The Effect of Social Media Marketing Content on Consumer

Engagement: Evidence from Facebook

Dokyun Lee

Kartik Hosanagar

Harikesh S. Nair

The Wharton School

The Wharton School

Stanford GSB

First version: September 2013. This version: Oct 2014.

Abstract
We investigate the effect of social media content on customer engagement using a large-scale field
study on Facebook. We content-code more than 100,000 unique messages across 800 companies engaging
with users on Facebook using a combination of Amazon Mechanical Turk and state-of-the-art Natural
Language Processing algorithms. We use this large-scale database of content attributes to test the effect
of social media marketing content on subsequent user engagement defined as Likes and comments
with the messages. We develop methods to account for potential selection biases that arise from
Facebooks filtering algorithm, EdgeRank, that assigns messages non-randomly to users. We find that
inclusion of persuasive content like emotional and philanthropic content increases engagement with
a message. We find that informative content like mentions of prices, availability, and product features
reduce engagement when included in messages in isolation, but increase engagement when provided
in combination with persuasive attributes. Persuasive content thus seems to be the key to effective
engagement. Our results inform content design strategies in social media, and the methodology we
develop to content-code large-scale textual data provides a framework for future studies on unstructured
natural language data such as advertising content or product reviews.
Keywords: consumer engagement, social media, advertising content, marketing communication, large-scale
data, natural language processing, selection, Facebook, EdgeRank.

We thank seminar participants at the ISIS Conference (Jan 2013), Mack Institute Conference (Spring 2013), SCECR Conference (Summer 2013), WITS Conference (Dec 2013), and INFORMS Conference (Oct 2014) for comments, and a collaborating
company that wishes to be anonymous for providing the data used in the analysis. The authors gratefully acknowledge the
financial support from the Jay H. Baker Retailing Center and Mack Institute of the Wharton School and the Wharton Risk
Center (Russell Ackoff Fellowship). The authors also thank David Bell, Jonah Berger, Cexun Jeffrey Cai, Pradeep Chintagunta,
Pedro Gardete, Arun Gopalakrishnan, Raghuram Iyengar, Carl Mela, Navdeep Sahni, Olivier Toubia and Christophe Van den
Bulte for their helpful feedback. All errors are our own.

Introduction

Social networks are increasingly taking up a greater share of consumers time spent online. As a result,
social media which includes advertising on social networks and/or marketing communication with social
characteristics is becoming a larger component of firms marketing budgets. Surveying 4,943 marketing
decision makers at U.S. companies, the 2013 Chief Marketing Officer survey (www.cmosurvey.org) reports
that expected spending on social media marketing will grow from 8.4% of firms total marketing budgets in
2013 to about 22% in the next five years. As firms increase their social media activity, the role of content
engineering has become increasingly important. Content engineering seeks to develop content that better
engages targeted users and drives the desired goals of the marketer from the campaigns they implement.
This raises the question: what content works best? The most important body of academic work on this
topic is the applied psychology and consumer behavior literature which has discussed ways in which the
content of marketing communication engages consumers and captures attention. However, most of this work
has tested and refined theories about content primarily in laboratory settings. Surprisingly, relatively little
has been explored systematically about the empirical consequences of advertising and promotional content
in real-world, field settings outside the laboratory. Despite its obvious relevance to practice, Marketing
and advertising content is also relatively under emphasized in economic theory. The canonical economic
model of advertising as a signal (c.f. Nelson (1974); Kihlstrom and Riordan (1984); Milgrom and Roberts
(1986)) does not postulate any direct role for ad content because advertising intensity conveys all relevant
information about product quality in equilibrium to market participants. Models of informative advertising
(c.f. Butters (1977); Grossman and Shapiro (1984)) allow for advertising to inform agents only about price
and product existence yet, casual observation and several studies in lab settings (c.f. Armstrong (2010);
Berger (2012)) suggest that advertisements contain much more information and content beyond prices. In
this paper, we explore the role of content in driving consumer engagement in social media in a large-scale field
setting. We document the kinds of content used by firms in practice. We show that a variety of emotional,
philanthropic, and informative advertising content attributes affect engagement and that the role of content
varies significantly across firms and industries. The richness of our engagement data and the ability to
content code social media messages in a cost-efficient manner enables us to study the problem at a larger
scale than much of the previous literature on the topic.
Our analysis is of direct relevance to industry in better understanding and improving firms social media
marketing strategies. Many industry surveys (Ascend2, 2013; Gerber, 2014) report that achieving engagement on large audience platforms like Facebook is one of the most important social media marketing goals for
consumer-facing firms. Social media marketing agenciess financial arrangements are increasingly contracted
on the basis of the engagement these agencies promise to drive for their clients. In the early days of the
industry, it was thought that engagement was primarily driven by the volume of users socially connected
to the brand. Accordingly, firms aggressively acquired fans and followers on platforms like Facebook by
investing heavily in ads on the network. However, early audits of the data (e.g., Creamer 2012) suggested
2

that only about 1% of an average firms Facebook fans show any engagement with the brand by Liking,
sharing, or commenting on messages by the brand on the platform. As a result, industry attention shifted
from acquisition of social media followers per se, to the design of content that achieves better reach and
engagement amongst social media followers. In a widely reported example that reflects this trend (WSJ,
2012), General Motors curtailed its annual spending of $10M on Facebooks paid ads (a vehicle for acquiring
new fans for the brand), choosing instead to focus on creating content for its branded Facebook Page, on
which it spent $30M. While attention in industry has shifted towards content in this manner, industry still
struggles with understanding what kinds of content work better for which firms and in what ways. For
example, are messages seeking to inform consumers about product or price attributes more effective than
persuasive messages with humor or emotion? Do messages explicitly soliciting user response (e.g., Like this
post if ...) draw more engagement or in fact turn users away? Does the same strategy apply across different
industries? Our paper systematically explores these kinds of questions and contributes to the formulation of
better content engineering policies in practice.1
Our empirical investigation is implemented on Facebook, which is the largest social media platform in the
world. As alluded to above, many top brands now maintain a Facebook page from which they serve posts
and messages to connected users. This is a form of free social media marketing that has increasingly become
a popular and important channel for marketing. Our data comprises information on about 100,000 such
messages posted by a panel of about 800 firms over a 11-month period between September 2011 and July
2012. For each message, our data also contains time-series information on two kinds of engagement measures
Likes and comments observed on Facebook. We supplement these engagement data with message
attribute information that we collect using a large-scale survey we implement on Amazon Mechanical Turk
(henceforth AMT), combined with a Natural Language Processing algorithm (henceforth NLP) we build
to tag messages. We incorporate new methods and procedures to improve the accuracy of content tagging
on AMT and our NLP algorithm. As a result, our algorithm achieves great accuracy, recall, and precision
under 10-fold cross validation for almost all tagged content profiles.2 We believe the methods we develop
will be useful in future studies analyzing other kinds of advertising content and product reviews.
Our data has several advantages that facilitate a detailed study of content. First, Facebook messages
have rich content attributes (unlike say, Twitter tweets, which are restricted in length) and rich data on user
engagement. Second, Facebook requires real names and, therefore, data on user activity on Facebook is often
more reliable compared to other social media sites. Third, engagement is measured on a daily basis (panel
data) by actual message-level engagement such as Likes and comments that are precisely tracked within a
1 As of December 2013, industry-leading social media analytics firms such as Wildfire (now part of Google) do not offer
detailed content engineering analytics connecting a wide variety of social media content with real engagement data. Rather,
to the best of our knowledge, they provide simpler analytics such as optimizing the time-of-the-day or day-of-the-week to post
and whether to include pictures or videos.
2 The performance of NLP algorithms are typically assessed on the basis of accuracy (the total % correctly classified), precision
(out of predicted positives, how many are actually positive), and recall (out of actual positives, how many are predicted as
positives). An important tradeoff in such algorithms is that an increase in precision often causes decrease in recall or vice versa.
This tradeoff is similar to the standard bias-variance tradeoff in estimation.

closed system. These aspects make Facebook an almost ideal setting to study the role of content for this
type of marketing communication.
Our strategy for coding content is motivated by the psychology, marketing and economic literatures on
advertising (see Cialdini (2001); Bagwell (2007); Berger (2012); Chandy et al. (2001); Vakratsas and Ambler
(1999) for some representative overviews). In the economics literature, it is common to classify advertising as
informative (shifting beliefs about product existence or prices) or persuasive (shifting preferences directly).
The basis of informative content is limited to prices and/or existence, and persuasive content is usually treated
as a catch-all without finer classification. Rather than this coarse distinction, our classification follows the
seminal classification work of Resnik and Stern (1977), who operationalize informative advertising based
on the number of characteristics of informational cues (see Abernethy and Franke, 1996 for an overview of
studies in this stream). Some criteria for classifying content as informative include details about products,
promotions, availability, price, and product related aspects that could be used in optimizing the purchase
decision. Following this stream, any product oriented facts, and brand and product mentions are categorized
as informative content. Following suggestions in the persuasion literature (Cialdini, 2001; Nan and Faber,
2004; Armstrong, 2010; Berger, 2012), we classify persuasive content as those that broadly seek to influence
by appealing to ethos, pathos, and logos strategies. For instance, the use of a celebrity to endorse a product
or attempts to gain trust or good-will (e.g., via small talk, banter) can be construed as the use of ethos
appeals through credibility or character and a form of persuasive advertising. Messages with philanthropic
content that induce empathy can be thought of as an attempt at persuasion via pathos an appeal to a
persons emotions. Lastly, messages with unusual or remarkable facts that influence consumers to adopt a
product or capture their attention can be categorized as persuasion via logos an appeal through logic. We
categorize content that attempt to persuade and promote relationship building in this manner as persuasive
content. Though we believe we consider a larger range of content attributes than the existing literature, it
is practically impossible to detail the full range of possible content profiles produced on a domain as large
as Facebook (or in a data as large as ours). We choose content profiles that reflect issues flagged in the
existing academic literature and those that are widely used by companies on Facebook. We discuss this in
more detail in Section 2.
Estimation of the effect of content on subsequent engagement is complicated by the non-random allocation
of messages to users implemented by Facebook via its EdgeRank algorithm. EdgeRank tends to serve to users
messages that are newer and are expected to appeal better to his/her tastes. We account for the selection
induced by EdgeRank by developing a semi-parametric correction for the filtering it induces. One caveat
to the correction is that it is built on prior (but imperfect) knowledge of how EdgeRank is implemented.
In the absence of additional experimental/exogenous variation, we are unable to address all possible issues
with potential nonrandom assignment perfectly. We view our work as a large-scale, and relatively exaustive
exploratory study of content variables in social media that could be the basis of further rigorous testing
and causal assessment, albeit at a more limited scale. A fully randomized large-scale experiment that

provides a cross-firm and cross-industry assessment like provided here may be impossible or cost-prohibitive
to implement, and hence, we think a large-scale cross-industry study based on field data of this sort is
valuable.
Our main finding from the empirical analysis is that persuasive content drives social media engagement
significantly. Additionally, informative content tends to drive engagement positively only when combined
with such content. Persuasive content thus seem to be the key to effective content engineering in this
setting. This finding is of substantive interest because most firms post messages with one content type
or other, rather than in combination. Our results suggest therefore that there may be substantial gains to
content engineering by combining characteristics. The empirical results also unpack the persuasive effect into
component attribute effects and also estimate the heterogeneity in these effects across firms and industries,
enabling fine tuning these strategies across firms and industries.
Our paper adds to a growing literature on social media. Studies have examined the the diffusion of usergenerated content (Susarla et al., 2012) and their impact on firm performance (Rui et al., 2013; Dellarocas,
2006). A few recent papers have also examined the social media strategies of firms, focusing primarily on
online blogs and forums. These include studies of the impacts of negative blog messages by employees on
blog readership (Aggarwal et al., 2012), blog sentiment and quality on readership (Singh et al., 2014), social
product features on consumer willingness to pay (Oestreicher-Singer and Zalmanson, 2013), and the role of
active contributors on forum participation (Jabr et al., 2014). We add to this literature by examining the
impact of firms content strategies on user engagement.
An emerging theoretical literature in advertising has started to investigate the effects of content. This
includes new models that allow ad content to matter in equilibrium by augmenting the canonical signaling
model in a variety of ways (e.g. Anand and Shachar (2009)) by allowing ads to be noisy and targeted;
Anderson and Renault (2006) by allowing ad content to resolve consumers uncertainty about their matchvalue with a product; and Mayzlin and Shin (2011) and Gardete (2013) by allowing ad content to induce
consumers to search for more information about a product). Our paper is most closely related to a small
empirical literature that has investigated the effects of ad content in field settings. These include Bertrand
et al. (2010) (effect of direct-mail ad content on loan demand); Anand and Shachar (2011); Liaukonyte et al.
(2013) (effect of TV ad content on viewership and online sales); Tucker (2012a) (effect of ad persuasion on
YouTube video sharing) and Tucker (2012b) (effect of social Facebook ads on philanthropic participation).
Also related are recent studies exploring the effect of content more generally (and not specifically ad content)
including Berger and Milkman (2012) (effect of emotional content in New York Times articles on article
sharing) and Gentzkow and Shapiro (2010) (effect of newspapers political content on readership). Relative
to these literatures, our study makes two main contributions. First, from a managerial standpoint, we show
that while persuasive ad content especially emotional and philanthropic content positively impacts
consumer engagement in social media, informative content has a negative effect unless it is combined with
persuasive content attributes. This can help drive content engineering policies in firms. We also show how

the effects differ by industry type. Second, none of the prior studies on ad content have been conducted at the
scale of this study, which spans a large number of industries. The rigorous content-tagging methodology we
develop, which combines surveys implemented on AMT with NLP-based algorithms, provides a framework
to conduct large-scale studies that analyze the content of marketing communication.
Finally, the reader should note we do not address the separate but important question of how engagement
affects product demand and firms profits so as to complete the link between ad-attributes and those outcome
measures. First, the data required for the analysis of this question at a scale comparable to this study are still
not widely available to researchers. Second, as mentioned, firms and advertisers care about engagement per se
and are willing to invest in advertising for generating engagement, rather than caring only about sales. This is
consistent with our view that advertising is a dynamic problem and a dominant role of advertising is to build
long-term brand-capital for the firm. Even though the current period effects of advertising on demand may
be small, the long-run effect of advertising may be large, generated by intermediary activities like increased
consumer engagement, increased awareness and inclusion in the consumer consideration set. Thus, studying
the formation and evolution of these intermediary activities like engagement is worthwhile in order to
better understand the true mechanisms by which advertising affects outcomes in market settings. We note
other papers such as Kumar et al. (2013); Goh et al. (2013); Rishika et al. (2013); Li and Wu (2014); Miller
and Tucker (2013) as well as industry reports (comScore, 2013; Chadwick-Martin-Bailey, 2010; 90octane,
2012; HubSpot, 2013) have linked the social media engagement measures we consider to customer acquisition,
sales, and profitability metrics.

Data

Our dataset is derived from the pages feature offered by Facebook. The feature was introduced on Facebook
in November 2007. Facebook Pages enable companies to create profile pages and to post status updates,
advertise new promotions, ask questions and push content directly to consumers. The left panel of Figure 1
shows an example of Walmarts Facebook Page, which is typical of the type of pages large companies host
on the social network. In what follows, we use the terms pages, brands, and firms interchangeably. Our data
comprises posts served from firms pages onto the Facebook profiles of the users that are linked to the firm
on the platform. To fix ideas, consider a typical message (see the right panel of Figure 1): Pretty cool seeing
Andy giving Monfils some love... Check out what the pros are wearing here: http://bit.ly/nyiPeW. 3 In
this status update, a tennis equipment retailer starts with small talk, shares details about a celebrity (Andy
Murray and Gael Monfils) and ends with link to a product page. Each such message is a unit of analysis in
our data.
3 Retailer

picked randomly from an online search; not necessarily from our data.

Figure 1:

(Left) Example of a firms Facebook Page (Walmart). (Right) Example of a firms message and subsequent user
engagement with that message (Tennis Warehouse). Example is not necessarily from our data.

2.1
2.1.1

Data Description
Raw Data and Selection Criteria

To collect the data, we partnered with an anonymous firm, henceforth referred to as Company X that
provides analytics services to Facebook Page owners by leveraging data from Facebooks Insights. Insights is
a tool provided by Facebook that allows page owners to monitor the performance of their Facebook messages.
Company X augments data from Facebook Insights across a large number of client firms with additional
records of daily message characteristics, to produce a raw dataset comprising a message-day-level panel of
messages posted by companies via their Facebook pages. The data also includes two consumer engagement
metrics: the number of Likes and comments for each message each day. These metrics are commonly used in
industry as measures of engagement. They are also more granular than other metrics used in extant research
such as the number of fans who have Liked the page. Also available in the data are the number of impressions
of each message per day (i.e., the total number of users the message is exposed to). In addition, page-day
level information such as the aggregate demographics of users (fans) who Liked the page on Facebook or have
ever seen messages by the page are collected by Company X on a daily level. This comprises the population
of users a message from a firm can potentially be served to. We leverage this information in the methodology
we develop later for accounting for non-random assignment of messages to users by Facebook. Once a firm
serves a message, the messages impressions, Likes, and comments are recorded daily for an average of about
30 days (maximum: 126 days).4 The raw data contains about a million unique messages by about 2,600
unique companies.
4 A vast majority of messages do not get any impression or engagement after 7 days. After 15 days, virtually all engagements
and impressions (more than 99.9%) are accounted for. Unfortunately, reliable tabulation of shares is not available in the data.

The reader should note that as of this writing, our data is the most complete observational data available
outside of Facebook the data includes details such as demographics of page fans and engaged fans, which
cannot be scraped by outsiders (but are essential for correcting for EdgeRank ) but are available only to
the page owners via Facebooks Application Programming Interface. Our data also includes daily snapshots
of message-level engagement that Facebook does provide to page owners (Page owners must take snapshots
themselves if they want this data). These daily snapshots generate the within-message variation that enables
the panel analysis in our paper. Finally, page-owners do not have access to data on performance of any
messages by other pages, unlike our dataset which spans a large number of companies across sectors.
We clean the data to reflect the following criteria:
Only pages located in the US, and,
Only messages written in English, and,
Only messages with complete demographics data.
After cleaning, the data span 106,316 unique messages posted by 782 companies (including many large
brands) between September 2011 and July 2012. This results in about 1.3 million rows of message-level daily
snapshots recording about 450 million page fans responses. Removing periods after which no significant
activity is observed for a message reduces this to 665,916 rows of message-level snapshots (where activity
is defined as either impressions, Likes, or comments). The companies in our dataset are categorized into
6 broader industry categories following Facebooks page classification criteria: Celebrities & Public Figure
(e.g., Roger Federer), Entertainment (e.g., Star Trek), Consumer Products & Brands (e.g., Tesla Motors),
Organizations & Company (e.g., WHO), Websites (e.g., TED), Local Places & Businesses (e.g., MoMA).
2.1.2

Content-coded Data

We use a two-step method to label content. First, we contract with workers through AMT and tag 5,000
messages for a variety of content profiles. Subsequently, we build an NLP algorithm by combining several
statistical classifiers and rule-based algorithms to extend the content-coding to the full set of 100,000 messages. This algorithm uses the 5,000 AMT-tagged messages as the training data-set. We describe these
methods in more detail later in the paper.
The content in Facebook messages can be categorized as informative, persuasive, or both. Some messages
inform consumers about deals and discounts about products, while other messages seek to connect with
consumers on a personal level to promote brand personality, form relationships and are social in nature. We
call the first type informative content, and the second persuasive content. Some messages do both at the
same time by including casual banter and product information simultaneously (e.g., Are you a tea person
or a coffee person? Get your favorite beverage from our website: http://www.specific-link-here.com).
Table 1 outlines the finer classification of the attributes we code up, including precise definitions, summary statistics, and the source for coding the attribute. In Table 1, the 8 variables: BRANDMENTION,
8

DEAL, PRICECOMPARE, PRICE, TARGET, PRODAVAIL, PRODLOCATION, and PRODMENTION

are informative. These variables enable us to assess the effect of search attributes, brand, price, and product
availability information on engagement. The 8 variables: REMFACT, EMOTION, EMOTICON, HOLIDAYMENTION, HUMOR, PHILANTHROPIC, FRIENDLIKELY, and SMALLTALK are classified as persuasive.
These definitions include emotional content, humor, banter, and philanthropic content.
Informative content variables are identified using the seminal work by Resnik and Stern (1977), which
provides an operational guideline to identify informative content. Their work provides fourteen evaluative
content criteria to identify informative content that includes content such as product price and availability. Our persuasive content are identified mostly from existing consumer behavior research. For example,
emotional and humorous content have been identified as drivers of word-of-mouth and of viral marketing
tactics (Porter and Golan, 2006; Berger, 2012, 2011; Berger and Milkman, 2012). Philanthropic content
has been studied in the context of advertising effectiveness (Tucker, 2012b). Similarly, Berger and Schwartz
(2011) documented that the interestingness of content such as mentions of remarkable facts is effective in
generating word-of-mouth. For a survey of papers motivating our choice of persuasive variables, see Berger
(2012). While not fully exhaustive, we have attempted to cover most variables that are 1) highlighted by
prior research to be relevant, 2) commonly discussed in the industry.
Besides these main variables of interest, controls and content-related patterns noted as important in industry reports are profiled. We include these content categories to investigate more formally considerations
laid out in industry white papers, trade-press articles, and blogs about the efficacy of message attributes in
social media engagement. It includes content that explicitly solicits readers to comment or includes blanks
for users to fill out (thus providing an explicit option to facilitate engagement). Additionally, characteristics
like whether the message contained photos, website links, and the types of the page-owner (e.g., business
organization versus celebrity) are also coded. Other message-specific characteristics and controls include
metrics such as message length in characters and SMOG (Simple Measure of Gobbledygook), an automatically computed reading complexity index that is used widely. Higher values of SMOG implies a message is
harder to read. Table 2 shows sample messages taken from Walmarts page in December 2012 and shows
how we would have tagged them. The reader should note that some elements of content tagging and classification are necessarily subjective and based on human judgement. We discuss our methods (which involve
obtaining agreement across 9 tagging individuals) in section 2.2. All things considered, we believe this is one
of the most comprehensive attempts at tagging marketing communication related content in the empirical
literature.
2.1.3

Data Descriptive Graphics

This section presents descriptive statistics of the main stylized patterns in the data. The first thing we would
like to report is what kinds of content are used by firms (this may be useful for instance, for a researcher
interested in studying the role of specific content profiles in Facebook, who would like to know what content

Variable

Description

Source

Mean

SD

Min

Max

TAU ( )

Time since the post release (Day)

Facebook

6.253

3.657

16

LIKES

Number of Likes post has obtained

Facebook

48.373

1017

324543

COMMENTS

Number of Comments post has obtained

Facebook

4.465

78.19

22522

IMPRESSIONS

Number of times message was shown to users

Facebook

9969.2

129874 1

4.5107

SMOG

SMOG readability index (higher means harder to read)

Computed

7.362

2.991

25.5

MSGLEN

Message length in characters

Computed

157.41

134.54

6510

HTTP

Message contains a link

Computed

0.353

0.478

QUESTION

Message contains questions

Computed

0.358

0.479

BLANK

Message contains blanks (e.g. My favorite artist is __)

Computed

0.010

0.099

ASKLIKE

Explicit solicitation for Likes (e.g. Like if ...)

Computed

0.006

0.080

ASKCOMMENT

Explicit solicitation for Comments

Computed

0.001

0.029

Persuasive
REMFACT

Remarkable fact mentioned

AMT

0.527

0.499

EMOTION

Any type of emotion present

AMT

0.524

0.499

EMOTICON

Contains emoticon or net slang (approximately 1000

Computed

0.012

0.108

scraped from web emoticon dictionary e.g. :D, LOL)

HOLIDAYMENTION

Mentions US Holidays

Computed

0.006

0.076

HUMOR

Humor used

AMT

0.375

0.484

PHILANTHROPIC

Philanthropic or activist message

AMT

0.498

0.500

FRIENDLIKELY

Answer to question: Are your friends on social media

AMT

0.533

0.499

AMT

0.852

0.355

likely to post message such as the shown?

SMALLTALK

Contains small talk or banter (defined to be content other

than about a product or company business)
Informative

BRANDMENTION

Mentions a specific brand or organization name

AMT+Comp

0.264

0.441

DEAL

Contains deals: any type of discounts and freebies

AMT

0.620

0.485

PRICECOMPARE

Compares price or makes price match guarantee

AMT

0.442

0.497

PRICE

Contains product price

AMT+Comp

0.051

0.220

TARGET

Message is targeted towards an audience segment (e.g.

AMT

0.530

0.499

AMT

0.557

0.497

AMT

0.690

0.463

0.146

0.353

demographics, certain qualifications such as Moms)

PRODAVAIL

Contains information on product availability (e.g. stock

and release dates)

PRODLOCATION

Contains information on where to obtain product (e.g.

link or physical location)

PRODMENTION

Specific product has been mentioned

AMT+Comp

MSGTYPE

Categorical message type assigned by the Facebook

Facebook

- App

application related messages

Facebook

0.099

0.299

- Link

link

Facebook

0.389

0.487

- Photo

photo

Facebook

0.366

0.481

- Status Update

regular status update

Facebook

0.140

0.347

- Video

video

Facebook

0.005

0.070

PAGECATEGORY

Page category closely following Facebooks categorization

Facebook

- Celebrity

Singers, Actors, Athletes etc

Facebook

0.056

0.230

- ConsumerProduct

consumer electronics, packaged goods etc

Facebook

0.296

0.456

- Entertainment

Tv shows, movies etc

Facebook

0.278

0.447

- Organization

non-profit organization, government, school organization

Facebook

0.211

0.407

- PlaceBusiness

local places and businesses

Facebook

0.071

0.257

- Website

page about a website

Facebook

0.088

0.283

Table 1:

Variable Descriptions and Summary for Content-coded Data: To interpret the Source column, note that

Facebook means the values are obtained from Facebook, AMT means the values are obtained from Amazon Mechanical
Turk and Computed means it has been either calculated or identified using online database resources and rule-based methods
in which specific phrases or content (e.g. brands) are matched. Finally, AMT+Computed means primary data has been
obtained from Amazon Mechanical Turk and it has been further augmented with online resources and rule-based methods.

10

Sample Messages

Content Tags

Cheers!

BRANDMENTION, SMALLTALK,

Let Welchs help ring in the New Year.

HOLIDAYMENTION, EMOTION
Marias mission is helping veterans and their families find employment.

PHILANTHROPIC, SMALLTALK,

Like this and watch Marias story.

ASKLIKE, HTTP

http://walmarturl.com/VzWFlh

SMALLTALK, QUESTION,

On a scale from 1-10 how great was your Christmas?

HOLIDAYMENTION
Score an iPad 3 for an iPad2 price!
iPad 3.

Now at your local store, $50 off the

Plus, get a $30 iTunes Gift Card.

Offer good through 12/31 or

PRODLOCATION, PRODAVAIL,
PRICE

while supplies last.

Theyre baaaaaack!
local Walmart.

PRODMENTION, DEAL,

Now get to snacking again.

Find Pringles Stix in your

EMOTION, PRODMENTION,
BRANDMENTION,
PRODLOCATION

Table 2:

Examples of Messages and Their Content Tags: The messages are taken from 2012 December messages on
Walmarts Facebook page.

variables are used a lot by firms). Table 1 reports on the mean proportion of messages that have each
content characteristic. One can see that messages with videos, product or holiday mentions or emoticons are
relatively uncommon, while those with smalltalk and with information about where to obtain the product
(location/distribution attributes) are very common. Figure 2 reports on the co-occurrence of the various
attributes across messages. The patterns are intuitive. For instance, emotional and philanthropic content
co-occurs often, so does emotional and friend-like content, as well as content that describes product deals and
availability. Figure 2 also shows via a solid line how content types are clustered across messages.5 We see
that persuasive content types and informative content types are split into two separate clusters, suggesting
that firms typically tend to use one or the other in their messages. Later in the paper, we show evidence
suggesting that this strategy may not be optimal. Figure 3 shows the percentage of messages featuring a
content attribute split by industry category. We represent the relative percentages in each cell by the size of
the bubbles in the chart. The largest bubble is SMALLTALK for the celebrities category (60.4%) while the
smallest is PRICECOMPARE for the celebrities category (0%). This means that 6 in 10 messages by celebrity
pages in the data have some sort of small talk (banter) and/or content that does not relate to products or
brands; and that there are no messages by celebrity owned pages that feature price comparisons. Remarkable
facts (our definition) are posted more by firms in the entertainment category and less by local places and
businesses. Consistent with intuition, consumer product pages and local places/businesses post the most
about products (PRODMENTION), product availability (PRODAVAIL), product location (PRODLOC),
and deals (DEAL). Emotional (EMOTION) and philanthropic (PHILAN) content have high representation
in pages classified as celebrity, organization, and websites. Similarly, the AMT workers identify a larger
portion of messages posted by celebrity, organization and website-based pages to be similar to messages by
friends.
5 Clustered

with hierarchical clustering.

11

Figure 2: Co-occurrence of Attribute Characteristics Across messages. Shades in upper triangle represent correlations. Numbers in lower triangle represent the same correlations in numerical form in 100-s of units (range -100,+100). For e.g.,
the correlation in occurrence of smalltalk and humor across messages is 0.26 (cell [3,2]). The dark line shows the separation
into 2 clusters. Persuasive content and informative content attributes tend to form two separate clusters.

Industry Category VS Message Content Appearance Percentage

Websites

14

11

13

50

24

22

10

39

17

PlacesBusiness

10

40

53

39

18

36

31

Organization

21

12

16

14

50

44

28

17

Entertainment

17

12

48

46

24

18

ConsumerProduct

10

39

53

19

11

36

37

12

13

19

60

33

27

11

remfact

emotion

emoticon

holiday

humor

philan

friendlikely

smalltalk

brandmention

deal

pricecompare

price

target

prodavail

prodloc

prodmention

The labels on the bubbles are the percentages

Celebrity

Figure 3:

Bubble Chart of Broader Industry Category vs Message Content: Each bubble represents the percentage
of messages within a row-industry that has the column-attribute. Computed for the 5000 tagged messages. Larger and lighter
bubbles imply higher percentage of messages in that cell. Percentages do not add up to 100 along rows or columns as any given
message can have multiple attributes included in it. The largest bubble (60.4%) corresponds to SMALLTALK for the celebrity
page category and the smallest bubble (0%) corresponds to PRICECOMPARE for the celebrity category.

12

Log(Imp+1) VS Tau (time since post release) boxplot

Log(Imp+1)

15

10

0
1

10

11

12

13

14

15

16

Tau

Log(Comment+1) VS Tau (time since post release) boxplot

Log(Like+1) VS Tau (time since post release) boxplot

12

10

10
8
Log(Like+1)

Log(Comment+1)

6
4

0
1

10

11

12

13

14

15

16

Tau

Figure 4: :

10

11

12

13

14

15

16

Tau

Box Plots of Log(engagement+1) vs Time since message Release: Three graphs show the box plots

of (log) impressions, comments and Like vs. respectively. Both comments and Likes taper to zero after two and six days
respectively. Impressions take longer. After 15 days, virtually all engagements and impressions (more than 99.9%) are accounted
for. There are many outliers.

We now discuss the engagement data. Figure 4 shows box plots of the log of impressions, Likes, and
comments versus the time (in days) since a message is released ( ). Both comments and Likes taper off to
zero after two and six days respectively. The rate of decay of impressions is slower. Virtually all engagements
and impressions (more than 99.9%) are accounted for within 15 days of release of a message.
Figure 5 shows the average number of Likes and comments by message type (photo, link, etc.) over the
lifetime of a message. Messages with photos have the highest average Likes (94.7) and comments (7.0) over
their lifetime. Status updates obtain more comments (5.5) on average than videos (4.6) but obtain less Likes
than videos. Links obtain the lowest Likes on average (19.8) as well as the lowest comments (2.2). Figure
6 shows the same bar plots split across 6 industry categories. A consistent pattern is that messages with
photos always obtain highest Likes across industries. The figure also documents interesting heterogeneity in
engagement response across industries. The patterns in these plots echo those described in reports by many
market research companies such as Wildfire and comScore.
Figure 7 presents the average number of Likes and comments by content attribute. Emotional messages
obtain the most number of Likes followed by messages identified as likely to be posted by friends (variable:
FRIENDLIKELY). Emotional content also obtain the highest number of comments on average followed by
SMALLTALK and FRIENDLIKELY. The reader should note these graphs do not account for the marketsize (i.e., the number of impressions a message reached). Later, we present an econometric model that
incorporates market-size as well as selection by Facebooks filtering algorithm to assess user engagement
more formally.

13

Average number of likes and comments obtained

over lifetime by message type

Average Count

80

60
comment
like

40

20

0
link

app

status update

video

photo

Figure 5: Average Likes and Comments by Message Type: This figure shows the average number of Likes and comments
obtained by messages over their lifetime on Facebook, split by message type.

comment

400

like

300
200
100

comment

80

like

60
40
20

10
0

comment

60

like
40
20
0

link

te
pda deo
to
app tus u
vi
pho
a
t
s

te
pda deo
to
vi
pho

app tus u
sta

Websites
comment
Average Count

20

link

Organization
Average Count

like

30

like

te
pda deo
to
app tus u
vi
pho
a
st

link

comment

40

10

ConsumerProduct
Average Count

comment

0
te
pda deo
to
app tus u
vi
pho
a
st

link

Figure 6:

PlacesBusiness
Average Count

500

Entertainment

100
Average Count

Average Count

Celebrity

150

like

100
50
0

link

te
pda deo
to
app tus u
vi
pho
a
t
s

link

te
pda deo
to
vi
pho

app tus u
sta

Average Likes and Comments by Message Type by Industry: This figure shows the average number of

Likes and comments obtained by messages over their lifetime split by message type for each industry.

14

Average number of likes and comments obtained

over lifetime by message content

Average Count

150

100

comment
like

50

Figure 7:

prodloc

prodmention

prodavail

price

target

deal

pricecompare

smalltalk

brandmention

philan

friendlikely

humor

holiday

emotion

emoticon

remfact

Average Likes and Comments by Message Content:This figure shows the average number of Likes and

comments obtained by messages over their lifetime split by message content.

2.2

Amazon Mechanical Turk (AMT)

We now describe our methodology for content-coding messages using AMT. AMT is a crowd sourcing marketplace for simple tasks such as data collection, surveys, and text analysis. It has now been successfully
leveraged in several academic papers for online data collection and classification. To content-code our messages, we create a survey instrument comprising of a set of binary yes/no questions we pose to workers (or
Turkers) on AMT. To ensure high quality responses from the Turkers, we follow several best practices
identified in literature (e.g., we obtain tags from at least 9 different Turkers choosing only those who are
from the U.S., have more than 100 completed tasks, and an approval rate more than 97%. We also include
an attention-verification question.) Please see the appendix for the final survey instrument and the complete
list of strategies implemented to ensure output quality.
Figure 8 presents the histogram of Cronbachs Alphas, a commonly used inter-rater reliability measure,
obtained for the 5, 000 messages.6 The average Cronbachs Alpha for our 5, 000 tagged messages is 0.82
(median 0.84), well above typically acceptable thresholds of 0.7. About 87.5% of the messages obtained an
alpha higher than 0.7, and 95.4% higher than 0.6. For robustness, we replicated the study with only those
messages with alphas above 0.7 (4,378 messages) and found that our results are qualitatively similar.
At the end of the AMT step, approximately 2, 500 distinct Turkers contributed to content-coding 5, 000
messages. This constitutes the training dataset for the NLP algorithm used in the next step.

2.3

Natural Language Processing (NLP) for Attribute Tagging

We use NLP techniques to label message content from Facebook messages using the AMT-labeled messages as
the training data. Typical steps for such labeling tasks include: 1) breaking the sentence into understandable
building blocks (e.g., words or lemmas) and identifying sentence-attributes similar to what humans do when
6 Recall, there are at least 9 Turker inputs per message. We calculate a Cronbachs Alpha for each message by computing
the reliability across the 9 Turkers, across all the content classification tasks associated with the message.

15

Cronbachs Alphas for 5,000 Tagged

Messages Among 9+ Inputs

150

Counts

100

50

0
0.0

0.2

0.4

0.6

0.8

1.0

Cronbachs Alpha

Figure 8: Cronbachs Alphas for 5,000 Messages: This bar graph shows the inter-rater reliability measure of Cronbachs
Alpha among at least 9 distinct Turkers inputs for each 5,000 messages. The mean is 0.82 and the median is 0.84. We replicated
the study with only those above 0.7 and found the result to be robust.

reading; 2) obtaining a set of training sentences with labels tagged from a trusted source identifying whether
the sentences do or do not have a given content profile (in our case, this source comprise the 5000 AMTtagged messages); 3) using statistical tools to infer which sentence-attributes are correlated with content
outcomes, thereby learning to identify content in sentences. When presented with a new set of sentences,
the algorithm breaks the sentence down to building blocks, identifies sentence-level attributes, and assigns
labels using the statistical models that were fine-tuned in the training process. We summarize our method
here briefly. A detailed description of the algorithms employed is presented in the Appendix.
The use of NLP techniques has been gaining traction in business research due to readily available text
data online (e.g., Netzer et al. (2012); Ghose et al. (2012); Geva and Zahavi (2013)), and there are many
different techniques. Our NLP methods closely mirror cutting edge multi-step methods used in the financial
services industry to automatically extract financial information from textual sources (e.g., Hassan et al.
(2011)) and are similar in flavor to winning algorithms from the recent Netflix Prize competition.7 The
method we use combines five statistical classifiers with rule-based methods via heterogeneous ensemble
learning. Statistical classifiers are binary classification machine learning models that take attributes as
input and output predicted classification probabilities.8 Rule-based methods usually use large data sources
(a.k.a dictionaries) or use specific if-then rules inputted by human experts, to scan through particular words
or occurrences of linguistic entities in the messages to generate a classification. For example, in identifying
brand and product mentions, we augment our AMT-tagged answers with several large lists of brands and
products from online sources and a company list database from Thomson Reuters. Further, to increase the
range of our brand name and product database, we also ran a separate AMT study with 20,000 messages in
7 See

http://www.netflixprize.com.
use a variety of different classifiers in this step including logistic regression with L1 regularization (which penalizes
the number of attributes and is commonly used for attribute selection for problems with many attributes; see (Hastie et al.,
2009)), Naive Bayes (a probabilistic classifier that applies Bayes theorem based on presence or absence of features), and support
vector machines (a gold-standard algorithm in machine learning that works well for high dimensional problems) with L1 and L2
regularization and various kernels including linear, radial basis function, and polynomial kernels. We also utilize class-weighted
classifiers and resampling method to account for imbalance in positive and negative labels.
8 We

16

With Ensemble Learning (The

Best Performing Algorithm)
Accuracy Precision
REMFACT
EMOTION
HUMOR
PHILANTHROPIC
FRIENDLIKELY
SMALLTALK
DEAL
PRICECOMPARE
TARGETING
PRODAVAILABILITY
PRODLOCATION

Table 3:

0.94
0.97
0.98
0.97
0.94
0.85
0.94
0.99
0.98
0.96
0.97

Recall

0.99
0.99
1
0.99
0.99
0.88
0.99
0.99
0.99
0.99
0.99

0.68
0.87
0.90
0.85
0.68
0.80
0.65
1
0.89
0.76
0.90

Without Ensemble Learning

(Support Vector Machine version
1 + Rule-based)
Accuracy Precision
Recall
0.88
0.94
0.97
0.93
0.90
0.78
0.90
0.99
0.95
0.91
0.87

0.99
0.98
1
0.99
0.99
0.34
1
1
0.99
1
1

0.33
0.65
0.14
0.62
0.41
0.28
0.43
0.85
0.71
0.10
0.11

Performance of Text Mining Algorithm on 5000 Messages Using 10-fold Cross-validation: This table

presents metrics for performance of the classification algorithms used. The left 3 columns show the metrics for the final
algorithm which combines classifiers via ensemble learning methods while the right 3 columns shows the metrics for a support
vector machine algorithm. Notice that the support vector machine classifier tends to have low recall and high precision. Naive
Bayes tends to have high recall but low precision. Classifiers on their own cannot successfully overcome the standard precisionrecall tradeoff (if one is higher, the other is lower). But combining many different classifiers with ensemble learning can increase
both precision and recall. We obtain similar results for negative class labels.

which we asked AMT Turkers to identify any brand or product name included in the message. We added
all the brand and product names we harvested this way to our look-up database. We then utilize rule-based
methods to identify brand and product mentions by looking up these lists. Similarly, in identifying emoticons
in the messages, we use large dictionaries of text-based emoticons freely available on the internet.
Finally, we utilize ensemble learning methods that combine classifications from the many classifiers and
rule-based algorithms we use. Combining classifiers is very powerful in the NLP domain since a single statistical classifier cannot successfully overcome the classic precision-recall tradeoff inherent in the classification
problem. The final combined classifier has higher precision and recall than any of the constituent classifiers.
Assessment We assess the performance of the overall NLP algorithm on three measures, viz., accuracy,
precision, and recall (as defined in Footnote 2) using 10-fold cross-validation. 10-fold cross-validation is
computationally intensive and makes it harder to achieve higher accuracy, precision and recall, but we find
using the criterion critical to obtaining the external validity required for large scale classification. Table 3
shows these metrics for different content profiles. The performance is extremely good and comparable to
performance achieved by the leading financial information text mining systems (Hassan et al., 2011). We
also report the improvement of the final ensemble learning method relative to using only a support vector
machine classifier. As shown, the gains from combining classifiers are very substantial. We obtain similar
results for negative class labels.
As a final point of assessment, note that several papers in the management sciences using NLP methods
implement unsupervised learning which does not require human-tagged data. These techniques use existing
databases such as WordNet (lexical database for English) or tagged text corpus (e.g, tagged Brown Corpus)
to learn content by patterns and correlations. Supervised NLP instead utilizes human-taggers to obtain
17

a robust set of data that can be used to train the algorithm by examples. While unsupervised NLP is
inexpensive, its performance is significantly poor compared to that of supervised NLP algorithms like the
ones implemented here. Finally, To the best of our knowledge, the NLP method used in this paper that
uses ensemble learning to combine several statistical classifiers and rule-based methods, has not been used
in business research journals.9 Further, several current implementations of NLP do not utilize the strict
bar of utilizing the 10-fold cross-validation criterion. We believe one of the contributions of this paper is to
demonstrate how to utilize AMT in combination with ensemble learning techniques, to implement supervised
NLP in business research to produce robust and cost-efficient NLP algorithms that perform well at the scale
required for empirical work. We believe the method will be useful in future studies on unstructured natural
language data such as advertising content or product reviews. For interested readers, a detailed step-by-step
description of our NLP algorithms training and classification procedures is presented in the Appendix.

Empirical Strategy

Our empirical goal is to investigate the effect of message ad content on subsequent customer engagement.
Engagement the y-variable is observed in the data; and content the x-variables has been tagged
as above and is also observed. If messages are randomly allocated to users, the issue of assessing the
effect of message-content on engagement is straightforward; one simply projects y on x. Unfortunately, a
complication arises because Facebooks policy of delivery of messages to users is non-random: users more
likely to find a message appealing are more likely to see the message in their newsfeed, a filtering implemented
via Facebooks EdgeRank algorithm. The filtering implies a selection problem in estimation of the effect of
message-characteristics on engagement if we see that messages with photos are more likely to be commented
on by users, we do not know if this is the effect of including a photo in a message, or whether Facebook is more
likely to show messages with photos to users who are more likely to comment on them. To our knowledge, the
issue has been ignored in the literature on social media analysis so far.10 We address the selection issue via a
two-step procedure, first by building a semi-parametric model of EdgeRank that delivers an estimate of the
expected number of impressions a message is likely to receive, and then, by incorporating this model to run
a selectivity-corrected projection of Likes and comments on message characteristics in the second-stage. For
the first-stage, we exploit the fact that we observe the aggregated decisions of Facebook to serve impressions
to users, and that EdgeRank is based on three variables as revealed by Facebook: Type, Tie, and Time.11
Type (z) refers to the type of message. Facebook categorizes message-type into 5 classes: status update,
photo, video, app, or link.
9 Although there exist business research papers combining statistical classifiers and rule-based algorithms, to our knowledge,
none utilize ensemble learning methods which we find are critical in increasing accuracy, precision, and recall.
10 We discuss later in this section why other sources of confounds (like direct targeting by firms) are second-order in this
setting, compared to the selection induced by EdgeRank-based filtering.
11 As disclosed first at the 2010 f8 conference. See http://whatisEdgeRank.com for a brief description of EdgeRank. For the
duration of our data collection, this EdgeRank specification holds true.

18

 Tie (hijt ) refers to the affinity score between page j (company) and the Facebook user i (viewer of the
message) at time t which is based on the strength and frequency of the interaction history between the
user and the page.
Time ( ) refers to the time since the message.
Our dataset contains direct observations on the variables Type and Time. We do not have individual-level
data on a users history with pages to model tie strengths. However, we exploit the fact that we have access
to demographics data on the set of users who could potentially have been shown a message, versus who were
actually shown the message. The difference reflects the selection by EdgeRank, which we utilize as a proxy
measure of Tie-strength based targeting. Since we do not know the exact functional form of EdgeRanks
targeting rule, we work with a semi-parametric specification, utilizing flexible splines to capture the effect
of EdgeRank. At the end of this step, we thus develop a flexible approximation to EdgeRanks targeting.
In the second step, we can then measure the effect of ad content on Likes and comments, by controlling for
the non-random targeting using our first-stage model. Figure 9 shows the empirical strategy visually. One
advantage of directly modeling EdgeRank this way, is that we are also able to predict which message would
eventually reach users in addition to handing selection. This has auxiliary managerial value for advertisers.

Figure 9:

Impression-Engagement Funnel: Facebooks EdgeRank chooses subset of Page fans to show messages released

by the page and fans whove seen the message engage with the message based on content and type. EdgeRank is modeled with
a generalized additive model and the final engagement is estimated through aggregate logistic regression. Details of estimation
are in Sections 3.1 and 3.2.

3.1

First-stage: Approximating EdgeRanks Assignment

We represent message ks type in a vector zk , the time since message k was released in k , and the history of
user is past engagement with company j on Facebook in a vector hijt . Table 4 summarizes the notation.
(d)

To understand our procedure, let nkjt denote the number of users of demographic type d = 1, .., D who
(d)

(d)

were shown message k by firm j at time t. We refer to nkjt as impressions. We observe nkjt directly, and nkjt
is indirectly reported in the data and can be reverse-engineered from Company Xs reports. A description of
(d)

this procedure is provided in appendix. Let Njt denote the total number of users of demographic type d for
(d)

firm j on day t to whom the message can potentially be delivered. Njt is directly observed in the data, and
19

comprises all users of demographics d who have Liked the firm on Facebook. To be clear, note that Liking
a message is different from Liking a page Liking a page provides the firm that maintains that page an
(d)

opportunity to serve its messages to that user via Facebooks Newsfeed. Njt is a count of all such users.
(d)

Now, note that by EdgeRanks assignment rule, the aggregated impressions for demographic type d, nkjt ,
(d)

is an (unknown) function of liked-fans Njt , the tie strength between users within demographic bucket d and
(d)

the posting firm, hijt , the type of message zk , and time since message release k ,
(d)

(d)

(d)

(1)

E(nkjt ) = g(Njt , hijt , zk , k )

We do not observe individual-level data on each users i! s interaction with every message which could be
the basis of estimating Equation (1). Instead, we can construct the aggregated number of impressions and
liked-fans within a set of demographic buckets in the data. To use this variation as a source of approximating
EdgeRank, we approximate the RHS of Equation (1) as,
(d)

(d)

(d)

(2)

E(nkjt ) gd (Njt , 1j , zk , k )
(d)

where, we use a firm-demographic bin specific fixed effect, 1j , to capture the effect of user history. This
approximation would literally be true if all individuals within demographic bucket d had the same history
with firm j. In practice, this is not the case, and this may induce approximation errors into the procedure,
because additional history-heterogeneity within demographic buckets is not modeled (or is assumed into the
error term). This is a caveat to our analysis. Access to individual-level data could be the basis of improving
this procedure and relaxing this assumption. We view Equation (2) as a flexible approximation that allows
us to leverage the observed variation in firm-level impressions across demographics, while enabling us to
include firm and demographic-level fixed effects into a procedure that best approximates EdgeRank based
on what we as researchers (and firms) know about Facebooks filtering algorithm. We will also estimate the
right-hand function gd (.) separately for each demographic bucket, in effect allowing for slope heterogeneity
Notation

Description

User

Firm

message

Time (day)

zk

message ks media type (5 options: photo, video, status update, app, link)

Time since message k was released

hijt

History of user is past engagement with firm j

g(.)

EdgeRank score approximating function

nkjt

(d)

Impressions of message k by page j at time t by users in demographics bin d

(d)

Number of users of demographics bin d who Liked page j as of time t

Njt

(d)

Intercept term for each demographics d

.(d)

Parameters in EdgeRank approximation for demographics bin d

Table 4: User-level Setup Notation

20

in demographics in addition to intercept heterogeneity across demographics.

The next step relates to approximating the function gd (.). Since we do not know the exact functional
form of the above selection equation, we approximate the function semiparametrically via a Generalized
Additive Model (GAM) (c.f., Hastie and Tibshirani (1990)). The GAM is a generalized linear model with
additive predictors consisting of smoothed (e.g. interpolation and curve fitting) covariates. The GAM fits
the following flexible relationship between a set of covariates X and dependent variable Y ,
(E(Y |X1 , X2 , ..., Xp )) = + s1 (X1 ) + s2 (X2 ) + ... + sp (Xp )
where is a link function (e.g. gaussian, poisson, gamma), and s1 , s2 , ...sp are nonparametric smoothing
functions such as cubic splines or kernel smoothers. We model the EdgeRank selection equation for each
demographic d as the following,
5
!
"
#
(d) (d)
(d) (d)
(d)
(d)
(d)
(d)
hd log(nkjt + 1) = 0 + 1j + 2 Njt + s1 (Njt ; 3 ) +
4r I (zk = r)

(3)

r=2

16
#

(d)

(d)

5r I (k = r) + $kjt

r=2

(d)

where, hd gd1 (.) is the identity (Gaussian) link function, 0

mographic, d, and

is an intercept term unique to each de-

(d)
1j

is a firm-demographic fixed effect that captures the tie strength between the firm j

12

Njt is the number of fans of demographic d for firm j at time t and denotes the

and demographics d.

(d)

potential audience for a message. s1 is a cubic spline smoothing function, essentially a piecewise-defined
function consisting of many cubic polynomials joined together at regular intervals of the domain such that
the fitted curve, the first and second derivatives are continuous. We represent the interpolating function
%
&
$q
(d)
(d) (d)
(d)
s1 (.) as a linear combination of a set of basis functions b (.) and write: s1 (Njt ; 3 ) = r=3 br Njt 3r ,
(d)

where the br (.) are a set of basis functions of dimension q to be chosen and 3. are a set of parameters
to be estimated. We follow a standard method of generating basis functions, br (.), for the cubic spline
interpolation as defined in Wood (2006). Fitting the spline also requires choosing a smoothing parameter,
which we tune via generalized cross-validation. We fit all models via the R package mgcv described in Wood
(2006).
Finally, we include dummy variables for message-type (zk ) and for each day since release of the message
(k ; up to 16 days), to capture the effect of message-type and time-since-release semiparametrically. These
are allowed to be dspecific. We collect the set of parameters to be estimated for each demographic bucket
in a vector, .(d) , which we estimate by GAM estimation. The estimated parameter vector, denoted .(d) , d =
1, .., D, serves as an input to the second stage of the estimation procedure.
(d)

12 We

also tried Poisson and Negative Binomial link functions (since nkjt is a count variable), as well as the identity link
function without logging the y-variable. Across these specifications, we found the identity link function with log (y) resulted
in the best fit, possibly due to many outliers. We also considered specifications with numerous interaction of the covariates
included, but found they were either not significant or provided trivial gains in the R2 .

21

3.2

Second-stage: Modeling Engagement Given Message Assignment

We operationalize engagement via two actions, Likes and comments on the message. The selection problem
is that users can choose to Like or comment on a message only if they are served impressions, which
generates non-random censoring because impression assignment is endogenous to the action. We address
the censoring by including a correction for the fact that a user is shown a message non-randomly, estimated
(d) denotes the fitted estimate from the first-stage of the expected
semiparametrically as above. Suppose
kjt
number of impressions of message k for firm j amongst users of type d at time t,
%
&
(d) = gd N (d) , zk , k ; (d)

jt
kjt

(4)

We model the probability that users of type-d will Like a message given the full set of message characteristics,
Mkt , as logistic with parameters = (d , )d=1..D ,
d (Mkt ; ) =

1
1 + e(d +Mkt )

(5)

The parameter vector, , is the object of inference in the second stage.13

We will estimate by fitting the model to explain Qkjt , the observed number of Likes of the message in
each period in the data. To see the intuition for how our correction works in the estimation, note that we
can aggregate Equation (5) across users, so that the expected number of Likes is,

E(Qkjt ; )

D
#

(d)

kjt

d=1

'

1
1 + e(d +Mkt )

(6)

(d) are treated as known from the first-stage (Equation 4). The right-hand side is a weighted sum
with
kjt
of logit probabilities of Liking a message. Intuitively, the decision to Like a message is observed by the
researcher only for a subset of users who were endogenously assigned an impression by FB. The selection
(d) serve as weights that reweigh the probability of Liking to account for the fact that those users
functions
kjt
were endogenously sampled, thereby correcting for the non-random nature of message assignment when
estimating the outcome equation.
We could use the expectation in Equation (6) as the basis of an estimation equation. Instead, for efficiency,
we estimate the parameter vector by maximum likelihood. To set up the likelihood, note the expected
number of impressions of message k for firm j at time t across all demographic buckets is simply the sum,
kjt =

D
#
d=1

&
%
(d)

gd Njt , zk , k ; (d)

(7)

We can obtain an estimate of the implied probability that an impression picked at random from the pool is
13 Allowing to be d-specific as well in Equation (5) is conceptually straightforward. Unfortunately, this results in parameter
proliferation and trouble with convergence; hence we settled for a more limited specification with d-specific intercepts.

22

of type-d,
(dkt =

(d)

kjt
kjt

(8)

Thus, the probability (Mkt ; ) that an impression picked at random from the pool will Like the message
given a guess of , is,

(Mkt ; ) =

D
#

Pkt (d) Pkt (Like|d) =

d=1

D
#

(dkt d (Mkt ; )

(9)

d=1

Intuitively, with probability Pkt (d) = (dkt an impression is of type-d, and with probability P (Like|d) =
d (Mkt ; ), an impression will Like the message conditional on being type-d; hence the unconditional probability a random impression will Like the message is the sum-product of these marginals and conditionals
across all D types.
The number of Likes is a count variable for which we specify a Binomial likelihood. Accordingly, the
kjt assigned impressions are observed to Like the message, and that
probability that Qkjt out of the
kjt Qkjt of the remaining impressions are observed not to, is binomial with probability, (Mkt ; ),

kjt , (Mkt ; ))
Qkjt Binomial(

(10)

kjt as given, then delivers estimates

Maximizing the implied binomial likelihood across all the data, treating
of . The intuition for the selection correction here is the same as that encapsulated in Equation (6). We
can repeat the same procedure using the number of comments on the message as the dependent variable
so as the recover the effect of message-characteristics on commenting as well. This two-step procedure thus
delivers estimates of the effects of message-characteristics on the two outcomes of interest. Standard errors
are obtained by bootstrapping both steps 1 and 2 over the entire dataset.
Discussion of Identification Identification in the model derives from two sources. First, we exploit
the observed discrepancy in demographic distributions between the set of individuals to whom a message
could have been served, versus those who were actually served. The discrepancy reflects the filtering by
EdgeRank. Our first stage essentially projects this discrepancy onto message-type, time-since-release, page
and demographic characteristics in a flexible way. This essentially serves as a quasi control function that
corrects for the selectivity in the second stage (Blundell and Powell, 2003), where we measure the effect of
message characteristics on outcomes. The second source of identification arises from exploiting the implied
restriction that the rich set of AMT-content-coded attributes affect actual engagement, but are not directly
used by EdgeRank to assign messages to users. This serves as an implicit exclusion that helps address
selection. The only message-characteristic used by EdgeRank for assignment is zk , which is controlled for.
Thus, any systematic correlation in outcomes with AMT-content-coded characteristics, holding zk fixed, do
not reflect selection-related considerations. One caveat is the control for selection does depend on assumptions

23

we made about EdgeRank based on what is known publicly. We used a flexible first-stage specification so
as to be as robust as possible to these assumptions. Notwithstanding these aspects, to the best of our
knowledge, the full details of EdgeRank are not known to any firm or researcher. In our view, a perfect
solution to the selection problem is unlikely to be achieved without full knowledge of Facebooks targeting
rule.
We concentrated on EdgeRank-induced selection as the main difficulty in inference since we believe the
specifics of the Facebook environment makes several other sources of confounds second-order compared
to the effect of EdgeRank. For instance, one concern may be that firms may target content directly to
specific users on Facebook. This is unlikely because in contrast to Facebooks banner advertisements, the
Facebook page environment does not allow companies to target specific audiences (the only factor that can
be controlled is the time-of-day of release of the message). Rather, all targeting is implicitly implemented
by Facebook via EdgeRanks filtering. Another story may be that firms observe that a particular type of
content receives significant engagement, and subsequently start posting similar content. Thus, new content
reflects past engagement. Our data shows significant within-variation in the attributes of messages launched
over time by a given firm, which is inconsistent with this story which would predict instead significant withinfirm persistence in these attributes. A final concern is that engagement is driven by unmeasured message
characteristics that co-occur with included message characteristics. To the extent that these unmeasured
message characteristics drive engagement, they represents unobservables that are potentially correlated with
included message characteristics and generate an omitted variables problem. This concern is plausible, but
is second order in our view to the extent that we have included a very rich set of message characteristics.
Our approach to this problem has been to convert unobservables into observables by collecting direct data
on a relatively exhaustive set of message-characteristics.

4
4.1

Results
First-Stage

The first-stage model, as specified in Equation 3, approximates EdgeRanks message assignment algorithm.
We run the model separately for each of the 14 age-gender bins used by Facebook. These correspond to two
gender and seven age bins. For a given bin, the model relates the number of users of demographic type d who
were shown message k by firm j at time t to the message type (zk ), days since message ( ), and tie between
(d)

the firm and the user. Table 5 presents the results. The intercepts (0 ) indicate that messages by companies
in our dataset are shown most often to Females ages 35-44, Females 45-54, and Males 25-34. The lowest
number of impressions are for the 65+ age group. In our model, tie between a user and a firm is proxied by
a fixed-effect for each firm-demographic pair. This implies 800 14 fixed effects corresponding to 800 firms
and 14 demographic bins. Due to space constraints, we do not present all the estimated coefficients. Table
5 presents the coefficients for two randomly chosen firms. The first is a new-born clothing brand and the
24

second is a protein bar brand. For ease of visualization, these fixed effects are shown graphically in Figure 10
(only the statistically significant coefficients are plotted). For messages by the the new-born clothing brand,
the most impressions are among from females in the age-groups of 25-34, 18-24, and 35-44. Among males,
ages 25-34 receive the most number of impressions. For messages by the protein bar brand, impressions are
more evenly distributed across the different demographic bins, with the Male 18-24 group receiving the most
impressions. These estimated coefficients are consistent with our expectations for the two brands.
Female
Intercept
Page 1 fixed effect - new

F 13-17
5.528***
-0.210

F 18-24
6.071***
2.458***

F 25-34
6.446***
2.685***

F 35-44
7.165***
1.544**

F 45-54
7.209***
0.888

F 55-64
6.133***
0.813

F 65+
4.887***
0.489

born clothing brand

Page 2 fixed effect -

-0.573***

1.285***

1.466***

0.928***

0.016

1.671***

1.518***

protein bar brand

Message Type - App is the base
Link
0.010
Photo
0.253***
Status Update
0.100***
Video
0.033

0.045***
0.318***
0.161***
0.041

0.063***
0.340***
0.175***
0.061**

0.042***
0.309***
0.152***
0.041

0.051***
0.297***
0.152***
0.021

0.051***
0.267***
0.129***
0.024

0.048***
0.249***
0.114***
0.030

(d)

Njt (Fan Number)

2.0

1.8

7.2

1.9

1.9

3.8

8.5

106 ***

106 ***

106 ***

105 ***

105 ***

105 ***

105 ***

s(Njt ) significance
R-Squared

***
0.78

***
0.78

***
0.77

***
0.78

***
0.78

***
0.78

***
0.77

Intercept
Page 1 fixed effect - new

M 13-17
5.486***
0.156

M 18-24
6.118***
0.932

M 25-34
7.075***
1.673**

M 35-44
6.635***
1.082

M 45-54
6.125***
0.722

M 55-64
5.151***
0.209

M 65+
4.011***
0.111

born clothing brand

Page 2 fixed effect -

1.867***

2.423***

0.907***

0.670***

1.158***

1.575***

1.502***

protein bar brand

Message Type - App is the base
Link
-0.005
Photo
0.226***
Status Update
0.077***
Video
0.014

0.025***
0.284***
0.124***
0.039

0.033***
0.295***
0.126***
0.044*

0.034***
0.277***
0.120***
0.031

0038***
0.254***
0.106***
0.016

0.049***
0.230***
0.103***
0.007

0.030***
0.212***
0.084***
0.023

(d)

Male

(d)

Njt (Fan Number)

3.6

1.0

6.7

2.5

3.8

5.2

2.3

106 ***

106 ***

106 ***

105 ***

105 ***

105 ***

104 ***

(d)

s(Njt ) significance
***
***
***
***
***
***
R-Squared
0.79
0.80
0.79
0.78
0.78
0.77
*App is the base for message type. Significance Level: *** <0.001 ** < 0.01 * 0.05

***
0.76

Table 5:

EdgeRank Model Estimates: This table presents the coefficients obtained from 14 generalized additive models
for EdgeRank, calculated for each demographic bin. There are 14 demographic (gender-age) bins provided by Facebook. F13-17
means all females in the age between 13 and 17. Time since message ( ), and page-level fixed effects are not included in the

table and presented graphically separately.

The estimates for message type are roughly the same in all demographic bins. For all demographics,
the photo type has the highest coefficient (around 0.25) suggesting that photos are preferred to all other
media types by EdgeRank. This is likely because users have historically engaged better with photos causing
Facebook to show photos more often. The next most preferred message type is the status update with
coefficients averaging around 0.12 followed by videos and links. The baseline message type, apps, is the
message type that is least preferred by EdgeRank. The rank ordering of coefficients for message type do

25

not strictly follow the rank ordering of number of messages released by firms, which is shown in Table
1. Whereas links are posted more often, photos get more impressions relative to messages of other types,
clearly highlighting the role of EdgeRank. Days since message ( ) are not presented in Table 5 due to
space constraints. However, Figure 11 presents a box plot of the coefficients for across all 14 demographic
bins. All coefficients are negative and significant and also more negative for higher values of , implying
that EdgeRank prefers to show more recent messages. Finally, the coefficients for number of fans,

(d)

Njt ,

are positive and significant but they have relatively low magnitude. This is because our model includes a
smoothed term of the number of fans,

s(Njt ),
(d)

which soaks up both the magnitude and nonlinearity. The

smoothed fan-numbers are all significant.

The generalized additive model of EdgeRank recovers coefficients that make intuitive sense and are
consistent with claims made in several industry reports (e.g. that photos have the highest EdgeRank weight).
Further, the model fit appears to be good especially given that we have used generalized cross-validation to
guard against overfitting.

New Born Clothing Brand

Protein Bar Brand

Figure 10:

male 65+

male 5564

male 4554

male 3544

male 2534

male 1824

male 1317

female 65+

female 5564

female 4554

female 3544

female 2534

female 1824

female 1317

Pagelevel fixedeffect from GAM

Pagelevel fixedeffect estimates from GAM

across 14 demographic bins

Page-level Fixed effect Estimates from Generalized Additive Model Across 14 Demographic Bins:

This bar graph shows two randomly chosen page-level fixed effect estimates from the EdgeRank models. Only the statistically
significant estimates are shown. New born clothing brands are positively significant for 18-24 female, 25-34 female, 35-44 female,
and 25-34 male. Protein bar brands have the highest fixed effect among 18-24 male demographics.

Coefficient of Taus from Edgerank Model

Tau (time since post release) Coefficients from Edgerank Model (GAM)
2

10

11

12

13

14

15

16

Tau

Figure 11:

Time Since message Release ( ) Coefficients Box plot Across Demographics: This box plot shows the

coefficients on across all the demographics bin. = 1 is the base case and every coefficients are significant at the highest level
of p < 0.001.

26

Variable

Comment

Like

Constant

-6.913***(0.002)

-4.671***(0.001)

Persuasive

0.053***(0.001)

0.061***(0.000)

Informative

-0.143***(0.001)

-0.068***(0.000)

Persuasive Informative

0.012***(0.000)

0.003***(0.000)
0.009

McFadden R-sq.

0.015

Nagelkerke R-sq.

0.015

0.009

Log-likelihood

-4208220.431

-33678695.014

Deviance

8012471.987

66409947.187

AIC

8416448.861

67357398.028

665916

665916

Significance *** 0.001 ** 0.01 * 0.05 . 0.1

Table 6:

Persuasive vs Informative: Logistic regression for {Comment, Like} with composite summary variables for

persuasive and informative content.

4.2

Second-Stage

In the second-stage, we measure the effect of content characteristics on engagement using our selectivitycorrected model from the first-stage. All results in this section are based on an analysis of the entire
set of over 100,000 messages (i.e. the 5,000 AMT-tagged messages as well as the messages tagged using
NLP). The results for only the 5,000 AMT-tagged messages are qualitatively similar and are presented in
the appendix. To present the results in a simple way, we first create two composite summary variables
corresponding to persuasive content and informative content. Persuasive (informative) composite variables
are created by adding up the content variables categorized as persuasive (informative) in Table 1. To
be clear, the persuasive variable is obtained by adding values of REMFACT, EMOTION, EMOTICON,
HOLIDAYMENTION, HUMOR, PHILANTHROPIC, FRIENDLIKELY, and SMALLTALK resulting in a
composite variable ranging from 0 to 8. The informative composite variable is obtained by adding values
of BRANDMENTION, DEAL, PRICECOMPARE, PRICE, TARGET, PRODAVAIL, PRODLOCATION,
and PRODMENTION resulting in a composite variable ranging from 0 to 8. Table 6 shows the result of
logistic regression on engagement with these composite variables and interaction of those two variables as
the x-s.
We find that inclusion of more persuasive content has a positive and statistically significant effect on
both types of engagement; further, inclusion of more informative content reduces engagement. Interestingly,
the interaction between persuasive and informative content is positive, implying that informative content
increases engagement in the presence of persuasive content in the message. This results suggests broad
guidelines for marketers: persuasive content in isolation is preferred to purely informative ones. Further,
mixing persuasive and informative content should be made a basis of content engineering for improving
engagement with consumers on this medium.
Table 7 presents the results of aggregate logistic regression with the full list of content variables. We
present results for both engagement metrics (Likes/comments) as well as for models with and without the
27

EdgeRank correction. We exclude the 16 estimated coefficients from the table since they are all negative
and statistically significant just as in the EdgeRank model in Figure 11. We also exclude demographic fixed
effects for space. Scanning through the results, we observe that the estimates are directionally similar, in
most cases, with and without EdgeRank correction. However, the magnitudes often change. For example,
consider the coefficients for message type Photo. In the model without EdgeRank correction, Photos are very
likely to get comments (coefficient = 0.867) and Likes (coefficient = 1.011). After EdgeRank correction, the
results are similar but the magnitude of the effect drops. This makes sense because we know that EdgeRank
prefers Photos. Similarly, Status Updates continue to be more likely (than apps) to get comments and Likes
but the effect size is smaller after EdgeRank correction. In some instances, there are directional changes for
some coefficients. For example, the result that links are more likely to get Likes/comments relative to apps
changes sign after EdgeRank correction. This highlights the importance of EdgeRank correction, an issue
that most industry reports (e.g., Wildfire 2012) often overlook. For example, most industry reports ordering
of engaging media type often list status update to be more engaging than videos. While we find this to be
true before EdgeRank correction for Likes, we find that this is reversed after the EdgeRank correction.
We find that high reading complexity (SMOG) decreases both Likes and comments whereas shorter
messages (MSGLEN) are Liked and commented on more, albeit with a small effect size. Having links
(HTTP) is worse for engagement whereas asking questions (QUESTION) significantly increase comments
but at the cost of Likes. Using blanks in the message to encourage comments has a similar effect of increasing
comments but hurting Likes. Interestingly, while the odds ratio of comments increases by 69% if a message
asks a question, it increases by 200% if blanks are included suggesting that blanks are more effective than
questions if the goal is to increase comments. Asking for Likes increase both Likes and comments, whereas
asking for comments increase comments but at the cost of Likes. It is clear that even these simple content
variables impact user engagement.
The next 16 variables in the table are the persuasive and informative content variables. Figure 12 charts
the coefficients for these variables in a bar graph and demonstrates the sharp difference between persuasive
and informative content types. Looking at comments, a striking pattern is that most informative contents
have a negative impact whereas persuasive contents have a positive impact. The informative content variables
with the most negative impact are PRICE, DEAL, and PRODMENTION. The persuasive content variables
with the most positive impact are EMOTION and PHILANTHROPIC. Interestingly, HOLIDAYMENTION
discourages comments.14 One possible explanation is that near holidays, all Facebook pages indiscriminately
mention holidays, leading to dulled responses. For example, during Easter, the occurrence of holiday mention
jumped to nearly 40% across all messages released that day compared to the average occurrence of about 1%.
Looking at Likes, fewer persuasive content variables have positive impact but the results are qualitatively
similar to that for comments. Among persuasive contents, EMOTION has the most positive impact on Likes
whereas EMOTICON has the most negative impact. Most informative content variables continue to have a
14 We

checked for correlation with other contents to investigate this matter but no correlation was over 0.02.

28

negative impact (i.e., reduce engagement), with PRICE and DEAL having the most negative impact. The
results also highlight that there exist some differences between impact on Likes versus Comments.
Figure 13 shows the results on content effects by industry. Only the statistically significant results are
graphed and all results are EdgeRank-corrected. The coefficients are very different across industries both
in magnitude and, for some variables, in direction. For example, emotional and philanthropic content has
the most positive impact on Facebook pages of type Organizations which include non-profits, educational
organizations and religious groups. Further, while mentioning holidays has a negative impact on engagement
for most industry types, it has a positive impact on engagement for Organizations. Similarly, looking at
informative content, we observe that variables such as Price, Product Availability, and Product Mentions
generally have a negative impact on engagement for most industry types, but have a positive impact for
industry type Celebrity. Users seem more forgiving of celebrity pages endorsing products and sharing price
information.
Comparing Figures 3 and 13 also provides interesting comparisons of what each industry is currently
posting and what users engage with. For example, pages of types Places and Businesses, Entertainment, and
Consumer Products do not post emotional content much though Figure 13 shows that emotional content
induce higher Likes and Comments. Similarly, while Places and Business pages tend to post more of deal
content, only Consumer Product pages seem to be benefiting from the deal content (in terms of obtaining
more comments). Places and Businesses pages also post larger percent of product availability content while
only the Consumer Product and Celebrity pages benefit from inclusion of such content.
Robustness

We run a variety of alternative specifications to assess the robustness of our results. Estimates

of these alternative specifications are presented in the Appendix. First, we replicate the results using only
the set of 5,000 messages directly coded up by the Amazon Mechanical Turkers. Second, we assess the extent
to which the parameters are stable when we drop subsets of attributes. Third, we include additional checks
as added robustness against selection. Our added checks use residuals from the first-stage as a control
(d)

function in the second-stage. To see this, note the residuals in Equation 3, $kjt , represent unobserved
reasons that users in demographic bucket d would be more likely to be targeted a message k by EdgeRank.
As robustness, we ask whether our results on the effect of message attributes change when we control for
these unobservable drivers of attractiveness of each bucket for that message. To do this, note that from our
(d)

first-stage, we can obtain an estimate of the residual, denoted $kjt . We re-run our second stage estimation
(d)

including the estimated

$kjt -s as covariates in Mkt in Equation 5. We can interpret the revised results as the
effect of message characteristics on engagement after controlling for the unobserved attractiveness of each
bucket for that message. Results from these alternative models show that the main qualitative features of
our results are robust across these specifications.

29

NO ER COMMENT
Constant

12.309*** (0.197)

OR

ER COMMENT

OR

14.083*** (0.142)

NO ER LIKE

OR

-7.383*** (0.089)

ER LIKE

OR

13.504*** (0.065)

SMOG

-0.045*** (0.000)

0.956

-0.066*** (0.000)

0.936

-0.029*** (0.000)

0.971

-0.057*** (0.000)

0.945

MSGLEN

0.000 (0.000)

1.000

-0.000*** (0.000)

1.000

-0.000*** (0.000)

1.000

-0.000*** (0.000)

1.000

HTTP

-0.484*** (0.002)

0.616

-0.324*** (0.002)

0.723

-0.353*** (0.000)

0.703

-0.180*** (0.000)

0.835

QUESTION

0.449*** (0.001)

1.567

0.527*** (0.001)

1.694

-0.292*** (0.000)

0.747

-0.185*** (0.000)

0.831

BLANK

0.942*** (0.003)

2.565

1.099*** (0.003)

3.001

-0.716*** (0.002)

0.489

-0.625*** (0.002)

0.535

ASKLIKE

0.002 (0.010)

1.002

0.178*** (0.010)

1.195

0.456*** (0.003)

1.578

0.501*** (0.003)

1.650

ASKCOMMENT

0.779*** (0.021)

2.179

0.710*** (0.021)

2.034

-0.090*** (0.011)

0.914

-0.282*** (0.011)

0.754

REMFACT

-0.019*** (0.002)

0.981

0.010*** (0.002)

1.010

-0.060*** (0.001)

0.942

-0.035*** (0.001)

0.966

EMOTION

0.203*** (0.002)

1.225

0.257*** (0.002)

1.293

0.201*** (0.001)

1.223

0.257*** (0.001)

1.293

EMOTICON

0.118*** (0.004)

1.125

-0.053*** (0.004)

0.948

-0.132*** (0.001)

0.876

-0.214*** (0.001)

0.807

HOLIDAYMENTION

-0.493*** (0.014)

0.611

-0.352*** (0.014)

0.703

-0.323*** (0.004)

0.724

-0.136*** (0.004)

0.873

HUMOR

0.023*** (0.002)

1.023

0.082*** (0.002)

1.085

-0.044*** (0.000)

0.957

0.012*** (0.000)

1.012

PHILANTHROPIC

0.147*** (0.002)

1.158

0.140*** (0.002)

1.150

0.008*** (0.001)

1.008

0.001 (0.001)

1.001

FRIENDLIKELY

0.002 (0.002)

1.002

-0.022*** (0.002)

0.978

0.073*** (0.001)

1.076

0.058*** (0.001)

1.060

SMALLTALK

0.045*** (0.002)

1.046

-0.074*** (0.002)

0.929

-0.052*** (0.001)

0.949

-0.121*** (0.001)

0.886

BRANDMENTION

0.000 (0.002)

1.000

0.077*** (0.002)

1.080

-0.031*** (0.000)

0.969

0.004*** (0.000)

1.004

DEAL

-0.163*** (0.002)

0.850

-0.168*** (0.002)

0.845

-0.198*** (0.001)

0.820

-0.200*** (0.001)

0.819

PRICECOMPARE

-0.031*** (0.001)

0.969

0.005*** (0.001)

1.005

-0.037*** (0.000)

0.964

-0.040*** (0.000)

0.961

PRICE

-0.104*** (0.005)

0.901

-0.319*** (0.005)

0.727

-0.187*** (0.001)

0.829

-0.400*** (0.001)

0.670

TARGET

-0.016*** (0.002)

0.984

-0.073*** (0.002)

0.930

0.030*** (0.001)

1.030

-0.041*** (0.001)

0.960
0.937

Persuasive

Informative

PRODAVAIL

-0.067*** (0.002)

0.935

-0.060*** (0.002)

0.942

-0.109*** (0.001)

0.897

-0.065*** (0.001)

PRODLOCATION

-0.054*** (0.002)

0.947

0.009*** (0.002)

1.009

0.063*** (0.001)

1.065

0.107*** (0.001)

1.113

PRODMENTION

-0.050*** (0.002)

0.951

-0.148*** (0.002)

0.862

0.077*** (0.001)

1.080

-0.007*** (0.001)

0.993

Message Type - App is the base

-Link

0.177*** (0.003)

1.194

-0.238*** (0.003)

0.788

0.126*** (0.001)

1.134

-0.374*** (0.001)

0.688

-Photo

0.867*** (0.003)

2.380

0.519*** (0.003)

1.680

1.011*** (0.001)

2.748

0.651*** (0.001)

1.917

-Status Update

1.146*** (0.003)

3.146

0.818*** (0.003)

2.266

0.478*** (0.001)

1.613

0.060*** (0.001)

1.062

-Video

-0.106*** (0.009)

0.899

0.466*** (0.009)

1.594

-0.200*** (0.003)

0.819

0.341*** (0.003)

1.406

Industry Category - Celebrity is the base

-ConsumerProduct

0.171*** (0.002)

1.186

-0.319*** (0.002)

0.727

-0.369*** (0.001)

0.691

-0.813*** (0.001)

0.444

-Entertainment

0.362*** (0.002)

1.436

0.418*** (0.002)

1.519

-0.291*** (0.001)

0.748

-0.256*** (0.001)

0.774

-Organization

0.485*** (0.002)

1.624

0.281*** (0.002)

1.324

-0.004*** (0.001)

0.996

-0.212*** (0.001)

0.809

-PlaceBusiness

0.429*** (0.005)

1.536

0.021*** (0.005)

1.021

-0.639*** (0.002)

0.528

-1.109*** (0.002)

0.330

-Websites

0.012*** (0.003)

1.012

0.074*** (0.003)

1.077

0.088*** (0.001)

1.092

0.118*** (0.001)

1.125

McFadden R-sq.

0.271

0.207

0.32

Nagelkerke R-sq.

0.271

0.207

0.321

0.239
0.241

Log-likelihood

-2,446,467.133

-3,423,466.377

-14,108,100.91

-25,950,910.53

Deviance

4,488,295.547

6,443,162.73

27,268,539.27

50,955,992.81

AIC

4,893,058.266

6,847,056.753

28,216,325.82

51,901,945.06

665,916

665,916

665,916

665,916

Significance *** 0.001 ** 0.01 * 0.05 . 0.1

Table 7:

Aggregate Logistic Regression Results For Comments and Likes: This table presents the aggregate logistic

regression on comments and Likes for both EdgeRank-corrected (ER) and uncorrected (NO ER) for all data. OR means Odds
ratio and shows the odds ratio for the estimates left of the column.

30

Logistic Regression Coefficients of Message Contents for Comments

Logistic Regression Coefficients of Message Contents for Likes

0.2
Logistic Regression Coefficients

Logistic Regression Coefficients

0.2

0.0

0.2

edgerank
0.4

0.0

0.2
edgerank

no edgerank

no edgerank

prodloc

prodmention

prodavail

price

target

deal

pricecompare

smalltalk

brandmention

philan

friendlikely

humor

holiday

emoticon

remfact

emotion

prodloc

prodmention

prodavail

price

target

deal

pricecompare

smalltalk

brandmention

philan

friendlikely

humor

holiday

emoticon

remfact

emotion

0.4

Figure 12:

message Characteristic Coefficients for Comments and Likes: These bar graphs show the coefficients of
logistic regression for both EdgeRank corrected and uncorrected models. Only the significant coefficients are plotted.

Logistic Regression Result across Industry (Comment)

Logistic Regression Result across Industry (Like)

0
1

organization

1
0

placesbusiness

1
0
1

entertainment

entertainment

0
1

0
1

organization

0
1

placesbusiness

Logitstic Regression Coef

consumerproduct

1
consumerproduct

1
0
1

1
website

1
0

website

Logitstic Regression Coef

celebrity

1
celebrity

SMOG
msglen
http
question
blank
asklike
askcomment
remfact
emotion
emoticon
holiday
humor
philan
friendlikely
smalltalk
brandmention
deal
pricecompare
price
target
prodavail
prodloc
prodmention
Type:link
Type:photo
Type:StatusUpdate
Type:video

SMOG
msglen
http
question
blank
asklike
askcomment
remfact
emotion
emoticon
holiday
humor
philan
friendlikely
smalltalk
brandmention
deal
pricecompare
price
target
prodavail
prodloc
prodmention
Type:link
Type:photo
Type:StatusUpdate
Type:video

Variables

Variables

Figure 13:

Logistic Regression by Industry (Comments and Likes): This bar graphs show the coefficients of logistic

regression for EdgeRank-corrected model. Only the significant coefficients (p<0.05) are graphed. In the Like plot on the right,
the coefficient for ASKCOMMENT for websites is -4.8 but zoomed in to optimize the clarity of the graph.

4.3

Out-of-Sample Prediction & Managerial Implications

Finally, we informally assess the extent to which the models we develop may be used as an aid to content
engineering, and to predict the expected levels of engagement for various content profiles a firm may consider
for a potential message it could serve to users. We present an illustration set of out-of-sample prediction of
engagement with real messages. Our intent is to give the reader a rough sense for the use of our estimates
31

as a tool to assess expected engagement for hypothetical content bundles.

Suppose a firms content marketing team has developed multiple alternative messages. The marketing
team may be interested in choosing a message keeping in mind a specific engagement goal. To illustrate
the above, we choose three messages released around the same time outside our sample. To emphasize that
our second-stage model of engagement has predictive power, we choose these to be for the same firm, of the
same message type and having roughly the same number of impressions (i.e., we are taking out the effect
of EdgeRank ). Table 8 shows the messages, the actual lifetime engagement realized for those messages, and
the content tags we generated for those messages. For each message, we use the coefficients from our second
stage to predict the expected engagement. In the Content Coef: column, we present the latent linear
index of the logistic function obtained by multiplying the coefficients for the engagement model (Table 7,
EdgeRank corrected) with indicator variables for whether each type of content attribute is present in these
messages, and then adding these up. In the last two columns we present the predicted and actual ranks for
the three messages in terms of their engagement. We see that the match is very good and that the model
can be useful to select the message that maximizes engagement.
Now imagine that a firm starts with the second message as the base creative. Content engineering for this
message using the model is straightforward. For instance, if the marketer is assessing the potential impact of
adding a philanthropic aspect and asking user to like the post to message two, we can determine that it will
increase the latent linear index for comments from 0.731 to 0.731 + 0.140 + 0.178 = 1.049, which increases
the predicted comments rank of this message to 1, and increases the predicted odds ratio for comments by
37%. Similarly, if the marketer is considering asking for comments explicitly, this will increase the number
of comments for the message obtained by increasing the latent linear index from 0.731 to 0.731 + 0.710 =
1.441. In this sense, the model is able to aid the assessment of the anticipated engagement from various
possible content bundles in a straightforward way.
Sample Messages
{Actual Comments, Actual Likes}

Content Tags Content

Comments
Coef
Rank
{Com,Likes}

Likes
Rank

Dont forget in celebration of hitting over

HTTP, DEAL, {-0.596,-0.552} Actual:3

PRODLOCAPredicted:3
TION,
PRODAVAIL,

Actual:2

70,000 we are giving all our awesome fans

{exclusively} the "employee discount" take
20% off your entire order on our website
{http://anonymized.com} with the code:
SOMECODE and it is good until 3/16/12.
some shopping on us :)

Predicted:2

EMOTICON

Enjoy

{12, 83}

Who is ready for a givvveeeawayyyyy?!

(35 mins from now!)

:)

:)

{132, 438}
COMPLETE THIS SENTENCE: Crafting is best with
________.

EMOTION,

{0.731,-0.142}

EMOTICON,
QUESTION
BLANK,
SMALLTALK

{1.025,-0.746}

Actual:2

Actual:1

Predicted:2

Predicted:1

Actual:1

Actual:3

Predicted:1

Predicted:3

{416, 72}

Table 8:

Predicted versus Actual Engagement Ranking for Three Illustrative messages: Note: we anonymized

some parts of the messages for presentation.

32

Conclusions and Implications

We show through a large-scale study that content engineering in social media has a significant impact on user
engagement as measured by Likes and comments for messages. Our analysis shows that persuasive content,
such as emotional and philanthropic content, has a positive impact on engagement. This suggests that firms
gain from sharing their brand personality and information about their social initiatives on social media.
Further, we find that product informative content has a negative impact on user engagement. This presents
a challenge to marketers who seek to build a large following on social media and who seek to leverage that
following to disseminate information about new products and promotions. One takeaway from our study
is that these strategies work when product informative content is combined with persuasive content. In
addition, our results are moderated by the industry type suggesting that there is no one-size-fits-all content
strategy and that firms need to test multiple content strategies.
Because of the scale of our study (nearly 800 firms and more than 100,000 messages analyzed), we believe
our results generalize and have broad applicability. Nonetheless, it is important to recognize several limitation
of our study. First, we note that the results from any study on consumer response to content depend on
the mix of content used in the study. For example, we find that messages mentioning holidays, especially
by consumer product companies, have a negative effect on engagement. This may be due to excessive use of
holiday messages by firms. It is possible that the effect may be positive if firms use these kinds of messages in
moderation. Similarly, we find that emotional messages have a positive impact on engagement. Here again,
it is possible this effect may reduce in the future if firms start using emotional content excessively. Hence, it
is important to interpret our results in the context of the content mix used by firms and redo the analysis
in the event of large-scale changes in the content mix used by firms.
We used two metrics for user engagement, namely Likes and comments on messages. Other measures
worth considering, include whether users share messages with friends, visit the websites of firms posting
messages, or buy more products from these firms. Our use of Likes and comments is motivated both by
the widespread use of these metrics as marketing goal in social media settings, and also the availability
of data. Future studies that evaluate other measures of interest can add value, particularly in validating
the generalizability of our findings and in exploring mechanisms underpinning the effects we describe. As
noted in the introduction, we do not address the question of how engagement affects product demand and
firms profits so as to complete the link between ad-attributes and those outcome measures. Such data are
still not widely available at the scale needed for this study. Although it is not the focus of our study, it
is worth highlighting that several extant studies have studied the link between Facebook engagements and
sales (albeit at a smaller scale). For example, based on randomized studies, comScore (2012) reports a 38%
lift in purchase for fans exposed to Starbucks advertising on Facebook through Facebook Pages or Facebook
paid advertising. Similarly, studies such as Kumar et al. (2013); Goh et al. (2013); Rishika et al. (2013);
Li and Wu (2014); Miller and Tucker (2013) show that social media can be used to generate growth in
sales, and ROI, consumer participation, retention, and profitability, connecting social media metrics such as
33

comments to financial metrics.

The competition for consumer attention across media outlets is intense, especially on social media platforms. Consumers, in turn, are overwhelmed by the proliferation of online content, and it seems clear that
marketers will not succeed without engineering this content for their audience. We hope this study contributes to improve content engineering by firms on social media sites and, more generally, creates interest
in evaluating the effects of content on consumer engagement.

References
90octane: 2012, New Facebook Content and ConversionStrategies Generate Engagement and Leads for
Gates Automotive. Technical report, 90octane.
Abernethy, A. M. and G. R. Franke: 1996, The Information Content of Advertising: A Meta-Analysis.
Journal of Advertising 25(2), 117.
Aggarwal, R., R. Gopal, R. Sankaranarayanan, and P. V. Singh: 2012, Blog, blogger, and the firm: Can
negative employee posts lead to positive outcomes?. Information Systems Research 23(2), 306322.
Anand, B. and R. Shachar: 2009, Targeted advertising as a signal. Quantitative Marketing and Economics
7(3), 237266.
Anand, B. and R. Shachar: 2011, Advertising the Matchmaker. Rand Journal of Economics 42(2), 205245.
Anderson, S. and R. Renault: 2006, Advertising Content. American Economic Review March, 93113.
Armstrong, S.: 2010, Persuasive Advertising. Palgrave Macmillan.
Ascend2: 2013, Marketing Strategy Report: Social Media. Technical report, Ascend2 and Research Underwriters.
Bagwell, K.: 2007, Chapter 28: The Economic Analysis of Advertising. Handbook of IO 3, 17011844.
Bennett, P. N.: 2006, Building Reliable Metaclassifiers for Text Learning. Ph.D. thesis, CMU.
Berger, J.: 2011, Arousal Increases Social Transmission of Information. Psychological Science 22(7), 89193.
Berger, J.: 2012, Word-of-Mouth: A Reveiew and Directions for Future Research. Work in Progress.
Berger, J. and K. L. Milkman: 2012, What Makes online Content Viral?. Journal of Marketing Research
49(2), 192205.
Berger, J. and E. Schwartz: 2011, What Drives Immediate and Ongoing Word-of-Mouth?. Journal of
Marketing Research 48(5), 869880.
Bertrand, M., D. Karlan, S. Mullianathan, E. Shafir, and J. Zinman: 2010, Whats Advertising Content
Worth? Evidence from a Consumer Credit Marketing Field Experiment. pp. 263306.
Bird, S., E. Klein, and E. Loper: 2009, Natural Language Processing with Python. OReilly Media.
Blundell, R. and J. Powell: 2003, Endogeneity in Nonparametric and Semiparametric Regression Models,
Vol. II of Econometric Monograph Series 36. Cambridge University Press.
Brieman, L.: 1996, Bagging predictors. Machine Learning 24, 123140.
Buhrmester, M., T. Kwang, and S. D. Gosling: 2011, Amazons Mechanical Turk: a New Source of Inexpensive, Yet High-Quality, Data?. Psychological Science 6(1).
Butters, G.: 1977, Equilibrium Distributions of Sales and Advertising Prices. Review of Economic Studies
44(3), 46591.

34

Chadwick-Martin-Bailey: 2010, Consumers Engaged Via Social Media Are More Likely To Buy, Recommend.
Chandy, R., G. Tellis, D. Macinnis, and P. Thaivanich: 2001, What to Say When: Advertising Appeals in
Evolving Markets. Journal of Marketing Research 38, 399414.
Cialdini, R.: 2001, Influence: Science and Practice. Needham Heights, MA: Allyn and Bacon.
comScore: 2012, The Power of Like 2. Technical report, comScore.
comScore: 2013, The Power of Like Europe. Technical report, comScore.
Creamer, M.: 2012, Study Only 1 percent of Facebook Fans Engage With Brands. AdAge.
Dellarocas, C.: 2006, Strategic manipulation of internet opinion forums: Implications for consumers and
firms. Management Science 52(10), 15771593.
Freund, Y. and R. E. Schapire: 1995, A decision-theoretic Generalization of on-line learning and an application to boosting. Computational Learning Theory: EUROCOLT pp. 2337.
Gardete, P. M.: 2013, Cheap-Talk Advertising and Misrepresentation in Vertically Differentiated Markets.
Marketing Science.
Gentzkow, M. and J. Shapiro: 2010, What Drives Media Slant? Evidence from U.S. Newspapers. Econometrica 78(1).
Gerber, S.: 2014, 9 Ways Social Media Marketing Will Change in 2014.
Geva, T. and J. Zahavi: 2013, Empirical evaluation of an automated intraday stock recommendation system
incorporating both market data and textual news. Decision Support Systems.
Ghose, A., P. G. Ipeirotis, and B. Li: 2012, Designing Ranking Systems for Hotels on Travel Search Engines
by Mining User-Generated and Crowd-Sourced Content. Marketing Science 31(3).
Goh, K.-Y., C.-S. Heng, and Z. Lin: 2013, Social media brand community and consumer behavior: Quantifying the relative impact of user-and marketer-generated content. Information Systems Research 24(1),
88107.
Grossman, G. and C. Shapiro: 1984, Informative Advertising with Differentiated Products. Review of
Economic Studies 51(1), 6381.
Hassan, M. H., B. S. Vikas, and H. Fiorletta: 2011, Accurate Information Extraction for Quantitaive
Financial Events. CIKM.
Hastie, T., R. Tibshirani, and J. Friedman: 2009, The Elements of Statistical Learning: Data Mining,
Inference, and Prediction. Springer.
Hastie, T. and R. J. Tibshirani: 1990, Generalized Additive Models. Chapman and Hall CRC.
HubSpot: 2013, State of Inbound Marketing. Technical report, HubSpot.
Ipeirotis, P. G., F. Provost, and J. Wang: 2010, Quality Mangagement on Amazon Mechanical Turk.
HCOMP.
Jabr, W., R. Mookerjee, Y. Tan, and V. S. Mookerjee: 2014, LEVERAGING PHILANTHROPIC BEHAVIOR FOR CUSTOMER SUPPORT: THE CASE OF USER SUPPORT FORUMS.. MIS Quarterly
38(1).
Jurafsky, D. and J. H. Martin: 2008, Speech and Language Processing. Prentice Hall.
Kihlstrom, R. and M. Riordan: 1984, Advertising as a Signal. Journal of Political Economy 92(3), 42750.
Kumar, V., V. Bhaskara, R. Mirchandani, and M. Shah: 2013, Creating a Measurable Social Media Marketing Strategy: Increasing the Value and ROI of Intangibles and Tangibles for Hokey Pokey. Marketing
Science 32(2), 194212.
35

Li, X. and L. Wu: 2014, Measuring Effects of Observational Learning and Social-Network Word-of-Mouth
(WOM) on the Sales of Daily-Deal Vouchers. SSRN Working Paper.
Liaukonyte, J., T. Teixeira, and K. Wilbur: 2013, TV Advertising Content and Online Shopping. (in
progress).
Mason, W. A. and S. Suri: 2012, Conducting behavioral research on Amazons Mechanical Turk. Behavior
Research Methods 44(1), 123.
Mayzlin, D. and J. Shin: 2011, Uninformative Advertising as an Invitation to Search. Marketing Science
30(4), 666685.
Milgrom, P. and J. Roberts: 1986, Price and Advertising Signals of Product Quality. Journal of Political
Economy 94(4), 796821.
Miller, A. R. and C. Tucker: 2013, Active social media management: the case of health care. Information
Systems Research 24(1), 5270.
Nan, X. and R. J. Faber: 2004, Advertising Theory: Reconceptualizing the Building Blocks. Marketing
Theory 4, 730.
Nelson, P.: 1974, Advertising as Information. Journal of Political Economy 82(4), 72954.
Netzer, O., R. Feldman, J. Goldenberg, and M. Fresko: 2012, Mine Your Own Business: Market-Structure
Surveillance Through Text Mining. Marketing Science 31(3), 521543.
Oestreicher-Singer, G. and L. Zalmanson: 2013, CONTENT OR COMMUNITY? A DIGITAL BUSINESS
STRATEGY FOR CONTENT PROVIDERS IN THE SOCIAL AGE.. MIS Quarterly 37(2).
Paolacci, G., J. Chandler, and P. G. Ipeirotis: 2010, Running Experiements on Amazon Mechcanical Turk.
Judgement and Decision Making 5(5).
Porter, L. and G. J. Golan: 2006, From Subservient Chickens to Brawny Men: a Comparison of Viral
Advertising to Television Advertising. Journal of Interactive Advertising 6(2).
Resnik, A. and B. L. Stern: 1977, An Analysis of Informative Content in Television Advertising. Journal
of Marketing 41(1).
Rishika, R., A. Kumar, R. Janakiraman, and R. Bezawada: 2013, The effect of customers social media participation on customer visit frequency and profitability: an empirical investigation. Information systems
research 24(1), 108127.
Rui, H., Y. Liu, and A. Whinston: 2013, Whose and what chatter matters? The effect of tweets on movie
sales. Decision Support Systems 55(4), 863870.
Sheng, V., F. Provost, and P. G. Ipeirotis: 2007, Get another label? Improving data quality and data
mining using multiple, noisy labelers. 14th ACM SIGKDD pp. 614622.
Singh, P. V., N. Sahoo, and T. Mukhopadhyay: 2014, Seeking variety: A dynamic model of employee blog
reading behavior. forthcoming in Information Systems Research.
Snow, R., B. OConnor, D. Jurafsky, and A. Y. Ng: 2008, Cheap and fast - but is it good? Evaluating
non-expert annotations for natural language tasks. Empirical Methods in Natural Language Processing.
Susarla, A., J.-H. Oh, and Y. Tan: 2012, Social networks and the diffusion of user-generated content:
Evidence from YouTube. Information Systems Research 23(1), 2341.
Tucker, C.: 2012a, Ad Virality and Ad Persuasiveness. MIT Working paper.
Tucker, C.: 2012b, Social Advertising. Working paper, MIT.
Vakratsas, D. and T. Ambler: 1999, How Advertising Works: What do we really know?. Journal of
Marketing 63(1).
36

Wildfire: 2012, Edgerank and Your Brand: An Analysis of Messaging Engagement on Facebook by Business
Category. Technical report, Wildfire.
Wood, S. N.: 2006, Generalized Additive Models: An Introduction with R. CRC press.
WSJ: 2012, GM Says Facebook Ads Dont Pay Off.
Xu, L. and A. Krzyzak: 1992, Methods of Combining Multiple Classifiers and Their Applications to Handwriting Recognition. IEEE Transactions on systems.
Zadrozny, B. and C. Elkan: 2002, Transforming Classifier Scores into Accurate Multiclass Probability
Estimates. KDD.

37

Appendix 1: Survey Instrument

Figure 14: Survey Form Used in Amazon Mechanical Turk

38

Appendix 2: Amazon Mechanical Turk - Robust Content Extraction

Following best-practices in the literature, we employ the following strategies to improve the quality of
classification by the Turkers in our study.
1. For each message, at least 9 different Turkers inputs are recorded. We obtain the final classification
by a majority-voting rule.
2. We restrict the quality of Turkers included in our study to comprise only those with at least 100
reported completed tasks and 97% or better reported task-approval rates.
3. We use only Turkers from the US so as to filter out those potentially not proficient in English, and
to closely match the user-base from our data (recall, our data has been filtered to only include pages
located in the US).
4. We refined our survey instrument through an iterative series of about 10 pilot studies, in which we
asked Turkers to identify confusing or unclear questions. In each iteration, we asked 10-30 Turkers
to identify confusing questions and the reasons they found those questions confusing. We refined the
survey in this manner till almost all queried Turkers stated no questions were confusing.
5. To filter out participants who were not paying attention, we included an easily verifiable test question
does the message have a dollar sign ($)?. Responses from Turkers that failed the verification test are
dropped from the data.
6. In order to incentivize workers, we awarded additional bonuses of $2-$5 to the top 20 workers with
exceptional accuracy and throughput.
7. On average, we found that message tagging took a little over 3 minutes and it typically took at least
20 seconds or more to completely read the tagging questions. We defined less than 30 seconds to be
too short, and discarded any message tags with completion times shorter than that duration to filter
out inattentive Turkers and automated programs (bots).
8. Once a Turker tags more than 20 messages, a couple of tagged samples are randomly picked and
manually examined for quality and performance. This process identified about 20 high-volume Turkers
who completed all surveys in less than 10 seconds and tagged several thousands of messages (there
were also Turkers who took time to complete the surveys but chose seemingly random answers). We
concluded these were automated programs. These results were dropped, and the Turkers hard blocked
from the survey, via the blocking option provided in AMT.
We believe our methodology for content-classification has strong external validity. The binary classification
task that we serve to the AMT Turkers in our study is relatively simpler than the more complex tasks for
which AMT-based data have been employed successfully in the literature. The existing AMT literature
39

has documented evidence that several of the strategies implemented above improves the quality of the data
generated (Mason and Suri (2012); Ipeirotis et al. (2010); Paolacci et al. (2010)). Snow et al. (2008) show
that combining results from a few Turkers can produce data equivalent in quality to that of expert labelers
for a variety of text tagging tasks. Similarly, Sheng et al. (2007) document that repeated labeling of the type
we implement wherein each message is tagged by multiple Turkers, is preferable to single labeling in which
one person tags one sentence. Finally, evaluating AMT based studies, Buhrmester et al. (2011) concludes
that (1) Turkers are demographically more diverse than regular psychometric studies samples, and (2) the
data obtained are at least as reliable as those obtained via traditional methods as measured by psychometric
standards such as Cronbachs Alpha, a commonly used inter-rater reliability measure.

40

Appendix 3: NLP Algorithm

This section provides detailed outline of the algorithm used in the paper. Figure 15 shows the process
visually.
Training The Algorithm
1. The raw textual data of 5, 000 messages in the training sample are broken down into basic building
blocks of sentences using stop-words removal (removing punctuation and words with low information
such as the definite article the), tokenization (the process of breaking a sentence into words, phrases,
and symbols or tokens), stemming (the process of reducing inflected words to their root form, e.g.,
playing to play), and part-of-speech tagging (determining part-of-speech such as nouns). For reference see Jurafsky and Martin (2008). In this process, the input to the algorithm is a regular sentence
and the output is an ordered set of fundamental linguistic entities with semantic values. We use a
highly regarded python NLP framework named NLTK (Bird et al., 2009) to implement this step.
2. Once the messages are broken down as above, an algorithm extracts sentence-level attributes and
sentence-structure rules that help identify the included content. Some examples of sentence-level
attributes and rules include: frequent noun words (bag-of-words approach), bigrams, the ratio of partof-speech used, tf-idf (term-frequency and inverse document frequency) weighted informative word
weights, and whether a specific key-word is present rule. For completeness, we describe each of
these in Table 9. The key to designing a successful NLP algorithm is to figure out what we (humans)
do when identifying certain information. For example, what do we notice about the sentences we
have identified as having emotional content? We may notice the use of certain types of words, use
of exclamation marks, the use of capital letters, etc. At the end of this step, the dataset consists
of sentence-level attributes generated as above (the x -variables), corresponding to a series of binary
(content present/not-present) content labels generated from AMT (the y-variables).
3. For each binary content label, we then train a classification model by combining statistical and rulebased classifiers. In this step, the NLP algorithm fits the binary content label (the y-variable) using
the sentence-level attributes as the x -variables. For example, the algorithm would fit whether or not
a message has emotional content as tagged by AMT using the sentence attributes extracted from the
message via step 2. We use a variety of different classifiers in this step including logistic regression with
L1 regularization (which penalizes the number of attributes and is commonly used for attribute selection
for problems with many attributes; see (Hastie et al., 2009)), Naive Bayes (a probabilistic classifier
that applies Bayes theorem based on presence or absence of features), and support vector machines
(a gold-standard algorithm in machine learning that works well for high dimensional problems) with
different flavors of regularization and kernels 15 . To account for imbalance in positive and negative class
15 We

tried support vector machines with L1 and L2 regularization and various kernels including linear, radial basis function,

41

labels in some content, we utilized combination of class-weighted classifiers and resampling methods.
4. To train the ultimate predictive classifier, we use ensemble methods to combine results from the multiple
statistical classifiers we fit in step 3. The motivation for ensemble learning is that different classifiers
perform differently based on underlying characteristics of data or have varying precision or recall in
different locations of the feature vector space. Thus, combining them will achieve better classification
output either by reducing variance (e.g. Bagging (Brieman, 1996)) or reducing bias (e.g. Boosting
(Freund and Schapire, 1995)). Please see Xu and Krzyzak (1992); Bennett (2006) for further reading on
ensemble methods. This step involves combining the prediction from individual classifiers by weightedmajority voting, unweighted-majority voting, or a more elaborate method called isotonic regression
(Zadrozny and Elkan, 2002) and choosing the best performing method in terms of accuracy, precision
and recall for each content profiles. In our case, we found that support vector machine based classifiers
delivered high precision and low recall, while Naive Bayes based classifiers delivered high recall but
low precision. By combining these, we were able to develop an improved classifier that delivers higher
precision and recall and in effect, higher accuracy. Table 10 shows the improvement of the final
ensemble learning method relative to using only one support vector machine. As shown, the gains from
combining classifiers are substantial. We obtain similar results for negative class labels.
5. Finally, we assess the performance of the overall NLP algorithm on three measures, viz., accuracy,
precision, and recall (as defined in Footnote 4) using the 10-fold cross-validation method. Under this
strategy, we split the data randomly into 10 equal subsets before the step 2. One of the subsets is used as
the validation sample, and the algorithm trained on the remaining 9 sets. This is repeated 10 times, each
time using a different subset as the validation sample, and the performance measures averaged across
the 10 runs. The use of 10-fold cross-validation reduces the risk of overfitting and increases the external
validity of the NLP algorithm we develop. Note, 10-fold cross-validation of this sort is computationally
intensive and impacts performance measures negatively and is not implemented in some existing papers
in business research. While the use of 10-fold cross-validation may negatively impact the performance
measures, it is necessary to increase external validity. Table 10 shows these metrics for different content
profiles. The performance is extremely good and comparable to performance achieved by the leading
financial information text mining systems (Hassan et al., 2011).
6. We repeat steps 2-5 until desired performance measures are achieved.
Tagging New Messages
1. For each new messages repeat steps 1-2 described above.
2. Use the ultimate classifier developed above to predict whether a particular type of content is present
or not.
and polynomial kernels. For more details, refer to Hastie et al. (2009).

42

One can think of this NLP algorithm as emulating the Turkers collective opinion in content-coding.

Figure 15:

Diagram of NLP Training and Tagging Procedure: This diagram shows the steps of training the NLP
algorithm and using the algorithm to tag the remaining messages. These steps are described in Appendix 3.

Rules and Attributes

Description

Bag of Words

Collects all the words and frequency for a message. Different variations include
collecting top N most occurring words.

Bigram

A bigram is formed by two adjacent words (e.g. Bigram is, is formed are bigrams).

Ratio of part-of-speech

Part-of-speech (noun, verb, etc) ratio in each message.

TF-IDF weighted informative word

Term-Frequency and Inverse Document Frequency weighs each word based on their
occurrence in the entire data and in a single message.

Specific Keywords

Specific keywords for different content can be collected and searched. e.g.,
Philanthropic messages have high change of containing the words donate and help.
For brand and product identification, large online lists were scraped and converted into
dictionaries for checking.

Frequency of different punctuation

Counts the number of different punctuations such as exclamation mark and question

marks

mark. This helps to identify emotion, questions, appearance of deals etc.

Count of non-alphanumerics

Counts the number of characters that are not A-Z and 0-9.

Table 9:

A Few Examples of Message Attributes Used in Natural Language Processing Algorithm

43

With Ensemble Learning (The

Best Performing Algorithm)

Without Ensemble Learning

(Support Vector Machine version
1 + Rule-based)
Accuracy Precision
Recall

Accuracy

Precision

Recall

REMFACT
EMOTION

0.94
0.97

0.99
0.99

0.68
0.87

0.88
0.94

0.99
0.98

0.33
0.65

HUMOR
PHILANTHROPIC

0.98
0.97

1
0.99

0.90
0.85

0.97
0.93

1
0.99

0.14
0.62

FRIENDLIKELY

0.94

0.99

0.68

0.90

0.99

0.41

SMALLTALK
DEAL

0.85
0.94

0.88
0.99

0.80
0.65

0.78
0.90

0.34
1

0.28
0.43

PRICECOMPARE

0.99

0.99

0.99

0.85

TARGETING
PRODAVAILABILITY

0.98
0.96

0.99
0.99

0.89
0.76

0.95
0.91

0.99
1

0.71
0.10

PRODLOCATION

0.97

0.99

0.90

0.87

0.11

Table 10:

Performance of Text Mining Algorithm on 5000 Messages Using 10-fold Cross-validation: This table
presents metrics for performance of the classification algorithms used. The left 3 columns show the metrics for the final algorithm
which combines classifiers via ensemble learning method while the right 3 columns show the metric for a support vector machine
algorithm. Notice that the support vector machine classifier tends to have low recall and high precision. Naive Bayes tends
to have high recall but low precision. Classifiers on their own cannot successfully overcome precision-recall tradeoff (if one is
higher, one is lower). But combining many different classifiers with ensemble learning can increase both precision and recall.
We obtain similar results for negative class labels.

44

Appendix 4: Obtaining the Number of Impressions

(d)

We discuss our procedure for constructing nkjt , the number of impressions for each message k of firm j in
(d)

day t split by demographic bin d for use in the EdgeRank correction model. As mentioned above, nkjt is
not directly reported by Company X (or made available to page-owners by Facebook Insights). Instead,
(d)

Company X reports njt , the number of impressions for all messages associated with firm j in demographic
(d)

bucket d on day t, which is essentially nkjt summed across all k associated with j. In addition, we observe
nkjt , the total number of impression obtained by a specific post k by firm j on day t. To assess how we
may split this across the various demographic buckets, we checked the extent to which pages release different
types of messages over time. The bulk of impressions for a message occur within the first week of its release.
(d)

Hence, the total impressions for a page on a given day out of a specific demographic bucket, njt , reflects the
aggregate impressions to users in that bucket of all messages released by that firm over the past one week.
Since EdgeRank allocates messages to users by message-type, if the firm releases the same type of messages
(d)

(i.e., photos, videos, status updates, apps or links) over a weeks duration, then the the distribution of njt

across the various k messages released by firm j within the past week may be roughly the same. In other
words, the distribution of demographics of the impressions of all messages released by a firm in the past
week should be the same if all those released messages are similar.
To check informally if this is the case, we picked a random sample of 10, 000 page-7-day combinations
from our data. For each combination, we collated all the messages released by that page during that 7-day
window and tabulated the type of these messages (i.e., photos, videos, status updates, apps or links). We
then construct two concentration metrics, C1 , the proportion of messages released by that page during that
7-day window that belong to the highest type bucket, and C2 , the proportion of messages released by that
page during that 7-day window that belong to the highest and second-highest type bucket. C1 and C2
are analogous to top-firm and top-two-firm concentration ratios used in industry-concentration studies, and
measure the extent to which the messages released by a page in a given 7-day period are spread across types.
If all messages released by a page during that 7-day window are of the same type, C1 and C2 will be both 1.
The spread away from 1 thus indicates higher variation in message-types released by an average Facebook
page over a weeks duration. Table 11 reports on the distribution of C1 and C2 we computed in this manner.
Looking at Table 11, we find that the median C1 is .71 (mean .72) and the median C2 is 1.0 (mean .94).
Most pages seem to be releasing at-most 2-types of messages within a week window, and more than 2/3rd of
messages released by an average page in an average week are of the same type. Given this, we assume that
(d)

(d)

(d)

nkjt is split into nkjt with the same distribution as given by njt . We construct the variable nkjt in the left
hand-side of the EdgeRank correction equation 2 in this manner. We replicated the study with different split
(d)

distributions such as 7-day aggregated distribution of njt , kernel smoothed distribution, and the results
were similar.
The method is not without its limitations. We view it as a practical way to deal with the lack of datareporting by Facebook, while exploiting the variation embedded in the observed impressions and to correlate
45

C1
C2

Min.

1st Qu.

Median

Mean

3rd Qu.

Max.

0.250
0.500

0.535
0.889

0.706
1.000

0.719
0.942

0.915
1.000

1.000
1.000

Table 11:

Distribution of the Top (C1 ) and Top-two (C2 ) Concentration Ratios of the type of messages Served
by a Facebook Page over a Randomly picked 7-day period

it with the observed variation in the the potential market for each message in each demographic bucket
%
&
(d)
(d)
Njt . The method produces potential measurement error in the dependent variable, nkjt in the EdgeRank
correction stage. Measurement error in the dependent variable is absorbed into the RHS unobservables and

is usually less of a concern unless it is correlated with the unobservables. The fact that we include page-fixed
(d)

effects separately for each demographic (1j in Equation 3) mitigates concerns to a large extent that these
unobservables may systematically be correlated with included characteristics.

46

Appendix 5: Result for AMT-Tagged & Different Models

NO ER COMMENT
Constant

-5.431*** (0.047)

OR

ER COMMENT

OR

-5.682*** (0.041)

NO ER LIKE

OR

-3.326*** (0.014)

ER LIKE

OR

-4.199*** (0.012)

SMOG

-0.098*** (0.001)

0.907

-0.125*** (0.001)

0.882

-0.045*** (0.000)

0.956

-0.066*** (0.000)

0.936

MSGLEN

-0.000*** (0.000)

1.000

-0.000*** (0.000)

1.000

-0.000*** (0.000)

1.000

-0.000*** (0.000)

1.000

HTTP

-0.253*** (0.005)

0.776

-0.143*** (0.005)

0.867

-0.097*** (0.002)

0.908

-0.025*** (0.002)

0.975

QUESTION

0.164*** (0.004)

1.178

0.191*** (0.004)

1.210

-0.325*** (0.001)

0.723

-0.220*** (0.001)

0.803

BLANK

0.739*** (0.012)

2.094

0.755*** (0.012)

2.128

-1.009*** (0.010)

0.365

-1.098*** (0.010)

0.334

ASKLIKE

-0.376*** (0.033)

0.687

-0.336*** (0.033)

0.715

-0.100*** (0.009)

0.905

-0.072*** (0.009)

0.931

ASKCOMMENT

0.927*** (0.049)

2.527

0.272*** (0.049)

1.313

0.111*** (0.020)

1.117

-0.406*** (0.020)

0.666

Persuasive
REMFACT

0.105*** (0.006)

1.111

0.145*** (0.005)

1.156

-0.063*** (0.002)

0.939

-0.032*** (0.002)

0.969

EMOTION

0.159*** (0.006)

1.172

0.135*** (0.006)

1.145

0.084*** (0.002)

1.088

0.106*** (0.002)

1.112

EMOTICON

-0.383*** (0.026)

0.682

-0.361*** (0.026)

0.697

-0.416*** (0.008)

0.660

-0.192*** (0.008)

0.825

HOLIDAYMENTION

-1.465*** (0.033)

0.231

-1.293*** (0.033)

0.274

-0.433*** (0.006)

0.649

-0.266*** (0.006)

0.766

HUMOR

-0.086*** (0.012)

0.918

0.163*** (0.012)

1.177

-0.275*** (0.004)

0.760

0.019*** (0.004)

1.019

PHILANTHROPIC

0.142*** (0.007)

1.153

0.138*** (0.007)

1.148

0.182*** (0.002)

1.200

0.168*** (0.002)

1.183

FRIENDLIKELY

-0.159*** (0.005)

0.853

-0.130*** (0.005)

0.878

0.206*** (0.002)

1.229

0.159*** (0.002)

1.172

SMALLTALK

-0.005 (0.004)

0.995

-0.051*** (0.004)

0.950

0.022*** (0.001)

1.022

-0.028*** (0.001)

0.972

Informative
BRANDMENTION

-0.179*** (0.004)

0.836

-0.228*** (0.004)

0.796

-0.141*** (0.001)

0.868

-0.224*** (0.001)

0.799

DEAL

-0.151*** (0.008)

0.860

0.011 (0.007)

1.011

-0.478*** (0.002)

0.620

-0.302*** (0.002)

0.739

PRICECOMPARE

0.688*** (0.193)

1.990

0.302 (0.193)

1.353

-0.498*** (0.103)

0.608

-0.857*** (0.103)

0.424

PRICE

-0.051*** (0.014)

0.950

-0.500*** (0.014)

0.607

-0.316*** (0.005)

0.729

-0.712*** (0.005)

0.491

TARGET

0.188*** (0.010)

1.207

0.076*** (0.010)

1.079

0.130*** (0.003)

1.139

0.059*** (0.003)

1.061

PRODAVAIL

-0.281*** (0.007)

0.755

-0.360*** (0.007)

0.698

0.073*** (0.002)

1.076

-0.022*** (0.002)

0.978

PRODLOCATION

-0.151*** (0.006)

0.860

-0.015** (0.005)

0.985

-0.239*** (0.002)

0.787

-0.080*** (0.002)

0.923

PRODMENTION

-0.170*** (0.005)

0.844

-0.349*** (0.005)

0.705

0.291*** (0.001)

1.338

0.095*** (0.001)

1.100

Message Type - App is the base

-Link

0.226*** (0.008)

1.254

-0.051*** (0.008)

0.950

0.004 (0.003)

1.004

-0.310*** (0.003)

0.733

-Photo

0.633*** (0.008)

1.883

0.379*** (0.008)

1.461

0.718*** (0.002)

2.050

0.418*** (0.002)

1.519

-Status Update

1.325*** (0.009)

3.762

1.148*** (0.009)

3.152

0.671*** (0.003)

1.956

0.553*** (0.003)

1.738

-Video

-0.191*** (0.027)

0.826

-0.447*** (0.027)

0.640

-0.636*** (0.011)

0.529

-0.928*** (0.011)

0.395

Industry Category - Celebrity is the base

-ConsumerProduct

-0.024*** (0.007)

0.976

-0.368*** (0.007)

0.692

-0.541*** (0.002)

0.582

-0.761*** (0.002)

0.467

-Entertainment

0.014* (0.006)

1.014

0.221*** (0.006)

1.247

-0.472*** (0.002)

0.624

-0.169*** (0.002)

0.845

-Organization

0.222*** (0.007)

1.249

0.072*** (0.007)

1.075

-0.163*** (0.002)

0.850

-0.161*** (0.002)

0.851

-PlaceBusiness

0.600*** (0.014)

1.822

0.339*** (0.014)

1.404

-0.698*** (0.007)

0.498

-0.869*** (0.007)

0.419

-Websites

-0.076*** (0.009)

0.927

0.152*** (0.008)

1.164

0.099*** (0.002)

1.104

0.398*** (0.002)

1.489

McFadden R-sq.

0.288

0.239

0.295

0.214

Nagelkerke R-sq.

0.288

0.239

0.296

0.216

Log-likelihood

-299060.434

-428096.493

-1728807.768

-3,119,513.948

Deviance

552293.46

810390.763

3365047.493

6,146,513.932

AIC

598244.869

856316.985

3457739.535

6,239,151.895

38706

38706

38706

38706

Significance *** 0.001 ** 0.01 * 0.05 . 0.1

Table 12: Aggregate Logistic Regression Results For Comments and Likes (5000 Messages):
This table presents the aggregate logistic regression on comments and Likes for both EdgeRank-corrected
(ER) and uncorrected (NO ER) for 5000 messages data tagged by Turkers. OR means Odds ratio and shows
the odds ratio for the estimates left of the column.
47

NO ER COMMENT
Constant

-8.122*** (0.004)

OR

ER COMMENT

OR

-7.129*** (0.004)

NO ER LIKE

OR

-5.168*** (0.001)

ER LIKE

OR

-4.205*** (0.001)

SMOG

-0.047*** (0.000)

0.954

-0.059*** (0.000)

0.943

-0.031*** (0.000)

0.969

-0.049*** (0.000)

0.952

MSGLEN

0.000*** (0.000)

1.000

-0.000*** (0.000)

1.000

-0.000*** (0.000)

1.000

-0.000*** (0.000)

1.000

HTTP

-0.545*** (0.002)

0.580

-0.301*** (0.002)

0.740

-0.388*** (0.000)

0.678

-0.085*** (0.000)

0.919

QUESTION

0.488*** (0.001)

1.629

0.581*** (0.001)

1.788

-0.284*** (0.000)

0.753

-0.186*** (0.000)

0.830

BLANK

1.002*** (0.003)

2.724

1.186*** (0.003)

3.274

-0.709*** (0.002)

0.492

-0.538*** (0.002)

0.584

ASKLIKE

0.080*** (0.010)

1.083

0.259*** (0.010)

1.296

0.494*** (0.003)

1.639

0.598*** (0.003)

1.818

ASKCOMMENT

0.514*** (0.021)

1.672

0.413*** (0.021)

1.511

-0.244*** (0.011)

0.783

-0.557*** (0.011)

0.573

REMFACT

-0.026*** (0.002)

0.974

0.017*** (0.002)

1.017

-0.055*** (0.001)

0.946

-0.015*** (0.001)

0.985

EMOTION

0.213*** (0.002)

1.237

0.217*** (0.002)

1.242

0.218*** (0.001)

1.244

0.191*** (0.001)

1.210

EMOTICON

0.171*** (0.004)

1.186

0.101*** (0.004)

1.106

-0.056*** (0.001)

0.946

-0.087*** (0.001)

0.917

HOLIDAYMENTION

-0.485*** (0.014)

0.616

-0.378*** (0.014)

0.685

-0.329*** (0.004)

0.720

-0.162*** (0.004)

0.850

HUMOR

0.030*** (0.001)

1.030

0.094*** (0.002)

1.099

-0.049*** (0.000)

0.952

0.047*** (0.000)

1.048

PHILANTHROPIC

0.182*** (0.002)

1.200

0.261*** (0.002)

1.298

0.019*** (0.001)

1.019

0.157*** (0.001)

1.170

FRIENDLIKELY

0.008*** (0.002)

1.008

-0.053*** (0.002)

0.948

0.076*** (0.001)

1.079

0.005*** (0.001)

1.005

SMALLTALK

0.059*** (0.002)

1.061

-0.111*** (0.002)

0.895

-0.061*** (0.001)

0.941

-0.172*** (0.001)

0.842

Persuasive

Informative
BRANDMENTION

0.013*** (0.002)

1.013

0.084*** (0.002)

1.088

-0.015*** (0.000)

0.985

0.043*** (0.000)

1.044

DEAL

-0.134*** (0.002)

0.875

-0.245*** (0.002)

0.783

-0.183*** (0.001)

0.833

-0.357*** (0.001)

0.700

PRICECOMPARE

-0.038*** (0.001)

0.963

0.008*** (0.001)

1.008

-0.033*** (0.000)

0.968

-0.020*** (0.000)

0.980

PRICE

-0.031*** (0.005)

0.969

-0.303*** (0.005)

0.739

-0.185*** (0.001)

0.831

-0.459*** (0.001)

0.632

TARGET

-0.021*** (0.002)

0.979

-0.135*** (0.002)

0.874

0.036*** (0.001)

1.037

-0.120*** (0.001)

0.887

PRODAVAIL

-0.090*** (0.002)

0.914

-0.003 (0.002)

0.997

-0.119*** (0.001)

0.888

0.038*** (0.001)

1.039

PRODLOCATION

-0.056*** (0.002)

0.946

-0.043*** (0.002)

0.958

0.072*** (0.001)

1.075

0.034*** (0.001)

1.035

PRODMENTION

-0.087*** (0.002)

0.917

-0.128*** (0.002)

0.880

0.076*** (0.001)

1.079

0.039*** (0.001)

1.040

Message Type - App is the base

-Link

0.201*** (0.003)

1.223

-0.301*** (0.003)

0.740

0.116*** (0.001)

1.123

-0.389*** (0.001)

0.678

-Photo

0.876*** (0.003)

2.401

0.318*** (0.003)

1.374

1.036*** (0.001)

2.818

0.475*** (0.001)

1.608

-Status Update

1.118*** (0.003)

3.059

0.666*** (0.003)

1.946

0.458*** (0.001)

1.581

-0.015*** (0.001)

0.985

-Video

0.641*** (0.009)

1.898

0.108*** (0.009)

1.114

0.218*** (0.003)

1.244

0.203*** (0.003)

1.225

Industry Category - Celebrity is the base

-ConsumerProduct

0.068*** (0.002)

1.070

-0.062*** (0.002)

0.940

-0.446*** (0.001)

0.640

-0.544*** (0.001)

0.580

-Entertainment

0.420*** (0.002)

1.522

0.660*** (0.002)

1.935

-0.260*** (0.001)

0.771

-0.067*** (0.001)

0.935

-Organization

0.497*** (0.002)

1.644

0.631*** (0.002)

1.879

-0.005*** (0.001)

0.995

0.110*** (0.001)

1.116

-PlaceBusiness

0.381*** (0.005)

1.464

0.234*** (0.005)

1.264

-0.726*** (0.002)

0.484

-0.933*** (0.002)

0.393

-Websites

0.218*** (0.003)

1.244

0.380*** (0.003)

1.462

0.183*** (0.001)

1.201

0.314*** (0.001)

1.369

1st Stage Residual

-0.000*** (0.000)

1.000

0.000*** (0.000)

1.000

-0.000*** (0.000)

1.000

0.000*** (0.000)

0.000

McFadden R-sq.

0.229

0.199

0.298

0.270

Nagelkerke R-sq.

0.230

0.199

0.298

0.272

Log-likelihood

-2575187.42

-3460249.739

-14563950.18

-24911467.22

Deviance

4745694.741

6516530.605

28179842.5

48875491.61

AIC

5150474.839

6920599.479

29128000.36

49823034.45

665916

665916

665916

665916

Significance *** 0.001 ** 0.01 * 0.05 . 0.1

Table 13: Aggregate Logistic Regression Results For Comments and Likes (All Messages with
the 1st Stage Residuals as a Control Function): This table presents the aggregate logistic regression
on comments and Likes for both EdgeRank-corrected (ER) and uncorrected (NO ER) for all messages data.
OR means Odds ratio and shows the odds ratio for the estimates left of the column. This second-stage model
includes first-stage residuals as a control function.
48

Variable

Intercept only

Constant
-7.052***(0.001)
SMOG
MSGLEN
HTTP
QUESTION
BLANK
ASKLIKE
ASKCOMMENT
Message Type - App is the base
-Link
-Photo
-Status Update
-Video
Industry Category - Celebrity is the base
-ConsumerProduct
-Entertainment
-Organization
-PlaceBusiness
-Websites
FRIENDLIKELY
REMFACT
EMOTION
EMOTICON
HOLIDAYMENTION
HUMOR
PHILANTHROPIC
SMALLTALK
BRANDMENTION
DEAL
PRICECOMPARE
PRICE
TARGET
PRODAVAIL
PRODLOCATION
PRODMENTION

McFadden R-sq.
Nagelkerke R-sq.
Log-likelihood
Deviance
AIC
N

-4267283.759
8130598.643
8534569.517
665916

Controls

Friendlikely

Persuasive

Informative

All (No demo controls)

-6.906***(0.003)
-0.065***(0.000)
-0.000***(0.000)
-0.406***(0.001)
0.543***(0.001)
1.172***(0.003)
0.222***(0.010)
0.508***(0.021)

-6.952***(0.004)
-0.065***(0.000)
-0.000***(0.000)
-0.400***(0.001)
0.546***(0.001)
1.168***(0.003)
0.229***(0.010)
0.529***(0.021)

-6.975***(0.004)
-0.065***(0.000)
-0.000***(0.000)
-0.393***(0.001)
0.557***(0.001)
1.187***(0.003)
0.220***(0.010)
0.465***(0.021)

-6.861***(0.004)
-0.067***(0.000)
-0.000***(0.000)
-0.390***(0.002)
0.541***(0.001)
1.144***(0.003)
0.223***(0.010)
0.506***(0.021)

-6.889***(0.004)
-0.068***(0.000)
-0.000***(0.000)
-0.355***(0.002)
0.564***(0.001)
1.146***(0.003)
0.221***(0.010)
0.469***(0.021)

-0.388***(0.003)
0.376***(0.003)
0.656***(0.003)
0.406***(0.009)

-0.389***(0.003)
0.371***(0.003)
0.654***(0.003)
0.406***(0.009)

-0.394***(0.003)
0.365***(0.003)
0.644***(0.003)
0.410***(0.009)

-0.376***(0.003)
0.372***(0.003)
0.645***(0.003)
0.395***(0.009)

-0.370***(0.003)
0.373***(0.003)
0.641***(0.003)
0.398***(0.009)

-0.402***(0.002)
0.514***(0.002)
0.391***(0.002)
-0.087***(0.005)
0.123***(0.002)

-0.392***(0.002)
0.519***(0.002)
0.393***(0.002)
-0.082***(0.005)
0.134***(0.002)
0.064***(0.001)

-0.371***(0.002)
0.537***(0.002)
0.409***(0.002)
-0.052***(0.005)
0.150***(0.003)
-0.051***(0.002)
-0.065***(0.002)
0.217***(0.002)
0.109***(0.004)
-0.391***(0.014)
0.022***(0.001)
0.054***(0.002)
-0.055***(0.002)

-0.374***(0.002)
0.514***(0.002)
0.400***(0.002)
-0.065***(0.005)
0.136***(0.003)

-0.347***(0.002)
0.529***(0.002)
0.408***(0.002)
-0.021***(0.005)
0.182***(0.003)
-0.006***(0.002)
0.014***(0.002)
0.256***(0.002)
0.121***(0.004)
-0.388***(0.014)
0.072***(0.002)
0.174***(0.002)
-0.086***(0.002)
0.081***(0.002)
-0.172***(0.002)
-0.006***(0.001)
-0.317***(0.005)
-0.071***(0.002)
-0.064***(0.002)
0.011***(0.002)
-0.151***(0.002)
Demographic proportion
variables excluded
0.171
0.172
-3570184.37
6736399.867
7140466.74
665916

0.071***(0.002)
-0.124***(0.002)
0.043***(0.001)
-0.368***(0.005)
0.049***(0.002)
0.028***(0.002)
0.023***(0.002)
-0.186***(0.002)

0.161
0.161
0.165
0.161
0.162
0.165
-3612205.117
-3610817.72
-3597891.814
6820441.36
6817666.566
6791814.753
7224474.234
7221701.44
7195863.627
665916
665916
665916
Significance *** 0.001 ** 0.01 * 0.05 . 0.1

0.164
0.165
-3598824.088
6793679.301
7197730.175
665916

Table 14: Logistic Regression EdgeRank-Corrected Estimates Model Comparison (Comments)

49

Variable
Intercept only
Constant
-4.662***(0.000)
SMOG
MSGLEN
HTTP
QUESTION
BLANK
ASKLIKE
ASKCOMMENT
Message Type - App is the base
-Link
-Photo
-Status Update
-Video
Industry Category - Celebrity is the base
-ConsumerProduct
-Entertainment
-Organization
-PlaceBusiness
-Websites
FRIENDLIKELY
REMFACT
EMOTION
EMOTICON
HOLIDAYMENTION
HUMOR
PHILANTHROPIC
SMALLTALK
BRANDMENTION
DEAL
PRICECOMPARE
PRICE
TARGET
PRODAVAIL
PRODLOCATION
PRODMENTION

McFadden R-sq.
Nagelkerke R-sq.
Log-likelihood
Deviance
AIC
N

-33968732.078
66990021.316
67937466.156
665916

Controls
-3.929***(0.001)
-0.057***(0.000)
-0.001***(0.000)
-0.214***(0.000)
-0.198***(0.000)
-0.603***(0.002)
0.576***(0.003)
-0.379***(0.011)

Friendlikely
-3.979***(0.001)
-0.057***(0.000)
-0.000***(0.000)
-0.207***(0.000)
-0.195***(0.000)
-0.612***(0.002)
0.583***(0.003)
-0.359***(0.011)

Persuasive
-3.931***(0.001)
-0.057***(0.000)
-0.001***(0.000)
-0.205***(0.000)
-0.181***(0.000)
-0.592***(0.002)
0.576***(0.003)
-0.444***(0.011)

Informative
-3.921***(0.001)
-0.060***(0.000)
-0.000***(0.000)
-0.205***(0.000)
-0.198***(0.000)
-0.602***(0.002)
0.564***(0.003)
-0.404***(0.011)

All (No demo controls)

-3.892***(0.001)
-0.061***(0.000)
-0.000***(0.000)
-0.189***(0.000)
-0.175***(0.000)
-0.596***(0.002)
0.565***(0.003)
-0.465***(0.011)

-0.517***(0.001)
0.572***(0.001)
-0.062***(0.001)
0.244***(0.003)

-0.519***(0.001)
0.567***(0.001)
-0.067***(0.001)
0.245***(0.003)

-0.518***(0.001)
0.561***(0.001)
-0.083***(0.001)
0.252***(0.003)

-0.511***(0.001)
0.566***(0.001)
-0.067***(0.001)
0.225***(0.003)

-0.505***(0.001)
0.561***(0.001)
-0.083***(0.001)
0.231***(0.003)

-0.948***(0.001)
-0.172***(0.001)
-0.165***(0.001)
-1.305***(0.002)
0.028***(0.001)

-0.937***(0.001)
-0.168***(0.001)
-0.164***(0.001)
-1.301***(0.002)
0.038***(0.001)
0.072***(0.000)

-0.941***(0.001)
-0.174***(0.001)
-0.160***(0.001)
-1.292***(0.002)
0.019***(0.001)
0.051***(0.001)
-0.065***(0.001)
0.250***(0.001)
-0.049***(0.001)
-0.187***(0.004)
-0.009***(0.000)
-0.072***(0.001)
-0.112***(0.001)

-0.939***(0.001)
-0.189***(0.001)
-0.171***(0.001)
-1.294***(0.002)
0.042***(0.001)

-0.932***(0.001)
-0.193***(0.001)
-0.171***(0.001)
-1.275***(0.002)
0.041***(0.001)
0.080***(0.001)
-0.021***(0.001)
0.260***(0.001)
-0.020***(0.001)
-0.183***(0.004)
0.009***(0.000)
0.002*(0.001)
-0.146***(0.001)
0.021***(0.000)
-0.207***(0.001)
-0.047***(0.000)
-0.471***(0.001)
-0.019***(0.001)
-0.073***(0.001)
0.138***(0.001)
0.012***(0.001)
Demographic proportion
variables excluded
0.201
0.203
-27221231.997
53495021.154
54442561.994
665916

0.012***(0.000)
-0.166***(0.001)
-0.036***(0.000)
-0.504***(0.001)
0.050***(0.001)
-0.027***(0.001)
0.134***(0.001)
-0.004***(0.001)

0.191
0.191
0.192
0.193
-27584573.527
-27566219.315
54221704.214
54184995.789
55169211.054
55132504.629
665916
665916
Significance *** 0.001 ** 0.01

0.195
0.197
-27424516.699
53901590.557
54849113.398
665916
* 0.05 . 0.1

0.196
0.197
-27419099.696
53890756.552
54838281.393
665916

Table 15: Logistic Regression EdgeRank-Corrected Estimates Model Comparison (Likes)

50