusersmightnotonly(binge-)watchthefocalanimeduringa weekend, but also show more engagement with the media franchise during the same weekend, i.e., that the weekend dummy might not be a valid exclusion variable. Descriptive statistics in Table 2.3 provide evidence for the validity of my instrument. My data show that, even if a franchise series is available, the mean number of days until a user starts watching a franchise series from when she ﬁnishes the focal anime is signiﬁcantly greater that two days.

56 Table 2.4: Weekday Frequencies of Start Dates

Weekday Percent Percent Non-Binge-Watch Binge-Watch Monday 14.64 14.22 Tuesday 13.96 13.51 Wednesday 13.73 13.17 Thursday 13.49 13.08 Friday 13.51 13.57 Saturday 14.84 15.89 Sunday 15.84 16.55

Equations (2.1) and (2.3) and equations (2.2) and (2.3) constitute bivariate probit models. I estimate the models using Simulated Maximum Likelihood Estimation (SMLE). The likelihood of each model is given by + N 1 J L = P (y =1 X, ⇥, ⌃,↵) P (Binge =1 X, ⇥, ⌃,↵0) d↵ d↵0d⌃, (2.4) ij | i · ij | i i i i=1  j=1 Y Z1 Y where ⇥= ↵,↵0,,0,⌘, ,,0,,0,,0,⇢ is the vector of coecients to be estimated { } and X = Binge ,C , C ,w ,G ,T represents the vector of covariates. { ij ij ij ij j } e 2.5 Results

I start by discussing the results for North America (Table 2.5).10 The lower half of Table 2.5 shows the results from the probit models capturing the decision to binge. Across the four probit regressions, the coecient estimates have the expected signs and most of them are significant: my exclusion variable, a weekend dummy, has the expected significant positive e↵ect in three of the four regressions. The lower the popularity rank of the focal season (i.e., a better rank), the more likely it is that a user will binge it. More and longer episodes also increase the probability of binge-watching. Lastly, I find evidence for a significant amount of unobserved heterogeneity across users.

10I show the results from univariate probit regressions ignoring the potential endogeneity of the binge- watching decision in Table D.1 in Appendix D.

57 Table 2.5: Results - North America

Franchises UGC Franchise Available Franchise Not Available Forum Post Recommendation (i) (ii) (iii) (iv)

Engagement Equation Binge Dummy -0.139** -0.151* -0.535*** -0.590 (0.047) (0.073) (0.134) (0.308) Sequel Dummy 1.399*** 0.762*** (0.021) (0.028) Binge Dummy Sequel Dummy 0.152*** 0.117*** ⇥ (0.023) (0.034) Own Rating of Focal Season 0.140*** 0.176*** 0.048*** 0.050* (0.004) (0.008) (0.008) (0.023) Own Rating Dummyb -0.954*** -0.119 -0.005 (0.038) (0.091) (0.323) Popularity Rank of Focal Seasona -0.100*** -0.081*** -0.047*** 0.004 (0.004) (0.01) (0.009) (0.031) Community Rating of Focal Season -0.153*** -0.103*** -0.047* -0.142 (0.011) (0.022) (0.024) (0.085) WaitTimeUntilFranchiseSeriesAvailable -0.160*** When Started Watching Focal Seasona (0.007) NumberofEpisodesofFocalSeason -0.038** 0.200*** 0.370*** 0.278** (0.010) (0.025) (0.029) (0.092) Duration of an Episodea -0.193*** 0.354*** 0.210* 0.225 (0.036) (0.084) (0.085) (0.281) Constant 1.718*** -1.536*** -2.360*** -1.972 (0.154) (0.343) (0.37) (1.218) Variance of User Random E↵ects 0.425*** 0.571*** 0.758*** 0.229*** (0.016) (0.035) (0.059) (0.051) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes

Binge Decision Equation WeekendDummy 0.061*** 0.037* 0.050** 0.067 (0.010) (0.017) (0.017) (0.055) Popularity Rank of Focal Seasona -0.063*** -0.070*** -0.035*** -0.048 (0.005) (0.009) (0.008) (0.028) Community Rating of Focal Season -0.009 0.029 0.037 0.032 (0.012) (0.02) (0.02) (0.071) NumberofEpisodesofFocalSeason 0.233*** -0.251*** 0.294*** 0.326*** (0.012) (0.024) (0.023) (0.077) Duration of an Episodea 0.281 -0.164 0.630*** 0.780* (0.050) (0.084) (0.082) (0.356) Constant -2.853*** 0.546 -4.458*** -4.875*** (0.197) (0.336) (0.353) (1.316) Variance of User Random E↵ects 1.147*** 0.886*** 0.651*** 0.617*** (0.046) (0.053) (0.06) (0.143) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes Error Correlation 0.120** 0.071 0.298** 0.416 (0.032) (0.048) (0.099) (0.266) NumberofObservations 164,666 44,346 57,965 6,224 AIC 308,517.224 92,990.404 82,869.404 8,609.047 BIC 309,648.543 93,973.479 83,864.807 9,329.817 Log Likelihood -154,145.612 -46,382.202 -41,323.702 -4,197.524 Standard errors in parentheses. * p<0.05, ** p<0.01, *** p<0.001 a Measured on logarithmic scale. b Not estimated in all models due to collinearity.

58 The top half of Table 2.5 shows the results of the probit regressions describing users’ engagement actions. In general, compared to other types of franchise extensions, sequels have a higher probability of being adopted. The e↵ects of binge-watching on users’ franchise adoptions (next season and other franchise extensions) di↵er depending on the type of franchise extension and on whether the franchise was available at the time of the binge: if it was, binge-watching significantly increases the probability of watching the next season (at p<0.05), while it decreases the probability of watching other franchise extensions (at p<0.01).11 These mixed findings on the e↵ects of binge-watching are consistent with my expectation that the “flow” created by binge-watching is closely related to the story line involving the main characters in the binged series. Previous research indicates that users who experience the flow are more likely to repeat their behaviors or even become addicted in order to stay in the flow (e.g., Kubey and Csikszentmihalyi 2002; Chou and Ting 2003). Among the di↵erent kinds of franchise extensions, sequels, i.e., next seasons, are the ones that continue the same story line of and share the same main characters with the prequel or previous season. Other franchise extensions may have a di↵erent story line or center around di↵erent characters (e.g., “Better Call Soul,” as a spin o↵of “Breaking Bad,” follows the story of a lawyer who was a secondary character in “Breaking Bad.”).12 As a result, the natural way for users to continue the flow after bingeing a season is to watch the next season. The significant negative e↵ect of binge-watching on other franchise extensions can be attributed to the net e↵ect of enjoyment or immersion in the flow and the physical/mental burden of bingeing (Grøntved and Hu 2011; Matrix 2014; Sung et al. 2015). For sequels, the flow is strongly continued in sequels. As a result, despite the fatigue, a user is more likely

11The e↵ect of binge-watching on other franchise extensions is captured by the binge-dummy. The e↵ect of binge-watching on sequels is the combined e↵ect of the binge dummy and the interaction terms between the binge and sequel dummies. In the text, I describe the combined e↵ect in terms of its size, direction, and signiﬁcance when referring to the e↵ects of binge-watching on sequels.

12Other examples are “Frasier” as a spin-o↵of “Cheers,” “Joey” as a spin-o↵of “Friends” and “The Good Fight” as a spin-o↵of “The Good Wife.”

59 to adopt the sequel. In the case of other franchise extensions, however, the ﬂow is disrupted by a new story line introduced in the extension. Consequently, a user has less motivation to overcome the fatigue and to spend more time watching shows.

If the franchise extension is not available at the time of the binge, binge-watching significantly decreases the probability of adopting another franchise series (at p<0.05), while it does not a↵ect the adoption of sequels. The ﬂow can only continue if the next season of the series is available at the time when users adopt the focal season.

Next, I discuss the e↵ects of binge-watching on the production of UGC. For both forum posts and recommendations, I find that bingeing decreases the probability that a user produces content. The e↵ect is significant is at p<0.001 for forum posts and at p<0.10 for recommendations (p<0.10 not shown in Table 2.5). This result can also be explained by bingers’ inclination to stay in the flow. Therefore they tend to avoid any activities that distract them from watching. This avoidance tendency is also manifested in (Schweidel and

Moe, 2016) where the authors ﬁnd that binge-watchers are less responsive to advertisements compared to non-binge-watchers. Another plausible explanation is that, compared to binge- watchers, it takes non-binge-watchers a longer time and more viewing sessions to complete an anime season. The more frequent (yet not as deep) interaction with the media product may lead to a higher likelihood to generate product-related UGC.

My control variables related to the focal anime (own rating, rating dummy, popularity rank, community rating) have consistent e↵ects across the four regressions and are mostly significant. The wait time until the franchise extension becomes available in regression (ii) has a significant negative e↵ect, i.e., the longer users have to wait for a franchise extension to become available, the less likely they are to watch it. And lastly, I find a significant amount of unobserved heterogeneity across users.

60 Lastly, I discuss the results for the other geographic regions included in my data: South

America, Europe, Asia, and Oceania.13 Table 2.6 shows the coecient estimates for the binge e↵ects for all engagement measures and all continents.14 The complete sets of results showing all coecient estimates can be found in Appendix ??. Across the five regions, consumers are significantly more likely to watch sequels than other franchise extensions. Consistent with the results for North America, binge-watching significantly increases the adoption probability of sequels (if those are available) in Asia (at p<0.001) and the estimates are directionally consistent but insignificant for the other continents (see footnote 11 for a description of how these e↵ects were calculated). If the next season is not available, the e↵ect of binge-watching on the adoption probability of sequels is insignificant across all regions. And lastly, the e↵ect of binge-watching on the adoption of other franchise extensions is consistently negative (if significant).

2.6 Robustness Checks

I assess the robustness of my main data pattern in Appendix F. Recall that I deﬁne a user as having binged if she watches more than 3 hours of an anime series per day. In Appendix F,

Iexploretwoalternativedefinitionsofbinge-watching,namely,havingwatchedmorethan2 hours and having watched more than 4 hours of an anime series per day. Tables F.1 and F.2 show the probabilities of producing UGC and watching franchise extensions under these two alternative definitions of binge-watching. Overall, I find data patterns under the alternative definitions that are similar to the data patterns under my main definition.

13Note that my data are much smaller for South America, Oceania, and Asia than for North America and Europe.

14I show the corresponding results from univariate probit regressions ignoring the potential endogeneity of the binge-watching decision in Table D.2 in Appendix D.

61 Table 2.6: E↵ects of Binge-Watching Across Di↵erent Regions

Franchises UGC Available Not Available Forum Posts Recommendation North America Binge Dummy -0.139** -0.151* -0.535*** -0.590 (0.047) (0.073) (0.134) (0.308) Sequel Dummy 1.399*** 0.762*** (0.021) (0.028) Binge Dummy Sequel Dummy 0.152*** 0.117*** ⇥ (0.023) (0.034)

South America Binge Dummy -0.204* 0.045 -0.568 -0.830 (0.100) (0.061) (0.376) (0.486) Sequel Dummy 1.238*** 0.744*** (0.039) (0.035) Binge Dummy Sequel Dummy 0.205*** 0.012 ⇥ (0.044) (0.078)

Europe Binge Dummy -0.131* -0.091 -0.298* -0.002 (0.038) (0.087) (0.129) (0.076) Sequel Dummy 1.228*** 0.914*** (0.015) (0.032) Binge Dummy Sequel Dummy 0.198*** 0.032 ⇥ (0.016) (0.044)

Asia Binge Dummy 0.124 -0.288 -0.563 0.148 (0.087) (0.200) (0.324) (0.143) Sequel Dummy 1.273*** 0.904*** (0.038) (0.065) Binge Dummy Sequel Dummy 0.172*** -0.023 ⇥ (0.040) (0.075)

Oceania Binge Dummy -0.205 -0.038 -1.038 -1.553 (0.151) (0.295) (0.723) (1.662) Sequel Dummy 1.265*** 0.876*** (0.060) (0.100) Binge Dummy Sequel Dummy 0.330*** -0.014 ⇥ (0.060) (0.013)

Standard errors in parentheses. * p<0.05, ** p<0.01, *** p<0.001

62 2.7 Limitations and Future Research

There are several limitations to my research. First, a media franchise can also include merchandising items that are available for purchase, such as posters, co↵ee mugs, toys, and trading card games. In my data, I do not observe (o✏ine) purchases of such ancillary products. It is left for future research to investigate whether the viewing modus of bingeing a↵ects (o✏ine) purchases. Second, even though I provide evidence for the validity of my data, measurement error in my binge-watching variable remains a potential concern. Measurement error might be due to the self-reported character of the data or due to the usage of the date of the last update in cases in which the ﬁnish date is missing (see Data Section). It is well- known that measurement error in an independent variable leads to attenuation bias, i.e., a bias of the coecient towards zero. Thus my results should be interpreted as a lower bound of the e↵ects of binge-watching.

Third, some shows have a higher probability of being binged than others. While I quantify the e↵ects of variables such as weekend or ratings on the probability that a user binges, I do not model the e↵ects of di↵erent creative content. It is left for future research to study whether and how di↵erent characteristics such as story line characteristics, episode openings and endings make a show more or less bingeable. And lastly, di↵erent methods or channels of watching such as online streaming websites, streaming platforms, DVDs, or piracy websites might produce varying degrees of bingeing behavior. Channels deploy di↵erent interfaces, advertising methods, and sequential watching strategies, which can inﬂuence binge-watching behavior.

2.8 Conclusion

With the introduction of video-on-demand services during the last decade, binge-watching has become very common among TV viewers. An open empirical question is whether the

63 viewing modus has implications for user engagement compared to the traditional, linear way of watching TV. Built on extant literature, I argue that binge-watchers would want to stay in the “ﬂow,” a state of concentrated focus created by binge-watching. In this paper, using novel data coming from an online anime platform containing information on individual users’ adoptions of di↵erent animes and their user-generated content, I quantify the e↵ects of binge-watching on consumers’ engagement with a media franchise as related to user-generated content and the adoption of franchise extensions. My paper thus adds to the small but rapidly growing body of literature on consumers’ digital media consumption as well as on the online streaming industry. To the best of my knowledge, my paper is the

ﬁrst systematic empirical examination of the e↵ects of binge-watching on user engagement with a media franchise.

My results show that binge-watching decreases the production of UGC. For the video elements of the media franchise, the e↵ect of binge-watching crucially depends on both the type and the availability of franchise extensions at the time of watching the focal anime series. If the franchise extension is available, binge-watching increases the probability that a user watches the next season, while it has the opposite e↵ect for other franchise extensions.

If the franchise extension is not available, binge-watching decreases the probability of other franchise extensions being adopted and has no e↵ect on the adoption of sequels. My results are directionally consistent across the ﬁve continents if the coecients are signiﬁcant.

My results o↵er the following important managerial implications for TV channels and online streaming platforms. First, binge-watching can boost viewership of subsequent seasons

(sequels). However, the availability of the subsequent season plays a crucial role. Companies have started to recognize this by making prior seasons available (for binge-watching) shortly before the release of the next season. Figure 2.8 shows several examples from Netﬂix.

Second, binge-watching does not boost viewership of all franchise extensions. Which franchise extensions would beneﬁt from a bingeable prior season depends on whether the

64 Figure 2.8: Examples of Release Dates on Netflix franchise extension would help continue the flow viewers experience when bingeing the prior season. Franchise extensions that di↵er significantly in story lines and/or main characters may not attract binge-watchers of the prior season. The general lackluster performance of spin-o↵s speaks to the importance of staying close to the successful original series when developing franchised extensions.15

Third, online streaming networks such as Netflix have been aggressive in expanding their services beyond the home country. My study provides first empirical evidence regarding the similarities and di↵erences in consumers’ media consumption and engagement behaviors across five continents. Specifically, I find that the e↵ect of binge-watching are directionally consistent across the di↵erent regions. These findings provide valuable information that helps online steaming companies decide to what extent their content strategy in general and

15Wikipedia lists 1,142 TV spin-o↵s on its website (https://en.wikipedia.org/wiki/List of television spin- o↵s). Only 135 spin-o↵s (12%) ran for 5 or more seasons. 413 spin-o↵s (36%) ran for one season or less.

65 content release timing strategy in particular should be customized to accommodate local consumers’ preferences.

66 CHAPTER 3

A MODEL OF NETWORK DYNAMICS: TIE FORMATION, PRODUCT

ADOPTION, AND CONTENT GENERATION

3.1 Introduction

(Online) social networks have become an indispensable part of many consumers’ everyday lives. By the end of 2016, there were 2.3 billion active social network users worldwide1 with the average user having more than 5 accounts across di↵erent platforms.2 Despite these astonishing numbers, new (online) social networks are still being started every month.3 The goal of social networks is to connect people. The number of connections users have can grow very fast especially in the online environment. For example, Instagram users saw, on average, a monthly growth of 16% in the number of their followers in 2017.4 As the size and connectedness of social networks grow, people are increasingly sharing information and communicating with each other through these networks. For example, every minute,

Facebook users shared 1.3 million pieces of content and Twitter users sent 0.35 million tweets in 2016.5

Research has shown that social networks greatly facilitate information dissemination through social learning (e.g., Duan et al. 2009; Katona et al. 2011; Christakis and Fowler

2013). Social networks provide a natural platform for users to produce and publish content

(e.g., posts in discussions forums, product reviews, re-shares) related to various activities

1https://www.statista.com/statistics/278414/number-of-worldwide-social-network-users/

2https://insight.globalwebindex.net/social-q4-2014

3For example, Mastodon, a micro-blogging service, was founded in October 2016 and had 81,000 users by July 2017 (https://mastodon.social/about/more).

4https://socialpilot.co/blog/151-amazing-social-media-statistics-know-2017/

5https://socialpilot.co/blog/125-amazing-social-media-statistics-know-2016/

67 they engage in (e.g., Toubia and Stephen 2013) and to disseminate these contents to others beyond their own personal network (e.g., Moe and Trusov 2011). Moreover, these opinions and experiences shared by friends or other users through social networks have a signiﬁcant e↵ect on network users’ decisions in a wide range of product markets (e.g., Bramoull´eet al.

2009; Carrell et al. 2009; Aral and Walker 2011; Ameri et al. 2016). However, with the exception of (Toubia and Stephen, 2013), all these results were found for mature social networks. It is an open empirical question whether they also hold for platforms in which the social network is still evolving and changing.

In this paper, I study the endogenous formation of an online social network in which users can become friends, adopt products, and produce user-generated content (UGC). Despite their signiﬁcant role in today’s society, very little is known about how (online) social networks develop and evolve and, in particular, how and with whom people become friends. An intriguing aspect of making friends in many online social networks is that people do not know each other’s real identities. As a result, an individual’s behaviors and opinions as observed by other people in the online environment are the main factors inﬂuencing friendship formation decisions in these types of networks. If online behaviors and opinions are likely drivers of friendship formations, studying online networks provides valuable insights into important factors that drive people’s friend making decisions.

For platform owners – since online social networks mostly rely on advertising revenue for proﬁtability and thus for platform survival – it is crucial to understand how social networks evolve and what drives users’ friendship formations and activity levels (e.g., production of UGC). This knowledge would allow them to devise e↵ective stimulation and seeding strategies that leverage friends’ social inﬂuence to promote (more) friendships and other activities on the platform. For example, previous literature has shown that seeding to more connected users, i.e., users with more friends, is the most e↵ective strategy to encourage the di↵usion of a behavior (e.g., Hinz et al. 2011; Aral et al. 2013). However, these studies

68 investigated mature (static) social networks. Thus their finding may or may not hold true in evolving networks in which a stimulation intervention is very likely to also change individuals’ friendship networks and increase connectedness. As a result, to accurately evaluate the e↵ectiveness of seeding strategies in evolving networks, one also needs to take the endogenous formation of new ties into account. I do so in a counterfactual in this paper. The extant empirical studies on network formation typically examine changes in the network structure at a macro level, but are insucient in providing a micro-foundation for network tie formations from an individual actor’s perspective (see Jackson 2008 and Toivonen et al. 2009 for a review). A notable exception is a small set of studies that are known as strategic network formation models (Christakis et al. 2010; Snijders et al. 2010). In both papers, (Snijders et al., 2010) and (Christakis et al., 2010), the authors model future states of a network based on characteristics of the existing network state (current ties). However, the models proposed in these two papers are not appropriate for the study of network formation in an online environment for the following reason: Both papers have data on at most a few snapshots of the network and not the continuous development of the network. As a result, the authors resort to simulating network states between observed snapshots. Due to the simulated network states, besides changes in network structure, the authors are not able to capture the e↵ects of time-varying variables such as users’ activities on tie formation. While (Snijders et al., 2010) and (Christakis et al., 2010) explain the formation of ties via static user characteristics such as age or gender, it is possible that users’ time-varying behaviors and actions are also significant drivers of their friendship formation decisions. This problem is aggravated in online environments where an individual’s observed online activities and opinions user such as product adoptions or UGC may even be the more important drivers of friendship formation decisions since personal characteristics such as age or gender are either unavailable or non-verifiable. Furthermore, while people’s activities and opinions influence the friendships they form, their future activities and opinions are also subject to the influence of their friends. The latter

69 is often termed social inﬂuence, peer or network e↵ects in the literature (e.g., Sacerdote 2001;

Katona et al. 2011; Iyengar et al. 2011). However, the endogeneity of network formation makes it a challenging task to correctly identify network e↵ects (Manski 1993). Previous literature has suggested several approaches to deal with this challenge, e.g., instrumental variables (e.g., Bramoull´eet al. 2009; De Giorgi et al. 2010), correlated group e↵ects (e.g., Lee

2007; Lee et al. 2010; Ma et al. 2014), randomness/exogenous shocks (Sacerdote 2001; Tucker

2008), experiments (e.g., Aral and Walker 2011), individual-speciﬁc unobserved preferences

(e.g., Nair et al. 2010; Trusov et al. 2010), and co-evolution models (e.g., Snijders et al. 2007).

In this paper, I build on and extend the last two approaches: I account for individual-speciﬁc unobserved preferences of performing an action and for the interdependence among actions, while explicitly modeling the evolution of the network to which an individual belongs.

My data come from a special interest online community website for animes (Japanese cartoons) called MyAnimeList.net. This website provides a gathering place for anime fans from all over the world to interact with each other and to form friendships. Since anime fandom is a special interest and anime fans are scattered around the world, the online channel naturally becomes the main venue through which anime fans interact with each other. This implies that most users of MyAnimeList.net do not know each other before forming their friendship ties, and that the actions they observe on the website are the main drivers of the friendship decisions — making this platform an ideal environment for my research inquiry.

I take advantage of this novel data set that documents both the continuous development of the network, i.e., which individuals become friends with each other and when that happens, and all users’ entire activity histories on the platform, i.e., anime watching and

UGC posts. Users report anime watching through personal watch lists, i.e., lists of animes that they have watched together with the dates of doing so, and exchange information and opinions, i.e., produce UGC, through publishing discussion forum posts, anime reviews, and recommendations on the platform. I observe the dates of all UGC posts as well. Access

70 to these data containing the complete network evolution and all of users’ actions on the platform allows us, unlike (Christakis et al., 2010) and (Snijders et al., 2010), to model the friendship network development without the need to simulate the state of the network at each point in time and thus I are able to measure the e↵ects of time-varying variables such as anime watching or UGC posts on the probability that two individuals become friends.

I model the endogenous formation of a social network and the occurrence of two types of online activities, namely, product adoptions and content generation, over time. More specifically, each day, a user makes three types of decisions: (i) with whom to become friends, (ii) whether to watch any anime, and (iii) whether to make a UGC post. All decisions are made with a utility-maximizing framework. I model friendship tie formation between two individuals as non-cooperative decisions. Each individual maximizes her own friendship formation utility which depends on the attractiveness of the potential friend and the similarity between the pair. A friendship tie is formed if and only if both users agree to it. A user’s utilities of participating in either product adoptions or content generation are functions of past online activities of the same type and of the user’s friendship network which can a↵ect her actions through peer e↵ects.

Furthermore, to capture any common shocks unobserved by the researcher which might result in simultaneity, I include time fixed e↵ects and also allow the error terms in the utility functions (for friendship formations, product adoption, and content generation) to be correlated within a day. To tease apart homophily from peer e↵ects, I estimate individual-specific intrinsic propensities to watch animes and to post UGC. To estimate these individual-specific unobserved preferences in the utility functions for these two online activities, I use observations before and after a user has made friends. I further incorporate individual-specific propensities to make friends in the friendship formation utility to capture any benefits and costs that users associate with making friends. The three utility functions are connected in three ways: through observed variables, through correlated error terms, and through corre-

71 lated individual-specific unobserved preferences. By explicitly accounting for the interdepen- dences between network formation and individuals’ online activities in the modeling and by using data from both before and after individuals make friends to estimate individual-specific observed preferences, I are able to provide a clean identification of peer e↵ects.

My results for friendship formation reveal the relative importance of users’ activities versus users’ friendship networks in friend making decisions. I find that the friendship networks of other users with whom a focal user is not friends (yet), both in terms of how many friends other users have and how many common friends other users have with the focal user, are more important than other users’ activities. In addition, even in (anonymous) online networks similar demographics matter. I further find significant e↵ects of a user’s friends on the user’s in-site activities, i.e., product adoptions and the production of UGC.

However, I do not find any spill-over e↵ects of one type of activity of friends on the other type of activity of a focal users, i.e., friends’ product adoptions (content generation) do not influence the focal user’s content generation (product adoptions). Lastly, while having more friends does not make a user more active, having active friends does increase a user’s activity level due to the positive social influence.

Iusemyresultstosimulatevariouscounterfactualscenariostoassessthee↵ectiveness of various seeding and stimulation strategies in increasing users’ activities on the website.

Contrary to previous studies investigating static networks (e.g., Hinz et al. 2011; Aral et al.

2013), my results for evolving networks reveal that seeding to well-connected users, i.e., users with many friends, is not always the best strategy to increase users’ UGC activities on the platform. This ﬁnding might be due to well-connected users generally becoming friends with other well-connected users who are not necessarily the ones who publish the most posts or watch the most animes. Further, I ﬁnd that not accounting for the endogenous network formation in evolving networks when assessing the e↵ectiveness of seeding strategies leads, on average, to an underestimation of seeding e↵ectiveness by 10%.

72 The contribution of this paper is three-fold. First, my paper contributes to the strategic network formation literature by quantifying the e↵ects of individuals’ time-varying actions on friendship formation and their relative importance compared to users’ static characteristics such as age or gender. This richer specification is much needed when describing the network development in an online environment. Second, my paper contributes to the social learning and network e↵ects literature by providing a novel approach to identify the influence of friends’ activities. Specifically, I account for the latent homophily by explicitly modeling the choice of friends and by incorporating individuals’ intrinsic preferences for performing actions that are identified absent of their friends’ influence. And lastly, to the best of my knowledge, this paper is one of the first papers to model the strategic co-evolution of a friendship network along with users’ actions within the network. By understanding the interdependent dynamics of network formation and online activities, my model yields important insights regarding strategies for companies and network platform owners to e↵ectively stimulate user engagement.

The remainder of this paper is organized as follows: In the next section, I discuss the relevant literature. In Section 3.3, I describe my data and in Sections 3.4 to 3.6, I introduce my model, estimation approach, and identiﬁcation strategy. I present and discuss my estimation and simulation results in Sections 3.7 and 3.8. In the following section, I examine limitations of the current work and opportunities for future research. Finally, I conclude by summarizing my ﬁndings in Section 3.10.

3.2 Relevant Literature

In this section, I review three relevant streams of literature on network formation, peer e↵ects, and seeding strategies in social networks.

73 3.2.1 Network Formation

Researchers have studied network formation using three main modeling approaches: nodal attribute models, exponential graph models, and strategic network formation models. The

first category of models explains the existence of ties and the resulting network structure via similarities among pairs of individuals (e.g., Ho↵et al. 2002; Boguñáet al. 2004). However, this modeling approach explains the status quo of a network, i.e., who is/is not friends with whom, rather than its evolution. Further, it fails to take the structure of network connections into account. Exponential graph models explain the network development based on structural patterns such as triangular connections or transitivity, but do not provide insights into the mechanisms that drive individuals’ tie formation decisions (e.g., Katona and

M´ori 2006; Mele 2017). These models are well suited for making predictions but not for causal inferences and therefore they are not conducive for counterfactual analyses. To overcome these shortcomings, strategic network formation models, the most recently developed modeling approach among the three, have taken the perspective of individual actors’ utility maximizations when explaining the evolution of a network, and allow them to depend on the existing state of the network (e.g., Hanaki et al. 2007). Strategic network formation models are also known as network evolution models (Toivonen et al. 2009) or actor based models

(Snijders et al. 2010) in the economics literature.6 This paper falls into the last category of modeling approaches.

(Christakis et al., 2010) and (Snijders et al., 2010) are two empirical papers in the strategic network formation literature that are most closely related to my study. Both papers develop dynamic models to explain network formation through the e↵ects of the existing network state (current ties) on individuals’ future friendship tie formation decisions in the network.

(Christakis et al., 2010) propose a structural model in which each person’s utility from

6I refer the interested reader to (Jackson, 2008) and (Toivonen et al., 2009) for a detailed comparison of the three categories of models.

74 forming a friendship tie with another person is a function of the number of friends the other person has as well as the distance or overlap between their respective friendship networks. In their model, given the chance of meeting based on a distribution of random sequences, each pair of individuals decide whether to become friends. However, due to the data constraint of observing only one snapshot of the network, (Christakis et al., 2010) only incorporate static user characteristics such as age and gender in the utility function besides the existing network state. They apply their model to a network of 669 students with 1,541 friendship ties and ﬁnd that, while having common friends is important for friendship formation, people are less likely to become friends with popular individuals. They also ﬁnd that people prefer to become friends with people who are similar to themselves in terms of characteristics such as age or gender.

Unlike (Christakis et al., 2010), (Snijders et al., 2010) use a continuous Markov process to model the formation of the network as individuals decide whether to consider another individual as a friend. They use a data set that contains a few snapshots of a network.

Although having more observations of the network gives (Snijders et al., 2010) more ﬂexibility in modeling than (Christakis et al., 2010), due to not knowing the continuous sequence of tie formations, they are still unable to incorporate the e↵ects of individuals’ time-varying actions on their utilities for forming ties. They test their model using an empirical setting of a classroom of 32 students and 6 snapshots of the network and ﬁnd that individuals prefer bi-directional ties to un-reciprocated uni-directional ties. In addition, students tend towards networks with transitive connections that are not singled out from the rest of the network.

In terms of demographic characteristics, they ﬁnd that males are more popular as friends, but that similarity of genders is not important.

My paper complements and enriches the strategic network formation literature in two important ways. First, by observing the continuous evolution of the network, I do not need to rely on assumptions to simulate the current network state. Second, I directly observe

75 each individual’s entire action history in the network, which enables us to explicitly account for the e↵ects of individuals’ time-varying actions in their friendship formation decisions.

The latter improvement is especially important since I study network formation in an online environment. In such an environment, individuals’ behaviors and opinions as observed by other individuals are likely to be among the most important drivers of friendship tie decisions.

3.2.2 Peer E↵ects

Previous research has shown that social networks greatly facilitate information dissemination and product di↵usion (e.g., Duan et al. 2009; Katona et al. 2011; Christakis and Fowler

2013). These networks provide a platform for consumers to produce and publish content, for example, to post in discussions forums, to write product reviews or to re-share content (e.g.,

Toubia and Stephen 2013). Further, past research has also shown that friends inﬂuence each other’s product adoption decisions through reviews and ratings they post online (e.g., Aral and Walker 2011; Ma et al. 2014; Ameri et al. 2016).

Making a casual inference of friends’ inﬂuence, however, is a challenging task (Manski

1993). Multiple social phenomena can confound the identification and inference of social outcomes (Shalizi and Thomas 2011; Hartmann et al. 2008). Among these phenomena, homophily is probably the most challenging one to be accounted for. Homophily refers to the observation that individuals tend to become friends with similar individuals. Due to the similarity among friends, they exhibit the same behavior without one necessarily influencing the other. Di↵erent approaches have been proposed by previous literature to account for these issues including the use of instrumental variables (e.g., Bramoulléet al. 2009;

De Giorgi et al. 2010; Claussen et al. 2014), the incorporation of individual-speciﬁc unobserved tastes/preferences (e.g., Nair et al. 2010; Trusov et al. 2010), controlling for correlated group e↵ects (e.g., Lee 2007; Lee et al. 2010; Ma et al. 2014), the use of exogenous shocks to peers (e.g., Tucker 2008) or exogenous randomness (e.g., Sacerdote 2001), experiments

76 (e.g., Aral and Walker 2011), and network co-evolution models (Snijders et al. 2007; Badev 2013). In my paper, I develop a network co-evolution model that explicitly incorporates individual-specific unobserved preferences to control for homophily. The most notable among the co-evolution models was proposed by (Snijders et al., 2007). (Snijders et al., 2007) propose a stochastic model in which both the network structure and individuals’ actions evolve simultaneously in a dynamic process: individuals are selected at random rates and each selected individual decides whether to make a change in her friendship ties or whether to perform an action of interest or whether to do neither. There are several limitations to (Snijders et al., 2007). First, since individuals cannot change both their ties and their actions at the same point in time, simultaneity between tie formation and other actions is not accounted for. Second, although (Snijders et al., 2007) capture homophily using friendship selection and similarity indices, latent homophily (arising from the similarity among friends in their unobserved intrinsic preferences) remains a confounding factor that may bias the e↵ect of friends’ influence. Third, due to the randomness in the decision timing, the e↵ects of exogenous time-varying factors cannot be identified and, as a result, simultaneity between actions of friends cannot be controlled for. In this paper, I overcome these limitations by proposing a structural model of individuals’ concurrent decisions to both form ties and to perform activities at each point in time. This allows us to capture the e↵ects of any time-varying variable while controlling for the simultaneity across these decisions. Furthermore, I are able to account for the latent homophily by explicitly modeling the choice of friends. In addition, by observing users’ actions before and after they make friends, I estimate individuals’ intrinsic preferences for actions absent of their friends’ influence and therefore provide a better identification of peer e↵ects.

3.2.3 Seeding

Seeding refers to the determination of whom to target for motivational stimulation with the goal of triggering large information cascades, adoptions, or other types of actions. The most

77 commonly studied seeding strategies are based on network metrics such as “in/out-degree centrality”7 or “betweenness centrality.”8 For example, (Hinz et al., 2011) compare three seeding strategies — stimulating high-degree, low-degree, and high betweenness individuals with random seeding — in terms of their e↵ectiveness in increasing adoption. They find that seeding to well-connected individuals is the most successful strategy in increasing participation in viral marketing campaigns. Similarly, (Aral et al., 2013) examine the e↵ectiveness of di↵erent seeding strategies under varying levels of homophily (self-tendency to adopt) and under the influence of friends on adoptions. They examine the e↵ectiveness of seeding to high-degree individuals, dense regions of the network, and hubs unlikely to adopt against the e↵ectiveness of random seeding and find that high-degree and dense region targeting generally perform better. They further show that seeding more than 0.2% of the population is wasteful because the gain from their adoptions is lower than the gain from their natural adoptions. (Katona, 2013) studies seeding strategies in a theoretical framework and shows that highly connected influencers are valuable only if they cover consumers who are not connected to many other influencers. Many of the previous studies have focused on adoption behavior. However, for the growing number of online social platforms, a continuous engagement of its users with the website may be more important than one-time adoption behavior. (Trusov et al., 2010) focus on the activity levels of users within an online platform instead of adoption behavior. The authors develop an approach to determine which of a focal user’s friends have a significant influence on the focal user’s activity level. (Trusov et al., 2010) use their results to examine the e↵ect of any change in the focal user’s behavior on the behavior of those linked to him and find that, on average, only 20% of a focal user’s friends are influencing the focal user.

7In-degree and out-degree centralities refer to the number of incoming and outgoing ties, respectively, of an individual in a network.

8Betweenness centrality of an individual refers to the number of shortest chains of links that connect all pairs in a network and include that individual.

78 Following (Trusov et al., 2010), I add to the existing knowledge on the e↵ectiveness of seeding strategies on user engagement in online social platforms. Going beyond (Trusov et al., 2010), I examine di↵erent seeding strategies that are not only based on network metrics, as is common in the literature, but can also depend on users’ actions on the website. Furthermore, by modeling the co-evolution of friendship formations and users’ actions over time, I not only capture the immediate e↵ect of friends’ inﬂuence on each other, but also capture how that e↵ect propagates into the further development of the network and into future actions of users.

3.3 Data

My data come from MyAnimeList.net. This website is a consumption-related online community (Kozinets 1999) where online interactions are based upon shared enthusiasm for a specific consumption activity. MyAnimeList.net was created to allow anime (Japanese cartoons) fans to gather and to share their excitement and opinions about animes. The website has developed into one of the most popular platforms for anime fans over the years. Users of the website create a profile page when they join the website. On their profile page, users can share some information about themselves (e.g., age, gender, or location) and create a list of the animes they have watched or are watching (which I refer to as “watch list” throughout this paper). The website also provides a forum where users can share information and exchange opinions about animes with each other. In addition, users have the option to become friends, which makes it easier for them to access their friends’ pages and to be notified about their friends’ activities, similar to bookmarking and to RSS functions in web browsers. Anime fandom is a special interest and not very common. As a result, fans use special interest communities such as MyAnimeList.net to connect with other fans. This implies that most users in my data meet their friends for the first time on the website under study and their interactions are happening within the website. Furthermore, this website is a worldwide

79 is,ueswojie h est eoeMrh20 r ieyt d ahohra friends as other each add to likely are 2006 March before website the joined who users First, end the by 11,500 and 2007 July of beginning the 2007. at members of 2,700 450 to about had rapidly website grew the number function time, this the in 2007, and point 16, that March At on added. members later, was of year friendships number a forming the About of domain, the 3.1). public time, Figure a (see in quickly to point grow transfer that to its At started After shape. users. current 300 its was take about it had to 2006, website started 6, April and On domain public domain. private a a to as moved however, 2004, in started first was website The Sample the Estimation to confined 3.3.1 mostly are users the among interactions and platform. validates meetings further frequently observation users that This that see assumption countries. can erent ↵ di my I from users pages. other profile with About their friends on globe. become locations the their reveal around users countries the and of cities half erent ↵ di from users attracts and community Ifocusonuserswhojoinedthewebsiteinthesecondhalfof2007,mainlyfortworeasons.

iue31 ae sr ondMyAnimeList.Net Joined Users Dates 3.1: Figure Density 01jan2004 0 .001 .002 .003 01jan2006 01jan2008 01jan2010 80 Join Date 01jan2012 01jan2014 01jan2016 based on past interactions. To put it di↵erently, had they had the option of adding friends before, they would have done so. And second, it might have taken existing members some time to learn about this new feature. Therefore, I start my study period about three months after the introduction of the friendship feature.

Studying daily friendship formation among all the users who joined between July and

December 2007 is, however, computationally impossible since the data set would include over

7billionpair-dayobservations.Onepotentialsolutionistoshortentheobservationperiod.

However, this approach would result in insucient variation in the dependent variables.

Figure 3.2 shows the distributions of the number of days in between activities of each type.

In about 50% of the cases, users add a friend and publish a post more than a month after their last action in the corresponding area. In 40% of the cases, users watch an anime more than a month after the last watched anime.

A second potential solution is to aggregate the observations to the weekly level. However, aggregation of observations leads to information loss on the dependent variables and the sequence of actions. On top of that, I observe that in more than 20% of the cases, users make more than one friend in a week. Anime watching and content generation also happen more than once a week in about 25% and 10% of the cases, respectively. Consequently, aggregating data to the weekly level would force us to model the sequence of users’ actions within a week. As a result, similar to previous studies on strategic network formation

(Christakis et al. 2010, Snijders et al. 2010), the degree to which I could capture the e↵ects of time-varying activities of users on their friendship making would be restricted.

Athirdpotentialsolutionistosamplefromthenetwork.Iimplementthissolutionusing snowball sampling, which is the common sampling method in the network literature. Figure

3.3 visualizes my sampling strategy. First, I draw a random sample of 400 users (“core users”) out of about 8,800 users who joined the website in the second half of 2007, and then include all of their friends in my estimation sample. Note that the friends of the 400 core

81 (a) Friend Addition (b) Anime Watching 2 6 1.5 4 1 Percentage of Cases Percentage of Cases 2 .5 0 0 0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180 Number of Days Between Adding Friends Number of Days Between Watching Animes

(c) Content Generation 8 6 4 Percentage of Cases 2 0 0 20 40 60 80 100 120 140 160 180 Number of Days Between Writing Posts

Figure 3.2: Number of Days Between Activities users can also be core users or they can be users not in the core. For example, in Figure 3.3, users 1 and 2 are both core users. User 2 is friends with user 1 who is another core user and with users 3 and 4 who are not core users. I term all users who are not core users themselves, but friends with a core user “non-core users.” This second set of users includes 986 users.9

Thus, my estimation sample contains 1,386 users (400 core users and 986 non-core users).

9The set of non-core users includes 732 users who joined after July 2007 and the 254 users joined before July 2007.

82 Figure 3.3: User Sampling Strategy

In the estimation, I model all anime adoptions and all UGC production activities for both core and non-core users. For friendship formation, I model all potential ties among core users (e.g., between users 1 and 2 in Figure 3.3), all potential ties among core and non-core users (e.g., between users 1 and 3 in Figure 3.3), and all potential ties among non-core users (e.g., between users 3 and 4 in Figure 3.3). However, I assume that non- core users’ friendship formations with outside users, i.e., users who are neither core nor non-core users, are exogenous. This exogeneity assumption means that I do not model them becoming friends, but I do take their friendship into account when creating friendship related independent variables.

Given that activities of users in the three areas can be correlated, missing a portion of the network formation for these 986 non-core users can lead to bias in my estimates of the friendship decisions. Note that since I model all actions of core and non-core users and incorporate the e↵ects of friendships with outside users on non-core users’ anime adoptions and UGC production, this bias is mainly a concern for the estimation of the friendship decision. To alleviate this concern, I estimate separate coecients for the 400 core users (for

83 whom I have their complete tie formations) and for the 986 non-core users (for whom I do not model the portion of the friendship network that includes users outside of my sample).

Aconcernwithsnowballsamplingistheoversamplingofactiveusers.Thisconcernis alleviated by controlling for unobserved heterogeneity among users. Furthermore, I estimate separate coecients for the core and non-core users in the friendship utility. And lastly, since I draw a random sample of users and include the friendship network of those users in my sample, for a focal user, the other 399 randomly drawn users and their friends who are not friends with that focal user are a random representative sample of the whole network.

As a result of this randomness and the incorporation of separate parameters, I believe that my estimates are unbiased for the core 400 users.

The observation period is 184 days between July 1 and December 31, 2007. However, I have fewer observations for users who joined after July 1, 2007. On average, I observe each user for 140 days.

3.3.2 Data Description

Within my sample of 1,386 users, I observe 5,038 ties out of 947,155 possible ties being formed during the observation period and about 68 million daily observations of possible pairs. Figure 3.4 shows the states of the network for snapshots at days 1, 60, 120, and 184.

The nodes represent individual users in the network, and the links between nodes represent friendships ties. Furthermore, the color of a node reﬂects the quantity of a user’s UGC production and the size of a node reﬂects the number of animes a user watched. The color of the nodes becomes darker as users publish more posts on the website and the size of the nodes increases as users watch more animes. As expected, the nodes become darker, bigger, and more connected over time. The larger and darker nodes are also associated with more links, suggesting the interdependence between users’ friendship formation and other activities.

84 (a) Network - Day 1 (b) Network - Day 60

Figure 3.4: Network Co-Evolution Over Time Lines Between Nodes Indicate Friendship Ties. Node Size Increases with More Animes Watched. Node Color Darkens with More Posts Written.

85 Table 3.1: Descriptive Statistics

Mean Std. Dev. Min Median Max N Age 19.29 5.20 12 18.5 78 1088 Gender (% Females) 38.5 Gender (% Males) 54 Gender(%NotSpeciﬁed) 7.5

Core Users: NumberofActiveDays 12.62 15.43 0 7 101 400 NumberofFriendAddingDays 2.55 4.69 0 1 59 400 NumberofAnimeWatchingDays 9.19 10.37 0 5 68 400 NumberofContentGeneratingDays 2.47 9.17 0 0 95 400

PercentageofActiveDays 17.76 17.07 0 12.90 100 400 PercentageofFriendAddingDays 3.80 6.19 0 1.63 40 400 PercentageofAnimeWatchingDays 13.55 14.11 0 9.70 85.71 400 PercentageofContentGeneratingDays 2.56 7.72 0 0 61.69 400

FriendAddingIntervalinDays 44.26 36.05 1 34 156 12,414 AnimeAddingIntervalinDays 25.64 28.20 1 14 138 20,680 PostAddingIntervalinDays 32.07 29.33 1 25 109 6,330

Non-Core Users: NumberofActiveDays 24.06 24.42 1 16 181 986 NumberofFriendAddingDays 5.93 6.40 1 4 48 986 NumberofAnimeWatchingDays 15.55 16.65 0 10 121 986 NumberofContentGeneratingDays 6.31 17.57 0 0 181 986

PercentageofFriendAddingDays 22.49 18.36 .54 17.95 100 986 PercentageofAnimeWatchingDays 6.85 8.31 .54 4.23 80 986 PercentageofContentGeneratingDays 14.46 13.38 0 10.87 82.22 986 PercentageofContentGeneratingDays 4.72 12.01 0 0 98.37 986

FriendAddingIntervalinDays 48.41 40.36 1 37 181 78,996 AnimeAddingIntervalinDays 24.82 29.86 1 13 163 84,133 PostAddingIntervalinDays 47.98 48.96 1 28 182 40,287

Table 3.1 summarizes key statistics of my data. In terms of demographics, 78% of users report their age and are, on average, 19 years old and 94% of users report their gender with

39% being female and 53% being male.

Figure 3.5 demonstrates how activities of users who joined the website in second half of

2007 change over time from the day they joined the website. Figures 3.5a shows a decreasing trend in making new friendship ties. Since one of the beneﬁts of having friends is a cost

86 (a) Friend Addition (b) Anime Watching .8 10 8 .6 6 .4 4 .2 2 Average Number of Friends Added Average Number of Animes Watched 0 0 0 30 60 90 120 150 0 30 60 90 120 150 Days of Membership Days of Membership

0 30 60 90 120 150 Days of Membership

Figure 3.5: Average Activity Levels Over Time Since Joining (New Users) reduction in learning about the website and new animes, users are more likely to add friends shortly after they join the website. Figures 3.5b and 3.5c show the activity trends for anime watching and content generating. Both graphs reveal a rather constant trend over time.10

Users can engage in multiple types of activities simultaneously. On average, my core users have 2.6 active days in terms of friend adding, 9 active days in terms of anime watching, and 2.5 active days in terms of post writing (see Table 3.1). In total, they have 12.6 days

10Note that the high number of animes shortly after joining is mainly due to users adding animes that they watched before joining the website to their watch lists.

87 Figure 3.6: Percentage of Observations with Certain Activities Conditional on Performing at least one Activity in which they participate in at least one type of activity. To put it di↵erently, on average, core users are active on about 18% of the days during the study period. Figure 3.6 shows a

Venn diagram of the joint probabilities of each type of activity conditional on engaging in at least one type of activity. Users are active in only one area in 85.81% of the cases. Users are active in two and three of the areas of interest in 13.36% and 0.83% of cases, respectively.

I observe similar a pattern for non-core users albeit with generally higher average activity amounts (see lower half of Table 3.1).

Lastly, Figure 3.7 shows histograms of individual users’ daily activity intensities conditional on them being active. In more than 80% of the cases, users add only one friend on an active day. Similarly, in about 60% of the cases, users watch only one anime per active day. However, the content generation intensity is higher: users publish one post per active day in about half of the cases and publish 2 or 3 posts per active day in about 20% and 10% of the cases, respectively. Based on this data pattern, I make the simplifying assumption of modeling the anime watching and content generation as binary indicator variables, i.e., I

88 (a) Number of Friends Added in a Day (b) Number of Animes Watched in a (Truncated at 100) Day (Truncated at 100) 60 80 60 40 40 Percentage of Cases Percentage of Cases 20 20 0 0 0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35 Friends in A Day Animes in A Day

Figure 3.7: Number of Activities in Each Area Per Day model whether a user watches an anime or makes a post, but not the number of watched animes or posts.11

3.4 Model

In this section, I describe how I jointly model a user’s decisions to form ties, adopt products, and generate UGC. For friendship formations, I model whether and with whom users become

11Since I model the decision of a user to become friends with each of the other users as separate independent decisions, even if users make more than one friend in a day, my model captures that.

89 friends, while for the activities in the other two areas, namely, anime adoptions and content generation, I only model users’ decisions to participate in an activity or not.

3.4.1 Tie Formation

IstartbydescribinghowImodeltieformationamongusersovertime.Ineachtimeperiod

(day), a user makes decisions whether to become friends with any other user with whom she is not friends yet. Since there are usually many users with whom the individual is not friends yet, at each point in time, a user can become friends with multiple users. Note that

I model a user’s tie formation decisions for each possible friendship pair in each time period and not whether a user makes a friend or not in a time period.

Suppose the website contains i =1,...,N users and these users can become friends with other users during t =1,...,T time periods.12 Let M denote the adjacency matrix of the network which shows the stati of ties between each pair of individuals i and j with j =1,...,N and i = j. m equals 1 if i and j are friends at time t and 0 otherwise. Ties 6 ij,t in the network are bi-directional, i.e., mij,t = mji,t.Furthermore,bothusershavetoagree to become friends. In my data, I do not observe users’ “requests” for friendship with other users, only the formation of ties upon mutual agreement. Therefore, my model describes the decision of both users to become friends regardless of who ﬁrst requested the friendship.13

The decision of two users to become friends depends on the utilities both individuals

m derive from becoming friends (see e.g., Christakis et al. 2010). Let us deﬁne Ui,t as individual

m i’s utility of becoming friends with individual j in time period t and Uj,t as individual j’s

12Note that t is the calendar day and not the day since a user joined the website.

13I do not model the dissolution of friendship ties, i.e., once users become friends, they stay friends. This is due to a limitation of my data: if two users “unfriend” each other, they appear as non-friends. I do not believe unfriending is a common action among users and given that the observation period is 184 days, I view not modeling friendship dissolution as a minor limitation.

90 utility of becoming friends with individual i in time t. User i’s utility function is given by

m m m m Ui,t = f(A , R ,✏ ). (3.1)

The utility individual i derives from forming a tie with individual j depends on the attractiveness of user j as judged by her past actions, Am,andthesimilaritybetweenuser i and user j, Rm. ✏m captures the part of the utility of user i at time t that is observed by the user but not by the researcher.14

Iassumethatindividualshavemyopicutilities,i.e.,individualsdonotanticipatefuture states of the network and only care about the current state of the network when deciding to form a tie. They do not take future links of themselves or the other party into consideration when making the decision to become friends. The assumption of myopic utility is appropriate for large networks in which individuals can meet numerous other individuals at each point in time and the number of future states of the network increases exponentially. Furthermore, users are not limited in the number of ties they can make in an online friendship network.

As a result, users independently and non-strategically form ties if the utility of such ties is positive. This ensures that the network formation will have a unique equilibrium.15

Imodelthetieformationbetweentwousersasanon-cooperativedecision,i.e.,each pair’s decision to become friends only depends on the observed pair-speciﬁc variables and decisions of di↵erent pairs of users are conditionally independent. A tie between i and j is formed if and only if both parties decide that it is beneﬁcial for them to become friends, i.e.,

m m mij,t =1 i↵ Ui,t > 0andUj,t > 0. (3.2)

14Throughout the paper, I use the superscript m to refer to variables associated with friendship utility.

15I refer the interested reader to (Jackson, 2008) for an extensive discussion of equilibria in networks.

91 3.4.2 Product Adoption and Content Generation

Next, I describe how I model a user’s activities on the website. I study users’ activities in two broad areas, namely, product adoption and content generation. Let Ak ,k Product i,t 2{ Adoption (pa), Content Generation (cg) deﬁne user i’s activity at time t.Ifuseri adopts } pa an anime at time t,thenAi,t equals 1 and 0 otherwise. If user i posts on the website at time cg t,thenAi,t equals 1 and 0 otherwise. pa cg User i’s utilities from watching animes, Ui,t ,andproducingcontent,Ui,t ,aregivenby

pa pa pa pa Ui,t = g(A ,F ,✏ ) (3.3) cg cg cg cg Ui,t = h(A ,F ,✏ ), where the utilities depend on a user’s past actions, Apa and Acg,aswellasontheinﬂuence of a user’s friendship network, F pa and F cg. ✏pa and ✏cg describe the part of the utility that is observed by the user but not by the researcher.

3.4.3 Integrating All Actions

I now present the full model, integrating user i’s actions in all three areas.

U m = f(Am, Rm,✏m) j =1...N, i= j i,t 8 6 pa pa pa pa Ui,t = g(A , F ,✏ ) (3.4)

cg cg cg cg Ui,t = h(A , F ,✏ ).

Some variables, unobserved by the researcher, might inﬂuence all three types of decisions ausermakes.16 For example, having watched an anime, a user might be excited to discuss the anime with other members in the forum section of the website. To accommodate simultaneity

16If I were to assume that the decision a user makes regarding one action is independent of the user’s decision regarding actions in the other areas, each of the decisions in the three areas could be estimated separately.

92 among these three decisions, I allow the three error terms in Equation (3.4) to be correlated, i.e., 2 m ⇢m,pa ⇢m,cg

2 2 3 ⌃= ⇢m,pa pa ⇢pa,cg . (3.5) 6 7 6 2 7 6⇢m,cg ⇢pa,cg cg 7 6 7 As is well known, I need to set one4 element in the covariance5 matrix to 1 for identiﬁcation reasons (McCulloch and Rossi 1994). Thus, I normalize ⌃11,i.e.,thevarianceoftheerror term in the friendship utility, to 1.

3.4.4 Utility Speciﬁcations

Next, I present details of the utility speciﬁcations for the speciﬁc context of my data. I model the utility of user i forming a tie with user j as

mij m m m m m m m m Ui,t = ↵i + Aj,t 1 + Rij,t 1 + Cit + ✏i,t, where

t 1 t 1 m pa cg (3.6) Aj,t 1 = mij,t 1 + Aj,⌧ + Aj,⌧ i=1 ⌧=0 ⌧=0 Xi=j X X 6 m m pa D Rij,t 1 = Rij,t 1 + Rij,t 1 + Rij . m ↵i captures user i’s intrinsic preference for making friends and follows a normal distri-

m m bution with mean↵ ¯ and standard deviation ↵m .ThevariableAj,t 1 describes user j’s attractiveness as a potential friend. It only depends on j’s attributes and reﬂects how popu-

m lar and/or knowledgeable user j is. I operationalize Aj,t 1 as the number of user j’s actions t 1 pa in the three areas, i.e., j’s number of friends, mij,t 1,numberofadoptedanimes, Aj,⌧ , i=1,i=j ⌧=0 6 t 1 P cg P and number of posts in the UGC part of the website, Aj,⌧ .Theﬁrstvariabledescribes ⌧=0 the utility gained from becoming friends with popular usersP (Tong et al. 2008; Langlois et al.

2000), while the latter two represent the utility gained from information sharing and learning

93 from friends who are knowledgeable about animes (Watson and Johnson 1972; Brandtzæg and Heim 2009).

m The variable Rij,t 1 is tie-speciﬁc and captures the similarity between individual i and m m individual j. Rij,t 1 includes the number of common friends, Rij,t 1,thenumberofcommon pa animes, Rij,t 1,attimet 1, and demographic similarity between user i and user j in D terms of age, gender, and country of origin, Rij .Providingsuchdemographicinformationis optional for users. However, the presence of such information may signal honesty and thus increase the credibility and perceived utility gained from forming a friendship (Lampe et al.

D 2007). Rij additionally includes three dummy variables that indicate whether age, gender, or location information of both individual i and individual j are available.17

m Cit contains several variables whose e↵ects I control for. The first variable is the length of time (in days) since user i joined the website. As revealed in Figure 3.5, new users are more likely to add friends compared to experienced users since having friends in the beginning helps to reduce learning costs associated with navigating the website. Second, I also include weekend dummies. Third, I also include a dummy variable indicating whether use j was active on the platform during the previous week. This variable captures user j’s visibility.18 Fourth, I include time fixed e↵ects to address common shocks that might result in simultaneity of friendship actions across users. I operationalize these time fixed e↵ects as

19 m week dummies. And lastly, I assume that ✏i,t follows normal distribution.

17To address the potential bias in the estimation results due to not modeling the formation of the full m m D network of non-core users, I further estimate separate coecients of Aj,t 1, Rij,t 1, and Rij for core and non-core users.

18 m m In addition to user i’s preference for friendship with user j, both Aj,t 1 and Rij,t 1 also capture the degree to which user j is visible to user i. Unlike previous studies in the strategic network formation literature (e.g., Christakis et al. 2010; Snijders et al. 2010) which model the meeting and the decision to become friends separately, I follow the conventional approach in the choice model literature and model the combined e↵ect of visibility and preference in the utility.

19While it would be desirable to include daily dummy variables, for computational reasons (see Section 3.5), I are not able to do so as the number of additional parameters to be estimated (552 = 184 days 3 activities) would be too large and the estimation would take many months to converge. ⇥

94 pa User i’s utility from watching an anime, Ui,t ,isgivenby

pa pa pa pa pa pa pa pa pa Ui,t = ↵i + Ai,t 1 + Fi,t 1 + Cit + ✏i,t, where

pa pa (3.7) Ai,t 1 = Ai,t 1 pa pa cg Fi,t 1 = mij,t 1 + Aj,t 1 + Aj,t 1, j=1 j mij,t 1=1 j mij,t 1=1 Xj=i 2{ X } 2{ X } 6

pa where ↵i represents user i’s intrinsic tendency to watch animes and is assumed to follow a

pa pa normal distribution with mean↵ ¯ and standard deviation ↵pa . Ai,t 1 captures the state dependence of anime watching and is operationalized as a dummy variable which equals 1 if

pa user i watched an anime at t 1 and 0 otherwise. Next, Fi,t 1 captures the e↵ects of user i’s friendship network on user i’s actions. It includes user i’s number of friends, mij,t 1, j=1,j=i pa 6 the number of animes watched by all of user i’s friends at t 1, Aj,t 1,andthenumberofP j mij,t 1=1 cg 2{ } posts written by all of i’s friends at t 1, Aj,t 1.PreviousliteraturehasshownthathavingP j mij,t 1=1 2{ } more friends might directly increase theP level of social activities of network users (Toubia and Stephen 2013; Shriver et al. 2013).20 In addition, the number of animes watched by all of user i’s friends captures the direct inﬂuence of friends’ activities on user i’s product adoptions, while the number of posts written by all of user i’s friends reﬂects the spill-over e↵ect of friends’ activities in post publishing over user i’s activity in anime watching.

pa Furthermore, Cit contains several variables whose e↵ects I control for. It includes a

pa weekend dummy and week ﬁxed e↵ects. And lastly, I assume that ✏i,t is normally distributed.

cg Similarly, user i’s utility from writing a post, Ui,t ,isgivenby

20A potential explanation for this e↵ect is the image or prestige utility users gain from performing social activities within a network. (Toubia and Stephen, 2013) ﬁnd that, aside from the intrinsic utility users derived from posting on social media, the image these activities create for users also motivated them to perform these activities. They also found that image-related utility was more dominant for users with more friends.

95 cg cg cg cg cg cg cg cg cg Ui,t = ↵i + Ai,t 1 + Fi,t 1 + Cit + ✏i,t, where

t 1 cg cg pa (3.8) Ai,t 1 = Ai,t 1 + Ai,⌧ ⌧=0 X cg pa cg Fi,t 1 = mij,t 1 + Aj,t 1 + Aj,t 1, j=1 j mij,t 1=1 j mij,t 1=1 Xj=i 2{ X } 2{ X } 6 pa where ↵i is user i’s intrinsic tendency to produce content and follows a normal distribu-

cg cg tion with mean↵ ¯ and standard deviation ↵cg . Ai,t 1 represents user i’s past actions and cg contains two variables. The ﬁrst variable, Ai,t 1,capturesstatedependenceandisopera- tionalized as a dummy variable indicating whether user i wrote a post at t 1. In addition, user i’s past anime watching behavior may also inﬂuence her posting decision because a user who watches more animes may have more things to talk about. Therefore, I also include the

t 1 pa cumulative number of animes watched by user i,⌃⌧=0Ai,⌧ 1,asacovariateinEquation3.8. cg The variable Fi,t 1 describes the e↵ects of user i’s friendship network on user i’s actions and pa is deﬁned in a similar manner as Fi,t 1 in Equation (7), i.e., it includes user i’s number of pa friends, mij,t 1,thenumberofanimeswatchedbyallofi’s friends at t 1, Aj,t 1,and j=1 j mij,t 1=1 j=i 2{ } P6 cg P the number of posts written by all of user i’s friends at t 1, Aj,t 1. As in the previous j mij,t 1=1 cg 2{ } cg equation, Cit contains a weekend dummy and week ﬁxed e↵ects.P And lastly, ✏i,t follows a normal distribution.21

3.5 Estimation

Given the conditional independence assumption of user i’s decision to become friends with each user j (as discussed in Section 3.4.1) and given the need for mutual agreement to become friends, the probability of a tie forming between individual i and individual j is given by

21All continuous variables in the three utility functions are incorporated in the form of natural logarithms.

96 P (m =1)=P (U m > 0) P (U m > 0). (3.9) ij,t i,t · j,t Then the likelihood of user i becoming friends with user j at time t is given by

m m 1 m 1 mij,t 1 ij,t ij,t Lij,t ↵ ,↵ ,✏ ,✏ = [P (mij,t =1)] [1 P (mij,t =1)] , (3.10) | i j i j m pa cg⇥ ⇤m where ↵i = ↵i ,↵i ,↵i and ↵j is deﬁned similarly. Note that Lij,t ↵ ,↵ ,✏ ,✏ conditions { } | i j i j on the two users not being friends before time t through the exponent 1 mij,t 1.Given the above equation, the likelihood of all friendship formations of user i in time period t is denoted by

N m m 1 m 1 mij,t 1 ij,t ij,t Li,t ↵,✏ = [P (mij,t =1)] [1 P (mij,t =1)] i = j | 6 j=1 (3.11) Y ⇥ ⇤

where ↵ = ↵ ,...,↵ and ✏ = ✏ ,...,✏ . { 1 N } { 1 N } The likelihoods for the other two types of activities, i.e., product adoption and content generation, at time t are given by

pa pa Apa pa 1 Apa i,t i,t Li,t ,↵ ,✏ =[P (Ai,t =1)] [1 P (Ai,t =1)] | i i (3.12) cg cg Acg cg 1 Acg i,t i,t Li,t ,↵ ,✏ =[P (Ai,t =1)] [1 P (Ai,t =1)] . | i i Combining equations (3.11) and (3.12), the joint likelihood of user i’s actions at time t is denoted by

+ 1 pa pa pa Ai,t pa 1 Ai,t Li,t ↵ = [P (Ai,t =1)] [1 P (Ai,t =1)] | Z1 cg Acg cg 1 Acg [P (A =1)] i,t [1 P (A =1)] i,t · i,t i,t (3.13) N m 1 m 1 mij,t 1 (P (m =1)] ij,t [1 P (m =1)] ij,t d✏i = j. · ij,t ij,t 6 j=1 Y ⇥ ⇤ 97 The full likelihood can be written as

+ + T N 1 1 cg Apa pa 1 Apa L = [P (A =1)] i,t [1 P (A =1)] i,t i,t i,t t=1 i=1 Z1 Z1 Y Y cg Acg cg 1 Acg [P (A =1)] i,t [1 P (A =1)] i,t (3.14) · i,t i,t N m 1 m 1 mij,t 1 [P (m =1)] ij,t [1 P (m =1)] ij,t d✏d↵ · ij,t ij,t j=i+1 Y ⇥ ⇤ and the log likelihood of the model given by

+ + T N 1 1 cg Apa pa 1 Apa LL =log [P (A =1)] i,t [1 P (A =1)] i,t i,t i,t t=1 i=1 Z1 Z1 Y Y cg Acg cg 1 Acg [P (A =1)] i,t [1 P (A =1)] i,t · i,t i,t N m 1 m 1 mij,t 1 [P (m =1)] ij,t [1 P (m =1)] ij,t d✏d↵. · ij,t ij,t j=i+1 Y ⇥ ⇤ (3.15)

IestimatemymodelusingSimulatedMaximumLikelihood(SMLE).Toestimatethefull covariance matrix of user random e↵ects, I take 30 draws from a standard normal distribution for each user and each activity and use the Cholesky decomposition of the covariance matrix.

Similarly, to estimate the full covariance matrix of the three error terms, I take 30 draws from a standard normal distribution for each user and each activity in each time period and use the Cholesky decomposition of the error covariance matrix in the estimation.

For computational reasons, the conventional approach to estimate a model via MLE and

SMLE involves taking the logarithm of the model likelihood in order to convert an extremely- small-in-value product of probabilities to a sum of the logarithms of these probabilities. This approach cannot be applied to the likelihood of my model for three reasons. First, recall that, at any time t,theerrortermsinthethreeutilityfunctionsarecorrelated(seeEquation

98 (5)).22 Therefore the integral taken over ⌃has to include user i’s likelihood of all three activities at time t.Second,recallthattheprobabilityofafriendshipformationdependson both user i’s and user j’s utilities for the tie formation, i.e., a friendship is only formed if both users derive positive utilities from doing so (see Equation (2)). Since at each time t,allusers can become friends with any other user with whom they are not friends yet, all friendship formation decisions of all users at time t are connected through the error terms in users’ friendship formation utilities. In other words, due to the second reason, the integral over ⌃ has to to include all friendship formation probabilities of all users at time t.Combiningthe

ﬁrst and second reason, it is evident that the integral over ⌃has to include the probabilities of all actions of all users at time t.

Third, recall that my model includes time-invariant individual-speciﬁc preferences for

m pa cg each type of activity, i.e., ↵i , ↵i ,and↵i .Therefore,foreachuserandeachtypeof activity, the integral over ↵ has to include all activities of that type over all time periods.

Given that the first two reasons necessitate that the integral over ⌃contains the probabilities of all actions of all users at each time t and given that the third reason necessitates that the integral over ↵ contains all probabilities over all time period for a specific type of activity and a specific user, the integrals over both ⌃and ↵ have to contain the probabilities of all actions of all users over all time periods (see Equation (14)). As a result of these three issues, when I take the logarithm of the model likelihood, I cannot convert the product of the probabilities into a sum of the logarithms of these probabilities (see Equation (15)).

This poses a problem for common computing technologies since the likelihood is the product of a very large number of probabilities and too small in magnitude to be detected.23 To make the likelihood estimation computationally tractable, I use a transformation of the

22 m pa Another reason is that the individual-speciﬁc intrinsic propensities for each type of activity, ↵i , ↵i , cg and ↵i , are correlated as well.

23For the interested reader, the likelihood is given by the product of over 136,000,000 probabilities.

99 logarithm of sum of variables to a function of the logarithm of those variables. Details of the transformation are presented in Appendix G.

To speed up the estimation, I use OpenBLAS as the system BLAS (Basic Linear Algebra

Subprograms), tensorization of large matrices and parallel computing methods to run the estimation program. Due to the large size of the data and parallelization, I cannot run the estimation code on conventional computing systems.24 Iutilizeseverallargememorysuper- computing servers including the Texas Advanced Computing Center (TACC) and Jetstream cloud-computing (Stewart et al. 2015; Towns et al. 2014).25

3.6 Identiﬁcation

The set of parameters to be estimated is given by ↵¯m, ↵¯pa, ↵¯cg, ⌃↵,m,pa,cg,m,pa,cg, { m,pa,cg, ⌃ . The identification of m,pa,cg,m,pa,cg is standard. } { } The mean intrinsic propensities, ↵¯m, ↵¯pa, ↵¯cg , are identified by the average user behavior { } across users and across times. The covariance matrix of the user random e↵ects, ⌃↵,is identified by the variation in average activity levels across users. In contrast, the covariance matrix of the error terms, ⌃, is identified by the variation in the co-occurance of activities on the same day.

The parameter m captures the e↵ects of common factors for each pair of users and is identified by the variation in the percentage of common friends and the percentage of common animes among di↵erent pairs. The parameters pa,cg capture friends’ influence { } on a user’s actions and are identified by changes in average behavior of friends. Lastly, conditional on ⌃, the three utilities are separately identified since each action is modeled as afunctionofotheractionsintheprevioustimeperiod.

24The estimation code requires at least 350GB of RAM.

25It takes more than 3 weeks to estimate a model with 49 parameters on a super computer utilizing 32 CPU cores using my data containing 68 million observations.

100 Separating homophily from influence is a challenging task (Manski 1993). Recall that homophily refers to friends behaving in a similar manner due to their similar preferences and not because of one influencing the other. The similarity in unobserved preferences, if unaccounted for, can lead to correlated errors which, in turn, lead to upward biased estimates of friends’ influence. I address this issue by incorporating unobserved time-invariant

m pa cg components, ↵i , ↵i ,and↵i ,inauser’sdecisionstoformties,toadoptanimes,andto generate content (similar approach as in Nair et al. 2010 and Trusov et al. 2010). Since I model the incidence of users’ actions and not the speciﬁc taste for which product to adopt or what type of content to generate, homophily only plays a role in the frequency level of users’ actions, i.e., whether they perform an action on each day. For example, two friends are similar to each other if both tend to publish a lot of posts. In my model, this unobserved

m pa heterogeneity in the propensity to perform each of the three actions is captured by ↵i , ↵i ,

cg m pa cg and ↵i .Furthermore,↵i , ↵i ,and↵i are assumed to be time-invariant since levels of homophily are unlikely to change during the relatively short time span of my observation period. Moreover, since many of the users are new to the network, the latent propensity is identiﬁed not only by the variation in behavior after any friendship formation, but also by behavior before any ties are formed, i.e., when friends’ inﬂuence is absent. And lastly, to capture correlation among a user’s intrinsic propensities to perform the three types of activities (i.e., make friends, watch animes, and produce UGC), I allow for correlations among

m pa cg ↵i , ↵i ,and↵i .

3.7 Results

I present the estimation results in Table 2.5. As discussed in Section 3.3.1, I estimate separate coecients for core and non-core users. The estimation results for non-core users are also presented in a separate column in Table 2.5. In the following, I focus on discussing the results for the core users. Column (i) in the Table 2.5 contains the results for a model in which the

101 decisions about the three types of actions of making friends, watching animes, and publishing posts are made independently of each other. Column (ii) presents the parameter estimates for a model in which I allow these three decisions to be correlated, but there is no unobserved heterogeneity among users. Lastly, column (iii) depicts the results for my full model in which Iallowforbothcorrelatederrorsandunobservedheterogeneityamongconsumers.Overall, the results across the three di↵erent specifications are consistent. Potential simultaneity among the three types of actions a user might engage in each day is captured through the correlated errors. I find significant correlations among two of the three actions.26 Ialso find significant coecients for the std. deviations of the individual-specific random e↵ects suggesting the presence of heterogeneity in intrinsic propensities across users. However, the correlations among the random e↵ects are insignificant. Ifirstdiscusstheparameterestimatesforthefriendshipformationutilityforthe400 core users in my sample. User j’s number of friends, j’s number of watched animes, and j’s number of written posts represent the attractiveness of user j as a potential friend for user i. I find a significant positive e↵ect of the number of friends user j has indicating that users gain utility from becoming friends with well-connected users. This finding stands in contrast to the findings in (Christakis et al., 2010) who find that students are less likely to become friends with popular students. A potential explanation for this result might be the unique context of the online environment. Next, I find that user j’s cumulative number of watched animes and his cumulative number of posts have significant positive e↵ects on friendship formation. This result is consistent with the notion that domain knowledge and information sharing are the primary incentives for friendship tie formations in my specific empirical context. Common friends and common animes capture the similarity between two users. As expected, I find a significant positive e↵ect for common friends implying that users are

26The correlation between product adoption and UGC production is insigniﬁcant.

102 Table 3.2: Results

(i) (ii) (iii) Independent Homogenous MainModel Core Non-Core Core Non-Core Core Non-Core

Friendship Formation

Attractiveness a j’s Number of Friends by t 1 0.426⇤⇤⇤ 0.097⇤⇤⇤ 0.426⇤⇤⇤ 0.097⇤⇤⇤ 0.426⇤⇤⇤ 0.098⇤⇤⇤ (0.010) (0.010) (0.010) (0.010) (0.009) (0.009) a j’s Number of Watched Animes by t 1 0.008 -0.036⇤⇤⇤ 0.007 -0.037⇤⇤⇤ 0.007⇤⇤⇤ -0.037 (0.010) (0.010) (0.011) (0.011) (0.001) (.) a j’s Number of Written Posts by t 1 0.060⇤⇤⇤ -0.046⇤⇤⇤ 0.061⇤⇤⇤ -0.045⇤⇤⇤ 0.062⇤⇤⇤ -0.045⇤⇤⇤ (0.010) (0.011) (0.010) (0.011) (0.010) (0.010) Similarity Number of Friends in Common with j 0.106⇤⇤⇤ -0.172⇤⇤⇤ 0.106⇤⇤⇤ -0.173⇤⇤⇤ 0.106⇤⇤⇤ -0.173⇤⇤⇤ by t 1a (0.023) (0.026) (0.023) (0.026) (0.023) (0.026) Number of Animes in Common with j -0.008 0.006 -0.008 0.008 -0.008 0.008 by t 1a (0.012) (0.014) (0.013) (0.014) (0.010) (0.010) Dummy for Whether Both i and j -0.577⇤⇤⇤ 0.516⇤⇤⇤ -0.577⇤⇤⇤ 0.517⇤⇤⇤ -0.577⇤⇤⇤ 0.517⇤⇤⇤ Indicate Their Country (0.085) (0.087) (0.084) (0.087) (0.080) (0.082) Dummy for Whether i and j Are 0.180⇤⇤⇤ -0.096 0.180⇤⇤⇤ -0.097 0.180⇤⇤⇤ -0.097 from Same Country (0.051) (0.054) (0.051) (0.054) (0.048) (0.050) Dummy for Whether Both i and j -0.327⇤⇤⇤ 0.216⇤⇤⇤ -0.327⇤⇤⇤ 0.216⇤⇤⇤ -0.327⇤⇤⇤ 0.216⇤⇤⇤ IndicateTheirAge (0.040) (0.045) (0.041) (0.045) (0.039) (0.044) Dummy for Whether i and j Are 0.221⇤⇤⇤ -0.047 0.221⇤⇤⇤ -0.047 0.223⇤⇤⇤ -0.047 Within5yearsofAge (0.033) (0.038) (0.033) (0.038) (0.032) (0.037) Dummy for Whether Both i and j -0.168⇤⇤⇤ 0.085 -0.167⇤⇤⇤ 0.085 -0.167⇤⇤⇤ 0.086 Indicate Their Gender (0.047) (0.051) (0.048) (0.051) (0.046) (0.050) Dummy for Whether i and j 0.128⇤⇤⇤ -0.077⇤⇤⇤ 0.128⇤⇤⇤ -0.077⇤⇤⇤ 0.128⇤⇤⇤ -0.077⇤⇤⇤ HavetheSameGender (0.026) (0.029) (0.026) (0.029) (0.025) (0.028)

Control Variables Dummy for Whether j was Active from 0.100⇤⇤⇤ 0.157⇤⇤⇤ 0.101⇤⇤⇤ 0.157⇤⇤⇤ 0.101⇤⇤⇤ 0.157⇤⇤⇤ t 7tot 1(0.030)(0.032)(0.030)(0.032)(0.029)(0.031) Dummy for Whether t Is a Weekend -0.494⇤⇤⇤ 0.519⇤⇤⇤ -0.495⇤⇤⇤ 0.518⇤⇤⇤ -0.495⇤⇤⇤ 0.518⇤⇤⇤ (0.075) (0.075) (0.075) (0.076) (0.073) (0.073) a Number of Membership Days by t -0.562⇤⇤⇤ -0.562⇤⇤⇤ -0.562⇤⇤⇤ (0.003) (0.005) (0.004) Dummy for i Being Non-Core User -0.941⇤⇤⇤ -0.940⇤⇤⇤ -0.941⇤⇤⇤ (0.087) (0.087) (0.076) Dummy for i Havingjoinedbefore 0.203⇤⇤⇤ 0.203⇤⇤⇤ 0.204⇤⇤⇤ July2007 (0.008) (0.009) (0.008) Constant -1.574⇤⇤⇤ -1.575⇤⇤⇤ -1.575⇤⇤⇤ (0.081) (0.085) (0.071) Standard Deviation of Friendship 0.025⇤⇤⇤ 0.021⇤⇤⇤ Random E↵ect (0.005) (0.006) Week Dummies yes yes yes

Model Summary Statistics NumberofObservations 69,020,774 69,020,774 69,020,774 AIC 287,850.051 287,884.646 287,835.360 BIC 288,588.329 288,622.924 288,702.056 LogLikelihood -143,879.025 -143,896.323 -143,863.680 Standard errors in parentheses.

⇤ p<0.05, ⇤⇤ p<0.01, ⇤⇤⇤ p<0.001 a Measured on logarithmic scale.

103 Table 3.2 continued (i) (ii) (iii) Independent Homogenous MainModel Core Non-Core Core Non-Core Core Non-Core Anime Watching a Number of Friends by t 1 -0.030⇤⇤⇤ -0.028⇤⇤⇤ -0.028⇤⇤⇤ (0.005) (0.005) (0.005) a Number of Animes Watched by Friends in t 1 0.082⇤⇤⇤ 0.082⇤⇤⇤ 0.082⇤⇤⇤ (0.005) (0.005) (0.006) Number of Posts Published by Friends in t 1a 0.003 (0.006) Dummy for Whether i Watched an Anime 0.872⇤⇤⇤ 0.872⇤⇤⇤ 0.872⇤⇤⇤ in t 1(0.011)(0.011)(0.011) Dummy for Whether t Is a Weekend 0.051⇤⇤⇤ 0.051⇤⇤⇤ 0.051⇤⇤⇤ (0.009) (0.009) (0.009) Constant -1.328⇤⇤⇤ -1.328⇤⇤⇤ -1.328⇤⇤⇤ (0.008) (0.008) (0.008) StandardDeviationofAnime 0.011⇤⇤⇤ 0.010⇤ Adoption Random E↵ect (0.004) (0.004) Week Dummies yes yes yes

Content Generation Number of Friends by t 1a 0.013 0.014 0.013 (0.007) (0.007) (0.009) Number of Animes Watched by Friends in t 1a 0.000 (0.008) a Number of Posts Published by Friends in t 1 0.208⇤⇤⇤ 0.208⇤⇤⇤ 0.208⇤⇤⇤ (0.008) (0.008) (0.008) Dummy for Whether i Published a Post in t 11.706⇤⇤⇤ 1.706⇤⇤⇤ 1.706⇤⇤⇤ (0.017) (0.017) (0.017) a Number of Animes Watched by t 1 0.032⇤⇤⇤ 0.032⇤⇤⇤ 0.032⇤⇤⇤ (0.004) (0.004) (0.004) Dummy for Whether t Is a Weekend 0.023 0.024 0.024 (0.014) (0.014) (0.014) Constant -2.201⇤⇤⇤ -2.201⇤⇤⇤ -2.201⇤⇤⇤ (0.019) (0.019) (0.019) Standard Deviation of UGC Production Random E↵ect 0.013⇤⇤⇤ 0.016⇤⇤ (0.007) (0.006) Week Dummies yes yes yes

Error Correlation Matrix FriendshipErrorStandardDeviation 1 1 1

AdoptionErrorStandardDeviation 0.009⇤⇤⇤ 0.014⇤⇤⇤ 0.013⇤⇤⇤ (0.004) (0.004) (0.004) UGCErrorStandardDeviation 0.005 0.012⇤⇤⇤ 0.011⇤⇤⇤ (0.006) (0.005) (0.004) Correlation between Friendship and Adoption 0.729⇤⇤⇤ 0.716⇤⇤⇤ (0.215) (0.226) Correlation between Friendship and UGC -0.874⇤⇤⇤ -0.865⇤⇤⇤ (0.219) (0.237) CorrelationbetweenAdoptionandUGC -0.449 -0.397 (0.409) (0.436) Random E↵ects Correlation Matrix CorrelationbetweenFriendshipandAdoption -0.426 (0.290) Correlation between Friendship and UGC 0.051 (.) CorrelationbetweenAdoptionandUGC -0.085 (0.241) Standard errors in parentheses.

⇤ p<0.05, ⇤⇤ p<0.01, ⇤⇤⇤ p<0.001 a Measured on logarithmic scale. 104 more likely to connect with friends of friends. Having more friends in common serves as asignalfortheunobservedmatchbetweenindividuali and individual j. This ﬁnding is in line with results in the previous literature (e.g., Aral et al. 2009; Shalizi and Thomas

2011) suggesting that “birds of a feather flock together.” The coecient for the number of common animes, however, is insignificant. In terms of demographic similarities, I find positive coecients for user i and user j being from the same country, being close in age, and having the same gender if both individuals reveal this information. However, the coecients for dummies indicating whether both users provided the information are negative. In other words, knowing demographic information about each other only increases the chance of a friendship if both users are similar in those characteristics. Otherwise, it actually hurts the chance of forming a friendship tie.

Comparing the magnitudes of the e↵ects of user j’s friends versus user j’s activities on the platform using marginal e↵ects, I ﬁnd the following: The two largest drivers of user i’s utility of forming a friendship with user j are user j’s number of friends and the number of friends user i and user j have in common. User j’s online activities, i.e., his number of watched animes and his number of written posts, only play a secondary role.

And lastly, I find significant e↵ects for all my control variables. The coecient for the dummy variable indicating whether user j showed any activity during the previous week is positive and significant. One likely reason is the increased visibility and awareness of user j that comes with activity. Next, the weekend dummy has a significant negative coecient. Users are less likely to make friends on weekends. And lastly, as expected, user i’s utility of making friends declines with the length of her membership on the platform since, especially shortly after joining, having friends helps to reduce the learning costs associated with navigating the website.

Next, I discuss my results for anime watching. I ﬁnd a negative signiﬁcant e↵ect for the cumulative number of friends implying that having more friends does not make a user

105 more active. The number of animes watched by friends reflects the influence friends have on a user’s anime watching behavior. I find a significant positive coecient for the e↵ect of the number of animes watched by friends the previous day indicating the existence of peer e↵ects. However, I do not find any spill-over e↵ects of friends’ posting behavior on a user’s anime watching. My results also reveal a positive state dependence in anime watching. Users are more likely to watch an anime if they did so on the previous day. And lastly, the coecient for the weekend dummy is positive and significant implying that users are more likely to watch animes on weekends. A potential explanation is that users might have more free time during weekends. Inowdescribemyfindingsrelatedtocontentgeneration.Ifindaninsignificantcoecient for the cumulative number of friends indicating that users with more friends do not publish more content. I also find evidence for friends’ influence on users’ content generation behavior. The number of posts published by friends has a significant positive e↵ect on a user’s content generation decision. However, there is no spill-over e↵ect. The number of animes watched by friends does not have a significant e↵ect on a user’s UGC production. In addition, there is evidence of positive state dependence in content generation, i.e., I find a significant positive e↵ect of a user’s posting on her posting behavior the following day. One likely reason is the conversation/discussion nature of content generation. Other users can post a reply or comment on a post published by the focal user and/or the user herself might respond by writing another post. Additionally, as discussed in Section 3.4.4, I model a user’s utility of UGC as a function of the number of animes watched by that user. I find a significant positive e↵ect of the number of animes a user has watched. The more animes a user watched, the more likely it is that the user publishes content (likely about the watched anime). This provides another reason for the existence of state dependence. When a user watches animes, she is likely to want to talk about them. This interest in talking might last for a few days and state dependence captures this e↵ect. And lastly, I do not find a significant e↵ect for the weekend dummy.

106 To summarize, my results for friendship formation reveal that both the attractiveness of and similarity with a potential friend matter. Further, the number and overlap of users’ friends are more important drivers of friendship formation than product adoption and content generation activities. And lastly, even in (anonymous) online networks having similar demographics matters. With regard to online activities of anime watching and content generation, I ﬁnd evidence of signiﬁcant peer e↵ects. Having friends who watch many animes and post a lot makes a user more likely to do the same. However, simply having many friends does not result in more activity.

3.8 Counterfactual

For companies operating social networks, advertising revenue represents their primary source of income. In 2015, the industry had revenues of over $25 billion through advertisements.27

Advertising revenues depend on site trac: the more active users are, the more ads can be shown to them. In addition, having more active users can increase the appeal of the website to non-users and lead to continuous growth of the user base. Therefore, it is in platform owners’ best interest to motivate users (or a subset of users if stimulating all users is not feasible) to increase their in-site activities.

The existence of peer e↵ects within social networks implies that an increase in a user’s activity level can have a cascading e↵ect on the user’s friends and friends of friends and so on. Previous literature has also shown that seeding to more connected users is the most e↵ective way of increasing the total number of product adoptions within a community (e.g.,

Hinz et al. 2011; Aral et al. 2013). However, the studies in this area assume a static network structure that does not evolve over time. While this result may hold true for mature networks where a static assumption applies, a stimulation intervention in evolving networks is very

27https://www.statista.com/statistics/271406/advertising-revenue-of-social-networks-worldwide/

107 likely to also lead to a change in the structure of the network due to the possibility of newly formed ties. As a result, to understand di↵usion patterns in evolving networks, one needs to take the evolving ties in the network into account as well. By modeling the co-evolving friendship network and users’ actions under their friends’ inﬂuence, I capture the cascading e↵ects of stimulating users to conduct more activities of a speciﬁc type on future states of the network and users’ future activity levels.

Using my estimation results, I examine the e↵ects of stimulating di↵erent types of users and di↵erent types of in-site activities. More precisely, I assume that the platform can trigger an increase in any of the three activities of making friends, watching animes, and generating content by using a recommendation system. For example, the platform can recommend a user to become friends with some other users, to adopt some speciﬁc animes, or to participate in forum discussions that are active and related to the user’s past adoptions or posts. Although

Idonotobservetheloginorpageviewactivitiesofauserand,asaresult,cannotdirectly translate the changes in activity levels to changes in ad viewership, as long as users are not spending less time on each activity compared to before the stimulation, an increase in the total activity level will also lead to an increase in the time spent on the website. Furthermore, an increase in the activity level is observable by other users and non-users of the website and therefore can lead to activity cascades as well as a growing user base.

For the simulations, I use the state of the network for all users in my sample on day 150.

Note that on day 150, only 1,194 out of the 1,386 users in my sample were members of the website. The remaining users joined the website sometime between day 151 and 184 and their actions are simulated from the time they join. Furthermore, as discussed in Section

3.3.1, I simulate the actions of all core users and all non-core users going forward until day

184. However, I take the actions of friends of non-core users as exogenous and adjust the relevant independent variables for non-core users. Furthermore, I only compare the changes in activity levels of core users for whom I estimated the model based on their full network.

108 3.8.1 How to Increase In-Site Activities through Platform-Wide Stimulation System?

In the ﬁrst set of simulations, I examine and compare the overall activity level of all core users due to the implementation of di↵erent platform-wide stimulation systems where taking the evolving network structure due to the stimulation into account. Here I sum the three di↵erent activities to make the overall activity level as a proxy for the total site trac or total time spent on the site. In each scenario, I increase one type of activity (friendship ties, product adoptions or UGC generation) among all core and all non-core users by one standard deviation on day 150 and simulate users’ behavior going forward until day 184. Figure 3.8 plots daily overall activity levels and the number of active users among all core users. Out of the three types of stimulation strategies, stimulating users to watch more animes leads to the highest overall activity level among users, 35% higher than the result of the stimulation strategy to have users make more friends and two times higher than the result of the UGC generation stimulation. However, in terms of the number of users who engage in at least one activity, the stimulation strategy to encourage users to make more friends slightly outperforms product adoption stimulation strategy (by 1%). These ﬁndings imply that out of the three recommendation systems the platform can implement (i.e., to recommend friends, animes, or forum discussion topics), the anime recommendation system that increases users’ anime watching is the most e↵ective in driving the overall site trac while the friending recommendation system that increases the number of friendships is the most e↵ective in producing more active users.

3.8.2 How to Increase In-Site Activities through Seeding?

Previous literature has found that not all users have the same degree of inﬂuence on their friends (e.g., Manchanda et al. 2008; Iyengar et al. 2011). Further, users can also have varying degrees of activity in di↵erent areas. For example, a user might make many friends

109 (b) Number of Active Users Per (a) Number of Activities Per Day Under Day Under Di↵erent Recommendation Di↵erent Recommendation Strategies Strategies 100 150 80 100 60 40 50 Number of Activities Number of Active Users 20 0 0 150 155 160 165 170 175 180 185 150 155 160 165 170 175 180 185 Day Day

Baseline Friend Adoption UGC Baseline Friend Adoption UGC

Figure 3.8: Number of Activities and Active Users Per Day Under Di↵erent Recommendation Strategies (Colors Represent Seeding in Indicated Areas of Activities.) or publish many posts but only watch few animes. Consequently, carefully choosing whom to target and which type of activity to stimulate are crucial for platform owners. In the second set of simulations, I examine the e↵ectiveness of di↵erent seeding strategies in increasing tie formations and UGC production within the network. For these simulations, I select the top/bottom 50 most/least active users (about 15% of core users) among the core users based on their activity levels in making friends, adopting animes, producing UGC or conducting any activity.28 Next, I increase the activity level for these selected users in one of the three activities by one standard deviation in the beginning of day 150 and simulate the network evolution and users’ activities until day 184.

IreporttheresultsinTable3.3.ThenumbersinTable3.3representthepercentage changes in users’ activities compared to the baseline scenario without any stimulation. First,

Idiscussthee↵ectivenessofseedingstrategiesinincreasingthenumberoffriendships.The two best strategies are to target the bottom 50 users based on their anime watching and

28For the selection of users in the “any activity” category, I use the sum of the standardized activity levels in the three areas.

110 Table 3.3: Counterfactual Results

Changes in SELECTION Activity SEEDING Activity Friendships in % UGC in %

Top 50 Selection Friend 4.91 0.87 Friendship Anime 0.00 7.36 Post 2.39 0.15

Friend 6.97 1.56 Anime Watching Anime 0.86 0.97 Post 1.59 0.23

Friend 5.11 2.22 UGC Anime 0.33 15.92 Post 1.26 0.00

Bottom 50 Selection Friend 5.44 1.26 Friendship Anime 1.20 1.88 Post 1.79 0.71

Friend 34.06 2.22 Anime Watching Anime 0.47 17.53 Post 7.04 0.23

Friend 0.40 0.01 UGC Anime 0.47 0.00 Post 0.40 0.00

to stimulate their friend making. The number of formed friendships increases by 34% when this strategies is applied. In general, motivating users to make friends is the best strategy to increase friendship ties in a network. Furthermore, promoting UGC production is more e↵ective than promoting adoptions to increase connectedness within the network.

The best two strategies for increasing UGC production on the website are selecting the top

50 users in terms of UGC production and the bottom 50 users in terms of anime watching and stimulating them to watch more animes. These two strategies result in a 16% and

111 18% increase in UGC production, respectively. My simulation results show that, if platform owners want to increase UGC on their website, promoting post publishing behavior is far less e↵ective that promoting product adoptions. Furthermore, my results reveal that selecting users based on their number of friends is not always the best seeding strategy for ﬁrms.

This result is also in line with (Katona et al., 2011)’s ﬁnding that average inﬂuential power of individuals decreases with their total number of contacts. Although in comparing the e↵ectiveness of selecting the top 50 and the bottom 50 users in terms of their number of friends, selecting the better connected users generally results in more UGC, this strategy is still far less e↵ective (by about 150%) than selecting the top 50 users based on their UGC production and stimulate them to watch more animes.

Lastly, I examine by how much the e↵ectiveness of seeding strategies is underestimated when the endogenous network formation is not accounted for. To do so, I re-run the counterfactual scenarios discussed in this section, but do not allow users to form new friendships.

I ﬁnd that not accounting for the endogenous network formation leads, on average, to an underestimation of seeding e↵ectiveness by 10%.

The best two strategies for increasing UGC production on the website are to select the top 50 users in UGC production and the bottom 50 users in anime watching and to have them watch more animes. These two strategies result in a 16% and 18% increase in

UGC production, respectively. My simulation results show that, if platform owners want to increase UGC on their website, promoting post publishing behavior is far less e↵ective that promoting anime watching. Furthermore, my results reveal that targeting the most connected users is not always the best seeding strategy for ﬁrms. I ﬁnd that well-connected users in the network tend to become friends with other well-connected users who are not necessarily the ones who would produce a lot of UGC or watch many animes. Because of this non-overlap of users with a lot of friends and users with a lot UGC, selection based on the metric of number of friends is not the best strategy. This result is in line with (Katona

112 et al., 2011) in which the authors find that average influential power of individuals decreases with their total number of contacts. Although seeding the more connected users generally leads to more UGC production than seeding the less connected users, this strategy is still far less e↵ective (by about 150%) than targeting the top 50 users in UGC production and stimulating them to watch more animes. Lastly, I examine to what extent the e↵ectiveness of seeding strategies is underestimated when the endogenous network formation is not accounted for. To do so, I re-run the counterfactual scenarios discussed in this section, but do not allow users to form new friendship ties. I find that not accounting for the endogenous network formation leads, on average, to an underestimation of seeding e↵ectiveness by 10%.

3.9 Limitations and Future Research

There are several limitations to my research. First, I only observe a friendship if both users agree to become friends. In other words, I do not observe the ﬁrst request for the friendship and the potential rejection of that request. This is a limitation of my data. As a result, I cannot separately identify whether an increase in a user’s number of friends is due to that user’s tendency to form friendships or due to an increase of attractiveness of that user to other users. Second, in my data, I do not observe tie dissolution and thus assume everlasting friendships. Although, due to the small cost of friendship ties for users, I do not believe unfriending is a frequent action in the network under study, it is still possible for users to break their friendship ties. This unfriending behavior in itself is interesting and can provide additional insights into network formation dynamics. Third, in this paper, I model whether users post something on the website or watch an anime and not where and how many posts users write or what particular anime they watch. Studying the details of each action can shed further light on the co-evolution process of users’ friendship formations and concurrent actions and is left for future research. Fourth, I

113 do not model platform growth in my relatively short observation period, i.e., I do not model users’ joining behavior and assume it is exogenous. However, in the long run, popularity of aplatformintermsofitsuserbaseandrichnessofitscontentcanchangetherateofusers joining the website. And lastly, in this paper, I do not consider the content of users’ posts.

Longer or more detailed posts may imply the writer is more knowledgeable and thus more attractive. However, studying the e↵ects of characteristics of users’ UGC is left for future research.

3.10 Conclusion

Idevelopastructuralmodelfortheco-evolutionofindividuals’friendshiptieformationsand their concurrent online activities (product adoptions and production of user-generated content) within a social network. Explicitly modeling the endogenous formation of the network and accounting for the interdependence between decisions in these two areas (friendship formations and concurrent online activities) provides a clean identiﬁcation of peer e↵ects and of important drivers of individuals’ friendship decisions. I estimate my model using a novel data set capturing the continuous development of a network and users’ entire action histories within the network.

My results reveal that, compared to a potential friend’s product adoptions and content generation activities, the total number of friends and the number of common friends this potential friend has with the focal individual are the most important drivers of friendship formation. Further, while having more friends does not make a person more active, having more active friends does increase a user’s activity levels in terms of both product adoptions and content generation through peer e↵ects. Via counterfactuals I assess the e↵ectiveness of various seeding and stimulation strategies in increasing website trac while taking the endogenous network formation into account. Contrary to previous studies (e.g., Hinz et al.

114 2011; Aral et al. 2013), I ﬁnd that seeding to users with the most friends is not always the best strategy to increase users’ activity levels on the website.

115 APPENDIX A

ORIGINAL AND FINAL DATA

In this appendix, I compare the characteristics of the eligible population of 40,000 users to the (i) data set containing all 380,000 users from the same network and to a (ii) data set containing nearly 80,000 users from the same network who have the same “platform lifetime,” i.e., joined the platform before July 2011, as my eligible population. For users to be included in the last data set (ii), I also require them to show some type of activity after the release of the last anime under study in January 2014, i.e., that they add at least one anime to their watch list (any anime; not necessarily one of the animes under study). Comparing my eligible population not only to all users from the same network, but also to users with the same platform lifetime is necessary since some of the descriptive statistics such as the number of friends change with users’ length of membership on the platform. The di↵erence between the eligible population and the population of users with the same lifetime is that the latter also includes users who showed little to no activity during the study period, i.e., added fewer than 10 animes to their watch list (any anime; not necessarily one of the animes under study). Panel A of Table A.1 shows the descriptive statistics for the population of 380,000 users, Panel B of Table A.1 displays the descriptive statistics for the same-platform-lifetime population of nearly 80,000 users, and Panel C of Table A.1 shows the descriptive statistics for the eligible population of 40,000 users. In terms of demographics (age and gender), users in Panels B and C are similar. Users in Panel C compared to users in Panel B watch, on average, more animes per year and adopt more animes among those under study. However, this di↵erence is to be expected since Panel B includes users with little to no activity during the study period. Lastly, conditional on adopting, users in Panels B and C adopt animes at the same time. Comparing Panels A and C, I ﬁnd that newer users are less likely to report their gender. Users in Panel A have fewer friends which is likely due to them having been

116 a member of the platform for a signiﬁcantly shorter time period. Further, users in Panel A watch fewer animes per year and adopt fewer animes among those under study than users in Panel C – this di↵erence is due to Panel A including users who show little to no activity. Lastly, users in Panel A watch animes conditional on adoption about 4 weeks later than users in Panel C. Iconcludethatusersinmyeligiblepopulationarerepresentativeactiveusersonthe platform.

117 Table A.1: Descriptive Statistics

Mean Std. Dev. Min Median Max N

PANEL A: Population of 380,000 Users Age 23 7 12 22 84 218,130 Gender (% Females) 26 377,644 Gender (% Males) 40 377,644 Gender(%NotSpeciﬁed) 34 377,644 NumberofFriends 9 26 1 33,731 377,644 AverageNumberofAnimesAdoptedperYear 55 65 0 32 330 377,644 Number of Animes Adopted Among AnimesUnderStudy 11 15 0 3 103 377,644 AdoptionWeek(ConditionalonAdoption) 21 14 1 17 52 1,368,846

PANEL B: Same-Platform-Lifetime Population Age 24 5 12 23 84 25,500 Gender (% Females) 33 76,069 Gender (% Males) 45 76,069 Gender(%NotSpeciﬁed) 22 76,069 NumberofFriends 13 33 1 43,077 76,069 AverageNumberofAnimesAdoptedperYear 45 58 0 28 2,633 76,069 Number of Animes Adopted Among Animes Under Study 7 14 0 0 103 76,069 AdoptionWeek(ConditionalonAdoption) 17 13 1 14 52 738,969

PANEL C: Eligible Population Age 23 6 11 23 84 24,584 Gender (% Females) 35 39,652 Gender (% Males) 52 39,652 Gender(%NotSpeciﬁed) 13 39,652 NumberofFriends 18 31 1 91,720 39,652 AverageNumberofAnimesAdoptedperYear 76 68 1 58 2,144 39,652 Number of Animes Adopted Among AnimesUnderStudy 17 18 0 8 103 39,652 AdoptionWeek(ConditionalonAdoption) 16 13 1 13 52 614,048

118 APPENDIX B

VARIABLE (RE-)CONSTRUCTION

In this appendix, I describe the process through which I (re-)constructed several variables used in the estimation of the main model and robustness checks, namely, the WOM and OL variables from the community network. For the WOM and OL variables from the community network, this re-construction was necessary because MyAnimeList.net only shows the current levels of these variables, but not historical values. Thus I had information on the number of adoptions, the rank based on the number of adoptions, the average rating, the rank based on the average rating, and the number of ratings and forum posts from the community network in March 2015 (start date of the data collection), but not earlier to that.

Community OL

The variable “Cumulative Number of Community Users Who Adopted” is my main measure of OL from the community network and, as the name indicates, captures the cumulative number of users from the community network who have adopted the anime. I used my complete collected data containing the adoption histories of 380,000 users to re-construct the cumulative number of users who adopted each anime for each week. The variable “Community Rank” is my alternative measure of OL from the community network and captures the weekly rank of an anime among all the animes on the website based on the cumulative number of adoptions by all users. Note that a lower rank is a “better” rank. On MyAnimeList.net, this popularity rank is explained as:

“This popularity is measured according to the number of users who have the title in their list. The more users that have the title shown in their Anime or Manga list, the higher it will be ranked.”

119 Iusedmycompletecollecteddatacontainingtheadoptionhistoriesofnearly380,000users to re-construct the weekly rank data based on users’ adoption behavior using the following steps: First, for each anime on the website and each week, I calculated the cumulative number of users who had adopted the anime. Second, for each week, I sorted all animes in a decreasing order based on the cumulative number of adoptions. Thus, for each week, the position of each anime in the sorted list indicates the rank of that anime among all animes.

To test the accuracy of the re-constructed rank data, I compared the re-constructed ranks of several randomly selected animes in the ﬁrst week of March 2015 to the ranks provided by the website at that point in time. The comparison showed that I are able to closely recover anime ranks.

Community WOM Valence

The variable “Community Rating” is my main measure of WOM valence from the community network and captures the average of user ratings for an anime from all users in the community network. I re-constructed the community rating for each week using the ratings for an anime from all users who had adopted and rated the anime among the nearly 380,000 users. For each week, I calculated the average rating based on the ratings submitted by all users by that week. Then I compared the re-constructed values of the “Community Rating” variable for several randomly selected animes in the ﬁrst week of March 2015 to those shown on the website at the same point in time and found my re-constructed “Community Ratings” to be close to those shown on the website.

The variable “Community Rating Rank” is my alternative measure of WOM valence from the community network and describes the rank of an anime based on ratings submitted by the whole community. MyAnimeList.net reports the method used to calculate these weighted ranks as:

120 “Only scores where a user has completed at least 1/5 of the anime/manga are calculated. Example: If you watched a 26 episode series, this means you would had to have watched at least 5 episodes (26/5.2)=5. We’re using 5.2 instead of 5 so I get a whole number for “most” series. The formula used is: W eightedRank(WR)=(v/(v + m)) S +(m/(v + m)) C ⇤ ⇤ S = Average score for the Anime (mean). v = Number of votes for the Anime = (Number of people scoring the Anime). m = Minimum votes/scores required to get a calculated score (currently 50 scores required). C = The mean score across the entire Anime DB.”

Iappliedthisformulaandcalculatedweeklyweightedranksforallanimes.Totestthe accuracy of the re-constructed rank data, I compared the re-constructed ranks of several randomly selected animes in the ﬁrst week of March 2015 to the ranks provided by the website at that point in time. The comparison showed that I are able to closely recover anime ranks.

Community WOM Volume

The variable “Community Number of Ratings and Forum Posts” is my measure of WOM volume from the community network and, as the name indicates, captures the cumulative number of ratings and forum posts published about an anime. Since each rating and forum post is dated, i.e., has a calendar date, I re-constructed this variable by calculating the cumulative number of ratings and forum posts for each anime and each week.

121 APPENDIX C

ROBUSTNESS CHECKS

Table C.1: Robustness Checks

(i) (ii) (iii) (iv) Alternative Operationalizations for ... Personal OL Community OL Community WOM Adoption Time

Word-of-Mouth Friends’Av.RatingDummy -0.0043*** -0.0029*** -0.0035*** -0.0038*** (0.0000) (0.0002) (0.0002) (0.0002) Friends’Av.RatingInteraction 0.0003*** 0.0001*** 0.0001*** 0.0002*** (0.0001) (0.0000) (0.0000) (0.0000) Friends’ No. Ratings and Forum Postsa 0.0056*** 0.0033*** 0.0039*** 0.0035*** (0.0002) (0.0001) (0.0001) (0.0001) Community Rating 0.0216*** 0.0136*** 0.0074*** (0.0001) (0.0002) (0.0002) Community Rating Ranka -0.0030*** (0.0001) Community Number Ratings and Forum Postsa 0.0002*** 0.0012*** 0.0011*** 0.0004*** (0.0000) (0.0000) (0.0000) (0.0000) Observational Learning Cum. Number of Friends Who Adopteda 0.0022*** 0.0035*** 0.0030*** (0.0001) (0.0001) (0.0001) Cum. PercentageofFriendsWhoAdopted 0.0025*** (0.0001) Cum. Number of Community Users Who Adopteda 0.0030*** 0.0021*** 0.0023*** (0.0001) (0.0001) (0.0001) Community Ranka -0.0056*** (0.0001) Other Parameters Number of Animes Watched During the Weeka 0.0078*** 0.0078*** 0.0078*** 0.0011*** (0.0000) (0.0000) (0.0000) (0.0000) Number of Online Newsa 0.0001*** 0.0001*** 0.0000 -0.0001** (0.0000) (0.0000) (0.0000) (0.0000) SeasonFinaleDummy -0.1192*** -0.0909*** -0.0997*** -0.1271*** (0.0010) (0.0044) (0.0012) (0.0010) Season Finale Dummy Community Rating 0.0019*** 0.0024*** 0.0023*** ⇥ (0.0001) (0.0001) (0.0001) Season Finale Dummy Community Rating Rank -0.0012*** ⇥ (0.0001) Season Finale Dummy CommunityNumberof 0.0137*** 0.0132*** 0.0150*** 0.0141*** Ratings and Forum Posts⇥ a (0.0002) (0.0003) (0.0003) (0.0002) Season Finale Dummy Cum.Numberof 0.0007** -0.0001 0.00017*** Community Users Who⇥ Adopteda (0.0002) (0.0003) (0.0002) Season Finale Dummy Community Ranka -0.0030*** ⇥ (0.0000)

User-Anime Fixed E↵ects Yes Yes Yes Yes User-Release Week Fixed E↵ects Yes Yes Yes Yes Calendar Week Fixed E↵ects Yes Yes Yes Yes AdjustedR-Squared 0.1238 0.1240 0.1234 0.1258 NumberofObservations 21,853,295 21,853,295 21,853,295 21,808,418 Standard errors in parentheses * p<0.05, ** p<0.01, *** p<0.001 a Measured on logarithmic scale.

122 APPENDIX D

PROBIT RESULTS

Table D.1: Probit Results - North America

Franchises UGC Franchise Available Franchise Not Available Forum Post Recommendation (i) (ii) (iii) (iv)

Binge Dummy 0.038** -0.004 -0.077*** -0.017 (0.012) (0.031) (0.023) (0.067)

Sequel Dummy 1.263*** 0.780*** (0.009) (0.019)

Binge Dummy Sequel Dummy 0.143*** 0.031 ⇥ (0.020) (0.040)

Own Rating of Focal Season 0.137*** 0.170*** 0.043*** 0.042* (0.003) (0.007) (0.007) (0.018)

Own Rating Dummy

Popularity Rank of Focal Seasona -0.090*** -0.082*** -0.055*** -0.068** (0.004) (0.009) (0.008) (0.025)

Community Rating of Focal Season -0.151*** -0.114*** -0.001 0.010 (0.010) (0.022) (0.021) (0.065)

WaitTimeUntilFranchiseExtensionAvailable -0.156*** When Started Watching Focal Seasona (0.006)

Number of Episodes of Focal Seasona -0.035*** 0.239*** 0.299*** 0.164** (0.009) (0.025) (0.019) (0.063)

Duration of an Episodea -0.206*** 0.525*** 0.148* 0.115 (0.033) (0.090) (0.072) (0.213)

Constant 0.700*** -2.112*** -2.326*** -2.256* (0.137) (0.358) (0.295) (0.879)

Log Variance of User Random E↵ects -0.137 (0.358) -0.474*** -1.808*** (0.026) (0.044) (0.045) (0.167)

Genre Dummies Yes Yes Yes Yes

YearDummies Yes Yes Yes Yes NumberofObservations 173,025 36,859 57,523 8,661 AIC 186,931.685 42,740.038 36,282.44 3,795.95 BIC 187,505.173 43,225.385 36,775.23 4,170.48 LogLikelihood -93,408.842 -21,313.019 -18,086.22 -1,844.98 Standard errors in parentheses. * p<0.05, ** p<0.01, *** p<0.001 a Measured on logarithmic scale.

123 Table D.2: E↵ects of Binge-Watching Across Di↵erent Regions

Franchises UGC Franchise Available Franchise Not Available Forum Post Recommendation (i) (ii) (iii) (iv)

North America Binge Dummy 0.038** -0.004 -0.077*** -0.017 (0.012) (0.031) (0.023) (0.067) Sequel Dummy 1.263*** 0.780*** (0.009) (0.019) Binge Dummy Sequel Dummy 0.143*** 0.031 ⇥ (0.020) (0.040)

South America Binge Dummy 0.017 0.045 0.071 -0.202 (0.024) (0.061) (0.076) (0.238) Sequel Dummy 1.159*** 0.744*** (0.016) (0.035) Binge Dummy Sequel Dummy 0.199*** 0.012 ⇥ (0.041) (0.078)

Europe Binge Dummy 0.024** 0.029 -0.035 -0.025 (0.009) (0.022) (0.023) (0.070) Sequel Dummy 1.172*** 0.741*** (0.007) (0.014) Binge Dummy Sequel Dummy 0.192*** -0.004 ⇥ (0.015) (0.028)

Asia Binge Dummy 0.020 -0.004 0.064 0.074 (0.023) (0.055) (0.047) (0.113) Sequel Dummy 1.218*** 0.829*** (0.019) (0.037) Binge Dummy Sequel Dummy 0.167*** -0.016 ⇥ (0.038) (0.069)

Oceania Binge Dummy -0.043 0.145 0.024 -0.471 (0.033) (0.094) (0.065) (0.291) Sequel Dummy 1.206*** 0.826*** (0.025) (0.061) Binge Dummy Sequel Dummy 0.318*** -0.006 ⇥ (0.057) (0.122)

Standard errors in parentheses. * p<0.05, ** p<0.01, *** p<0.001

124 APPENDIX E

COMPLETE RESULTS FOR OTHER CONTINENTS

Table E.1: Results - South America

Franchises UGC Franchise Available Franchise Not Available Forum Post Recommendation (i) (ii) (iii) (iv)

Engagement Equation Binge Dummy -0.204* 0.045 -0.568 -0.830 (0.100) (0.061) (0.376) (0.486) Sequel Dummy 1.238*** 0.744*** (0.039) (0.035) Binge Dummy Sequel Dummy 0.205*** 0.012 ⇥ (0.044) (0.078) Own Rating of Focal Season 0.152*** 0.164*** 0.047* 0.080 (0.008) (0.013) (0.023) (0.108) Own Rating Dummyb

Popularity Rank of Focal Seasona -0.099*** -0.062*** 0.062* 0.307 (0.008) (0.017) (0.031) (0.262) Community Rating of Focal Season -0.091*** -0.060 -0.128 -0.378 (0.019) (0.039) (0.080) (0.61) WaitTimeUntilFranchiseSeriesAvailable -0.179*** When Started Watching Focal Seasona (0.012) NumberofEpisodesofFocalSeason -0.028 0.325*** 0.456*** 0.414 (0.018) (0.048) (0.087) (1.026) Duration of an Episodea 0.080 1.068*** -0.084 3.895 (0.069) (0.184) (0.253) (6.398) Constant -0.615* -4.374*** -1.675 -12.993 (0.282) (0.699) (1.081) (20.372) Variance of User Random E↵ects 0.460*** 0.540*** 0.960*** 0.050 (0.034) (0.046) (0.21) (0.168) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes

Binge Decision Equation WeekendDummy 0.099*** 0.030 0.054 -0.405 (0.020) (0.037) (0.064) (0.428) Popularity Rank of Focal Seasona -0.040*** 0.011 -0.007 -0.896 (0.009) (0.020) (0.031) (0.588) Community Rating of Focal Season 0.078*** 0.167*** 0.143 1.249 (0.023) (0.045) (0.079) (1.903) NumberofEpisodesofFocalSeason 0.196*** 0.084 0.250** 0.251 (0.022) (0.053) (0.080) (2.157) Duration of an Episodea 0.229* 0.210 0.742* -10.587 (0.099) (0.208) (0.353) (17.041) Constant -3.501*** -3.815*** -6.085*** 27.652 (0.391) (0.812) (1.458) (49.485) Variance of User Random E↵ects 1.173*** 0.851*** 0.999*** 1.123*** (0.094) (0.263) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes Error Correlation 0.145* 0.000 0.507 0.000 (0.068) (0.000) (0.334) (0.000) NumberofObservations 49,592 12,561 6,077 184 AIC 86,828.260 21,995.063 7,892.839 455.723 BIC 87,850.404 22,798.405 8,617.763 722.563 Log Likelihood -43,298.130 -10,889.531 -3,838.419 -144.862 Standard errors in parentheses. * p<0.05, ** p<0.01, *** p<0.001 a Measured on logarithmic scale. b Not estimated in all models due to collinearity.

125 Table E.2: Results - Europe

Franchises UGC Franchise Available Franchise Not Available Forum Post Recommendation (i) (ii) (iii) (iv)

Engagement Equation Binge Dummy -0.131* -0.091 -0.298* -0.002 (0.038) (0.087) (0.129) (0.076) Sequel Dummy 1.228*** 0.914*** (0.015) (0.032) Binge Dummy Sequel Dummy 0.198*** 0.032 ⇥ (0.016) (0.044) Own Rating of Focal Season 0.173*** 0.170*** 0.056*** 0.123*** (0.003) (0.009) (0.008) (0.022) Own Rating Dummyb -1.209*** (0.088) Popularity Rank of Focal Seasona -0.085*** -0.067*** -0.047*** -0.041 (0.003) (0.011) (0.009) (0.029) Community Rating of Focal Season -0.128*** -0.012 -0.062** -0.210** (0.008) (0.026) (0.023) (0.080) WaitTimeUntilFranchiseSeriesAvailable -0.157*** When Started Watching Focal Seasona (0.008) NumberofEpisodesofFocalSeason -0.017* 0.280*** 0.331*** -0.050 (0.007) (0.031) (0.026) (0.086) Duration of an Episodea -0.108*** 0.644*** 0.179* -0.304 (0.025) (0.099) (0.076) (0.276) Constant 0.055 -2.516*** -2.627*** 0.726 (0.104) (0.431) (0.337) (1.180) Variance of User Random E↵ects 0.420*** 0.543*** 0.524*** 0.142*** (0.012) (0.038) (0.042) (0.027) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes

Binge Decision Equation WeekendDummy 0.054*** 0.048* 0.047** 0.058 (0.007) (0.022) (0.015) (0.050) Popularity Rank of Focal Seasona -0.045*** -0.032* -0.040*** -0.036 (0.003) (0.013) (0.007) (0.025) Community Rating of Focal Season 0.063*** 0.082** 0.098*** 0.073 (0.009) (0.029) (0.018) (0.066) NumberofEpisodesofFocalSeason 0.184*** 0.147*** 0.207*** 0.383*** (0.008) (0.032) (0.019) (0.074) Duration of an Episodea 0.272*** 0.150 0.785*** 0.502 (0.034) (0.129) (0.073) (0.326) Constant -3.174*** -2.957*** -5.168*** -4.503*** (0.137) (0.521) (0.330) (1.267) Variance of User Random E↵ects 1.019*** 1.008*** 0.619*** 0.565*** (0.033) (0.075) (0.055) (0.000) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes Error Correlation 0.100* 0.063 0.174* 0.000 (0.025) (0.056) (0.088) (0.000) NumberofObservations 305,262 34,276 68,513 5,709 AIC 554,671.635 63,873.605 87,177.215 7,267.988 BIC 555,883.333 64,836.016 88,182.040 7,946.267 Log Likelihood -277,221.818 -31,822.802 -43,478.607 -3,531.994 Standard errors in parentheses. * p<0.05, ** p<0.01, *** p<0.001 a Measured on logarithmic scale. b Not estimated in all models due to collinearity.

126 Table E.3: Results - Asia

Franchises UGC Franchise Available Franchise Not Available Forum Post Recommendation (i) (ii) (iii) (iv)

Engagement Equation Binge Dummy -0.124 -0.288 -0.563 0.148 (0.087) (0.200) (0.324) (0.143) Sequel Dummy 1.273*** 0.904*** (0.038) (0.065) Binge Dummy Sequel Dummy 0.172*** -0.023 ⇥ (0.040) (0.075) Own Rating of Focal Season 0.154*** 0.155*** 0.122*** -0.001 (0.008) (0.017) (0.024) (0.048) Own Rating Dummyb

Popularity Rank of Focal Seasona -0.090*** -0.078*** 0.032 0.084 (0.008) (0.021) (0.025) (0.068) Community Rating of Focal Season -0.071*** -0.067 -0.261*** -0.247 (0.021) (0.044) (0.065) (0.224) WaitTimeUntilFranchiseSeriesAvailable -0.158*** When Started Watching Focal Seasona (0.016) NumberofEpisodesofFocalSeason -0.025 0.467*** 0.384*** 0.539** (0.020) (0.063) (0.079) (0.191) Duration of an Episodea -0.216* 1.540*** 0.518 -1.759 (0.085) (0.277) (0.317) (1.655) Constant -0.049 -6.118*** -2.850* 4.901 (0.328) (1.006) (1.155) (5.796) Variance of User Random E↵ects 0.427*** 0.579*** 0.780*** 0.064 (0.032) (0.082) (0.145) (0.042) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes

Binge Decision Equation WeekendDummy 0.037* 0.019 0.056 0.043 (0.019) (0.037) (0.040) (0.132) Popularity Rank of Focal Seasona -0.043*** -0.053* -0.053* -0.137 (0.009) (0.021) (0.021) (0.071) Community Rating of Focal Season 0.056* 0.115** 0.084 -0.020 (0.022) (0.044) (0.049) (0.234) NumberofEpisodesofFocalSeason 0.292*** 0.208*** 0.369*** 0.768*** (0.022) (0.054) (0.061) (0.216) Duration of an Episodea 0.520*** 0.558* 0.187 -0.285 (0.121) (0.284) (0.281) (1.689) Constant -4.320*** -4.222*** -3.621*** -1.877 (0.444) (1.027) (1.047) (6.106) Variance of User Random E↵ects 1.187*** 0.885*** 0.804*** 0.790*** (0.093) (0.135) (0.182) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes Error Correlation 0.095 0.198 0.481 0.000 (0.058) (0.146) (0.293) (0.000) NumberofObservations 41,774 11,751 10,162 808 AIC 78,942.647 22,679.711 15,678.752 1,513.000 BIC 79,910.330 23,497.969 16,459.204 1,958.983 Log Likelihood -39,359.324 -11,228.855 -7,731.376 -661.500 Standard errors in parentheses. * p<0.05, ** p<0.01, *** p<0.001 a Measured on logarithmic scale. b Not estimated in all models due to collinearity.

127 Table E.4: Results - Oceania

Franchises UGC Franchise Available Franchise Not Available Forum Post Recommendation (i) (ii) (iii) (iv)

Engagement Equation Binge Dummy -0.205 -0.038 -1.038 -1.553 (0.151) (0.295) (0.723) (1.662) Sequel Dummy 1.265*** 0.876*** (0.060) (0.100) Binge Dummy Sequel Dummy 0.330*** -0.014 ⇥ (0.060) (0.129) Own Rating of Focal Season 0.190*** 0.204*** 0.092** 0.221 (0.013) (0.028) (0.031) (0.156) Own Rating Dummyb

Popularity Rank of Focal Seasona -0.096*** -0.070* 0.075* -0.998 (0.012) (0.032) (0.035) (0.832) Community Rating of Focal Season -0.180*** -0.191* -0.286** -15.734 (0.031) (0.076) (0.106) (8.256) WaitTimeUntilFranchiseSeriesAvailable -0.154*** When Started Watching Focal Seasona (0.024) NumberofEpisodesofFocalSeason -0.013 0.386*** 0.612*** -32.050 (0.028) (0.095) (0.168) (19.397) Duration of an Episodea -0.104 1.491** 0.602 -155.213 (0.106) (0.504) (0.380) (98.063) Constant 0.136 -5.113** -2.829 859.468 (0.433) (1.766) (1.514) (509.339) Variance of User Random E↵ects 0.483*** 0.720*** 0.890** NA (0.055) (0.149) (0.313) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes

Binge Decision Equation WeekendDummy -0.039 0.124* -0.026 0.940 (0.026) (0.058) (0.068) (0.746) Popularity Rank of Focal Seasona -0.047*** -0.059 -0.031 0.120 (0.012) (0.031) (0.033) (0.945) Community Rating of Focal Season 0.070* 0.128 -0.009 -3.184 (0.032) (0.074) (0.086) (7.728) NumberofEpisodesofFocalSeason 0.177*** 0.096 0.358** 2.358 (0.030) (0.083) (0.109) (3800.451) Duration of an Episodea 0.190 -0.136 0.383 -70.462 (0.151) (0.320) (0.403) (1864.914) Constant -3.230*** -2.202 -3.988* 233.623 (0.571) (1.271) (1.730) (3595.251) Variance of User Random E↵ects 1.001*** 0.703*** 0.762* 0.000 (0.121) (0.166) (0.331) (0.000) Genre Dummies Yes Yes Yes Yes Year Dummies Yes Yes Yes Yes Error Correlation 0.103 0.129 0.849 0.524 (0.098) (0.208) (0.771) (1.29) NumberofObservations 22,713 4,995 4,940 143 AIC 41,728.857 9,072.505 8,004.589 370.712 BIC 42,628.294 9,795.802 8,707.142 625.517 Log Likelihood -20,752.428 -4,425.252 -3,894.294 -99.356 Standard errors in parentheses. * p<0.05, ** p<0.01, *** p<0.001 a Measured on logarithmic scale. b Not estimated in all models due to collinearity.

128 APPENDIX F

ALTERNATIVE CLASSIFICATIONS OF BINGE-WATCHING

Table F.1: Probability of Engagement Action – Binge-Watching Deﬁnition: More than 2 Hours

Next Season Other Franchises UGC Available Not Available Available Not Available Forum Posts Recommendation North America Non-Binge-Watch 0.677 0.543 0.276 0.308 0.146 0.058 Binge-Watch 0.733 0.542 0.292 0.294 0.112 0.052

South America Non-Binge-Watch 0.709 0.560 0.339 0.345 0.119 0.033 Binge-Watch 0.765 0.556 0.360 0.332 0.078 0.048

Europe Non-Binge-Watch 0.693 0.550 0.321 0.325 0.098 0.044 Binge-Watch 0.760 0.535 0.336 0.322 0.079 0.050

Asia Non-Binge-Watch 0.687 0.552 0.304 0.325 0.113 0.034 Binge-Watch 0.755 0.534 0.307 0.296 0.084 0.044

Oceania Non-Binge-Watch 0.670 0.562 0.284 0.347 0.141 0.045 Binge-Watch 0.748 0.553 0.298 0.331 0.115 0.035

129 Table F.2: Probability of Engagement Action – Binge-Watching Deﬁnition: More than 4 Hours

Next Season Other Franchises UGC Available Not Available Available Not Available Forum Posts Recommendation North America Non-Binge-Watch 0.693 0.546 0.283 0.311 0.139 0.057 Binge-Watch 0.710 0.526 0.279 0.280 0.103 0.045

South America Non-Binge-Watch 0.724 0.562 0.347 0.350 0.113 0.036 Binge-Watch 0.739 0.536 0.340 0.319 0.069 0.033

Europe Non-Binge-Watch 0.712 0.549 0.326 0.331 0.094 0.046 Binge-Watch 0.741 0.520 0.324 0.308 0.073 0.044

Asia Non-Binge-Watch 0.708 0.552 0.310 0.322 0.105 0.038 Binge-Watch 0.745 0.514 0.292 0.284 0.081 0.041

Oceania Non-Binge-Watch 0.688 0.558 0.291 0.349 0.135 0.042 Binge-Watch 0.749 0.564 0.287 0.324 0.117 0.039

130 APPENDIX G

LIKELIHOOD DERIVATION

In this section, I explain the estimation techniques used to estimate the likelihood. Recall that the full likelihood is

+ T N 1 cg Apa pa 1 Apa L = (P (A =1)) i,t (1 P (A =1)) i,t i,t i,t t=1 i=1 Z1 Y Y cg Acg cg 1 Acg (P (A =1)) i,t (1 P (A =1)) i,t (G.1) · i,t i,t N m 1 m 1 mij,t 1 (P (m =1)) ij,t (1 P (m =1)) ij,t .d⌃.d↵. · ij,t ij,t j=i+1 Y ⇥ ⇤ The log-likelihood is calculated by taking log of the likelihood

+ T + N 1 1 cg Apa pa 1 Apa LL =log (P (A =1)) i,t (1 P (A =1)) i,t i,t i,t t=1 i=1 Z1 Y Z1 Y cg Acg cg 1 Acg (P (A =1)) i,t (1 P (A =1)) i,t · i,t i,t N m 1 m 1 mij,t 1 (P (m =1)) ij,t (1 P (m =1)) ij,t .d⌃.d↵. · ij,t ij,t j=i+1 Y ⇥ ⇤ (G.2)

First, I explain how I estimate the integral using simulation. First, for each user-time combination, I draw R random draws from standard normal distribution, which will be used as error terms in Cholesky decomposition of covariance of actions. Next, for each set of user-time draws out of the R draws, each probability is calculated at its drawn value and the

ﬁnal value of the expression inside the integral is calculates. Lastly, I take average over the

R ﬁnal values of each drawn set. Given the law of large numbers and the aforementioned process, the log-likelihood is given by

131 R T N 1 cg Apa pa 1 Apa LL =log (P (A =1)) i,t (1 P (A =1)) i,t R i,t i,t r=1 t=1 i=1 X h Y Y cg Acg cg 1 Acg (P (A =1)) i,t (1 P (A =1)) i,t · i,t i,t N mij,t 1 mij,t 1 mij,t 1 (P (mij,t =1)) (1 P (mij,t =1)) · j=i+1 ir Y ⇥ ⇤ R = log R +log Q , r r=1 X (G.3) where the calculated values of the expression inside integration for each r =1...R are denoted as Q1 ...QR. Since the number of probabilities being multiplied in each drawn set N(N 1) is large (N N )andeachprobabilityisasmallnumber,theﬁnalvaluecalculated ⇤ ⇤ 2 for each set will be extremely small, and most likely not processed properly by computer. In order to bypass this limitation, I use the following transformation

N N (log ai log a0) log ai =logaa +log 1+ e . (G.4) i=0 i=1 ! Thus I can transit theX log into the sum, i.e. X

R (log Qr log Q1) LL = log R +logQ +log 1+ e , (G.5) 1 r=2 where X

log Q =

T N Apa log(P (Acg =1))+(1 Apa) log(1 P (Apa =1)) i,t · i,t i,t · i,t t=1 i=1 X X  +Acg log(P (Acg =1))+(1 Acg ) log(1 P (Acg =1)) i,t · i,t i,t · i,t N

+ (1 mij,t 1) [mij,t log(P (mij,t =1))+(1 mij,t) log(1 P (mij,t =1))] . · · · j=i+1 X ⇥ (G.6)⇤

132 REFERENCES

Aarseth, E. (2006). The culture and business of cross-media productions. Popular communication 4 (3), 203–211.

Ameri, M., E. Honka, and Y. Xie (2016). Word-of-mouth, observational learning, and product adoption: Evidence from an anime platform. Working Paper.

Aral, S., L. Muchnik, and A. Sundararajan (2009). Distinguishing inﬂuence-based contagion from homophily-driven di↵usion in dynamic networks. Proceedings of the National Academy of Sciences 106 (51), 21544–21549.

Aral, S., L. Muchnik, and A. Sundararajan (2013). Engineering social contagions: Optimal network seeding in the presence of homophily. Network Science 1 (02), 125–153.

Aral, S. and D. Walker (2011). Creating social contagion through viral product design: A randomized trial of peer inﬂuence in networks. Management Science 57 (9), 1623–1639.

Badev, A. (2013). Discrete games in endogenous networks: Theory and policy. Working Paper.

Bandiera, O. and I. Rasul (2006). Social networks and technology adoption in northern mozambique. The Economic Journal 116 (514), 869–902.

Banerjee, A. V. (1992). A simple model of herd behavior. The Quarterly Journal of Eco- nomics 107 (3), 797–817.

Bikhchandani, S., D. Hirshleifer, and I. Welch (1992). A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy 100 (5), 992–1026.

Bogu˜n´a, M., R. Pastor-Satorras, A. D´ıaz-Guilera, and A. Arenas (2004). Models of social networks based on social distance attachment. Physical Review E 70 (5), 056122.

Bollinger, B. and K. Gillingham (2012). Peer e↵ects in the di↵usion of solar photovoltaic panels. Marketing Science 31 (6), 900–912.

Bowden, J. L.-H. (2009). The process of customer engagement: A conceptual framework. Journal of Marketing Theory and Practice 17 (1), 63–74.

Brakus, J. J., B. H. Schmitt, and L. Zarantonello (2009). Brand experience: What is it? how is it measured? does it a↵ect loyalty? Journal of Marketing 73 (3), 52–68.

Bramoull´e, Y., H. Djebbari, and B. Fortin (2009). Identiﬁcation of peer e↵ects through social networks. Journal of Econometrics 150 (1), 41–55.

133 Brandtzæg, P. B. and J. Heim (2009). Why people use social networking sites. In Interna- tional Conference on Online Communities and Social Computing,pp.143–152.Springer.

Brodie, R. J., L. D. Hollebeek, B. Juric, and A. Ilic (2011). Customer engagement: conceptual domain, fundamental propositions, and implications for research. Journal of Service Research 14 (3), 252–271.

Brown, J. J. and P. H. Reingen (1987). Social ties and word-of-mouth referral behavior. Journal of Consumer Research 14 (3), 350–362.

Cai, H., Y. Chen, and H. Fang (2009). Observational learning: Evidence from a randomized natural ﬁeld experiment. Technical Report 3.

Calder, B. J., E. C. Malthouse, and U. Schaedel (2009). An experimental study of the relationship between online engagement and advertising e↵ectiveness. Journal of Interactive Marketing 23 (4), 321–331.

Carrell, S., R. Fullerton, and J. West (2009). Does your cohort matter? measuring peer e↵ects in college achievement. Journal of Labor Economics 27 (3), 439–464.

Cha, J. (2013). Predictors of television and online video platform use: A coexistence model of old and new video platforms. Telematics and Informatics 30 (4), 296–310.

Cha, J. and S. M. Chan-Olmsted (2012). Substitutability between online video platforms and television. Journalism & Mass Communication Quarterly 89 (2), 261–278.

Chen, Y., Q. Wang, and J. Xie (2011). Online social interactions: A natural experiment on word of mouth versus observational learning. Journal of Marketing Research 48 (2), 238–254.

Chevalier, J. A. and D. Mayzlin (2006). The e↵ect of word of mouth on sales: Online book reviews. Journal of Marketing Research 43 (3), 345–354.

Chintagunta, P. K., S. Gopinath, and S. Venkataraman (2010). The e↵ects of online user reviews on movie box oce performance: Accounting for sequential rollout and aggregation across local markets. Marketing Science 29 (5), 944–957.

Chou, T.-J. and C.-C. Ting (2003). The role of ﬂow experience in cyber-game addiction. CyberPsychology & Behavior 6 (6), 663–675.

Christakis, N. and J. Fowler (2013). Social contagion theory: examining dynamic social networks and human behavior. Statistics in Medicine 32 (4), 556–577.

Christakis, N., J. Fowler, G. Imbens, and K. Kalyanaraman (2010, May). An empirical model for strategic network formation. Working Paper.

134 Claussen, J., B. Engelst¨atter, and M. Ward (2014). Susceptibility and inﬂuence in social media word-of-mouth. ZEW-Centre for European Economic Research Discussion Paper (14- 129).

Csikszentmihalyi, M. (1997). Finding ﬂow: The psychology of engagement with everyday life.BasicBooks.

De Giorgi, G., M. Pellizzari, and S. Redaelli (2010). Identiﬁcation of social interactions through partially overlapping peer groups. American Economic Journal: Applied Eco- nomics 2 (2), 241–275.

Dellarocas, C., Z. Xiaoquan, and N. F. Awad (2007). Exploring the value of online product reviews in forecasting sales: The case of motion pictures. Journal of Interactive Market- ing 21 (4), 23–45.

Devasagayam, R. (2014). Media bingeing: A qualitative study of psychological inﬂuences. Proceedings of the Marketing Management Association,40.

Duan, W., B. Gu, and A. Whinston (2009). Informational cascades and software adoption on the internet: An empirical investigation. MIS Quarterly 33 (1), 23–48.

Duan, W., B. Gu, and A. B. Whinston (2008). Do online reviews matter? an empirical investigation of panel data. Decision Support Systems 45 (4), 1007–1016.

Fafchamps, M., M. Soderbom, and M. vanden Boogaart (2016, May). Adoption with social learning and network externalities. Working Paper 22282, National Bureau of Economic Research.

Garrahan, M. (2014). The rise and rise of the hollywood ﬁlm franchise.

Godes, D. and D. Mayzlin (2004). Using online conversations to study word-of-mouth communication. Marketing Science 23 (4), 545–560.

Gonzales, D. (2014). A look back at the rise of the hollywood mega franchise.

Grøntved, A. and F. B. Hu (2011). Television viewing and risk of type 2 diabetes, cardio- vascular disease, and all-cause mortality: a meta-analysis. Jama 305 (23), 2448–2455.

Hanaki, N., A. Peterhansl, P. S. Dodds, and D. J. Watts (2007). Cooperation in evolving social networks. Management Science 53 (7), 1036–1050.

Hartmann, W., P. Manchanda, H. Nair, M. Bothner, P. Dodds, D. Godes, K. Hosanagar, and C. Tucker (2008). Modeling social interactions: Identiﬁcation, empirical methods and policy implications. Marketing Letters 19 (3-4), 287–304.

135 Heckman, J. J. (1978). Dummy endogenous variables in a simultaneous equation system. Econometrica 46 (4), 931–959.

Hernandez, B. A. (2014). All ‘breaking bad’ episodes are now on netﬂix.

Herzenstein, M., U. M. Dholakia, and R. L. Andrews (2011). Strategic herding behavior in peer-to-peer loan auctions. Journal of Interactive Marketing 25 (1), 27 – 36.

Hinz, O., B. Skiera, C. Barrot, and J. U. Becker (2011). Seeding strategies for viral marketing: An empirical comparison. Journal of Marketing 75 (6), 55–71.

Ho↵, P. D., A. E. Raftery, and M. S. Handcock (2002). Latent space approaches to social network analysis. Journal of the American Statistical Association 97 (460), 1090–1098.

Ho↵man, D. L. and T. P. Novak (1996). Marketing in hypermedia computer-mediated environments: Conceptual foundations. Journal of Marketing 60 (3), 50–68.

Hoyer, W. D., R. Chandy, M. Dorotic, M. Kra↵t, and S. S. Singh (2010). Consumer cocre- ation in new product development. Journal of Service Research 13 (3), 283–296.

Iyengar, R., C. V. den Bulte, and T. W. Valente (2011). Opinion leadership and social contagion in new product di↵usion. Marketing Science 30 (2), 195–212.

Jackson, M. (2008). Social and Economic Networks. Princeton University Press.

Jenner, M. (2015). Binge-watching: Video-on-demand, quality tv and mainstreaming fandom. International Journal of Cultural Studies.

Katona, Z. (2013). Competing for inﬂuencers in a social network. Working Paper.

Katona, Z. and T. F. M´ori (2006). A new class of scale free random graphs. Statistics & Probability Letters 76 (15), 1587–1593.

Katona, Z., P. P. Zubcsek, and M. Sarvary (2011). Network e↵ects and personal inﬂuences: The di↵usion of an online social network. Journal of Marketing Research 48 (3), 425–443.

Kozinets, R. V. (1999). E-tribalized marketing?: The strategic implications of virtual communities of consumption. European Management Journal 17 (3), 252–264.

Kubey, R. and M. Csikszentmihalyi (2002). Television addiction. Scientiﬁc American 286 (2), 74–81.

Lampe, C. A., N. Ellison, and C. Steinﬁeld (2007). A familiar face (book): proﬁle elements as signals in an online social network. In Proceedings of the SIGCHI conference on Human factors in computing systems,pp.435–444.ACM.

136 Langlois, J. H., L. Kalakanis, A. J. Rubenstein, A. Larson, M. Hallam, and M. Smoot (2000). Maxims or myths of beauty? a meta-analytic and theoretical review. Psychological Bulletin 126 (3), 390–423.

Lee, L.-f. (2007). Identiﬁcation and estimation of econometric models with group interactions, contextual factors and ﬁxed e↵ects. Journal of Econometrics 140 (2), 333–374.

Lee, L.-f., X. Liu, and X. Lin (2010). Speciﬁcation and estimation of social interaction models with network structures. The Econometrics Journal 13 (2), 145–176.

Li, X. and L. M. Hitt (2008). Self-selection and information role of online product reviews. Information Systems Research 19 (4), 456–474.

Liu, Y. (2006). Word of mouth for movies: Its dynamics and impact on box oce revenue. Journal of Marketing 70 (3), 74–89.

Lovett, M. J. and R. Staelin (2016). The role of paid and earned media in building entertainment brands: Reminding, informing, and enhancing enjoyment. Marketing Science 35 (1), 142–157.

Ma, L., R. Krishnan, and A. Montgomery (2014). Latent homophily or social inﬂuence? an empirical analysis of purchase within a social network. Management Science 61 (2), 454–473.

Maddala, G. S. (1983). Limited-dependent and qualitative variables in econometrics.Cam- bridge university press.

Manchanda, P., Y. Xie, and N. Youn (2008). The role of targeted communication and contagion in product adoption. Marketing Science 27 (6), 961–976.

Manski, C. (1993). Identiﬁcation of endogenous social e↵ects: The reﬂection problem. The Review of Economic Studies 60 (3), 531–542.

MarketCast (2013). Marathon tv: How binge-viewing is changing the way we watch.

Matrix, S. (2014). The netﬂix e↵ect: Teens, binge watching, and on-demand digital media trends. Jeunesse: Young People, Texts, Cultures 6 (1), 119–138.

McCulloch, R. and P. E. Rossi (1994). An exact likelihood analysis of the multinomial probit model. Journal of Econometrics 64 (1), 207–240.

Mele, A. (2017). A structural model of dense network formation. Econometrica 85 (3), 825–850.

Merriam-Webster.com (2017). Binge-watching, https://www.merriam-webster.com/ dictionary/binge-watch.

137 Moe, W. and M. Trusov (2011). The value of social dynamics in online product ratings forums. Journal of Marketing Research 48 (3), 444–456.

Mollen, A. and H. Wilson (2010). Engagement, telepresence and interactivity in online consumer experience: Reconciling scholastic and managerial perspectives. Journal of Business Research 63 (9–10), 919–925.

Nair, H. S., P. Manchanda, and T. Bhatia (2010). Asymmetric social interactions in physician prescription behavior: The role of opinion leaders. Journal of Marketing Research 47 (5), 883–895.

Nambisan, S. and P. Nambisan (2008). How to proﬁt from a better ‘virtual customer environment’. MIT Sloan Management Review 49 (3), 53–61.

Netflix (2013). Binge watching is the new normal, https://media.netflix.com/en/press- releases/netflix-declares-binge-watching-is-the-new-normal.

Pena, L. L. (2015). Breaking binge: exploring the e↵ects of binge watching on television viewer reception. Ph. D. thesis, Syracuse University.

Pittman, M. and K. Sheehan (2015). Sprinting a media marathon: Uses and gratiﬁcations of binge-watching television through netﬂix. First Monday 20 (10).

Sacerdote, B. (2001). Peer e↵ects with random assignment: Results for dartmouth room- mates. The Quarterly Journal of Economics 116 (2), 681–704.

Schweidel, D. A. and W. W. Moe (2016). Binge watching and advertising. Journal of Marketing 80 (5), 1–19.

Shalizi, C. R. and A. Thomas (2011). Homophily and contagion are generically confounded in observational social network studies. Sociological Methods & Research 40 (2), 211–239.

Shannon-Missal, L. (2013). Americans taking advantage of ability to watch tv on their own schedules.

Shriver, S. K., H. S. Nair, and R. Hofstetter (2013). Social ties and user-generated content: Evidence from an online social network. Management Science 59 (6), 1425–1443.

Snijders, T., J. Koskinen, and M. Schweinberger (2010, 6). Maximum likelihood estimation for social network dynamics. The Annals of Applied Statistics 4 (2), 567–588.

Snijders, T., C. Steglich, and M. Schweinberger (2007). Modeling the coevolution of networks and behavior, pp. 41–71. Lawrence Erlbaum Associates Publishers.

Snijders, T., G. van de Bunt, and C. Steglich (2010). Introduction to stochastic actor-based models for network dynamics. Social Networks 32 (1), 44–60.

138 Statistica (2016a). Statistics and facts about the ﬁlm industry.

Statistica (2016b). Value of the global entertainment and media market from 2011 to 2020.

Stewart, C., T. Cockerill, I. Foster, D. Hancock, N. Merchant, E. Skidmore, D. Stanzione, J. Taylor, S. Tuecke, G. Turner, M. Vaughn, , and N. Ga↵ney (2015). Jetstream: a self- provisioned, scalable science and engineering cloud environment. In XSEDE Conference: Scientiﬁc Advancements Enabled by Enhanced Cyberinfrastructure.

Sun, M., X. M. Zhang, and F. Zhu (2012). To belong or to be di↵erent? evidence from a large-scale ﬁeld experiment in china. Technical report.

Sung, Y. H., E. Y. Kang, and W.-N. Lee (2015). A bad habit for your health? an exploration of psychological factors for binge watching behavior. 65th International Communication Association, San Juan.

TiVo (2015). Original streamed series top binge viewing survey for ﬁrst time.

Toivonen, R., L. Kovanen, M. Kivel¨a, J.-P. Onnela, J. Saram¨aki, and K. Kaski (2009). A comparative study of social network models: Network evolution models and nodal attribute models. Social Networks 31 (4), 240–254.

Tong, S. T., B. Van Der Heide, L. Langwell, and J. B. Walther (2008). Too much of a good thing? the relationship between number of friends and interpersonal impressions on facebook. Journal of Computer-Mediated Communication 13 (3), 531–549.

Toubia, O. and A. Stephen (2013). Intrinsic vs. image-related utility in social media: Why do people contribute content to twitter? Marketing Science 32 (3), 368–392.

Towns, J., T. Cockerill, M. Dahan, I. Foster, K. Gaither, A. Grimshaw, V. Hazlewood, S. Lathrop, D. Lifka, G. D. Peterson, R. Roskies, J. R. Scott, and N. Wilkins-Diehr (2014). Xsede: Accelerating scientiﬁc discovery. Computing in Science & Engineering 16 (5), 62– 74.

Trusov, M., A. V. Bodapati, and R. E. Bucklin (2010). Determining inﬂuential users in internet social networks. Journal of Marketing Research 47 (4), 643–658.

Tucker, C. (2008). Identifying formal and informal inﬂuence in technology adoption with network externalities. Management Science 54 (12), 2024–2038.

Van Doorn, J., K. N. Lemon, V. Mittal, S. Nass, D. Pick, P. Pirner, and P. C. Verhoef (2010). Customer engagement behavior: Theoretical foundations and research directions. Journal of Service Research 13 (3), 253–266.

139 Vivek, S. D., S. E. Beatty, and R. M. Morgan (2012). Customer engagement: Exploring customer relationships beyond purchase. Journal of Marketing Theory and Practice 20 (2), 122–146.

Wang, J., A. Aribarg, and Y. F. Atchad´e(2013). Modeling choice interdependence in a social network. Marketing Science 32 (6), 977–997.

Watson, G. B. and D. W. Johnson (1972). Social Psychology: Issues and insights.Lippincott Philadelphia.

Wei, Y. (2016). The similarity network of motion pictures. Working Paper.

Wilde, J. (2000). Identiﬁcation of multiple equation probit models with endogenous dummy regressors. Economics letters 69 (3), 309–312.

Wooldridge, J. M. (2010). Econometric analysis of cross section and panel data (2nd ed.). MIT press.

Zhang, J. (2010). The sound of silence: Observational learning in the u.s. kidney market. Marketing Science 29 (2), 315–335.

Zhang, J. and P. Liu (2012). Rational herding in microloan markets. Management Sci- ence 58 (5), 892–912.

Zhang, J., Y. Liu, and Y. Chen (2015). Social learning in networks of friends versus strangers. Marketing Science 34 (4), 573–589.

Zhang, Y., E. T. Bradlow, and D. S. Small (2013). New measures of clumpiness for incidence data. Journal of Applied Statistics 40 (11), 2533–2548.

Zhang, Y., E. T. Bradlow, and D. S. Small (2015). Predicting customer value using clumpiness: From rfm to rfmc. Marketing Science 34 (2), 195–208.

Zhang, Y. and D. Godes (2013). Learning from online social ties. Working Paper.

Zhu, F. and X. M. Zhang (2010). Impact of online consumer reviews on sales: The moderating role of product and consumer characteristics. Journal of Marketing 74 (2), 133–148.

140 BIOGRAPHICAL SKETCH

Mina Ameri was born in Tehran, Iran. After completing her schoolwork at Tizhooshan high school in 2006, Mina entered Sharif University in Tehran, Iran, and received a BS in Electrical Engineering in 2010. She then entered the Graduate School of Sharif University and obtained her MS with a concentration in marketing in 2013. Mina entered the Management Science Program of the Naveen Jindal School of Management at The University of Texas at Dallas in August 2013 to obtain her PhD in Management Science with a concentration in Marketing.

141 CURRICULUM VITAE

Mina Ameri

Naveen Jindal School of Management Website: www.minaameri.com University of Texas at Dallas Email: [email protected]

Education

Ph.D. in Marketing, Naveen Jindal School of Management, University of Texas at Dallas (expected May 2018) M.Sc. in Marketing, Sharif University of Technology (February 2013) B.Sc. in Electrical Engineering, Sharif University of Technology (September 2010) Research Interest Social Networks, Social Learning, User Engagement, Digital Entertainment Submitted Papers

Word-of-Mouth, Observational Learning, and Product Adoption: Evidence from an Anime Network, with E. Honka and Y. Xie, 2nd Round Revise & Resubmit at Marketing Science Best Paper Award Winner at Texas PhD Conference, 2016

The Eect of Binge-Watching on Media Franchise Engagement, with E. Honka and Y. Xie, Reject & Resubmit at Journal of Marketing Research Working Papers

A Structural Model of Network Dynamics: Tie Formation, Product Adoption, and Content Generation, with E. Honka and Y. Xie Work in Progress

Learning, Companionship, and Status: Why Do People Make Friends Online? Awards and Honors NSF cloud computing award of $8,000, 2017 Fellow, AMA-Sheth Doctoral Consortium, University of Iowa, 2017 Fellow, Marketing Science Doctoral Consortium, University of Southern California, 2017 Fellow, UH Marketing Doctoral Symposium, 2016 Best Paper Award Winner with $1,000 prize at Texas PhD Conference, 2016 Graduate Research Fellowship, University of Texas at Dallas, 2016-2017 Doctoral Student Scholarship, Jindal School of Management, University of Texas at Dallas, 2013-current Recognized as Exceptional Talent, Iran’s National Elites Foundation and Iran’s Ministry of Science, Research and Technology, 2010 Ranked 20th among more than 50,000 participants, Iranian National University Entrance Exam for graduate programs, 2010 Ranked 41st among more than 1 million participants, Iranian National University Entrance Exam for undergraduate programs, 2006 Conference Presentations “The Eect of Binge-Watching on Media Franchise Engagement” INFORMS Marketing Science Conference (ISMS) June, 2017

“A Structural Model of Network Dynamics: Tie Formation, Product Adoption, and Content Generation” Texas Marketing Faculty Research Colloquium March, 2017

“Word-of-Mouth, Observational Learning, and Product Adoption: Evidence from an Anime Network” UH Marketing Doctoral Symposium April, 2016 Texas PhD Conference March, 2016

“New Product Growth in Multi-Generation Multi-Brand Markets” Invited Discussant, UT Dallas FORMS Conference February, 2016

Computer Skills

Python, R, Matlab, C/C++, Perl, SAS, Stata, VB Conference Participation and Training

UT Dallas Frontiers of Research in Marketing Science (FORMS) 2014, 2015, 2016, 2017 Summer Institute in Competitive Strategy (SICS) 2016 Quantitative Marketing and Economics (QME) 2014, 2016 Quantitative Marketing and Structural Econometrics Workshop 2015 Teaching Experience

Instructor: PhD Seminar on Web Scraping May, 2017 Guest Lecturer at UCLA Anderson School of Management

PhD Workshop on Web Scraping at UT Dallas April, 2017

Principles of Marketing (Undergraduate) Fall 2016 Class size: 49 Course Evaluation: 4.6/5

Teaching Assistant: Applied Econometrics: Fall 2015 Introduction to Sales: Summer 2015, Summer 2017 Brand Management: Spring 2015 Advertising and Promotional Strategies: Spring 2015 Marketing Research: Spring 2014, Fall 2014 Interactive and Digital Marketing: Summer 2014 Database Marketing: Spring 2014 Principles of Marketing: Fall 2013, Fall 2015, Summer 2017 Related Coursework Marketing Special Topics in Marketing - Research Design Brian Ratchford Special Topics in Marketing - Consumer Choice Models Brian Ratchford Special Topics in Marketing - Industrial Organization Dmitry Kuksov Special Topics in Marketing - Data Analysis B.P.S. Murthi Special Topics in Marketing - Consumer Search Models Brian Ratchford Special Topics in Marketing - Industrial Organization Ram Rao Special Topics in Marketing - Market Design Ernan Haruvy Special Topics in Marketing - Bayesian Dynamic Models Norris Bruce

Statistics Approaches to Statistical Inference John Wiorkowski Probability and Stochastic Processes Shun-Chen Niu

Economics Econometrics I Donggyu Sul Econometrics II Dong Li Special Topics in Econometric and Spatial Analysis Dong Li Game Theory Gary Bolton Advanced Managerial Economics Kyle Hyndman

References

Elisabeth Honka Ying Xie Assistant Professor of Marketing Associate Professor of Marketing Anderson School of Management Naveen Jindal School of Management University of California Los Angelos University of Texas at Dallas 110 Westwood Plaza 800 West Campbell Rd Los Angeles, CA 90024 Richardson, Tx, 75080 Email: [email protected] Email: [email protected] Phone: +1 (310) 825-0296 Phone: +1 (972) 883-5839 Brian Ratchford Charles and Nancy Davidson Chair in Mar- keting Naveen Jindal School of Management University of Texas at Dallas 800 West Campbell Rd Richardson, Tx, 75080 Email: [email protected] Phone: +1 (972) 883-5975