You are on page 1of 23

Article

Journal of Marketing Research


2021, Vol. 58(3) 417-438
Happiness Begets Money: Emotion ª American Marketing Association 2021
Article reuse guidelines:

and Engagement in Live Streaming sagepub.com/journals-permissions


DOI: 10.1177/00222437211002477
journals.sagepub.com/home/mrj

Yan Lin, Dai Yao, and Xingyu Chen

Abstract
Live streaming offers an unprecedented opportunity for content creators (broadcasters) to deliver their content to consumers
(viewers) in real time. In a live stream, viewers may send virtual gifts (tips) to the broadcaster and engage with likes and chats free
of charge. These activities reflect viewers’ underlying emotion and are likely to be affected by the broadcaster’s emotion. This
article examines the role of emotion in interactive and dynamic business settings such as live streaming. To account for the
possibility that broadcaster emotion, viewer emotion, and viewer activities influence each other, the authors estimate a panel
vector autoregression model on data at the minute level from 1,450 live streams. The results suggest that a happier broadcaster
makes the audience happier and begets intensified viewer activities, in particular tips. In addition, broadcasters reciprocate viewer
engagement with more smiles. Further analyses suggest that these effects are pronounced only after a live stream has been active
for a while, and they manifest only in streams by broadcasters who have more experience, receive more tips, or are more popular
in past live streams. These results help platforms and broadcasters optimize marketing interventions such as broadcaster emotion
enhancement in live streaming and quantify the financial returns.

Keywords
digital marketing, emotion, engagement, live streaming, social media, text analysis, video analysis
Online supplement https://doi.org/10.1177/00222437211002477

Live streaming, as a new channel that enables users to interact the live stream to consume the content. During the content
with each other over the internet in real time, has been gaining consumption process, they also engage with the broadcaster
traction among major social media platforms. For example, in various ways that help shape the content generation process.
YouTube took a major step in April 2011 by launching live Some engagement activities are nonmonetary, such as chatting
streaming functionality to allow its partners to broadcast live via text with the broadcaster or other viewers in the stream and
content on the platform to its vast global audience (Garrahan sending likes to the stream. Another engagement activity is
2011). Twitter, which initiated microblogging, expanded its tipping, where viewers send the broadcaster virtual tokens or
business in May 2015 to live streaming with a dedicated app gifts they have purchased from the platform with real money.
called Periscope (Kharpal 2015). Facebook followed in August Viewers can choose when and how much to tip the broadcaster
2016, first experimenting with its Facebook Watch video at any point in time during a live stream, paralleling the pay-
streaming service among selected users and then providing it what-you-want (PWYW) business model, which has been
to all users in January 2017 (Bell 2016). In addition to these widely adopted by firms and rigorously scrutinized by aca-
social network titans, which had already accrued a sizable user demics (e.g., Chen, Koenigsberg, and Zhang 2017; Kim, Nat-
base before venturing into live streaming, other companies ter, and Spann 2009; Jung et al. 2014; Schmidt, Spann, and
have started businesses from scratch and developed into giants Zeithammer 2015). Viewers often engage in these monetary
in this market. A notable example is Twitch, which was and nonmonetary activities multiple times throughout their
launched as a spin-off of the now-defunct Justin.tv in June stays in a live stream. As major social platforms continue to
2011 to focus on video game live streaming and quickly
exploded. By February 2014, it had become the fourth-largest Yan Lin is Assistant Professor, Shenzhen Audencia Business School, Shenzhen
source of internet traffic in the United States, which led to its University, China (email: linyan@szu.edu.cn). Dai Yao is Assistant Professor,
NUS Business School, National University of Singapore, Singapore (email:
acquisition by Amazon in August 2014. dai.yao@nus.edu.sg). Xingyu Chen (corresponding author) is Associate
A typical live stream features one broadcaster who creates Professor, College of Management, Shenzhen University, China (email:
and delivers content to the audience in real time. Viewers join celine@szu.edu.cn).
418 Journal of Marketing Research 58(3)

embrace this technology-enabled new channel, and the busi- opposed to offline business contexts, live streaming as an
nesses using it are expected to prosper, it has become critical to online business facilitates more detailed data collection. For
understand the interactions and coevolution of viewer and example, the video clip of a live stream can be recorded and
broadcaster behaviors in such an interactive and dynamic efficiently analyzed to extract the emotion of the broadcaster
environment. using state-of-the-art machine learning algorithms (e.g., deep
The extant research on PWYW has usually studied its neural networks). We can also probe the viewers’ emotions by
implementation in offline business contexts, including restau- analyzing their chats in the live stream. These novel and
rants, theaters, shopping malls, and museums, and has exam- unstructured data enable us to link the emotions of content
ined various drivers for customers to make payments to product providers (broadcasters) to viewers’ payment decisions in the
or service providers under this pricing practice.1 Customers context of live streaming at the aggregate level, which other
may decide to pay nonzero prices due to their personal traits PWYW applications have not yet examined.
(e.g., altruism), habits, emotion, satisfaction, and loyalty In this research, we augment the PWYW literature by exam-
(Gneezy et al. 2010; Kim, Natter, and Spann 2009). Customers ining and quantifying the effect of emotion in interactive and
may also be motivated or incentivized to make payments by dynamic business contexts such as live streaming. Specifically,
fairness concerns (Chen, Koenigsberg, and Zhang 2017; Kim, we test whether broadcasters’ emotions influence viewers’
Natter, and Spann 2009), social norms (Feldhaus, Sobotta, and emotions, and vice versa. We further examine the effects of
Werner 2018; Jung et al. 2014; Jung, Perfecto, and Nelson both broadcaster and viewer emotions on viewer engagement,
2016), reference prices (Jung et al. 2014; Jung, Perfecto, and including both free (i.e., chatting and liking) and paid (i.e.,
Nelson 2016; Kim, Kaufmann, and Stegemann 2014; Park, tipping) activities and viewer participation (i.e., the number
Nam, and Lee 2017), and social distance between themselves of viewers). In addition, we investigate the effects of viewer
and the receiver (Kim, Kaufmann, and Stegemann 2014), engagement and participation on broadcaster emotion. To
among other reasons. Researchers have also begun to examine account for autocorrelation of the time series and the possibility
the application of PWYW in online settings. For example, that all these key variables are likely to influence each other,
Elberse and Bergsman (2008) studied the band Radiohead’s we employ a panel vector autoregression model (PVAR; for
use of PWYW for its 2007 album. Chen, Koenigsberg, and details, see Love and Zicchino [2006]) to capture the interde-
Zhang (2017) discussed online platforms using PWYW, such pendent and dynamic nature of this business context and to
as the Humble Bundle website using it to sell software. In quantify and disentangle both the immediate and relatively
addition, Spann, Stich, and Schmidt (2017) explored the pos- longer-lasting effects of one variable (e.g., broadcaster emo-
sibility of adopting PWYW in open access publishing.2 tion) on other variables.
Live streaming differs from existing PWYW applications in The PVAR model is estimated on data from a popular live
three major aspects. First, the detailed tipping behaviors of all streaming platform in China that ranked among the top 30 in
viewers are publicly observable in live streaming, whereas in
the market as of July 2019.3 The data comprise 1,450 live
other PWYW applications, the amount of payment is usually
streams collected over a two-week period in January and Feb-
private. The public announcement of tipping in live streaming
ruary 2018. For each live stream, we recorded the number of
makes it possible for viewers both to derive social image–based
viewers, amount of tips, chat details, and number of likes at the
utility and to free ride other viewers. For those who tip the
minute level. We also obtained video clips of all the streams,
broadcasters, the public announcement of their tips allows
which enabled us to transcribe the scripts of each broadcaster
them to be perceived by their peers as rich and generous. For
and recognize their facial expressions. Our main model con-
those who appreciate the performance, seeing that others have
cerns six key variables that are typical in a live stream and are
already tipped allows them to free ride other viewers’ tips
measured as follows: BCEmotion is the average of the broad-
without making their own contribution. Second, in addition
caster emotion across all the seconds in a minute, where the
to tipping, viewers can engage in nonmonetary activities, such
broadcaster’s emotion is calibrated on the basis of facial
as liking and chatting. These activities are influenced by and
expression in each second; ViewerEmotion is the viewer emo-
may influence broadcaster behaviors, giving rise to a dynamic
tion calibrated from all the viewer chats pooled together during
system in which multiple key variables coevolve together. To
measure the impact of one variable on another, we must a minute; AmountTip, LenCmt, and NumLikes are three vari-
account for all interactions among the variables rather than ables of viewer engagement activities in a minute that capture
examining each relationship individually. In addition, as the amount of tips, the count of words in the comments, and the
number of likes, respectively; and NumViewers is the number
of viewers recorded at the end of a minute.
1
Note that the literature on PWYW differs from that on tipping in service
marketing (which has been summarized well in Azar [2020]), as the former
3
examines the payment decisions customers make under PWYW, whereas the The platform ranked between 25th and 30th as of July 2019, as reported by
latter examines payment decisions on top of the normal prices that firms set for China Internet Weekly and eNet, both of which are large media outlets focused
their products or services. on internet business in China. Their report about the rankings is available at
2
For a good review of the literature on PWYW as well as other customer http://www.enet.com.cn/article/2019/0716/A20190716948154.html (in
participative pricing mechanisms, see Spann et al. (2018). Chinese).
Lin et al. 419

Our study yields interesting results regarding live streaming


and the application of PWYW to this business context. First,
we find that positive broadcaster emotion induces positive
viewer emotion, and similarly, positive viewer emotion
enhances broadcaster emotion, suggesting that the two parties
influence each other (Russell 2003), leading to emotional
empathy between them (Escalas and Stern 2003). Second, we
find that as the broadcaster becomes happier, all viewer activ-
ities, including tips, chats, and likes, intensify. Third, we find
that viewer emotion has positive effects on all viewer activities
as well. In general, viewers send more tips, supply more com-
ments, and press the “Like” button more often when they are
experiencing a happier mood. These results are consistent with
findings in the literature on emotional contagion in business
settings and its effects on either customer attitudes toward the
products (e.g., Howard and Gengler 2001) or behaviors such as
providing public services or donations (e.g., Liang, Chen, and
Lei 2016). Fourth, by using the PVAR model, we can quantify
the effects among key variables over a relatively longer period
(e.g., over 10 or 20 minutes). These relatively longer-lasting
effects on emotional contagion, customer attitudes, or customer
behaviors have not yet been investigated in the extant literature
to the best of our knowledge. For example, we find that broad-
caster emotion influences viewer behaviors, in particular
viewer tips, even after 20 minutes. Fifth, estimating the PVAR
model with rolling windows (e.g., Pauwels and Hanssens 2007)
in which fixed-length periods are used leads to the finding that Figure 1. A screenshot of an ongoing live stream.
these effects occur only after a live stream has been active for a
while and not from the very beginning. Finally, we find strong
enthrall the audience and garner new viewers. Platforms allow
broadcaster heterogeneity: only broadcasters who had more
viewers to converse using text but not voice, which augments
experience, received more tips, or were more popular in past
the content in the live stream. Meanwhile, viewers watch and
live streams can leverage their smiles to increase viewer tips.
consume the content that is continuously generated in the
The remainder of the article is structured as follows. We first
stream. Viewers do not pay an entry fee for a live stream but
introduce the live streaming setting briefly and relate our study
are encouraged to send virtual gifts (i.e., tips) to reward the
to several research streams in the literature. We then discuss the
broadcaster. In addition to this tipping behavior, viewers can
details of the data and methodology, present the main results,
also express their attitude or gratitude by sending likes into the
and conduct several robustness tests. Finally, we discuss the
stream. Figure 1 provides a snapshot of a typical ongoing live
implications of our study for practitioners as well as the limita-
stream from viewers’ perspective.
tions and potential directions for future research.

Research Background Relevant Literature


Our article is relevant to several streams of research. First, it is
The Live Streaming Setting related to the growing research on live streaming. A large body
In a typical live stream, a broadcaster sets up the video stream, of this literature focuses on understanding the motivational
which is essentially a virtual showroom hosted on a live factors of viewers’ various behaviors (e.g., watching, chatting,
streaming platform. All ongoing live streams are usually dis- tipping) in this context. A general conclusion from existing
played in chronological, alphabetical, or another order. Upon research is that users’ motivations are rather complex, includ-
seeing the stream, any platform user can join it free of charge, ing both intrinsic and extrinsic factors. For example, Hilvert-
which means that the user enters the virtual showroom and is Bruce et al. (2018) identified six motivations that explain
immediately exposed to whatever is happening inside. The viewer engagement activities: social interaction, sense of com-
broadcaster interacts with existing and new viewers in this munity, meeting new people, entertainment, seeking informa-
virtual space in real time. Viewers may interact with each other tion, and seeking support. Hu, Zhang, and Wang (2017) found
as well. The broadcaster generates various types of content in that viewers’ community identification with both broadcasters
line with their expertise (e.g., hosting talk shows, singing, dan- and their peer viewers is associated with their watching inten-
cing, playing musical instruments, playing video games) to tion. Some research has examined motivations associated with
420 Journal of Marketing Research 58(3)

tipping behavior, which is the primary source of income for perspectives of the broadcaster and the platform, is nontransi-
both broadcasters and platforms. For example, Wohn, Free- ent and can last beyond 20 minutes.
man, and McLaughlin (2018) conducted an interview-based Our article is also linked to research that examines the roles
study to show that viewers give money for both intrinsic of the emotions of service providers and customers in business
(e.g., emotional attachment, interaction) and extrinsic (e.g., settings. Previous research has established that people may
rewarding the broadcaster for the entertainment received, vicariously share the emotions of others in social interactions
encouraging future content provision) reasons. Most of these and has termed such sharing “emotional empathy” (Escalas and
studies were qualitative in nature or used surveys for data col- Stern 2003). Previous research has also examined the contagion
lection. In a few studies that collected data in the field, engage- of emotion and its impact on decision makers’ attitudinal
ment activities were examined independently (e.g., Lu et al. responses and behaviors. For example, Howard and Gengler
2020), which ignores the interdependencies and dynamics of (2001) found, in a laboratory setting, that receivers can mimic
broadcasters’ and viewers’ various activities within a live the smiles of senders and experience a happy emotion, which
stream.4 Our study contributes to this literature by examining further leads them to have a positive attitudinal bias toward a
the role of the emotion of the service provider (i.e., the broad- product. Pugh (2001) noted that customers can catch the pos-
caster) in facilitating the activities of the customers (i.e., the itive affect of frontline workers, resulting in higher evaluations
viewers) at the aggregate level. In addition, our modeling of service quality. Hennig-Thurau et al. (2006) discovered that
framework enables us to quantify not only the immediate the authenticity of employee emotion rather than the extent of
effects of broadcaster emotion but also the relatively longer- employee smiling can influence customer emotion and, subse-
lasting effects, which have been lacking in previous research. quently, customer assessments of a service interaction and their
Our study is also closely related to research scrutinizing the relationship with the service provider. Liang, Chen, and Lei
PWYW pricing model in various business contexts. A central (2016) proposed an “inspiration-helping” hypothesis and estab-
question in the PWYW literature and the related voluntary lished that positive emotion can inspire people to donate; when
market payment literature concerns the factors underlying will- positive emotion was used in tandem with negative emotion, its
ingness to pay and the amount paid (Natter and Kaufmann persuasive power was even stronger. Using large and real-life
2015). The extant research has uncovered several factors, such field data, this research adds to the literature by documenting
as social norms (e.g., Regner 2015), social image (e.g., Lu et al. the existence of emotional contagion in the burgeoning busi-
2020), and status (Gravert 2017), for voluntary paying beha- ness context of live streaming. It also documents the effects of
viors. Nevertheless, this literature has overlooked at least two the emotions of both the broadcaster and the viewers on viewer
factors. First, voluntary paying behavior occurs not indepen- engagement activities in a live stream, in particular viewer
dently but in combination with other customer behaviors—and, tips. In addition, whereas extant studies have examined only
on some occasions, firm behaviors as well. Investigating pay- the immediate effects of emotion, our research investigates
ing behavior in isolation could potentially lead to a biased both the immediate and the relatively longer-lasting effects
understanding of its triggers. To this end, we take into account of emotion.
variables on both the broadcaster and viewer sides that may
affect viewer tips, and we provide evidence that broadcaster
emotion, an important factor on the broadcaster side, signifi- Data and Methodology
cantly affects the payment of viewer tips. Second, most busi- We obtained our data from an online live streaming service
ness contexts that use PWYW and have been examined by provider in China, which is affiliated with a large user-
academics have a static setting in which each customer makes generated content (UGC) platform and started in 2016 when
the payment decision once. However, in business contexts such live streaming was taking off in China. The UGC community is
as live streaming, viewers interact with the broadcaster in real home to people who like to share funny and embarrassing
time during their entire stay in the live stream. Thus, engaging stories about themselves and others and is one of the largest
activities take place throughout a live stream, and multiple of its kind (Chen et al. 2019). The live streaming service is also
tipping events might occur. It is thus also critical to understand popular and ranked among the top 30 in this market as of July
whether the motivational factors drive continuing payments 2019. The data contain details of 1,450 live streaming sessions
over time under PWYW, and the extant research remains silent randomly drawn from all active sessions that started between 8
on this question. Due to the nature of the data in the current and 10 P.M. on the platform each day over two weeks from
study (i.e., at the aggregate level only), we could not analyze January 31 to February 13, 2018.5 For each live stream, we
individual viewer activities to provide a conclusive assessment. collected the amount of viewer tips, details of viewer chats, and
Nonetheless, our results suggest that the effect of broadcaster number of viewer likes in each minute. We also recorded
emotion on attracting viewer tips in live streams, from the videos of all the live streams, which enabled us to analyze the

4 5
We have conducted extensive literature reviews on each research stream The platform suggested the two-hour sampling time window because both
related to our study and present the results in three literature review tables in broadcasters and viewers are most active during this time period, and their
Web Appendix A. behaviors during this time period are of primary interest to the platform.
Lin et al. 421

content of the live stream. The length of the live streams, Viewer emotion. We collected all live comments posted by the
plotted in Figure WB1 in Web Appendix B, is highly skewed. audience in each minute. We used the Linguistic Inquiry and
While some sessions lasted almost 800 minutes, the majority Word Count (LIWC) text analysis tool (Pennebaker et al. 2015)
lasted less than 120 minutes (highlighted in orange). We there- to measure the sentiment of the live comments and denoted it as
fore focused on the first 120 minutes of the sessions.6 viewer emotion. The LIWC program reads a text and compares
Our main analysis was conducted at the minute level. For each word in the text with built-in dictionaries. We chose this
each minute in each live stream, we assembled structured data program because it is convenient to use and has proven useful
such as the amount of tips, the count of words in the comments, in a wide range of studies in marketing (e.g., Ertimur and
and the number of likes sent by the audience. We also used Coskuner-Balli 2015; Ransbotham, Lurie, and Liu 2019; Yaz-
machine learning algorithms to analyze the unstructured data, dani, Gopinath, and Carson 2018) and other disciplines (e.g.,
including video clips of the live streams and the detailed viewer Goes, Lin, and Yeung 2014; Li 2008; Yin, Bond, and Zhang
chats, and constructed moment-to-moment measures of the 2014, 2017). The program originally had built-in dictionaries
emotions of both broadcasters and viewers (Zhang, Wang, and only for English texts, which were subsequently translated into
Chen 2020). Our final data have an unbalanced panel structure, other languages, including the simplified Chinese version suit-
wherein cross-sectional units are live streams and longitudinal able for our business context.8 A well-known problem for Chi-
units are time in minutes, due to the varying durations of dif- nese, as opposed to English, is that Chinese words in a text are
ferent live streams. Next, we discuss the process to construct concatenated to each other. Therefore, we had to segment the
the focal variables and provide evidence of the interdependen- texts and add white spaces between words before performing
cies of the variables. We then establish the stationarity of the any follow-up natural language processing tasks. This process
data and test the endogeneity of the focal variables, which helps is commonly referred to as Chinese segmentation (Peng, Feng,
us determine both whether a PVAR model is appropriate for the and McCallum 2004). To segment the texts, we used the PKU-
data and which variables should be included as dependent vari- SEG toolkit, which provides a word segmenter tailored for
ables. We also determine the optimal number of lags for the online and informal language (Luo et al. 2019).9 We then used
PVAR model. LIWC to extract viewer emotion from the segmented words.
The LIWC results include two numbers representing the extent
of positive and negative emotion exhibited in the text. The
results also include an integrated variable called Tone, which
Variables and Summary Statistics combines the two numbers and indicates a more positive emo-
Broadcaster emotion. We measured emotion using the broadcas- tion when it is greater than 50 (Pennebaker et al. 2015). We
ter’s facial expressions. We employed a pretrained deep neural used this integrated variable in our analysis and named it
network that produces seven probabilities for an input image, ViewerEmotion.
each representing the likelihood that the image shows a partic-
ular type of emotion out of seven types: happiness, surprise, Viewer tips. Viewers can tip the broadcaster during their stay in
disgust, anger, sadness, fear, and neutrality. We used the deep the live stream. The tips are denominated in the platform’s
neural network to calibrate the probability of the broadcaster electronic currency and can be exchanged for real money at a
being happy in every second and used the second-level data to rate of approximately 100 tips for 1 Chinese yuan (CNY). Tips
construct the broadcaster emotion (BCEmotion) at the minute from live streams usually constitute a significant source of
level in our main analysis.7 This variable measures the average income for broadcasters. When a broadcaster exchanges tips
probability of the broadcaster being happy across all seconds in for cash with the platform, a higher exchange rate is used (e.g.,
a minute, as happiness is deemed the most important type of 150 tips for 1 CNY), enabling the platform to take a share of the
emotion in the business context of live streaming and is the revenue. We counted the number of tips the broadcaster
focal variable of the article. We employed a host of alternative received in each minute and divided it by 100 to obtain a
approaches to construct this key variable of broadcaster emo- measure of the real money paid by the viewers (AmountTips).
tion and test the robustness of our main results, including mea-
suring it at the half-minute level rather than the minute level,
loading into it both happiness and sadness, loading into it both 8
The details of this toolkit are available at http://dictionaries.liwc.net.
9
valence and arousal, and constructing it on the basis of the PKUSEG is available at https://github.com/lancopku/pkuseg-python. Other
broadcasting scripts rather than the broadcasters’ facial expres- popular Chinese segmentation tools include the Stanford Word Segmenter
(https://nlp.stanford.edu/software/segmenter.shtml), Jieba Segmenter (https://
sions. We discuss details of these tests in the “Robustness github.com/fxsjy/jieba), and THULAC (http://thulac.thunlp.org/). These tools
Checks” section. differ primarily in the underlying programming languages and scope of the
natural language processing tasks to which they can be applied. For Chinese
word segmentation, the performance of the different tools is comparable (for
6
Using different cutoff values (such as 90 or 150 minutes) generated comparison results, see http://thulac.thunlp.organdhttps://github.com/
qualitatively consistent results, which are available from the authors upon lancopku/pkuseg-python). We selected PKUSEG because of our familiarity
request. with it as well as its capability of dealing with words that are commonly
7
The details of the algorithm and check of its reliability appear in Web used online (e.g., “bgm” for “background music,” “233” for laughing, “lol”
Appendix C. for “laughing out loud”).
422 Journal of Marketing Research 58(3)

Table 1. Descriptive Statistics and Correlations.

Mean SD Min Max 1 2 3 4 5 6

1. BCEmotion .103 .140 .000 1.000 1.000


2. ViewerEmotion .256 .356 .000 1.000 .081** 1.000
3. AmountTips .169 3.098 .000 396 .026** .023** 1.000
4. NumLikes .007 .022 .000 .662 .018** .093** .012** 1.000
5. LenCmt .013 .017 .000 1.116 .103** .299** .023** .141** 1.000
6. NumViewers 4.797 5.010 1.000 119 .088** .212** .053** .303** .406** 1.000

**p < .01.


Notes: N(obs) ¼ 105,453 (only from minute 3 to minute 120).

Number of likes. In addition to tipping, viewers can tap the Model-Free Evidence
“Like” button to engage with the broadcaster. Tapping the
Table 1 also displays the correlations among the six key vari-
“Like” button is not only free of charge but also an effortless
ables constructed in the context. As the table shows, all the
way to express viewer attitude, appreciation, or other feelings
correlations are positive and significant (p < .01), suggesting
about the broadcaster and the content. We counted the number
that these variables likely move together and influence each
of likes received by the broadcaster in a minute and divided it
other. Specifically, there is a significant correlation between
by 1,000 for scaling purposes (NumLikes).
broadcaster and viewer emotion (corr ¼ .081), and there are
also significant correlations between broadcaster emotion
Length of viewer comments. We also measured another engage- and viewer engagement activities, in particular, viewer tips
ment activity of the viewers: viewer chatting, which is free of (corr ¼ .026). In addition, the number of viewers is signifi-
charge to viewers but requires them to exert some effort to cantly correlated with each of the other variables.
compose the chat entries. In particular, we counted the number To understand these correlations more intuitively and to
of words in the live comments by the audience in a minute and probe the dynamics within the same stream as well as the
divided it by 1,000 for scaling purposes (LenCmt). differences across streams, we display the evolution processes
of three key variables (BCEmotion, ViewerEmotion, and
AmountTips) over time and up to 120 minutes in three sample
Number of viewers. Finally, another variable that can be
sessions in Figure 2. A few observations can be made about the
retrieved and might interact with others in this context is the
plots. First, there is a pattern of positive correlation between
number of viewers. The platform helped us record the number
broadcaster emotion and viewer emotion. Viewers in a live
of viewers at the end of each minute in a live stream (Num- stream are generally happier when the broadcaster is happier,
Viewers). Note that this number does not include the viewers and similarly, the broadcaster is happier when the audience is
who appeared in the live stream during the minute but left pleased. Second, viewer tips are scattered across the entire
before the number was recorded.10 duration of a live stream and exhibit a few spikes. These spikes
Table 1 shows the descriptive statistics and correlations for seem to either be triggered by heightened broadcaster emotion
these variables across all minutes of all live streams. It is worth (and viewer emotion) a few minutes earlier (e.g., at about min-
mentioning that viewer engagement activities are generally ute 20 in the first stream and about minute 50 in the second
highly skewed, which gives rise to popular live streams attract- stream) or make the broadcaster (and the viewers themselves)
ing large amounts of tips, likes, and chats. As we show in happier in subsequent minutes (e.g., at about minute 40 in the
Figure WB2 in Web Appendix B, while 30% of live streams third stream). This leads to the third observation that the influ-
did not receive a single tip, the top 5% could derive more than ences among the variables vary over time in each live stream
70,000 in tips (or 700 CNY).11 In addition, there were approx- and across different live streams.12
imately 5 viewers on average in each minute across all minutes These items of model-free evidence motivate us to account
and all live streams, with a maximum of approximately 120 for the interrelationships among these time series to quantify
concurrent viewers. the influence of one variable on another under the framework
of the PVAR model. The PVAR model is an extension of the
10
vector autoregression (VAR) model that has been widely used
The platform could have recorded the number of viewers at the end of each
second in a given minute and taken the average across all the seconds as the
to model interdependencies among multiple time series (Sims
number of viewers for that minute. It also could have recorded the detailed 1980). It assumes that each dependent variable is a function of
entry and exit timing of each viewer and calibrated the number of viewers in a
minute. These alternative approaches were not adopted by the firm because
12
they entailed higher data collection costs. In additional preliminary analysis, we regress each of the key variables on
11
In additional tests (discussed subsequently), we used the logarithm of the the remaining five variables as well as all the lagged variables, using
number of tips and removed the top 5% of the live streams. Both analyses fixed-effects models. The results are presented in Web Appendix D and lend
generated results that were qualitatively consistent with our main analysis. additional support to the strong correlations among these key variables.
Lin et al. 423

Figure 2. Dynamics of variables in sample live streams.

its own past values and the past values of all other dependent Stationarity and Endogeneity
variables (as well as the control variables, if any). The VAR
model has been further extended to panel data (Holtz-Eakin, Following existing applications of the PVAR model (e.g.,
Newey, and Rosen 1988). Previous studies have shown the Abrigo and Love 2016), we began our empirical analysis with
success of applying both VAR and PVAR models to business unit-root tests to check the stationarity of the key endogenous
contexts such as online word of mouth and social media (Chae variables in our panel data. Generalized method of moments
et al. 2016; Stephen and Galak 2012; Trusov, Bucklin, and (GMM) estimators suffer from the weak instruments problem
Pauwels 2009; Zhang et al. 2012), corporate social responsi- if endogenous variables contain unit roots (Abrigo and Love
bility (Kang, Germann, and Grewal 2016), service crises (Gij- 2016; Binder, Hsiao, and Pesaran 2005). Because our panel data
senberg, Van Heerde, and Verhoef 2015), and sales promotion were unbalanced, certain popular panel unit-root tests, such as
(Srinivasan et al. 2004). Our data comprise 1,450 live streams the Levin–Lin–Chu test and Breitung test, could not be used, as
with detailed records of the key variables in each minute of they require strongly balanced panel data. We thus resorted to
each stream. Such a panel structure enables us to control for the Fisher-type approach for testing unit roots in panel data
stream-specific heterogeneity in the key variables by the use of (Kang, Germann, and Grewal 2016). In its execution, this algo-
fixed effects (Kang, Germann, and Grewal 2016; Love and rithm conducts a unit-root test for each individual stream and
Zicchino 2006). then merges the p-values from the tests to conclude whether a
424 Journal of Marketing Research 58(3)

Table 2. Results of Panel Unit Root Tests. PVAR model for the data. Given the set of six endogenous
variables and P time lags and the absence of other covariates,
Inverse
Variable Chi-Squared p-Value Conclusion
the model can be written as follows:
2 3 2 3
BCEmotioni;t 2 p 3 BCEmotioni;tp
BCEmotion 41,854 .000 Stationary 6 7 b11 bp12 bp13 bp14 bp15 bp16 6 7
6ViewerEmotion 7 6 bp 6ViewerEmotion 7
ViewerEmotion 60,659 .000 Stationary 6 i;t 7 6 21 bp22 bp23 bp24 bp25 bp26 7
76 i;tp 7
6 7 6 p p 7
6 7
6 7 6 7
AmountTips 51,762 .000 Stationary 6 AmountTipsi;t 7 Xp 6 b
6 31
bp32 bp33 bp34 bp35 b36 76 AmountTips
76 i;tp 7
6 7¼ p¼1 6 bp
7
NumLikes 56,568 .000 Stationary 6
6 NumLikesi;t 7
7 6 41 bp42 bp43 bp44 bp45 bp46 76
76 NumLikes 7
7
6 7 6 p p 7 6 i;tp
7
LenCmt 36,070 .000 Stationary 6 7 4 b51 bp52 bp53 bp54 bp55 b56 56 7
4 LenCmti;t 5 p 4 LenCmti;tp 5
NumViewers 21,990 .000 Stationary bp61 bp62 bp63 bp64 bp65 b66
NumViewersi;t NumViewersi;tp
Notes: Number of panels ¼ 1,450; average panel length ¼ 75. H0: All panels þ mi þ eit :
ð1Þ
contain unit roots; Ha: At least one panel is stationary.
In this model, the coefficient matrix of b denotes the para-
specific variable in the panel data passes the unit-root test. As we meters to be estimated that capture the effects of the lagged
show in Table 2, the null hypothesis of having a unit root is endogenous variables, mi captures the panel fixed effects, and
rejected for all six variables. Note that this approach leads to the eit is a vector of the idiosyncratic error terms that are assumed
conclusion that the panel data pass the unit-root test if at least to be zero centered.14
one stream passes the test, or all streams do not have a unit root. To complete the model, we need to determine the optimal
This approach does not guarantee that none of the streams have a number of lags (i.e., the P parameter in the model). A conven-
unit root. In an additional test presented in Web Appendix E, we tional way to select the optimal lag is to use the moment and
performed unit-root tests for the key variables for each stream model selection criteria (MMSC) proposed by Andrews and Lu
separately and checked the results individually. We observed (2001) for GMM models (Abrigo and Love 2016). The criteria
that all the variables passed the unit-root tests for 681 of these are based on the J-statistics of overidentifying restrictions for
1,450 live streams (about 47.0%).13 GMM models (Hansen 1982). These criteria are analogous to
For the PVAR model to accurately capture the influences various information criteria, such as Akaike information criter-
among the endogenous variables, we first had to determine ion (AIC), Bayesian information criterion (BIC), and Hannan–
which variables are endogenous and to be included on the left- Quinn information criterion (HQIC), which are commonly
hand side of the equation, and which variables are not endogen- used to select the best models using maximum likelihood esti-
ous and can be included as control variables. Intuitively, in such mators. Ideally, the optimal lag is the one that minimizes
a dynamic and interactive environment as a live stream, we MMSC BIC , MMSC AIC , and MMSC HQIC together. Table 4
would expect all six variables to be endogenous. Empirically, presents the results of the selection process for our data. As the
the Granger causality test is widely used to establish causality in information criteria in the table suggest, the values of both
time series data (Granger 1969). The basic idea underlying this MMSC AIC and MMSC HQIC are minimized at two lags.
causality concept is that if one variable X Granger-causes Furthermore, MMSC BIC is quite small at two lags and is close
another variable Y, then observing the history of X will help to the minimal value, which is obtained at one lag. We thus
explain the residual variances in Y after accounting for Y’s own used two lags for our main analysis. Subsequently, we discuss
history. A common practice is to apply the test to data with a analyses using two additional models with different numbers of
history, starting from one period and extending to an arbitrary lags to show the robustness of the results.
number of periods (e.g., Trusov, Bucklin, and Pauwels 2009). In
the context of live streaming, there is no a priori knowledge of
how long the history should be. Thus, we performed the causality Empirical Findings
tests for each pair of variables, starting with 1 lag and stopping at
We applied the PVAR model to the live streaming data and
20 lags, beyond which there was no substantive change to any
obtained the coefficient estimates (i.e., the b s), which are
test. We report the smallest p-value from these tests in each cell
displayed in Table 5. Previous research has highlighted that it
of Table 3. As the table shows, Granger causality is established
for most of the pairs except that length of viewer comments does
not Granger-cause the amount of tips (p ¼ .100). 14
Note that all the endogenous variables are stationary rather than trend
stationary; thus, there is no need to include the time trends, which are
usually included as exogenous variables, in our PVAR model (e.g., Trusov,
Model Specification Bucklin, and Pauwels 2009). We could not add seasonal dummies either (e.g.,
Trusov, Bucklin, and Pauwels 2009; Zhang et al. 2012) due to the short time
The presence of endogeneity in all the variables in the data and span of data collection. In addition, minute dummies are not appropriate
the stationarity of these time series justify the choice of the because of the differences across live streams (Figure 2). Because the
endogenous variables may Granger-cause each other, and the influences
among them (both immediate and relatively longer lasting) are most often
13
Further analysis using only the live streams for which all the key variables the primary focus, it is not unusual to see VAR or PVAR models without
pass the unit-root tests generated qualitatively the same results presented in the exogenous variables in the literature (e.g., Gijsenberg, Van Heerde, and
main text. These tests are available from the authors upon request. Verhoef 2015).
Lin et al. 425

Table 3. Results of Granger Causality Tests.

DV (Column Name) Is Caused by (Row Name) BCEmotion ViewerEmotion AmountTips NumLikes LenCmt NumViewers

BCEmotion — .001** .023* .034* .007** .005**


ViewerEmotion .000** — .050* .022* .000** .000**
AmountTips .000** .002** — .000** .002** .000**
NumLikes .005** .000** .014* — .000** .000**
LenCmt .000** .000** .100 .001** — .000**
NumViewers .000** .000** .014* .000** .000** —

*p < .05.
**p < .01.
Notes: DV ¼ dependent variable. Tests were done with up to 20 lags, and we report the smallest p-values of the chi-square tests across the models with 20
different lags.

Table 4. MMSC for Optimal Lag Selection. Drawing on a particular IRF, we could calculate the cumu-
lative IRF to assess the total response of one endogenous vari-
Lag MMSCBIC MMSCHQIC MMSCAIC able to a shock on another variable up to a given time period.
1 1,256.37 309.22 105.02 We illustrate both GIRFS and cumulative IRFs in Figure 3,
2 1,144.49 434.13 123.45 using four pairs of variables to delineate the key relationships
3 781.68 308.11 100.98 between emotion and engagement in live streaming. We use a
4 394.65 157.86 54.30 time window of 20 minutes to illustrate the IRFs to be consis-
Notes: MMSCBIC, MMSCHQIC, and MMSCAIC are moment model selection cri-
tent with the previously performed Granger causality tests. The
teria analogous to the widely used likelihood-based model selection criteria: the choice of 20 minutes also helps shed light on these effects in
Bayesian information criterion (BIC), the Hannan–Quinn information criterion relatively longer time periods given the dynamic nature of live
(HQIC), and the Akaike information criterion (AIC) (Andrews and Lu 2001). streaming. We draw the 95% confidence intervals in all the
plots with Monte Carlo simulations using 1,000 samples for
each response. To facilitate discussion, we also present the
is difficult to interpret the full impact of one variable on another cumulative responses of all variables in relatively shorter time
from the estimated coefficients alone (Sims 1980). To capture periods (i.e., 2 and 5 minutes) as well as longer time periods
all the possible routes for one variable to influence another, (i.e., 10 and 20 minutes) in Table 6.
researchers usually resort to using impulse response functions
(e.g., Kang, Germann, and Grewal 2016; Trusov, Bucklin, and
Pauwels 2009; Zhang et al. 2012) to capture both direct and
Main Results for Emotion and Engagement
indirect influences.
An impulse response function (IRF) illustrates how the Panels A and B in Figure 3 visualize the responses of viewer
shock of one standard deviation on one variable affects another and broadcaster emotion, respectively, to shocks in the emotion
variable over time, with other variables in the system remaining of the other. Panels C and D show the responses of viewer tips
unchanged. We adopted the generalized IRF (GIRFs; Pesaran and number of viewers to shocks in broadcaster emotion,
and Shin 1998; Trusov, Bucklin, and Pauwels 2009) rather than respectively. Within each panel, the first plot displays the
the orthogonalized alternative. The former approach imposes GIRF, and the second shows the cumulative GIRF. Note that
no arbitrary causal ordering among the endogenous variables in a six-equation PVAR model, there are 36 IRFs and their
and allowed us to calibrate the effects of simultaneous shocks cumulative versions. Because each IRF takes into consider-
by making direct use of the residual variance–covariance ation not only the direct response of an endogenous variable
matrix of the PVAR model (Dekimpe and Hanssens 1999).15 to the shock of another variable but also the indirect responses
As discussed in Pesaran and Shin (1998), both approaches shed (e.g., via a third variable that also responds to the shock), we
some light on the short- and long-term effects of changes in one can examine and discuss each panel in isolation from the rest.
variable on another for any nondiagonal error variance–covar- Broadcaster and viewer emotion. We first investigated the
iance matrix. In addition, IRFs calculated with the two responses of viewer and broadcaster emotion to shocks in the
approaches coincide in the cases for the first equation of the emotion of the other (i.e., Panels A and B, respectively), which
model and deviate only in those for other equations. For our sheds light on whether the two parties in a live stream influence
application, we adopted GIRFs due to the lack of obvious each other in terms of emotion. Note that the standard deviation
causal ordering among the six key variables. of viewer emotion is much larger than that of broadcaster emo-
tion (.354 vs. .140, as shown in Table 1).
Comparing the first plot in each of the two panels leads us to
15
For exact specifications of the two approaches, see Equations 7 and 10 in two observations. First, for a relatively smaller shock on broad-
Pesaran and Shin (1998). caster emotion (i.e., .140), viewer emotion is immediately
426 Journal of Marketing Research 58(3)

Table 5. Estimation Results from the PVAR Model.

Dependent Variable

BCEmotioni,t ViewerEmotioni,t AmountTipsi,t NumLikesi,t LenCmti,t NumViewersi,t

BCEmotioni;t1 .4496** .0408** .3478* .0031** .0008 .1010


(.0082) (.0146) (.1364) (.0009) (.0006) (.0910)
BCEmotioni;t2 .2477** .0242 .1333 .0017* –.0001 –.0304
(.0069) (.0135) (.1078) (.0008) (.0005) (.0800)
ViewerEmotioni;t1 .0175** .1180** .0623 .0004 .0013** .0967**
(.0012) (.0048) (.0420) (.0003) (.0002) (.0247)
ViewerEmotioni;t2 .0150** .0661** .0513 .0007** .0003 –.0087
(.0012) (.0045) (.0401) (.0003) (.0002) (.0217)
AmountTipsi;t1 .0003** .0008 .0987** –1.82e5 –4.94e6 .0158**
(.0001) (.0004) (.0286) (2.51e5) (1.32e5) (.0045)
AmountTipsi;t2 –4.92e5 –.0002 .0294 –4.49e5 1.52e5 .0054
(.0001) (.0004) (.0200) (2.55e5) (1.18e5) (.0033)
NumLikesi;t1 .0271 .3274** –.0224 .4188** .0088** 2.9239**
(.0171) (.0702) (.7165) (.0146) (.0025) (.4346)
NumLikesi;t2 –.0019 .0156 .6147 .0671** .0020 –.8631*
(.0179) (.0675) (.8377) (.0110) (.0025) (.4371)
LenCmti;t1 .3044** 1.5958** .7393 .0270** .3296** 9.6095**
(.0500) (.2044) (1.4827) (.0080) (.0396) (1.3582)
LenCmti;t2 .1464** .8096** 1.4561 .0047 .1922** –1.5022*
(.0408) (.1779) (1.3035) (.0075) (.0254) (.7515)
NumViewersi;t1 .0052** .0088** .0448* .0009** .0006** .6607**
(.0005) (.0016) (.0195) (.0001) (.0001) (.0217)
NumViewersi;t2 .0028** .0016 .0042 .0004** –2.68e5 .1953**
(.0003) (.0011) (.0138) (.0001) (.0001) (.0155)

Stream fixed effects Yes


N(obs) 104,440
*p < .05.
**p < .01.
Notes: Robust standard errors clustered by streams are in parentheses.

boosted by approximately .01, whereas for even a relatively These findings might be rationalized as follows. First, the
larger shock on viewer emotion (i.e., .354), broadcaster emo- broadcaster as the host of the live stream always stays in the
tion is immediately increased by less than .005. This suggests stream, while viewers may enter and exit the live stream, result-
that in the short term (i.e., immediate or within a few minutes), ing in different compositions across different minutes. Second,
broadcaster emotion has a relatively larger influence on viewer they play different roles in the live stream: the broadcaster is
emotion than the other way around. Second, the first plot in the provider of the content and may actively engage with view-
Panel A clearly shows that the response of viewer emotion to ers in live streaming, whereas viewers collectively are the con-
the shock on broadcaster emotion becomes 0 at approximately sumers of the content and often follow whatever happens in
minute 10, while the first plot in Panel B suggests that the the stream.
response of broadcaster emotion to viewer emotion is nonzero
even at minute 20. This suggests that the former is only short- Broadcaster emotion draws viewer tips. Viewer tips are important
lived, whereas the latter is relatively longer lasting. These to broadcasters and the platform, as they are the primary rev-
observations are further confirmed by the second plot in each enue source. The plots in Panel C of Figure 3 and the results in
panel. Specifically, the cumulative response of viewer emotion Table 6 demonstrate the possibility of broadcasters manipulat-
to the shock of broadcaster emotion (i.e., the second plot in ing their emotion, or smiles, to draw more viewer tips. In
Panel A) seems to become satiated quickly, hovering at a level particular, the shock on broadcaster emotion (i.e., .140) is
of approximately .06 after minute 15 and ultimately reaching found to boost viewer tips by .189 CNY in just two minutes.
.062 by minute 20 (the first row and second column in the last To put the number into perspective, the average amount of tips
panel of Table 6), while the cumulative response of broadcaster per minute is only .167 CNY across all the streams (Table 1).
emotion to the shock of viewer emotion (i.e., the second plot in The first plot in Panel C suggests that this effect becomes
Panel B) reaches approximately .093 by minute 20 (the second nonsignificant after approximately 7 minutes and becomes
row and first column in the last panel of Table 6, up from .073 almost 0 by minute 20. The cumulative effect displayed on the
by minute 10) and is not yet satiated. second plot in the same panel increases to .331 CNY in 10
Lin et al. 427

Figure 3. Selected IRFs and cumulative IRFs over time.

minutes and further to .371 CNY in 20 minutes (column 3 in increases the number of viewers by only .05 immediately and
Table 6). These observations indicate that happier broadcasters approximately .164 in 2 minutes (the first row and last column
can potentially draw more viewer tips, which aligns with find- in the first panel in Table 6). Although this cumulative effect
ings in previous studies that positive emotion leads to a positive increases to .538 in 10 minutes, it becomes nonsignificant by
evaluation of service quality (e.g., Pugh 2001) and facilitates minute 20, suggesting that broadcaster emotion increases the
customer behaviors such as providing public services or dona- number of viewers only temporarily and marginally (as
tions (e.g., Liang, Chen, and Lei 2016). opposed to five viewers on average per minute across all the
streams). These results, together with the results for the effects
Broadcaster emotion increases audience size. The mere presence of broadcaster emotion on viewer tips, also suggest that the
of viewers may also be valuable to broadcasters and the plat- mechanism for broadcaster emotion to increase viewer tips is
form, as it helps generate traffic and can be leveraged for likely by increasing the tips by each individual viewer rather
sponsored advertisement displays. In addition, a nonpaying than increasing the audience size. Subsequently, we conduct
viewer today may become a paying viewer tomorrow. Thus, additional tests where viewer engagement activities are mea-
the effect of broadcaster emotion on audience size also war- sured per capita, to confirm this conjecture and demonstrate the
rants some discussion. The plots in Panel D of Figure 3 and the robustness of our main results.
results in Table 6 (the last column) suggest an overall weak and
only temporal effect of broadcaster emotion on audience size. Brief discussion of other effects. Table 6 also displays other effects
In particular, the shock on broadcaster emotion (i.e., .140) that are present in the context of live streaming. Detailed results
428 Journal of Marketing Research 58(3)

Table 6. Immediate and Relatively Longer-Lasting Effects.

Response/Impulse BCEmotion ViewerEmotion AmountTips NumLikes LenCmt NumViewers

Up to 2 Minutes

BCEmotion .191** .026** .189** .002** .001** .164**


(.002) (.003) (.033) (2.62e4) (1.54e4) (.032)
ViewerEmotion .023** .395** .085** .001** .003** .255**
(.001) (.003) (.031) (2.29e4) (1.43e4) (.025)
AmountTips .009** .007** 3.482** –1.68e4 5.78e5 .204**
(.001) (.002) (.114) (1.62e4) (9.16e5) (.028)
NumLikes .008** .018** .010 .030** 4.90e4** .321**
(.001) (.002) (.027) (4.97e4) (1.03e4) (.021)
LenCmt .020** .090** .045 .001** .019** .553**
(.001) (.004) (.031) (2.25e4) (.001) (.039)
NumViewers .030** .052** .241** .006** .004** 3.500**
(.002) (.006) (.063) (.001) (3.20e4) (.081)

Up to 5 Minutes

BCEmotion .272** .040** .268** .003** .002** .323**


(.005) (.007) (.066) (.001) (4.08e4) (.110)
ViewerEmotion .048** .416** .134** .003** .005** .503**
(.003) (.005) (.051) (4.90e4) (3.20e4) (.075)
AmountTips .016** .011** 3.528** 2.17e4 3.43e4* .406**
(.002) (.003) (.135) (2.68e4) (1.60e4) (.065)
NumLikes .018** .030** .048 .036** .001** .608**
(.002) (.004) (.047) (.001) (2.19e4) (.062)
LenCmt .050** .129** .126* .003** .024** 1.130**
(.004) (.010) (.063) (.001) (.002) (.122)
NumViewers .092** .105** .443** .013** .008** 5.863**
(.008) (.015) (.157) (.002) (.001) (.315)

Up to 10 Minutes

BCEmotion .328** .053** .331** .005** .002** .538*


(.010) (.011) (.095) (.001) (.001) (.261)
ViewerEmotion .073** .428** .177** .004** .005** .807**
(.005) (.008) (.069) (.001) (.001) (.173)
AmountTips .027** .017** 3.551** .001** .001** .633**
(.003) (.004) (.141) (4.22e4) (2.56e4) (.121)
NumLikes .033** .040** .085 .038** .002** .949**
(.004) (.006) (.062) (.001) (3.78e4) (.139)
LenCmt .091** .157** .215* .006** .026** 1.853**
(.010) (.016) (.100) (.001) (.002) (.295)
NumViewers .197** .174** .704** .023** .013** 8.519**
(.022) (.030) (.283) (.004) (.002) (.861)

Up to 20 Minutes

BCEmotion .357** .062** .371** .006** .003* .779


(.018) (.015) (.119) (.002) (.001) (.503)
ViewerEmotion .093** .438** .215** .005** .006** 1.122**
(.011) (.011) (.089) (.001) (.001) (.350)
AmountTips .038** .023** 3.576** .002** .001** .858**
(.006) (.006) (.148) (.001) (4.18e4) (.217)
NumLikes .051** .050** .123 .039** .003** 1.292**
(.009) (.009) (.080) (.002) (.001) (.286)
LenCmt .133** .179** .300* .009** .028** 2.598**
(.021) (.023) (.144) (.002) (.003) (.626)
NumViewers .329** .248** .988* .033** .018** 11.150**
(.058) (.055) (.443) (.007) (.005) (1.985)

*p < .05.
**p < .01.
Notes: NumLikes and LenCmt are measured in thousands. Standard deviations from Monte Carlo simulations are in parentheses.
Lin et al. 429

at each minute level can be visualized in corresponding IRFs expose the relative impact of other variables on the focal vari-
and are not reported herein. In brief, we find the following: (1) able over time.16 Formally, the revised formula is as follows:
Broadcaster emotion helps to boost the number of viewer likes. Xt
(2) In addition to broadcaster emotion, viewer emotion exerts c ðlÞ 2
l¼0 ij
yij ðtÞ ¼ X 6 Xt ; i; j ¼ 1; 2; . . . ; 6; j 6¼ i;
some influence on all viewer engagement activities as well as
k¼1; k6¼i
c ðlÞ 2
l¼0 ik
audience size. In fact, because broadcaster emotion does not
directly affect audience size (the first and second rows in the ð2Þ
last column in Table 6), it ultimately could boost the audience
size because of the mediation effects via viewer emotion and where c i j ð lÞ is the GIRF of variable i to the one-standard-
other variables. (3) Viewer engagement activities and audience deviation shock of variable j in time period l. y i j ð tÞ thus
size show strong and relatively longer-lasting effects on each captures the relative contribution of the shock of variable j to
other. In particular, audience size has a strong influence on all the total variances in variable i across all the time periods from
viewer activities and broadcaster emotion. 0 to t, excluding what are induced by the shock of variable i
itself. Note that for any variable i, the sum of the relative
P
weights continues to be 1—that is, 6j¼1; j6¼ i y i j ð tÞ ¼ 1.
Determinants of Emotion and Engagement Figure 4 shows the GFEVDs over time for the four variables
We have thus far discussed each response of one endogenous reported in Figure 3 (i.e., BCEmotion, ViewerEmotion,
variable to the shock on another separately. To understand the AmountTips, and NumViewers), which provides a glimpse of
relative impact of different variables on a focal variable in the not only which variables are the most significant determinants
PVAR model, generalized forecast error variance decomposi- of a focal variable (excluding itself) but also how the relative
tion (GFEVD) can be employed. For each particular l-period importance evolves over time. Regarding broadcaster emotion
ahead, the forecast error variance in the focal variable j is (the top-left plot), we can see that viewer emotion is a signif-
calibrated as the summation of the variance engendered by icant determinant that accounts for approximately 30% of the
each error or impulse in the system (for orthogonalized forecast error variances that other variables contribute in the
impulses, see Lütkepohl [2005, Equation 2.3.35, p., 63]) or the first few minutes. Over time, this contribution decreases but
summation of not only the variances but also the covariances still accounts for over 10% by minute 20 (when 30.4% of the
(for generalized impulses, see Pesaran and Shin [1998, the variance in broadcaster emotion is explained by all the other
formulas above Section 3 on p. 20]). If, for each particular variables than broadcaster emotion itself). In addition, viewer
period t, we take into account the forecast error variances for tips contribute to the variance in broadcaster emotion in the
any period from period 0 to the current period, then the total first two minutes, but as time passes, they become a negligible
forecast error variances of variable j can also be calculated by factor in boosting broadcaster emotion. Consistent with what
first summing up the variances induced by each impulse from we discussed previously, the number of viewers has a large
period 0 to t and then summing up these variances across all the impact on broadcaster emotion as well as on other variables
impulses. In this way, we are able to quantify the relative and (other plots in Figure 4), and the relative impact tends to
dynamic impact of the impulses pertaining to all the endogen- strengthen over time. Examining the GFEVDs for viewer emo-
ous variables on a particular focal variable. As noted in prior tion (the top-right plot), we find that broadcaster emotion mat-
studies (Nijs, Srinivasan, and Pauwels 2007; Pesaran and Shin ters, and its relative impact is almost constant across different
1998), we can use either IRF or GIRF for the decomposition periods. In addition, the length of viewer chats significantly
tasks, depending on whether a specific causal ordering can be contributes to variances in viewer emotion because viewer
established. We use GIRF to compute GFEVD for our main emotion is deciphered from viewer chats; thus, the two natu-
interest in understanding the drivers of the key variables on rally could be highly correlated.
both the broadcaster side and the viewer side and because of An important finding from the GFEVDs for viewer tips (the
the lack of an obvious causal ordering. bottom-left plot) is that audience size and broadcaster emotion
Prior studies commonly determine a specific time period for are the two most significant contributors. At t ¼ 0, each
computing GFEVD by when the VAR system is stabilized accounts for approximately 50% of the variance. As time
(e.g., 26 weeks in Nijs, Srinivasan, and Pauwels [2007]). In
the context of live streaming, we have no prior knowledge of 16
how many time periods are needed to fully capture the dynamic Note that all the discussions in this section, without explicit notice, are about
the forecast error variances that are the net of those induced by the focal
effects. Thus, in accordance with the number of periods used to variables themselves, and the percentages should be interpreted only within
compute GIRF, we also compute GFEVD for up to 20 minutes. the rest of the variables. We present in Web Appendix F the GFEVD plots
Because the GIRFs of each variable to the shock of itself are using the standard formula, where the GIRFs of each variable to the shock of
often larger than the GIRFs to shocks of other variables (in itself are also included. An alternative approach to compare the variables with
particular, the former are mostly one order of magnitude higher only small contributions to a focal variable is to simply plot these contributions
calculated from the standard formula against time (see, e.g., Pauwels et al.
than the latter when t ¼ 0, as reported in Table WF1 in Web 2004), which is de facto equivalent to the approach used in the main text.
Appendix F), we exclude the GIRFs of each variable to the The plots generated by this alternative approach also appear in Web
shock of itself in the standard formulas so that we can visually Appendix F.
430 Journal of Marketing Research 58(3)

Figure 4. GFEVDs for key variables (excluding variances induced by focal variable).
Notes: GFEVD provides estimate of the proportion of the variance of a variable in the system with forecasts in t steps that can be explained by the generalized
shock in each variable. In these plots, the other variables (excluding the focal variable) together explain 30.4%, 6.6%, .9%, and 7%, respectively, of the variance of the
focal variable with forecasts in 20 minutes.

passes, the impact of the number of viewers rises, while that of an important marketing variable that broadcasters and plat-
broadcaster emotion diminishes. Nevertheless, broadcaster forms could leverage in live streaming. A natural question that
emotion continues to drive more than 20% of the variance in immediately follows is when more emotion, or more smiles,
viewer tips for a time window of 20 minutes. This result high- should be shown by the broadcaster in a live stream.17 Figure 2
lights the importance of broadcaster emotion management in shows that the spikes of viewer tips seem to happen during the
live streaming as well as its feasibility and potential in this course of a stream (e.g., at about minute 20, 25, 45, and 95 for
particular business context. The GFEVDs for audience size (the the first stream; at about minute 35, 45, and 50 for the second
bottom-right plot) suggest that the contributions of other vari- stream; at about minute 40 for the third stream), but not at the
ables are relatively stable over time, with viewer chats being beginning or near the end, even if broadcaster emotion in these
the most important contributor, followed by other viewer activ- periods sometimes is heightened (e.g., near minute 120 in the
ities. The influence of broadcaster emotion is marginal and first stream). This implies that the relationship between emo-
accounts for less than 10% of the variance in audience size tion and tips might be stronger during some period in the
across all time periods. streaming session. To shed light on the dynamic effect of
broadcaster emotion on viewer tips, as well as other effects
Dynamic Effects Within Streams in our PVAR model, we conducted an exploratory moving
window analysis (Pauwels and Hanssens 2007). This algorithm
Our results show that broadcaster emotion in live streaming can
help generate more viewer tips and, to a lesser extent, boost
audience size after all other possible influences in the system 17
We thank an anonymous reviewer for raising the question and pointing us to
are controlled. This suggests that broadcaster emotion can be the analysis to address the question.
Lin et al. 431

Figure 5. Dynamic effects within streams from moving window analysis.

uses a fixed time window to conduct the PVAR analysis and from 0 at the beginning of a live stream, and neither is signif-
starts with the first time period. By rolling the time window to icantly different from 0 approaching the end of the stream. The
later time periods, with older data points being dropped and relatively larger confidence intervals during the last few min-
new data points being added to the analysis, the algorithm can utes are caused by reduced sample sizes. The only time period
potentially capture the “time path” (Pauwels and Hanssens when we observe a significantly positive coefficient estimate
2007, p. 309) of the estimated effects. for the former (i.e., the effect of broadcaster emotion on viewer
To conduct a moving window analysis, the size of the time tips) is from minutes 40–60, roughly in the middle of a typical
window needs to be predetermined. With no prior knowledge live stream. Coincidently, the estimated coefficient of the latter
of the length of data needed to identify the effects in PVAR (i.e., the effect of viewer tips on broadcaster emotion) is
models or criteria for optimal selection of the length of data, we significantly positive during the period roughly from minute
resorted to using multiple window sizes and checked whether 40 to minute 70.
the results were consistent. While a wider time window is more
capable of identifying the effects because there are more data
points, it also precludes the possibility of shedding light on the Heterogeneous Effects Across Streams/Broadcasters
dynamics of the effects approaching the end of a live stream. Figure 2 also suggests that there is substantive heterogeneity
For example, for a live stream lasting 120 minutes and a mov- among different live streams or broadcasters hosting the
ing window size of 20 minutes, the last analysis that can be streams. In particular, although the extent of happiness of the
conducted is at minute 100, using data from this minute to the broadcaster in the first stream is less than that of the broad-
end. We ended up with several window sizes in the current caster in the third, the first-stream broadcaster is able to gen-
exploratory analysis, including 20, 30, and 40 minutes.18 The erate more tips than the broadcaster in the third because we
results are consistent across the different choices of window observe more tip spikes in the first stream, which is also
sizes and vary across different effects in the PVAR model. In slightly longer than the third stream. This heterogeneity may
Figure 5, we plot the estimated coefficients for two effects, that be caused by broadcaster characteristics or the presence of
of broadcaster emotion on viewer tips (the first plot) and that of different viewers. Next, we provide some evidence that broad-
viewer tips on broadcaster emotion (the second), using a win- caster characteristics can explain the observed heterogeneity in
dow size of 30 minutes. For both plots, the x-axis represents the different streams to some extent.19
starting minute of the time window, and the y-axis represents The platform helped us collect some variables of the broad-
the estimated coefficient (and the 95% confidence intervals). caster before each live stream, including the number of streams
As the plots show, both coefficients are statistically indifferent they had hosted (termed Experience), the total number of tips
received across all previous streams (termed Fame), and the
18
number of followers on the platform (termed Popularity). All
All the variables for all the time windows pass the panel unit-root tests, and three characteristics reflect broadcasters’ different levels of
the results are available upon request. However, these results do not preclude
the possibility that a unit root may exist in a particular variable in a particular
maturity or seniority within the platform. We first median split
stream and with a particular time window. The results of the moving window the broadcasters according to each of the three variables to
analysis thus need to be further ascertained by conducting extensive unit-root
tests on all the variables in all the stream for all the time windows separately.
19
We acknowledge the absence of this test as a potential caveat of the current We thank an anonymous reviewer for raising the question and asking us to
approach. investigate further.
432 Journal of Marketing Research 58(3)

Table 7. Results from Split Sample Analyses.

Dependent Variable: AmountTipsi,t

BCEmotioni;t1 .0849 .5787* .1641 .5189* .0333 .6207*


(.1556) (.2280) (.1183) (.2387) (.0874) (.2653)
BCEmotioni;t2 .0472 .2457 .1264 .1453 .1767* .0925
(.1237) (.1794) (.0961) (.1878) (.0894) (.2042)
ViewerEmotioni;t1 .0081 .1258 .0390 .0758 .0546 .0637
(.0412) (.0731) (.0307) (.0675) (.0313) (.0716)
ViewerEmotioni;t2 .0072 .1146 .0387 .0588 .0706 .0337
(.0364) (.0703) (.0335) (.0640) (.0506) (.0608)
AmountTipsi;t1 .1563** .0733* .1407** .0945** .0789* .1028**
(.0294) (.0317) (.0419) (.0306) (.0361) (.0339)
AmountTipsi;t2 .0236 .0298 .0575* .0265 .0328 .0285
(.0278) (.0260) (.0276) (.0213) (.0218) (.0237)
NumLikesi;t1 .9746 .7804 .0665 .0311 .2114 .1318
(.8472) (1.0944) (.4244) (1.1924) (.3831) (1.2418)
NumLikesi;t2 .4316 .7174 .3210 .8023 .3496 .8019
(.9353) (1.2913) (.6925) (1.3473) (.6637) (1.3928)
LenCmti;t1 .7658 .7274 .8161 1.7393 .7490 1.4877
(1.4706) (2.8608) (1.1568) (2.5053) (1.4784) (2.1967)
LenCmti;t2 .6890 2.5373 .6519 1.9624 .9528 1.6456
(.9981) (2.7519) (.8174) (2.2245) (.9343) (1.9890)
NumViewersi;t1 .0467 .0390 .0576* .0350 .0672* .0296
(.0261) (.0286) (.0292) (.0252) (.0297) (.0261)
NumViewersi;t2 .0022 .0119 .0154 .0031 .0006 .0048
(.0149) (.0232) (.0188) (.0187) (.0182) (.0197)

Categorization variable Experience Fame Popularity

Less More Less More Less More

Stream fixed effects Yes Yes Yes Yes Yes Yes


N(obs) 53,774 50,666 47,508 56,932 48,015 56,425

*p < .05.
**p < .01.
Notes: Experience is measured by the number of streams a broadcaster has hosted, fame is measured by the total number of tips received the broadcaster has
received, and popularity is measured by the number of followers the broadcaster has on the platform. Robust standard errors clustered by streams are in
parentheses. Only coefficient estimates for AmountTips are reported.

generate three dummy variables (1 ¼ yes, 0 ¼ no). We then lag) directly and immediately influence viewer tips. These
reran the PVAR model on the split live streams (Table 7).20 results together provide converging evidence that only savvy
Finally, we created interaction terms of broadcaster emotion broadcasters can leverage their emotion to generate more
and each of the three dummy variables and reestimated the viewer tips in their live streams.
PVAR models with each of these interaction terms (Table 8).
Because these dummy variables were defined with broadcaster
characteristics, we refrained from also interacting them with
Robustness Checks
variables pertaining to the audience. The results in Table 7 In this section, we employ three sets of robustness checks to
show that broadcaster emotion could directly and immediately validate and corroborate our main analysis. We address the
influence viewer tips in a live stream only when the broadcaster following issues that may threaten the results of our main anal-
is more senior, operationalized in three different ways (i.e., ysis: (1) measurement of broadcaster emotion; (2) additional
columns 2, 4, and 6). These results are corroborated in Table 8. data issues related to skewness and extreme observations,
Across all the model specifications, the interaction terms of the excess zeros in viewer tips, and stationarity of variables; and
dummy variables and broadcaster emotion (with a one-minute (3) (mis)specification of the model, including alternative sets
of dependent variables, different numbers of time lags, and
restricted versions of the main model in which broadcasters are
20 assumed to be unable to respond to viewer chats and viewer
All the variables in all the split samples pass the panel unit-root tests. The
results in this section also hold if only the 681 streams are used, where all the emotion immediately. The results from 15 robustness tests lend
variables in all the streams pass the individual unit-root tests. These results are additional support to the main conclusion of our study. All the
available upon request. results appear in Web Appendix G.
Lin et al. 433

Table 8. Results from Moderation Analyses.

Dependent Variable: AmountTipsi,t


BCEmotioni;t1 .0961 .1461 .1116 .1306 .0193 .0067
(.1485) (.1428) (.0985) (.1032) (.0869) (.0828)
BCEmotioni;t2 .1345 .0330 .1357 .0973 .1438 .1686y
(.1079) (.1159) (.1085) (.0857) (.1099) (.0868)
BCEmotioni;t1  Morei .5073* .4080y .4547y .4186y .6310* .6541**
(.2377) (.2179) (.2359) (.2138) (.2576) (.2247)
BCEmotioni;t2  Morei .2038 .0738 .0470
(.1731) (.1725) (.1806)

Categorization variable Experience Fame Popularity


defining Morei

Stream Fixed Effects Yes


N(obs) 104,440
y
p < .1.
*p < .05.
**p < .01.
Notes: Morei is defined following a median split of each variable and takes 1 for streams with a relatively larger value and 0 otherwise. Robust standard errors
clustered by streams are in parentheses. Only coefficient estimates of the interaction terms for AmountTips are reported.

Measurement of Emotion PVAR model and examined the effects of these two types of
This study concerns the interdependencies and influences valence rather than a single type (Table WG1.3). Fourth,
among broadcaster emotion, viewer emotion, viewer engage- because emotion comprises both valence and arousal, we con-
ment activities and participation in live streams. Here, the two solidated happiness and surprise, which represent valence and
variables pertaining to emotion are obtained from unstructured arousal, respectively, into a single variable of emotion (Table
data (i.e., videos and texts) using machine learning algorithms. WG1.4). In relation to these analyses, we also conducted a
placebo test that employed neutrality as the only emotion in
Thus, having an accurate and consistent measure of the two
the model (Table WG1.5) and found that when broadcaster
variables is of ultimate importance. In our main analysis,
emotion is neutral, its influence on viewer tips is statistically
broadcaster emotion is calibrated by facial expressions, but
indifferent from 0.
viewer emotion is constructed on the basis of live comments
aggregated across all viewers. It is likely that our results may be
susceptible to inconsistency in constructing the two variables, Other Potential Data Issues
to the transient nature of the broadcaster showing a happy face, In addition to the data issues pertaining to the measurement of
and even to the reliability of the machine learning algorithm emotion, our results may also suffer from other potential issues.
used to extract emotion from facial expressions. We conducted four additional tests to demonstrate the robustness
Correspondingly, we conducted four different robustness of our main results. First, as Table 1 suggests, some variables in
tests and found the results of all the tests to support the main our data are highly skewed, which could bias the estimates. We
findings. First, we leveraged the vocal information embedded followed the common practice to take the logarithm of each
in the recordings (i.e., what was said by the broadcasters) to variable (plus one to avoid creating nonsense data) and used the
extract broadcaster emotion. The process involved two steps: new variables in the PVAR model (Table WG2.1). Second, the
transcribing the video clips to obtain the scripts (note that only top 5% of live streams in terms of the total viewer tips received
the broadcaster speaks in a live stream) and then using the same by the broadcasters (Figure WB2 in Web Appendix B) attract a
approach for constructing viewer emotion (i.e., segmenting disproportionately larger amount of viewer tips. To ensure that
words, then applying LIWC) to construct broadcaster emotion the results were not primarily driven by these few streams, we
(Table WG1.1).21 Second, rather than structuring all the data at removed them and reestimated the model on the basis of the
the minute level, we recalculated everything at the half-minute remaining data (Table WG2.2). Third, as Figure 2 suggests, the
level so that not only emotion but also viewer activities were occurrence of viewer tipping appears to be much less frequent
calibrated every 30 seconds rather than every minute (Table (approximately 6% of the minutes) than other key variables,
WG1.2). Third, we included both happiness and sadness in the which may lead to our main results suffering from excess zeros.
To alleviate this concern, we ranked the streams by the fre-
21 quency of occurrence of tips and retained the first quarter of
We performed the transcription task using the speech recognition service of
Baidu, the top search engine in China. The company is also a leading firm in streams where tips occurred most frequently. In these streams,
speech recognition, in particular for the Chinese language. The paid service is viewer tips on average occurred in approximately 17% of the
available at https://ai.baidu.com/tech/speech/asr. minutes, and the issue of excess zeros was less of a concern
434 Journal of Marketing Research 58(3)

(Table WG2.3). Lastly, in conducting panel unit-root tests, we Discussion and Conclusion
followed common practice in PVAR models by stating our null
hypotheses, as all panels contain a unit root. This is quite a loose Summary of Empirical Findings
criterion because the nonexistence of a unit root of the data for In this research, we estimate a PVAR model using a data set of
any single live stream would allow the whole panel data to pass 1,450 live sessions from a popular live streaming platform in
the unit-root test. We thus unpacked the unit-root test for the China to understand the dynamic and interdependent broad-
panel data and conducted the test for each variable in each caster and viewer behaviors in this novel business context. In
stream. The number of streams that passed the unit-root test for the main model, we study the potential emotional contagion
all the variables is 681 (or 47% of all 1,450 streams). Details between broadcasters and viewers. We also investigate the
appear in Web Appendix E. The results of estimating the model impact of emotion of both parties on viewers’ engagement
on data from only these 681 streams are qualitatively unchanged activities, including the amount of tips they collectively send,
(Table WG2.4). the length of comments, and the number of likes they generate
together. Because we measure collective viewer behaviors, the
number of viewers is relevant, and we include it in the model.
Model (Mis)specification In addition to the main analysis, we conduct multiple robust-
ness tests to corroborate the findings.
Finally, our model may suffer from potential misspecification We uncover several interesting results pertaining to the driv-
issues. Our final set of tests show that model misspecification is ers of emotion and the impact of emotion in data-rich environ-
less a concern in our study. One such concern is whether the ments such as live streaming (Wedel and Kannan 2016). First,
number of viewers should be treated as endogenous. While we find that the emotions of broadcasters and viewers influence
viewers in a live stream observe what happens inside, viewers each other. However, the broadcaster’s emotional response to
outside the stream may decide to join for a host of reasons. viewer emotion, compared with viewers’ response to broad-
Some may be related to the fame of the broadcaster, while caster emotion, is relatively larger in terms of both immediate
others may be irrelevant. It is thus worthwhile to evaluate effects and relatively longer-lasting effects. Second, broadcas-
whether the key results from our main analysis hold without ter emotion helps attract viewer tips (a measure of viewer
the number of viewers as an endogenous variable. We ran such behavior) and stimulate viewer liking and chatting (measures
a model specification, with the number of viewers included related to viewer attitude). Our research thus contributes to the
only as an exogenous control variable (Table WG3.1). A literature by examining the effect of emotion on both attitude
related matter is that the aggregated viewer activities are obvi- and behavior, in contrast to prior research that examined the
ously affected by the audience size. It was thus of interest to effects of emotion on the two constructs in isolation. The find-
operationalize the viewer activity variables as per capita rather ings indicate that broadcaster emotion also helps increase the
than aggregated to help shed light on the behaviors of a repre- audience size. Third, the PVAR model enables us to measure
sentative viewer. Performing such an analysis also enabled us and quantify the effects of one variable on another in the sys-
to identify and potentially differentiate two mechanisms for tem in both a relatively shorter time period (e.g., two or five
broadcaster emotion to have a positive effect on viewer tips: time periods; each unit is a minute in our context) and a rela-
by increasing the tips at the individual viewer level, by increas- tively longer time period (e.g., 10 or 20 minutes). In live
ing the audience size, or both. We thus transformed the three streaming, for example, we find that the cumulative effect of
engagement activities (AmountTips, NumLikes, LenCmt) to a broadcaster emotion on viewer tips is statistically positive and
per capita basis and reestimated the model with the new vari- substantive even in 20 minutes. These relatively longer-lasting
ables (Table WG3.2). Another concern relates to the choice of effects, pertaining to emotion in particular, have so far been
the number of time lags for the PVAR model. As determined by absent in the literature. Fourth, our results also highlight the
different MMSC values, our main analysis estimated the model dynamics of these effects across time in a live stream. Specif-
with two lags. It is, however, also true that MMSC BIC favors a ically, we find that the effects are pronounced during the course
model with only one lag, and models with two and three lags, of a live stream but not at the beginning or near the end. Finally,
based on both MMSC HQIC and MMSC AIC , are not very dif- our results suggest that only broadcasters who have more expe-
ferent. We thus estimated alternative models with only one lag rience, receive more tips, or are more popular in past live
and with three lags (Table WG3.3). Finally, processing the streams can leverage their emotion to attract more viewer tips.
information from viewers, in particular, viewer chats and relat-
edly viewer emotion embedded in the chats, may take time.
Correspondingly, broadcasters in live streams may not be able Managerial Implications
to process the information timely. We thus estimated restricted Our results shed light on the behaviors of the main players in
PVAR models where immediate effects of viewer chats and live streaming (i.e., broadcasters and viewers) and have strong
viewer emotion on broadcaster emotion were arbitrarily set implications for live streaming platforms and broadcasters who
to 0, which effectively assumed that broadcasters do not are direct beneficiaries of the business. The results also inform
respond to these variables in the immediately following period viewers of the best strategy for engaging in a live stream as well
(Table WG3.4–3.6). as potential pitfalls.
Lin et al. 435

Our study can be leveraged by live streaming platforms and spending and better plan the replenishment of the virtual gifts
broadcasters to quantify the financial returns on marketing in their accounts.
interventions (e.g., emotion enhancement). Prior research on
PWYW, while illuminating, has focused largely on the motiva- Limitations and Future Research
tional factors (e.g., Feldhaus, Sobotta, and Werner 2018; Kim,
Natter, and Spann 2009; Regner 2015) for tipping behaviors in Our research joins other studies to advance research on PWYW
static and isolated settings and cannot be readily applied to the to online settings. It is important to systematically examine and
dynamic and interactive live streaming setting. Our results understand business contexts under PWYW pricing where cus-
enable simple back-of-the-envelope calculations of the number tomer tipping, chatting and liking, and the emotions of both
of additional tips to expect in a live stream, conditional on customers and providers of products and/or services influence
change in any variable in the ambience. For example, the first each other and coevolve in an integrated manner. Our research
plot in Panel C of Figure 3 suggests that if a broadcaster’s takes a step in this direction estimating a PVAR model on data
happiness is increased by one standard deviation (i.e., .140 in from live streaming. The new explorations expose our study to
Table 1), they will immediately be able to attract .1 CNY, a few constraints and limitations that we call for future research
representing a 60% increase in what they receive (i.e., .167 to address.
in Table 1). The cumulative effects reported in the second plot First, our results are based on data from a single live stream-
in Panel C of Figure 3 and in Table 6 suggest that this number ing platform in China, which primarily hosts talk shows and
increases as time passes (e.g., to .189 CNY in 2 minutes and to casual chats. As live streaming continues to embrace games
.331 CNY in 10 minutes). and e-commerce, it is an empirical question whether our results
Our finding that broadcaster emotion helps attract viewer are replicable in other live streaming platforms.22 For live
tips also provides suggestions for solving the so-called cold stream platforms that cater primarily to game fans, we expect
start problem (Wang et al. 2018), which manifests in the con- our main results to hold. Viewers in game-playing live streams
text of live streaming as the difficulty that broadcasters face in are likely to be even more involved, which allows them to
generating the first virtual gift. Obviously, rather than relying discern the broadcaster’s emotion and easily synchronize with
on the random arrival of virtual gifts to kick off the vicarious the broadcaster. For live streaming channels operating on
process, a broadcaster can seed the process with a smile and e-commerce platforms and used as an alternative channel to
allow their own efforts to pay off with time. In addition, our demonstrate products, our main results may not apply. Some
results show that such a strategy may not be as effective during viewers on these platforms may be more interested in the prod-
the first few minutes and should be used after a live stream has ucts being demonstrated than in the broadcaster in a live
been active for a while. Another related result is that all the stream. Thus, it may be less likely that these viewers will tip
viewer-side variables, except the number of viewers, are less the broadcaster for product demonstration alone. In terms of
effective than broadcaster emotion in driving up tips (the cultural differences, our results, based on data from a live
bottom-left plot in Figure 4). Live streaming platforms some- streaming platform in a collectivist society, may not apply
times need to rationalize their allocation decisions for an often directly to societies with a different cultural orientation.
limited budget, on the broadcaster side, the viewer side, or both Second, we use established machine learning algorithms to
sides, to achieve the best return on investment. This finding conduct text analysis to calibrate the sentiment in chats for both
suggests that if maximizing viewer tips is the primary goal, viewers and broadcasters and use video analysis to extract the
platforms may be better off if they train their broadcasters to emotion of broadcasters based on their facial expressions.
smile more in live streams rather than incentivizing viewers to These algorithms have proven reliable for these tasks. Never-
tip in order to foster a sustainable and prosperous business. theless, the richness of live stream videos may not necessarily
Our study can be used by viewers to guide their behaviors on be captured by the facial expressions of broadcasters alone, and
live streaming platforms as well. First, our results suggest that sentiment derived from chats may not necessarily align with
viewers’ presence matters in live streams. The audience size is the true emotion of broadcasters and viewers. As machine
a strong driver of positive broadcaster emotion. In addition, learning algorithms become more versatile, we expect that
viewer engagement activities, such as chats and tips, greatly future research could expand the metrics from emotion
improve broadcaster emotion (the top-left plot in Figure 4). embedded in facial expressions and sentiment based on text
Second, despite tips being a useful device to reward a broad- to much more comprehensive sets of variables that machines
caster in live streaming, who subsequently may deliver better can readily and automatically collect.
content, viewers should be reminded that the effect of their tips Third, our examination of viewer behavior is at the aggre-
on improving broadcaster emotion seems to come into play gate level, which is determined by the data made available to
only during the course of a live stream rather than at the begin- the current study and the choice of model (i.e., PVAR). Thus,
ning or near the end (i.e., from minute 40 to minute 70, as
suggested by the second plot in Figure 5). This finding provides 22
For example, a recent report by iResearch, a large marketing research and
useful information about the timing under which viewers may consulting firm, provides a glimpse of the success of live streaming in
want to tip to maximize the utility they derive from engage- combination with ecommerce. The preview of the report is accessible at
ment in the live stream. It also helps viewers rationalize their http://www.iresearchchina.com/content/details8_55544.html.
436 Journal of Marketing Research 58(3)

all of our results should be interpreted as aggregated viewer Application to Dynamic Panel Data Models,” Journal of Econo-
behaviors or the behavior of an average viewer in live stream- metrics, 101 (1), 123–64.
ing. Relatedly, while our results show that broadcaster emotion Azar, Ofer H. (2020), “The Economics of Tipping,” Journal of Eco-
is a key driver of viewer tips at the aggregate level, due to the nomic Perspectives, 34 (2), 215–36.
lack of individual viewer–level metrics, we could not verify Bell, Karissa (2016), “Facebook Is Finally Bringing Live Streaming to
whether broadcaster emotion indeed influences the tipping Everyone,” Mashable (January 28), https://mashable.com/2016/
decisions of each individual viewer—and, if so, how. Future 01/28/facebook-live-video-how-to.
research could delve into the questions explored in our study Binder, Michael, Cheng Hsiao, and M. Hashem Pesaran (2005),
with fine-grained individual-level data and investigate viewers’ “Estimation and Inference in Short Panel Vector Autoregressions
heterogeneous behaviors and decision-making processes to with Unit Roots and Cointegration,” Econometric Theory, 21 (4),
better understand the drivers in this or other similar contexts. 795–837.
In this case, PVAR models may not continue to work due to Chae, Inyoung, Andrew T. Stephen, Yakov Bart, and Dai Yao (2016),
issues such as relatively shorter panels and thus the existence of “Spillover Effects in Seeded Word-of-Mouth Marketing Cam-
unit roots in the data. Correspondingly, new methodologies paigns,” Marketing Science, 31 (1), 89–104.
may have to be invented to address the complex interactions Chen, Xingyu, Xing Li, Dai Yao, and Zhimin Zhou (2019), “Seeking
between individual viewers and broadcasters, including the the Support of the Silent Majority: Are Lurking Users Valuable to
same viewer visiting different streams and/or different viewers UGC Platforms?” Journal of the Academy of Marketing Science,
visiting the same stream at the same time. 47 (6), 986–1004.
Chen, Yuxin, Oded Koenigsberg, and Z. John Zhang (2017), “Pay-as-
Acknowledgments You-Wish Pricing,” Marketing Science, 36 (5), 780–91.
The authors thank a large live streaming platform in China for help Dekimpe, Marnik G. and Dominique M. Hanssens (1999), “Sustained
with the collection of the data and extensive discussions about the Spending and Persistent Response: A New Look at Long-Term
industry. They are also thankful to the JMR review team for the con- Marketing Profitability,” Journal of Marketing Research, 36 (4),
structive and developmental review process. The authors also thank 397–412.
the participants at the 2019 Conference on Artificial Intelligence, Elberse, Anita and Jason Bergsman (2008), Radiohead: Music at Your
Machine Learning, and Business Analytics at Temple University; the Own Price (A). Brighton, MA: Harvard Business Publishing.
18th Pre-ICIS Workshop on e-Business; the 2020 Marketing Science Ertimur, Burçak and Gokcen Coskuner-Balli (2015), “Navigating the
Conference; Peking University; Tsinghua University; Shanghai Jiao- Institutional Logics of Markets: Implications for Strategic Brand
tong University; City University of Hong Kong; and Hong Kong
Management,” Journal of Marketing, 79 (2), 40–61.
Polytechnic University for their valuable feedback.
Escalas, Jennifer Edson and Barbara B. Stern (2003), “Sympathy and
Empathy: Emotional Responses to Advertising Dramas,” Journal
Author Contributions
of Consumer Research, 29 (4), 566–78.
Yan Lin and Dai Yao contributed equally and are listed in alphabetical Feldhaus, Christoph, Tassilo Sobotta, and Peter Werner (2018),
order. “Norm Uncertainty and Voluntary Payments in the Field,” Man-
agement Science, 65 (4), 29–37.
Associate Editor Garrahan, Matthew (2011), “YouTube Launches Live Streaming,”
Randolph Bucklin Financial Times (April 8), https://www.ft.com/content/753e646e-
6208-11e0-8ee4-00144feab49a.
Declaration of Conflicting Interests Gijsenberg, Maarten J., Harald J. van Heerde, and Peter C. Verhoef
The author(s) declared no potential conflicts of interest with respect to (2015), “Losses Loom Longer than Gains: Modeling the Impact of
the research, authorship, and/or publication of this article. Service Crises on Perceived Service Quality over Time,” Journal
of Marketing Research, 52 (5), 642–56.
Funding Gneezy, Ayelet, Uri Gneezy, Leif D. Nelson, and Amber Brown
The author(s) disclosed receipt of the following financial support for (2010), “Shared Social Responsibility: A Feld Experiment in
the research, authorship, and/or publication of this article: This Pay-What-You-Want Pricing and Charitable Giving,” Science,
research is supported by the Start-up Research Grant of Shenzhen 329 (5989), 325–27.
University (00001109), the NUS Research Grant (R-316-000-104- Goes, Paulo B., Mingfeng Lin, and Ching-man Au Yeung (2014),
133), the Natural Science Foundation of Guangdong Province, China “‘Popularity Effect’ in User-Generated Content: Evidence from
(2019WQNCX106), and the National Natural Science Foundation of Online Product Reviews,” Information Systems Research, 25 (2),
China (71872115). 222–38.
Granger, Clive W.J. (1969), “Investigating Causal Relations by
References Econometric Models and Cross Spectral Methods,” Econometrica,
Abrigo, Michael R.M. and Inessa Love (2016), “Estimation of Panel 37 (3), 424–38.
Vector Autoregression in Stata,” Stata Journal, 16 (3), 778–804. Gravert, Christina (2017), “Pride and Patronage: Pay-What-You-Want
Andrews, Donald W.K. and Biao Lu (2001), “Consistent Model and Pricing at a Charitable Bookstore,” Journal of Behavioral and
Moment Selection Procedures for GMM Estimation with Experimental Economics, 67, 1–7.
Lin et al. 437

Hansen, Lars Peter (1982), “Large Sample Properties of Generalized Lütkepohl, Helmut (2005), New Introduction to Multiple Time Series
Method of Moments Estimators,” Econometrica, 50 (4), 1029–54. Analysis. Berlin: Springer Science and Business Media.
Hennig-Thurau, Thorsten, Markus Groth, Michael Paul, and Dwayne Natter, Martin and Katharina Kaufmann (2015), “Voluntary Market
D. Gremler (2006), “Are All Smiles Created Equal? How Emo- Payments: Underlying Motives, Success Drivers and Success
tional Contagion and Emotional Labor Affect Service Potentials,” Journal of Behavioral and Experimental Economics,
Relationships,” Journal of Marketing, 70 (3), 58–73. 57, 149–57.
Hilvert-Bruce, Zorah, James T. Neill, Max Sjöblom, and Juho Hamari Nijs, Vincent R., Shuba Srinivasan, and Koen Pauwels (2007),
(2018), “Social Motivations of Live-Streaming Viewer Engage- “Retail-Price Drivers and Retailer Profits,” Marketing Science,
ment on Twitch,” Computers in Human Behavior, 84, 58–67. 26 (4), 473–87.
Holtz-Eakin, Douglas, Whitney Newey, and Harvey S. Rosen (1988), Park, Sangkon, Sohyun Nam, and Jungmin Lee (2017), “Charitable
“Estimating Vector Autoregressions with Panel Data,” Econome- Giving, Suggestion, and Learning from Others: Pay-What-You-
trica, 56 (6), 1371–95. Want Experiments at a Coffee Shop,” Journal of Behavioral and
Howard, Daniel J. and Charles Gengler (2001), “Emotional Contagion Experimental Economics, 66, 16–22.
Effects on Product Attitudes,” Journal of Consumer Research, 28 Pauwels, Koen and Dominique M. Hanssens (2007), “Performance
(2), 189–201. Regimes and Marketing Policy Shifts,” Marketing Science, 26
Hu, Mu, Mingli Zhang, and Yu Wang (2017), “Why Do Audiences (3), 293–311.
Choose to Keep Watching on Live Video Streaming Platforms? An Pauwels, Koen, Jorge Silva-Risso, Shuba Srinivasan, and Dominique
Explanation of Dual Identification Framework,” Computers in M. Hanssens (2004), “New Products, Sales Promotions, and Firm
Human Behavior, 75, 594–606. Value: The Case of the Automobile Industry,” Journal of Market-
Jung, Minah H., Leif D. Nelson, Ayelet Gneezy, and Uri Gneezy ing, 68 (4), 142–156.
(2014), “Paying More When Paying for Others,” Journal of Per- Peng, Fuchun, Fangfang Feng, and Andrew McCallum (2004),
“Chinese Segmentation and New Word Detection Using Condi-
sonality and Social Psychology, 107 (3), 414–31.
tional Random Fields,” in Proceedings of the 20th International
Jung, Minah H., Hannah Perfecto, and Leif D. Nelson (2016),
Conference on Computational Linguistics: Association for Com-
“Anchoring in Payment: Evaluating a Judgmental Heuristic in
putational Linguistics. Stroudsburg, PA: Association for Compu-
Field Experimental Settings,” Journal of Marketing Research, 53
tational Linguistics, 562–68.
(3), 354–68.
Pennebaker, James W., Ryan L. Boyd, Kayla Jordan, and Kate
Kang, Charles, Frank Germann, and Rajdeep Grewal (2016),
Blackburn (2015), “The Development and Psychometric Proper-
“Washing Away Your Sins? Corporate Social Responsibility, Cor-
ties of LIWC 2015,” https://repositories.lib.utexas.edu/handle/
porate Social Irresponsibility, and Firm Performance,” Journal of
2152/31333.
Marketing, 80 (2), 59–79.
Pesaran, H. Hashem and Yongcheol Shin (1998), “Generalised
Kharpal, Arjun (2015), “Twitter Launches Live Broadcasting App
Impulse Response Analysis in Linear Multivariate Models,” Eco-
Periscope,” CNBC (March 26), https://www.cnbc.com/2015/03/
nomics Letters, 58 (1), 17–29.
26/twitter-launches-live-broadcasting-app-periscope.html.
Pugh, S. Douglas (2001), “Service with a Smile: Emotional Contagion
Kim, Ju-Young, Katharina Kaufmann, and Manuel Stegemann (2014),
in the Service Encounter,” Academy of Management Journal, 44
“The Impact of Buyer-Seller Relationships and Reference Prices (5), 1018–27.
on the Effectiveness of the Pay What You Want Pricing Mechan- Ransbotham, Sam, Nicholas H. Lurie, and Hongju Liu (2019),
ism,” Marketing Letters, 25 (4), 409–23. “Creation and Consumption of Mobile Word of Mouth: How Are
Kim, Ju-Young, Martin Natter, and Martin Spann (2009), “Pay What Mobile Reviews Different?” Marketing Science, 38 (5), 773–92.
You Want: A New Participative Pricing Mechanism,” Journal of Regner, Tobias (2015), “Why Consumers Pay Voluntarily: Evidence
Marketing, 73 (1), 44–58. from Online Music,” Journal of Behavioral and Experimental Eco-
Li, Feng (2008), “Annual Report Readability, Current Earnings, and nomics, 57, 205–14.
Earnings Persistence,” Journal of Accounting and Economics, 45 Russell, James. A. (2003), “Core Affect and the Psychological Con-
(2), 221–47. struction of Emotion,” Psychological Review, 110 (1), 145–72.
Liang, Jianping, Zengxiang Chen, and Jing Lei (2016), “Inspire Me to Schmidt, Klaus M., Martin Spann, and Robert Zeithammer (2015),
Donate: The Use of Strength Emotion in Donation Appeals,” Jour- “Pay What You Want as a Marketing Strategy in Monopolistic and
nal of Consumer Psychology, 26 (2), 283–88. Competitive Markets,” Management Science, 61 (6), 1217–36.
Love, Inessa and Lea Zicchino (2006), “Financial Development and Sims, Christopher A. (1980), “Macroeconomics and Reality,” Econ-
Dynamic Investment Behavior: Evidence from Panel VAR,” Quar- ometrica, 48 (1), 1–48.
terly Review of Economics, 46 (2), 190–210. Spann, Martin, Lucas Stich, and Klaus M. Schmidt (2017), “Pay What
Lu, Shijie, Dai Yao, Xingyu Chen, and Rajdeep Grewal (2020), “Do You Want as a Pricing Model for Open Access Publishing,” Com-
Large Audiences Generate Greater Revenues Under Pay What You munications of the ACM, 60 (11), 29–31.
Want: Evidence from a Live Streaming Platform,” working paper. Spann, Martin, Robert Zeithammer, Marco Bertini, Ernan Haruvy,
Luo, Ruixuan, Jingjing Xu, Yi Zhang, Xuancheng Ren, and Xu Sun Sandy D. Jap, Oded Koenigsberg, et al. (2018), “Beyond Posted
(2019), “PKUSEG: A Toolkit for Multi-Domain Chinese Word Prices: The Past, Present, and Future of Participative Pricing
Segmentation,” arXiv preprint. Mechanisms,” Customer Need and Solution, 5 (1/2), 121–36.
438 Journal of Marketing Research 58(3)

Srinivasan, Shuba, Koen Pauwels, Dominique M. Hanssens, and 2018 CHI Conference on Human Factors in Computing Systems.
Marnik G. Dekimpe (2004), “Do Promotions Benefit Manufactur- New York: Association for Computing Machinery, 1–13.
ers, Retailers, or Both?” Management Science, 50 (5), 617–29. Yazdani, Elham, Shyam Gopinath, and Steve Carson (2018),
Stephen, Andrew T. and Jeff Galak (2012), “The Effects of Traditional “Preaching to the Choir: The Chasm Between Top-Ranked
and Social Earned Media on Sales: A Study of a Microlending Reviewers, Mainstream Customers, and Product Sales,” Marketing
Marketplace,” Journal of Marketing Research, 49 (5), 624–39. Science, 37 (5), 838–51.
Trusov, Michael, Randolph E. Bucklin, and Koen Pauwels (2009), Yin, Dezhi, Samuel D. Bond, and Han Zhang (2014), “Anxious or
“Effects of Word-of-Mouth Versus Traditional Marketing: Find- Angry? Effects of Discrete Emotions on the Perceived Helpfulness
ings from an Internet Social Networking Site,” Journal of Market- of Online Reviews,” MIS Quarterly, 38 (2), 539–60.
ing, 73 (5), 90–102. Yin, Dezhi, Samuel D. Bond, and Han Zhang (2017), “Keep Your
Wang, Xiaodong, Ye Tian, Rongheng Lan, Wen Yang, and Xinming Cool or Let It Out: Nonlinear Effects of Expressed Arousal on
Zhang (2018), “Beyond the Watching: Understanding Viewer Perceptions of Consumer Reviews,” Journal of Marketing
Interactions in Crowdsourced Live Video Broadcasting Services,” Research, 54 (3), 447–63.
in Proceedings of IEEE Transactions on Circuits and Systems for Zhang, Kaifu, Theodoros Evgeniou, V. Padmanabhan, and Emile
Video Technology. Piscataway, NJ, IEEE, 3454–68. Richard (2012), “Content Contributor Management and Network
Wedel, Michel and P.K. Kannan (2016), “Marketing Analytics for Effects in a UGC Environment,” Marketing Science, 31 (3), 433–47.
Data-Rich Environments,” Journal of Marketing, 80 (6), 97–121. Zhang, Qiang, Wenbo Wang, and Yuxin Chen (2020), “In-
Wohn, Donghee Yvette, Guo Freeman, and Caitlin McLaughlin Consumption Social Listening with Moment-to-Moment Unstruc-
(2018), “Explaining Viewers’ Emotional, Instrumental, and Finan- tured Data: The Case of Movie Appreciation and Live Comments,”
cial Support Provision for Live Streamers,” in Proceedings of the Marketing Science, 39 (2), 285–95.
Copyright of Journal of Marketing Research (JMR) is the property of American Marketing
Association and its content may not be copied or emailed to multiple sites or posted to a
listserv without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.

You might also like