Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Essays on consumer conversations in social media
(USC Thesis Other)
Essays on consumer conversations in social media
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Running head: CONSUMER CONVERSATIONS IN SOCIAL MEDIA 1
ESSAYS ON CONSUMER CONVERSATIONS IN SOCIAL MEDIA
by
ABHISHEK BORAH
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree of
DOCTOR OF PHILOSOPHY
(BUSINESS ADMINISTRATION)
May 2013
Copyright 2013 Abhishek Borah
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 2
Dedication
This dissertation would not have been possible without the support of the three people
who matter most to me…..
My mother, for giving me all the love;
My father, for giving me all the courage;
My wife, Pongkhi, for her persistence, her faith in me, and blessing me with her angelic
presence, encouragement, and love for the last 14 years;
Thank you very much! I owe this effort to you three!
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 3
Acknowledgments
I want to express my sincerest thanks to the people who have supported me in completing
this dissertation.
My advisor, Prof. Gerard J. Tellis, has been my guiding light in my academic career till
now. His kindness, humility, insightful knowledge of the field of marketing and business
management, and unrelenting passion for research are traits any advisee would dream of. I have
been very lucky to have him as my mentor. He has wholeheartedly encouraged and supported
both my wife and me here at USC. I cannot be more thankful to him for his time and patience
with me.
My dissertation committee members – Donna Hoffman, Lan Luo and Thomas Valente
have helped me in developing and improving my essays. They have been always available to me
with their valuable time and thought.
The people at USC – Professors S. Siddarth, Rex Kovacevich, Dennis Rook, Allen
Weiss, and Roger Moon have always supported me and I thank them for this. Michelle Lee,
Elizabeth Mathew, Ruth Joya, and Yvonne King have always been there when I needed
administrative help.
My friends and colleagues at USC have been there whenever I needed time off academics
or when I needed advice during my studies. I would especially like to thank Pankaj Mishra.
My close friends – Gautam Kadian, Amlan Goswami, and Tom Cirillo - Thank you for
being there over the years.
My favorite musicians for getting me through times when I needed them.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 4
Table of Contents
List of Tables ......................................................................................................... 7
List of Figures ........................................................................................................ 9
Abstract ................................................................................................................ 10
Chapter 1. Introduction and Outline .................................................................... 11
Motivation ..................................................................................................... 11
Outline........................................................................................................... 11
Objective ....................................................................................................... 12
Chapter 2. Is All That Twitters Gold? ................................................................. 14
Introduction ................................................................................................... 14
Conceptual Background ................................................................................ 17
Theory ...................................................................................................... 17
Indirect route ........................................................................................ 17
Social spillover effect .................................................................. 18
Information effect ........................................................................ 19
Direct route .......................................................................................... 20
Method .......................................................................................................... 22
Research Design....................................................................................... 22
Sampling of firms ................................................................................ 22
Sampling of time .................................................................................. 22
Data Collection ........................................................................................ 23
Measures .................................................................................................. 26
Measures of Twitter metrics ................................................................ 26
Measures of stock returns .................................................................... 26
Measures of control variables .............................................................. 28
Analysts’ forecasts ....................................................................... 28
Advertising ................................................................................... 28
Key firm developments ................................................................ 28
Model ............................................................................................................ 29
Test for Endogeneity using Granger Causality ........................................ 29
Test for Unit Root and Cointegration ...................................................... 30
Vector Auto-Regressive Model with Exogenous Variables .................... 31
Generalized Impulse Response Function ................................................. 32
Generalized Forecast Error Variance Decomposition ............................. 32
Results ........................................................................................................... 33
Descriptive Analysis ................................................................................ 33
Granger Causality Tests To Check for Endogeneity ............................... 37
Test for Unit Root and Cointegration ...................................................... 38
Dynamics of Twitter Metrics on Stock Returns ...................................... 39
Relative Importance of Twitter Metrics ................................................... 42
Effect of Control Variables on Twitter Metrics ....................................... 43
Robustness Analysis ................................................................................ 43
Effects of other media variables ........................................................... 43
Effects of other Twitter metrics. ........................................................... 47
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 5
Portfolio analysis... ............................................................................... 48
Firm tweets as endogenous... ................................................................ 48
Discussion ..................................................................................................... 49
Summary of Findings ............................................................................... 49
Discussion of Key Issues ......................................................................... 50
Why does the effect of Twitter metrics wear-in immediately? ............ 50
Why is Retweet the most important metric of Twitter? ....................... 51
What type of content does Twitter capture? ........................................ 52
Why is valence of tweets not important? ............................................. 53
Implications.............................................................................................. 53
Limitations and Further Research ............................................................ 54
Chapter 3. Halo Effects in Social Media ............................................................. 55
Introduction ................................................................................................... 55
Theory ........................................................................................................... 58
Definitions................................................................................................ 59
Why does Perverse Halo Occur? ............................................................. 59
Method .......................................................................................................... 62
Research Design....................................................................................... 62
Industry context ................................................................................... 62
Sampling of brands .............................................................................. 64
Timeframe ............................................................................................ 65
Quasi-Experimental design .................................................................. 65
Data Collection of Online Conversations ................................................ 65
Measures of Endogenous Variables ......................................................... 67
Measures of online conversation ......................................................... 67
Measure of citations of print about recalls ........................................... 68
Measures of Control Variables ................................................................ 68
Recalls .................................................................................................. 69
ABC news coverage ............................................................................. 69
Negative events in Toyota’s acceleration crisis ................................... 70
Advertising ........................................................................................... 70
New product announcements ............................................................... 71
Key developments ................................................................................ 71
Model ............................................................................................................ 71
Empirical strategy .................................................................................... 71
Why Vector Auto-Regressive framework? .............................................. 73
Test for Unit Root and Cointegration ...................................................... 73
Vector Auto-Regressive Model with Exogenous Variables .................... 74
Generalized Impulse Response Function ................................................. 76
Generalized Forecast Variance Decomposition ....................................... 77
Results ........................................................................................................... 78
Descriptive Results .................................................................................. 78
Testing for Evolution ............................................................................... 79
Estimation of VARX Model .................................................................... 79
Estimates of VARX for Japanese brands ............................................. 81
Estimates of VARX for American and Japanese brands ..................... 86
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 6
Relative Importance of Concerns and Citations of Print ......................... 89
Effect on Market Capitalization ............................................................... 89
Robustness Analyses ................................................................................ 90
Positive conversations .......................................................................... 90
Brand size............................................................................................. 91
Presence of Toyota ............................................................................... 92
Car type ................................................................................................ 92
Omitted variable................................................................................... 92
Discussion ..................................................................................................... 93
Summary of Findings ............................................................................... 93
Discussion of Key Issues ......................................................................... 94
Why does perverse halo exist? ............................................................. 94
What is the mechanism behind perverse and reverse halo? ................. 95
Why do these halo effects have a short wear-in? ................................. 96
Implications................................................................................................... 97
Limitations .................................................................................................... 99
References .......................................................................................................... 100
Appendix A: Classification Accuracy of Support Vector Algorithm ................ 113
Appendix B: Descriptive Statistics for Chapter 2 .............................................. 115
Appendix C: Model Results by Brand of Chapter 2 .......................................... 121
Appendix D: Details of Classification Algorithm of Chapter 3 ........................ 124
Appendix E: Descriptive Statistics of Chapter 3 ............................................... 126
Appendix F: Robustness Results of Chapter 3 .................................................. 127
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 7
List of Tables
Table 1. Steps for Creation of Twitter Metrics ............................................................................. 24
Table 2. Volume of Tweets on Weekends and Weekdays ............................................................ 35
Table 3. Weekday to Weekend Ratio for Various Metrics ........................................................... 36
Table 4. Result of Granger Causality Test of Twitter Metrics on Stock Returns ......................... 38
Table 5. Result of Granger Causality Test of Stock Returns on Twitter Metrics ......................... 38
Table 6. Summary of Unit Root Tests of the Endogenous Variables ........................................... 39
Table 7. Duration of Impact of Twitter Metrics on Stock Returns ............................................... 40
Table 8. Impact of Twitter Metrics on Stock Returns .................................................................. 41
Table 9. Description of Other Media Variables ............................................................................ 44
Table 10. Impact of Retweet and Other Media on Stock Returns ................................................ 45
Table 11. Impact of Other Media on Retweet ............................................................................... 46
Table 12. Augmented Dickey Fuller Test by Brand ..................................................................... 80
Table 13. Phillips-Perron Test by Brand ....................................................................................... 80
Table 14. Interpretation of VAR coefficients ............................................................................... 82
Table 15. Full VARX Coefficient Matrix for Japanese Brands .................................................... 83
Table 16. Full VARX Coefficient Matrix for Toyota and Chrysler ............................................. 87
Table A1. Illustration of Confusion Matrix ................................................................................ 113
Table A2. Classification Accuracy of Support Vector Machine Algorithm ............................... 114
Table B1. Correlation of Key Variables ..................................................................................... 120
Table B2. Descriptive Statistics of Key Variables ...................................................................... 120
Table C1. Granger Causality Tests by Brand of Twitter Metrics on Stock Returns .................. 121
Table C2. Granger Causality Tests by Brand of Stock Returns on Twitter Metrics .................. 121
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 8
Table C3. Augmented Dickey Fuller Test by Brands ................................................................ 122
Table C4. Short and Cumulative Impact of Twitter Metrics on Stock Returns .......................... 122
Table C5. Short and Cumulative Impact of Retweet and Other Media on Stock Returns . ....... 123
Table E1. Descriptive Statistics .................................................................................................. 126
Table E2. Correlation of Key Variables ..................................................................................... 126
Table F1. Reduced VARX Matrix for Negative Sentiment among Japanese Brands ................ 127
Table F2. Reduced VARX Matrix for Negative Sentiment between Toyota and Chrysler ....... 127
Table F3. Reduced VARX Matrix for Scaled Negative Sentiment among Japanese Brands ..... 128
Table F4. Reduced VARX Matrix for Scaled Negative Sentiment for Toyota and Chrysler .... 128
Table F5. VARX Coefficient Matrix for Only Honda and Nissan ............................................. 129
Table F6. Reduced VARX Matrix for Luxury Brands ............................................................... 129
Table F7. Reduced VARX Coefficient Matrix for Non-Luxury Brands .................................... 130
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 9
List of Figures
Figure 1. Conceptual framework of how tweets relate to stock returns 17
Figure 2. Time series plot of twitter metrics for all brands 34
Figure 3. Time series plot of twitter metrics and stock price 37
Figure 4. Number of automobile recalls in USA 56
Figure 5. Graph of recalls and concerns for Toyota, Honda and Nissan 72
Figure 6. Impulse response from Toyota’s concerns to Honda’s concerns 84
Figure 7. Impulse response from Honda’s citations to Nissan’s concerns 85
Figure 8. Impulse response from Toyota’s citations to Toyota’s concerns 85
Figure 9. Impulse response from Toyota’s citations to Chrysler’s concerns 88
Figure 10. Impulse response from Chrysler’s to Toyota’s concerns 88
Figure B1. Time series plot of twitter metrics for Best Buy 115
Figure B2. Time series plot of twitter metrics for Canon 115
Figure B3. Time series plot of twitter metrics for Dell 116
Figure B4. Time series plot of twitter metrics for iPhone 116
Figure B5. Time series plot of twitter metrics for Kindle 117
Figure B6. Time series plot of twitter metrics for PlayStation 117
Figure B7. Time series plot of twitter metrics for Priceline 118
Figure B8. Time series plot of twitter metrics for Sandisk 118
Figure B9. Time series plot of twitter metrics for TiVo 119
Figure B10. Time series plot of twitter metrics for Xbox 119
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 10
Abstract
Consumer conversations in social media provide a rich forum for market research. They
are spontaneous, passionate, information rich, and live. These conversations can be informative
of firm performance. In fact, the ubiquity and availability of such conversations at higher
disaggregate temporal levels (e.g., minutes) than sales and market shares mean that such
conversations can provide an early diagnosis of firm performance and an early warning and
diagnosis of potential problems.
The second chapter of this dissertation first examines if conversations about brands in
Twitter, i.e., tweets, can affect stock returns of firms that own the brands. The results indicate
that there is a significant positive relationship between Twitter metrics and stock returns. The
word of mouth metric “retweet” affects and explains stock returns the most and the effect wears-
in immediately. These results imply that managers should monitor conversations in Twitter and
include it as part of their marketing research. Moreover, the focus on should be on monitoring
word of mouth rather than volume and valence.
The third chapter of this dissertation explores the effect of one brand’s product recall on
online conversations of a rival. The results suggest that consumers make similar inferences for
brands that belong to the same country and opposing inferences for brands that belong to a
different country. These perverse and reverse halo effects have a short wear-in and a modest
wear-out. These results imply that brands from the same country should keep an eye on each
other’s recalls while brands could emphasize their strengths and uniqueness when a brand from
another country is under a recall crisis.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 11
Chapter 1: Introduction and Outline
Motivation
“If you make customers unhappy in the physical world, they might each tell 6 friends. If
you make customers unhappy on the Internet, they can each tell 6,000 friends.”
Jeff Bezos, Founder, Chairman and CEO of Amazon.com
Nowadays, nearly four in five active Internet users engage in social media such as
Facebook, Twitter, and YouTube (Nielsen, 2011). In 2012, 1.43 billion users worldwide spent
time in a social networking site, which translates into 20% of the world’s population (Arno,
2012). These users leave a trail of conversations in social media sites. These conversations are
spontaneous, passionate, information rich, and live.
These conversations can provide a rich forum for market research. The ubiquity and
availability of such conversations at higher disaggregate temporal levels (e.g., minutes) than
sales and market shares mean that such conversations can provide an early diagnosis of firm
performance and an early warning of potential problems.
Specifically I can ask questions such as: Do consumer conversations in social media sites
provide signals about the firm’s performance? Can a product recall for one brand increase the
volume of negative online consumer conversations for another brand?
I explore each of these questions in chapters 2 and 3.
Outline
The first chapter of the dissertation summarizes the motivation, outline, and objectives of
the research. Chapters 2 and 3 are two separate essays examining if (1) conversations in Twitter
can be a leading indicator of stock returns, and (2) negative online conversations about one brand
in a product recall can increase or decrease negative online conversations of another brand.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 12
Objective
The key objective of the dissertation is to evaluate if consumer conversations in social
media sites can be leading indicators of firm performance. Both essays are related to the same
substantive field of social media. In one essay, I look at conversations about the focal firm and
the effect of these conversations on the firm’s performance. In the other essay, I look at the effect
of a firm’s conversations on its rivals’ conversations. My findings and implications are targeted
to marketing managers, social media managers, chief marketing officers, investors, and
researchers of social media and online word of mouth. These findings show the strategic
diagnosticity of online conversations.
In this section, I briefly present the implications and the techniques used in each study.
In chapter 2, I examine if conversations about brands in Twitter affect the stock market
returns of the firm. I collect around 9 million tweets for ten brands such as iPhone, Xbox, Kindle.
I use text mining and natural language processing techniques (i.e. Support Vector Machine) to
create measures of volume, valence, and word of mouth for Tweets. I collect and control for
other media such as citations of print (e.g., Wall Street Journal), influential blogs, and consumer
forums.
The findings have the following implications. First, managers should monitor
conversations in Twitter and include it as part of their marketing research. Second, managers
should monitor the Retweet metric. Third, managers could target users with huge follower bases
to Retweet positive information about their brand. Fourth, investors can use the information in
Twitter in their trading strategies.
In chapter 3, I examine if negative conversations about brands hurt or help rival brands. I
test the existence of such effects in a unique sample of around 6000 online daily conversations of
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 13
Japanese and American automobiles during a series of product recalls, which includes the Toyota
acceleration crises in winter 2009-10. I use the Vector AutoRegressive model exploiting the
natural experiment in recalls. I collect data from thousands of social media sites and citations of
print. I use text mining analysis to uncover the mechanism of why negative conversations of one
brand can hurt or help a rival.
My findings have the following implications. First, brands from the same country should
keep a close eye on rival’s recall events. Second, brands from a different country to the recalled
brand could emphasize their strengths and uniqueness when the recalled brand is under crisis.
Third, managers need to monitor conversations in social media during product recalls. Fourth,
marketing managers of the recalled brand need to focus on managing both mass media (e.g.
print) and social media. Fifth, investors can use online conversations in their trading strategies.
I use a variety of methods in the essays. I use the Vector Support Machine Algorithm for
creating the valence metrics in essay 1, Text Mining (Lexicon) analysis in essay 2, financial 4-
factor model in both essays using a daily event study, Portfolio analysis in essay 1, and Vector
Autoregressive models in both essays.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 14
Chapter 2: Is All That Twitters Gold? Market Value of Brand Conversations in Social Media
Introduction
Currently, nearly four in five active Internet users engage in social media such as
Facebook, Twitter, and YouTube (Nielsen, 2011). This behavior of consumers has led marketers
to increasingly adopt social media. eMarketer estimates 88% of US businesses will be involved
in social media marketing in 2012 (Williamson, 2010). Marketers are participating in social
media in many ways, such as by advertising their brands on social network sites and creating
Twitter accounts, YouTube channels, Facebook pages, and official blogs. Additionally,
marketers are also engaging in monitoring what consumers and prospects are saying within
social media about their brands. Some scholars attribute Barack Obama’s victory in the
Democratic primary to his skillful use of social media (Deighton & Kornfield, 2009).
In this “attention economy” (Davenport & Beck, 2002), where human attention is a
scarce commodity, digital conversations in social media provide a rich forum for consumer
research. These conversations allow firms to gauge consumer attitude, attachment, and advocacy
for their products and brands. Moreover, such consumer feedback is an important source of
information on firm performance as it reflects the truth about the firm and involves real
conversations laden with emotion.
Twitter, one of the most popular sites within social media, provides a cornucopia of
digital conversations called ‘tweets’ about brands. Launched on July 13th 2006, the site is a
messaging service that has characteristics of social networks, blogging, and texting. As of
October 2011, the site has a growing user base of around 254 million users and serves around
250 million tweets per day originating from both desktop and mobile devices (Bennett, 2011).
Among social network sites in the U.S. including Facebook, Twitter saw the highest percentage
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 15
growth in its mobile audience over the past year (Comscore, 2011). Twitter’s 140 character limit
captures instantaneous feelings and sentiments, which may not be possible in other realms of
user generated content. Moreover, the directed network structure of Twitter allows a rich
measure of word of mouth (retweet) of a brand.
Only, three published studies have attempted to link digital conversations in Twitter with
firm performance. Two recent papers demonstrate the usefulness of conversations in Twitter to
forecast movie revenues (Asur & Huberman, 2010; Rui, Whinston, & Winkler, 2010). Another
recent study (Bollen, Mao, & Zeng, 2011) shows that Twitter feeds can predict the Dow Jones
Industrial Average (DJIA). Bollen, Mao, & Zeng’s (2011) study is promising and raises the
following opportunities. First, it focuses on aggregated tweets, with or without brand mentions,
and averaged stock prices across firms. It would be interesting to determine if such results hold
for individual firms, where variation in conversations and stock prices is much greater. Second, it
does not control for risk free rate, market-wide shocks, media coverage, and analysts’ forecasts,
which can affect returns (Tetlock, 2007). It would be useful to evaluate if tweets indeed affect
returns accounting for other explanatory factors. Third, it uses only one metric of tweets, mood
(valence), ignoring metrics such as interest (volume) and word of mouth (e.g., retweet). Some
commentators have expressed doubts whether defining success by number of Twitter followers
and YouTube views has any merit (Baker, 2009). It would be worthwhile to investigate how
valence holds up against other conversation metrics, especially retweet (word of mouth).
Fourth, it does not explore the dynamics of the relationship between tweets and stock
price. It would be valuable to know how fast or slowly the stock market reflects information in
tweets (wearin) and how that effect decays (wearout). Thus, the current study seeks to answer the
following questions:
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 16
• Do conversations about brands in Twitter affect the stock market returns of the firm
that own those brands?
• If so, among the various metrics of Twitter conversations, which metric best captures
this effect?
• What are the dynamics of this relationship in terms of wearin and wearout?
In order to answer the queries above, I collect data for around 9 million conversations for
ten brands from Twitter. I focus on stock market returns because it is of utmost importance to
firms, is widely available at a disaggregate level, and reflects the consensus forecast of millions
of investors about the financial health of a firm (Srinivasan & Hanssens, 2009).
I use text mining and natural language processing techniques to create measures of
volume, valence, and word of mouth of Twitter conversations. Using data at a daily level for a
period of 6 months from October 2009, I evaluate whether the Twitter metrics have an impact on
the stock market returns of firms. I use Vector-AutoRegressive models with exogenous variables
(VARX) to capture these effects. The exogenous variables I use are analysts’ forecasts,
advertising expenditures, and media coverage of key firm developments such as innovations,
mergers, announcement of earnings, client contracts, strategic alliances, lawsuits, and changes in
key executives. In the robustness analysis, I further control for citations of print, influential
blogs, and consumer forums.
The rest of the chapter is organized as follows. The second section presents the theory,
the third section explains the method, the fourth section describes the model, and the last two
sections present the results and discussion.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 17
Conceptual Background
This section builds the conceptual framework for the relationship between tweets and
stock market returns.
Theory
I propose that tweets can affect stock returns either through an indirect or a direct route.
Figure 1 depicts these routes. I next elaborate on each of these routes.
Figure 1. Conceptual framework of how tweets relate to stock returns
Indirect route. Studies using both offline and online data show that word-of-mouth
influences the purchasing behavior of prospective consumers (e.g., Bowman & Narayandas,
2001 for offline; Liu, 2006 for online). Twitter is currently the most immediate measurable
medium for consumer word-of-mouth. Thus, Twitter may serve as a medium for consumers to
spread information about a brand, which influences the sales (and consequently the earnings) of
the firm. Studies have established the influence of earnings on stock returns (e.g., Barth, Beaver,
& Landsman, 2001).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 18
In this indirect route, Twitter metrics may affect stock returns due to two types of user
reactions: the social spillover effect and the information effect. I describe each of these reactions
below.
Social spillover effect. In Twitter, users “follow” or subscribe to another user’s tweets.
Unlike other social network sites such as Facebook, in Twitter, a user A “following” user B does
not imply that B “follows” A. This unidirectional network structure suggests that users are
susceptible to be influenced by the tweets of those they follow. Social influence theory suggests
that individuals get swayed to change their attitude and behavior due to other people (Kelman,
1958). This change is induced by the identification and connection a user has with someone they
like or respect (Kelman, 1958). “Following” can be interpreted as a way of identifying and
connecting with the “followed” user. Indeed, researchers find that the direction of a tie matters in
influencing behavior (Christakis & Fowler, 2007). By design, a signed-in user receives every
tweet from the users they follow.
1
Thus, an increase in the volume of tweets about a high quality
brand from a user can increase the brand’s awareness in the user’s “follower” network. This
increase in awareness could further influence the “followers” to buy the brand.
Another way through which the “social spillover effect” occurs is when a follower
endorses a user’s original tweet and the follower’s followers are influenced to buy the brand.
This effect is enabled by the retweet feature in Twitter. Social media commentators suggest that
retweet is the most important feature of Twitter (Israel, 2009). It is a social gesture indicating
endorsement of an idea or opinion. Users often acquire information indirectly via retweets than
directly from those they “follow” (Kwak, Lee, Park, & Moon, 2010). The act of retweeting does
1
Users are also exposed to a “reply” to a previous tweet from a user who they don’t follow, but one who follows the
user they follow. Since they do not follow the user who “replies”, the possibility of them getting influenced by the
content in the “reply” is much lower.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 19
not happen automatically in Twitter. It requires active involvement from users. Users first get
exposed to the tweet and then decide whether or not to endorse the tweet. In a structural sense,
retweeting is the Twitter analog to email forwarding (Boyd, Golder, & Lotan, 2010). If a tweet is
endorsed i.e., retweeted, information diffuses beyond adjacent users to the rest of the network
(Kwak et al., 2010). Computer scientists, Kwak, Lee, Park, & Moon (2010), find that a retweet
reaches an average of 1000 users no matter the number of followers of the first user. They find
that once retweeted, a tweet gets retweeted almost instantly on subsequent hops. Thus, if there is
an increase in the volume of retweets (i.e., higher number of endorsements) about a brand, more
users are likely to reason that the brand is of superior quality, and be influenced to buy the brand.
In summary, two metrics capture related dimensions of the social spillover effect: volume
captures the follower’s awareness to a user’s tweet while retweet captures a follower’s
endorsement of the original tweet.
Information effect. The “information effect” occurs when a follower’s buying behavior
is influenced by the content of a user’s tweet. In this context, the follower receives the message,
processes it, and makes a decision based on it. A positive tweet about a brand from a user could
convey positive information about the brand and lead to the conclusion that the brand is of
superior quality and worth buying. On the other hand, a negative tweet would convey negative
information about the brand and discourage buying. Thus, followers of that user may be
persuaded to buy the brand if there are more positive than negative tweets and dissuaded to buy
the brand if there are more negative than positive tweets. Responses of such followers would
translate into sales, earnings, and ultimately stock prices.
Thus the volume of tweets, retweets, and valence of tweets may be indicative of the
effects of Twitter on stock returns through its impact on the future sales and earnings of the firm.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 20
In this path, the relationship between Twitter metrics and stock returns would reveal itself
gradually over time, depending on how long it takes for the impact of Twitter to affect sales,
earnings, and stock returns. In practice, depending on the reporting of sales and earnings in a
market, this indirect effect could take weeks or months to be noticed and measured.
Direct route. The direct route is due to investor’s direct reactions to tweets. This reaction
occurs because investors observe the tweets, retweets, or valence and infer that they will affect
stock prices over time in the manner described above. In other words, investors use these metrics
as a signal about the brand’s quality, which they cannot directly observe. This pathway could
affect stock returns relatively quickly, perhaps in days.
But why would investors even consider the information in tweets?
Prior research has shown that investors of a firm seek extra information from sources
other than the firm. Investors react to information from journalists (Mitchell & Mulherin, 1994;
Tetlock, 2007), experts (Tellis & Johnson, 2007), and consumers (Tirunillai & Tellis, 2012). A
primary reason for this information search lies in the information asymmetry between the firm
and investors (see Healy & Palepu, 2001). Though firms reduce the asymmetry by providing
reports, these are either sporadic (e.g., press releases) or infrequent (e.g., sales reports). Thus,
investors continuously look for any novel information about the firm. Investors could consult
media stories, analyst reports, or expert reviews. However, I argue that the low frequency of such
reports (at most monthly) only whets the appetite of investors. Indeed, a recent paper by
Tirunillai and Tellis (2012) finds that online consumer reviews are a leading indicator of stock
returns. These reviews are available at a high frequency (daily). So, can tweets contain useful
information?
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 21
Tweets may possess valuable information to investors for the following reasons. First, on
aggregate, tweets are written at a much higher frequency than consumer reviews or forum
discussions. Every second 2,893 tweets are sent through Twitter’s servers (Bennett, 2011). This
high rate can be attributed to the number of active Twitter users. In October 2011, Compete.com
estimated that Twitter has 37 million unique users.
2
The sheer volume of tweets from users, each
with a unique perspective, can provide new information about brands. Second, the simplicity of
the user interface, short word limit, and the option of tweeting through mobile devices (40%
users’ tweets come through mobile devices) enable users to react and tweet instantaneously to
life events (Arthur, 2011). Such conversations may not be available in other online media. Third,
Twitter facilitates the phenomena of users sharing links to varied content on the internet (e.g.,
videos, blog posts, news articles, music streams). Rao (2010) estimates that 25% of all tweets
contain links. Such links increase the likelihood that tweets contain useful information. For these
reasons, investors may conclude that information on tweets represent a good source of the
“wisdom of crowds” (Surowiecki, 2004) about the latent quality of the brand. Thus, investors
may process the number of tweets, retweets, and valence as a signal of the brand’s quality and
use this inference for their trading decisions.
Investors could leverage several publicly available tools (e.g., Search.Twitter.com,
TweetBeep.com) or private tools (e.g., Radian6, Visible Technologies) to monitor what people
are saying about brands on Twitter. In fact, evidence is mounting that investors are consulting
Twitter in making their investment decisions. For example, companies like Bloomberg and
Thomson Reuters track tweets about firms to assist their Wall Street customers (Bowley, 2010).
2
http://siteanalytics.compete.com/twitter.com/, Retrieved November 30, 2011
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 22
Method
This section describes the research design, data collection, and measures.
Research Design
The next subsections explain the sampling of firms and time.
Sampling of firms. I select firms for this study on several criteria to ensure the validity
and reliability of the study. First, the brand names in the study do not overlap with commonly
used words. Thus, I do not select brands with names such as "Target" or "GAP," which have
common noun equivalents. Second, the firms (which own the brands) are listed on the U.S. stock
exchanges (NASDAQ/NYSE/AMEX). This criterion is because my focal dependent variable is
the stock market returns. Third, the brand did not undergo any rebranding exercise. Fourth, the
brands have a sizeable number of tweets at a daily level across the time period of the study.
Brands from categories such as digital products, high tech products, and e-commerce satisfy this
criterion. Fifth, the brand constitutes a major portion of the firm’s sales. This criterion ensures
that tweets about a brand would be a strong indicator about the firm’s performance. This
criterion only pertains to brand names which are different from the firm’s name. Based on these
criteria, I select the following brands for my study: BestBuy, Canon, Dell, iPhone, Kindle,
PlayStation, Priceline, Sandisk, TiVo, and Xbox.
Sampling of time. I focus on the period from October 1st 2009 to March 30th for two
reasons. First, prior to October 2009, use of Twitter was fairly low around 5 million users. This
low usage leads to sparseness in the number of tweets per day for some brands, for the period
prior to October 1st 2009. I estimate my models using daily data. Higher levels of aggregation
such as weekly or monthly may lead to biased estimates (Tellis & Franses, 2006). Second, in
order to make the data handling manageable, I focus on a period of 6 months. Even in this
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 23
limited time, the 10 brands have around 9 million tweets, which I text-mine and classify as
positive, negative, and neutral using natural language processing techniques.
Data Collection
I obtain the tweets for the brands from a corpus of tweets scraped by a third party data
provider. The full scrape consists of 35 million users, 500 million tweets, and 1 billion
relationships between users. This scraping is a random scrape containing 20% of the universe of
tweets in between the period March 2006 and March 2010. Any tweet that contains the brand
names: Xbox, PlayStation, iPhone, Kindle, TiVo, Canon, Dell, Sandisk, Priceline, Best Buy are
provided to me from the third party data provider. Due to the sheer volume of tweets, it is not
efficient to collect or process the data manually. Therefore, I resort to text mining and natural
language processing techniques for the creation of Twitter metrics (e.g., Tirunillai & Tellis,
2012; Archak, Ghose, & Ipeirotis, 2011).
To arrive at the valence of the tweet, I use the support vector machine classification
algorithm. I next explain the steps for creating the valence.
Each tweet is a row which contains the text of the tweet and the timestamp of when the
tweet was written. Table 1 provides the outline for the valence extraction.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 24
Table 1
Steps for Creation of Twitter Metrics
Data processing step Description of step
Creation of dictionary Create a dictionary, which forms the basis for the
valence classification. I create the dictionary using a
combination of tweets and various dictionaries. I use a
subset of tweets from my corpus (15,000 tweets) and
dictionaries such as Urban dictionary, Harvard’s General
Inquirer, Twictionary, Roget’s Thesaurus, and Miriam-
Webster
Formation of training sample Pre-process the 15,000 tweets into positive, negative and
neutral. Two human coders manually classify the tweets
intro positive, negative, and neutral
Removal of urls and user-ids Remove universal resource locators (URLs) and user-ids
in the tweets
Conversion of emoticons Convert the emoticons into positive and negative
meanings (e.g. as positive, as negative). I consult
various online sources to convert the various types of
emoticons
Removal of punctuation and numeric
characters
Remove punctuation and numeric characters except
exclamation and question marks. The exclamation and
question marks were replaced by placeholders “EXM”
and “QSM” respectively
Tokenization of tweets Tokenize tweets to individual words or phrases
Removal of stop-words Remove stop-words (e.g. I, or)
Stemming of words Stem words (convert to base form: e.g., love, loved,
loving, etc. stemmed to “love”)
Classification of valence Classify the tweets using Support Vector Machine
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 25
For the final step in Table 1, I employ the Support Vector Machine classification
algorithm. This algorithm is a semi-parametric classification technique. It has been widely used
in classification problems where the predictive validity is very important (Cui & Curry, 2005).
Joachims 2002 has shown that it ensures a high reliability in text classification. I adopt this
technique to classify the tweets as positive, negative and neutral. I use the dictionary and training
sample explained in Table 1 for the classification. Given a training set of instance-label pairs
) , (
i i
y x where
i
x is the instance and
i
y is the valence category (positive, negative and neutral);
l ,......... 2 , 1 i = where
n
i
R ε x and ε y , } 1 , 1 {
l
the support vector machines (SVM) (Boser, Guyon,
& Vapnik, 1992; Cortes & Vapnik, 1995) require the solution of the following optimization
problem:
. 0 ξ
, ξ 1 ) b ) x ( φ w ( y to subject
ξ C w w
2
1
min
i
i i
T
i
l
1 i
i
T
ξ b, w,
≥
- ≥ +
∑
+
=
(1)
Here training vectors
i
x are mapped into a higher (maybe infinite) dimensional space by
the function φ . SVM finds a linear separating hyperplane with the maximal margin in this higher
dimensional space. C > 0 is the penalty parameter of the error term. Furthermore,
) x ( φ ) x ( φ ) x , x ( K
j
T
i j i
= is called the kernel function. I use the libsvm package via Matlab for
the classification (Chang & Lin 2001). The details of the classification accuracy are in Appendix
A.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 26
Measures
This section explains the Twitter metrics, measures of stock returns, and the control
variables.
Measures of Twitter metrics. I use three Twitter metrics for my empirical investigation.
Two metrics are related to brand awareness (Hoffman & Fodor, 2010). Volume is the number of
tweets about the brand in a day and Positive to Negative ratio is the ratio of the number of
positive tweets to the number of negative tweets about the brand in a day. The Positive to
Negative ratio is created using the Support Vector Machine algorithm explained in the data
collection section. I find that across the 10 brands, 29% of the tweets are classified as positive,
10% are classified as negative, and 61% are classified as neutral. This number of neutral tweets
compares well to a number of academic and commercial work that have classified the valence in
Twitter into positive, negative, and neutral. For example, Kouloumpis, Wilson, & Moore (2011)
find 57% neutral tweets in their Hash-tagged tweets and iSieve Corporation find 57% neutral
tweets in their manually annotated dataset.
3
Moreover, note that when the training dataset is
manually classified, I find 52% neutral, 32% positive, and 16% negative tweets. The third
metric is a word-of-mouth measure (Hoffman & Fodor, 2010): Retweet is the number of tweets
retweeted about the brand in a day. My dataset does not explicitly label retweets. Thus, I
preprocess the data and detect retweets using patterns such as RT@[usename],
retweet@[usename], and retweeting@[usename]. I apply the patterns in a case-insensitive
manner.
Measure of stock returns. The event here is change in Tweets. However, due to the rich
data available, I can track these changes daily. Thus, I use a daily event study to measure the
3
www.i-sieve.com
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 27
effect of information in Twitter on stock returns (e.g., Tirunillai & Tellis, 2012). I first compute
the daily returns, as the buy and hold returns - emulating an investor purchasing the stock and
holding it for a day, then selling it the following day, accounting for the dividends, if any. I then
calculate the abnormal stock returns using the Carhart Four Factor model, which combines the
three factor approach of asset pricing (Fama & French, 1993) and the momentum factor (Carhart,
1997).
I use the CRSP (Center for Research in Security Prices) database to obtain the data for
stock prices for the 10 firms in the sample. I obtain the Fama-French factors and the momentum
factor from Kenneth French’s website.
4
it t i t i t i t rf, mt i i t rf, it
ε MOM m HML h SMB s ) R (R β α R R + + + + − + = − (2)
Here i stands for firm, t stands for time,
it
R denotes the returns for firm i on day t,
t rf,
R is
the risk-free rate of return (thirty day treasury bill), R
mt
stands for returns on corresponding daily
equally-weighted market index on day t, SMB denotes the returns on a portfolio of small stocks
minus returns on large stocks, HML stands for returns on a portfolio of stocks with high book-to-
market ratio minus the returns on a portfolio of stocks with low book-to-market ratio and MOM
stands for Carhart’s Price-Momentum Factor that captures the one-year momentum. The level
residual from the regression in equation (2) is the measure of “stock returns” that I use in this
chapter. This measure has eliminated the part of stock returns which could be explained by the
four factors (the SMB, HML, MKT, and MOM). I control for the serial correlation of the residuals
using the conditional Fama French approach (Luo, 2009):
(3)
4
http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/
it t it
γ ρε ε + =
−1
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 28
For each brand, I then match the “stock returns” with the Twitter metrics on a daily basis
to form the time series data for the vector autoregressive models.
Measures of control variables. I describe the measures of the control variables that
could possibly influence the results: analysts’ forecasts, advertising, and key firm developments.
Analysts’ forecasts. I measure analysts’ forecasts by the median of the analysts’
consensus forecast of the earnings. I obtain the data from the Institutional Brokers' Estimate
System (I/B/E/S) database.
Advertising. I measure a brand’s advertising by the daily dollar spend for the brand in
television stations in the United States of America. I obtain the data from the Kantar Stradegy
database.
Key firm developments. I measure the firm’s key developments by counting and
aggregating all key firm events. I aggregate more than 100 such events. Some of the events
include announcement of earnings and dividends, new product announcements, product changes,
innovations, mergers, client contracts, strategic alliances, lawsuits, SEC inquiries, expansions,
reorganizations, corporate governance changes, transaction milestones and rumors, changes in
key executives and labor relations. I obtain the data from Capital IQ’s key developments
database. Capital IQ has a Key Developments feature.
This feature provides categorized news and corporate event data. Experienced data
analysts from Capital IQ “continuously monitor, aggregate and tag information from over 20,000
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 29
news sources in addition to regulatory filings, transcripts, investor presentations and company
websites”. The Key Developments feature tracks over 100 company event types.
5
Model
This section explains the Vector Auto-Regressive modeling framework, which I use to
estimate the relationship between tweets and stock returns (Dekimpe & Hanssens, 1995, 1999). I
describe the test for endogeneity using Granger Causality test, test for unit root and
cointegration, the Vector Auto-Regression with Exogenous variables (VARX), the Generalized
Impulse Response Function and the Generalized Forecast Error Variance Decomposition.
Test for Endogeneity Using Granger Causality
Stock returns could affect Tweets when consumers tweet about unusual performance in
stock prices and up or down movements in stock prices. Thus, I test for endogeneity among the
variables using the Granger Causality test. The Granger Causality examines if there is a temporal
causality between two variables (Granger, 1969). I perform an array of Granger causality tests on
each pair of key variables. I test if tweets “Granger cause” stock returns, i.e., do any of the
Twitter metrics predict stock returns even after controlling for the lag of stock returns. I run the
Granger Causality test separately for each of the three Twitter metrics with stock returns. My
null hypothesis is that the Twitter metric does not “Granger cause” stock returns. I next test if
stock returns “Granger cause” any of the Twitter metrics. My null hypothesis is that stock returns
does not “Granger cause” the Twitter metric. If I find that any of the Twitter metrics “Granger
cause” stock returns and stock returns “Granger causes” any of the Twitter metrics, that would
imply that these variables are endogenous. I would need to capture these endogenous
5
Event types including SEC inquiries, expansions, reorganizations, client announcements, corporate governance,
dividends, earnings, transaction milestones and rumors, executive/board changes, litigations, labor relations, product
announcements, investor relations, strategic alliances, mergers and acquisitions, etc.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 30
relationships and complex interactions in a full dynamic system of multiple equations. A wrong
choice for the number of lags may erroneously conclude the absence of “Granger causality”
(e.g., Hanssens, 1980). Thus, I run the Granger Causality tests up to 25 days and report the
results of the lag that has the highest significance. The focus is not on whether X “Granger
causes” Y at a specific lag but whether one can rule out that X “Granger causes” Y at any lag
(Trusov, Bucklin, & Pauwels, 2009). Specifically I estimate the following model for each pair of
variables (e.g., Y
t
represents the stock returns and X
t
represents any one of the Twitter metrics)
t
p
l
l t l
p
l
l t l t
X Y Y ε λ β α + + + =
∑ ∑
=
−
=
−
1 1
(4)
Equation (4) estimated with all free parameters provides the unrestricted version. The
restricted version constrains all the coefficients related to X
t
to 0 i.e.,
l
λ
=0. I implement the
Granger Causality test using the asymptotic results of these constraints with the test statistic
distributed as a chi-square (p) variable. The null hypothesis states that all coefficients are 0
which implies no evidence of Granger causality.
Test for Unit Root and Cointegration
The Unit Root test examines if the variables in my dataset are evolving or are stationary.
I use the Augmented Dickey-Fuller (ADF) test recommended by Enders (1995). This test is
carried out under the null hypothesis that there exists a unit root. If the variables are integrated in
their levels, I run the Johansen’s test for cointegration (Johansen, 1995). One must test for the
possible presence of cointegration when one examines the relationship between two integrated
variables (Granger & Newbold, 1974). The Johansen test is a procedure, which examines
whether, if variables evolve, they evolve together. If I find evidence that the variables evolve
together, this finding implies that the variables do not return to their mean levels after some
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 31
period. Based on the ADF and cointegration test, I choose the proper specification for the
endogenous variables that enter the VARX model. If I find that the variables are not evolving
and are not cointegrated, I run the VARX model using levels of the variables. If I find that the
variables are evolving and are not cointegrated, I run the VARX model using first differences of
the evolving variables. If I find that the variables are evolving and are cointegrated, I use error
correction models wherein I adjust with an error term.
Vector Auto-Regressive Model with Exogenous Variables (VARX)
I estimate the relationships among the Twitter metrics and stock returns using the Vector
Auto-Regressive model with Exogenous variables (VARX). If I find that the Twitter metrics and
stock returns are endogenous, this implies that these metrics are explained by both past variables
of themselves (autoregressive carryover effects) and past variables of each other (cross effects
from Twitter metrics to Stock returns or vice versa).
The VARX model enables me to handle such endogeneity and other biases such as
omitted variables and auto-correlations. Below is the specification that I estimate for each brand.
t
t
t
t
RTW
PNR
VOL
STR
=
RTW
PNR
VOL
STR
α
α
α
α
+
t
t
t
t
RTW
PNR
VOL
STR
*
*
*
*
δ
δ
δ
δ
+
∑
=
L
l
l l l l
l l l l
l l l l
l l l l
1
4 , 4 3 , 4 2 , 4 1 , 4
4 , 3 3 , 3 2 , 3 1 , 3
4 , 2 3 , 2 2 , 2 1 , 2
4 , 1 3 , 1 2 , 1 1 , 1
β β β β
β β β β
β β β β
β β β β
−
−
−
−
l t
l t
l t
l t
RTW
PNR
VOL
STR
+ (5)
EXO * θ +
t RTW
t
PNR
t VOL
t STR
,
,
,
,
ε
ε
ε
ε
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 32
where STR, VOL, PNR and RTW denote stock returns, volume, positive to negative ratio,
and retweet respectively. t is the deterministic-trend and the control variables are denoted by
EXO (1 holiday dummy, 1 key developments variable, 1 advertising spending variable, 1 analyst
forecast variable). The holidays include the three days around Halloween, the three weeks
around Thanksgiving, the three weeks around Christmas, the week after New Year, three days
around Martin Luther King day, and three days around Valentine’s day. Σ is the covariance
matrix of the residuals. I estimate the model using the Newey-West standard errors (Newey &
West, 1987). These standard errors are Heteroskedasticity and Auto-Correlation consistent.
Generalized Impulse Response Function (GIRF)
Based on the parameters of the VARX model, I model the dynamics of the system using
the Generalized Impulse Response Function (Pesaran & Shin, 1998). The Generalized Impulse
Response function is invariant to the temporal ordering of the variables in the VARX system.
This function serves two purposes for my analysis. First, they allow me to quantify the net result
of a “shock” to each of the Twitter metrics on stock returns relative to their baselines. Second,
they allow me to measure the lag number of days before the effect of Twitter metrics on stock
returns reaches its peak (i.e., wearin) and the number of days when the effect reaches its
asymptote. To evaluate the accuracy of my GIRF estimates, I compute the confidence intervals
using the Monte Carlo simulation approach with 250 runs (Benkwitz, Lütkepohl, & Wolters,
2001).
Generalized Forecast Error Variance Decomposition (GFEVD)
I assess the relative contribution of each of the Twitter metrics on stock returns using the
Generalized Forecast Error Variance Decomposition technique (Koop et al. 1996; Pesaran &
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 33
Shin 1998). Analogous to a ‘dynamic R2’, the Generalized Forecast Error Variance
Decomposition calculates 100% of the variation in a response variable that can be attributed to
contemporaneous and past values of all endogenous variables. The GFEVD is invariant to the
temporal ordering of the variables in the VARX system. The decomposition determines to what
extent the Twitter metrics contribute to the deviation in stock returns from its baseline
expectations. The relative importance of the endogenous variables in the VARX is established
based on GFEVD values at 15 days, which reduces sensitivity to short-term fluctuations. To
establish the statistical significance of the GFEVD estimates, I compute the confidence intervals
using Monte Carlo simulation approach with 250 runs (Benkwitz et al., 2001).
Results
This section presents descriptive analysis, test of endogeneity using Granger Causality
test, test for unit root, effects of Twitter metrics on returns, relative importance of Twitter
metrics, and robustness analysis.
Descriptive Analysis
Figure 2 depicts the trend of the three Twitter metrics in my study’s timeframe for all
brands (i.e. Oct 1
st
2009 to Mar 30
th
2010).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 34
Figure 2. Time series plot of Twitter metrics for all brands
I create the line graphs by summing up the volume of tweets and retweets and averaging
the positive to negative ratio for all the 10 brands in my sample. Note the substantial variation in
all the three metrics across time. The mean positive to negative ratio of tweets increases on
Christmas Eve reaching a peak on New Year’s Eve. This result indicates that people tweet more
positive messages about brands during the holidays. The line graphs by individual brands are in
Appendix B (Figure B1 to B10). There is substantial variation and a number of spikes in Twitter
metrics across the brands. However, I do observe that the positive to negative ratio of tweets
increases on Christmas eve for Best Buy, Dell, iPhone, Kindle, Sandisk, Priceline, and Xbox.
Similarly, I find an increase in the positive to negative ratio of tweets for Canon, Dell, iPhone,
Kindle, PlayStation, Priceline, Sandisk, TiVO, and Xbox on Thanksgiving Day and the day after.
These results may be because people are in a positive mood during holidays, which one would
expect intuitively. Thus, these results partly validate partly my classification technique.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 35
I find that users tweet equally on weekdays and weekends for four brands: Best Buy,
Canon, Sandisk, and Xbox (See Table 2). Column 5 in Table 2 provides the ratio of the weekday
to weekend tweets. Users tweet more on weekdays than on weekends for six brands: Dell,
iPhone, Kindle, PlayStation, Priceline, and TiVo. For example for iPhone, users post 34,898
tweets on weekdays while users post 27,742 tweets on weekends.
Table 2
Volume of Tweets on Weekends and Weekdays
Brands Weekday Weekend All days
Weekday to
weekend ratio
Best Buy 159 160 160 0.995
Canon 87 83 86 1.052
Dell 428 350 406 1.223
iPhone 34898 27742 32842 1.258
Kindle 2612 1907 2409 1.37
PlayStation 1808 1271 1654 1.422
Priceline 72 50 66 1.464
Sandisk 138 120 133 1.154
TiVo 441 324 407 1.361
Xbox 7102 7156 7118 0.993
Total 4775 3916 4528 1.219
Table 3 provides the ratios of weekday to weekend for retweets, positive tweets, negative
tweets, and neutral tweets. Users retweet equally on weekdays and weekends for three brands:
Best Buy, Canon, and Dell. Users post positive tweets equally on weekdays and weekends for
six brands: Best Buy, Canon, Dell, iPhone, TiVo, and Xbox. Users post negative tweets equally
on weekdays and weekends for three brands: Best Buy, Canon, and Xbox. These results indicate
that there is variation among brands in users’ tweets on weekdays and weekends.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 36
Table 3
Weekday to Weekend Ratio for Various Metrics
Brands Retweet
Positive
tweet
Neutral
tweet
Negative tweet
Best Buy 1.08 0.98 1.03 0.84
Canon 1.05 1.04 1.03 1.10
Dell 1.17 1.14 1.26 1.29
iPhone 1.37 1.17 1.31 1.27
Kindle 1.53 1.21 1.33 1.69
PlayStation 2.24 1.25 1.37 1.35
Priceline 2.27 1.50 1.47 1.38
Sandisk 1.91 1.41 1.07 1.57
TiVo 1.56 1.17 1.53 1.35
Xbox 1.51 0.92 0.98 1.00
Overall 1.40 1.13 1.24 1.24
I find that tweets capture firm related news well. Figure 3 depicts the relation between
Twitter metrics and stock returns when Amazon announced that it extended the battery life of
Kindle by 85 % and added a native PDF Reader.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 37
Figure 3. Time series plot of selected Twitter metrics and stock price for Amazon
The number of tweets, retweets, and positive tweets rise on the day of the announcement
and continue till the next day. Most importantly, the stock price rises on the next day. This figure
suggests that there exists a relationship between Twitter and stock returns. The subsequent
subsections systematically examine if there is such a relationship between Twitter metrics and
stock returns. The correlations among the key variables (Stock Returns, Volume of Tweets,
Retweets, and Positive to Negative Ratio) are reported in Appendix B (Table B1). The mean and
standard deviation of these key variables are reported in Appendix B (Table B2).
Granger Causality Tests To Check for Endogeneity
I find that each of the three Twitter metrics significantly “Granger causes” stock returns
at 25 lags (see Table 4).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 38
Table 4
Result of Granger Causality Test of Twitter Metrics on Stock Returns
Independent variable Stock returns
(p-values at 25 lags)
Volume of tweets
0.00
Volume of retweets
0.01
Positive to negative ratio
0.01
Note. All values are the average p-values of the Chi-Square statistic. Significant values are indicated in boldface.
The Granger Causality tests by brand are in Appendix C (Table C1).
I find that stock returns significantly “Granger causes” volume of tweets and retweets at
25 lags (see Table 5).
Table 5
Result of Granger Causality Test of Stock Returns on Twitter Metrics
Dependent Variable Stock returns
(p-values at 25 lags)
Volume of tweets
0.05
Volume of retweets
0.02
Positive to negative ratio
0.06
Note. All values are the average p-values of the Chi-Square statistic. Significant values are indicated in boldface.
The Granger Causality tests by brand are in the Appendix C (Table C2).
Because I reject the null hypothesis that stock returns does not “Granger cause” volume
of tweets and retweets, the variables are endogenous and must be analyzed as a full dynamic
system such as with a VARX model.
Test for Unit Root and Cointegration
I perform the Augmented Dickey Fuller (ADF) test to check if any of the Twitter metrics
and stock returns is evolving or stationary. I use the iterative procedure recommended by Enders
(2004) to determine if a time trend should be included in this test. For all Twitter metrics across
all brands, except for the volume and retweet metric for Dell, I can reject the null hypothesis of a
unit root at the 5% confidence level (see Table 6 for the summary).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 39
Table 6
Summary of Unit Root Tests of the Endogenous Variables
Twitter metric ADF test (critical
values)
Volume of tweets -13.99
Volume of retweets -6.51
Positive to negative ratio -9.25
Stock returns -8.71
Note. Significant values are indicated in boldface. All the values of ADF are significant at 5% levels. The critical
value at 5% level is -3.44. I use the iterative procedure recommended by Enders (2004, page 213-214) to determine
if a time trend should be included in the ADF test. The ADF tests by brand are in Appendix C (Table C3).
Since the volume and retweet metrics of Dell evolve in levels, I run cointegration tests
(Johansen, 1995) to check if the combination of Twitter metrics and stock returns show evidence
of cointegration over my study’s timeframe. I find no evidence of cointegration and thus use the
first difference of the volume and retweet metrics in the VARX model for Dell. The first
difference of the volume and retweet metric is stationary. For stock returns, I can safely reject the
null hypothesis of a unit root at the 5% level.
Dynamics of Twitter Metrics on Stock Returns
Because each of the Twitter metric significantly “Granger” causes stock returns, I
estimate the VARX model in Equation 5, where each of the three Twitter metrics and stock
returns are endogenous. I run the VARX model separately for each brand. The optimal lag order
is 1 for all the VARX models as per the (Schwartz’s) Bayesian Information Criteria. My results
are not affected by the presence of any residual correlation, non-normality of residuals, and
heteroskedasticity. Further, the VARX model is estimated with Heteroskedasticity and
Autocorrelation consistent estimator which accounts for any potential serial correlation. I use the
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 40
simulations of the Generalized Impulse Response Function to show the short-term and long-term
effect of Twitter metrics on stock returns (Pesaran & Shin, 1998; Dekimpe & Hanssens, 1999).
Utilizing the VARX parameter estimates, the generalized impulse response function tracks the
over-time impact of a unit shock (one standard deviation) in any one of the Twitter metrics on
stock returns. I define the short-term impact as the time period when the effect of Twitter metrics
on stock returns reaches its peak. I find that the effect of Twitter metrics reaches its peak in the
first time period (see Table 7).
Table 7
Duration of Impact of Twitter Metrics on Stock Returns
Twitter metric Wear-in
(in days)
Wear-Out
(in days)
Time to
asymptote
(in days)
Volume of tweets 1 6.6 7.6
Volume of retweets 1 6.5 7.5
Positive to negative
ratio
1 4.7 5.7
I define the duration of long-term or cumulative impact as the time period from the start
of the effect till the time the accumulated impact of Twitter metrics on stock returns reaches its
asymptote (see Tirunillai & Tellis, 2012 for a similar definition). For both the volume of tweets
and retweets, this duration is the 8
th
day while for positive to negative ratio it is the 6
th
day. I
expect this time to decline in the future due to investor’s increasing their awareness and
processing of Tweets. I follow standard practice in finance and accounting literatures and express
the short-term and cumulative impact in basis points. One basis point is one-hundredth of a
percentage.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 41
Table 8 presents the results of the Twitter metrics on stock returns, averaged across the
10 brands.
Table 8
Impact of Twitter Metrics on Stock Returns
Twitter Metric Short-Term
(in basis
points)
Cumulative
(in basis
points)
V olume of Tweets 12.3 14.9
V olume of Retweets 16.8 15.8
Positive to Negative
Ratio
8.5 7.0
Note. Significant values are indicated in boldface. Values presented are median values. 1 basis point = 1 hundredth
of a percentage. Results by individual brands are in Appendix C (Table C4).
Among the Twitter metrics, retweets has both the highest short-term (16.8 basis points,
p<.01) and accumulated (15.8 basis points, p<.01) impact on stock returns. Volume of tweets
also has a significant short-term (12.3 basis points, p<.01) and accumulated impact (14.9 basis
points, p<.01) on stock returns. The positive to negative ratio has no significant short-term (8.5
basis points, n.s.) and accumulated impact on stock returns (7.0 basis points, n.s.). These results
are after controlling for other variables such as earnings, new product announcements,
advertising, and analyst forecasts.
Though these effects appear small in basis points, they have a big impact when translated
in dollar value. All else remaining the same, a unit shock to the volume of retweets can raise
about $22.2 million to the average market capitalization in the short term while a unit shock to
the volume of tweets can raise about $16.3 million to the average market capitalization in the
short term. Note that these dollar values exclude the cost of trading. There is a possibility that
transaction costs may be high enough so that the gains made from trading are lost.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 42
Next, I compare the effects from Twitter metrics on stock returns and from stock returns
on Twitter metrics. Since, I find a presence of reverse causality, i.e., stock returns “Granger
cause” Twitter metrics, I evaluate whether the effect from Twitter to stock returns is stronger
than the reverse. I use the arc elasticity measure to compare the effects (Trusov et al., 2009
6
). I
do not test for positive to negative ratio since its effect on stock returns is not significant. I find
short-term elasticities of 3.3 and 1.4 on stock returns of volume of tweets and retweets,
respectively. On the other hand, I find short-term elasticities of stock returns of 0.003 on volume
of tweets and 0.003 on retweets. I find similar results for the cumulative elasticities. These
results indicate that the effect of Twitter on stock returns is much stronger than the reverse.
Relative Importance of Twitter Metrics
The relative importance of each of the Twitter metrics is assessed using the Generalized
Forecast Error Variance Decomposition (GFEVD). I derive the GFEVD estimates from the
model estimated in VARX specification in equation 5. I establish the statistical significance of
the GFEVD estimates by running Monte Carlo simulations with 250 runs. I find that the volume
of retweets explains 1.8% of the variation in stock returns on the 10
th
day while volume of tweets
explains 1.3% of the variation in stock returns on the 10th day. By the 15th day, volume of
tweets explains 1.4% while volume of retweets continues to explain 1.8% of the variation in
stock returns. These results indicate that Retweet is the most important metric influencing stock
6
,
Y
X Y
X
arc
×
∆
=
σ
η where
arc
η = arc elasticity, Y ∆ = impulse response of response (dependent) variable,
X
σ =
std. deviation of shocked (independent) variable, X =Mean value of shocked variable, Y = mean value of response
variable (see Trusov et al., 2009).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 43
returns. I do not look at the relative importance of positive to negative ratio of tweets as it is not
significant.
Effect of Control Variables on Twitter Metrics
The impact of the volume of tweets and retweets on stock returns holds even after I
control for three variables: analysts’ forecasts, advertising, and key firm developments. I next
describe the effects of these three control variables on Twitter metrics. I report the estimated
elasticity values that I calculate from the VARX model in equation 5. I find that a 1% increase in
a firm’s key development significantly increases the volume of tweets by 2.4% (p<.05). This
result suggests that Twitter users tweet about the brand when the firm undergoes a key
development such as new product announcements, product changes, earnings announcements,
etc. However, neither retweets nor the positive to negative ratio of tweets are affected by the
firm’s key development. Advertising and analyst forecasts do not influence any of the Twitter
metrics.
7
Robustness Analysis
I carry out three tests to establish the robustness of the results.
Effects of other media variables. I next report the results of a VARX model, which
contains the following endogenous variables: volume of retweets, volume of citations of print,
volume of influential blogs, volume of consumer discussions, and stock returns. This model
evaluates if the effect of retweets holds after I incorporate the effects of other media, which too
can affect stock returns. Researchers in marketing and finance find that citations of print and
7
I also control for tweets from the brand’s PR dept. within Twitter (e.g., http://twitter.com/#!/BESTBUY) and
media coverage by publishers within Twitter (e.g., http://twitter.com/#!/Gizmodo). I obtain the data from the
Twitter website for the former and from Newstex for the latter. My findings remain the same after the inclusion of
these two exogenous variables.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 44
message boards can influence stock returns (Tetlock, 2007; Das & Chen, 2007). The definition
and sources of these other media variables are in Table 9.
Table 9
Description of Other Media Variables in Extended VARX Model
Endogenous variable Source Definition/Operationalization
Citations of print Factiva and Lexis-
Nexis (also used in
Tirunillai & Tellis,
2012)
I define citations of print as the number of articles in
print media in each day. I carry out a search for any
article that mentions the brand name in Lexis-Nexis
and Factiva in my sample. Across both the sources, I
search all newspapers or newswires except non-US
newspapers that mention the brand name on any given
day. I use Lexis-Nexis’ relevancy score feature to
ensure that I only select articles that are relevant and
not accidental mentions. I identify an article as relevant
if Lexis-Nexis gives a relevancy score of 60% or
above. Similarly, I use Factiva’s company tag feature,
which indicates if the results of a search are relevant to
ensure that I select only pertinent articles.
Influential blogs Newstex I define influential blogs as number of blogs written by
authoritative publishers in each day. Newstex’s
Authoritative Content feature enables me to select
blogs from news organizations and corporate blogs, as
well as respected independent experts and thought
leader blogs, which includes blogging sites such as
Gawker.com, Mashable.com, b5media.com, and
consumerist.com
Consumer discussions Google Groups (also
used in Stephen &
Galak, 2012)
I define consumer discussions as the number of posts
on internet community forums that mention the brand
name in each day. Google Groups supports discussion
groups, which include many Usenet newsgroups.
Similar to the Twitter metrics, I run the Augmented Dickey Fuller (ADF) test on the
volume of citations of print, influential blogs, and the consumer discussions about the brands. I
can safely reject the null hypothesis of unit root at the 5% confidence level for each of these
variables for all the brands (see Appendix C, Table C3).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 45
I run the VARX model separately for each brand. I find that my results are not affected
by the presence of any residual correlation, non-normality of residuals, and heteroskedasticity.
Moreover, the VARX model is estimated with Heteroskedasticity and Autocorrelation consistent
estimator, which accounts for any potential serial correlation. The optimal lag order is 1 for the
VARX models as per the BIC criterion. Table 10 presents the results averaged across the 10
brands.
Table 10
Impact of Retweet and Other Media on Stock Returns
Twitter metric Short-term (in
basis points)*
Cumulative
(in basis
points)*
Wear-in
(in days)**
Wear-out
(in days)**
Time to
asymptote
(in days)**
V olume of retweets 9.15 9.6 1.8 7.4 8.4
Volume of citations
of print
17.85 15.55 1.1 7.3 8.3
Volume of influential
blogs
11.85 11.25 1.4 7.1 8.1
Volume of consumer
discussions
0.15 1.85 1.6 6.6 7.6
Note. Significant values are indicated in boldface; *Median Values reported; **Average Values reported; 1 basis
point = 1 hundredth of a percentage; Results by individual brands are in Appendix C (Table C5).
The volume of retweets has a significant short-term (9.15 basis points, p<.05) and
accumulated impact on stock returns (9.6 basis points, p<.05). I find that citations of print has the
highest short-term (17.85 basis points, p<.05) and accumulated (15.55 basis points, p< .05)
impact on stock returns. Influential blogs has a significant short-term (11.85 basis points, p<.05)
and accumulated impact on stock returns (11.25 basis points, p<.05). The volume of consumer
discussions neither has a significant short-term (0.15 basis points, n.s.) nor a significant
accumulated impact on stock returns (1.85 basis points, n.s.). Thus, the significant short-term and
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 46
accumulated effect of retweets persists even after I control for other media. I find that the wear-
in time for citations of print is lesser than retweet and other media (see Table 10). The
accumulated effect of retweets, citations of print, influential blogs, and consumer discussions
reach an asymptote around the 8
th
day.
I next describe the effect of citations of print, influential blogs, and consumer discussions
on retweets. I find citations of print and influential blogs have a significant short-term impact on
retweets (see Table 11).
Table 11
Impact of Other Media on Retweet
Other media Short term
(in number of retweets)
Cumulative
(in number of
retweets)
Wear-in
(in days)
Citations of print 67.47 61.97 1
Influential blogs 106.61 139.28 1.6
Consumer discussions 7.23 1.74 1.6
Note. Significant values are indicated in boldface. *Median Values reported
Citations of print has a short-term impact of 67 (p<.05) while influential blogs has a
short-term impact of 107 (p<.05). This means that a shock on citations of print leads to an
increase of approximately 67 retweets while a shock on influential blogs leads to an increase of
approximately 107 retweets. Citations of print has the shortest wear-in time (1 day) in impacting
retweets while the effect of influential blogs on retweets reaches its peak in 1.6 days. These
results suggest that current events in the news are retweeted immediately while blogs have a
slight delay to get tweeted and retweeted. Consumer discussions do not have any significant
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 47
impact on retweets. The accumulated impact of citations of print and influential blogs on
retweets is 62 (p<.05) and 139 (p<.05) respectively.
Effects of other Twitter metrics. I analyze the effect of two other Twitter metrics on
stock returns besides the volume, retweets and positive to negative ratio of tweets. I include each
of these metrics one at a time as an endogenous variable along with the volume, retweet, positive
and negative ratio of tweets, and stock returns. First, I evaluate the Speed of Tweets metric.
Twitter’s simple interface enables users to quickly tweet through mobile devices and computers.
The speed at which users tweet, measures interest about an event or brand. For example, soccer
fans sent a Twitter record of 13,684 tweets per second during the 2012 semi-final of the
Champions League after the final goal in the match (Mogg, 2012). I create this measure by
calculating the average hourly rate of tweets in a day. Since the Speed of Tweets metric is highly
correlated with volume of tweets metric, I replace volume of tweets with the Speed of Tweets in
my model.
I find the results remain the same. Retweet retains its rank as the most important Twitter
metric. Retweet has both the highest short-term (16.3 basis points, p<.01) and accumulated (15.7
basis points, p<.01) impact on stock returns. Speed of Tweets has a significant short-term (12.0
basis points, p<.01) and accumulated impact (14.5 basis points, p<.01) on stock returns. The
positive to negative ratio has no significant short-term (8.3 basis points, n.s.) and no significant
accumulated impact on stock returns (7.6 basis points, n.s.).
Next, I assess the effect of the Positive to Negative ratio of Retweets metric. I create this
measure by using the Support Vector Machine algorithm explained in the data collection section.
I find the impact of this metric is not significant for both the short (3.9 basis points, n.s.) and
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 48
long term (9.2 basis points, n.s.). The impact of retweets remains highest followed by volume of
tweets and positive to negative ratio of tweets.
Portfolio analysis. I run a portfolio analysis to explore investment opportunities using
my results. I test if an investor can reap positive returns using the retweet metric for investing. I
use the calendar-time portfolio method for this purpose (e.g., Jaffe, 1974; Mitchell & Stafford,
2000; Sorescu, Shankar, & Kushwaha, 2007). In this method, I form a portfolio to include firms,
which have a high number of retweets. Specifically, I use the retweet rate, the ratio of the volume
of retweets to the volume of tweets, as my focal metric. Since each brand can have a different
retweet rate, I first normalize the data. I divide each brand’s daily retweet rate by its maximum
retweet rate. I include a firm in the portfolio if the normalized daily retweet rate for the brand is
greater than 50%. The portfolio’s composition changes daily as firms are added or deleted based
on the normalized daily retweet rate. After constructing the portfolio, I measure the abnormal
returns to the portfolio using the Fama-French and Carhart four-factor model specified in
equation 2.
As the number of firms in the portfolio changes each day, I use the weighted least squares
method to estimate equation 2. This method puts more weight to the days in which the portfolio
contains more firms (Sorescu et al., 2007). The estimate of α in equation 2, give the returns to
the portfolio beyond the effects predicted by the four factors. The returns of the retweet portfolio
are positive and significant ( α =0.24%, p<.05). This result corroborates the finding from the
VARX models.
Firm tweets as endogenous. I run VARX models including firm’s tweets as an
endogenous variable. My effects may be biased if I don’t include firm’s response in Twitter.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 49
Firm tweets are endogenous because firms may respond to an increase in negative tweets and
consumers may respond positively due to the firm’s response in Twitter.
I find the results remain the same. Retweet retains its rank as the most important Twitter
metric. Retweet has both the highest short-term (16.37 basis points, p<.01) and accumulated
(15.6 basis points, p<.01) impact on stock returns. Volume of tweets has a significant short-term
(12.44 basis points, p<.01) and accumulated impact (13.51 basis points, p<.01) on stock returns.
The positive to negative ratio has no significant short-term (9.49 basis points, n.s.) and
accumulated impact on stock returns (7.14 basis points, n.s.).
Discussion
Twitter is one of the most popular social media sites. This study seeks to find out whether
conversations in Twitter about brands affect the stock market returns of the firms that own those
brands. If so, which metrics best captures this effect and what are the dynamics of this
relationship. This section summarizes the findings, discusses some key issues, suggests
implications, and lists the limitations.
Summary of Findings
The key findings of the study are the following:
• Twitter metrics have a significant positive relationship on stock returns. The impact
of Twitter metrics on stock returns exists after I control for key firm developments,
analysts’ forecasts, and advertising. The significant short-term and cumulative effect
of retweets persists even after I incorporate effects of citations of print, influential
blogs, and consumer discussions.
• The effect of volume of tweets and retweets on stock returns wears-in immediately
reaching an asymptote around the 8th day.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 50
• The effect of Twitter metrics on stock returns is much stronger on stock returns than
the reverse.
• Of all Twitter metrics, retweet has the strongest relationship with stock returns.
Retweets has the highest short-term and cumulative impact on stock returns (in basis
points) and explain stock returns the most (using the GFEVD estimates).
• The volume of citations of print and influential blogs has a significant short-term and
cumulative impact on retweets. However, citations of print are retweeted at a faster
rate than influential blogs.
Discussion of Key Issues
This section addresses 4 key questions emerging from the results: Why does the effect of
Twitter metrics wear-in immediately? Why is retweet the most important metric of Twitter?
What type of content does Twitter capture? Why is valence of tweets not important?
Why does the effect of Twitter metrics wear-in immediately? I find the effect of
Twitter metrics on stock returns wears-in immediately (i.e., within a day). This result suggests
that the direct route, and not the indirect, is at play.
The indirect route suggests that tweets affect sales and later sales affects earnings, profits,
and stock prices. Thus tweets and stock prices would be positively related. However, this path
(sales to profits to stock price) would take weeks or months to materialize, because sales are only
available at the weekly or monthly level for most categories.
On the other hand, in the direct path, investors respond directly and immediately to the
values of tweets and retweets, using them as a signal about the brand’s underlying quality.
Investors may infer that the increase in the volume of tweets and retweets is due to consumers
spreading awareness about the brand. Investors may envisage that over time, this increase in
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 51
awareness will lead to an increase in sales (and consequently profits as hypothesized in the
indirect path). This thinking could lead them to immediately make trades using the information
obtained from Twitter.
Indeed, investors are increasingly enabled by social media monitoring firms to collect
and extract information from social media. Such information extraction, which is generally time
consuming and requires sophisticated methods, by social media monitoring firms decreases an
individual trader’s costs of processing digital conversations and may allow them to gain from
same day trades. For example, Bloomberg continuously monitors tweets and alerts its customers
if a lot of people are suddenly sending tweets about any company (Bowley, 2010). Moreover,
even institutional traders are increasingly using Twitter in their investment decisions. Recently,
Derwent Capital Markets, a London based hedge fund, launched a £25 million hedge fund that
will use tweets to help guide its trading strategy (Telegraph, 2011).
Why is Retweet the most important metric of Twitter? I find that retweets are more
strongly related to stock returns than the volume of tweets and valence. A number of social
media experts have suggested that retweet is the most important metric in Twitter (e.g., Israel,
2008; Owyang, 2008). Indeed, retweets can be a stronger signal than tweets about a brand’s
quality due to the following reasons. First, a retweet is more salient than a non-retweeted tweet
because it stands out from ordinary tweets. Second, a retweet diffuses beyond the original user’s
network to the rest of the network (Kwan et al., 2010). Thus, a retweet reaches many more users
than a non-retweeted tweet and increases awareness of the brand. Third, a retweet is more
credible because it is prescreened by the follower. This aspect of retweets may make the
follower’s recipients more likely to read the content as it is relayed by someone familiar
(Malhotra, Kubowicz Malhotra, & See, 2011). Fourth, a retweet enables users to access varied
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 52
content in the World Wide Web much more than a non-retweeted tweet (Boyd et al., 2010).
Boyd, Golder, & Lotan (2010) find that 52% of retweets contain a URL.
8
Investors may reason
that retweet is the most important metric based on a combination of social media expert opinions
and their own judgment.
What type of content does Twitter capture? I find that key developments of the firm
such as new product announcements, product changes, and earnings announcements significantly
increase the volume of tweets. This result suggests that important events of a focal firm stimulate
tweets. For example, I find huge spikes in volume of tweets about the iPhone on 27
th
Jan, 2010.
On that day, Apple unveiled the iPad and released the Software Development Kit for iPhone OS
version 3.2. In my robustness analysis, I find that retweets significantly react to citations of print
and influential blogs on any third party discussion of the brand.
9
For example, I find an increase
in the number of retweets about the iPhone on 9th Dec, 2009. On that day, Oppenheimer and
Piper Jaffray analysts speculated about the iPad’s launch and iPhone’s potential deal with
Verizon Wireless respectively. This speculation was widely covered by media and influential
blogs.
However, I do not find any evidence that consumer discussions about brands in forum
sites significantly affect brand conversations in Twitter. This result suggests that conversations in
Twitter are related to topical issues covered in news media and influential blogs and are not
related to conversations in forum sites. The control of content in a forum site is more
8
I add retweets with links as another endogenous variable in my model which includes volume, retweets, positive to
negative ratio, and stock returns as endogenous variables. The Akaike Information Criteria for this 5-variable model
is larger than the model 4-variable model, which excludes retweets with links. I also do not find any significant
difference in the log likelihood values of a 4-variable model which includes retweets and another 4- variable model,
which replaces retweets with retweets with links. This result suggests that retweets with links does not provide extra
information to explain stock returns.
9
Note that in the robustness analysis, I drop the key developments as an exogenous variable because media citations
can include coverage of a firm’s key development events such as acquisitions, new product announcements, etc.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 53
decentralized (i.e., group-based) than a Twitter user’s account (i.e., single-user). The access to
content in a forum site is generally inaccessible to outside users and the content is more
unfocussed, unpredictable, and varied than a Twitter user’s content. These differences could lead
to the un-relatedness of the conversations between forum sites and Twitter. Moreover, I do not
find any evidence that offline advertising significantly influences conversations in Twitter.
Why is valence of tweets not important? I find that positive to negative ratio of tweets
does not significantly affect stock market performance of firms. There are two possible reasons
for this finding. First, investors might find it easier to monitor volume and retweets than valence.
Second, the frequency of neutral tweets in Twitter is quite high. These neutral tweets either
contain no positive or negative word or contain only URLs. Boyd et al. (2010) find that 22% of
tweets include a URL.
Implications
This study has the following implications.
First, marketing managers should definitely monitor conversations in Twitter and include
it as part of their marketing research. I find that conversations in Twitter are not noise and
contain information about firm performance. This information is beyond other sources of
information such as citations of print, influential blogs, forums, and analyst forecasts.
Second, managers can obtain the precise impact of tweets on stock returns using the
impulse response function derived from the Vector AutoRegressive model.
Third, my results suggest that marketing managers should monitor the retweet metric.
Monitoring a metric like retweet is much easier than sentiment measures such as positive to
negative ratio. In general, firms have to develop natural language processing capabilities in-
house or seek help from consultants to monitor sentiments in social media. However, examining
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 54
the number of retweets only involves counting the number of times a tweet is retweeted. Tools
such as Retweetist, Retweetradar, and Retweetrank can easily enable firms to know such
statistics.
Fourth, managers should find out what drives consumers to retweet. The model suggests
that managers can increase retweets by announcing firm developments such as product launches
and promptly relaying such announcements to journalists and influential bloggers.
Fifth, investors can use the information in Twitter in their trading strategies. I find that
information in Twitter affects stock returns even after controlling for analyst forecasts, key
developments of the firm, citations of print, influential blogs and forum discussions.
Limitations and Further Research
This study has several limitations that can be basis of future research. First, I had to
restrict my sample to 10 brands and 6 months due to the sheer volume of tweets. It would be
worthwhile to investigate the generalizability of the results. Second, it would be useful to explore
how tweets affect sales and how sales affects returns. I am unable to obtain sales data at the daily
level. Third, I could not obtain review data for the brands from sites such as Amazon.com and
epinions.com and test whether tweets affect returns controlling for reviews (e.g., Tirunillai &
Tellis, 2012). Such an enterprise is time consuming and requires considerable effort. However, I
collect other media data such as blogs and consumer discussions. Such data are likely to be
correlated with reviews.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 55
Chapter 3: Halo Effects in Social Media: Do Recalls of One Brand Hurt or Help Rival Brands
Introduction
In this “attention economy” (Davenport & Beck, 2002), where human attention is a
scarce commodity, online conversations in social media provide a rich forum for consumer
research. These conversations can be construed as the “wisdom of crowds” (Surowiecki, 2004).
They are spontaneous, passionate, information rich, and live. They affect consumer behavior
because consumers trust comments from other consumers more than any other factor (Blackshaw
& Nazarro, 2006). These conversations can drive sales (Dellarocas, Zhang, & Awad, 2007; Asur
& Huberman, 2010) and stock market performance (Tirunillai & Tellis, 2012). The high
visibility and impact of such conversations imply that they can be catastrophic for negative
events. Indeed, researchers find that bad news about brands travels fast in social media and that
negative chatter is more informative of firm performance than positive chatter (Chevalier &
Mayzlin, 2006; Kwak et al., 2010; Tirunillai & Tellis, 2012).
Product recalls are one of the most frequent negative events that firms face in the current
marketplace. Firms from various industries such as food, toy, automobile, and drugs encounter
product recalls. The number of product recalls has increased substantially over the last two
decades (see Figure 4 showing increase in automobile recalls) and is likely to rise in the future
(Dawar & Pillutla, 2000).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 56
Figure 4. Number of automobile recalls in USA (Source: NHTSA)
In 2010 alone, National Highway Traffic Safety Administration (NHTSA) reports that
more than 20 million vehicles were recalled. These recalls may lead to losses in sales and market
value for the recalling firm. Toyota’s recalls due to its unintended acceleration crises in the
winter of 2010 led the firm to lose 16% in sales and a drop of USD $12 billion in market value
(Rubel, Naik, & Srinivasan, 2011). Besides financial losses, product recalls can damage a firm’s
reputation, trust, and brand equity (Dawar & Pillutla, 2000; Rhee & Haunschild, 2006).
Prior research on product recalls mostly focuses on the recalled brand (see Kalaignanam,
Kushwaha, & Eilert, 2013 for an excellent overview of previous research on product recalls).
Researchers find that product recalls negatively affect the recalled firm’s sales, decrease the
firm’s advertising effectiveness, and erode the firm’s market value (Chu, Lin, & Prather, 2005;
Rhee & Haunschild, 2006; Van Heerde, Helsen, & Dekimpe, 2007; Cleeren, Dekimpe, &
Helsen, 2008; Chen, Ganesan, & Liu, 2009).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 57
However, scarce prior research exists on the effect of recalls among rival brands except a
forthcoming paper by Cleeren, Van Heerde, & Dekimpe (2013). This study differs from that
study in focusing on online conversations of brands as the dependent variable, in the context of
consumer durables, at the daily level of temporal aggregation, while analyzing detailed dynamics
of the relationship.
More generally, analyzing the effect of recalls on online consumer conversations offers
the following advantages: First, my study focuses on online conversations as the dependent
variable. Online conversations are now widely available and provide an early warning of the
impact on firm performance. Prior research has shown that online conversations are leading
indicators of sales (Dellarocas et al., 2007; Asur & Huberman, 2010) and stock market
performance (Tirunillai & Tellis, 2012). Second, my study analyzes the effects at a very high
temporal frequency (e.g., daily level). This level of temporal frequency means that firms have a
prompt feedback to crises, which gives them the option of a quick response to the crises. Third,
my study explores the dynamics of the relationship between crises and consumers conversations.
It can show how quickly or slowly a crisis affects a firm (wearin) and how that effect decays
over time (wearout).
In particular, the current study seeks to answer the following questions:
• Does perverse halo exist from online conversations? That is, do negative online
conversations of one brand increase negative online conversations of rival brands?
• Does reverse halo exist from online conversations? That is, do negative online
conversations of one brand decrease negative online conversations of rival brands?
• Are perverse and reverse halo effects from online conversations affected by the
brands’ country of origin?
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 58
• How quickly do these effects take to wearin and wearout? That is, what are the
dynamics of the effect?
• Does perverse and reverse halo exist from citations of print of the recalled brand to
rivals’ negative online conversations?
In order to answer the queries above, I obtain a unique dataset of online conversations for
three Japanese car brands and one American car brand. Using data at a daily level for a period of
470 days, which includes the Toyota recall crisis in the winter of 2010, I evaluate whether recall
of one brand increased or decreased a rival brand’s negative online conversations. I use Vector
AutoRegressive with exogenous variables (VARX) models to empirically identify these effects.
Moreover, I also evaluate the role of media in influencing the negative online conversations of
rivals.
My study makes the following contributions. First, mine is the first study to document the
effect of a recall on online conversations of the recalled brand and its rivals. Second, I theorize
and estimate two types of halo effects on rivals: perverse halo and reverse halo. Third, I capture
the dynamics of the effect of such recalls on online conversations of rivals. Fourth, I explore the
role of media in affecting such online conversations.
The rest of the paper is organized as follows. The second section presents the theory, the
third section explains the method, the fourth section describes the model, and the last two
sections present the results and discussion.
Theory
I first define the key terms used in the study. Next, I provide theory for why a recall for
one brand can increase negative online conversations of a rival.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 59
Definitions
I define two key terms relevant to the study: perverse halo and reverse halo. I define
perverse halo as the phenomenon whereby negative content about one brand increases negative
conversations for a rival. I define reverse halo as the phenomenon whereby negative content
about one brand decreases negative conversations for a rival.
Why Does Perverse Halo Occur?
The accessibility-diagnosticity theory proposed by Feldman and Lynch (1988) suggests
that if a consumer thinks brand A is diagnostic, i.e., informative of brand B, the consumer will
use perceptions of brand A’s quality to infer quality of brand B. However, this inference occurs
only when both brands and their quality perceptions are accessible, i.e., retrievable from
memory, at the same time. Thus, the possibility of halo depends on the existence and strength of
association between brands in a consumer’s memory.
Prior research reveals that consumers have associative networks where information about
brands and their attributes reside in the consumer’s knowledge network as nodes (Collins &
Loftus, 1975; Janakiraman, Sismeiro, & Dutta, 2009). Brands are interconnected in the
consumer’s mind through linkages between such nodes. The strength of the linkage between
nodes of two brands increases the accessibility of a rival. I posit that halo occurs when two
brands are similar. High similarity leads to an increase in both accessibility and diagnosticity
(Janakiraman et al., 2009). Halo effects occur in the positive domain. However, I posit that halo
effects would also hold in the negative domain if negative content for one brand causes negative
conversations for a rival. I refer to halo effects in the negative domain as perverse halo. Several
factors can increase perceptions of similarity between brands leading to perverse halo.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 60
The first factor is the “country of origin” of the brands. Research on “country of origin”
effect shows that country perceptions can lead to halo effects (Maheshwaran & Chen, 2006).
Prior research suggests that consumers might use “country” as an attribute and make similar
inferences for brands, which belong to the same country as the recalled brand and opposing
inferences for brands, which belong to a different country (Hong & Wyer, 1989, 1990).
Maheshwaran (1994) finds that both experts and novices use “country of origin” in product
evaluations when there is ambiguity in attribute information. Indeed, high ambiguity exists in
determining the root cause of automobile recalls. The Toyota acceleration crises exemplified this
ambiguity (Kane, Liberman, DiViesti, & Click, 2010). Moreover, Maheshwaran and Chen
(2006) find that “country of origin” effects occur more when consumers are angry. Anger among
consumers tends to be the dominating emotion in automobile recalls (Choi & Lin, 2009; Glor,
2010).
The second factor is the commonality of attributes between brands. An attribute such as
“safety”, a crucial attribute in a recall, is central to the automobile category. Indeed, Roehm and
Tybout (2006) find that scandals spillover from one brand to another if the scandal pertains to an
attribute that is strongly associated with a product category. Tversky’s (1977) “contrast” model
supports this view where similarity between brands increases due to the presence of common
attributes.
The third factor is the similarity of processes to develop a product. A recall for one brand
can erode trust in other brands in a category if the inadequacy of the production process is
perceived to be industry-wide (De Alessi & Staff, 1994). Consumers might assume that one
brand’s production process is similar to another. Indeed, in many cases, brands use the same
source and production processes to develop products. Due to common sourcing, manufacturing
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 61
defects in one brand may be present in another (Kraljic ,1983; Trevelen & Schweikhart, 1988;
Chopra & Sodhi, 2004; Hora, Bapuji, & Roth, 2011). For example, Kellogg’s recall of its peanut
butter sandwich crackers due to potential salmonella poisoning in early 2009 was followed by
recalls from Kroger, McKee Foods, and Perry’s Ice Cream. The peanut paste used in all these
brands had the same origin: Peanut Corporation of America. Such incidences of common
sourcing, which get high media coverage, may increase consumers’ belief that a rival may have
similar issues as the recalled brand.
Thus, a recall of one brand can increase a rival’s negative online conversations if the rival
and the recalled brand are perceived to be similar in the minds of the consumer.
These effects are initiated by the announcement of the recall. The announcement of the
recall can be voluntary where the firm recalls without any external persuasion. Or it can be
involuntary where the firm is pressured by the National Highway Transport Safety
administration (NHTSA) to recall. Consumers learn about the recall directly from sources such
as the firm, NHTSA, and news media if they miss the firm or NHTSA announcement. Prior
research finds that recalls damage the firm’s reputation in the consumer’s minds (Dawar, 1998;
Dawar & Pillutla, 2000). Such damage leads consumers to post negative conversations about the
recalled brand in message boards, forums, blogs, and review sites. Indeed, consumer use of
social media is pervasive and mobile. Nowadays, nearly four in five active Internet users engage
in social media (Nielsen, 2011). These negative conversations about the recalled brand then
affect a rival’s negative conversations.
Here, online conversations about the recalled brand appear to the consumers as “negative
news” spread over days. Consumers educate themselves about brands, products, and services
through online conversations that inform their purchase and loyalty behavior (Blackshaw &
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 62
Nazzaro, 2006). In fact, conversations from fellow consumers are seen as more credible and
trustworthy than media stories, company’s public relations, and advertising activities (Herr,
Kardes, & Kim, 1991; Bickhart & Schindler, 2001; Allsop, Bassett, & Hoskins, 2007;
Blackshaw & Nazzaro, 2006).
10
Researchers find that negative conversations travel faster
(Allsop et al., 2007; Kwok et al., 2010) and are more potent in influencing customers than
positive conversations (Arndt, 1967; Mizerski, 1982).
I posit that after the recall announcement, consumers react by posting negative
conversations about the recalled brand. These negative conversations could range from the faulty
attribute (“safety”) to the overall quality of the brand. In fact, these conversations may provide
scoops and insights, which the recall announcement missed. For example, an affected consumer
might indicate a reason for the faulty vehicle. Subsequently, unaware consumers, who miss the
announcement of the recall read these negative conversations and infer that a similar rival can
have similar problems. Or unaffected consumers, who at first don’t infer a similar rival’s quality
negatively, over time, change their minds.
Method
This section describes the research design, data collection, and measures.
Research Design
The next subsections explain the industry context, sampling of brands, timeframe, and
quasi-experimental design of the study.
Industry context. I select the U.S. automobile industry to analyze the effect of recalls for
several reasons. First, this industry has a high frequency of recalls. The high rate provides a
10
We test if conversations about a focal brand matter more in explaining a rival’s online conversations than the focal
brand’s media citations by using the Generalized Forecast Error Variance Decomposition technique.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 63
decent number of recall events and their media coverage for my analysis. Between 1966 and
2010, firms and the National Highway Transport Safety Administration (NHTSA) have recalled
more than 470 million vehicles (Schepp, 2011). The recall system was introduced in the USA in
1966 through the National Traffic and Motor Vehicle Safety Act. This was done to remove
potentially dangerous vehicles from the road and solve safety issues. Recalls have increased
since 1990 (Refer Fig. 1) peaking in 2004 (30.8 million vehicles). Reasons for these increases,
among others, include complexity of cars, changes in the regulatory environment, and common
sourcing (Peters, 2005; Bae & Benítez-Silva, 2012).
Second, the automobile industry gives me an ample number of online conversations per
day as there are numerous social media sites dedicated to it (e.g., forums:
www.automotiveforums.com, blogs: www.thetruthaboutcars.com, review sites:
www.edmunds.com). Other industries with frequent recalls have a lesser quantity of online
conversation activity compared to the automobile industry. The rich data in the automobile
industry enables analysis at the daily level. Such disaggregate temporal analysis is essential to
get deep insights into dynamics and avoid biased estimates (Tellis & Franses, 2006).
Third, the automobile industry is of considerable economic significance. It represents 3%
of US’s GDP and accounts for 1 out of 7 jobs in the US economy (Pauwels, Silva-Risso,
Srinivasan, & Hanssens, 2004; Kalaignanam et.al, 2013).
Automobile recalls can be voluntary or involuntary. Firms recall on their own when they
discover potential faults or hazards in their vehicles. These recalls are initiated even if there are
no official consumer complaints. In fact, when a firm learns of a defect, the firm is required by
U.S. law to start a recall, even if it is not pressed by the NHTSA. 2010 saw the highest number of
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 64
voluntary recalls. 14.9 out of 20.3 million vehicles were sent back to dealerships for repairs that
were not part of a NHTSA investigation.
Other firms need a push from the NHTSA to recall. Generally, this process involves four
elements. “Screening” involves a review of all consumer complaints by the Office of Defects
Investigation (ODI), a body within the NHTSA, to check if an investigation is warranted.
“Petition analysis” involves an examination of all consumer petitions where a consumer asks the
NHTSA to start an investigation for a vehicle defect. Based on these two elements, the ODI
evaluates if enough evidence exists for a full investigation. The “Investigation” process involves
a Preliminary Evaluation (PE) and an Engineering Analysis (EA). If these investigations reveal
that the firm is guilty, the NHTSA sends a notice to the firm. This notice gives the firm an
opportunity to argue against the verdict or provide new evidence. This is the last chance the firm
gets to voluntarily recall. If the firm does not recall, the NHTSA sends a notice to the firm to
start the recall process.
11
This makes the recall involuntary. Next, the firm informs their
customers and announces the recall in the media. “Recall Management” involves monitoring the
efficacy of the recall such as ensuring that the firm is fixing the problem quickly and efficiently.
I select both voluntary and involuntary recalls in my empirical analysis.
Sampling of brands. I select brands on several criteria to empirically test my theory.
First, these brands have a sufficient number of recalls. This criterion is because I need
conversations and media coverage about one brand’s recall to test the perverse halo effect in
online conversations. Second, these brands should be major brands and are at the auto firm level
(e.g., Toyota) rather than automaker (e.g., Lexus) or model level (e.g., Corolla). This criterion
11
If the firm does not agree with NHTSA’s decision, the matter is taken to the courts. While this scenario seldom
occurs, this does delay the recall process considerably (Bae & Benítez-Silva, 2012).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 65
ensures that I have a sizeable number of online conversations at a daily level across the time
period of the study. Third, some brands should belong to one country and some to another. This
criterion allows me to test if there is perverse halo due to “country of origin”. Based on these
criteria, I select the following brands for my study: Toyota, Honda, Nissan, and Chrysler. These
brands constitute four out of the top five brands with the most recalls in 2010. Toyota led the
number of recalled units followed by General Motors, Honda, Nissan, and Chrysler (Jensen
2011). Thus these brands (i.e., auto firms) include their automakers. Toyota: Toyota, Lexus, and
Scion. Honda: Honda and Acura. Nissan: Nissan and Infiniti. Chrysler: Chrysler, Dodge, Jeep,
and Ram.
Timeframe. I focus on the period from Jan 1st 2009 to April 15th 2010 as this period
witnessed a high number of recalls. In 2010, more than 20 million vehicles were recalled.
Furthermore, I ensure that my timeframe overlaps with the Toyota recall crisis as it allows me to
test if Toyota’s recalls went beyond financial losses, i.e., did it also hurt its own and its rival’s
online perceptions?
Quasi-Experimental design. I exploit the natural experiment in recalls. I use recalls as a
random shock to the system. This shock leads to a big increase in the number of negative online
conversations for the recalled brand. But in the absence of perverse or reverse halo (the null
hypothesis), recalls should not affect the negative online conversations of rivals. Thus, the
“recalls” variable works as a quasi-experimental manipulation and my design constitutes a
repeated natural event or quasi-experiment.
Data Collection of Online Conversations
I obtain the online conversations from a third party data provider. The firm uses its
proprietary software to mine and code these conversations using techniques such as Natural
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 66
Language Processing (NLP), machine learning, text mining, and statistical analysis. The online
conversations span postings about the four brands on Internet discussion forums, consumer
rating websites, Usenet newsgroups and listservs, blogs, and social networking sites such as
Twitter.
The firm’s algorithm quantifies the content of these conversations by generating tag data
(similar to “coding”) on three dimensions at the sentence level: subject, attribute, and valence.
For example, for a conversation with a sentence such as: “one cannot be safe in a Toyota car”,
the subject is Toyota, the attribute is safety and the valence is negative. The algorithm also
considers other inherent attributes of online conversations in its classification such as the URL,
author information, post time, etc. Moreover, to get accuracy, the algorithm goes beyond
keyword-based technology. In keyword-based technology, any conversation is divided into a list
of words without any stemming (e.g., love, loved, loving, etc. stemmed to “love”) and any
consideration of their meaning (e.g., “stock” can mean “company share”, “stored goods”,
“broth”). This implies that in classifications, the algorithm returns only those conversations with
words written exactly as the user writes them.
The third party data provider’s algorithm identifies: Implicit Subject (e.g. “This car has
great fuel economy.”). Though the subject is not explicitly mentioned, the 3rd party’s solution
picks it up. Implicit Product Feature/Issue (e.g., “This car is too small.”). Here the attribute
“size” is not mentioned but the 3rd party’s solution will detect it. Situation-dependent Valence
(e.g., “The battery life in this car is ‘long’” is a positive conversation whereas “This journey is
‘long’” is a negative conversation). The details of the third party’s classification algorithm are in
Appendix D.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 67
I check the accuracy of the algorithm’s classification with the help of two research
assistants. For this purpose, I randomly select 500 conversations from the total corpus of
negative conversations. Two research assistants independently read each conversation and
classify the conversation as positive, negative, and neutral. The inter-rater agreement is 86%. I
find a classification accuracy of 80%, i.e., 80% of the conversations classified as negative by the
algorithm are also negative as per both Research Assistants.
Measures of Endogenous Variables
This section explains the measures of the endogenous variables in the model: online
conversations and citations of print about recalls.
Measures of online conversation. I use the number of negative online conversations
about the brand’s recall attribute as the measure of online conversation for the three Japanese
brands. By recall attribute, I mean conversations related to the brand’s recall. I use the number of
negative online conversations about the brand’s acceleration attribute as the online conversation
metric for the American brand (Chrysler). I use negative conversations about the acceleration
attribute and not the recall attribute for Chrysler since negative conversations about recall are not
available for Chrysler. I use the term concerns to mean negative conversations about either the
recall or the acceleration attribute.
Why negative conversations? I use the number of negative conversations because prior
research suggests that negative information is less ambiguous (Birnbaum, 1972; Wyer, 1973) and
more diagnostic (e.g., Herr et al., 1991) than positive information. In fact, researchers find that
consumers give greater weight to negative information compared with positive information of
equal intensity in the formation of overall evaluations (Ahluwalia, 2002; Herr et al., 1991; Klein,
1996; Wyer, 1973). Moreover, prior research finds that negative conservations are more
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 68
informative of firm performance than positive conservations (Chevalier & Mayzlin, 2006; Kwak
et al., 2010; Tirunillai & Tellis, 2012).
Measure of citations of print about recalls. I measure citations of print about recalls as
the number of articles in print media, which cover the brand’s recall. I carry out a search for any
article that mentions the brand name and its recall using Lexis-Nexis and Factiva. Across both
the sources, I search all newspapers or newswires except non-US newspapers that mention the
brand name and its recall on any given day. I use Lexis-Nexis’ relevancy score feature to ensure
that I only select articles that are relevant and not chance mentions. I identify an article as
relevant if Lexis-Nexis gives a relevancy score of 60% or above (see Tirunillai & Tellis, 2012).
Similarly, I use Factiva’s company tag feature, which indicates if the results of a search are
relevant.
I use citations of print as an endogenous variable in my model as the agenda setting
theory argues that consumers regard an issue as important due to the saliency (as per the rate and
prominence of coverage) of the issue in the media (Lippmann, 1922; McCombs & Shaw, 1972).
Citations of print about a brand can prime consumers about the brand’s quality by repeating the
recall news. This priming can lead consumers to write negative conservations not only for the
recalled brand but also other brands. Moreover, it is possible that journalists read about the
brand’s recall in blogs, forums, and review sites, which in turn inform their journalistic pieces.
Thus, online conversations of a brand can trigger citations of print about the brand and its rivals.
Measures of Control Variables
I describe the measures of the control variables and why they could possibly influence the
results: Recalls, ABC news coverage, negative events in Toyota’s acceleration crisis, advertising,
new product announcements and key developments.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 69
Recalls. I use the total number of recalled units in each recall as the measure of recalls. I
use the Office of Defects Investigation’s (ODI) database of recalls to identify the dates, models,
brands, and units involved in each recall. This database captures all recalls, voluntary or
involuntary. The database covers all vehicle and equipment recalls attributable to the brand’s
manufacturing process, for which the brand has official responsibility. I do not have data on the
severity of the recall as it is no longer provided by the NHTSA. The severity or hazard level of
the recall included four levels. This score was provided by NHTSA until it stopped in 2001. I
match the recalled model with each of the four brands (e.g., Toyota Corolla Sedan matched with
Toyota). To confirm the details and date of the recall, I consult automobile sites (such as
www.cars.com, www.autoblog.com) and teaching cases (such as Greto, Schotter, & Teagarden,
2010).
I use recalls as a control variable because the recall event leads to an increase in negative
conversations for the recalled brand. I run robustness checks to test if recalls is endogenous. I run
Granger Causality tests till 20 lags and don’t find substantial evidence that concerns significantly
“Granger Cause” recalls.
ABC news coverage. I measure ABC News coverage by counting the number of times
Toyota’s (including its sub-brands) recall was mentioned in ABC news programs. I used the
Lexis-Nexis database to obtain the ABC transcripts. I text-mine the transcripts using a keyword-
based technology (selecting specific keywords such as Toyota, recall, and various brands and
models of Toyota) to identify and count the number of times Toyota’s recall was mentioned in
each day across the study timeframe.
I use ABC News coverage as a control variable because ABC ran several investigative
reports on Toyota’s acceleration crisis that tarnished Toyota’s image. The reports blamed Toyota
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 70
even though significant doubts existed in determining the recall’s root cause. For example, Brian
Ross exposé titled “Expert: Electronic Design Flaw Linked to Runaway Toyotas” showed
Toyota vehicles had a defect in their electronic system (Ross, 2010). Later investigations
clarified that this report was a hoax (Cook, 2010). The selective framing of reports could lead to
an increase in concerns not only for Toyota but also for other Japanese brands.
Negative events in Toyota’s acceleration crisis. I measure negative events related to
Toyota’s acceleration crisis by an indicator variable (1: Negative event; 0: No negative event) on
the day the event occurred. I examine content related to the crisis in the Lexis-Nexis and Factiva
databases and use the Toyota: The Accelerator Crisis case (Greto et al., 2010) to identify the
dates. I control for negative events, such as Consumer Reports dropping two of Toyota’s cars in
its annual list of top pick cars and US transportation secretary LaHood criticizing Toyota’s
response to the pedal problems. I use these events as a control variable because they could
stimulate concerns for Toyota and possibly other Japanese brands. Note that these events include
Toyota’s response to the crisis such as the Toyota president and CEO Akio Toyoda apologizing
for the car recalls and Toyota’s announcement to install a brake override system for the
acceleration issue.
Advertising. I measure a brand’s advertising by the daily dollar spend for the brand in
television stations in the United States of America. I obtain the advertising data from the Kantar
Stradegy database. I deflate the advertising spend by the monthly consumer price index. I use
advertising as a control variable because brands may advertise in response to an increase in
concerns. However, prior research finds a decrease in the recalled brand’s own advertising
elasticities (Van Heerde et al., 2007).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 71
New product announcements. I measure new product announcements by counting the
number of times a brand introduced a new product. I used the firm’s website and the Capital IQ
database to collect the data. I use new product announcements as a control variable because it
may reduce the number of negative online conversations about recall or acceleration (i.e.,
concerns) due to consumers’ enthusiasm about new cars.
Key developments. I measure key developments, which include firm’s press releases, by
counting the number of times a brand underwent a key development such as earnings
announcements, acquisitions, strategic alliances, awards, etc. I use key developments as a control
variable because they may affect the online conversations. I obtain the key developments data
from the firm’s website and the Capital IQ database.
Model
This section first explains the empirical strategy that I use to identify the effects. Next, I
explain why I use the Vector Auto-Regressive (VAR) modeling framework to estimate the
relationship among concerns and citations of print about recalls of the various brands (Dekimpe
& Hanssens, 1995, 1999). Next I explain the various steps in the VAR framework such as the
test for unit root and cointegration, the Vector Auto-Regression with Exogenous variables
(VARX), the Generalized Impulse Response Function, and the Generalized Forecast Error
Variance Decomposition.
Empirical Strategy
Figure 5 shows a snippet of the number of concerns for the three Japanese brands across
the timeframe of my study. I show only some recalls and not all for legibility reasons. The solid
arrows below the x-axis indicate recalls (e. g, recall dates). The arrow sizes suggest the size of
the recall. Note the variation in the time and size.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 72
I identify the effects of one brand’s recall on another by 1) exploiting the separation of
the recall dates across brands, and 2) using the variation in the number of units recalled for each
brand. In case of an overlap of recall dates, the variation in recalled units between the two brands
enables me to estimate the effects. I have only two instances where two brands recall on the same
date.
I run two different models. I do this since there is a mismatch between the attributes of
the Japanese brands and the American brand. I use the Vector Auto-Regressive modeling
framework in both models. The first focuses only on Japanese brands (Model 1). Thus, I use
Figure 5. Graph of recalls and concerns for Toyota, Honda and Nissan
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 73
negative conversations about recall attribute as the online conversation metric, and citations of
print about the recall as the endogenous variables in the first model.
In the second model, I focus on only the Toyota and Chrysler brands (Model 2). I use
negative conversations about the acceleration attribute as the online conversation metric for
Toyota and Chrysler. I use citations of print about acceleration as the endogenous variables in
this model. Note that for Toyota, I use the control variable of recalls related to its acceleration.
Why Vector Auto-Regressive Modeling Framework?
I use the Vector Auto-Regressive (VAR) framework for three reasons. First, it allows
estimation of causality among a set of variables (endogenous variables) via use of their lagged
values. Second, it ensures robustness of the model to issues of non-stationarity, spurious
causality, endogeneity, serial correlation, and reverse causality (Granger & Newbold, 1986).
Third, it enables estimation of the long term or cumulative effects of causal variables using the
impulse response functions (Tirunillai & Tellis, 2012; Nijs, Srinivasan, & Pauwels, 2007).
Test for Unit Root and Cointegration
The Unit Root test examines if the endogenous variables in my dataset are evolving or
are stationary. I use the Augmented Dickey-Fuller (ADF) test recommended by Enders (1995).
This test is carried out under the null hypothesis that there exists a unit root. Though, the ADF
test is the most widely used unit root test in marketing, the results of the test may be biased if
there is heteroscedasticity in the error term (Maddala & Kim, 1998). If there is heteroscedasticity
in the error term, the Phillips-Perron (PP) test is the appropriate unit root test (Phillip & Perron,
1988).
If the variables contain a unit root, (Johansen, 1995), I test for the possible presence of
cointegration (Granger & Newbold, 1974). That is, I test whether, if variables evolve, they
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 74
evolve together. For this purpose, I run the Johansen’s test for cointegration. If I find evidence
that the variables evolve together, I use the error correction model.
Based on the ADF, Phillips-Perron, and cointegration test, I choose the proper
specification for the endogenous variables that enter the VARX model. If I find that the variables
are not evolving and are not cointegrated, I run the VARX model using levels of the variables. If
I find that the variables are evolving and are not cointegrated, I run the VARX model using first
differences of the evolving variables. If I find that the variables are evolving and are
cointegrated, I use vector error correction models. Here, the model is formulated using first
differences of the variables. However, the model is adjusted on the right hand side with the
difference between the lags of the two evolving variables so that the variables move together in
the long run.
Vector Auto-Regressive Model with Exogenous Variables (VARX)
I estimate the relationships among concerns and citations of print about recalls of the
various brands using the Vector Auto-Regressive model with exogenous variables (VARX).
Citations of print may be endogenous because the media may report about a brand’s recall due to
the increase in negative conversations about the recall. The endogeneity implies that the recalls,
and citations of print are explained by both past variables of themselves (autoregressive
carryover effects) and past variables of each other (cross effects from citations of print of Brand
A to concerns of Brand B or vice versa). For ease of exposition, below is the model specification
using levels of the variables for the model with only the Japanese brands (Model 1). Note that
this specification may either include some variables in differenced form or may change when I
employ the Vector error correction model. The exact specification is determined by the unit root
result.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 75
=
t
t
t
t
t
t
MedNis
MedHon
MedToy
ConNis
ConHon
ConToy
+
+
+
+
+
+
t
t
t
t
t
t
MedNis MedNis
MedHon MedHon
MedToy MedToy
ConNis ConNis
ConHon ConHon
ConToy ConToy
*
*
*
*
*
*
δ α
δ α
δ α
δ α
δ α
δ α
+
∑
=
L
l
l l l l l l
l l l l l l
l l l l l l
l l l l l l
l l l l l l
l l l l l l
1
6 , 6 5 , 6 4 , 6 3 , 6 2 , 6 1 , 6
6 , 5 5 , 5 4 , 5 3 , 5 2 , 5 1 , 5
6 , 4 5 , 4 4 , 4 3 , 4 2 , 4 1 , 4
6 , 3 5 , 3 4 , 3 3 , 3 2 , 3 1 , 3
6 , 2 5 , 2 4 , 2 3 , 2 2 , 2 1 , 2
6 , 1 5 , 1 4 , 1 3 , 1 2 , 1 1 , 1
β β β β β β
β β β β β β
β β β β β β
β β β β β β
β β β β β β
β β β β β β
−
−
−
−
−
−
l t
l t
l t
l t
l t
l t
MedNis
MedHon
MedToy
ConNis
ConHon
ConToy
+ (6)
X * θ +
t MedNis
t MedHon
t MedToy
t ConNis
t ConHon
t ConToy
,
,
,
,
,
,
ε
ε
ε
ε
ε
ε
Here ConToy, ConHon, and ConNis denote concerns for Toyota, Honda and Nissan
respectively. MediaToy, MediaHon, and MediaNis denote citations of print about recall for
Toyota, Honda and Nissan respectively. The coefficients
3 , 1 2 , 1
β and β estimate the perverse halo
effect from Honda and Nissan on Toyota respectively. The coefficients
3 , 2 1 , 2
β and β estimate the
perverse halo effect from Toyota and Nissan on Honda respectively. The coefficients
2 , 3 1 , 3
β and β estimate the perverse halo effect from Toyota and Honda on Nissan respectively.
The vector X comprises the p control variables: Recalls, ABC news coverage, negative events in
Toyota’s acceleration crisis, advertising, new product announcements and key developments. I
add 3 additional controls to the vector X----- 1) day of the week dummies to control for day
effects, 2) year dummy, where 2009 is 0 and year 2010 is 1, to control for the higher number of
recalls, concerns, and media coverage in 2010 than in 2009, 3) Holiday dummy (Halloween,
Thanksgiving, Christmas, New Year, Martin Luther King day, Labor Day, Memorial Day, etc.)
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 76
to control for holiday effects. t is the deterministic-trend variable, which captures the effect of
omitted, gradually changing variables. θ and , β , δ , α are the parameters to be estimated and
t
ε
are white noise residuals, which are distributed as N (0, ). I estimate the models using a
Feasible Generalized Least Squares (FGLS) estimator that purges autocorrelation and
heteroscedasticity.
Generalized Impulse Response Function (GIRF)
Based on the parameters of the VARX model, I compute the total effects using the
impulse response function. Since there are numerous direct and indirect effects through
numerous lags, closed form estimation of total effects is not possible. The impulse response
functions serve two purposes for my analysis. First, they allow me to quantify the total effect of a
“shock” to one independent variable on the future values of a dependent variable incorporating
the direct and indirect effects. Impulse response functions provide the difference between two
forecasts: one based on an information set, which does not take the “shock” into account, and
another which takes the “shock” into account (Dekimpe & Hanssens, 2004). Second, they allow
me to measure the dynamics (wearin and wearout) of the effect of one variable on another. I use
the Generalized Impulse Response Function (Pesaran & Shin, 1998) to identify the effects. The
Generalized Impulse Response function is invariant to the temporal ordering of the variables in
the VARX system. This method uses information in the residual variance-covariance matrix of
equation (6) to capture the contemporaneous effects.
The Generalized Impulse Response Function estimates of a unit shock on independent
variable i on value of dependent variable j is:
6 ... 2 , 1 ,i,j=
σ
Σe Π e
= GIRF
ii
i T
'
j
ij,t
(7)
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 77
Here, ∑ is the covariance matrix of residuals,
j i
e e , are column vectors,
t
Π is the
coefficient matrix of the residuals at time t of the infinite moving average representation of
equation 6,
ii i
σ δ = is the unit shock to independent variable i. To evaluate the accuracy of my
GIRF estimates, I compute the confidence intervals using the bootstrapping approach with 1000
runs (Benkwitz et al., 2001). The procedure for computing the confidence intervals is:
1. Estimate equation (6), and save the residuals
t
ε ˆ , which contains the residuals for each
of the six equations in equation (6)
2. Compute centered residuals , ˆ ,..... ˆ
. . 1
ε ε ε ε − −
T
where
.
ε =
∑
−
t
T ε ˆ
1
, and generate
bootstrap residuals
* *
1
,...
T
ε ε by randomly drawing with replacement from the
centered residuals
3. Compute bootstrap time series recursively
T t Y Y c Y
t p t p t t
,..... 1 ,
ˆ
..
ˆ
ˆ
* * *
1 1
*
= + Π + + Π + =
− −
ε , where ( ) ( )
0 1
*
0
*
1
,...., ,...., Y Y Y Y
p p + − + −
=
4. Reestimate parameters based on the bootstrapped time series
5. Based on parameter estimates obtained in step 4, calculate a bootstrap version of the
statistic of interest
6. Repeat steps (2) to (5) 1000 times
Generalized Forecast Variance Decomposition (GFEVD)
I assess the relative contribution of each of the metrics on concerns (i.e., how much of
brand A’s concerns and citations of print explain a rival B’s concerns) using the Generalized
Forecast Error Variance Decomposition technique (Koop, Pesearan, & Potter, 1996; Pesaran &
Shin, 1998). Analogous to a ‘dynamic R2’, the Generalized Forecast Error Variance
Decomposition calculates how much of the variation in a response variable can be attributed to
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 78
each endogenous variable. The GFEVD is invariant to the temporal ordering of the variables in
the VARX system. The decomposition determines to what extent the concerns of brand A
contribute to the deviation in concerns of brand B from its baseline expectations. To establish the
statistical significance of the GFEVD estimates, I compute the confidence intervals using
bootstrapping approach similar to the GIRF. The Generalized Forecast Error Variance
Decomposition estimates of independent variable i on dependent variable j till time T is:
∑
∑
T
0 = t
j
'
t t
'
j
T
0 = t
2
i t
'
j
2 -
ii
T ij,
e Π ∑ Π e
) Σe Π (e ) σ (
= GFEVD , t= 1, 2,…T (8)
Results
This section presents descriptive results, tests for unit root and cointegration, results of
the VARX model, the variance decomposition, the effect on market capitalization, and the
robustness analysis.
Descriptive Results
Figure 5 shows the pattern of recalls and concerns during the timeframe of the study for
Japanese automobile brands. The solid arrows below the horizontal axis show recalls and other
events related to Toyota’s recall (e.g., ABC news investigation report on Nov 3rd). The size of
the arrows reflects the size of the recall. Not all recalls are shown due to space limits. Concerns
seem to correlate with recalls. There are spikes in Toyota’s concerns for its large recalls.
Similarly, there are spikes in Honda’s concerns for its large recalls.
Note the steep rise in the number of concerns for Toyota from Jan 21st 2010. It takes
about two months for the concerns to die down and return to their previous level. There are
minor spikes in the number of concerns for Honda during that week. Other recall events in the
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 79
graph increase not only the recalled brand’s concerns, but also a rival’s concerns. For example,
Honda’s recall on March 16th increases concerns for both Honda and Toyota.
Nevertheless the curves exhibit substantial variation over time. Moreover, it is difficult to
determine statistical effects or causality from all of these graphical associations. The VARX
framework will enable us to rigorously test if such associations are causal in the sense of Granger
causality.
The descriptive statistics and correlation matrix for the concerns and citations of print are
provided in Table E1 and E2 in the Appendix.
Testing for Evolution
Do concerns and citations of print for these brands evolve, i.e., do these metrics not revert
back to a base level but wander in one direction or another? I begin with the Augmented Dickey
Fuller (ADF) test to check if any of the endogenous variables are evolving. I use the iterative
procedure recommended by Enders (2004) to determine if a time trend or intercept should be
included in this test. I test stationarity in concerns and citations of print of all four brands. None
of the variables has a unit root as per the ADF test (See Table 12). The Phillips-Perron test
reveals that none of the variables evolve (See Table 13). Thus, I proceed with estimating the
VARX models (Models 1 & 2) in levels. This is the specification in Equation 6.
Estimation of VARX Model
As stated earlier, for Chrysler, I have data on only negative conversations about its
acceleration attribute but not the recall attribute. So I run one VARX model on only the Japanese
brands (Model 1) and another VARX model on only Toyota and Chrysler (Model 2). I report the
results of these models in separate sub-sections.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 80
Table 12
Augmented Dickey Fuller Test by Brand (Critical Values)
Brands Citations of print
about recall
concerns Acceleration
concerns
Citations of
print about
acceleration
Toyota -3.9 -2.24** -4.26 -3.95
Honda -8.83 -15.04 n.a. n.a.
Nissan -4.41 -19.69 n.a. n.a.
Chrysler n.a. n.a. -11.56 -14.95
Note. All the values of ADF are significant at 5% levels. The critical value with intercept and trend at 5% level is -
3.44; The critical value with no intercept and no trend at 5% level is 1.94; ** No intercept and trend as per Enders
iterative procedure (2004)
Table 13
Phillips-Perron Test by Brand (Critical Values)
Brands Citations of print
about recall
Concerns Acceleration
concerns
Citations of
print about
acceleration
Toyota -7.42 -9.5 -9.11 -5.44
Honda -10.41 -15.04 n.a. n.a.
Nissan -14.39 -19.63 n.a. n.a.
Chrysler n.a. n.a. -18.16 -14.84
Note. All the values of Phillips-Perron are significant at 5% levels. The critical value at 5% level is -3.42
The optimal lag order is 1 for all the VARX models as per the (Schwartz’s) Bayesian
Information Criteria. My results are not affected by the presence of any residual correlation, non-
normality of residuals, and heteroskedasticity. Further, the VARX model is estimated using a
Feasible Generalized Least Squares (FGLS) estimator that purges autocorrelation and
heteroscedasticity. I also estimate the VARX, model using an Ordinary Least Square regression
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 81
accounting for Heteroskedasticity and potential serial correlation with the Newey-West
Estimator. My results are robust to both methods. I report the results of the FGLS estimator.
I use the simulations of the Generalized Impulse Response Function to show the
cumulative effects of the independent variables (Pesaran & Shin, 1998; Dekimpe & Hanssens,
1999). Utilizing the VARX parameter estimates, the generalized impulse response function
tracks the over-time impact of a unit shock (one standard deviation) to one independent variable
on a dependent variable. I define the wear-in time as the time-to-peak effect of an independent
variable on a dependent variable. I define the total impact as the accumulated impact of the
impulse response function (of the effect of concerns of brand A on concerns of brand B) till the
accumulated impulse response function reaches its asymptote. Most of the accumulated effect of
concerns of one brand on concerns of another reaches the asymptote within 6 days. I report the
estimates of concerns and citations of print on concerns in term of elasticities using the arc
elasticity measure.
12
Estimates of VARX for Japanese brands (Toyota, Honda, Nissan). Note the important
distinction in these results from those presented in the next subsection is that these results are for
all Japanese brands that have a perception of the same country of origin. The VARX model
provides estimates of all the endogenous and exogenous variables on each endogenous variable.
Thus the matrix of these effects is quite large and can be complex to interpret. Table 14 provides
a simple case of two brands (A, B) and two endogenous variables (concerns and citations of print
about recall). Note, that some of the key off diagonal elements provide estimates of the halo,
12
,
Y
X Y
X
arc
×
∆
=
σ
η where
arc
η = arc elasticity, Y ∆ = impulse response of response (dependent) variable,
X
σ
= std. deviation of shocked (independent) variable, X =Mean value of shocked variable, Y = mean value of
response variable (see Trusov et al., 2009).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 82
because they capture the effect of an independent variable from brand A on a dependent variable
from Brand B. This VARX model also provides estimate of carryover (past values of a variable
on its current value), direct effects (past effects of one brand’s citations of print about recalls on
its own concerns), feedback (past effects from concerns on citations of print about recalls) and
reaction (past effects from citations of print about recalls from one brand on citations of print
about recalls of a rival). In the interest of parsimony, I will not discuss these four estimates.
Table 14
Interpretation of VAR coefficients (2 Brands A, B & 2 Variables, Concerns and Citations of
print)
Dependent Variable
Effect
Cause
Brand A
concerns
Brand B
concerns
Brand A
citations of
print
Brand B
citations of
print
Lagged
Independent
Variable
Brand A
concerns
Carryover
Halo
A→B
Feedback
direct A→A
Feedback react
A→B
Brand B
concerns
Halo B→A Carryover
Feedback react
B→A
Feedback direct
B→B
Brand A
citations of print
Direct
A→A
Halo
A→B
Carryover React A→B
Brand B
citations of print
Halo B→A
Direct
B→B
React B→A Carryover
Note: → indicates the direction of the effect from Cause to the Effect. The hypothesized perverse halo effects are
indicated using red shading.
Table 15 presents the results of Model 1 where I focus on the three Japanese brands.
Toyota’s concerns have a significant impact on Honda’s concerns and Nissan’s concerns.
Honda’s concerns have a significant impact on Toyota’s concerns and Nissan’s concerns.
Nissan’s concerns have a significant impact on Toyota’s concerns and Honda’s concerns. Note
that Toyota and Honda are dominant brands in the category while Nissan is a subdominant
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 83
brand. These results indicate that perverse halo effects exist from concerns of one brand to
another brand’s concerns both from dominant brands, such as Toyota, to others and from sub-
dominant brands, such as Nissan, to others. Thus, concerns seem to be quite viral among the
three Japanese brands causing halo in both directions. The wear-in time for these perverse halo
effects is 1 day.
Table 15
Full VARX Coefficient Matrix for Japanese Brands (All Estimates are Cumulative Effects)
Toyota
concerns
Honda
concerns
Nissan
concerns
Toyota
citations
of print
Honda
citations
of print
Nissan
citations
of print
Toyota concerns
1.05 0.2 0.06 0.89 0.9 0.65
Honda concerns
0.3 1.19 0.14 0.09 1.9 -0.07
Nissan concerns
0.04 0.15 1.04 -0.03 0.12 0.29
Toyota citations
of print
0.53 0.17 -0.02 1.08 1.02 0.76
Honda citations
of print
0.03 0.18 0.03 0.04 0.83 0.09
Nissan citations
of print
0.02 -0.01 0.06 0.06 0.1 0.85
Note: Significant values are indicated in boldface. Red shading indicates perverse halo. Significant effects are in
bold font. The cumulative effects are in elasticities. I follow prior VAR literature (Sims & Zha, 1999; Pauwels et al.,
2004; Trusov et al., 2009) to assess whether each impulse response value is significantly different from zero.
Figure 6 illustrates the results of the effect of Toyota’s concerns on Honda’s concerns.
Toyota’s concerns increase Honda’s concerns immediately with an accumulated elasticity of 0.2.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 84
Figure 6. Accumulated impulse response from Toyota’s concerns to Honda’s concerns
Toyota’s citations of print about recalls have a significant impact on Honda’s concerns
but don’t significantly affect Nissan’s concerns. Honda’s citations of print about recalls have a
significant impact on both Toyota’s and Nissan’s concerns. Nissan’s citations of print about
recalls neither significantly affect Toyota’s nor Honda’s concerns. These results indicate that
perverse halo effects of citations of print are stronger from dominant brands than sub-dominant
brands, such as Nissan. Figure 7 illustrates the results of the effect of Honda’s citations of print
about recalls on Nissan’s concerns. A one-unit shock in Honda’s citations of print about recalls
has an immediate impact on Nissan’s concerns with a peak (wear-in) in the first day. The effect
then wears out over the next few days to long term equilibrium, resulting in an accumulated
elasticity of 0.03.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 85
Figure 7. Accumulated impulse response from Honda’s citations of print to Nissan’s concerns
Citations of print about a focal brand’s recall significantly increase the focal brand’s
concerns. Figure 8 illustrates the results of the effect of Toyota’s citations of print about recalls
on Toyota’s concerns. Citations of print about Toyota’s recall immediately increase Toyota’s
concerns with an accumulated elasticity of 0.53. This result indicates that media damages the
online perceptions of the recalled brand.
Figure 8. Accumulated impulse response from Toyota’s citations of print to Toyota’s concerns
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 86
I next explain the other effects in Model 1. I find that Toyota’s concerns have a
significant impact on its own citations of print and on Honda’s and Nissan’s citations of print.
Honda’s concerns have a significant impact on its own and Toyota’s citations of print but don’t
significantly affect Nissan’s citations of print. Nissan’s concerns have a significant impact on its
own and Honda’s citations of print but don’t significantly affect Toyota’s citations of print.
These results indicate that journalists read online conversations about recalled brands and use it
to cover the recalled brand and its rivals, in print.
I find that Toyota’s citations of print significantly increase Honda’s and Nissan’s
citations of print. Honda’s citations of print significantly increase Toyota’s and Nissan’s citations
of print. Nissan’s citations of print significantly increase Honda’s and Toyota’s citations of print.
These results suggest that journalists write about a rival after referring a recalled brand’s print
articles.
Estimates of VARX for American and Japanese brands (Toyota and Chrysler). Note
that the important distinction between this analysis and the prior one is that Toyota and Chrysler
differ in terms of country of origin. Table 16 presents the results of the model where I focus on
Toyota and Chrysler.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 87
Table 16
Full VARX Coefficient Matrix for Toyota and Chrysler (All Estimates are Cumulative Effects)
Toyota
concerns
Chrysler
concerns
Toyota
citations of
print
Chrysler
citations of
print
Toyota concerns
1.24 -0.04 1.11 0.06
Chrysler concerns
-0.48 1.02 0.06 -0.01
Toyota citations of
print 1.2 0.09 2.23 0.01
Chrysler citations
of print 0.09 0.00 0.01 1.12
Note: Significant values are indicated in boldface. Red shading indicates perverse halo. Blue shading indicates
reverse halo. The cumulative effects are in elasticities.
Similar to the prior results of the effect of one brand’s citations of print on its rival’s
concerns, I find that Toyota’s citations of print increase Chrysler’s concerns and Chrysler’s
citations of print increase Toyota concerns. Thus, perverse halo effects from citations of print
occur even between brands from two different countries. The wear-in time for these effects is 1
day. Figure 9 shows the results of the effect of Toyota’s citations of print on Chrysler’s concerns.
A one-unit shock to Toyota’s citations of print increases Chrysler’s concerns peaking (wear-in)
on the first day. The effect then wears out over the next few days resulting in an accumulated
elasticity of 0.09. I next look at the cumulative impact of concerns of one brand on concerns for
another. Surprisingly, I find that Toyota’s concerns significantly decrease Chrysler’s concerns
and Chrysler’s concerns significantly decrease Toyota’s concerns.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 88
Figure 9. Accumulated impulse response from Toyota’s citations of print to Chrysler’s concerns
This is evidence of reverse halo. This result runs counter to the hypothesis of a perverse
halo. I elaborate on this phenomenon of reverse halo in the discussion section. The wear-in time
for these reverse halo effects is 2 days. Figure 10 illustrates the results of the effect of Chrysler’s
concerns on Toyota’s concerns. Chrysler’s concerns have a negative impact on Toyota’s
concerns resulting in a cumulative elasticity of -0.48.
Figure 10. Accumulated impulse response from Chrysler’s concerns to Toyota’s concerns
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 89
I next explain the other effects in Model 2. I find that Toyota’s concerns have a
significant impact on its own citations of print and on Chrysler’s citations of print. Chrysler’s
concerns have a significant impact on Toyota’s citations of print but don’t significantly affect its
own citations of print. These results indicate that journalists read online conversations about
recalled brands and use it to cover rivals from another country.
I find that Toyota’s citations of print do not significantly affect Chrysler’s citations of
print and Chrysler’s citations of print do not significantly affect Toyota’s citations of print. These
results suggest that journalists don’t tend to link rivals from two countries.
Relative Importance of Concerns and Citations of Print
The relative importance of each independent variable on a dependent variable is assessed
using the Generalized Forecast Error Variance Decomposition (GFEVD). I derive the GFEVD
estimates from the model estimated in VARX specification in equation 6. I establish the
statistical significance of the GFEVD estimates by bootstrapping with 1000 runs.
I find that concerns of a focal brand explain more variation in a rival brand’s concerns
than citations of print of the focal brand. The higher importance of concerns than citations of
print continues from the 1
st
till the 6
th
day, when the effects reach their asymptote.
Effect on Market Capitalization
I evaluate the effect of perverse halo and reverse halo on a firm’s market capitalization.
For perverse halo, I estimate the effect of a unit shock to Toyota’s concerns on Honda’s market
capitalization. I proceed in the following manner. Since I have the estimated impulse response of
a unit shock to Toyota's concerns on Honda's concerns from Model 1, I need to find the impulse
response of a unit shock to Honda's concerns on Honda's market capitalization thereby linking
the two impulse responses. Thus, I first estimate Honda’s abnormal returns using the Fama
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 90
French and Carhart Method (see Tirunillai & Tellis, 2012). Second, I run a VARX model that
includes Honda’s concerns and abnormal returns as endogenous variables. I control for variables,
which can affect Honda's abnormal returns and concerns such as analyst forecasts, offline
advertising, citations of print, new product announcements, acquisitions, and earnings
announcements. I calculate the impulse response of a unit shock to Honda's concerns on Honda's
abnormal stock returns (in basis points). Third, I calculate the value of a unit shock to Toyota's
concerns on Honda's abnormal returns. I do this by linking the impulse response from Toyota
concerns on Honda concerns with the impulse response from Honda concerns on Honda’s
abnormal returns. Fourth, the effect on Honda’s abnormal return is multiplied by the average
number of outstanding shares and the average share price over the 470 days of my sample. I find
that a unit shock to Toyota’s concerns erodes about $3 million from Honda’s average market
capitalization over just 6 days following the concerns. For reverse halo, I estimate the effect of a
unit shock to Chrysler’s concerns on Toyota’s market capitalization. I follow the same procedure
as above to estimate the effect. I find that a unit shock to Chrysler’s concerns benefits Toyota’s
average market capitalization by about $2.5 million over just 6 days following the concerns.
Note that these dollar values exclude the cost of trading. There is a possibility that transaction
costs may be high enough so that the gains made from trading are lost.
Robustness Analyses
I carry out five tests to establish the robustness of the results. I account for positive
conversations, brand size, presence of Toyota, car type, and omitted variable.
Positive conversations. It is possible that when one brand has a recall an increase in the
rival’s volume of negative conversations is accompanied by an increase in the rival’s volume of
positive conversations. For example, not accounting for such positive conversations may
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 91
upwardly bias my estimates of perverse halo. Thus, instead of using the volume of negative
conversations as the measure of online conversation, I use a negative sentiment measure. I
measure negative sentiment as the difference between the volume of negative conversations and
positive conversations. I use the term concerns to mean negative sentiment about either the recall
or the acceleration attribute.
Table F1 in Appendix F shows the results of the model for the three Japanese brands. For
easy reading and space constraints, I only provide the values of the perverse halo effect from
concerns of one brand on a rival’s concerns. As before, I find significant perverse halo effects
from one brand’s concerns on its rival’s concerns. However, Nissan’s concerns and Toyota’s
concerns do not significantly affect each other. Table F2 in Appendix F shows that reverse halo
effects exists between Toyota and Chrysler. Thus, my results are robust controlling for positive
conversations.
Brand size. Since the number of online conversations and citations of print can be
influenced by the size of the brand, I scale the negative sentiment measure by the sales of the
brand in the prior year. I obtain the data for the brand’s sales from the Wards Automotive
Yearbook. I use the term concerns to mean the scaled negative sentiment about either the recall
or the acceleration attribute. Table F3 in Appendix F shows the results of the model for the three
Japanese brands. For easy reading and space constraints, I only provide the values of the
perverse halo effect from concerns of one brand on a rival’s concerns. Similar to prior results, I
find perverse halo effects from one brand’s concerns on its rival’s concerns. The only non-
significant effect is from Nissan to Toyota. Table F4 in Appendix F shows that reverse halo
effects exists between Toyota and Chrysler. Thus, my results are robust controlling for positive
conversations and brand size.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 92
Presence of Toyota. Since Toyota underwent a one-of-a-kind of crises during the latter
half of 2009 and early 2010, it is possible that the results may be affected by the presence of
Toyota. I thus estimate the model focusing on only the Japanese brands removing Toyota. Table
F5 in Appendix F shows that perverse halo effects significantly exist between Honda’s concerns
and Nissan’s concerns.
Car type. Since I group luxury and non-luxury vehicle brands into one (e.g., automaker
Acura and Honda as one brand Honda), I run the risk of biasing the halo effects. Consumers may
react differently when a luxury car brand has a recall compared to a non-luxury car brand. Thus,
I estimate two models focusing on the Japanese brands. For the Japanese brands, I estimate one
model focusing on the luxury type (Lexus, Acura, Infiniti) and another model on non-luxury type
(Toyota, Honda, Nissan). Table F6 and F7 in Appendix F show the perverse halo effects from
luxury and non-luxury types respectively. I find evidence of perverse halo in both types.
However, the perverse halo effect appears to be more prevalent in luxury brands than non-luxury
brands. Lexus’ concerns increase Infiniti’s concerns and vice versa, however Toyota’s and
Nissan’s concerns don’t affect each other. This result could be consumer’s having higher
expectation about luxury cars than non-luxury cars. I don’t divide Chrysler into a luxury
(Chrysler) and non-luxury segment (Comprises Dodge, Jeep, Ram) because consumers are
ambivalent about Chrysler’s status as a luxury brand (Plain dealer wire services, 2009).
Omitted variable. Because I find that Toyota’s concerns in Model 2 are affected by
citations of print and concerns about Chrysler, I include Chrysler’s citations of print and
concerns as a control variable in Model 1. Note, I don’t incorporate Chrysler’s citations of print
and concerns with those of the three Japanese brands in Model 1 because there is a discrepancy
in the attribute. I don’t have online conversations about Chrysler’s recall attribute. Similarly, I
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 93
include Honda’s and Nissan’s citations of print and concerns as control variables in Model 2. I
don’t have online conversations about Honda’s and Nissan’s acceleration attribute. My results
are robust after the inclusion of these variables for both model 1 and 2.
Discussion
Product recalls are one of the most common events that firms face. This study seeks to
find out whether recalls for one brand can help or hurt rivals. Instead of focusing on metrics such
as sales and market share, I focus on negative online conversations, which are real time and
available at highly disaggregate temporal levels (e.g., hours, days). If rivals are affected, the
study researches when are they hurt or helped, and what are the routes and dynamics of these
relationships. This section summarizes the findings, discusses some key issues, suggests
implications, and lists the limitations.
Summary of Findings
The key findings of the study are the following:
• Perverse halo exists from online conversations, i.e., negative online conversations
about one brand increases negative online conversations for a rival. However,
perverse halo from online conversations occurs for only brands from the same
country.
• Reverse halo exists from online conversations i.e., negative conversations about a
brand decreases negative conversations for a rival. However, reverse halo occurs for
only brands from different countries.
• These halo effects suggest that consumers are affected by brands’ country of origin.
Consumers use “country” as an attribute and make similar inferences for brands,
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 94
which belong to the same country and opposing inferences for brands, which belong
to a different country.
• These perverse and reverse halo effects have a short wear-in of 1 day and a modest
wear-out of about 6 days. However, each of these effects (shocks) reoccur and last for
6 days.
• Perverse halo is widespread from citations of print to online conversations. Citations
of print for the recalled brand increases negative conversations for a rival irrespective
of the recalled brand’s country of origin.
These results suggest that same country brands face quintuple jeopardy online: from its
own recall and citations of print and from a rival’s recall, online conversations, and citations of
print.
Discussion of Key Issues
This section addresses 3 key questions emerging from the results: Why does perverse
halo exist? What is the mechanism behind perverse and reverse halo? Why do these perverse and
reverse halo effects have a short wear-in?
Why does perverse halo exist? I find substantial evidence of perverse halo from both
citations of print and concerns of one brand to a rival’s concerns. This effect could be attributed
to three reasons. First, consumers may use “country” as an attribute and make similar inferences
for brands, which belong to the same country as the recalled brand (Maheshwaran, 1994).
Indeed, I find that perverse halo effects occur in both citations of print and concerns only among
the Japanese brands. Second, consumers might view a focal attribute in a recall as common
among brands. Common attributes increase perceptions of similarity among brands (Tversky,
1977). Indeed, safety, which is a focal attribute in a recall, is a common attribute in the
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 95
automobile category. Third, consumers might assume that brands use the same manufacturing
processes to develop a product.
What is the mechanism behind perverse and reverse halo? Prior research finds that
“country of origin” effects occur more when consumers are angry (Maheshwaran & Chen, 2006).
Consumers are prone to carry out more heuristic processing than systematic processing when an
emotion such as “anger” overpowers them. An emotion such as anger leads consumers to have
more “country of origin” thoughts.
In order to test this supposition, I carry out a text mining analysis on the online
conversations. I first create a dictionary of emotion words (e.g., fear, cheerful, angry). I focus on
three emotions: anger, sadness, and fear. I use the WordNet Affect list of emotion words (Anger,
Disgust, Fear, Joy, Sadness, and Surprise) to classify the online conversations (Strapparava &
Mihalcea, 2007; Strapparava & Valitutti, 2004). WordNet Affect list of emotion words has been
used for categorizing news headlines previously (Strapparava & Mihalcea, 2007). I supplement
the WordNet list with emotion words denoting anger, fear, and sadness from the General
Inquirer, a well-known quantitative content analysis program and Opinion Lexicon (Liu, 2004), a
list of positive and negative opinion words or sentiment words for English. General Inquirer has
been used in quite a few papers in Finance and Opinion Lexicon has been used in multiple text
mining studies.
Note the dictionary contains words in their un-stemmed versions (e.g., scared, scare).
Thus, the list is exhaustive of all possible versions of a word (e.g., love, loving, loves). After
assembling the dictionary of words, I run a keyword based search to find if the content of the
concerns included any of the emotion words.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 96
I find that online conversations have significantly (p<.01) more anger than sadness and
fear. These results strengthen the argument that “country of origin” effects maybe at work.
Besides the text mining, two research assistants read and classified a random sample of 500
conversations into anger, fear and sadness. I find results similar to the text mining analysis.
Why do these perverse and reverse halo effects have a short wear-in? I find the
perverse and reverse halo effects have a short wear-in. This result suggests that consumers
quickly respond in social media in associating a rival with the recalled brand. Consumer use of
social media is pervasive and mobile. Nowadays, nearly four in five active Internet users engage
in social media (Nielsen, 2011). Moreover, nearly 40% of social media users access social media
content from their mobile phones (Nielsen, 2011). Always on the go, consumers could react and
post to a recall event directly by venting negative opinions about the recalled brand and similar
rivals, accessed from memory. These consumers would spread their concerns fast because they
want to save others from harm or because they are angry (Allsop et al., 2007). Other consumers
who are constantly engaging with friends in social networks or peers in forums might read about
concerns about the recalled brand. These concerns are more personal, credible, and vivid than
media reports (Herr et al., 1991; Allsop et al., 2007). After reading through these concerns,
consumers would access a rival brand from their memory and instantly voice a concern about a
similar brand or drop a concern about a dissimilar brand (e.g., a brand from a different country as
the recalled brand). Indeed, 60% of social media users write reviews (Nielsen, 2011).
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 97
Implications
This study has the following implications.
First, brands from the same country should keep an eye on rival’s recall events. I find that
an increase of 16 concerns for Toyota, which is one event (shock), can erode about $3 million
dollars from Honda’s market capitalization. However, each of these events (shocks) reoccurs and
lasts for 6 days. Thus, it is imperative for managers to continuously monitor their rival’s negative
events. As soon as a rival has a recall, firms should lie low and avoid comparisons with brands
undergoing a recall crisis thereby cutting off reflected failure (Snyder, Higgins, & Stucky, 1983).
Social comparison theory suggests that brands can protect their image or status by avoiding
comparisons with less reputable others (Snyder, Lassegard, & Ford, 1986; Washington & Zajac,
2005) or those undergoing a crisis. A denial strategy of stating how their sourcing,
manufacturing, designs, and scientific procedures have no link with the focal recall could
backfire for the rival (Siomkos & Shrivastava, 1993). Explicit statements of separation may lead
consumers to suspect a connection with the recalled brand.
Second, brands from a different country to the recalled brand could emphasize their
strengths and uniqueness when the recalled brand is under crisis (Hauser & Shugan, 1983).
Third, brands need to give more thought to the role of consumer opinions in determining
their rivals. This knowledge of consumer thinking will allow the brands to strategically deviate
from consumer perspectives (Kim & Tsai, 2012). If consumers think two brands are similar and
comparable, the innocent rival faces the danger of receiving negative feedback when the other
has a recall. Thus, brands may need to deviate from their current positioning and look unique.
Prior research has shown that comparative advertising increases consumers’ perceptions of
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 98
similarity between brands (Gorn & Weinberg, 1984; Snyder, 1992; Sujan & Dekleva, 1987; Kim
& Tsai, 2011).
Fourth, investors can use such online conversations in their trading strategies. For
example, I find that a unit shock to Toyota’s concerns erodes about $3 million from Honda’s
average market capitalization over the 6 days following the concerns. On the other hand, a unit
shock to Chrysler’s concerns benefits Toyota’s average market capitalization by about $2.5
million over the 6 days following the concerns.
Fifth, marketing managers of the recalled brand need to focus on managing both mass
media (e.g. print) as well as social media. During crisis situations, it is imperative for brands to
communicate with consumers in the right way, such as placating various concerns. Brands often
only focus on mass media as an external factor that would influence consumers. Thus, they adopt
communication strategies to manage mass media (Siomkos & Shrivastava, 1993). However, the
ubiquity of social media has created new challenges. Brands need to handle the spread of
information about product recalls in social media. Concerns for a brand can diffuse to a wider
audience in seconds and have high acceptances by fellow consumers. Thus, as a first step firms
should relay the information about the recall to all important social media sites, have a
comprehensive set of FAQs and ensure that all searches for information about the recall are
directed to one place (e.g., a microsite dedicated to the recall).
Next, brands need to listen to conversations in social media, preferably by the valence
(e.g., negative) and know the hashtags and keywords being used to discuss the recall. Identifying
the hashtags and keywords can enable managers to track mentions about the recall in the social
media space (Fisher, 2012). Brands can subsequently engage in a two-sided dialogue in these
important social media sites to allay specific concerns. This dialogue could mitigate the tide of
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 99
concerns that can diffuse beyond one network. For example, social media or online community
managers could provide clear information about the recall, steps taken to reduce the hazards and
address specific concerns directly either in their microsite or via their own blogs, social network
accounts (such as Facebook Groups, Twitter accounts, Facebook apps), and forums and address
concerns as they come up.
Limitations
This study has several limitations that can be basis for future research. First, I restricted
my focus to the automobile industry because of its high frequency of recalls and availability of
online conversations. It would be worthwhile to investigate the generalizability of the results to
other product categories and brands. Second, I analyze the effect of recalls at the auto firm (e.g.,
Toyota) or automaker (e.g., Lexus) level than model level (e.g., Corolla). I did this due to the
need of ample online conversations at a daily level. It would be useful to explore the effects of
recalls of one brand on a rival at the model level. Third, I do not categorize advertising by its
informational risk reduction (Byzalov & Shachar, 2004) or persuasive role. It would be
interesting to check if online conversations are affected by the type of advertising.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 100
References
Ahluwalia, R. (2002). How prevalent is the negativity effect in consumer environments? Journal
of Consumer Research, 29, 270-279.
Allsop, D. T., Bassett, B. R., & Hoskins, J. A. (2007). Word-of-mouth research: principles and
applications. Journal of Advertising Research, 47, 398-411.
Anderson, E. W. (1998). Customer satisfaction and word of mouth. Journal of Service Research,
1, 5-17.
Anderson, E. W., & Sullivan, M. W. (1993). The antecedents and consequences of customer
satisfaction for firms. Marketing Science, 12, 125-143.
Archak, N., Ghose, A., & Ipeirotis, P. (2011). Deriving the pricing power of product features by
mining consumer reviews. Management Science, 57, 1485-1509.
Arndt, J. (1967). Role of product-related conversations in the diffusion of a new product. Journal
of Marketing Research, 4, 291-295.
Arno, C. (2012). Worldwide social media usage trends in 2012. searchenginewatch.com.
Retrieved February 4, 2013, from http://searchenginewatch.com.
Arthur, C. (2011, February 14). Twitter chief dismisses talk of $10bn offer from Google as
'rumour'. guardian.co.uk. Retrieved December 1, 2011, from http://www.guardian.co.uk.
Asur. S., & Huberman, B. A. (2010). Predicting the future with social media.
arXiv:1003.5699v1.
Bae, Y-K., & Benítez-Silva, H. (2012). The effects of automobile recalls on the severity of
accidents. Economic Inquiry. doi: 10.1111/(ISSN)1465-7295
Barth, M., Beaver, W., & Landsman, W. (2001). The relevance of the value relevance literature
for financial accounting standard setting: another view. Journal of Accounting and
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 101
Economics, 31, 77-104.
Baker, S. (2009, December 14). Beware social media snake oil. BusinessWeek, 4159, 48-51.
Bickart, B., & Schindler, R. (2001). Internet forums as influential sources of consumer
information. In M.C. Gilly & J. Meyers-Levy (Eds.), NA - Advances in Consumer
Research Volume 28, in Valdosta, GA: Association for Consumer Research (pp. 134)
Birnbaum, M. H. (1972). Morality judgments: Tests of an averaging model. Journal of
Experimental Psychology, 93, 35-42.
Benkwitz, A., Lütkepohl, H. & Wolters, J. (2001). Comparison of bootstrap confidence intervals
for impulse responses of German monetary systems. Macroeconomic Dynamics, 5, 81-
100.
Bennett, S. (2011, October 18). Twitter: 250 million tweets per day and $8 billion value
confirmed (nut no IPO… yet). mediabistro.com. Retrieved November 30, 2011, from
http://www.mediabistro.com.
Blackshaw, P., & Nazzaro, M. (2006). Consumer-generated media (CGM) 101. Word-of-mouth
in the age of the web-fortified consumer. (2nd ed.) New York: BuzzMetrics.
Bollen, J., Mao, H., & Zeng, X-J. (2011). Twitter mood predicts the stock market. Journal of
Computational Science. 2, 1-8.
Bowley, G. (2010). Computers that trade on the news. nytimes.com. Retrieved November 22,
2011, from http://www.nytimes.com.
Bowman, D., & Narayandas, D. (2001). Managing customer-initiated contacts with
manufacturers: The impact on share of category requirements and word-of-mouth
behavior. Journal of Marketing Research, 38, 281-297.
Boser, B.E., Guyon, I., & Vapnik. V. (1992). A training algorithm for optimal margin classifiers.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 102
Fifth Annual Workshop on Computational Learning Theory (pp. 144–152). ACM Press.
Byzalov, D., & Shachar. R. (2004). The risk reduction role of advertising. Quantitative
Marketing and Economics, 2, 283-320.
Boyd, D., Golder, S., & Lotan, G. (2010, 5-8 Jan). Tweet tweet retweet: Conversational aspects
of retweeting on twitter. Proceedings of the 43
rd
Annual Hawaii International Conference
on Systems Science (HICSS-43). Kauai, HI: IEEE Computer Society.
Carhart, M. (1997). On the persistence of mutual fund performance. Journal of Finance, 52, 57–
82
Chen, Y., Ganesan, S., & Liu, Y. (2009). Does a firm's product recall strategy affect its financial
value? An examination of strategic alternatives during product-harm crises. Journal of
Marketing, 73 (6), 214-226.
Chevalier. J., & Mayzlin D. (2006). The effect of word of mouth on sales: online book reviews.
Journal of Marketing Research, 43, 345-354.
Christakis, N.A., & Fowler, J.H. (2007). The spread of obesity in a large social network over 32
years. New England Journal of Medicine, 357, 370-379.
Choi, Y., & Lin, Y-H. (2009). Consumer response to crisis: Exploring the concept of
involvement in Mattel product recalls. Public Relations Review, 35, 18-22.
Chopra, S., & Sodhi, M.S. (2004). Managing risk to avoid supply-chain breakdown. Sloan
Management Review, 46 (1), 53–61.
ComScore (2011, October 20). Social networking on-the-go: U.S. mobile social media audience.
comscore.com. Retrieved December 12, 2011, from http://www.comscore.com.
Chu, T.H., Lin, C.C., & Prather, L.J. (2005). An extension of security price reactions around
product recall announcements. Quarterly Journal of Business & Economics, 44 (3-4), 33-
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 103
47
Cleeren, K., Dekimpe, M. G., & Helsen, K. (2008). Weathering product-harm crises. Journal of
the Academy of Marketing Science, 362, 262-270
Cleeren, K., Van Heerde, H. J., & Dekimpe, M.G. (in press). Rising from the ashes: How brands
and categories can overcome product-harm crises. Journal of Marketing. doi:
10.1509/jm.10.0414
Collins, A.M., & Loftus, E.F. (1975). A spreading activation theory of semantic processing.
Psychological Review, 82, 407–428.
Cook, J. (2010, March 5). How ABC news’ Brian Ross staged his Toyota death ride.
gawker.com. Retrieved May 20, 2012, from http://gawker.com.
Cortes, C. , & Vapnik, V. (1995). Support vector networks. Machine Learning, 20, 273– 297.
Cui, D., & Curry, D. (2005). Prediction in marketing using the support vector machine.
Marketing Science, 24, 595-615.
Das, S., & Chen, M. (2007). Yahoo! for Amazon: Sentiment extraction from small talk on the
web. Management Science, 53, 1375.
Davenport, T.H., & Beck, J.C. (2002). The Attention economy: Understanding the new currency
of business. Cambridge, MA: Harvard Business Press.
Dawar, N. (1998). Product-Harm crises and the signaling ability of brands. International Studies
of Management & Organization, 28 (3), 109-119.
Dawar, N., & Pillutla, M. M. (2000). Impact of product-harm crises on brand equity: The
moderating role of consumer expectations. Journal of Marketing Research, 37, 215-226.
Deighton, J., & Kornfeld, L. (2009). Obama versus Clinton: The YouTube Primary. Harvard
Business School Cases, Retrieved October 15, 2011, from http://harvardbusinessonline.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 104
hbsp.harvard.edu/b01/en/common/item_detail.jhtml?id=509032
De Alessi, L., & Staaf, R.J. (1994). What does reputation really assure? The relationship of
trademarks to expectations and legal remedies. Economic Inquiry, 32, 477–485.
Dekimpe, M., & Hanssens, D.M. (1995). The persistence of marketing effects on sales.
Marketing Science, 14, 1-21.
Dekimpe, M. & Hanssens, D.M. (1999). Sustained spending and persistent response: A new look
at long-term marketing profitability. Journal of Marketing Research, 35, 397-412.
Dekimpe, M., & Hanssens, D.M. (2004). Persistence modeling for assessing marketing strategy
performance. In C. Moorman & D. Lehmann (Eds.), Assessing Marketing Strategy
Performance. Cambridge, MA: Marketing Science Institute.
Dellarocas, C., Zhang, X. (Michael), & Awad, N. F. (2007). Exploring the value of online
product reviews in forecasting sales: The case of motion pictures. Journal of Interactive
Marketing, 21(4), 23–45. doi:10.1002/dir.20087
Enders, W. (2009). Applied Econometric Times Series (3rd ed.). New York: Wiley.
Fama, E. F., & French, K. R. (1993). Common risk factors in the returns on stocks and bonds.
Journal of Financial Economics, 33, 3–56.
Feldman, J. M., & Lynch, J. G. (1988). Self-generated validity and other effects of measurement
on belief, attitude, intention, and behavior. Journal of Applied Psychology, 73, 421-435.
Fisher, T. (2012, April 20). Tips for using social media in product recall. socialmediatoday.com.
Retrieved June 24, 2012, from http://socialmediatoday.com.
Glor, J. (2010, January 29). Toyota recall fuels confusion, anger. cbsnews.com. Retrieved June
11, 2012, from http://cbsnews.com.
Granger, C. W. J. (1969). Investigating causal relations by econometric models and cross-
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 105
spectral methods. Econometrica: Journal of the Econometric Society, 37, 424–438.
Granger, C. W. J., & Newbold, P. (1974). Spurious regressions in econometrics. Journal of
Econometrics, 2, 111–120.
Gorn, G. J., & Weinberg, C. B. (1984). The impact of comparative advertising on perception and
attitude: Some positive findings. Journal of Consumer Research, 11, 719–727.
Greto, M., Schotter, A. & Teagarden, M.B. (2010). Toyota: The accelerator crisis. Thunderbird
School of Global Management Cases, Retrieved October 20, 2011, from
http://caseseries.thunderbird.edu/case/toyota-accelerator-crisis
Hanssens, D. M. (1980). Market response, competitive behavior, and time series analysis.
Journal of Marketing Research, 17, 470–485.
Hauser, J. R., & Shugan, S. M. (1983). Defensive marketing strategies. Marketing Science, 2,
319–360.
Herr, P. M., Kardes, F. R., & Kim, J. (1991). Effects of word-of-mouth and product-attribute
information on persuasion: An accessibility-diagnosticity perspective. Journal of
Consumer Research, 17, 454–462.
Healy, P. M., & Palepu, K. G. (2001). Information asymmetry, corporate disclosure, and the
capital markets: A review of the empirical disclosure literature. Journal of Accounting
and Economics, 31, 405–440.
Hoffman, D. L., & Fodor, M. (2010). Can you measure the ROI of your social media marketing?
MIT Sloan Management Review, 52(1), 41–49.
Hong, S. T., & Wyer Jr, R. S. (1989). Effects of country-of-origin and product-attribute
information on product evaluation: an information processing perspective. Journal of
Consumer Research, 16, 175–187.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 106
Hong, S. T., & Wyer Jr, R. S. (1990). Determinants of product evaluation: effects of the time
interval between knowledge of a product’s country of origin and information about its
specific attributes. Journal of Consumer Research, 17, 277–288.
Hora, M., Bapuji, H., & Roth, A. V. (2011). Safety hazard and time to recall: The role of recall
strategy, product defect type, and supply chain player in the US toy industry. Journal of
Operations Management, 29, 766–777.
Hu, M., & Liu, B. (2004). Mining and summarizing customer reviews. In Proceedings of the
Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining (pp. 168–177). Seattle, WA, USA: ACM.
Israel, S. (2008, October 4). The power of retweeting. redcouch.typepad.com/weblog/. Retrieved
August 18, 2011, from http://redcouch.typepad.com/weblog.
Jaffe, J. F. (1974). Special information and insider trading. Journal of Business, 47, 410–428.
Janakiraman, R., Sismeiro, C., & Dutta, S. (2008). Perception spillovers across competing
brands: a disaggregate model of how and when. Journal of Marketing Research, 46, 467-
481
Jensen, C. (2011, January 20). 2010 a record year for ‘Voluntary’ recalls. nytimes.com. Retrieved
March 2, 2012, from http://wheels.blogs.nytimes.com.
Johansen, S. (1996). Likelihood-based inference in cointegrated vector autoregressive models.
New York: Oxford University Press, USA.
Joachims, T. (2002). Learning to classify text using support vector machines: Methods, theory
and algorithms (Vol. 186). Norwell, MA, USA: Kluwer Academic Publishers.
Kalaignanam, K., Kushwaha, T., & Eilert, A. (2012). The impact of product recalls on future
product reliability and future accidents: Evidence from the automobile industry. Journal
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 107
of Marketing. doi: 10.1509/jm.11.0356
Kane, S., Liberman, E., DiViesti, T., & Click, F. (2010, October 25). Toyota sudden unintended
acceleration. safetyresearch.net. Retrieved May 14, 2012, from
http://www.safetyresearch.net/.
Kelman, H. C. (1958). Compliance, identification, and internalization: Three processes of
attitude change. The Journal of Conflict Resolution, 2, 51–60.
Kim, K. H., & Tsai, W. (2012). Social comparison among competing firms. Strategic
Management Journal, 33, 115–136.
Klein, J. G. (1996). Negativity in impressions of presidential candidates revisited: The 1992
election. Personality and Social Psychology Bulletin, 22, 288–295.
Koop, G., Pesaran, M. H., & Potter, S. M. (1996). Impulse response analysis in nonlinear
multivariate models. Journal of Econometrics, 74, 119-147.
Kouloumpis, E., Wilson, T., & Moore, J. (2011). Twitter Sentiment Analysis: The Good the Bad
and the OMG! In Proceedings of the International AAAI Conference on Weblogs and
Social Media. Retrieved January 10, 2013, from
www.aaai.org/ocs/index.php/ICWSM/ICWSM11/paper/view/2857/3251.
Kraljic, P. (1983). Purchasing must become supply management. Harvard Business Review,
61(5), 109–117.
Kwak, H., Lee, C., Park, H., & Moon, S. (2010). What is Twitter, a social network or a news
media? In Proceedings of the 19th international conference on world wide web (pp. 591-
600). Raleigh, North Carolina, USA: ACM.
Lippmann, W. (1997). Public opinion. New York, NY: Macmillan.
Liu, Y. (2001). Word-of-mouth for movies: Its dynamics and impact on box office revenue.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 108
Journal of Marketing, 70(3), 74–89.
Luo, X. (2009). Quantifying the long-term impact of negative word of mouth on cash flows and
stock prices. Marketing Science, 28, 148–165.
Maddala, G. S., & Kim, I. M. (1999). Unit roots, cointegration, and structural change (No. 4).
Cambridge, UK: Cambridge University Press.
Maheswaran, D. (1994). Country of origin as a stereotype: effects of consumer expertise and
attribute strength on product evaluations. Journal of Consumer Research, 21, 354–365.
Maheswaran, D., & Chen, C. Y. (2006). Nation Equity: Incidental Emotions in Country-of-
Origin Effects. Journal of Consumer Research, 33, 370–376.
Malhotra, A., Kubowicz Malhotra, C., & See, A. (2012). How to get your messages retweeted.
MIT Sloan Management Review, 53(2), 61-66.
McCombs, M. E., & Shaw, D. L. (1972). The agenda-setting function of mass media. Public
Opinion Quarterly, 36, 176–187.
Mitchell, M. L., & Mulherin, J. H. (1994). The impact of public information on the stock market.
The Journal of Finance, 49, 923–950.
Mitchell, M. L., & Stafford, E. (2000). Managerial decisions and long-term stock price
performance. The Journal of Business, 73, 287–329.
Mizerski, R. W. (1982). An attribution explanation of the disproportionate influence of
unfavorable information. Journal of Consumer Research, 9, 301–310.
Newey, W. K., & West, K. D. (1987). A simple, positive semi-definite, heteroskedasticity and
autocorrelation consistent covariance matrix. Econometrica: Journal of the Econometric
Society, 55, 703–708.
Nielsen (2011). State of the media: The social media report - Q3 2011. nielsen.com. Retrieved
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 109
November 14, 2011, from http://www.nielsen.com.
Nijs, V. R., Srinivasan, S., & Pauwels, K. (2007). Retail-price drivers and retailer profits.
Marketing Science, 26, 473–487.
Owyang, J. (2008, November 23). Retweet: The infectious power of word of mouth. web-
strategist.com. Retrieved November 12, 2011, from http://www.web-strategist.com.
Pauwels, K., & Srinivasan, S. (2004). Who benefits from store brand entry? Marketing Science,
23, 364–390.
Pauwels, K., Silva-Risso, J., Srinivasan, S., & Hanssens, D. M. (2004). New products, sales
promotions, and firm value: The case of the automobile industry. Journal of Marketing,
68(4), 142–156.
Pesaran, H. H., & Shin, Y. (1998). Generalized impulse response analysis in linear multivariate
models. Economics Letters, 58, 17–29.
Peters, J.W. (2005, October 26). More complex cars and stricter rules lead to more recalls.
nytimes.com. Retrieved May 30, 2012, from http://www.nytimes.com.
Phillips, P. C. B., & Perron, P. (1988). Testing for a unit root in time series regression.
Biometrika, 75, 335–346.
Plain dealer wire services (2009, October 5). Chrysler brand faces hard sell with luxury buyers,
analysts say. cleveland.com. Retrieved May 11, 2012, from http://www.cleveland.com.
Rao, L. (2010, September 14). Twitter seeing 90 million tweets per day, 25 percent contain links.
techcrunch.com. Retrieved May 30, 2011, from http://techcrunch.com.
Rhee, M., & Haunschild, P. R. (2006). The liability of good reputation: A study of product
recalls in the US automobile industry. Organization Science, 17, 101–117.
Roehm, M. L., & Tybout, A. M. (2006). When will a brand scandal spill over, and how should
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 110
competitors respond? Journal of Marketing Research, 43, 366–373.
Ross, B. (2010, February 22). UPDATED: Expert: Electronic design flaw linked to runaway
Toyotas. abcnews.go.com. Retrieved May 12, 2012, from http://abcnews.go.com/.
Rui, H., Whinston, A., & Winkler, E. (2009, November 30). Follow the tweets.
sloanreview.mit.edu. Retrieved October 7, 2010, from http://sloanreview.mit.edu.
Rubel, O., Naik, P. A., & Srinivasan, S. (2011). Optimal advertising when envisioning a product-
harm crisis. Marketing Science, 30, 1048–1065.
Schepp, D. (2011, January 19). It's not just Toyota: Auto recalls accelerate. dailyfinance.com.
Retrieved April 14, 2012, from http://www.dailyfinance.com.
Siomkos, G., & Shrivastava, P. (1993). Responding to product liability crises. Long Range
Planning, 26(5), 72–79.
Sims, C. A., & Zha, T. (2003). Error bands for impulse responses. Econometrica, 67, 1113–
1155.
Snyder, R. (1992). Comparative advertising and brand evaluation: Toward developing a
categorization approach. Journal of Consumer Psychology, 1, 15–30.
Snyder, C. R., Higgins, R. L., & Stucky, R. J. (1983). Excuses: Masquerades in search of grace.
New York, NY: Wiley.
Snyder, C. R., Lassegard, M. A., & Ford, C. E. (1986). Distancing after group success and
failure: Basking in reflected glory and cutting off reflected failure. Journal of Personality
and Social Psychology, 51, 382-388.
Sorescu, A., Shankar, V., & Kushwaha, T. (2007). New product preannouncements and
shareholder value: Don’t make promises you can’t keep. Journal of Marketing Research,
44, 468–489.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 111
Strapparava, C., & Valitutti, A. (2004). WordNet-Affect: an affective extension of WordNet. In
Proceedings of LREC (Vol. 4, pp. 1083–1086). Lisbon, Portugal: ELRA.
Strapparava, C., & Mihalcea, R. (2007). Semeval-2007 task 14: Affective text. Proceedings of
the 4th International Workshop on Semantic Evaluations (pp. 70-74). Prague, Czech
Republic: Association for Computational Linguistics.
Srinivasan, S., & Hanssens, D. (2009). Marketing and firm value: Metrics, methods, findings,
and future directions. Journal of Marketing Research, 46, 293-312.
Sujan, M., & Dekleva, C. (1987). Product categorization and inference making: Some
implications for comparative advertising. Journal of Consumer Research, 14, 372–378.
Stephen, A. T., & Galak, J. (2012). The effects of traditional and social earned media on sales: A
study of a microlending marketplace. Journal of Marketing Research, (49), 624-639.
Surowiecki, J. (2004). The wisdom of crowds. New York, NY: Knopf Doubleday Publishing
Group.
Tellis, G. J., & Franses, P. H. (2006). Optimal data interval for estimating advertising response.
Marketing Science, 25, 217–229.
Tellis, G. J., & Johnson, J. (2007). The value of quality. Marketing Science, 26, 758–773.
Tetlock, P. C. (2007). Giving content to investor sentiment: The role of media in the stock
market. The Journal of Finance, 62, 1139–1168.
Tirunillai, S., & Tellis, G. J. (2012). Does chatter really matter? Dynamics of user-generated
content and stock performance. Marketing Science, 31, 198–215.
Treleven, M., & Bergman Schweikhart, S. (1988). A risk/benefit analysis of sourcing strategies:
Single vs. multiple sourcing. Journal of Operations Management, 7(3-4), 93–114.
Trusov, M., Bucklin, R. E., & Pauwels, K. (2008). Effects of word-of-mouth versus traditional
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 112
marketing: Findings from an internet social networking site. Journal of Marketing, 73(5),
90-102.
Tversky, A. (1977). Features of similarity. Psychological Review, 84, 327-352.
Van Auken, S., & Adams, A. J. (1998). Attribute upgrading through across-class, within-
category comparison advertising. Journal of Advertising Research, 38, 6–16.
Van Heerde, H., Helsen, K., & Dekimpe, M. G. (2007). The impact of a product-harm crisis on
marketing effectiveness. Marketing Science, 26, 230–245.
Washington, M., & Zajac, E. J. (2005). Status evolution and competition: Theory and evidence.
Academy of Management Journal, 48, 282–296.
Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics Bulletin, 1, 80–83.
Williamson, D.A., (2010). Social media in the marketing mix: Budgeting for 2011. New York,
NY: eMarketer.
Wyer, R. S. (1973). Category ratings as“ subjective expected values”: Implications for attitude
formation and change. Psychological Review, 80, 446-467.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 113
Appendix A: Classification Accuracy of Support Vector Machine Algorithm of Chapter 2
I examine the accuracy of the Support Vector Machine classification algorithm using
three different measures. I first examine the “confusion matrix” to test for statistical significance
(Das & Chen, 2007). A confusion matrix presents the number of correct and incorrect predictions
made by the model compared with the labeled classifications in the test data. The matrix has 3
rows and 3 columns, where 3 is the number of classes. The rows represent the valence (positive,
negative, and neutral) of the tweets and the columns show how many of these tweets were
classified in each class: positive, negative, and neutral. The greater the weight of the diagonal in
the confusion matrix, the lesser is the confusion. Our null hypothesis is that the Support Vector
Machine algorithm has no classification ability for the test data. This implies that the rows and
columns of the matrix are independent of each other. I use the Chi-Square test to check for
statistical significance.
I next use the “precision” measure. Precision for the positive class is the number of true
positives (i.e. number of tweets correctly predicted as positive) divided by the total number of
tweets predicted as positive (i.e. the sum of true positives and false positives, which are tweets
incorrectly predicted as belonging to the positive class). This measure can be seen as a measure
of exactness. Since, I have three classes (positive, neutral and negative) I calculate precisions for
each class. I use the confusion matrix below to elucidate the formula.
Table A1
Illustration of Confusion Matrix
Predicted class
Labeled class
A B C
A TruePos
A
Error
AB
Error
AC
B Error
BA
TruePos
B
Error
BC
C Error
CA
Error
CB
TruePos
C
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 114
In the matrix above,
Precision for class A =
) (
CA BA A
A
Error Error TruePos
TruePos
+ +
(A1)
I next use the “recall” measure. Recall is defined as the number of true positives divided
by the total number of tweets that are labeled as positive (i.e. the sum of true positives and false
negatives, which are tweets that were not predicted as positive but should have been). This
measure can be seen as a measure of completeness. Similar to the precision measure, I calculate
recall for each class.
Using the matrix above,
Recall for class A =
) (
AC AB A
A
Error Error TruePos
TruePos
+ +
(A2)
The Chi-Square test indicates that the null hypothesis that the Support Vector Machine
algorithm has no classification ability for the test data can be safely rejected (p-value<.01). The
values of the precision and recall are in Table A2.
Table A2
Classification Accuracy of Support Vector Machine Algorithm
Positive Neutral Negative Average across classes
Precision 74.2% 75.5% 68.2% 72.7%
Recall 72.5% 81.0% 58.2% 70.3%
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 115
Appendix B: Descriptive Statistics for Chapter 2
Figure B1. Time series plot of Twitter metrics for Best Buy
Figure B2. Time series plot of Twitter metrics for Canon
Thanksgiving, 26
th
Nov
Thanksgiving, 26
th
Nov
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 116
Figure B3. Time series plot of Twitter metrics for Dell
Figure B4. Time series plot of Twitter metrics for iPhone
Thanksgiving, 26
th
Nov
Thanksgiving, 26
th
Nov
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 117
Figure B5. Time series plot of Twitter metrics for Kindle
Figure B6. Time series plot of Twitter metrics for PlayStation
Thanksgiving, 26
th
Nov
Thanksgiving, 26
th
Nov
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 118
Figure B7. Time series plot of Twitter metrics for Priceline
Figure B8. Time series plot of Twitter metrics for Sandisk
Thanksgiving, 26
th
Nov
Thanksgiving, 26
th
Nov
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 119
Figure B9. Time series plot of Twitter metrics for TiVo
Figure B10. Time series plot of Twitter metrics for Xbox
Thanksgiving, 26
th
Nov
Thanksgiving, 26
th
Nov
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 120
Table B1
Correlation of Key Variables
Stock returns Volume Retweet Positive to
negative ratio
Stock returns 1
Volume 0.12 1
Retweet 0.08 0.55 1
Positive to
negative ratio
0.03 0.03 0.19 1
Note. Significant values are in boldface.
Table B2
Descriptive Statistics of Key Variables
Stock returns Volume
Retweet Positive to
negative ratio
Mean Std. dev Mean Std. dev Mean Std. dev Mean Std. dev
Best Buy 0.02 1.46 160 113 18 15 4.0 2.8
Canon 0.04 1.33 86 58 29 40 12.3 15.5
Dell -0.08 1.79 406 207 116 144 4.3 3.1
iPhone 0.09 1.29 32,842 10,385 6,968 3,142 2.3 0.4
Kindle 0.24 3.05 2,409 1,773 430 438 2.9 1.0
PlayStation 0.15 1.61 1,654 1,178 263 368 2.8 1.7
Priceline 0.26 2.35 66 37 5 4 3.5 4.8
SanDisk 0.32 2.89 133 68 8 12 7.2 8.4
TiVo 0.39 5.99 407 294 54 44 1.8 0.5
Xbox 0.05 1.00 7,118 2,864 594 380 2.4 0.5
Total 0.15 2.67 4,528 10,266 848 2,287 4.3 6.7
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 121
Appendix C: Model Results by Brand of Chapter 2
Table C1
Granger Causality Tests by Brand of Twitter Metrics “Granger Causing” Stock Returns
p-values for 25 lags
Volume Retweet Positive to
negative Ratio
Best Buy 0.04 0.00 0.00
Canon 0.00 0.00 0.00
Dell 0.00 0.04 0.00
IPhone 0.00 0.00 0.00
Kindle 0.00 0.00 0.01
PlayStation 0.00 0.00 0.00
Priceline 0.00 0.00 0.04
SanDisk 0.00 0.00 0.00
TiVo 0.00 0.00 0.00
Xbox 0.00 0.01 0.01
Average 0.00 0.01 0.01
% significant 100 100 100
Note. The values shown here are the p-values of the chi-square test
Table C2
Granger Causality Tests by Brand of Stock Returns “Granger Causing” Twitter Metrics
p-values for 25 lags
Volume Retweet Positive to
negative Ratio
Best Buy 0.00 0.00 0.00
Canon 0.00 0.00 0.00
Dell 0.00 0.00 0.00
IPhone 0.00 0.00 0.00
Kindle 0.10 0.01 0.12
PlayStation 0.00 0.00 0.00
Priceline 0.00 0.01 0.28
SanDisk 0.35 0.17 0.00
TiVo 0.00 0.03 0.01
Xbox 0.00 0.00 0.23
Average 0.05 0.02 0.06
% significant 80 90 70
Note. The values shown here are the p-values of the chi-square test
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 122
Table C3
Augmented Dickey Fuller Test by Brand (Critical Values)
Brands Stock
rturns
Volume Retweet Positive/
negative
ratio
Citations
of print
Influential
blogs
Consumer
discussions
Best Buy -13.00 -5.50 -9.45 -11.71 -8.29 -9.83 -8.75
Canon -14.69 -5.14 -3.74 -4.53 -8.97 -12.38 -13.23
Dell -14.53 -17.79* -14.92* -3.02 -8.69 -10.61 -12.08
iPhone -11.66 -5.01 -4.64 -7.27 -7.76 -3.16 -13.6
Kindle -14.00 -8.34 -9.01 -7.82 -7.56 -9.94 -4.36
PlayStation -13.97 -8.09 -10.22 -9.6 -8.39 -8.56 -8.21
Priceline -15.54 -6.96 -11.19 -13.27 -8.4 -8.86 -6.67
SanDisk -13.73 -6.54 -11.47 -13.26 -8.52 -5.74 -3.4
TiVo -12.47 -6.48 -9.04 -10.66 -4.88 -6.89 -7.44
Xbox -14.98 -5.54 -5.4 -7.13 -5.68 -10.73 -9.16
Note: All the values of ADF are significant at 5% levels. The critical value at 5% level is -3.44
Table C4
Short and Cumulative Impact of Twitter Metrics on Stock Returns (Basis Points)
Brands Volume Retweet Positive to negative ratio
Short-term
Cumulative Short-term
Cumulative Short-term
Cumulative
Best Buy 6.76 0.00 15.75 5.99 17.32 5.26
Canon 11.34 3.38 5.72 0.00 5.53 -10.51
Dell 9.22 7.91 17.75 17.69 17.14 37.41
iPhone 2.89 0.00 3.12 0.02 11.51 18.69
Kindle 35.74 68.07 28.99 56.45 -11.69 -22.23
PlayStation 14.48 5.66 17.89 13.68 16.72 7.44
Priceline 84.58 97.14 18.70 13.97 -11.89 -1.23
SanDisk 5.64 52.15 0.01 58.92 30.61 37.62
TiVo 140.40 510.57 41.86 235.87 -30.21 49.30
Xbox 13.18 21.85 14.44 24.47 0.36 6.52
Note: I first test the normality of the short-term and long-term impulse responses for each metric using the Shapiro-
Wilk test. If I reject the null hypothesis of normality, I use the Wilcoxon signed-rank test (Wilcoxon, 1945). Else, I
use the T-test when reporting the significance of the median impact.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 123
Table C5
Short and Cumulative Impact of Retweet and Other Media on Stock Returns (Basis Points)
Brands Retweet Citations of print Influential blogs Consumer
discussions
Short-
term
Cumulative Short-
term
Cumulative Short-
term
Cumulative Short-
term
Cumulative
Best Buy 13.1 1.7 -10.3 -10.1 -1.2 -10.1 3.4 0.9
Canon 9.4 0.2 -2.2 0.7 14.4 13.4 -3.4 -5.7
Dell 7.3 0.6 -10.2 -12.0 -5.0 -6.0 8.1 2.8
iPhone 0.2 0.1 15.6 1.2 9.3 0.2 7.9 24
Kindle 27.3 69.3 67 119.1 61.1 119.1 -4.6 -5
PlayStation 8.9 0.0 78 136.1 61.9 122.1 -9 -5
Priceline 16.7 17.5 20.1 18.7 8.7 2.1 28.9 42.1
SanDisk 1.1 62 37.9 23.8 60.7 33.7 22.5 32
TiVo 16.3 117 187.4 169.4 156 418 -36.4 127.1
Xbox 6.7 20 9.8 12.4 -1.0 9.1 -3.1 -3.9
Note: I first test the normality of the short-term and long-term impulse responses for each metric using the Shapiro-
Wilk test. If I reject the null hypothesis of normality, I use the Wilcoxon signed-rank test (Wilcoxon, 1945). Else, I
use the T-test when reporting the significance of the median impact.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 124
Appendix D: Details of Classification Algorithm of Chapter 3
The third party’s classification algorithm determines the valence of an online
conversation by focusing on the situation where an attribute is used. For example, the word
“long” can indicate positive or negative valence on a product attribute depending on the attribute.
An online conversation (e.g., review, blog post, etc.) sentence may contain many product
attributes
( )
m 1
a ,.... a and many valence words
( )
n 1
w ,..... w
. An example of an attribute is “safety”
and examples of valence words are “good”, “bad”, etc.
The objective is to determine the overall valence expressed on each attribute
i
a in an
online conversation. Below are the steps involved in the algorithm:
1. Use a list of around 7000 positive, negative, and situation dependent valence words,
including phrases and idioms.
2. Partition each online conversation (e.g., review) into sentences (e.g., sentence ) that
form the online conversation.
3. Partition each sentence using words such as “but” and phrases such as “except that”.
If attribute
i
a is in sentence part
k
s , the valence score of
i
a in sentence part
k
s is
determined by the valence value of the word
j
w divided by the distance between the
word
j
w and the attribute
i
a . Thus, words far away from the attribute are given low
weights while words close to the attribute are given high weights. The valence value
of a word is +1 for positive word and -1 for negative word.
4. All valence scores for each sentence segment
k
s that contain the attribute are summed
to arrive at an overall valence score for sentence s . All valence scores for each
sentence s in a review or blog or forum post are summed to arrive at an overall
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 125
valence for the attribute in the review or blog or forum post. If the final score is
positive (negative), then the valence on attribute
i
a in a conversation (review, blog,
forum) is positive (negative). It is neutral otherwise.
As for the algorithm’s classification accuracy, the precision (p) value is 0.92 and the
recall (r) value is 0.91. Suppose there are positively labeled and negatively labeled conversations
and the goal is to accurately predict the positive label. Precision for the positive label is the
number of “true positives” (e.g., number of positively labeled conversations correctly predicted
as belonging to the positive label) divided by the sum of “true positives” and “false positives”
(e.g., number of negatively labeled conversations incorrectly predicted as belonging to the
positive label). This measure can be seen as a measure of exactness.
Recall for the positive label is defined as the number of “true positives” (e.g., number of
positively labeled conversations correctly predicted as belonging to the positive label) divided by
the sum of “true positives” and “false negatives” (e.g., number of positively labeled
conversations that were predicted as belonging to the negative label). This measure can be seen
as a measure of completeness. The F-score (2pr/ (p+r)) for the classification is 0.91. The F-score
is a measure of accuracy, which is the harmonic mean of precision and recall.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 126
Appendix E: Descriptive Statistics of Chapter 3
Table E1
Descriptive Statistics
Mean Min Max Std. Dev.
Toyota concerns 6.61 0 147 16.33
Honda concerns 3.36 0 44 4.01
Nissan concerns 1.18 0 10 1.35
Chrysler concerns 1.75 0 15 2.11
Toyota citations of
print
20.41 0 362 53.63
Honda Citations of
print
2.86 0 110 9.15
Nissan Citations of
print
1.18 0 37 3.50
Chrysler Citations of
print
0.5 0 13 1.47
Table E2
Correlation of Key Variables
Toyota
concerns
Honda
concerns
Nissan
concerns
Chrysler
concerns
Toyota
print
Honda
print
Nissan
print
Chrysler
print
Toyota concerns 1
Honda concerns
0.688
Nissan concerns
0.626 0.510
Chrysler concerns
0.396 0.258 0.423
Toyota citations of
print
0.830 0.544 0.437 0.319
Honda citations of
print
0.347 0.578 0.152 0.063 0.342
Nissan citations of
print
0.110 0.129 0.194 0.065 0.120 .149
Chrysler citations of
print
0.432 0.276 0.385 0.222 0.313 0.075 .093
Note: Significant values are in boldface
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 127
Appendix F: Robustness Results of Chapter 3
Table F1
Reduced VARX Coefficient Matrix for Negative Sentiment among Japanese Brands – Cumulative
Effect from Negative Sentiment (Brand A) to Negative Sentiment (Brand B)
Toyota
negative
sentiment
Honda
negative
sentiment
Nissan
negative
sentiment
Toyota
negative
sentiment
0.97
0.22
0.01
Honda
negative
sentiment
0.23
1.23
0.38
Nissan
negative
sentiment
0.00
0.03
1.03
Note: Significant values are indicated in boldface. Red shading indicates perverse halo effects.
Table F2
Reduced VARX Coefficient Matrix for Negative Sentiment between Toyota and Chrysler –
Cumulative Effect from Negative Sentiment (Brand A) to Negative Sentiment (Brand B)
Toyota
negative
sentiment
Chrysler
negative
sentiment
Toyota negative
sentiment 0.96 -0.05
Chrysler negative
sentiment -0.37 1.06
Note: Significant values are indicated in boldface. Blue shading indicates reverse halo effects.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 128
Table F3
Reduced VARX Coefficient Matrix for Scaled Negative Sentiment among Japanese Brands –
Cumulative Effect from Scaled Negative Sentiment (Brand A) to Scaled Negative Sentiment
(Brand B)
Toyota scaled
negative
sentiment
Honda scaled
negative
sentiment
Nissan scaled
negative
sentiment
Toyota scaled
negative
sentiment
0.95
0.23
0.24
Honda scaled
negative
sentiment
0.26
1.22
0.44
Nissan scaled
negative
sentiment
0.00
0.03
1.03
Note: Significant values are indicated in boldface. Red shading indicates perverse halo effects.
Table F4
Reduced VARX Coefficient Matrix for Scaled Negative Sentiment between Toyota and Chrysler –
Cumulative Effect from Scaled Negative Sentiment (Brand A) to Scaled Negative Sentiment
(Brand B)
Toyota scaled
negative
sentiment
Chrysler
scaled negative
sentiment
Toyota scaled
negative sentiment 0.96 -0.05
Chrysler
scaled negative
sentiment
-0.37 1.06
Note: Significant values are indicated in boldface. Blue shading indicates reverse halo effects.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 129
Table F5
VARX Coefficient Matrix for Only Honda and Nissan – Cumulative Effect from Concerns (Brand
A) to Concerns (Brand B)
Honda
citations
of print
Nissan
citations
of print
Honda
concerns
Nissan
concerns
Honda
citations
of print
1.47
0.5
0.28
0.06
Nissan
citations
of print
0.72
1.19
0.08
0.07
Honda
concerns
2.75 0.32 1.32 0.18
Nissan
concerns
0.13 0.46 0.19 1.05
Note: Significant values are indicated in boldface. Red shading indicates perverse halo effects. I find same results
using scaled negative sentiment measure.
Table F6
Reduced VARX Coefficient Matrix for Luxury Brands – Cumulative Effect from Concerns (Brand
A) to Concerns (Brand B)
Lexus
concerns
Acura
concerns
Infiniti
concerns
Lexus
concerns
1.15
0.04
0.13
Acura
concerns
0.1
1.1
0.41
Infiniti
concerns
0.13
0.06
0.99
Note: Significant values are indicated in boldface. Red shading indicates perverse halo effects. I find same results
using scaled negative sentiment measure.
CONSUMER CONVERSATIONS IN SOCIAL MEDIA 130
Table F7
Reduced VARX Coefficient Matrix for Non-Luxury Brands – Cumulative Effect from Concerns
(Brand A) to Concerns (Brand B)
Toyota
concerns
Honda
concerns
Nissan
concerns
Toyota
concerns
1.04
0.14
0.01
Honda
concerns
0.34
1.15
0.31
Nissan
concerns
0.03
0.14
1.2
Note: Significant values are indicated in boldface. Red shading indicates perverse halo effects. I find same results
using scaled negative sentiment measure.
Abstract (if available)
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Understanding virality of YouTube video ads: dynamics, drivers, and effects
PDF
Manipulating consumer opinion on social media using trolls and influencers
PDF
Essays on the turning points of the product life cycle
PDF
Essays on commercial media and advertising
PDF
Essays on the luxury fashion market
PDF
Essays on the role of entry strategy and quality strategy in market and consumer response
PDF
Essays on understanding consumer contribution behaviors in the context of crowdfunding
PDF
Three essays in derivatives, trading and liquidity
PDF
Share repurchases: how important is market timing?
PDF
Essays on information and financial economics
PDF
Essays on consumer returns in online retail and sustainable operations
PDF
Optimizing penalized and constrained loss functions with applications to large-scale internet media selection
PDF
Essays on the effect of cognitive constraints on financial decision-making
PDF
Essays on consumer product evaluation and online shopping intermediaries
PDF
Two essays on financial econometrics
PDF
Essays on development and health economics: social media and education policy
PDF
Creation and influence of visual verbal communication: antecedents and consequences of photo-text similarity in consumer-generated communication
PDF
Essays in asset pricing
PDF
Essays on information, incentives and operational strategies
PDF
Creating brand evangelists in the 21st century: using brand engagement through social media to develop brand loyalty in teens
Asset Metadata
Creator
Borah, Abhishek
(author)
Core Title
Essays on consumer conversations in social media
School
Marshall School of Business
Degree
Doctor of Philosophy
Degree Program
Business Administration
Publication Date
04/03/2015
Defense Date
02/26/2013
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
Halo,natural language processing,OAI-PMH Harvest,online chatter,social media,text mining,time series,Twitter
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Tellis, Gerard J. (
committee chair
), Hoffman, Donna L. (
committee member
), Luo, Lan (
committee member
), Valente, Thomas W. (
committee member
)
Creator Email
a.borah@gmail.com,aborah@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c3-231971
Unique identifier
UC11294803
Identifier
usctheses-c3-231971 (legacy record id)
Legacy Identifier
etd-BorahAbhis-1511.pdf
Dmrecord
231971
Document Type
Dissertation
Rights
Borah, Abhishek
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA
Tags
Halo
natural language processing
online chatter
social media
text mining
time series
Twitter