Binge watching and advertising

In this research, the authors decompose users’ viewing behavior into 1 whether the user continues the viewing session after each episode viewed, 2 whether the next episode viewed is from

David A Schweidel & Wendy W Moe

Binge Watching and Advertising How users consume media has shifted dramatically as viewers migrate from traditional broadcast channels toward online channels Rather than following the schedule dictated by television networks and consuming one episode of a series each week, many viewers now engage in binge watching, which involves consuming several episodes of the same series in a condensed period of time In this research, the authors decompose users’ viewing behavior into (1) whether the user continues the viewing session after each episode viewed, (2) whether the next episode viewed is from the same or a different series, and (3) the time elapsed between sessions Applying this modeling framework to data provided by Hulu.com, a popular online provider of broadcast and cable television shows, the authors examine the drivers of binge watching behavior, distinguishing between user-level traits and states determined by previously viewed content The authors simultaneously investigate users’ response to advertisements Many online video providers support their services with advertising revenue; thus, understanding how users respond to advertisements and how advertising affects subsequent viewing is of paramount importance to both advertisers and online video providers The results of the study reveal that advertising responsiveness differs between bingers and nonbingers and that it changes over the course of online viewing sessions The authors discuss the implications of their results for advertisers and online video platforms

Keywords: binge watching, online streaming video, digital advertising, digital media consumption

Media consumption has changed dramatically in recent

years Viewers have been moving away from

watching traditional broadcast channels and toward

online video consumption to gain more control over their

media consumption Traditional media consumption, whereby

viewers watch shows according to the schedule and sequence

in which the networks broadcast them, has gradually given

way to viewers determining their own viewing schedule

through digital video recorders or on-demand

program-ming (Littleton 2014) As a result of these trends, new

patterns of media consumption have emerged

Rather than consuming one episode of a series each week

in accordance with a typical television schedule, viewers may

opt to view several episodes of a single series in immediate

succession Surveys have revealed that a majority of

con-sumers prefer to watch multiple episodes of their favorite

programs in a single sitting (Pomerantz 2013) A Nielsen

(2013) studyfinds that 88% of Netflix users and 70% of Hulu

Plus users reported watching at least three episodes of the

same program in one day

In addition to individuals consuming more content, they

report doing so in a condensed period of time According to a

survey conducted by Netflix and Harris Interactive in 2013,

61% of adults who stream television shows at least once a

week reported that they regularly engage in“binge watching”

sessions that consist of two to three episodes of a single

television series in one sitting, with nearly three-quarters of

respondents having positive feelings about binge watching (Netflix 2013) In its 2014 Digital Democracy Survey, Deloitte reports that 31% of respondents engaged in binge watching at least once a week, with more than 40% of respondents age

14–25 engaging in the behavior weekly (Deloitte 2015) What is binge watching?1The Digital Democracy Survey

defined the activity as “watching three or more episodes of a

TV series in one sitting” (Deloitte 2015) Meanwhile, in the survey conducted by Netflix and Harris Interactive, nearly three-quarters of respondents defined binge watching as

“watching between 2–6 episodes of the same TV show in one sitting” (Netflix 2013) These studies focus on viewing within a single viewing session, but Netflix further reports that for a particular serialized drama, 25% of viewersfinished the 13-episode season within two days, and almost 50% did

so within one week (Jurgensen 2013) A similar pattern is also reported for a sitcom Across these reports, binge watching is characterized by two common elements First, there is a heavy rate of consumption, which may occur within a single session or across multiple sessions that occur within a short period of time Second, a key feature that distinguishes binge watching from heavy usage is that binge watching is char-acterized by consuming multiple episodes of the same series These two characteristics are consistent with the definition for

“binge watching” that Oxford Dictionaries added to its online version in 2014 (Oxford Dictionaries 2014, 2016):“watching multiple episodes of (a television program) in rapid suc-cession, typically by means of DVDs or digital streaming.” In this research, we define “binge watching” as the consumption

of multiple episodes of a television series in a short period

of time

David A Schweidel is Associate Professor of Marketing, Goizueta Business

School, Emory University (e-mail: dschweidel@emory.edu) Wendy W Moe

is Professor of Marketing, Robert H Smith School of Business, University of

Maryland (e-mail: wmoe@rhsmith.umd.edu) Rajkumar Venkatesan served

as area editor for this article 1We use the term“binge watching” to refer to the activity and

“bingers” to refer to those individuals who engage in the activity

Consistent with the reports noted previously, Google

Trends reveals a sharp increase in searches for “binge

watching” beginning in 2013, as illustrated in Figure 1

Although reports have acknowledged the trend toward binge

watching, we have little understanding of its implications

for online video platforms Many online video services are

supported by advertising revenues, as is the case in our

empirical context of Hulu.com But if viewers are immersed

in binge watching, are advertisements still effective?

Moreover, do the advertisements affect users’ subsequent

viewing behaviors? Elberse and Gupta (2009) report that

advertising on Hulu.com was more effective than advertising

on broadcast or cable television Yet the analysis does not

distinguish between users’ responsiveness to advertising

when users are engaged in binge watching and when they are

not If users who are binge watching are less responsive to

advertising, this may give advertising-supported online video

platforms pause in terms of encouraging such behavior

In this research, we propose a model of viewing behavior

and advertising response and apply it to data from Hulu.com,

a popular online video platform Our modeling framework

decomposes users’ viewing behavior into (1) the decision to

continue the viewing session after each episode, (2) whether

the next episode viewed is from the same or a different series,

and (3) the time elapsed between sessions, in an effort to

identify binge watching behavior and factors that affect this

behavior, including advertisements In addition to these three

components of viewing behavior, we simultaneously model users’ responsiveness to advertisements shown during epi-sodes This modeling approach allows us to examine how advertising is related to viewing behavior, in terms of the effect that binge watching has on advertising response, as well as how advertisements affect viewing behavior Our analysis provides empirical evidence in the context

of online video consumption that viewing begets more viewing (Kubey and Csikszentmihalyi 2002), suggesting that binge watching is at least in part a malleable behavior (in addition to being a user-specific tendency or trait) We also find that advertisements shown during a viewing session can deter binge watching behavior and in fact shorten the length of the viewing session Finally, wefind that when viewers engage in binge watching, they are less responsive

to advertising In particular, we show how advertising re-sponsiveness differs between users who have a propensity to engage in binge watching and those users who shift in and out

of binge watching states Thesefindings have significant implications for advertisers and online video platforms supported by advertising revenues

The remainder of this manuscript proceeds as follows

We next provide a review of related research We then de-scribe our data before presenting our modeling framework and model specification Finally, we present our results and conclude with a discussion of implications and directions for future work

FIGURE 1 Google Trend’s Index for “Binge Watching”

0

10

20

30

40

50

60

70

80

90

100

Week

Jan 2, 2011Mar 2, 2011May 2, 2011Jul 2, 2011Sep 2, 2011Nov 2, 2011Jan 2, 2012Mar 2, 2012May 2, 2012Jul 2, 2012Sep 2, 2012Nov 2, 2012Jan 2, 2013May 2, 2013Mar 2, 2013 Jul 2, 2013Sep 2, 2013Nov 2, 2013Jan 2, 2014Mar 2, 2014May 2, 2014Jul 2, 2014Sep 2, 2014Nov 2, 2014Jan 2, 2015Mar 2, 2015May 2, 2015Jul 2, 2015Sep 2, 2015Nov 2, 2015 Mar 2, 2016Jan 2, 2016

Related Research

In this section, we provide a brief review of the literature

related to binge watching We draw on multiple streams of

literature from behavioral, economic, and medical research

Our intention is not to engage in testing particular theories

using the observational data available to us Rather, we

provide this discussion to motivate our analysis of binge

watching behavior and our expectations for how the behavior

relates to advertising

What Is Binge Behavior?

The psychological and medical literature considers binge

behavior an addiction (e.g., Gold, Frost-Pineda, and Jacobs

2003), research into which has shown that individuals often

engage in such behaviors to escape reality Binge behavior,

in general, has been defined by psychological researchers

as an“excessive amount in a short time,” such as binge eating

or binge drinking (e.g., Heatherton and Baumeister 1991;

Leon et al 2007) This raises the question: Do such addictive

behaviors extend to television consumption?

Kubey and Csikszentmihalyi (2002) delve into this

ques-tion by examining the addictive nature of television and

comparing it to substance dependence The authors note that

electroencephalogram (EEG) studies of individuals watching

television have found that people“reported feeling relaxed

and passive” and reveal that they exhibited “less mental

stimulation” (p 76) In addition to proposing the association

of a relaxed feeling with viewing that continues throughout a

viewing session, the authors also contend that this association

is negatively reinforced by the stress that viewers experience

when the viewing session ends As Kubey and Csikszentmihalyi

(2002, p 77) note,“Viewing begets more viewing,” suggesting

that viewers exhibit a tendency to continue the viewing session

to maintain their current state of mind

In the online environment, this relates to the concept

of “flow” (e.g., Ghani and Deshpand´e 1994; Hoffman and

Novak 1996), which characterizes immersive experiences in

which the user is in a state of focused concentration, intrinsic

enjoyment, and time distortion Researchers have also linked

experiencingflow to addictive behaviors In the context of

video games, Chou and Ting (2003)find that individuals who

experience flow are more likely to become addicted They

also find evidence to suggest that experiencing flow is an

intermediary step through which repetitive behaviors

con-tribute to addictive behaviors

As Chou and Ting (2003) note, addictive behaviors have

been viewed in various ways depending on thefield of study

Economists have proposed the theory of rational addiction,

which posits that individuals who exhibit addictive behaviors

may be maximizing their utility and that past

consump-tion can have a substantial impact on the utility derived

from future consumption (e.g., Becker and Murphy 1988) In

the marketing literature, Gordon and Sun (2015) develop a

dynamic model of rational addiction to examine the impact

cigarette taxes on consumption behavior

Under rational addiction theory, binges can arise from

cyclical behavior (e.g., Becker and Murphy 1988; Dockner

and Feichtinger 1993) Using eating as an example, Becker

and Murphy (1988) describe individuals alternating between periods of overeating and dieting in order to enjoy consuming food while also maintaining their weight In the context of binge watching, we mayfind that users take longer to initiate a new viewing session after a binge experience because they may derive more utility from other activities

Advertising and Binge Watching Advertisements shown during a viewing session can be seen

as an interruption to the experience We can liken the effect to that of advertising interruptions during an online browsing session Previous studies have shown that online browsers frequently enter a state offlow (Hoffman and Novak 1996) Advertisements shown during these sessions interrupt the flow state and can adversely affect the browsing experience Along these lines, Moe (2006)finds that pop-up promotions that interrupt an online shopping session shorten the duration

of the session and encourage users to exit the site By the same token, we would expect that advertisements shown during a viewing session might interrupt the viewing experience and consequently contribute to an increase in viewers’ tendencies

to end the session

Binge watching behavior can also have an impact on advertising responsiveness As noted previously, research on binge and addiction behavior outside the context of binge watching has shown that users engage in addiction behaviors

as an escape from reality (e.g., Gold, Frost-Pineda, and Jacobs 2003) In other words, individuals engaged in a binge state are immersed in an alternate reality In the context

of binge watching, this alternate reality is created by the video content, and advertisements shown during these sessions can be seen as unwelcome reminders of the viewer’s true reality Thus, our expectation is that viewers engaged in binge watching will be less responsive to advertisements than viewers not engaged in binge watching because they prefer to remain immersed in the context of the series they are viewing

Modeling Framework

In this section, we conceptually describe our modeling framework before presenting the data and the methodological details of the model Advertising-supported platforms that provide streaming video content have an interest in two types

of behaviors of their users: viewing behavior and advertising responsiveness In this article, we simultaneously model viewing behavior and advertising responsiveness at the level

of the individual user while allowing the two to be related Our goal is to capture characteristics of viewing behavior that may indicate binge watching behaviors and relate those characteristics to how the user responds to advertising

To model viewing behavior, we consider users’ viewing decisions at the end of each episode they have viewed Consistent with prior research on live television viewing (e.g., Rust and Alpert 1984; Rust, Kamakura and Alpert 1992; Shachar and Emerson 2000), we decompose a viewing session into a series of choices made by users First, after each episode, we model users’ decisions to continue their viewing session by watching another episode (of any program) Second, we model users’ decisions to watch another episode of the same

program, an option not considered in studies of live television

viewing because this decision is facilitated by today’s

streaming video services Third, if a user decides to conclude

the current viewing session, we consider the time until the

user returns to the platform to begin a new viewing session

Finally, we simultaneously model the user’s response to any

advertising to which he or she is exposed Specifically, we

examine the number of advertisements on which a user clicks,

out of the total number of advertisements to which the user

was exposed during an episode, as a binomial process

Overall, our modeling framework allows us to examine

both how advertisements affect viewing and how viewing

behaviors affect the user’s response to advertisements In

modeling both advertising and viewing decisions, we

allow for heterogeneity across users and recognize that users’

tendencies for each behavior may be correlated In addition to

variation in users’ tendencies, we account for shifts in behavior

that reflect prior viewing and advertising responsiveness

Data

Data Description

The data for our empirical analysis consist of the video

viewing behavior of 9,873 registered users of Hulu.com from

February 28, 2009, to June 29, 2009.2While maintaining a

library including both movies and television programs, Hulu

com “had a brand promise that was clear and distinctive:

Hulu is where you go for network TV” (Hansell 2009) This

served as a point of differentiation compared with other

popular online video portals, such as YouTube, that were

populated primarily with user-generated content (e.g.,

Elberse and Gupta 2009) Due to this positioning,

dis-cussions have occurred in the popular press about Hulu

com’s impact on broadcast and cable television’s business

models (e.g., Learmonth 2009; Rose 2008, 2009) Following

its 2009 Super Bowl advertisement, online conversation

about Hulu increased more than 250% (Eshman 2009) In

April 2009, Hulu announced a deal to make content from

Disney available to Hulu users, including episodes of

prime-time hits such as Lost, Grey’s Anatomy, Desperate

House-wives, Ugly Betty, Samantha Who?, Scrubs, and Private

Practice, as well as content from ABC Family and Disney

Channel (Kilar 2009a) Thefive most popular shows on Hulu

in 2009 were Saturday Night Live, Family Guy, The Office,

The Simpsons, and Naruto Shippuden (a Japanese anime

series), and an episode of Family Guy was the most played

full episode (Kilar 2009b)

The data for each individual user consist of an event log

that indicates the videos viewed.3Each episode of a program

is divided into multiple segments The event log records the time at which each video segment began, as well as infor-mation about the video segment, including the title of the television series and episode, the season of the series to which the episode belongs, and the episode number within the season

In addition to the series and episodes that users viewed, the event log also contains information on the advertising to which users were exposed The advertising data include a time-stamp at which the advertisement was served to the user,

as well as the program and episode in which the advertise-ment aired Our advertising data also indicate whether an individual took action and clicked on the advertisement Over 1.1 million advertisement impressions were recorded in our data, with users clicking on 9,317 advertisements (.84% of the advertisements).4In Figures 2–4, we provide histograms that show the distribution of the number of episodes viewed

by users, the number of unique programs viewed by users, and the number of viewing sessions conducted by users, respectively

Although most users viewed several episodes, we find variation in the number of viewing sessions that users con-ducted While 25% of users conducted just one viewing session, approximately an equal proportion conducted ten or more viewing sessions We also see that the number of series that users viewed follows a bimodal distribution While 35%

of users viewed only one or two series, more than 25% of users viewed ten or more different series

We provide descriptive statistics based on users’ behav-iors across all viewing sessions in our data in the upper portion of Table 1 We define a viewing session as a period of video viewing separated by one hour or more of inactivity In the data, 9,873 users were responsible for 104,414 viewing sessions, an average of 10.58 sessions per user during the four-month data period We provide descriptive statistics about the composition of these sessions in the lower portion

of Table 1

Although Table 1 provides a summary of viewing behavior at the level of the session, such statistics do not shed light on the heterogeneity that exists across users They also

do not provide insight into the relationship between the volume of online video consumption and the content con-sumed To investigate these factors at the session level, we present the number of programs viewed conditional on the length of the session in Table 2

According to the session-level data presented in Table 1,

in at least 50% of sessions, viewers watched two or more episodes, and in at least 50% of the sessions, viewers con-strained their viewed episodes to a single series Additionally, the lower bound of the interquartile ranges presented in Table 2 is equal to one series, irrespective of the number of episodes viewed In other words, for each session length considered, at least 25% of sessions involved viewers watch-ing episodes from a swatch-ingle series These statistics suggest the prevalence of binge watching behavior in the data

2The data set employed in this study was previously employed by

Schwartz et al (2011) and Zhang, Bradlow, and Small (2013) We

refer interested readers to these studies for additional details of the

data We excluded data from 164 users who had sessions that

exceeded 24 hours in length, due to concerns about the veracity of

the data

3We use the terms“viewers” and “users” interchangeably

Typical of many panel data sets, our data set does not allow us to

distinguish among multiple individuals in the same household

4During the data period, all videos on the site could be viewed free

of charge, and the service was strictly ad-supported In June 2010, after the data period concluded, Hulu introduced the Hulu Plus subscription service

To further illustrate the prevalence of binge watching

be-havior in our data, we show the joint distribution of the number

of episodes and unique series viewed in a session in Table 3 We

see that 63.3% of viewing sessions consisted of a single series,

and 18.5% consisted of multiple episodes of a single series We

next consider the fraction of users who engaged in different types

of viewing sessions We consider three types of viewing

ses-sions: (1) single-episode sessions, (2) multiepisode sessions that

consist of episodes from a single series, and (3) multiseries

sessions that consist of episodes from multiple series Wefind

that 81.1% of users conducted at least one single-episode

ses-sion, 49.0% of users conducted at least one multiepisode viewing

session that consists of episodes from a single series, and 69.0%

users conducted at least one multiseries session

Taken together with Table 3, with nearly half of users viewing multiple episodes of a single program in a session, these statistics provide model-free evidence for the presence

of binge watching in our data While our exploratory analysis suggests that binge watching does occur, it does not enable

us to discern whether such behavior is driven by user-specific traits or by recent viewing behavior To disentangle these competing explanations, as well as to understand how this behavior is related to users’ advertising responsiveness, we develop a joint model of viewing behavior and advertising response We next describe the key variables in our empirical analysis before presenting our modeling framework Variable Specification

For our model, wefirst create a set of dependent variables that characterize the various components of users’ viewing and advertising response decisions These variables are user-level and time-varying from episode to episode The viewing decisions we are interested in reflect the length, variety of programming, and frequency of viewing sessions Because a viewing session is only observed when at least one episode is viewed, we characterize the length of a session according to a user’s decision of whether or not to continue the session after viewing an episode In other words, we construct a binary variable equal to 1 if the user views another episode and equal to 0 if the user chooses to end the viewing session (CONTINUE) To represent the variety of the viewing ses-sion, we again consider the user’s decision after viewing an episode Conditional on the user viewing another episode, we construct a binary variable equal to 1 if the next episode is from the same series as the episode just completed and equal

to 0 if it is from a different series (SAME) Finally, if the user chooses to end the viewing session, we then consider the frequency of viewing sessions by computing the time (in days) until the next viewing session (FREQUENCY) To model advertising click-through behavior, for each epi-sode, we count the number of advertisements on which a

FIGURE 3 Distribution of the Number of Series Viewed

0%

5%

10%

15%

20%

25%

30%

35%

40%

45%

50%

Series Viewed

FIGURE 4 Distribution of the Number of Viewing Sessions

0%

5%

10%

15%

20%

25%

30%

35%

40%

45%

50%

Viewing Sessions

FIGURE 2 Distribution of the Number of Episodes Viewed

0%

5%

10%

15%

20%

25%

30%

35%

40%

45%

50%

Episodes Viewed

user clicks (CLICKTHRU) out of the total number of

adver-tisements to which the user is exposed Taken together, this

yields the click-through rate for each episode

In addition, we construct a number of covariates that are

expected to influence the viewing and advertising

click-through decisions These covariates are designed to capture

the effects of temporal factors (time of day, day of week, etc.),

program variation (e.g., genre), and content to which a user

was previously exposed, including program content and

advertising We provide a list of the covariates employed in

our analysis in Table 4

First, we consider the effects of viewing variables that

capture previous viewing behavior Specifically, we

con-sider a breadth variable, which is measured as the number of

different series viewed in a session up until the point of the

behavioral decision being considered, and a depth variable,

which is measured as the number of episodes from the current

series viewed thus far in the viewing session We expect these

covariates to affect both the viewing decisions and

advertis-ing responsiveness We construct episode-by-episode breadth

and depth measures (BREADTH_EPISODE and DEPTH_

EPISODE) that are used to model within-session behaviors

(CONTINUE and SAME), as well as measures that summarize

breadth and depth within the previous session (BREADTH_ SESSION and DEPTH_SESSION) and that are used to model intersession durations (FREQUENCY) In addition to the breadth and depth measures, we include an indicator variable

to account for whether a user has viewed the current series in an earlier viewing session (PRIORVIEW)

Second, throughout these viewing sessions, users are exposed to advertisements Users’ exposures to advertisements vary with the amount of content they consume; that is, longer programs include more advertising than shorter programs

We construct a time-varying covariate, EXPOSURES, that represents the number of advertisements to which a user has been exposed in each episode Because viewers’ interactions with advertisements may affect their subsequent viewing decisions, we also consider the impact of the number advertisements on which the user has clicked in each episode, captured by the variable CLICKTHRU

Finally, we specify a number of control variables to capture the effects of content and temporal differences (time

of day, day of week, etc.) on both viewing and advertis-ing click-through decisions Drawadvertis-ing on the data provided

by Hulu.com, we identify 18 genres in our data and construct a series of 17 indicator variables (one for each genre, with action/

TABLE 1 Descriptive Statistics

By User

By Session

Intersession time (days) 3.73 9.84 86 [.05, 32.63] [.07, 7.93] [.33, 2.25] Notes: IQR = interquartile range, representing the range that de fines the middle 50% of observations.

TABLE 2 Unique Series Viewed, by Number of Episodes in the Session

Notes: IQR = interquartile range, representing the range that de fines the middle 50% of observations.

adventure being our baseline genre) We also construct a

covariate to represent whether the episode just viewed

was a seasonfinale (SEASON_FINALE) or a series finale

(SERIES_FINALE) Finally, we define a set of variables to

capture month, day-of-week, and time-of-day effects For day of

week, we differentiate between weekdays (Monday–Thursday),

Friday, Saturday, and Sunday, again constructing indicator

variables whereby weekdays are considered the baseline

For time of day, we divide the day into parts according to

Headen, Klompmaker, and Rust (1979) Specifically, we

construct indicator variables for early morning (7A.M.-10A.M.),

daytime (10A.M.-5P.M.), early fringe (5P.M.-8P.M.), and prime

time (8P.M.-11P.M.) The late fringe period of 11P.M.-7A.M.is

set as our baseline

Model

Model Specification Figure 5 provides a depiction of our modeling framework, including the user decisions we consider and the variables that

influence them Our model is made up of four decisions: (1) whether to continue the viewing session by viewing another episode (CONTINUE); (2) if the session is continued, whether to watch a video from the same or a different series (SAME); (3) how much time elapses until the next viewing session begins (FREQUENCY); and (4) whether to click on an advertisement (CLICKTHRU) We develop our model byfirst considering the intrasession viewing behaviors of whether to continue a viewing session and whether to watch the same series We then present TABLE 4

Variable Descriptions

BREADTH_EPISODEits Number of series previously viewed in session s by user i prior to video t; operationalized as

log(1+ number of series previously viewed) and mean-centered DEPTH_EPISODEits Number of episodes of the current series previously viewed in session s by user i prior to video

t; operationalized as log(1+ number of episodes previously viewed) and mean-centered BREADTH_SESSIONis Number of series viewed in session s by user i; operationalized as log(1+ number of series

viewed) and mean-centered DEPTH_SESSIONis The maximum number of episodes of a single series viewed by user i in session s;

operationalized as log(1+ number of episodes viewed) and mean-centered PRIORVIEWits Indicator variable for whether or not user i has previously viewed the same series as video t of

session s in an earlier viewing session EXPOSURESits Number of advertisements shown to user i in video t of session s

CLICKTHRUits Number of advertisements clicked by user i in video t of session s

SEASON_FINALEits Indicator variable for whether video t in session s viewed by user i is a seasonfinale SERIES_FINALEits Indicator variable for whether video t in session s viewed by user i is a seriesfinale MONTHits Month in which video t viewed by user i in session s began (March, April, May, June) DAYits Day of the week on which video t viewed by user i in session s began (Monday–Thursday,

Friday, Saturday, Sunday) DAYPARTits Day part in which video t viewed by user i in session s began (early morning, daytime, early

fringe, prime time, late fringe) GENREits Program genre of video t viewed by user i in session s (action/adventure, animation and

cartoons, comedy, drama, family, food and leisure, home and garden, horror and suspense, music, news and information, other, reality and game shows, sciencefiction, sports, talk and interviews, unknown, video games, web)

TABLE 3 Numbers of Episodes and Series Viewed, Across Sessions

Number of Series

Number of Episodes

the model component for intersession timing decisions Finally,

we describe our model for viewers’ advertising responsiveness

For user i who has viewed t videos in the current session s,

we model two binary decisions: (1) the decision to continue

session s by viewing another video (CONTINUE) and (2)

con-ditional on viewing another video, the decision to watch the

same series or a switch to a different program (SAME) As shown

in Figure 5, these decisions are affected by viewing variables

based on previous viewing activity, ad variables, and control

variables After user i views video t, we model the probability

with which the user continues the current session as pits:

logitðpitsÞ = ai+ q1DEPTH EPISODEits

+ q2DEPTH EPISODE2its

+ q3BREADTH EPISODEits

+ q4BREADTH EPISODE2its

+ q5PRIORVIEWits+ q6EXPOSURESits

+ q7CLICKTHRUits+ q8SEASON FINALEits

+ q9SERIES FINALEits+ 

26

j= 10

qjGENREits

+ 

29

j= 27

qjDAYits+ 

32

j= 30

qjMONTHits

+ 

36

j= 33

qjDAYPARTits, (1)

where ai is an individual-level intercept such that

ai= a + gi1, with gi1 a random effect with mean 0 that

captures variation across users We allow for nonlinear

effects of the viewing variables DEPTH_EPISODE and

BREADTH_EPISODE If we consider binge watching an addictive behavior whereby“viewing begets more viewing,”

as proposed by Kubey and Csikszentmihalyi (2002), we should expect DEPTH_EPISODE (captured by q1and q2)

to have a positive effect on pits Likewise, the impact of BREADTH_EPISODE (reflected by q3andq4) is expected

to be negative because continued viewing of a given pro-gram is more likely in such an addictive state in which BREADTH_EPISODE of viewing is low The coefficient q5 accounts for the effect of user i having viewed the current program prior to episode t of session s on the decision to continue the current viewing session The effects of ad vari-ables (EXPOSURES and CLICKTHRU) to which user i is exposed during episode t of session s on the decision to continue viewing session s are captured byq6andq7 If we assume that advertising breaks up theflow of a binge watching session and discourages further viewing, we should expect thatq6< 0 and q7< 0 The remaining coefficients (q8–q36) capture variation in the decision to continue the session that is related to control variables, including whether video t is a season (q8) or series (q9)finale, the genre viewed in video t (q10–q26), and day and time (q27–q36) at which video t is viewed

In the special case whereq = 0, then logit(pits)= aiand the length of the viewing session (in episodes) follows a shifted geometric distribution at the user level Following those who employ this individual-level model for discrete-time customer base analysis (e.g., Fader and Hardie 2009), we accommodate hetero-geneity across users We also allow the likelihood of continuing the viewing session to shift depending on the content that a user views

We employ a similar binary logit model for a user’s decision to continue viewing the same program, as opposed

to viewing a different program Conditional on user i

FIGURE 5 Depiction of Modeling Framework

Viewing Behavior

Continue Same Frequency

Do I continue or stop watching?

(pits)

Do I watch the same or a different series?

(q its )

How long before

I watch again?

(h(t))

Advertising Exposures

Advertising Clickthrough

Control Variables

• GENRE

• DAY

• MONTH

• DAYPART

• SEASON_FINALE

• SERIES_FINALE •• EXPOSURESCLICKTHRU

Viewing Variables

DEPTH_EPISODE DEPTH_EPISODE BREADTH_EPISODE BREADTH_EPISODE DEPTH_SESSION DEPTH_SESSION BREADTH_SESSION BREADTH_SESSION

•

•

•

•

• PRIORVIEW

Do I click on the ad? (r its )

Ad Variables

continuing viewing session s, we specify the probability with

which video t+ 1 is from the same series as video t as qits:

logitðqitsÞ = bi+ y1DEPTH EPISODEits

+ y2DEPTH EPISODE2its + y3BREADTH EPISODEits + y4BREADTH EPISODE2its + y5PRIORVIEWits+ y6EXPOSURESits + y7CLICKTHRUits

+ y8SEASON FINALEits + y9SERIES FINALEits

+ 

26

j= 10

yjGENREits+ 

29

j= 27

yjDAYits

+ 

32

j= 30

yjMONTHits+ 

36

j= 33

yjDAYPARTits, (2)

where the user-level intercept is specified as bi= b + gi2 As

in our specification of pits, our specification of qitsenables us

to distinguish the user’s general tendency to watch the same

program (throughgi2) from the impact of previously viewed

content (through y1–y5) If we assume that consuming

multiple episodes of the same series increases addictive

behavior, then we should expect to observe a tendency to

con-tinue viewing the same series as depth increases Likewise, if

limited breadth contributes to addictive viewing behavior focused

on a single series, then we are more likely to observe a tendency to

switch series as breadth increases The impact of ad variables

on this component of viewing behavior is reflected in y6and

y7 The coefficients y8-y36capture the effects of the control

variables that account for content and temporal differences

Whereas Equations 1 and 2 characterize the user decisions

within a single viewing session, our third model component

(FREQUENCY) looks across viewing sessions At the

con-clusion of viewing session s, we model the time until the start

of the next viewing session using a proportional hazard

model (e.g., Seetharaman and Chintagunta 2003) We

as-sume a Weibull distribution for the baseline hazard (e.g.,

Helsen and Schmittlein 1993; Schweidel, Fader, and Bradlow

2008; Seetharaman and Chintagunta 2003), which

accommo-dates increasing or decreasing hazards and nests the constant

exponential hazard process as a special case.5The

base-line Weibull hazard is given by

bðtÞ = luðltÞu-1

, (3)

where l > 0 and u > 0 The survival function can then be

written as a function of the hazard rate h(t), which is given by

h(t)= b(t)exp(Xis), where b(t) is the baseline Weibull hazard

rate and Xiscaptures the impact of covariates corresponding

to session s for user i As described in Equations 1 and 2 and

illustrated in Figure 5, we allow viewing variables, ad variables,

and control variables to affect how quickly a user returns to the website to begin a new viewing session The resulting survival function is then given by

SðtÞ = exp

-ðt 0

bðuÞexpðXisÞdu



= exp

-ðltÞu expðXisÞ, (4)

where

Xis= gi3+ w1DEPTH SESSIONis + w2DEPTH SESSION2is + w3BREADTH SESSIONis + w4BREADTH SESSION2is + w5EXPOSURESi $s + w6CLICKTHRUi $s + w7maxðSEASON FINALEi $sÞ + w8maxðSERIES FINALEi$sÞ + 

25

j= 9

wjGENREiNiss+ 

28

j= 26

+ 

31

j= 29

wjMONTHiNiss+ 

35

j= 32

wjDAYPARTiNiss, (5)

and Xiscaptures the impact of observed covariates and un-observed heterogeneity across users

The user-specific random effect gi3has a mean of 0 and captures unobserved differences across users As described

in Table 4, the measures of DEPTH_SESSIONisand BREADTH_ SESSIONisprovide measures of viewing depth and breadth that are calculated according to viewing behavior in session s For example,

if a user has viewed three episodes of one series and one episode of another, the maximum depth during the session is 3 If a user is addicted to a series and has watched many episodes of a given program in a session (i.e., DEPTH_SESSION is high), then we might expect the user to return to view more episodes in a relatively short amount of time, reflected by w1andw2 Alternatively, if users exhibiting binging behavior cycle between different activities (e.g., Becker and Murphy 1988), we might expect increased depth to increase the time until the next viewing session begins To account for the potential impact of ad variables on the time until user i begins session s+ 1, we aggregate the advertising exposure (w5) and advertisement clicks (w6) throughout session s by summing these variables across all videos that comprise the viewing session

We also account for the content viewed during session

s and the time of session s The coefficients w7and w8 account for the presence of a season or seriesfinale, re-spectively, in session s If an episode viewed in session s is

a seasonfinale, max(SEASON_FINALEits)= 1; otherwise, max(SEASON_FINALEits)= 0 We control for the genre of the final episode viewed by user i in session s (which we denote video Nis) to account for the most recently viewed content Similarly, we control for the time (day part, day of week, and month) at which video Nisis viewed

We use a binomial distribution to model the number of advertisements on which user i clicks in episode t of session s, according to the number of advertisements to which the user

is exposed We specify user i’s probability of clicking on

an advertisement during video t of session s, rits, using a

5We compared the proposed model that uses the Weibull distribution

as the baseline hazard with the model that employs the exponential

distribution as the baseline hazard Although we did notfind any

sub-stantive differences between the specifications, on the basis of the

log-marginal density, the model that uses the Weibull hazard betterfits the

data We therefore present the results corresponding to this model

binary logit model (e.g., Chatterjee, Hoffman, and Novak

2003; Dr`eze and Hussherr 2003; Hoban and Bucklin 2015;

Urban et al 2014), where rits is affected by the viewing

variables and control variables shown in Figure 5:

logitðritsÞ = di+ f1DEPTH EPISODEits

+ f2DEPTH EPISODE2its + f3BREADTH EPISODEits + f4BREADTH EPISODE2its + f5PRIORVIEWits + f6SEASON FINALEits + f7SERIES FINALEits

+ 

24

j= 8

fjGENREits+ 

27

j= 25

fjDAYits

+ 

30

j= 28

fjMONTHits+ 

34

j= 31

fjDAYPARTits, (6)

wheredi= d + gi4andgi4is a user-specific intercept; f1and

f2 capture the impact of the depth of viewing of the series

viewed in video t;f3andf4account for the breadth of viewing

that has occurred in the session prior to viewing video t; andf5

controls for the impact of prior exposure to the series In addition

to viewing variables, we also include control variables that

account for whether or not video t is a season (f6) or seriesfinale

(f7), the genre of video t (f8–f24), and when video t is viewed

(f25–f34).6If we assume that binge watching reduces viewers’

responsiveness to advertisements, we should anticipate that the

impact of depth on the click-through probability (captured byf1

andf2) is negative

To complete our model specification, Equation 7 provides the

joint likelihood function Let Yits= 1 when the next video that user i

chooses to view is from the same series as video t of session s, let

Yits= 2 when user i chooses to view a different program, and let

Yits= 3 when user i decides to end session s after viewing video t

Let disdenote the time between sessions s and s+ 1 for user i

Combining the intrasession viewing decisions (whether to continue

the session and whether to view the same series), episode-level

advertising response, and proportional hazard model of intersession

durations, the likelihood of user i’s behavior is given by

(7)

Li=



∏Si

s= 1



∏Nis t= 1

·ð1 - qitsÞ1 ðYits =2Þð1 - pitsÞ1 ðYits=3Þ

· ∏Si

s= 1

∏ Nis

t= 1

 EXPOSURESits CLICKTHRUits



its ð1 - ritsÞEXPOSURESits-CLICKTHRUits



·

∏

Si-1

s= 1

 S



dis+ 1 24



- SðdisÞ



SðT - diSiÞ

Thefirst line of Equation 7 captures intrasession viewing decisions to continue the current viewing session (with probability pits) and to view another episode of the same series, conditional on continuing the session (with probability

qits) If user i continues the session (Yits< 3), we employ a binary logit model for the decision to view the same or different content A user chooses to not continue the session after viewing episode Niswith a probability of 1- piNiss This episode-level behavior is modeled for each of the Siviewing sessions from user i The second line of Equation 7 accounts for user i’s response to advertising in video t of session s, using a binomial distribution with probability rits

The last line of Equation 7 models the intersession du-rations Because we do not know the exact time at which a viewing session ends, we treat the observed intersession times disas interval-censored data in which the contribution

of disto the likelihood Liis based on the difference between the survival function evaluated one hour (the cutoff used

to define the ends of sessions) after the end of the last episode

of session s and the survival function evaluated at the end

of the last episode of session s We also account for the right-censored duration observed between the end of our obser-vation period and the end of thefinal session observed from user i

We assume all users are heterogeneous across their viewing behavior and advertising responsiveness, with users’ tendencies being correlated We assume that users’ behav-ioral tendencies, reflected by the user-level random ef-fectsgi $, are distributed such thatgi $~ MVN(0,S) where 0

is a 4 · 1 vector of zeros and S is the covariance matrix The inclusion of user-specific random effects enables us to distinguish between unobserved heterogeneity that exists across users and state dependence arising from previously consumed content (e.g., breadth and depth of viewing) Neglecting to account for heterogeneity has also been shown

to yield misleading inferences (e.g., Hutchinson, Kamakura, and Lynch 2000; Roy, Chintagunta, and Haldar 1996) In addition to capturing differences across users for each of the behaviors we model, the correlated random effects also serve

to induce a correlation at the margin among the four com-ponents of our model in a manner consistent with prior research on brand choice (e.g., Bucklin and Gupta 1992), online behavior (e.g., Johnson et al 2004; Sismeiro and Bucklin 2004), and in-store behavior (e.g., Hui, Bradlow, and Fader 2009)

We assume an inverse Wishart prior for S and normal priors for thefixed effects.7We obtain posterior draws from the joint posterior using Markov chain Monte Carlo (MCMC) methods We ran two independent chains with different starting values until convergence For each chain, we removed thefirst 10,000 iterations as a burn-in period and then used the subsequent 5,000 iterations for inference

Is Binge Watching a Trait or a State?

Popular reports of binge watching confound two distinct explanations for the activity One explanation assumes that

6Note that DEPTH_EPISODEitsand BREADTH_EPISODEits, as

described in Table 4, are calculated based on user i’s viewing behavior

in session s prior to video t Thus, we avoid problems with simultaneity

7Becausel > 0 and u > 0, we assume diffuse normal priors for log (l) and log(u)