ArticlePDF Available

Abstract and Figures

Content platforms (e.g., newspapers, magazines) post several stories daily on their dedicated social media pages and promote some of them using targeted content advertising (TCA). Posting stories enables content platforms to grow their social media audiences and generate digital advertising revenue from the impressions channeled through social media posts’ link clicks. However, optimal scheduling of social media posts and TCA is formidable, requiring content platforms to determine what to post; when to post; and whether, when, and how much to spend on TCA to maximize profits. Social media managers lament this complexity, and academic literature offers little guidance. Consequently, the authors draw from literature on circadian rhythms in information processing capabilities to build a novel theoretical framework on social media content scheduling and explain how scheduling attributes (i.e., time of day, content type, and TCA) affect the link clicks metric. They test their hypotheses using a model estimated on 366 days of Facebook post data from a top 50 U.S. newspaper. Subsequently, they build an algorithm that allows social media managers to optimally plan social media content schedules and maximize gross profits. Applying the algorithm to a holdout sample, the authors demonstrate a potential increase in gross profits by at least 8%.
Content may be subject to copyright.
Article
Scheduling Content on Social Media:
Theory, Evidence, and Application
Vamsi K. Kanuri, Yixing Chen, and Shrihari (Hari) Sridhar
Abstract
Content platforms (e.g., newspapers, magazines) post several stories daily on their dedicated social media pages and
promote some of them using targeted content advertising (TCA). Posting stories enables content platforms to grow their
social media audiences and generate digital advertising revenue from the impressions channeled through social media posts’
link clicks. However, optimal scheduling of social media posts and TCA is formidable, requiring content platforms to
determine what to post; when to post; and whether, when, and how much to spend on TCA to maximize profits. Social
media managers lament this complexity, and academic literature offers little guidance. Consequently, the authors draw from
literature on circadian rhythms in information processing capabilities to build a novel theoretical framework on social media
content scheduling and explain how scheduling attributes (i.e., time of day, content type, and TCA) affect the link clicks
metric. They test their hypotheses using a model estimated on 366 days of Facebook post data from a top 50 U.S.
newspaper. Subsequently, they build an algorithm that allows social media managers to optimally plan social media content
schedules and maximize gross profits. Applying the algorithm to a holdout sample, the authors demonstrate a potential
increase in gross profits by at least 8%.
Keywords
circadian rhythms, content strategy, decision support system, genetic algorithm, social media
Online supplement: https://doi.org/10.1177/0022242918805411
More than 1.8 billion users worldwide spent an average of 118
min a day on social media in 2016 (Global Web Index 2016;
Mansfield 2016), and 77%of them actively engaged with
social media content through likes, comments, shares, and link
clicks (Statista 2016). Following this remarkable consumer
trend, content platforms (e.g., newspapers, sports websites,
magazines) frequently use social media to disseminate content
rapidly to their audiences (Kumar et al. 2016). ESPN.com, for
example, has more than 34 million Twitter page fans and posts
24 times per day, on average. People has approximately 6.8
million followers on its dedicated Facebook page and posts 28
stories per day, on average.
Building a social media following enables content platforms
to generate traffic ontheir own websites and increase their online
advertising revenue from impressions channeled through link
clicks of social media posts. However, content platforms are
struggling to develop profitable social media schedules to max-
imize website traffic originating from their social pages (CMO
Survey 2017; Collier 2017). To develop a profitable social media
schedule, a content platform must begin with the question, What
is the best time to post content on social media (i.e., timing)?
Moreover, social media websites allow content platforms to
advertise content in consumers’ social media news feed. Such
paid targeted content advertising (TCA) helps attract a new
audience base outside of a content platform’s current reach. This
raises a second question: When should content platforms sched-
ule advertised posts in correspondence with free posts (i.e., tim-
ing of TCA)? Furthermore, content platforms aim to design
content that better engages targeted users and drives users to
click on the posted stories (e.g., Lee, Hosanagar, and Nair
2018). In addition, when should content platforms schedule spe-
cific types of content (i.e., timing of content type)?
Existing social media management software platforms (e.g.,
Hootsuite, CoSchedule, Buffer, Tailwind, Post Planner, Sprout
Vamsi K. Kanuri is Assistant Professor of Marketing, Mendoza College of
Business, University of Notre Dame (email: vkanuri@nd.edu). Yixing Chen is
a doctoral student in Marketing, Mays School of Business, Texas A&M
University (email: y-chen@mays.tamu.edu). Shrihari (Hari) Sridhar is
Associate Professor of Marketing and Center for Executive Development
Professor, Mays School of Business, Texas A&M University (email: ssridhar@
mays.tamu.edu).
Journal of Marketing
1-20
ªAmerican Marketing Association 2018
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/0022242918805411
journals.sagepub.com/home/jmx
Social) do not offer a holistic solution to these questions.
1
Moreover, phone interviews with 15 social media professionals
of major content platforms (e.g., Dallas Morning News,News-
day,Baltimore Sun,Texas Tribune) indicated that they cur-
rently use simple rules of thumb to overcome the complexity
in social media content scheduling and indicated skepticism
about the profit-maximizing ability of their heuristics. In addi-
tion, barring anecdotal discussions (e.g., Collier 2017; Wolter
2017), social media scheduling has not been systematically
addressed in the academic literature, highlighting the urgency
to understand the drivers of effective social media scheduling
to justify return on social media investments (CMO Survey
2016; Mochon et al. 2017).
This study aims to address these shortcomings. Drawing on
the chronopsychology literature that shows that, for most people,
working memory availability is highest in the morning, lowest in
mid-afternoon, and moderate in the evening (Lupien et al. 2005),
we hypothesize that consumers’ desire to engage with content is
highest in the morning, moderate in the evening, and lowest in
the afternoon. Moreover, because the scarcity of working mem-
ory activates various actions intended to preserve working mem-
ory efficiency, we also hypothesize that the use of TCA and
content type (content with high-arousal emotions and content
requiring high cognitive processing) differentially affect link
clicks by time of day (morning, afternoon, and evening).
To test our hypotheses, we use data pertaining to 5,706 posts
on the Facebook page of a U.S. newspaper between December
31, 2014, and December 31, 2015. For robust identification of
our hypothesized effects, we consider strategic (nonrandom)
post allocation to consumers and account for endogeneity in
content platforms’ strategic decisions of content timing, con-
tent type, and TCA. We find strong support for our hypotheses,
thus empirically validating our framework.
Finally, we build and test an optimizer that incorporates
estimates from our econometric model to simultaneously deter-
mine the profit-maximizing mix of scheduling attributes (i.e.,
timing, content type, and TCA) over a given posting horizon.
We use a genetic algorithm to solve the implied multiobjective
large-scale optimization problem across several holdout peri-
ods. Our results indicate that the proposed solution can improve
the content platform’s profitability from its own digital adver-
tising by at least 8%.
Together, we make four contributions to marketing theory
and practice. First, we augment the burgeoning literature on the
drivers of social media content engagement (e.g., Akpinar and
Berger 2017; Berger and Milkman 2012; Toubia and Stephen
2013) by proposing time of day as a crucial driver of social
media content engagement. Our results imply that factoring
time-of-day effects into content scheduling is critical; for
example, posting content in the morning results in an 8.8%
(11.1%) increase in link clicks than doing so in the afternoon
(evening).
Second, we build a robust and replicable identification strat-
egy to demonstrate the impact of time of day, TCA, and content
type on link clicks. Specifically, we leverage (1) exogenous
shocks to content timing because of the nature of breaking
news and the institutional knowledge of the different functional
personnel responsible for crafting the Facebook message and
news article, and (2) the latent instrumental variables approach
to control for endogeneity. The combination of these methods
contributes to robust estimation of social media content
effectiveness.
Third, we show that time of day interacts with content type
and TCA to influence social media post performance and thus
add to the literature on paid social media advertising (e.g.,
Gong et al. 2017). For example, we show that employing TCA
in the afternoon generates 21%more link clicks compared with
doing so in the morning, and posting content that contains
high-arousal negative emotions in the afternoon is 1.6%less
effective in generating link clicks than in the morning. These
findings serve as guidelines for effective content scheduling
and allocation of marketing communication resources.
Fourth, in the spirit of contributing to both the rigor and
relevance of marketing literature (Kumar 2016), we present a
novel optimizer that works as a decision-support tool for social
media managers to profitably schedule content on social media.
Furthermore, we coded our algorithm using the genetic algo-
rithm feature in Microsoft Excel’s Solver, which greatly
enhances the managerial appeal of our proposed optimizer.
Next, we outline a theoretical framework to link key social
media scheduling attributes to postlevel performance metrics.
Subsequently, we describe our data and institutional context
and build an econometric model to validate our conceptual
framework. After discussing the results of our econometric
analysis, we describe our normative model as it pertains to
profit-maximizing social media schedules and illustrate an
application for our collaborating content platform. We con-
clude with a discussion of the key managerial takeaways and
possible extensions.
Theoretical Framework
Theoretical Extensions in the Social Media Content
Effectiveness Literature
Extant research on social media content effectiveness has
largely focused on how social media content characteristics
and TCA affect content engagement. For instance, prior
research has demonstrated that online content that evokes
high-arousal emotions leads to more virality (Berger and Milk-
man 2012) because it increases activation and elicits action-
related behaviors such as sharing and consumption (Gaertner
and Dovidio 1977). As such, content that elicits positive
(e.g., awe, amusement) or negative (e.g., anger, anxiety)
high-arousal emotions is more viral than content that does not.
1
Existing software can simultaneously post a firm’s content on multiple social
media platforms and allow managers to set up an inventory of posts at their
chosen time in the future, thereby saving significant time and increasing
efficiency. However, it lacks the prescriptive capability of suggesting what
content to post when and when to schedule TCA to maximize post link
clicks and implied advertising revenue.
2Journal of Marketing XX(X)
Likewise, content with high information value has been shown
to perform well online (Stieglitz and Dang-Xuan 2013) because
it elicits higher cognitive processing, which in turn fulfils con-
sumers’ self-enhancement goals (Wojnicki and Godes 2017)
and ability to generate social exchange value (Homans 1958).
Similarly, TCA is known to increase content engagement by
allowing content platforms to promote specific posts to broader
audiences on the basis of demographics, interests, and location
(Mochon et al. 2017). As such, TCA is a form of tailored
marketing communication that matches content with consu-
mers’ preferences and needs. Because content customization
increases the relevance of social media posts, TCA improves
content effectiveness by enhancing consumers’ propensity to
engage with social media content.
However, prior research has not explained how the efficacy
of psychological and cognitive traits embedded in social media
content can change during the day—a necessary input to under-
standing how to schedule content on social media. Similarly,
literature on TCA also falls short of explaining how the effec-
tiveness of TCA changes during the day. These limitations moti-
vate us to develop a novel framework around how diurnal
fluctuations in the psychological and cognitive traits embedded
in social media content and content targeting affect engagement.
Time-of-Day Effects in Social Behavior
What determines time-of-day effects in social behavior among
human beings? Research in chronopsychology has attributed
time-of-day effects to diurnal variation in an individual’s work-
ing memory availability and has found activation of inhibitory
processes to increase working memory efficiency during peri-
ods of low working memory availability. Working memory is a
“brain system that provides temporary storage and manipula-
tion of the information necessary for such complex cognitive
tasks as language comprehension, learning, and reasoning”
(Baddeley 1992, p. 556). It provides the necessary capabilities
of storing, retrieving, and processing immediate information.
For most people, working memory availability is highest when
they wake up in the morning, lowest in mid-afternoon, and
moderate in the evening (Lupien et al. 2005).
The availability of working memory affects an individual’s
psychological states and cognitive capabilities. For instance,
extant research has shown that high availability of working
memory in the morning is likely to make consumers more alert
(Tsaousis 2010), less creative (Giampietro and Cavallera 2007),
less innovative (Diaz-Morales 2007), and less pessimistic (Levy
1985) in the morning. More generally, diurnal variations in
working memory can cause sinusoidal cycles (or circadian
rhythms) in the level or intensity of people’s psychological states
and cognitive capabilities (Warner 1988). Such cycles can, in
turn, influence the perception of stimuli, judgments, and prefer-
ences (Hornik and Miniero 2009) and dictate consumers’ social
behavior (Dunlap, Loros, and DeCoursey 2004).
Research in chronopsychology has also attributed time-of-
day effects in social behavior to diurnal variation in inhibitory
processes that increase working memory efficiency. When
working memory availability decreases, the human brain auto-
matically activates several inhibitory processes to increase
working memory efficiency (Hasher, Lustig, and Zacks
2007). First, the brain gives preferential treatment to favorable
information triggered by external cues that can be easily refer-
enced from previously stored information (Myers et al. 2014).
For example, “you may be looking around your apartment for
your car keys and your phone simultaneously, holding templates
of both in your working memory as you scan your surroundings.
Suddenly the phone starts ringing, so you prioritize finding the
phone” (Myers, Stokes, and Nobre 2017, p. 450). Second, the
brain selectively inhibits processing new information that will
further drain working memory usage (Desimone and Duncan
1995). For example, when cortisol levels rise as a result of
anxiety, the human brain impairs the processing of visuospatial
information because it can further deteriorate working memory
availability (Shackman et al. 2006). Third, the brain minimizes
distracting tasks and tries to direct all cognitive resources to the
focal task (Hasher, Lustig, and Zacks 2007; Yoon, May, and
Hasher 1999). For example, when working memory availability
is reduced, “inhibitory mechanisms prevent irrelevant, off-task
information from entering working memory, thus limiting access
[of the working memory] to purely goal-relevant information”
(Yoon, May, and Hasher 1999, p. 91).
In the context of social media, consumers encode, process,
and decode social media posts in their working memory. Con-
sequently, consumers’ social media engagement (e.g., link
clicks) is reliant on their ability to process information in their
working memory. However, because the availability and effi-
ciency of working memory exhibit diurnal variations, we pur-
port that social media post performance is likely dependent on
diurnal variations in working memory.
We leverage the arguments on time of day, working memory
availability, and working memory efficiency to build a novel
theoretical framework for scheduling content on social media
(Figure 1). First, drawing on time of day and working memory
availability arguments, we hypothesize the main effect of time of
day and link clicks (H
1a–c
). Next, drawing on time of day and
working memory efficiency arguments, we posit that the main
effect of time of day on link clicks is moderated by content that
elicits positive and negative high-arousal emotions (H
2a–c
), con-
tent that requires higher cognitive processing (H
3a–c
), and TCA
(H
4a–c
).
Time of Day and Social Media Content Engagement
Conceptually, a day can be divided into four parts—morning,
afternoon, evening, and night—which we refer to as dayparts.
Because the majority (*98%in our empirical context) of social
media content is posted in the morning, afternoon, and evening
dayparts, we limit our theoretical discussion to these three day-
parts. Next, we present the arguments for our hypotheses (for an
overview of the logic employed, see Web Appendix W1).
For most social media content consumers, the availability of
working memory peaks in the morning. Higher availability of
working memory makes individuals more alert (Tsaousis
Kanuri et al 3
2010), attentive (Stopford et al. 2012), curious (Chamorro-
Premuzic and Furnham 2014), deliberative (Avery, Smillie,
and Fockert 2013), and information seeking in electronic envir-
onments (Marchionini 1997). However, as the day progresses,
people take on more tasks or accumulate more stress. Stress
causes cortisol levels to increase, which then impairs working
memory availability (Luethi, Meier, and Sandi 2009). Limited
availability of working memory limits people’s ability to pro-
cess new information and impairs their desire and ability to
engage with social media content.
Consequently, because working memory availability is
highest in the morning, lowest in mid-afternoon, and moderate
in the evening for most individuals (Lupien et al. 2005), we
theorize that the desire to engage with content will likely be
highest in the morning, moderate in the evening, and lowest in
the afternoon. As such, we posit:
H
1
: Ceteris paribus, (a) posting content in the afternoon
results in fewer link clicks than in the morning, (b) posting
content in the evening results in fewer link clicks than in the
morning, and (c) posting content in the afternoon results in
fewer link clicks than in the evening.
Time of Day and Effectiveness of Content with
High-Arousal Emotions
Previous research has concluded that online content that elicits
positive high-arousal emotions (e.g., awe, amusement) and
negative high-arousal emotions (e.g., anger, anxiety) receives
increased engagement owing to the activation of psychological
states (Berger and Milkman 2012). However, how does the
effectiveness of content with high-arousal emotions vary across
dayparts? As we have discussed, working memory availability
decreases from morning to afternoon and is moderate in the
evening. Therefore, in the evening (afternoon), when working
memory is more resource deprived than in the morning (eve-
ning), the brain selectively inhibits information that will further
drain working memory availability (Myers, Stokes, and Nobre
2017). Specifically, it focuses only on critical tasks achievable
with current working memory and filters out information that
could hinder it (Desimone and Duncan 1995). As such, inhibi-
tory mechanisms, which are responsible for the suppression of
irrelevant, off-task information, are activated when working
memory is struggling to process new information (Hasher, Lus-
tig, and Zacks 2007).
In the social media context, because content with high-
arousal emotions could further increase stress and cortisol lev-
els (Abercrombie, Speck, and Monticelli 2006; Kuhlmann,
Piel, and Wolf 2005; Tops et al. 2014),
2
which are known to
Gross Advertising Profits
Link Clicks
Targeted Content
Advertising (TCA)
Content type
High-arousal Emotions
Cognitive Processing
Time of Day
Control Variables
Content Platform Performance
Key Scheduling Attributes
Social Media Post Performance
Econometric Model
Normative Model and Algorithm
H1a, H1b, H1c
H4a, H4b, H4c
H2a, H2b, H2c
H3a, H3b, H3c
Figure 1. Conceptual Model.
2
Working memory identifies irrelevant information through textual cues (e.g.,
Kirschner 2002). When working memory processes information, it can identify
the emotions embedded within the content. Thereby, it can differentiate
between high-arousal and low-arousal information. As high arousal content
increases anxiety and cortisol levels, which further hinder the function of the
working memory when it is resource deprived, working memory signals the
brain to move away from such information (Kensinger and Corkin 2003). This
natural mechanism improves working memory efficiency.
4Journal of Marketing XX(X)
deplete working memory, we theorize that working memory
will deprioritize content with high-arousal emotions during peri-
ods of already constrained working memory. Because working
memory is most constrained in the afternoon, moderately con-
strained in the evening, and least constrained in the morning, we
conjecture that people will be less able to consumer content with
more high-arousal emotions when their working memory is
more depleted. As such, we theorize the following:
H
2
: Ceteris paribus, social media content with positive high-
arousal emotions (e.g., awe, amusement) and negative high-
arousal emotions (e.g., anger, anxiety) accumulate fewer
link clicks (a) in the afternoon than in the morning, (b) in
the evening than in the morning, and (c) in the afternoon
than in the evening.
Time of Day and Effectiveness of Content Requiring High
Cognitive Processing
Previous research has demonstrated that online content that
requires higher cognitive processing (e.g., insight, reason)
receives increased engagement because of its increased level
of cognitive involvement (Stieglitz and Dang-Xuan 2013).
However, how does the effectiveness of such content vary
across dayparts? As we have discussed, in the evening (after-
noon), when working memory is more resource deprived than
in the morning (evening), it inhibits irrelevant information by
minimizing distracting tasks and directing all available cogni-
tive resources to the focal task (Hasher, Lustig, and Zacks
2007; Yoon, May, and Hasher 1999). Inhibition fulfills two
crucial tasks in enhancing cognitive processing. First, it pre-
vents irrelevant and off-task information from entering the
working memory. Second, it deletes marginally relevant infor-
mation from working memory. Both tasks together minimize
competition from distracting information during information
encoding, retrieval, and processing in the working memory,
thereby increasing attention on focal information (Yoon, May,
and Hasher 1999) and improving analytical and cognitive pro-
cessing capabilities (Yoon and Lee 2004).
Within the social media context, individuals have better
inhibitory capabilities because the working memory is more
constrained. Thus, we theorize that the likelihood of people
consuming content (i.e., clicking on links) that requires higher
cognitive processing is highest in the afternoon, moderate in
the evening, and lowest in the morning. As such, we posit the
following:
H
3
: Ceteris paribus, social media content requiring higher
cognitive processing accumulates more link clicks (a) in the
afternoon than in the morning, (b) in the evening than in the
morning, and (c) in the afternoon than in the evening.
Time of Day and Effectiveness of TCA
Targeted content advertising enables content platforms to pro-
mote specific posts to broader audiences on the basis of demo-
graphics, interests, and location (Mochon et al. 2017).
Therefore, consistent with prior research, we expect a positive
association between TCA and link clicks.
However, how does TCA’s effectiveness vary across day-
parts? As we have discussed, working memory availability
decreasesfrommorningtoafternoonandthenmoderately
increases in the evening. Therefore, when working memory
is more resource deprived in the evening (afternoon) than in
the morning (evening), the brain prioritizes preferential infor-
mation and diverts available cognitive resources to this infor-
mation by biasing the receptive fields of neurons in the
information’s favor (Hillyard, Teder-Sa¨leja¨rvi, and Mu
¨nte
1998; Hillyard, Vogel, and Luck 1998). In neuropsychology
literature, this is commonly referred to as the “biased compe-
tition principle” (Desimone and Duncan 1995). However, to
activate preferential information processing, the working mem-
ory needs an external cue that can be easily referenced and
retrieved from long-term memory.
In the social media context, we theorize that TCA can serve
as an effective cue for the preferential processing of a social
media post. An individual’s interests and preferences are typi-
cally stored in his or her long-term memory and easily refer-
enced and retrieved to the working memory on demand
(Shiffrin and Atkinson 1969). When an individual is exposed
to TCA in the afternoon or evening, the working memory picks
it up as an external cue because TCA is sufficiently differen-
tiated from regular content in the news feed.
3
Subsequently, the
working memory prioritizes the advertised content over other
information in the news feed. Because TCA, by design, aligns
well with the individual’s interests and preferences, (s)he will
likely pull the template of the information from the long-term
memory, give it preferential processing, and engage with the
content (e.g., click on the post to read further). Thus, we expect
TCA to be most effective in the afternoon, moderately effective
in the evening, and least effective in the morning:
H
4
: Ceteris paribus, TCA on social media results in higher
link clicks (a) in the afternoon than in the morning, (b) in the
evening than in the morning, and (c) in the afternoon than in
the evening.
Empirical Analysis
Institutional Context and Social Media Metrics
Content platform description. We use data from a top 50 U.S.
newspaper (we refer to this as the “content platform” herein-
after) that generates revenue through print subscriptions, print
advertising, and digital advertising. The content platform has
been a local monopoly for several decades. It has a daily cir-
culation of *230,000 and weekend circulation of *336,000
and attracts *5.3 million monthly unique visitors to its
3
Social media sites are required by law to highlight advertised content within
the news feed. For instance, Facebook and LinkedIn explicitly identify TCA as
“sponsored” within an individual’s news feed. Such explicit identification
attracts attention and thus serves as an external cue (Samat, Acquisti, and
Babcock 2017).
Kanuri et al 5
website. The content platform reaches seven out of ten adults
with annual household incomes of $100,000 or more in the two
largest counties in its state.
Social media as a driver of website traffic. Like most U.S. news
organizations, the content platform views social media as a key
strategic lever to increase website traffic and digital advertising
revenue. The content platform has more than 350,000 fans on
its dedicated Facebook page and currently allocates more than
90%of its social media budget to Facebook. Each Facebook
post on the platform’s dedicated social media page includes a
web link to a corresponding full news story on the platform’s
website. Increasing website traffic through social media link
clicks helps the platform increase its digital advertising reve-
nue, as most advertisers pay for impressions. Digital advertis-
ing currently accounts for approximately 30%of the content
platform’s overall revenue and constitutes its fastest-growing
revenue source.
Need for a systematic strategy. In-depth interviews with the con-
tent platform’s social media manager, advertising director, and
content editor revealed that the firm currently employs ad hoc
rules of thumb, such as prioritize posting lifestyle and sports
news in the morning and waiting at least 30 min between posts,
to make daily scheduling decisions. While the content platform
realizes that arbitrary rules alleviate complexity, it desires a
model-based approach to maximize its digital advertising rev-
enues from impressions channeled through Facebook.
Uncontrollability of organic reach and focus on link clicks as the
dependent variable. We discuss two social media metrics that
drive traffic to the content platform’s website, organic reach
and link clicks, and comment on our choice of one metric over
the other. Organic reach is the total number of unique social
media users viewing the content platform’s posts in their news
feed for free. Maintaining a strong fan base helps maximize the
platform’s likelihood of engaging with its customers through
an unpaid distribution channel, in turn affecting brand equity
and word of mouth (Kumar et al. 2016; Naylor, Lamberton, and
West 2012). However, owing to increased competition in news
feed visibility, businesses have been experiencing a steady
decline in organic reach on Facebook (Boland 2014). Specifi-
cally, a high influx of posts from friends and other businesses a
user follows has pushed older posts to the bottom of the news
feed, making them less likely to gain exposure. Consequently,
Facebook instituted a relevance-based algorithm, EdgeRank, in
2014 to increase the exposure of relevant content to each Face-
book user. EdgeRank prioritizes stories on the basis of post
type (e.g., photo, video, link), affinity score between busi-
nesses’ dedicated Facebook page and users who view the
posted stories, and post recency (Constine 2014; Lee, Hosana-
gar, and Nair 2018).
However, because EdgeRank is a proprietary algorithm,
firms cannot determine whether their organic reach is due to
an individual’s choice to consume content or the algorithm’s
decision to show content to that individual. Thus, we do not
study organic reach but rather use total link clicks garnered by
each Facebook post on our collaborating content platform’s
dedicated page as our dependent variable. Unlike organic
reach, a link click is a deliberate action and reflects an individ-
ual’s revealed content preference. It also demonstrates an
instantaneous effect of post scheduling, thereby allowing the
firm to influence the metric.
TCA. Content platforms can also improve social media post
performance through TCA, commonly known as boosting
(Mochon et al. 2017). Facebook provides a content platform
with the opportunity to pay to reach users who are not sub-
scribed to the platform’s dedicated page on the basis of these
users’ demographics, interests, and location. When the content
platform boosts a post, it appears as an inline-ad on the news
feed of Facebook users who fit the targeting criterion. Thus,
TCA increases post engagement by reaching a wider audience
(Lovett and Staelin 2016). The higher engagement level that
results from TCA then prioritizes the post in the news feed of
Facebook users who are currently fans of the content platform,
further increasing link clicks.
Variable Operationalization
Link clicks. Our dependent variable “link clicks” refers to the
total number of clicks on the content platform’s link associated
with each Facebook post. Because link clicks are strictly pos-
itive, we use the logarithm of link clicks as the dependent
variable to alleviate distributional violations and account for
posts that receive abnormally high link clicks.
Time of day (dayparts). We specify four indicator variables to
capture time of day (dayparts) effects. Night (Daypart
1
) refers
to the period between 12:00 A.M. and 5:59 A.M., morning (Day-
part
2
) captures the period between 6:00 A.M.–11:59 A.M., after-
noon (Daypart
3
) is the period between 12:00 P.M. and 5:59 P.M.;
and evening (Daypart
4
) refers to the period between 6:00 P.M.
and 11:59 P.M. The respective indicator variable is equal to 1 if
a story’s posting time belongs to the daypart and 0 otherwise.
The baseline daypart is morning (i.e., Daypart
2
). Our collabor-
ating content platform is located in Pacific Time Zone, so our
time stamp corresponds to that time zone.
4
4
In our context, we have multiple sources of evidence to support that the
majority of the target audience (both readers and advertisers) are in the same
time zone. First, 99%of the subscribers to the print and online newspapers
come from one state located in the Pacific Time Zone. Second, Google Trends
data show that among the top 30 cities where searches of our collaborating
content platform are most popular, 27 cities are located in the Pacific Time
Zone.Third,98.5%of the print and online advertising revenue (in part
generated by redirecting to the online website from the Facebook page)
comes either from advertisers who are located only in the state or from local
advertising spend within the purview of local subsidiary of a national brand.
Finally, Audit Bureau of Circulation reports and the sales force pitch
documents of the content platform confirm that it competes locally by way
of its indicated presence in designated market areas. We thank an anonymous
reviewer for requesting this clarification.
6Journal of Marketing XX(X)
TCA. We use an indicator variable to capture whether a post is
advertised on Facebook (1 ¼yes, 0 ¼otherwise). The content
platform uses the same set of targeting filters (country ¼
United States, age range ¼22–65 years) and an identical TCA
budget of $100 across all advertised posts. Because we do not
observe variation in these two dimensions, we are only able to
assess the first-order effect of TCA (i.e., whether [¼1] or not
[¼0] a post was boosted) on link clicks.
High-arousal emotions. Following Berger and Milkman (2012),
we use an automated text analysis tool to quantify high-arousal
positive emotions (e.g., awe, amusement) and high-arousal
negative emotions (e.g., anger, anxiety) in Facebook posts. The
Linguistic Inquiry and Word Count (LIWC) program provides
the scale score of these two dimensions using LIWC2015 Dic-
tionary, which contains a list of 6,400 words, word stems, and
selected emoticons (Pennebaker et al. 2015).
Cognitive processing. Following Creswell et al. (2007) and Pen-
nebaker, Mayne, and Francis (1997), we use LIWC to calculate
the cognitive processing scale score (e.g., insight, causation).
The LIWC2015 Dictionary is an appropriate tool because it can
accommodate numbers, punctuation, short phrases, and informal
languages, allowing us to read the “netspeak common in social
media posts, and its internal reliability and external validity are
well supported in the literature (Pennebaker et al. 2015; Tausczik
and Pennebaker 2010). A sample of words and word stems of
three content types is available in Web Appendix W2.
Control variables. We include several control variables to
account for content- and environment-level heterogeneity.
First, we control for news topic categories. Our content plat-
form classifies its stories into eight categories: business, enter-
tainment, life, local, national, opinion, other, and sports. Each
content topic represents a substantive domain for the content
platform, with dedicated resources (e.g., editors, journalists),
and generates distinct costs per impressions from advertisers on
its website. Next, we control for the linear and quadratic terms
of interpost duration, operationalized as the minutes elapsed
between two subsequent posts. In addition, we include month
dummies and cluster standard errors by day of the week to
capture the unobserved temporal heterogeneity that might
influence a post’s link clicks (e.g., growth of the social media
platform, changes in external market conditions, popularity of
the newspaper industry). Finally, we control for content fea-
tures that might affect consumers’ perceptions, including mes-
sage length (i.e., word count) and text readability, which is
measured as the FOG index.
5
We present the notations of vari-
ables, measures, and data sources in Table 1.
Data Overview and Descriptive Statistics
Our data set comprises 5,706 individual posts from our content
platform’s dedicated Facebook page between December 31,
2014, and December 31, 2015. Our data are a snapshot of all
posts and the corresponding engagement on the content plat-
form’s Facebook page collected in June 2016. Therefore, all
posts in our data set reach their maximum lifetime engagement.
For each Facebook post, we observe the time stamp; original
link (a URL to the specific story on the content platform’s
website); message, title, and description of the post (for an
example, see Figure 2); whether the post is advertised; and key
performance indicators (e.g., link clicks).
On average, a post reaches 18,706 fans and obtains 967 link
clicks (for detailed descriptive statistics, see Web Appendices
W3 and W4). Both metrics exhibit considerable variation, with
organic reach ranging from 0 to 173,043 and link clicks ranging
from 0 to 152,448. On average, the interpost duration between
two posts is 68 min. Figure 3 shows noteworthy patterns in the
independent variables. Panel A shows that 2,040 stories (36%)
were posted in the morning, while 2,404 (42%), 1,135 (20%),
and 127 (2%) were posted in the afternoon, evening, and night
dayparts, respectively. Panel B shows that the majority of posts
are on local news (N ¼1,721, 30%), followed by sports
(N ¼1,026, 18%), and life (N ¼841, 15%). Next, among
518 targeted posts, we observe that 188 stories (36%) were
posted in the morning daypart, while 211 (41%), 111 (21%),
and 8 (2%) were posted in the afternoon, evening, and night
dayparts, respectively. Panel B also illustrates that sports stor-
ies (N ¼164) are among the most advertised, followed by local
(N ¼141) and life (N ¼64) stories. As Panel C shows, posted
stories have the highest level of positive high-arousal emotions
at night (4.53) and lowest level of positive high-arousal emo-
tions in the evening (2.85). However, posted stories have the
highest level of negative high-arousal emotions in the after-
noon (.49) and lowest level of negative high-arousal emotions
at night (.25). Finally, in Panel D, we see that posted stories
have the highest level of cognitive processing content in the
morning (7.67) and lowest level of cognitive processing con-
tent in the evening (7.31).
Econometric Model and Identification
We detail several empirical challenges that inhibit robust
model identification and subsequently present our correspond-
ing solutions (for a summary, see Web Appendix W5).
Strategic post allocation to consumers. As we have discussed,
organic reach is the total number of unique social media users
who view the content platform’s posts in their news feed for
free, and link clicks capture the number of users who clicked on
the post. However, the EdgeRank algorithm strategically deter-
mines whether users see the stories in their news feed and is
responsible for the organic reach a post obtains. Accordingly,
we model how our focal drivers affect link clicks by condition-
ing out this strategic behavior in two ways. First, we control for
5
The FOG index is the most commonly used metric to evaluate the lexical
complexity of texts. It indicates the number of years of formal education a
reader of average intelligence needs to understand text. Our results hold using
alternative measures such as the Flesch reading ease and Flesch–Kincaid
grade-level scores (Ghose and Ipeirotis 2011; Sridhar and Srinivasan 2012).
Kanuri et al 7
organic reach in the link clicks equation. By doing so, we model
the direct outcome of the EdgeRank algorithm (i.e., the number
of users who actually see the stories posted by the content plat-
form through unpaid distribution). Second, Facebook’s Edge-
Rank algorithm might display posts to consumers to induce
link clicks on the basis of characteristics other than those
included in organic reach. Therefore, we account for news topic,
month, and content features (i.e., message length and text read-
ability) to capture factors that induce strategic nonrandomness in
allocating posts. Thus, we have the following:
logðLink Click iÞ¼b0þb11 Night iþb12 Afternoon iþb13 Evening iþb2TCA iþb3Negemo arousal iþb4Posemo arousal i
þb5Cog process iþb61 Night iTCA iþb62 Afternoon iTCA iþb63 Evening iTCA i
þb71 NightiNegemo arousal iþb72 Afternoon iNegemo arousal i
þb73 Evening iNegemo arousal iþb81 Night iPosemo arousal i
þb82 AfternooniPosemo arousal iþb83 Evening iPosemo arousal i
þb91 NightiCog Process iþb92 Afternoon iCog Process iþb93 Evening iCog Process i
þb10 logðOrganic ReachiÞþY0Controls þei;
ð1Þ
Table 1. Variables, Notations, Measurements, and Data Sources.
Variable Notation Measurement Data Source
Dependent Variables
Link clicks log(Link Click
i
) Log of total number of clicks on content platform’s link associated
with each Facebook post
Facebook Insights
Independent Variables
Time of day Night
i
, Afternoon
i
,
Evening
i
1 if a story is posted in the corresponding daypart, 0 otherwise Facebook Insights
Targeted content
advertising
TCA
i
1¼yes, 0 ¼no Facebook Insights
High-arousal negative
emotions (message)
Negemo_arousal
i
LIWC scale score quantifying extent to which each message
evokes high arousal from negative emotions (e.g., anger)
Facebook Insights
High-arousal positive
emotions (message)
Posemo_arousal
i
LIWC scale score quantifying extent to which each message
evokes high arousal from positive emotions (e.g., amusement)
Facebook Insights
Cognitive processing
(message)
Cog_process
i
LIWC scale score quantifying extent to which each message
demands cognitive processing (e.g., insight)
Facebook Insights
Control Variables
Message length Number of words in Facebook message Facebook Insights
Message readability FOG index ¼.4 (average sentence length þ100 proportion
of difficult words)
Facebook Insights
Interpost duration Minutes elapsed between two subsequent posts Facebook Insights
News topic Categorical variable denoting eight topics: business, entertainment,
life, local, national, opinion, other, sports
Collaborating
Content Platform
Month Categorical variable with 12 values Facebook Insights
Organic reach log(Organic Reach
i
) Log of total number of unique people shown post through unpaid
distribution
Facebook Insights
Excluded Variables
Breaking tweets Breaking
ij
Average number of breaking tweets posted by Associated Press
and CNN Breaking News in each daypart
Twitter
High-arousal negative
emotions
(description)
D_negemo_arousal
i
LIWC scale score quantifying extent to which each description
evokes high arousal from negative emotions
Facebook Insights
High-arousal positive
emotions
(description)
D_posemo_arousal
i
LIWC scale score quantifying extent to which each description
evokes high arousal from positive emotions
Facebook Insights
Cognitive processing
(description)
D_cog_process
i
LIWC scale score quantifying extent to which each description
demands cognitive processing
Facebook Insights
Notes: “Message” refers to the text of the Facebook message; “description” refers to the text describing the news story.
8Journal of Marketing XX(X)
where i is the subscript for the Facebook post. b
11
,b
12
, and b
13
capture the effect of time of day on link clicks (with morning as
the baseline); b
2
captures the effect of TCA on link clicks; b
3
,
b
4
, and b
5
capture the effects of content types on link clicks;
b
61
,b
62
, and b
63
capture the three interactions between the time
of day dummies and TCA; and b
71
b
93
capture the nine inter-
actions between each of the time-of-day dummies and content
type. b
10
captures the effect of nonrandom post allocation on
link clicks, and a vector of covariates (Controls) is included.
Endogeneity of time of day. The social media manager of the
content platform is likely to decide the posting daypart strate-
gically drawing on private knowledge (e.g., expected number of
clicks), which we do not observe. This private knowledge cre-
ates a correlated unobservables problem because it influences
the posting daypart but resides in the error term. To alleviate
endogeneity bias from a correlated unobservables problem, we
use the control function approach (Petrin and Train 2010). Spe-
cifically, we estimate an auxiliary regression for posting deci-
sions in each daypart (i.e., the first stage). As a predictor in the
auxiliary regression, we need an excluded variable that meets
the relevance criterion (i.e., the excluded variables should be
correlated to the endogenous variable daypart) and the exclu-
sion restriction criterion (i.e., the excluded variables should not
be correlated to the shock in the dependent variable).
We use breaking news to identify our excluded variable. The
timing of breaking news is typically exogenous (e.g., the Air-
Asia crash), and content platforms such as newspapers push out
stories on such events as soon as possible to inform their audi-
ences. Thus, we collect all breaking news Twitter posts (tweets)
in 2015 reported by the Associated Press (@AP) and CNN
Breaking News (@cnnbrk), which receive a significantly larger
number of replies, shares (retweets), and likes compared with
regular tweets (p<.01) (for details, see Web Appendix W6).
The average number of breaking tweets posted by the Asso-
ciated Press and CNN Breaking News in a given daypart meets
the relevance criterion because more breaking events in a given
daypart (e.g., afternoon) affects the probability that our colla-
borating partner will post regular stories in the same daypart. In
other words, the original post planning in a given daypart is
more likely to be disrupted if the supply of breaking news in the
same daypart is higher. In Web Appendix W7, we present an
example showing how breaking news interrupts local newspa-
pers’ social media schedules. Here, our collaborating content
platform’s reporting of the AirAsia crash has pushed “life”
news that is unrelated to the crash and typically scheduled in
a given daypart down to the next time slot.
We validated this argument in interviews with a group of
social media professionals who work for content platforms,
including the Dallas Morning News,Newsday,Baltimore Sun,
Figure 2. News Story Content.
Notes: This figure is an example from the Associated Press (accessed January 22, 2018).
Kanuri et al 9
Texas Tribune, and others. Sample responses are as follows:
“We push breaking news out immediately and move the sched-
ule around as appropriate,” “We prioritize breaking news ahead
of the schedule,” and “We have a team of editors being ready to
get out urgent news at all times.” The first-stage results confirm
these intuitions (see Web Appendix W8).
The average number of breaking tweets in a corresponding
daypart also meets the exclusion restriction criterion because
breaking news events are external exogenous shocks (e.g., ter-
ror attacks, unexpected moves by North Korea) and are likely
uncorrelated with the anticipated link clicks of a news story
originally planned for the given daypart. Therefore, we esti-
mated the following first-stage model for each daypart:
Daypart
ij ¼a0þa1Breaking ij
þΛ1Controls þmij;and ð2aÞ
Daypart ij ¼1 if Daypart
ij >0;ð2bÞ
where Daypart
ij
is a binary variable indicating whether the
story i is posted in daypart j (j ¼1, 3, or 4 for night, afternoon,
or evening, respectively). Breaking
ij
is the average number of
breaking news tweets posted by the Associated Press and CNN
Breaking News in daypart j for each day in 2015. All other
covariates are as defined in Equation 1 to explain the likelihood
of posting in a given daypart. We then compute the inverse
Mills ratios (l
1i
,l
3i
,l
4i
) derived from each probit specification
and add them to Equation 1 to control for selection bias.
Endogeneity of content type. Similarly, social media manager is
likely to design each Facebook message strategically to induce
a larger number of link clicks drawing on private knowledge
(e.g., content types that elicit higher engagement) unobserved
us. This private knowledge creates a correlated unobservables
problem because it influences the content type of the Facebook
message but resides in the error term. For example, if the social
media manager receives a piece of relatively unbiased news to
be scheduled, (s)he may try to increase the arousal level in the
news by adding an anxiety-inducing spin to the content to
increase link clicks.
To address the endogeneity concern, we again use the con-
trol function approach (Petrin and Train 2010). We seek an
excluded variable that directly affects each of the three Face-
book message content types—the level of positive or negative
high-arousal emotions and level of cognitive processing
required—but only indirectly affects link clicks. We use each
of three content types in the story description as the excluded
AB
CD
127
2040
2404
1135
8
188 211 111
0
500
1000
1500
2000
2500
Night Morning Afternoon Evening
1721
1026
841
535 518 485 451
129
141 164 64 18 51 49 23 8
0
200
400
600
800
1000
1200
1400
1600
1800
4.53
2.90
3.24
2.85
0.25 0.47 0.49 0.42
0.00
0.50
1.00
1.50
2.00
2.50
3.00
3.50
4.00
4.50
5.00
Night Morning Afternoon Evening
High-arousal positive Emotions High-arousal negative emotions
7.59
7.67
7.44
7.31
7.10
7.20
7.30
7.40
7.50
7.60
7.70
7.80
Night Morning Afternoon Evening
Figure 3. Data Distribution. A: Distribution of Facebook Posts Across Dayparts. B: Distribution of Topic Categories. C: Arousal Level of
Content Across Dayparts. D: Level of Cognitive Processing Required Across Dayparts.
Notes: Black bars refer to all posts; gray bars refer to advertised posts.
10 Journal of Marketing XX(X)
variable for the corresponding content type in the Facebook
message (recall the difference between the Facebook message
and story description described in Figure 2).
Content types of story description meet the relevance criter-
ion because the Facebook message should carry the essence of
the story description, which is a summary of the original article.
In other words, the content of story description will explain, at
least partially, the content type of the Facebook message. We
confirmed this intuition by verifying the first-stage results (see
Web Appendix W8). Story description content types also meet
the exclusion restriction criterion because the story description
in the original article is not written by social media managers,
who might have expectations when engaging their audience,
but exogenously given by journalists or editors to social media
managers. Thus, we specify the following equations:
Negemo arousali¼g10 þg11 Dnegemo arousal i
þΛ2Controls þt1i;ð3aÞ
Posemo arousali¼g20 þg21 Dposemo arousal i
þΛ3Controls þt2i ;and ð3bÞ
Cog processi¼g30 þg31 Dcog process i
þΛ4Controls þt3i;ð3cÞ
where D_negemo_aoursal
i
, D_posemo_aoursal
i
, and D_cog_-
process
i
are scale scores of the three content types in the story
description, respectively. All other covariates are as previously
defined in Equation 1. The predicted residuals of c
t1i, c
t2i, and
c
t3i from Equations 3a, 3b, and 3c serve as effective control
variables to address the endogeneity concern.
Endogeneity of TCA. Finally, social media managers make TCA
decisions strategically in anticipation of a higher clicking prob-
ability or other factors unobservable to us. This strategic beha-
vior could render TCA endogenous to link clicks, because
correlated unobservables (e.g., expected future post perfor-
mance) drive both TCA decisions and content engagement.
Because of the lack of a clean exogenous TCA shifter in our
data set, we use a latent instrumental variables approach to
correct for possible endogeneity (Ebbes et al. 2005; Lee et al.
2015; Rutz, Bucklin, and Sonnier 2012). That is, we correct for
the endogenous regressor by introducing a discrete, unobserved
latent instrumental variable with m categories (m >1) that
partitions its variance into endogenous (possibly correlated
with the error term) and exogenous (uncorrelated with the error
term) components. Accordingly, we specify the following
equation:
TCA i¼yZiþt4i;ð4Þ
where i is the subscript for the post; TCA
i
denotes the endo-
genous TCA decision for post i; Z
i
is the unobserved discrete
instrument (uncorrelated with the error term in Equation 1);
and t
4i
refers to the error term, which is correlated with the
error term in Equation 1.
To obtain d
TCA i, we follow Wang, Gupta, and Grewal
(2017) and perform a latent class clustering, which splits TCA
i
into a manifest variable from a finite mixture of distributions.
For an m-cluster model, we can then predict every value of
TCA as
d
TCA i¼X
m
k¼1
ykpðCi¼kjTCA iÞ;ð5Þ
where y
1
,y
2
,...,y
m
is the latent cluster mean vector that
makes up TCA
i
;pð:Þis the predicted probability that a value
TCA belongs to cluster k. Using the Akaike information criter-
ion, we retain a two-cluster model. Because latent class mix-
tures, by definition, are computed by assuming that Z
i
is
uncorrelated with the error term in Equation 1, Z
i
is the unob-
served discrete instrument. Finally, we replace TCA
i
in Equa-
tion 1 with the predicted values of TCA from Equation 5
(d
TCA i) and add the error residual c
t4i as an additional control
variable (Wang, Gupta, and Grewal 2017). After correcting for
endogeneity of time of day, content type, and TCA, the full
model is specified as follows:
logðLink Click iÞ
¼b0þb11 Nightiþb12 Afternoon iþb13 Evening i
þb2d
TCA iþb3Negemo arousal iþb4Posemo arousal i
þb5Cog process iþb61 Night id
TCA i
þb62 Afternoonid
TCA iþb63 Evening id
TCA i
þb71 NightiNegemo arousal i
þb72 AfternooniNegemo arousal i
þb73 Evening iNegemo arousali
þb81 NightiPosemo arousal i
þb82 AfternooniPosemo arousal i
þb83 Evening iPosemo arousal i
þb91 NightiCog Process i
þb92 AfternooniCog Process i
þb93 Evening iCog Process iþb10 logðOrganic Reach iÞ
þY0Controls þd1l1i þd2l3i þd3l4i þd4c
t1i
þd5c
t2i þd6c
t3i þd7c
t4i þei
ð6Þ
where l1i;l3i ;l4i;c
t1i ;c
t2i ;c
t3i ;and c
t4i are terms correcting
for endogeneity, and all other covariates are as defined in
Equation 1.
Results
Table 2 presents the estimation results of Equation 6. We report
results from the auxiliary equations (Equations 2a–b, and 3a–c)
in Web Appendix W8. To compare afternoon with evening (as
opposed to using the morning daypart as the baseline), we
conducted a statistical test on the difference between b
12
and
b
13
, where the b
12
compares the effectiveness of afternoon with
Kanuri et al 11
morning, and b
13
compares the effectiveness of evening with
morning. Similarly, we also conducted statistical tests on the
differences between b
62
and b
63
,b
72
and b
73
,b
82
and b
83
, and
b
92
and b
93
.
Time of day. Our results suggest that posting content in the
afternoon results in fewer link clicks than in the morning
(b¼.104, p<.01), lending support to H
1a
. Furthermore,
posting content in the evening results in fewer link clicks than
in the morning (b¼.152, p<.01), lending support to H
1b
.
However, we do not find support for H
1c
, concerning the dif-
ferential impact of posting content in the evening and afternoon
(F ¼1.13, n.s.).
Content with high-arousal emotions. Consistent with prior
research (Berger and Milkman 2012), we find that content with
Table 2. Scheduling Attributes and Post Performance.
Log (Link Clicks)
Without Endogeneity Correction With Endogeneity Correction
Coef. SE Coef. SE
Night (12:00 A.M.–5:59 A.M.) .137 .100 .107 .106
Afternoon (12:00 P.M.–5:59 P.M.) .113*** .016 .104*** .014
Evening (6:00 P.M.–11:59 P.M.) .154*** .035 .152*** .037
TCA (1 ¼yes) .798*** .069 2.109*** .196
Night TCA .097 .126 .208* .098
Afternoon TCA .211** .062 .370*** .080
Evening TCA .040 .063 .187 .260
Negative emotions (message) .016** .005 .035 .022
Positive emotions (message) .002 .001 .029** .010
Cognitive processing (message) .002 .001 .013 .011
Night Negative emotions (message) .076** .024 .082** .025
Afternoon Negative emotions (message) .016** .005 .015* .006
Evening Negative emotions (message) .026* .013 .023 .015
Night Positive emotions (message) .004 .003 .003 .002
Afternoon Positive emotions (message) .001 .003 .001 .003
Evening Positive emotions (message) .001 .004 .000 .004
Night Cognitive processing (message) .003 .003 .003 .004
Afternoon Cognitive processing (message) .005** .002 .005** .002
Evening Cognitive processing (message) .007* .003 .007* .003
Log(organic reach) 1.742*** .061 1.740*** .063
Message length .003 .002 .003* .002
Message readability (FOG index) .007*** .001 .005** .002
Interpost duration .000 .000 .001 .001
Interpost duration
2
.000 .000 .000 .000
Local news dummy 1.275*** .159 .579*** .129
Business news dummy 1.310*** .177 .611*** .138
Sports news dummy 1.320*** .136 .662*** .109
Entertainment news dummy 1.413*** .176 .874*** .147
Life news dummy 1.454*** .177 .847*** .123
Opinion dummy 1.098*** .191 .369** .138
National news dummy 1.108*** .158 .465*** .088
TCA (LIV_error term) .100 .132
l
night
(inverse Mills ratio) .585 .615
l
afternoon
(inverse Mills ratio) 1.889 1.753
l
evening
(inverse Mills ratio) .031 .390
Negative emotions (residuals) .021 .020
Cognitive processing (residuals) .017 .011
Positive emotions (residuals) .031** .011
Intercept 11.945 .646 13.883*** 1.917
Day-of-week and month effects Yes Yes
Pseudo-R
2
.765 .773
N5,706 5,706
Notes: Standard errors in parentheses.
*p<.1.
**p<.05.
***p<.01.
12 Journal of Marketing XX(X)
high-arousal positive emotions is associated with higher link
clicks (b¼.029, p<.05). However, we do not find an asso-
ciation between high-arousal negative emotions and link clicks
(b¼.035, n.s.).
Interaction between time of day and content with high-arousal
emotions. We find that content with high-arousal negative emo-
tions garners fewer link clicks in the afternoon than in the
morning (b¼.015, p<.10), but we find no such evidence
for high-arousal positive emotions (b¼.001,n.s.).Therefore,
we find partial support for H
2a
. In addition, we do not find
support for H
2b
; that is, neither negative (b¼.023, n.s.) nor
positive (b¼.000, n.s.) high-arousal emotions garner fewer link
clicks in the evening than in the morning. In general, our lack of
support for positive high-arousal emotions could be because the
brain may not activate preferential treatment of information
when it encounters content with positive high-arousal emotions
because these emotions are less threatening to the working mem-
ory than content with negative high-arousal emotions. Finally,
with regard to the difference in the effectiveness between
emotion-filled content in the afternoon and evening dayparts
(H
2c
), we do not find significant differences for negative (F ¼
.750, n.s.) and positive (F ¼.170, n.s.) high-arousal emotions.
Interaction between time of day and content requiring cognitive
processing. We find evidence for a significant interaction
between timing and content that requires higher cognitive pro-
cessing. First, social media content based on higher cognitive
processing draws a larger number of link clicks in the afternoon
than in the morning (b¼.005, p<.05). This finding supports
H
3a
. Second, such social media content elicits higher link clicks
in the evening than in the morning (b¼.007, p<.10), lending
support to H
3b
. However, we do not find support for H
3c
, con-
cerning the differential impact of such social media content in
the afternoon and evening (F ¼.660, n.s.).
Interaction between time of day and TCA. We find that TCA
is more effective in the afternoon than morning (b¼.370,
p<.01), lending support to H
4a
. However, we do not find
support for H
4b
, which states that TCA is more effective in the
evening than morning (b¼.187, n.s.). In addition, we
observe that TCA is less effective at night than in the morning
(b¼.208, p<.10), likely because majority of the audience is
inactive at night. Finally, we do not find support for H
4c
, con-
cerning the differential effect of TCA in the evening and after-
noon (F ¼3.36, n.s.).
There could be several plausible explanations for the lack of
support for differences in reactions to content in the evening
versus in the afternoon. For instance, the stress levels among
the social media users in our sample could be consistent across
the afternoon and evening dayparts, resulting in identical work-
ing memory availability. Moreover, the difference in working
memory availability between afternoon and evening could be
less than the difference in working memory availability
between morning and afternoon, and morning and evening,
respectively.
Robustness Checks
Alternative definitions of daypart variables. There might be hetero-
geneity in how consumers view dayparts. Alternatively, we
redefine evening (daypart 4) to be between 6:00 P.M.and
9:59 P.M. and night (daypart 1) to be between 10:00 P.M. and
5:59 A.M. (i.e., sleep hours). Our results are robust to this alter-
native operationalization (see column 1, Web Appendix W9).
Lag error term. To further control for time-invariant unobserved
heterogeneity, we add a lagged error term (Jacobson 1990).
Note that we observe only one instance of performance metrics
for each of the 5,706 posts, so the lagged error term captures
unobserved heterogeneity that is time invariant and affects all
the posts uniformly. Results are robust to the addition of the
lagged error term (see column 2, Web Appendix W9).
Endogeneity of interpost duration. Interpost duration might also
represent a strategic decision by the social media manager. For
a story posted at a given time stamp, we use the number of
breaking tweets in the previous hour as the excluded variable
for interpost duration. Similar to our arguments in the identi-
fication section, the planned schedule is likely to be disrupted if
the supply of breaking news in the previous period is higher.
We confirm our intuition with the first-stage results. Our results
are robust to accounting for endogeneity in the interpost dura-
tion term (see column 3, Web Appendix W9).
Alternative solution for selection induced by Facebook’s EdgeRank
algorithm. Currently, we use the organic reach metric to account
for unobservable patterns in the exposure of social media con-
tent induced by the EdgeRank algorithm. Instead of organic
reach, one could also use number of impressions (i.e., the num-
ber of times when the content is displayed in a user’s news
feed) to account for patterns in the exposure of social media
content (Lee, Hosanagar, and Nair 2018). Thus, we use the log
of impressions (instead of the log of organic reach) as an alter-
native measure to correct for the Facebook algorithm. Again,
our results are robust to this alternative measure (see column 4,
Web Appendix W9).
Optimizing Scheduling Attributes, Post
Performance, and Firm Performance
Normative Model
The primary purpose of the econometric model was to illustrate
the theoretical linkages between the time of day, TCA and
content type, and link clicks. However, managers need a prac-
tical scheduling tool that recommends when to post (time of
day), whether to engage in TCA, and which content topic to
post at a certain time (e.g., sports, life, entertainment). We are
able to reconcile both the need for theory and practice in Equa-
tions 7–15, which contain estimates pertaining to time of day,
TCA, content type, content topic, and interpost duration.
However, from a social media manager’s standpoint, it is
not pragmatic to optimize the emotional and cognitive levels of
Kanuri et al 13
a post. Therefore, we deemphasize content type in the optimi-
zer and hold the emotional and cognitive levels at their respec-
tive median values.
6
Accordingly, in the normative model, we view the social
media manager’s objective as simultaneously determining time
of posting, interpost duration, and whether to employ TCA with
a capacity constraint on content topics and a constraint on the
number of posts that can be advertised. The objective of this
discrete optimization problem is represented as
max
fi;jgp¼X
8iX
8j
ðCPIiLink Clicks ijÞ Sj
()
cTCA:
ð7Þ
The objective function in Equation 7 represents the differ-
ence between revenue from social media scheduling and the
cost of TCA (cTCA). Revenue is obtained by multiplying link
clicks to the platform’s website from social media by the cost
per impression to advertise on the ith content topic of the plat-
form’s website (i ¼index of content topics 1–7) and S
j
,an
indicator variable capturing the decision to post a certain con-
tent topic in time slot j.
The cost of TCA is determined using the following equation:
cTCA ¼X
8jX
8i
TCA jContent Topic ij
CPC iLink Clicksij ;
ð8Þ
where TCA
j
is an indicator variable capturing the decision to
advertise a post in slot j, Content Topic
ij
represents whether the
social media manager has allocated content topic i in slot j,
CPC
i
indicates the average cost per click charged by Facebook
for topic i, and Link Clicks
ij
denotes the link clicks garnered by
topic i when posted in slot j.
Next, the social media manager must account for several
constraints as follows:
X
8j
Content Topic ij ¼Ci;ð9Þ
X
8i
Content Topicij 1;ð10Þ
TCA jX
8i
Content Topicij ;ð11Þ
Interpost Duration ij ¼
0ifk2f1;0g
ðTSk1TS kÞ30;otherwise
;
8
>
<
>
:ð12Þ
logðLink Click iÞ
¼b0þb11 Nightiþb12 Afternoon iþb13 Evening i
þb2d
TCA iþb3Negemo arousal iþb4Posemo arousal i
þb5Cog processiþb61 Night id
TCA i
þb62 Afternoonid
TCA iþb63 Evening id
TCA i
þb71 NightiNegemo arousal i
þb72 AfternooniNegemo arousal i
þb73 Evening iNegemo arousal i
þb81 NightiPosemo arousal i
þb82 AfternooniPosemo arousal i
þb83 Evening iPosemo arousal i
þb91 NightiCog Process i
þb92 AfternooniCog Process i
þb93 Evening iCog Processi
þb10 logðOrganic Reach iÞþY0Controls þd1l1i
þd2l3i þd3l4i þd4c
t1i þd5c
t2i þd6c
t3i þd7c
t4i ;
ð13Þ
X
8j
TCA jTCA Boosted;and ð14Þ
Sj2f0;1gContent Topic i2f0;1gTCA j2f0;1g:ð15Þ
Equation 9 ensures that the total number of posts within a
content topic i across all time slots sum to the number of stories
selected by the editor within the corresponding news topic.
Equation 10 ensures that the optimizer posts only one story per
time slot. Equation 11 ensures that the total number of stories
advertised is less than or equal to the total number of stories
available to be posted across all content categories. Equation 12
computes interpost duration. In particular, interpost duration is
assigned a value of 0 for the first post within the schedule;
otherwise, it is computed as the difference between the time
slot (TS) of the previous post and current post. Because each
time slot lasts 30 min, we multiply the difference by 30. Equa-
tion 13 uses the interpost duration, time of day, content topic,
daypart, and whether the firm decides to advertise the post (i.e.,
TCA), along with their respective regression weights, to predict
link clicks. We hold all other controls at their median values.
Equation 14 ensures that the total number of stories advertised
is less than or equal to the number of stories advertised in the
observed data.
Optimization Approach
The proposed optimizer presents a multidimensional, discrete,
nonlinear optimization problem for the social media manager.
For P
8i
Ciposts, the social media manager must decide which
time slots to select for each post, which posts to advertise, and
how many posts to advertise. For instance, assuming that there
are 25 30-minute slots (from 6 A.M.to6P.M.), the number of
6
For illustrative purposes, we also run the optimizer by holding the emotional
and cognitive levels of each post at low and high values, respectively.
14 Journal of Marketing XX(X)
ways the slots can be filled with rstories without replacement
is given by 25!/(25 r)!. If the content platform decides to post
one story from each content topic (seven stories in all), there
are more than 2.4 billion permutations. This is a conservative
estimate, as it excludes permutations for selecting stories to be
advertised. Consequently, although complete enumeration can
guarantee a global optimum solution, it is impractical and com-
putationally expensive. In fact, most discrete nonlinear combi-
natorial problems, such as product-line design problems in
marketing (e.g., Kanuri, Mantrala, and Thorson 2017), belong
to a special class of problems that are classified as NP-hard.
The global optimum to these problems is difficult to obtain
within polynomial time.
Therefore, we resort to heuristic techniques. Heuristics can
help with shrinking the problem space by applying well-
defined rules so the near-optimal solution can be found within
polynomial time. Depending on the formulation and complex-
ity within the Lagrange functions, one could use heuristics in
the attribute space, such as coordinate ascent, genetic algo-
rithm (GA), or simulated annealing; methods in the product
space, such as greedy heuristics, divide-and-conquer, or
product-swapping heuristics; or methods that evaluate par-
tially formed solutions, such as dynamic programming, beam
search, or nested partition heuristics. Belloni et al. (2008)
provide a comprehensive review of these techniques. We
choose the GA technique to implement our optimizer; GA
offers a superior ability to quickly arrive at a near-optimal
solution. Specifically, previous research has noted that GA
has a “higher probability of convergence to global optimum
solutions when data points are less, number of parameters is
large, the parameter space is multimodal, and the model is
inherently nonlinear” (Venkatesan, Krishnan, and Kumar
2004, p. 453). Because our parameter space is multidimen-
sional, nonlinear, and discrete-continuous, with gross profits
changing with content categories and time slots, GA is ideal.
Moreover, the availability of GA in Microsoft Excel enhances
its appeal, as one of our research goals is to develop a decision
support tool using a familiar interface for social media man-
agers. We provide additional details on the GA approach in
Web Appendix W10.
Profit-Maximizing Posting and TCA Schedule
Initial optimizer values. We use the coefficients of the estimated
model in Equation 6 to forecast link clicks. We obtain cost per
click and cost per impression values from our collaborating
content platform. The content platform’s costs per impression
for local, business, sports, entertainment, life, opinion, and
national stories are .06, .08, .08, .12, .10, .08, and .12 dollars,
respectively. The historical costs per click charged by Face-
book for local, business, sports, entertainment, lifestyle, opin-
ion, and national stories posted by our content platform are .04,
.07, .04, .05, .06, .03, and.07 dollars, respectively.
Establishing the baseline. We use posting and TCA schedules
from December 21–30, 2015, as the baseline for assessing
proposed optimizer’s performance. The baseline data, which
include 123 posts from seven content categories and 14 boosted
posts, constitutes our holdout sample. Table 3 illustrates the
distribution of posts across content categories. Cumulatively,
the posts in our holdout sample garner 49,920 link clicks,
which generates a gross profit of approximately $3,518 for the
content platform. This gross profit is a conservative estimate,
as it represents profit per advertisement on the firm’s website
and assumes a page depth (i.e., the number of pages a consumer
visits before exiting the website) of 1. Discussions with the
firm’s data analysts revealed that its webpages carry at least
five ads per page on average, and each visitor from Facebook is
believed to visit at least six pages before exiting. Factoring in
these average values would result in a gross profit of approx-
imately $105,540. However, because we do not have accurate
information on the total number of ads on a webpage for each
day, we restrict our comparison to gross profit per ad with the
assumption that page depth ¼1. Consequently, the observed
gross profit on each day between December 21 and December
30 serves as the baseline for evaluating the performance of the
content schedules predicted by our optimizer.
Results. We use the same starting values and stories as those
available to the social media manager between December 21,
2015, and December 30, 2015. We mimic the daily schedule of
a social media manager at our collaborating firm by allowing
Table 3. Optimizer Input.
Observed Posts across Topic Categories in the Holdout Sample
Local Business Sports Entertainment Life Opinion National Total # of Posts # of Posts Advertised
Monday 21-Dec 0 1 1 0 0 0 1 3 1
Tuesday 22-Dec 8 1 2 2 1 1 1 16 1
Wednesday 23-Dec 5 1 2 2 1 0 1 12 1
Thursday 24-Dec 3 2 2 2 2 2 1 14 1
Friday 25-Dec 2 0 3 1 0 0 1 7 1
Saturday 26-Dec 2 6 5 1 0 0 0 14 2
Sunday 27-Dec 2 2 7 1 0 1 2 15 1
Monday 28-Dec 3 2 5 0 2 0 0 12 0
Tuesday 29-Dec 4 2 1 1 3 1 3 15 2
Wednesday 30-Dec 4 2 5 0 2 1 1 15 4
Kanuri et al 15
the optimizer to create a schedule between 6 A.M. and 6 P.M.,
with 30-minute intervals. In addition, we restrict the optimizer
to the same number of TCAs as observed in the holdout sample
(see Table 3). Subsequently, we run the optimizer one day at a
time and document the predicted advertising revenue, adver-
tising cost, and gross profits for each day in the holdout sample.
Tables 4 and 5 summarize our results. As we have discussed,
we illustrate results from three scenarios in which the emo-
tional and cognitive levels of each post are held at their respec-
tive low, median, and high values. The proposed optimizer is
able to find a schedule that increases gross profits on every day
in the holdout sample. Across the ten-day period, the proposed
schedules generate $810.04, $901.86, and $4,004.91 in total
gross profits, which represent, on average, a 7.84%, 9.04%,
and 9.87%increase in daily gross profits from the baseline,
respectively.
The profit-maximizing schedules determined by our optimi-
zer look different from those in the baseline scenario. Table 5
compares the posting schedule predicted by our optimizer on
the last day in the holdout sample with the observed schedule.
As we can see, simply rearranging the posts without expending
additional resources can help the firm increase gross profits. In
summary, the optimizer increases profitability by reorganizing
the social media schedule to align content topic and timing with
performance and exploiting the benefit–cost trade-off to enable
simultaneous determination of TCA, along with content topic
and time of day.
Discussion
Content platforms have experienced a dramatic decline in print
advertising revenue and seek new practices to generate online
advertising revenue (Lambrecht and Misra 2017). One such
practice is to leverage social media channel to engage custom-
ers and direct traffic to websites. However, as we have dis-
cussed, a formidable challenge is to design a systematic
framework that enables social media managers to design
profit-maximizing social media schedules. This need is urgent
given practitioners’ call for effective scheduling strategies
(e.g., Collier 2017), sparse literature on scheduling content
on social media, and need for new knowledge in media
scheduling.
Accordingly, we fulfill this need in three steps. First, build-
ing on circadian rhythms literature, we provide novel insights
into how content effectiveness varies by the time of day, which
has typically been studied within the purview of behaviors such
as variety-seeking (Gullo et al. 2017), decision quality (Leone
et al. 2017), and risk-taking behavior (Wang and Chartrand
2015). Moreover, we offer a coherent theoretical framework
by theorizing how known drivers of social media engagement
(i.e., TCA and content type) interact with the time-of-day effect
to contribute to post performance. Second, we develop, esti-
mate, and validate a response model that simultaneously con-
siders attribute-based social media schedules involving time of
day, TCA, and content type using post-level data from a major
content platform. Third, we build a decision-support tool to
assist social media managers in profit-maximizing social media
content scheduling, and we show the profitability implications
over a finite planning horizon.
Managerial Takeaways
The estimates allow us to evaluate marginal effects of schedul-
ing attributes and thus conduct a set of what-if calculations.
From the calculations, we offer several key managerial take-
aways (note that we assume that each post attracts 967 link
clicks, the mean value in our data):
Takeaway 1: Timing of social media posts matters. Our estimates
on time-of-day effects suggest that, ceteris paribus, posting
stories in the morning generates approximately an 8.8%
(11.1%) increase in link clicks compared with posting stories
in the afternoon (evening). Assuming that page depth is 1 and
cost per impression is $0.06 (i.e., lowest observed return among
the seven categories), the 8.8%(11.1%) increase translates into
a gross profit of $25,529 ($32,201) for a content platform that
posts 5,000 free stories per year.
Takeaway 2: Deploy TCA at the right time. In the afternoon, on
average, TCA accumulates approximately a 21%increase in
link clicks compared with TCA in the morning. This 21%
increase translates into a $60,921 increase in advertising reve-
nue for a content platform that posts 5,000 stories per year. In
contrast, TCA at night, on average, decreases link clicks by
approximately 9.7%compared with TCA in the morning, lead-
ing to a loss of $28,140 in advertising revenue. These findings
contribute to the knowledge on boundary conditions of online
advertising effectiveness, such as personalization (Lambrecht
and Tucker 2013), obtrusiveness (Goldfarb and Tucker 2011),
and purchase funnel stage (Hoban and Bucklin 2015).
Takeaway 3: Post appropriate content type at the right time. Post-
ing social media content with negative high-arousal emotions
Table 4. Optimizer Results.
Low
(HANM, CP)
Median
(HANM, CP)
High
(HANM, CP)
% Increase in
Profits from
Observed Data
% Increase in
Profits from
Observed Data
% Increase in
Profits from
Observed Data
21-Dec 29.66% 23.01% 27.78%
22-Dec 18.46% 12.25% 13.34%
23-Dec 11.41% 1.41% 8.42%
24-Dec .27% 7.91% 6.92%
25-Dec 1.11% 5.19% .82%
26-Dec 3.70% 1.70% 16.63%
27-Dec 1.28% 4.31% 11.15%
28-Dec 1.74% .38% 1.89%
29-Dec 2.73% 24.67% 7.86%
30-Dec 8.02% 9.57% 3.93%
Ten-day average 7.84% 9.04% 9.87%
Notes: HANM ¼high-arousal negative emotions; CP ¼cognitive processing.
16 Journal of Marketing XX(X)
in the morning, on average, leads to a 1.6%(7.6%) increase in
link clicks compared with that in the afternoon (night). This
1.6%(7.6%) increase translates into $4,642 ($22,048) increase
in gross profits for the content platform that posts 5,000 stories
per year. Thus, we offer implications for online content virality
(Akpinar and Berger 2017) by underscoring the need to account
for content type depending on the time of the day. Specifically,
we suggest managers to weigh in on the interactions between
various content characteristics and day parts while designing
their social media message.
Takeaway 4: Timing reallocations pay off, even without budget
increases. Simply rearranging the posts without allocating addi-
tional budget for TCA can help the firm increase gross profits by
at least 8%on average over a ten-day horizon. This suggests that
our optimizer could be used as a decision-support tool to profit-
ably schedule content on social media without adding additional
resources. In fact, the managerial appeal of our scheduling tool,
which is developed in Microsoft Excel, significantly lowers the
hurdle of adoption of our prescriptive model within content plat-
forms. As such, 73%of managers we have interviewed have
explicitly expressed an interest in using our scheduling tool in
their operations. We provided an overview of an implementation
guide for managers in Web Appendix W11.
Takeaway 5: Spend advertising dollars wisely. Our analysis reveals
a nonlinear association between advertising spending (i.e.,
TCA costs) and gross profits. For instance, as we observe in
Web Appendix W12, additional spending on TCA will result in
only a marginal increase in gross profits, suggesting a concave
relationship between TCA and gross profits. Indeed, prior
research has shown that the relationship between increased
budgets on traditional media and optimized profits (conditional
on optimal allocation) is concave (e.g., Mantrala, Sinha, and
Zoltners 1992). Managers can use this finding to allocate bud-
gets effectively across multiple marketing communication
instruments including the TCA.
Limitations
Our work has some limitations that offer promising future research
avenues. First, our collaborating firm did not induce variation in
targeting filters while advertising content. Thus, we could not esti-
mate heterogeneity in TCA effectiveness with respect to those
filters. Future research could explore the role of targeting filters
on TCA effectiveness. Second, future research could explore the
effectiveness of TCA on the basis of topics discussed within the
content. Such fine-grained analysis could provide managers with
important guidelines on the allocation of TCA through the textual
characteristics of social media posts. Third, post-level data pre-
clude us from modeling how individuals allocate their time within
a daypart between Facebook browsing and other browsing activi-
ties. As individual data becomes increasingly available, future
research could address how other browsing options can affect
working memory allocation to Facebook content. Finally, we hope
managers and researchers use our econometric and optimization
model to generate empirical generalizations for other content plat-
forms (e.g., magazines, video sharing websites).
Acknowledgments
The authors thank the anonymous U.S. newspaper for sharing its Face-
book and financial data. They also thank the Reynolds Journalism
Institute at the University of Missouri for facilitating data collection.
Table 5. Sample Posting Schedule Predicted by Optimizer (December 30, 2015).
Current Schedule (Baseline)
Proposed Schedule
Low (HANM, CP) Median (HANM, CP) High (HANM, CP)
7:04:41 A.M. Local 6:00:00 A.M.Local 6:00:00 A.M. Local 10:00:00 A.M. Sports
7:27:11 A.M. National 8:30:00 Local 11:30:00 A.M. National 11:30:00 A.M. National
7:55:19 A.M.Business 9:00:00 A.M.Sports 12:00:00 Opinion 12:00:00 P.M. Opinion
8:34:41 A.M. Sports 11:30:00 A.M. National 12:30:00 Life 12:30:00 P.M. Life
9:53:26 A.M. Local 12:00:00 P.M. Opinion 1:00:00 P.M. Life 1:00:00 P.M.Life
10:32:49 A.M. Sports 12:30:00 P.M. Life 1:30:00 P.M. Sports 1:30:00 P.M. Sports
11:29:04 A.M. Opinion 1:00:00 P.M. Life 2:00:00 P.M. Sports 2:00:00 P.M. Sports
12:00:00 P.M. Local 2:00:00 P.M. Sports 2:30:00 P.M. Sports 2:30:00 P.M. Sports
12:33:45 P.M. Sports 2:30:00 P.M. Sports 3:00:00 P.M. Sports 3:00:00 P.M.Local
1:30:00 P.M.Business 3:00:00 P.M. Sports 3:30:00 P.M. Sports 3:30:00 P.M. Sports
2:00:56 P.M. Sports 4:00:00 P.M. Business 4:00:00 P.M. Business 4:00:00 P.M. Business
2:29:04 P.M. Life 4:30:00 P.M.Business 4:30:00 P.M.Business 4:30:00 P.M.Business
3:02:49 P.M. Life 5:00:00 P.M.Local 5:00:00 P.M.Local 5:00:00 P.M. Local
4:44:04 P.M.Local 5:30:00 P.M. Local 5:30:00 P.M.Local 5:30:00 P.M.Local
5:45:56 P.M.Sports 6:00:00 P.M. Sports 6:00:00 P.M. Local 6:00:00 P.M. Local
Ad revenue $197.35 (low), $214.82 (median), $922.34 (high) $211.65 $228.03 $1,020.01
Cost of TCA $100.93 (low), $109.88 (median), $470.73 (high) $107.50 $113.04 $550.64
Gross profit $96.42 (low), $104.94 (median), $451.60 (high) $104.15 $114.98 $469.36
% increase in profits from baseline 8.02% 9.57% 3.93%
Notes: Boldfaced values represent TCA posts. HANM ¼high-arousal negative emotions; CP ¼cognitive processing.
Kanuri et al 17
Area Editor
P.K. Kannan served as area editor for this article.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to
the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, author-
ship, and/or publication of this article.
References
Abercrombie, Heather C., Nicole S. Speck, and Roxanne M.
Monticelli (2006), “Endogenous Cortisol Elevations Are Related
to Memory Facilitation Only in Individuals Who Are Emotionally
Aroused,” Psychoneuroendocrinology, 31 (2), 187–96.
Akpinar, Ezgi, and Jonah Berger (2017), “Valuable Virality,” Journal
of Marketing Research, 54 (2), 318–30.
Avery, Rachel E., Luke D. Smillie, and Jan W. de Fockert (2013),
“The Role of Working Memory in Achievement Goal Pursuit,”
Acta Psychologica, 144 (2), 361–72.
Baddeley, Alan (1992), “Working Memory,” Science, 255 (5044),
556–59.
Belloni, Alexandre, Robert Freund, Matthew Selove, and Duncan
Simester (2008), “Optimizing Product Line Designs: Efficient
Methods and Comparisons,” Management Science,54(9),
1544–52.
Berger, Jonah, and Katherine L. Milkman (2012), “What Makes
Online Content Viral?” Journal of Marketing Research, 49 (2),
192–205.
Boland, Brian (2014), “Organic Reach on Facebook: Your Questions
Answered,” Facebook Business (June 5), https://www.facebook.
com/business/news/Organic-Reach-on-Facebook.
Chamorro-Premuzic, Tomas, and Adrian Furnham (2014), Personal-
ity and Intellectual Competence. New York: Psychology Press.
CMO Survey (2016), “Social Media Spending Triples but Falls Short
of Expectations,” (accessed July 10, 2018), https://cmosurvey.org/
2016/08/social-media-spending-triples-falls-short-expectations.
CMO Survey (2017), “Spending on Marketing Knowledge,” (accessed
July 10, 2018), https://cmosurvey.org/2017/06/spending-on-mar
keting-knowledge/.
Collier, Azure (2017), “How to Create a Social Media Posting
Schedule,” Constant Contact (April 28), https://blogs.constantcon
tact.com/social-media-posting-schedule/.
Constine, Josh (2014), “Why Is Facebook Page Reach Decreasing?
More Competition and Limited Attention,” Techcrunch (April 3),
https://techcrunch.com/2014/04/03/the-filtered-feed-problem/.
Creswell,J.David,SumanLam,Annette L. Stanton, Shelley E.
Taylor, Julienne E. Bower, and David K. Sherman (2007), “Does
Self-affirmation, Cognitive Processing, or Discovery of Meaning
Explain Cancer-related Health Benefits of Expressive Writing,”
Personality and Social Psychology Bulletin, 33 (2), 238–50.
Desimone, Robert, and John Duncan (1995), “Neural Mechanisms of
Selective Visual Attention,” Annual Review of Neuroscience,18
(1), 193–222.
ıaz-Morales, Juan Francisco (2007), “Morning and Evening-types:
Exploring Their Personality Styles,” Personality and Individual
Differences, 43 (4), 769–78.
Dunlap, Jay C., Jennifer J. Loros, and Patricia J. DeCoursey (2004),
Chronobiology: Biological Timekeeping. Sunderland, MA:
Sinauer.
Ebbes, Peter, Michel Wedel, Ulf Bo¨ckenholt, and Ton Steerneman
(2005), “Solving and Testing for Regressor-Error (In)Dependence
When No Instrumental Variables Are Available: With New Evi-
dence for the Effect of Education on Income,” Quantitative Mar-
keting and Economics, 3 (4), 365–92.
Gaertner, Samuel L., and John F. Dovidio (1977), “The Subtlety of
White Racism, Arousal, and Helping Behavior,” Journal of Per-
sonality and Social Psychology, 35 (10), 691–707.
Ghose, Anindya, and Panagiotis G. Ipeirotis (2011), “Estimating the
Helpfulness and Economic Impact of Product Reviews: Mining
Text and Reviewer Characteristics,” IEEE Transactions on Knowl-
edge and Data Engineering, 23 (10), 1498–1512.
Giampietro, Marina, and Guido M. Cavallera (2007), “Morning and
Evening Types and Creative Thinking,” Personality and Individ-
ual Differences, 42 (3), 453–63.
Global Web Index (2016), “Daily Time Spent on Social Networking
by Internet Users Worldwide from 2012 To 2016 (in Minutes),”
Statista (accessed July 10, 2018), https://www.statista.com/statis
tics/433871/daily-social-media-usage-worldwide/.
Goldfarb, Avi, and Catherine Tucker (2011), “Online Display Adver-
tising: Targeting and Obtrusiveness,” Marketing Science, 30 (3),
389–404.
Gong, Shiyang, Juanjuan Zhang, Ping Zhao, and Xuping Jiang (2017),
“Tweeting as a Marketing Tool: A Field Experiment in the TV
Industry,” Journal of Marketing Research, 54 (6), 833–50.
Gullo, Kelley, Jonah Berger, Jordan Etkin, and Bryan Bollinger
(2017), “Variety-Seeking and Time of Day,” working paper, 1–42.
Hasher, Lynn, Cindy Lustig, and R. T. Zacks (2007), “Inhibitory
Mechanisms and the Control of Attention,” in Variation in Work-
ing Memory, Andrew R. A. Conway, Christopher Jarrold, Michael
J. Kane, Akira Miyake and John N. Towse, eds. New York: Oxford
University Press, 227–49.
Hillyard, Steven A., Wolfgang A. Teder-Sa¨leja¨rvi, and Thomas F.
Mu
¨nte (1998), “Temporal Dynamics of Early Perceptual
Processing,” Current Opinion in Neurobiology, 8 (2), 202–10.
Hillyard, Steven A., Edward K. Vogel, and Steven J. Luck (1998),</