mba china | business school china | executive mba - the...

This article was downloaded by: [207.41.189.74] On: 29 September 2015, At: 08:12Publisher: Institute for Operations Research and the Management Sciences (INFORMS)INFORMS is located in Maryland, USA

Marketing Science

Publication details, including instructions for authors and subscription information:http://pubsonline.informs.org

The Squeaky Wheel Gets the Grease—An EmpiricalAnalysis of Customer Voice and Firm Intervention onTwitterLiye Ma, Baohong Sun, Sunder Kekre

To cite this article:Liye Ma, Baohong Sun, Sunder Kekre (2015) The Squeaky Wheel Gets the Grease—An Empirical Analysis of Customer Voice andFirm Intervention on Twitter. Marketing Science 34(5):627-645. http://dx.doi.org/10.1287/mksc.2015.0912

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial useor systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisherapproval, unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitnessfor a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, orinclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, orsupport of claims made of that product, publication, or service.

Copyright © 2015, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, managementscience, and analytics.For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org

http://pubsonline.informs.org

http://dx.doi.org/10.1287/mksc.2015.0912

http://pubsonline.informs.org/page/terms-and-conditions

http://www.informs.org

Vol. 34, No. 5, September–October 2015, pp. 627–645ISSN 0732-2399 (print) � ISSN 1526-548X (online) http://dx.doi.org/10.1287/mksc.2015.0912

© 2015 INFORMS

The Squeaky Wheel Gets the Grease—An Empirical Analysis of Customer Voice

and Firm Intervention on Twitter

Liye MaRobert H. Smith School of Business, University of Maryland, College Park, Maryland 20742, [email protected]

Baohong SunCheung Kong Graduate School of Business, New York, New York 10022, [email protected]

Sunder KekreTepper School of Business, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213, [email protected]

Firms are increasingly engaging with customers on social media. Despite this heightened interest, guidancefor effective engagement is lacking. In this study, we investigate customers’ compliments and complaints

and firms’ service interventions on social media. We develop a dynamic choice model that explicitly accountsfor the evolutions of both customers’ voicing decisions and their relationships with the firm. Voices are drivenby both the customers’ underlying relationships and other factors such as redress seeking. We estimate themodel using a unique data set of customer voices and service interventions on Twitter. We find that redressseeking is a major driver of customer complaints, and although service intervention improves relationships, italso encourages more complaints later. Because of this dual effect, firms are likely to underestimate the returnson service intervention if measured using only voices. Furthermore, we find an “error-correction” effect incertain situations, where customers compliment or complain when others voice the opposite opinions. Finally,we characterize the distinct voicing tendencies in different relationship states, and show that uncovering theunderlying relationship states enables effective targeting. We are among the first to analyze individual customerlevel voice dynamics and to evaluate the effects of service intervention on social media.

Keywords : service intervention; social media; microblogging; word of mouth; customer relationship; hiddenMarkov model; choice model

History : Received November 6, 2012; accepted January 27, 2015. Gerard J. Tillis served as guest editor-in-chiefand associate editor for this article. Published online in Articles in Advance May 27, 2015.

1. IntroductionThe proliferation of Internet social media has cre-ated significant excitement among marketers. Of par-ticular interest are sentiment analysis and complaintmanagement. A recent industry report shows thatmore than two-thirds of firms are already using socialmedia, with marketing and service being the toptwo functions (CRM Magazine 2012).1 Recognizingthe openness of social media platforms and the vast-ness of their information content, firms come to thesewebsites to gauge customer perceptions. Increasingly,firms are also moving from passive listening to activeservice intervention. Companies such as Dell, Ver-izon, and Comcast all have storefronts on popular

1 http://www.destinationcrm.com/articles/Columns/Departments/The-Tipping-Point/Using-Social-Media-for-Customer-Service-81584.aspx (accessed November 2, 2012).

platforms such as Twitter, some with dedicated ser-vice personnel who reach out to address customercomplaints. With these interventions, firms seek tostem negative sentiments and improve customer rela-tionships. Such interest in service intervention hasspawned a subindustry that supplies the relevanttools: tracking services for automatically extractingrelevant customer messages, text-mining services forclassifying messages into compliments or complaints,and aggregation algorithms for summarizing over-all sentiments. Sentiment indexes are supplied byboth large service providers, such as IBM and Thom-son Reuters, and smaller specialized marketing firms,such as sentimentmetrics.com and monitter.com.

The literature has long established the importanceof maintaining good relationships with customers(e.g., Zeithaml et al. 1993, Rust and Chung 2006, Net-zer et al. 2008). Grounded in the theory of exit andvoice (Hirschman 1970), numerous studies show that

627

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

mailto:[email protected]



http://www.destinationcrm.com/articles/Columns/Departments/The-Tipping-Point/Using-Social-Media-for-Customer-Service-81584.aspx



Ma, Sun, and Kekre: Analysis of Customer Voice and Firm Intervention on Twitter628 Marketing Science 34(5), pp. 627–645, © 2015 INFORMS

Table 1 Positioning and Contributions of This Paper in Internet WOM Literature

Individual consumerTopic Underlying drivers Brand/product level dynamics level dynamics

Creation of Internet WOM Anderson (1998), Hennig-Thurau et al. (2004),Schlosser (2005), Dellarocas and Narayan (2006),Ying et al. (2006), Berger and Schwartz (2011),Albuquerque et al. (2012), Berger and Milkman(2012), Yang et al. (2012), Chen and Kirmani(2015), Lovett et al. (2013), Toubia and Stephen(2013), etc.

Li and Hitt (2008), Wu and Huberman (2008),Moe and Trusov (2011), Godes and Silva(2012), Moe and Schweidel (2012), Schweideland Moe (2014)

Online sentiment analysis Schweidel and Moe (2014)

Online complaint managementThis Paper

Effect of Internet WOM Godes and Mayzlin (2004), Chevalier and Mayzlin (2006), Liu (2006), Duan et al. (2008), Godes and Mayzlin (2009),Trusov et al. (2009), Chintagunta et al. (2010), Sonnier et al. (2011), Stephen and Galak (2012),Tirunillai and Tellis (2012), etc.

effectively addressing customer complaints is crucialfor good customer relationships (Fornell and Wern-erfelt 1987, Blodgett et al. 1993, Blodgett and Ander-son 2000, Knox and van Oest 2014). However, existingcomplaint management research does not address theunique aspects of social media. Despite powerful toolsat their disposal, firms have only a limited under-standing of the dynamics of customer sentiment andthe effect of managing complaints on social media.Many questions remain open. What drives customersto compliment or complain on social media websites?Do customers’ voices reflect their underlying relation-ships with the firm, or are they also driven by otherfactors, and if so, how should the underlying relation-ships be assessed? Do service interventions on socialmedia improve customer sentiment, or do other fac-tors complicate matters? With limited resources, howshould a firm target its service interventions? Thesequestions are crucial for effective sentiment manage-ment on social media.

Compliments and complaints are specific typesof Internet word of mouth (WOM). Internet WOMhas substantial implications for sales and other mar-keting outcomes (Godes and Mayzlin 2004, Cheva-lier and Mayzlin 2006, Liu 2006, Duan et al. 2008,Godes and Mayzlin 2009, Trusov et al. 2009, Chinta-gunta et al. 2010, Stephen and Galak 2012, Tirunillaiand Tellis 2012, Kumar et al. 2013, Gopinath et al.2014, Tirunillai and Tellis 2014). Meanwhile, numer-ous studies show that WOM is motivated by manyunderlying factors, such as self-enhancement, emo-tions, social considerations, images, economic incen-tives, etc. (Anderson 1998, Hennig-Thurau et al. 2004,Schlosser 2005, Albuquerque et al. 2012, Lovett et al.2013, Toubia and Stephen 2013), and is also affectedby external factors, such as product characteristics,quality, and content type (Dellarocas and Narayan2006, Berger and Schwartz 2011, Berger and Milk-man 2012, Lovett et al. 2013). One implication of this

stream of literature is that WOM reflects but may notaccurately represent customers’ underlying opinions.Although the literature is rapidly growing, few stud-ies have investigated the dynamics of online senti-ment (with the exception of Schweidel and Moe 2014),and to the best of our knowledge, none has investi-gated firms’ service intervention. Furthermore, exist-ing studies on WOM creation either do not addressthe dynamics of WOM, or investigate the dynam-ics only at the product level (Li and Hitt 2008, Wuand Huberman 2008, Godes and Silva 2012, Moeand Schweidel 2012).2 These studies often use prod-uct review data that are product centric, with lowsocial interactivity among reviewers (Schweidel andMoe 2014). In contrast, popular social media plat-forms such as Facebook and Twitter are user cen-tric, with much higher levels of social interaction. Toaccurately gauge sentiment and evaluate the effectof service intervention, we need to understand howa customer’s relationship and voice evolve in thesesocial environments and in response to the firm’sactions. In other words, we need a dynamic anal-ysis at the individual customer level that explicitlyaccounts for the underlying relationship. This height-ened need and the gap in the literature motivate ourstudy (Table 1).

In this study, we investigate how customers’ com-pliments and complaints on a microblogging site aredriven by their relationships with the firm and bysocial factors at the site, and how the firm’s ser-vice intervention affects customers’ voices and rela-tionships. Recognizing that many factors other thanunderlying opinions also affect voices, we develop adynamic model to explicitly account for both voic-ing decisions and underlying relationships. We model

2 Although certain studies account for individual customer levelheterogeneity (Berger and Schwartz 2011, Godes and Silva 2012,Moe and Schweidel 2012), they do not allow for repeated interac-tions between a customer and a firm.

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

Ma, Sun, and Kekre: Analysis of Customer Voice and Firm Intervention on TwitterMarketing Science 34(5), pp. 627–645, © 2015 INFORMS 629

three key drivers of voicing decisions. First, customersmay compliment or complain to satisfy their intrinsicneeds. Second, they may do so to serve a social func-tion, e.g., to share their opinions with others. Finally,they may complain to seek redress, hoping to get theissues resolved. Meanwhile, we model a customer’sunderlying relationship with the firm and its dynamicevolution using a hidden-Markov model (Netzer et al.2008, Li et al. 2011). The relationship is reflected inthe customer’s intrinsic tendency to compliment andcomplain, and it also moderates the effects of socialand service factors. The firm’s service interventionaffects both voices and relationships.

We use a unique panel data set obtained from a For-tune 500 company. The data set contains the history ofindividual customers’ complaints, compliments, andgeneral chatter on Twitter, over an 11-month period.It also contains information on the firm’s serviceinterventions. We identify several underlying rela-tionship states with distinct voicing tendencies andother behavioral traits. More importantly, we find twoopposite effects of service intervention: on one hand,it does improve the firm’s relationship with the cus-tomer. On the other hand, we find that redress seekingis a major driver of complaints and that service inter-vention encourages even more complaints in the future.A major implication of this “squeaky wheel gets thegrease” effect is that sentiment will be underestimatedif we look at only customer voices. Firms, thus, needto uncover the underlying relationship to accuratelyevaluate the effect of service intervention. Further-more, we find that messages from friends on the web-site (defined as other users on the website whomthe focal customer follows) affect both relationshipsand voicing decisions. Although positive voices fromfriends improve a customer’s relationship with thefirm, their effect on the customer’s voice is nuanced.A significant error-correction or differentiation effectexists in certain situations and a tendency to conformexists in others. Finally, we show that uncovering theunderlying relationship enables effective targeting ofservice interventions.

We contribute to the literature in the following ways.First, we are among the first to explicitly model thedynamics of complaint and compliment decisions atthe individual customer level. Second, we are the firstto empirically investigate firms’ service interventionson social media and to show the nuanced effects ofthese interventions. Third, we separate the observedvoice from the underlying relationship, which allowsus to analyze how customers’ voices are driven bythe underlying relationship but are also influenced bysocial and service factors on microblogging websites.For industry managers, our study offers a frameworkto uncover underlying customer relationships, and todiscover ways to target customers more effectively. All

of these contributions advance our understanding ofcustomer engagement on social media, a topic withrapidly growing importance.3

2. Industry Background and DataOur data are obtained from a Fortune 500 telecom-munications firm that provides telecommunications,Internet, and wireless services, and that wishes toremain anonymous. With the growth of social media,customers increasingly post their comments about thefirm’s products and services online. The firm recog-nizes the importance of proactive customer engage-ment on social media. It uses a third-party trackingservice to automatically extract relevant customermessages from popular social media websites, suchas Twitter, Facebook, and CNET. Approximately 75%of the messages come from Twitter, which our studyfocuses on.4 Twitter is one of the most popularmicroblogging sites on the Internet. Created in 2006,it had more than 200 million active monthly usersin 2014. Users send messages that do not exceed140 characters in length, called “tweets.” Users canalso subscribe to other users’ tweets, known as“following” a user. These “following” links form adirected social network (in contrast to other popularsites, such as Facebook or LinkedIn, which have undi-rected networks) through which messages may prop-agate. The website has attracted great attention fromindustry and academia alike (see, e.g., Jansen et al.2009, Smith et al. 2012, Toubia and Stephen 2013).Twitter has been shown to host more brand-centralcontent than Facebook or YouTube (Smith et al. 2012).Among brand-related tweets, a great majority areneutral in sentiment, and in almost half, the brandis not the primary focus (Jansen et al. 2009). Senti-ment on Twitter can vary widely across brands andcan be quite negative (Smith et al. 2012). Further-more, user activities at the site are shown to be moti-vated by image-based utility more than by intrin-sic utility (Toubia and Stephen 2013). Given Twitter’s

3 Another paper that investigated sentiment analysis is Schwei-del and Moe (2014). Their approach controls for content, venue/audience, and self-selection of customers to venues. They derive abrand sentiment metric that correlates well with an offline brandtracking survey. Both their study and our study use latent con-structs to capture customer sentiment, although our study differsfrom theirs in a few notable aspects. First, we focus on customers’sentiment and voices at the individual customer level. Second, weexplicitly incorporate social influence on the website. Finally, andmost important, we investigate the effect of the firm’s service inter-vention on customer relationship and customer voice. We do notethat by accounting for multiple venues, Schweidel and Moe (2014)provide a broader view of social media than our study.4 Other than Twitter, the only website with a significant share isFacebook, accounting for 8.9% of messages. No other individualwebsite accounts for more than 1% of the messages. We do nothave social network data from Facebook.

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 2 Sample Messages in the Data Set

ComplimentsI love �the name of the firm’s service� its the bestThe Droid on �the firm’s name� is an excellent deterrent to being approached by �a competitor’s name� kiosk repsBut if you do want to switch consider �the firm’s name�, �the firm’s Twitter ID� is extremely responsive :D�the name of the firm’s service� is KICK A$$ GOOD! Even Iron Man good! ;)Tweeting from �the name of the firm’s service� TV. Love it!�the firm’s name� leads carriers in customer satisfaction, �a URL�

ComplaintsYour �a channel delivered through the firm’s service� feed is failing miserably! FIX IT PLEASE!! Missing the gameCan you find out why �a channel delivered through the firm’s service� fails every Caps game (it’s a �a channel delivered through the firm’s service� not

a �the name of the firm’s service� Issue, but still 0 0 0)I can’t get into �the firm’s name�.com/�the name of the firm’s service�. It tells me to sign up, I already have �the name of the firm’s service�. �the firm’s

name� site is a mess.�the firm’s name�, are you telling me that you can’t get someone to go to the next room at your office until Thursday?�the firm’s name�, I need to talk to techs at the #Morgantown office. Your CSRs refuse to do that. Unacceptable.If my modem had failed, I could just dash off to Best Buy for a replacement, but �the firm’s name� can’t even tell me that.

Neutral messages�the firm’s name� opens up �the name of the firm’s product� pre-orders �a URL�

where can i find free �the name of the firm’s product�??Breaking newS:: Jim Furyk Tops Brian Davis in Playoff �the name of the firm’s service� �a URL�

Today yankees vs. texas in the bronx and the game at 1:05 p.m. and is only on �the name of the firm’s service� 76 on io is 70�two phone products� appear in �the firm’s name� database?: What’s this? Two devices apparently of 0 0 0�The firm’s name� narrows 4G launch window down to �a date�

unique position in social media, research related tothis microblogging site is rapidly growing.

The tracking service uses a text-mining algorithmto classify each extracted message as negative, pos-itive, or neutral. A positive message is typically acompliment of the firm’s product or service, such as“�firm’s service name� was successfully installed, now myhome internet is blazingly fast.” A negative messageis typically a complaint, such as “the customer servicerep keeps passing the buck to the telesales rep and viceversa.”5 Table 2 lists examples of the messages con-tained in our data set. The majority of the messagesare classified as neutral (consistent with the findingsof Jansen et al. 2009). A team of service agents at thetelecom company focuses on customer engagementon social media. When a complaint is routed to thecompany by the tracking service, an agent tries torespond. The response may also be posted on the cor-responding website. Each instance of intervention isrecorded internally as a case. Our data set contains theinformation on these cases, including whether a com-plaint was responded to, the name of the agent whoresponded, and, in certain cases, the message content.

5 Automatic sentiment classification is common in the industry,given the high volume of data involved. Its use is also increasingin academic research (e.g., Sonnier et al. 2011 and Tirunillai andTellis 2012). According to the firm that provided the data, they alsoadded a manual step—if the agent at the firm did not believe amessage was classified correctly, the agent would manually correctit. We further verified the classification using a publically availabletool: sentiment140 (http://help.sentiment140.com/api). Classifica-tion using this tool matches those given by the firm for 84.7% ofpositive messages and 89.6% of negative messages, indicating ahigh degree of consistency.

About half of the customer complaints on the web-site were responded to (i.e., received service interven-tion). Given the fast pace of social media, complaintswere usually responded to on the same day they wereposted, often sooner.

2.1. Summary StatisticsThe data set contains all of the messages relevant tothe firm posted by customers on Twitter from Febru-ary 2010 to December 2010. For each message, weobserve the date it was posted, the account name ofthe customer, the message content, and the sentimentclassification. Also contained in the data set are thefirm’s service intervention records. We further aug-mented the data set by downloading the social net-work structure from Twitter. We refer to those whomthese customers follow on Twitter as their friends, andthose who follow these customers as their followers.We downloaded the list of friends of a randomlyselected subset of customers. This network informa-tion allows us to infer the messages each particu-lar customer’s friends sent, to which the customer ispotentially exposed.6

6 We used stratified random sampling. Specifically, 50% of cus-tomers in the sample are randomly selected from those who havevoiced five or more times in our data set, whereas the other 50%are from those who have voiced fewer than five times. For the lattergroup, an equal proportion is selected from those who voiced one,two, three, or four times. We performed disproportional insteadof proportional stratified sampling, both because the portion ofconsumers who voiced only one or two times is too high (93%combined), and because literature that focuses on individual leveldynamics tends to focus on “heavy users,” who provide more lon-gitudinal information for analysis (e.g., Erdem and Keane 1996).

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

http://help.sentiment140.com/api


Table 3 Summary Statistics by Customer

Mean SD Min Max

Number of compliments 2.47 3.51 0 27Number of complaints 3.05 2.68 0 20Number of neutral messages 30.21 26.41 0 234Number of interventions 1.42 1.57 0 14Number of compliments sent by 31.07 44.2 0 417

friendsNumber of complaints sent by 37.39 55.72 0 561

friendsMessage ratioa 0.465 0.165 0 1Number of followers 3,156.58 50,894.46 0 113551922Number of friends 1,030.94 1,422.49 0 51000Total customers 714Number of days 310

aMessage ratio is the ratio of compliments sent by friends (i.e., mes-sage ratio = compliments sent by friends/(compliments sent by friends +

complaints sent by friends)).

The sample data set used for our estimation con-tains the three types of messages of 714 customersover 310 days, and the messages each customer’sfriends sent on each day.7 The summary statistics arereported in Table 3. On average, customers postedslightly more complaints than compliments (3.05 ver-sus 2.47), as did their friends (37.39 versus 31.07).Note that this is markedly different from onlineproduct reviews, where ratings are typically positive(Chevalier and Mayzlin 2006). The ratio of compli-ments over the sum of compliments and complaintsis 0.465. The average number of followers and friendsper customer are 3,157 and 1,031, respectively. As istypical in social network data, both numbers are pos-itively skewed. About half of the complaints receivedservice intervention. According to the firm, slightpreference was given to customers with more fol-lowers. In the data set, those with an above-averagenumber of followers had 52% of their complaintsresponded to. In contrast, the response ratio for thosewith a below-average number of followers was 45%.

Table 4 reports the variables’ pairwise correlations.The number of compliments is slightly positively cor-related with that of complaints. The number is alsopositively correlated with the numbers of friends’compliments and complaints. This is preliminary evi-dence that customers’ voices are related to friends’messages. The number of complaints voiced by a cus-tomer is not correlated with friends’ messages, how-ever. This suggests either that complaints are lesssubject to network influence or that such influence

7 The data set provided by the firm contains more customers. How-ever, downloading the network data from Twitter for all of themwould take a prohibitively long time. Thus, our estimation dataset contains only a subset of these customers. Our original strat-ified sample yielded 1,000 customers. Because of technical issues,we could extract the network links for only 714 of them.

is more nuanced. Complaints and interventions arepositively correlated, as the former triggers the lat-ter. Interestingly, the number of compliments is alsoslightly positively correlated with interventions, indi-cating that service intervention possibly leads to morepositive customer voices. In summary, these statis-tics suggest that customers’ voicing behavior on socialmedia is closely related to that of friends and to firms’service interventions, which calls for a more detailedinvestigation of their interactions.

3. Model3.1. Theoretical FoundationWe first discuss the theoretical foundation of ourmodel. The key to our study is to recognize cus-tomers’ voices and underlying relationships as twodistinct constructs. We explicitly model the dynamicevolutions of both. Our modeling of the customer’svoicing decisions is grounded in the WOM creationliterature, and our modeling of the customer’s under-lying relationship with the firm draws from thecustomer relationship management (CRM) literature.The conceptual diagram of our model is shown inFigure 1.

Relationship0 Building and maintaining good rela-tionships with customers leads to stronger loyalty,more purchases, and higher customer lifetime value(e.g., Zeithaml et al. 1993, Rust and Chung 2006).Firms’ eagerness to manage sentiment on socialmedia is also driven by their quest for good customerrelationships. Conceptually, customers form their per-ceptions of (and relationships with) the firm based ontheir past direct and indirect experiences. These rela-tionships are often characterized using discrete statesthat evolve over time (Dwyer et al. 1987, Fournier1998, Aaker et al. 2004, Luo and Kumar 2013). Statetransitions may be triggered by interactions betweenfirms and customers (Netzer et al. 2008, Li et al.2011). The complaint management literature showsthat firms’ service interventions also shape customers’relationships (Blodgett et al. 1993, 1995). Furthermore,social media websites are unique in that they provideeasy access to others’ opinions, which also may affectcustomers’ relationships with firms.

Accordingly, in our model, customers form theirperceptions about the firm based on three sources ofinformation: their own experience of the firm’s prod-ucts and services, the firm’s responses to their com-plaints, and their exposure to the opinions expressedby others. First, relationships may change over time,depending on the customers’ own experiences. Thisappears random in the eyes of the researcher, as suchexperience is typically not observed. Second, afterlearning friends’ opinions about the firm, they may

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 4 Correlation Matrix of Observed Variables

Neutral Friends’ Friends’ MessageCompliments Complaints messages Interventions compliments complaints ratio ln(Followers + 1)

ComplimentsComplaints 00168Neutral messages 00409 −00005Interventions 00132 00653 −00025Friends’ compliments 00155 00040 00075 00065Friends’ complaints 00185 −00008 00118 00047 00953Message ratio −00079 00160 −00104 00038 00030 −00108ln(Followers + 1) 00009 00044 −00018 00072 00279 00236 −00031ln(Friends + 1) −00006 00058 −00039 00063 00544 00506 00006 00389

Figure 1 (Color online) Conceptual Framework

Voicedecision at t

Relationshipstate at t

t t + 1

Relationshipstate at t + 1

Selfexperience

Others’opinion

Firmintervention

Relationship

Voice Responseto others

Redressseeking

Voice decisionat t + 1

… …

Tendencyto talk

Evolution of relationship

Sources of discrepancy between voice and relationship

change their own perceptions of the firm. For exam-ple, hearing compliments may make a customer per-ceive more favorably of the firm. Finally, service inter-ventions may also change customers’ perceptions. Forinstance, if a firm intervenes every time a complaint ismade, customers may improve their perceptions anddevelop more favorable opinions of the firm.

Compliments and Complaints on Social Media0 Ourmodel accounts for three factors—intrinsic, social, andservice—that drive customer voices. First, customersmay compliment a firm or complain about it onsocial media to satisfy their intrinsic psychologicalneeds (Anderson 1998). Emotional desires, concernsfor others, self-enhancement, and the like all moti-vate customers to voice their opinions (Feick and Price1987, Hennig-Thurau et al. 2004, Toubia and Stephen2013). Our first utility component, the intrinsic util-ity, accounts for these factors. This intrinsic utility willreflect and be driven by customers’ underlying rela-tionships with the firm. The intrinsic utility also variesfrom person to person (Berger and Schwartz 2011,

Moe and Schweidel 2012, Toubia and Stephen 2013).For example, some customers may simply be moretalkative than others, and some customers may have astronger tendency than others to voice negative opin-ions (Goffman 1959, Richins 1984).

Second, as social media exposes customers to oth-ers’ voices, one customer’s voice may depend on whatothers have said (Schlosser 2005, Wu and Huberman2008, Godes and Silva 2012, Moe and Schweidel 2012).Customers may be more likely to voice their opin-ions if those opinions differ markedly from what oth-ers have said. In this case, they are “correcting” oth-ers’ mistakes or differentiating from others (Wu andHuberman 2008, Moe and Schweidel 2012). Alterna-tively, customers may be more likely to voice opinionssimilar to what others have said. In this case, theyare “conforming” to their social environment (Moeand Schweidel 2012). Our second utility component,social utility, captures such effects. Terminologies varyacross studies on WOM creation. Both the intrinsicand social utilities in our model can be mapped to

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


multiple factors in the existing literature.8 The key dis-tinction between intrinsic and social utility is that theformer depends on only the customers’ own charac-teristics, and the latter depends on others’ voices.

Furthermore, we include a third component, serviceutility, to account for customers’ intentions to seekredress. Literature shows that customers often com-plain in order to receive services from firms (Dayand Landon Jr 1976). Similarly, customers may sim-ply complain on social media in order to reach outto the firm, hoping to get their issues resolved. Thisservice utility is expected to depend on whether thefirm addressed the customer’s previous complaints.For example, if a firm consistently addressed a cus-tomer’s past complaints on social media, the customermay develop higher expectations and become morewilling to complain in the future.

Relationship and Voices. Customers’ relationshipswith a firm and their voices on social media are tworelated but distinct constructs. The underlying rela-tionship affects voices in several ways. The intrin-sic utility is primarily determined by and reflectsthe underlying relationship. To illustrate, customerswho have good relationships with the firm will havehigher intrinsic utility to compliment than to com-plain. The relationship can also moderate social andservice utility. Customers who do not have good rela-tionships with the firm, for instance, might not expectthat the firm will address their complaints. In thiscase, the service utility would be low. In anotherexample, customers who do not have good relation-ships with the firm may be likely to “correct” others’compliments. Accordingly, in our model the intrinsicutility reflects the customer’s underlying relationshipwith the firm, which also moderates the other twoutility components.

Effect of Service Intervention0 Modeling both rela-tionships and voicing decisions, we can analyze thepotentially opposing effects of a firm’s service inter-ventions. On one hand, by addressing customers’complaints, service interventions should improve thecustomers’ relationships with the firm (Hirschman1970, Fornell and Wernerfelt 1987, Blodgett et al.1995). On the other hand, a higher perceived like-lihood of success has been shown to encourage

8 Toubia and Stephen (2013), for example, compare intrinsic andimage-based utility. Hennig-Thurau et al. (2004) provide a typol-ogy of four factors: social interaction, economic incentive, altruism,and self-enhancement. Given our observational data, we would notknow, for example, whether a complaint is generated because of adesire to self-enhance or because of a concern for others, becausein both cases the customer may complain when others have not.Our data allow us to separate only the utility that depends on thesocial environment from the utility that does not.

complaints (Day and Landon Jr 1976). By raising cus-tomers’ expectations, a firm’s service interventionsmay very well encourage customers to complain morein the future. Service intervention, thus, may have apositive effect on the underlying customer relation-ships but a negative direct effect on the customers’voices. Because relationship underlies voices, serviceintervention also affects voices indirectly through therelationship. The net effect of service intervention onvoices, therefore, is not obvious ex ante.

In summary, our model builds on the WOM andCRM literature, and recognizes customers’ underly-ing relationships with the firm and their voices astwo distinct constructs. The dynamic evolutions ofboth of these constructs are explicitly accounted forin our model. We further account for the effects ofthe social media environment and the firm’s serviceinterventions, which may lead to a difference betweenrelationship and voice. Sentiment management wouldbe straightforward if customer voices strictly reflectedthe underlying relationship. However, the intrinsic,social, and service factors that also drive voicesmay have distorting effects. For example, pressureto conform might inhibit an angry customer’s com-plaint. The desire to seek redress, as another exam-ple, may induce a reasonably happy customer to com-plain more. By explicitly modeling relationship andvoice, we can probe beneath surface level observa-tions to better understand the underlying relation-ship and the nuanced effects of a firm’s serviceinterventions.

3.2. Voicing Decision and UtilityFormally, there are I customers of a firm. Time is dis-crete and indexed by t, t = 1121 0 0 0 1 T . Customers areconnected on the microblogging site Twitter. As statedin §2, we call the customers who follow a focal cus-tomer the customer’s followers, and those whom thecustomer follows the customer’s friends. In each timeperiod, a customer decides whether to voice a posi-tive, neutral, or negative message, or not to say any-thing about the firm. The decision of customer i attime t is denoted as

Dit =

3 voices a positive message2 voices a neutral message1 voices a negative message0 does not voice a message

(1)

We use a latent utility approach: At time t, cus-tomer i’s utilities of voicing a positive message (i.e., a

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


compliment), a neutral message, a negative message(i.e., a complaint), or saying nothing, are

Uit =

Ui3t = Ui3t + �i3t = �i +�i34Sit5+�i34Sit5NSit

+�i3t Dit = 3Ui2t = Ui2t + �i2t = �i + �i2t Dit = 2Ui1t = Ui1t + �i1t = �i +�i14Sit5+�i14Sit5N

Sit

+ �i4Sit5Iit + �i1t Dit = 1Ui0t = 0 + �i0t Dit = 00

(2)In Equation (2), the first two terms are the intrinsiccomponents. The baseline term, �i, enters the util-ity of all three voice types. It captures the intrinsic“talkativeness”—the desire to make the customer’svoice heard. The terms �i34Sit5 and �i14Sit5 account forthe intrinsic desires to compliment and complain (rel-ative to voicing a neutral message). They depend onthe customer’s underlying relationship with the firmat the time, denoted as Sit . A customer who holdsthe firm in high regard, for instance, would be moreapt to compliment the firm (i.e., �i34Sit5 would have ahigher value) than would another customer. We dis-cuss the modeling and evolution of Sit in §3.3.

The terms �i34Sit5NSit and �i14Sit5N

Sit in Equation (2)

represent the social utility of complimenting and com-plaining, respectively. The variable N S

it denotes thenetwork sentiment, which represents the sentiment con-tained in the messages the customer hears fromfriends on the website. Higher N S

it indicates a morepositive sentiment. We discuss the construction ofthis variable in §3.4. The coefficients �i34Sit5 and�i14Sit5 measure how network sentiment affects thecustomer’s voice. A positive coefficient �i34Sit5 meansthat the customer is more likely to compliment thefirm when the network sentiment is more positive.In other words, the customer tends to conform towhat others say. If the coefficient is negative, thenthe customer tends to “correct” others’ mistakes, or todifferentiate from others’ opinions. Similarly, a posi-tive �i14Sit5 means that the customer is more likely tocomplain when the network sentiment is more pos-itive, i.e., the customer tends to differentiate. Net-work sentiments do not necessarily affect compli-ments and complaints to the same extent. Havingthe two parameters allows for asymmetric effects. Forexample, �i34Sit5 < 0 and �i14Sit5 = 0 mean that hear-ing more positive messages makes a customer lesslikely to compliment but does not change the propen-sity to complain. This differs from the case where italso makes the customer more likely to complain, i.e.,�i14Sit5 > 0, even though in both cases the customerdemonstrates the error-correction tendency.

The fourth term of the utility of complaining,�i4Sit5Iit , is the service component. Customers maycomplain to get their issues resolved. This component

represents the expected utility from the firm’s serviceintervention. This expected utility should depend onhow often the firm has helped the customer in thepast. The variable Iit is the percentage of the cus-tomer’s past complaints to which the firm responded

Iit =

∑t−1�=1 I8Ii� = 19

∑t−1�=1 I8Di� = 19

0 (2a)

In Equation (2a), Ii� denotes the firm’s interven-tion decision: Ii� = 1 if the firm provided service inresponse to the complaint of customer i at time tand Ii� = 0 otherwise. To the customer, this variablereflects the likelihood, based on past experience, ofthe firm addressing the complaints. A positive coeffi-cient �i4Sit5 indicates that the more a customer’s com-plaints were addressed in the past, the more the cus-tomer expects to be helped, and the more prone tocomplain the customer becomes.

Finally, the mean utility of not saying anything isnormalized to zero. We assume that the error terms inEquation (2) follow i.i.d. extreme value distribution,leading to the logit choice probability

Pr4Dit = 35= exp4Ui3t5/(

∑3d=1 exp4Uidt5+ 1

)



)



)

Pr4Dit = 05= 1/(


)

0

(3)

We note that existing studies have used two-stage models, incidence and evaluation, to charac-terize WOM decisions (e.g., Ying et al. 2006, Moeand Schweidel 2012). A two-stage model can cap-ture empirical regularities such as the U -shapedresponse function of product reviews. However, sinceour model is state dependent, it is flexible enoughto admit such patterns. Furthermore, the intrinsictalkativeness parameter, �i, can be considered as rep-resenting the “incidence” aspect of WOM, i.e., howmuch a customer is generally prone to sending a mes-sage on Twitter. Because our focus is on customercomplaints and compliments, we use the more par-simonious multinomial model instead of a two-stagemodel. These two models are expected to yield simi-lar findings.

3.3. Relationship StatesWe model a customer’s underlying relationship withthe firm using a first-order discrete-time discrete-statehidden-Markov model (HMM) (e.g., Montgomeryet al. 2004, Du and Kamakura 2006, Moon et al. 2007,Netzer et al. 2008, Li et al. 2013). The state of cus-tomer i at time t is denoted as Sit , Sit ∈ 811 0 0 0 1K9,where K is the total number of states. An HMM isinvariant to permutation of states. For identification,

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


we order the states based on sentiment. The intrin-sic desire to complain is reflected by the coefficient�i14Sit5 in Equation (2). Thus, we order states from 1 toK, where the value of �i14Sit5 decreases with the state,with state 1 being the most negative. We account forthe evolution of relationship states using a customer-and time-specific state transition matrix, denoted as

Ait =

ait1111 0 0 0 ait111K0 0 0 0 0 0 0 0 0

ait1K11 0 0 0 ait1K1K

0 (4)

In Equation (4), ait1 s1 s′ is the probability that cus-tomer i transitions from state s at time t to state s′

at time t + 1. Since states are ordered from the mostnegative to the most positive, we model the statetransition using an ordered-logit model (Netzer et al.2008). To do so, we specify a set of threshold val-ues as boundaries between states and a link functionthat incorporate other factors that may influence statetransition. We denote these threshold values as �i1 ss′ ,where s ∈ 811 0 0 0 1K9 and s′ ∈ 811 0 0 0 1K−19, and denotethe dependent variable for the link function as yit . Thestate transition probabilities in Equation (4) can thenbe written as

ait1 s11 =exp4�i1 s1 − yit5

1 + exp4�i1 s1 − yit5

ait1 s1 s′ =exp4�i1 ss′ − yit5

1 + exp4�i1 ss′ − yit5−

exp4�i1 ss′−1 − yit5

1 + exp4�i1 ss−1′ − yit51

s′∈ 821 0 0 0 1K − 19

ait1 s1NS = 1 −exp4�i1 s1K−1 − yit5

1 + exp4�i1 s1K−1 − yit50

(5)Three factors drive the evolution of relationship

states: the customer’s own experience, the firm’s his-tory of service intervention, and the network senti-ment. These factors are incorporated in the link func-tion of the ordered-logit model

yit =�i0 +�i14Sit5Mit +�i24Sit5Iit + �t0 (6)

In Equation (6), �i0 represents the individual cus-tomer-specific tendency to transition among relation-ship states. Customers may differ in their generalexperiences with a firm’s product (e.g., quality willvary across different units of the same product) and,therefore, exhibit different state transition tendencies.Such factors are not observed by researchers, and arecaptured as unobserved heterogeneity using �i0. Thevariable Mit is the ratio of all positive messages in thepast. The variable Iit is the ratio of past firm inter-ventions, as defined in Equation (2a). The coefficients�i14Sit5 and �i24Sit5 are the coefficients on how friends’messages and the firm’s service interventions affect acustomer’s state transition. The coefficients are state

specific, as a factor may have different effects in dif-ferent states. The term �t in Equation (6) is a time-specific fixed effect that captures potential commonshocks. A major interruption in service, for instance,may make everyone more likely to move to a morenegative state. Finally, we denote the probability thata customer starts from state s as a0

s .

3.4. Network SentimentOur construction of the network sentiment variableN S

it closely follows standard industry practice anddraws on recent literature on online social networks.Sentiment monitoring on social media typically relieson a voice sentiment index, defined as a percentageof negative or positive messages. We denote the totalnumber of compliments and complaints voiced bycustomer i’s friends at t as mit = 8mP

i1 t1 mNi1 t9; then,

a straightforward construction of the network senti-ment variable would be the proportion of positivemessages at time t − 1

N Sit = mP

i1 t−1/4mPi1 t−1 + mN

i1 t−150 (7)

However, on a fast-paced microblogging site, usersreceive many messages every day and may be selec-tive in reading the messages. A customer may readcertain friends’ messages whereas routinely ignoringothers. Furthermore, although users may follow manyfriends, not all of these friends will equally influencethe focal user (Trusov et al. 2010).9 To account forthese potential differences in influence, we adopt themethodology of Trusov et al. (2010). We denote thefriends of customer i (i.e., whom customer i follows)as i11 i21 0 0 0 1 iGi, where Gi is the number of friendswhom customer i follows. We allow friends to havedifferent levels of influence on a focal customer, anddefine the influence-adjusted compliments and com-plaints actually received by the customer as follows:

mPit =

Gi∑

j=1

�ij I8Dij1 t = 39

mNit =

Gi∑

j=1

�ij I8Dij1 t = 190(8)

In Equation (8), I8 · 9 is the indicator function, and�ij represents friend j’s influence on customer i. Thevariables mP

it and mNit are, thus, the weighted sum of

compliments and complaints received by customer i,weighted by each friend’s influence level. Analogousto Equation (8), the network sentiment index is then

N Sit =mP

i1 t−1/4mPi1 t−1 +mN

i1 t−150 (9)

9 We thank an anonymous reviewer for this helpful suggestion.

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


We assume each dyad-specific �ij is an independentBernoulli random variable, where value 1 indicatesthat the friend is influential and 0 not

�ij ∼ Bernoulli4pi50 (10)

In Equation (10), pi is the probability that a friendis influential on customer i.

3.5. HeterogeneityOn social media websites in general and microblog-ging sites in particular, the number of followers acustomer has influences WOM behavior (Toubia andStephen 2013). For example, a customer with morefollowers may be more vocal. Furthermore, individ-uals also differ in their intrinsic voice tendencies,their responses to their social environments, and theirexpectations of the firm’s service interventions. Weadopt a hierarchical Bayesian framework to incorpo-rate both observed and unobserved heterogeneity inthe utility function. Each individual-specific parame-ter in the voicing utility function, generically denotedas �i, is modeled as

�i = �i0 + �1Fi0 (11a)

In Equation (11a), �i denotes each parameter inthe voicing utility function (�i, �i14Sit5, �i34Sit5,�i14Sit5, �i34Sit5, and �i4Sit55. The term �i0 is anindividual-specific intercept term. Fi = ln4Fi + 15 is thelog-transformed number of followers of the customer(Fi is the number of followers), and �1 captures thedifference across customers according to the numberof followers. A positive �1 indicates that the param-eter value is higher for those with more followers.To account for unobserved heterogeneity, we assumethat the individual-specific intercepts are drawn frompopulation level normal distributions

�i0 ∼N4�01�2�050 (11b)

Similarly, we account for unobserved heterogeneityacross customers on state transition. Each individual-specific parameter in the state transition equations,denoted as �i, is modeled as drawn from a population-level distribution

�i ∼N4�1�2� 50 (11c)

In Equation (11c), �i denotes each parameter inthe state transition equations (�i1 ss′ , �i0, �i14Sit5, and�i24Sit55. Note that only unobserved heterogeneity—not the observed heterogeneity on the number offollowers—is admitted to state transition parameters.This is because the number of followers is a char-acteristic specific to the microblogging website. It isexpected to affect voices on the website, but it is notclear whether the theoretical foundation exists for itto affect the underlying relationship.

3.6. Endogeneity, Identification, and EstimationAs pointed out by Manchanda et al. (2004), responsemodeling should account for the possibility of non-random marketing mix variables. We capture twopotential sources of such nonrandomness. First, it ispossible that service intervention decisions are madebased on characteristics of the customers who voicedcomplaints. Second, service intervention decisionsmay depend on time-specific factors. When there is aservice outage, for instance, both customer complaintsand the firm’s intervention efforts may intensify. Fol-lowing Manchanda et al. (2004), we model the inter-vention decision of the firm using a logistic regressionover the customer’s response parameters and time-specific effects

P4Iit = 1 �Dit = 15= exp4�it5/41 + exp4�it551 (12a)

where�it =ä′

i

⇀�+��t0 (12b)

In Equation (12b), äi is the vector of all individual-specific parameters (�i0 in Equation (11a), �i in Equa-tion (11c), and Fi5 and a constant term, and

⇀� is

the coefficient vector for the intervention decisionfunction. The term �t is the time-specific fixed effectthat accounts for common shocks, as shown in Equa-tion (6), and �� is the corresponding parameter.

For identification, we normalize the populationmean of �i0 in Equation (6) to 0.10 We estimate themodel using the Markov Chain Monte Carlo (MCMC)approach. More details of identification and estima-tion can be found in the online appendix (available assupplemental material at http://dx.doi.org/10.1287/mksc.2015.0912).

4. Empirical Result4.1. Model ComparisonWe first compare the model fit statistics of ourproposed model with those of several alternativeand benchmark models. The results are reported inTable 5. We use log-marginal density for model selec-tion (Newton and Raftery 1994, Chib 1995). We alsocalculate the overall hit rate, the hit rate when cus-tomers voiced messages, and the hit rate when cus-tomers complained or complimented, both in sampleand out of sample. We use the first 230 days of datafor estimation, and the remaining 80 days as the hold-out sample.

Model 1 is our proposed model. We estimated fouralternative specifications, from two to five states. Thethree-state version outperforms the others on log-marginal density, the out-of-sample log likelihood,

10 An equal change of �i0 and all �i1 ss′ will yield the same statetransition probabilities, hence the normalization.

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.




Table 5 Model Comparison

In sample Out of sample

Model LMD HR (%) HRV (%) HRCC (%) LL HR (%) HRV (%) HRCC (%)

Model 1–2 state −61118202 80035 15009 4098 −21143401 80060 14015 1071Model 1–3 state −60118605 80054 15057 6078 −21136703 80065 14017 1075Model 1–4 state −60158606 80032 15003 6040 −21143301 80055 13091 1073Model 1–5 state −60137200 80035 15018 7008 −21144309 80071 13098 1066Model 2 −62197705 80054 14037 2042 −21145402 80058 14023 1061Model 3 −61137709 80033 14090 6001 −21142005 80066 14004 1063Model 4 −60135309 80049 15042 5014 −21135204 80079 14002 1059Model 5 −61157405 80027 15003 5004 −21136506 80059 14014 1059Model 6 −63132505 79077 14042 1097 −21145504 80056 13016 1043Model 7 −63125409 80005 14046 1095 −21145308 80060 13024 1039Model 8 −63127304 79099 14044 1075 −21144007 80054 13016 1032

Note. LMD, Log marginal density; LL, log likelihood; HR, hit rate; HRV, hit rate for messages; HRCC, hit rate for compliments/complaints messages.

and the hit rate for compliments and complaints. Theoverall hit rate and the hit rate for voices are similaracross specifications. Considering these, we proceedwith the three-state version of our proposed model.

Model 2 is the “one-state” version of our model,which does not contain multiple underlying relation-ship states. It underperforms the proposed model(and also the two-state version) on log-marginal den-sity and on the hit rate for compliments and com-plaints. This confirms the importance of modelingthe underlying relationship states and the transitiondynamics. Models 3–5 are variations of the proposedmodel. They are estimated with three states to enabledirect comparison with the proposed model. Model 3makes state transitions conditional on the currentperiod network messages and service interventioninstead of the cumulative ones. It underperforms theproposed model. This suggests that state transitionis more gradual and cumulative than instantaneous.Model 4 excludes the service component from thevoicing utility function, and model 5 excludes bothsocial and service components. The proposed modeloutperforms model 4 on log-marginal density and thehit rate for compliments and complaints. Model 4also outperforms model 5. These suggest that bothsocial and service utilities are critical components thatshould be included. Models 6–8 are standard hierar-chical multinomial logit models that serve as outsidebenchmarks. Model 6 has only the intercept term andlinear term on the number of followers. Model 7 is anextension from model 6. It includes both the numberof complaints and the number of compliments heardby the customers in the utility functions. Model 8 isa variant of model 7, using the percentage of positivemessages in the utility function. The proposed modeloutperforms all of these benchmark models.11

11 All of the models have a similar overall hit rate and the hit ratefor voices. This is due to the nature of the data: in a given period,a customer is most likely not to say anything. When a customer

4.2. Parameter EstimatesVoice Utility0 We report the population mean esti-

mates of the voicing utility parameters in Table 6.The intercept of the intrinsic utility to voice a mes-sage of any type is −20532. This corresponds to onemessage about every 13 days, consistent with thedata. The intrinsic utilities of complimenting andcomplaining (�3 and �1 in Equation (2)) show thatcustomers are more likely to voice neutral messagesthan to voice either compliments or complaints—mostparameters are negative and statistically significant,with the utility to complain at the low state as theonly exception.12 Customers’ intrinsic likelihood tocomplain is almost 155 times the likelihood to compli-ment in the low state (0.284 versus −40760 for the twocoefficients). In contrast, customers in the higher stateare 61 times more likely to compliment than to com-plain (−00651 versus −40766). In the middle state, theintrinsic tendencies to compliment and to complainare similar. Considering these findings, we refer to thethree states more intuitively as the negative, neutral,and positive states.13 Customers have a higher desireto complain in the negative state than to compliment

does voice a message, it is more likely a neutral one than a com-pliment or complaint. Because our model focuses on complimentsand complaints, it is reasonable that these two hit rates do not varymuch across models. In contrast, the hit rates for compliments andcomplaints do differ significantly.12 We consider a coefficient “statistically significant” if the 95% cred-ible interval does not include zero, and two coefficients as statisti-cally different if their 95% credible intervals do not overlap. This isconsistent throughout the paper.13 Note that although the states are ordered according to their intrin-sic tendency to complain relative to voicing a neutral message, thisalone does not mean that the three states are necessarily negative,neutral, and positive. For example, if in all three states the intrin-sic tendency to complain is higher than it is to compliment, thenall should be referred to as negative states of different levels. Werefer to the three states as negative, neutral, and positive becausedoing so is consistent with the relative tendencies to complain andcompliment for each state.

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 6 Parameter Estimate—Voicing Utility

Voicing utility—Intercept

Intrinsic utility Relationship states

Voice −20532 (∗)

Negative Neutral Positive

Complaint (relative to neutral) 00284 4∗5 −30344 4∗5 −40766 4∗5

Compliment (relative to neutral) −40760 4∗5 −30223 4∗5 −00651 4∗5

Effect of network sentimentComplaint 00419 4∗5 −00073 00474 4∗5

Compliment −00106 4∗5 −00088 00426 4∗5

Effect of intervention 00291 00945 4∗5 10293 4∗5

∗The 95% credible interval does not include zero.

in the positive state. This suggests that the negativestate is more intense than the positive one, possiblybecause customers are prone to sending emotionalmessages when they have negative relationships withthe firm.

Looking at the social utility, we see that voic-ing decision is affected by friends’ messages in cer-tain situations. Customers in the negative state aremore likely to complain and less likely to compli-ment when they hear more positive messages fromfriends. (The coefficient for complaint is positive andfor compliment is negative, both statistically signif-icant.) This demonstrates a differentiating, or error-correcting, behavior in the negative state. In contrast,customers in the positive state demonstrate a mix ofconforming and differentiating tendencies. They aremore likely to compliment when others do, and alsomore likely to complain when hearing more compli-ments. (The coefficients for both complaint and com-pliment are positive and statistically significant.) Nosystematic conforming or differentiating tendency isevident for customers in the neutral state. (Neithercoefficient is statistically different from zero.) How-ever, such tendencies may still exist at the individualcustomer level, as the variance estimates reported inTable 7 show wide dispersions of these coefficientsacross customers.

The estimates of service utility coefficients also con-firm that redress seeking is a key driver of complaints.(The coefficients, also in Table 6, are positive for allthree states, 0.291, 0.945, and 1.293, and statisticallysignificant for the latter two.) Other things equal, acustomer’s past experience of receiving help from thefirm will encourage the customer to complain again.This suggests that customers will become more confi-dent of getting the firm’s help if it has responded totheir past complaints. This effect is more salient forthe positive and neutral states, suggesting that cus-tomers in those states have higher expectations of thefirm.

Table 7 Parameter Estimate—Voicing Utility Heterogeneity

Observed heterogeneity—Ln(No. of Followers + 1)


Voice 0.129 (∗)


Complaint 00051 00071 4∗5 −00058Compliment 00414 4∗5 00000 00017Effect of network sentiment

Complaint 00151 −00121 00309Compliment 00923 −00047 00239

Effect of intervention 00211 −00311 10344 4∗5

Unobserved heterogeneity—Population variance


Voice 1.052


Complaint 10169 00131 00107Compliment 00139 00158 00670Effect of network sentiment

Complaint 00366 00301 00328Compliment 00287 00344 00384

Effect of intervention 130841 60343 00570

∗The 95% credible interval does not include zero (variance parameters arepositive and are not marked).

We note a few additional results from estimates ofheterogeneity parameters, reported in Table 7. First,as expected, customers with more followers are morevocal on the website. (The coefficient is 0.129 and sta-tistically significant.) Next, in the negative state, cus-tomers with more followers are more likely to com-pliment. (The coefficient for compliment is positiveand statistically significant.) A potential explanationis that in the negative state, those with more fol-lowers are comparatively more “measured” in theirvoices, whereas those with fewer followers are more“extreme.” Note that this is on a relative basis only,as the tendency to compliment is low in general forthe negative state. Third, in the positive state, cus-tomers with more followers are more apt to respondto past interventions. (The coefficient is 1.344 and sta-tistically significant.) Finally, the unobserved hetero-geneity parameter estimates show wide dispersionsacross customers of their voicing tendencies, partic-ularly on their expectation of receiving service inter-ventions. (The variance of the service utility coeffi-cient is high for the negative and neutral state.)

State Transition0 Table 8 reports state transition pa-rameter estimates. From the negative, neutral, or pos-itive state, the unconditional probability of the cus-tomer remaining in the same state in the next timeperiod is 35.53%, 94.05%, or 65.83%, respectively (firstregion of Table 8). This suggests that the neutral and

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 8 Parameter Estimate—Relationship State

Relationship state and transition probability

From/To state Negative Neutral Positive

Unconditional transition probabilityNegative (%) 35053 63091 0056Neutral (%) 3025 94005 2070Positive (%) 0078 33039 65083

Initial state probabilityProbability (%) 9021 73036 17042

Conditional transition parameterNetwork sentiment 10261 4∗5 10091 4∗5 −0013Firm intervention 10044 4∗5 00587 4∗5 00422 4∗5

∗The 95% credible interval does not include zero.

positive states are relatively “sticky.” The neutral stateis more sticky than the positive state, whereas thenegative state is more transient. That both the positiveand negative states are more transient than the neu-tral state suggests a natural “reversion to the mean”over time. Furthermore, customers in the negative orpositive state are more likely to switch to the neu-tral state than to the opposite one, suggesting that therelationship is more likely to evolve gradually than tochange drastically.

When the firm addresses a customer’s complaintsmore actively, the customer is more likely to transitionto a more positive relationship state. (The coefficientsfor firm’s service intervention, reported in the thirdregion of Table 8, are positive and statistically signif-icant for all three states.) The mean estimates suggestthat, from the negative state, an intervention couldincrease the instantaneous probabilities of moving tothe neutral and positive states in the next period to82.19% and 1.57%. From the neutral state, an inter-vention could increase the probability of moving tothe positive state to 4.75%. The cumulative effect ofintervention over time is even larger, which we dis-cuss in §4.5. Meanwhile, when customers in the neg-ative or neutral state hear more positive messagesfrom others, they are also more prone to form morepositive perceptions of the firm. (The coefficients fornetwork sentiment are positive and statistically sig-nificant for these two states.) In summary, the state-transition parameter estimates show that customers’underlying relationships evolve over time naturallyand gradually, that the firm’s relationships with cus-tomers can indeed be improved through service inter-ventions, and that friends’ opinions also help shapecustomers’ perceptions of the firm.14

Service Intervention0 For the effect of service inter-vention, an interesting contrast can be made betweenvoicing utility and the underlying relationship. Al-though active service interventions would improve

14 We also estimated the same model using data aggregated at theweekly level. The results remain qualitatively the same.

Figure 2 (Color online) Histogram of Probabilities of BeingInfluenced by Friends

Fre

quen

cy

120

100

80

60

20

0

0 0.2 0.4

Probability of being influenced

0.6 0.8 1.0

the firm’s relationships with customers, they alsoraise the customers’ expectations of being helped.This makes them more likely to complain in thefuture. The firm needs to be cognizant of this dualrole, as it can lead to systematic differences betweenobserved voices and the underlying relationships.Similarly, although positive network opinions helpcustomers form positive perceptions of the firm, theseopinions can also trigger complaints due to a differ-entiation effect in certain situations.

In summary, the parameter estimates reveal mark-edly different behavioral tendencies in different un-derlying relationship states. Customers in the nega-tive state likely hold negative views about the firmand are prone to complain. They become even moreso when others praise the firm, eager to “correct”other people’s compliments. Customers in the neu-tral state likely hold neutral views about the firm anddo not exhibit a tendency to either compliment orcomplain about the firm. Redress seeking seems to bean important motivating factor for these customers tocomplain. Customers in the positive state are likelyto compliment the firm, reflecting positive relation-ships, and are likely to reinforce the complimentsvoiced by their friends. Redress seeking is also a keymotivating factor for customers in this state to com-plain. Within each relationship state, the behavior fur-ther varies across customers. Both friends’ messagesand the firm’s service interventions can induce cus-tomers to transition among these relationship states.Recognizing this diverse and state-dependent natureof voicing behaviors and understanding the intrica-cies of service intervention as well as friends’ mes-sages are crucial for successfully engaging customerson social media.

4.3. Network InfluencesAs discussed in §3, we estimate the probabilitiesof customers being influenced by their friends. Thehistogram of the posterior means of this influenceparameter is plotted in Figure 2. The average prob-ability across all customers is 0.469. This indicatesthat on average customers are influenced by slightly

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 9 Regression—Influence Probability and Followers

Regression 1 Regression 2

Dependent variable logit(prob. to be influenced) logit(prob. to influence)Independent variable ln(no. of friends + 1) ln(no. of followers + 1)Slope coefficient −00077 4∗∗∗5 00447 4∗∗∗5

R2 00016 00008

Note. Significance codes: “∗∗∗”, 0.001; “∗∗”, 0.01; “∗”, 0.05; “·”, 0.1.

less than half of those they follow. The histogram isslightly right skewed, with most customers havinga probability of lower than 0.8.15 From the posteriorof dyad-specific �ijs, we also compute the probabil-ity of each customer being influential on others in thesample. We then perform post-hoc analysis by run-ning two regressions. In the first, the dependent vari-able is the probability of a customer being influencedby those the customer follows, logit transformed. Theindependent variable is the number of friends of thecustomer, log transformed. The result, reported inTable 9, shows that as customers follow more peo-ple, the probability of them being influenced by eachindividual friend is lower. (The coefficient is negativeand statistically significant at the 0.001 level.) This isintuitive, as when a person follows more friends, theamount of time that can be spent on each friend’stweet is smaller. This lowers the probability of beinginfluenced. In the second regression, the dependentvariable is the probability of a customer being influ-ential on the customer’s followers, logit transformed.The independent variable is the number of followersthe customer has, log transformed. The result showsthat a customer with more followers is more likelyto be influential, again a reasonable result. (The coef-ficient is positive and statistically significant at the0.001 level.)

4.4. Voice and RelationshipA key aspect of our study is to separate the underly-ing relationships from the observed voices. We nowcompare and contrast these two constructs. To begin,Table 10 shows the empirical distribution of com-plaints and compliments over the underlying relation-ship states. States are recovered using the filteringapproach for HMM (Montgomery et al. 2004, Net-zer et al. 2008). We first note that voice and relation-ship are generally consistent: 70% of all complaints

15 In Trusov et al. (2010) the reported average probability of beinginfluenced is 0.22. The estimate here is higher, possibly because ofthe difference in context. Trusov et al. (2010) use log-in activities asindirect measures of usage, and the extent to which friends observesuch activities is unclear. In contrast, Twitter users explicitly choosewhom they follow, and their friends’ tweets will be pushed to them.It is not, therefore, surprising to get a higher estimate of influence inthe context of Twitter. An in-depth analysis of drivers of influenceon Twitter is left for future study.

Table 10 Complaints and Compliments by Relationship State

Relationship state (%)


Complaint 70000 26046 3054Compliment 0075 25003 74022

were made by customers in the negative state, and74% of compliments were made in the positive state.Furthermore, 26% of complaints and 25% of compli-ments were made by customers in the neutral state.Although the intrinsic desire to compliment or com-plain is low in the neutral state, customers spend themost time in this state, with social factors and redress-seeking effects also playing a role.16 Interestingly, 3.5%of complaints were also voiced by customers in thepositive state, considerably higher than the 0.75% ofcompliments in the negative state. This difference isdriven by redress seeking customers who have pos-itive perceptions of the firm but who may still com-plain. They expect the firm to address their issues,and they may even complain precisely because theyhold the firm in high regard, possibly because thefirm has been responsive before.

To further compare voice and relationship overtime, we compute two sentiment indexes, one usingthe observed voices and the other using relation-ship states. We term the former the “voice sentimentindex” and the latter the “relationship index.” Con-sistent with industry practice, the former is calcu-lated by dividing the number of compliments by thetotal number of compliments and complaints. Accord-ingly, the latter is calculated by dividing the numberof occurrences of positive relationship state by thetotal occurrences of positive and negative states. Wecompute both indexes at the weekly level to smoothout daily fluctuations. The two indices are plotted inFigure 3 (for the relationship index, both posteriormean and 2.5% and 97.5% credible interval are plot-ted), from which we make three observations. First,the two indexes generally track each other over time,confirming that the latent relationship is a key driverof customer voices. Second and more important, thevoice sentiment index is on average lower than therelationship index. This suggests that directly usingvoices may underestimate customer sentiment. Thediscrepancy is likely driven by redress seeking, whichdisproportionately encourages complaints. Anotherreason behind this discrepancy is that negative cus-tomers are more motivated to complain than posi-tive ones are to compliment. This mean difference

16 Although the population level intercept estimates do not showthe effect of social utility for the neutral state, at the individualcustomer level it does exist, as the population variances reportedin Table 8 show a wide dispersion of the related coefficients acrosscustomers.

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Figure 3 (Color online) Voice Sentiment Index and RelationshipIndex from the Data

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43

Week

Voice index (VI) Relationship index (RI)RI 2.5% RI 97.5%

between the observed voice and the underlying rela-tionship is a key takeaway for managers, one thatcannot be revealed without explicitly modeling therelationship. Third, the relationship index is much lessvolatile than the voice index, likely because customervoices are relatively sparse and depend on the socialmedia environment. Industry managers often tracknegative voices in real time and are concerned withcontrolling damage to a firm’s reputation. However,such concern may be misplaced or an overreaction.The relationship index shows a stable and gradualevolution of sentiment. Even the voice index, despitevolatility, remains stable over time. This is likely dueto the self-stabilizing effect in this environment, ascustomers often differentiate or correct others’ voices.By capturing the underlying trend in a more persis-tent fashion, the relationship index is a good alterna-tive to the voice index that managers typically rely on.Its lower volatility also makes it an easier-to-use tool.

4.5. Effectiveness of Firm’s Service InterventionDoes service intervention on social media improvea firm’s image? To what extent is the improvementreflected in customer voices? How should the firmmanage interventions effectively? To answer theseimportant but largely open questions, we investigatethe effect of a firm’s service intervention on voice andrelationship over time. We do so through simulation.For each individual complaint in the data set, we sim-ulate two paths, with the firm intervening in the firstbut not in the second. We then simulate forward fortwo weeks for both scenarios. We perform the simu-lation using individual MCMC draws, and computeboth the mean and credible intervals for the relevantmeasures.

The result of the simulation is plotted in Figure 4,where we show the mean improvement in voice andrelationship indexes through intervention, as well asthe 2.5% and 97.5% posterior quantiles. Three thingsare notable. First, service intervention has a significantpositive effect on both the voice sentiment and therelationship. The intervention improves the voice sen-timent index by 0.048 after two weeks (95% credibleinterval is −00001 to 0.088), and the relationship indexby 0.170 after two weeks (0.145 to 0.192). Second,the effect of intervention persists over time. The firstfew periods show increasing improvements on bothvoice and relationship indexes. The effect erodes to acertain extent after that, as the improvement dimin-ishes over time. However, the improvement does notregress to zero. This shows that interventions havea longer-term impact, as customers remember thefirm’s past responsiveness. Finally, and most impor-tant, the result again shows that using only observedvoices will underestimate the effect of intervention.The improvement of the relationship index is muchhigher than that of the voicing sentiment index. Thisdifference is also significant. (The two credible inter-vals do not overlap.)

The difference between the improvements in thevoice and relationship indices can be attributed tothe redress-seeking effect. To see this, we performtwo additional simulations to measure the partialeffects of intervention. In the first, we allow ser-vice intervention to affect relationship, but ignorethe redress-seeking effect (by fixing the service util-ity component to the value before the service inter-vention). In the second, we admit the redress-seekingeffect but ignore the effect on relationship (by fixingthe service intervention term in the state transitionequation). The changes to the voice index of thesetwo simulations, along with the overall improvementin the voice index as discussed earlier, are shown

Figure 4 (Color online) The Effect of Firm Intervention

–0.05

0

0.05

0.10

0.15

0.20

0.25

0.30

0.35

Impr

ovem

ent i

n vo

ice/

rela

tions

hip

inde

x

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Days after initial complaint

Voice index (VI) Relationship index (RI)VI 2.5% VI 97.5%RI 2.5% RI 97.5%

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Figure 5 (Color online) Decomposing the Effect of Intervention onVoice Index

–0.20

–0.15

–0.10

–0.05

0

0.05

0.10

0.15

0.20

0.25

Impr

ovem

ent i

n vo

ice

inde

x

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Days after initial complaint

VO 2.5%

Overalleffect (OE)

Relationshipimprovement (SO)

Redressseeking (VO)

OE 2.5% OE 97.5% SO 2.5%

SO 97.5% VO 97.5%

in Figure 5. If we consider only the effect of ser-vice intervention through improvement in relation-ship, the improvement stabilizes at 0.163 (0.139 to0.194), very close to the improvement in the rela-tionship index. However, if we consider only theredress-seeking effect of service intervention, then thevoice index reduces by 0.117 (−00148 to −00074), asthe customer subsequently complains more. There-fore, the redress-seeking effect offsets a significantportion of the would-be improvements in customervoices through improved relationship (almost two-thirds on average). This clearly illustrates the twoopposing effects that service intervention has on cus-tomer voices. And it underscores the importance ofunderstanding the nuances of customer voices andunderlying relationships.

Recovering the underlying relationship state alsoprovides an opportunity for targeting. To furtherinvestigate, we look at the effect of intervention condi-tional on the customer’s relationship state at the timeof complaint. Specifically, we break down complaintinstances into five quintiles based on the posterioron relationship state, from negative to positive. Theimprovements in voice and relationship indices forthe five quintiles are reported in Table 11. Althoughservice intervention improves voice and relationshipindices significantly for all quintiles, the improvementis higher for customers in the middle quintiles (thirdand fourth quintiles). Targeting the third or fourthquintile results in about 50% higher improvementthan targeting the first quintile. Customers in thethird or fourth quintiles are mildly negative. Firms,thus, should focus on the customers “on the mar-gin” between negative and neutral states.17 We also

17 At the time of complaint, customers are much more likely tobe in the negative state than in the positive one. The third andfourth quintiles are the “mildly negative” customers (posteriors

contrast this relationship state-based targeting witha common industry practice of targeting customerswith more followers. We break down customers intoquintiles based on the numbers of followers, andcalculate the improvement in voice and relationshipindices for each quintile. The results are also reportedin Table 11. The difference across quintiles by num-bers of followers is also quite significant for the voiceindex, largely driven by the observed heterogeneityof the service utility. However, the difference on therelationship index is muted, with all five quintilesgetting similar improvements. Hence the underlyingrelationship can be used as an effective targeting crite-rion, supplementing the popular criterion of numberof followers.

In summary, the managerial implications are three-fold. First, service intervention improves both cus-tomers’ voiced opinions and their relationships withthe firm. Although service intervention is a double-edged sword in that it also encourages complaints, itsindirect effect on voices through improved relation-ship states is stronger, and the net effect on customervoices is positive. Second, because of the dual effectsof intervention, measuring only the voice will under-estimate the return on service intervention. Third,recovering the underlying relationship enables effec-tive targeting of service interventions.

5. Conclusion and Future ResearchThe ease of complaining on social media platformsempowers customers to speak up. With the expec-tation of firm participation on such platforms, cus-tomers’ use of social media as a channel for complaintsis rapidly increasing. It is equally easy to complimentas to complain on social media. Happy customers whoappreciate the firm readily do so. Successfully engag-ing customers online through social media is quicklybecoming an imperative for business practitioners.Although anecdotes abound, a clear understanding ofcustomers’ voicing behavior on social media and ofthe effect of firms’ participation in this process is stilllacking. Our study fills this gap by explicitly modelingthe dynamics of customers’ voices and their underly-ing relationships with firms on social media. We alsoinvestigate the effects of service intervention in thisframework.

We show the importance of probing beneath thesurface of compliments and complaints to uncover theunderlying drivers of customer behavior. We find thatcustomers’ voices crucially depend on their under-lying relationships with the firm. Customers who

show slightly higher chance of being in the negative state than inneutral), whereas the first and second quintile are the more nega-tive customers (posteriors show a much higher chance of being inthe negative state).

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 11 ROI of Intervention by Relationship and Follower Quintiles

Target by Measure First quintile Second quintile Third quintile Fourth quintile Fifth quintile

Relationship Voice index 00042 00050 00053 00052 000434000125 4000175 4000225 4000275 4000285

Relationship index 00144 00171 00189 00195 001454000105 4000135 4000165 4000175 4000165

Number of followers Voice index 00026 00059 00056 00058 000444000435 4000165 4000145 4000205 4000395

Relationship index 00168 00170 00163 00182 001664000145 4000145 4000145 4000155 4000165

Notes. Average improvement in voice and relationship index through intervention after 14 days. Numbers in parentheses are standarderrors.

have negative, neutral, or positive relationships withthe firm have different inclinations to complain andto compliment. We further confirm that social fac-tors and service intervention also affect customervoices. These factors create a discrepancy betweenthe observed voices and the underlying relationships.This is especially true for redress-seeking motivation,which creates a negative bias in observed voices. Wealso provide a rich characterization of customers indifferent relationship states. For example, althoughredress seeking in general drives complaints, thiseffect is not evident for customers in the negativerelationship state. As another example, customers innegative relationship states demonstrate a strong ten-dency to correct friends’ messages. These detailedunderstandings are enabled by our focus at the indi-vidual customer level on the dynamics of both voicesand relationships.

These findings have significant implications for sen-timent and complaint management on social media.Industry managers tend to focus on positive or neg-ative voices at the surface level. However, we showthat customer voices are affected and biased by manyother factors. It is, therefore, vital to uncover theunderlying relationships of customers. Managers tendto fear the potential viral effect of complaints, butwe show that such concern may be misplaced. Weshow that the sentiment indexes are generally sta-ble and the social media environment is, to a cer-tain extent, self-stabilizing. Managers tend to targettheir service interventions based on the numbers offollowers a customer has. However, we show thattargeting based on the customer’s underlying rela-tionship could be more effective. More important,managers should understand that service interven-tion has opposing effects. When the firm respondsto a complaint, it also raises the customer’s expec-tations. Responding to complaints thus encourageseven more complaints. On the positive side, serviceintervention does help improve the customer’s under-lying relationship with the firm. The improved rela-tionship also leads to more positive voices, and the

positive effect outweighs the negative. It is importantfor practitioners to understand such nuances of ser-vice interventions in order to accurately evaluate theirimpact.

Several limitations of our study call for furtherresearch. First, the most direct and credible signal ofcustomer sentiment is retention and purchase. How-ever, as in many studies of social media, we do nothave such data, nor do we have an offline trackingsurvey for external validation as done by Schweideland Moe (2014). Such data can potentially providea stronger validation of our findings and tie themdirectly to marketing outcomes. Second, we do nothave customer demographic information other thanthe number of followers. It will be interesting to give aricher characterization of how voicing decisions differacross customer profiles. Similarly, it will be interest-ing to investigate how influence differs across friendprofiles. Third, our data are for a single microblog-ging site. Although our model is grounded in gen-eral WOM theories, it is possible that certain resultsare specific to the site. Our model can be adaptedwith minor adjustments to other types of venues. Itwill also be interesting to see to what extent the var-ious factors play a role on other platforms. Fourth,we focus on the dynamics at the individual customerlevel and take into account customer connections onsocial media only by including the direct effect ofmessages. Network structures of social media can becomplex and multilayered, a fact that calls for morefocused studies. Finally, firms may also benefit fromproactively engaging with customers on social media,rather than just reacting to complaints. As customersrespond favorably to service intervention, they mayvery well respond to other firm efforts, such as thank-ing customers for their compliments. This is an inter-esting topic for future study as firms expand theironline activities.

Supplemental MaterialSupplemental material to this paper is available at http://dx.doi.org/10.1287/mksc.2015.0912.

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


AcknowledgmentsThe authors thank David Godes, Wendy W. Moe, David A.Schweidel, the editor, the associate editor, and the anony-mous referees for comments on earlier drafts of this paper.The authors would also like to thank the participants at the2012 INFORMS Marketing Science Conference. The authorsacknowledge with appreciation financial support from theCarnegie Bosch Institute at Carnegie Mellon University forthis study.

ReferencesAaker J, Fournier S, Brasel SA (2004) When good brands do bad.

J. Customer Res. 31(June):1–16.Albuquerque P, Pavlidis P, Chatow U, Chen K-Y, Jamal Z (2012)

Evaluating promotional activities in an online two-sided mar-ket of user-generated content. Marketing Sci. 31(3):406–432.

Anderson EW (1998) Customer satisfaction and word of mouth.J. Service Res. 1(1):5–17.

Basuroy S, Chatterjee S, Ravid SA (2003) How critical are criticalreviews? The box office effects of film critics, star power andbudgets. J. Marketing 67(4):103–117.

Berger J, Milkman KL (2012) What makes online content viral?J. Marketing Res. 49(2):192–205.

Berger J, Schwartz E (2011) What drives immediate and ongoingword of mouth? J. Marketing Res. 48(5):869–880.

Blodgett JG, Anderson RD (2000) A Bayesian network model of thecustomer complaint process. J. Service Res. 2(4):321–338.

Blodgett JG, Granbois DH, Walters RG (1993) The effects ofperceived justice on complainants’ negative word-of-mouthbehavior and repatronage intentions. J. Retailing 69(4):399–428.

Blodgett JG, Wakefield KL, Barnes JH (1995) The effects of customerservice on consumer complaining behavior. J. Services Market-ing 9(4):31–42.

Chen Y-J, Kirmani A (2015) Posting strategically: The consumer asan online media planner. J. Consumer Psych. Forthcoming.

Chevalier JA, Mayzlin D (2006) The effect of word of mouth onsales: Online book reviews. J. Marketing Res. 43(3):345–354.

Chib S (1995) Marginal likelihood from the Gibbs output. J. Amer.Statist. Assoc. 90(432):1313–1321.

Chintagunta PK, Gopinath S, Venkataraman S (2010) The effects ofonline user reviews on movie box office performance: Account-ing for sequential rollout and aggregation across local markets.Marketing Sci. 29(5):944–957.

Day RL, Landon EL Jr (1976) Collecting comprehensive con-sumer compliant data by survey research. Adv. Consumer Res.3(1):263–268.

Dellarocas C, Narayan R (2006) A statistical measure of a popula-tion’s propensity to engage in post-purchase online word-of-mouth. Statist. Sci. 21(2):277–285.

Du RY, Kamakura WA (2006) Household life cycles and lifestylesin the United States. J. Marketing Res. 43(1):121–132.

Duan W, Gu B, Whinston AB (2008) Do online reviews matter?—Anempirical investigation of panel data. Decision Support Systems45(4):1007–1016.

Dwyer RF, Schurr PH, Oh S (1987) Developing buyer-seller rela-tionships. J. Marketing 51(2):11–27.

Erdem T, Keane MP (1996) Decision-making under uncertainty:Capturing dynamic brand choice processes in turbulent con-sumer goods markets. Marketing Sci. 15(1):1–20.

Feick LF, Price LL (1987) The market maven: A diffuser of market-place information. J. Marketing 51(1):83–97.

Fornell C, Wernerfelt B (1987) Defensive marketing strategy by cus-tomer complaint management: A theoretical analysis. J. Mar-keting Res. 24(4):337–346.

Fournier S (1998) Customers and their brands: Developing rela-tionship theory in customer research. J. Customer Res. 24(4):343–373.

Godes D, Mayzlin D (2004) Using online conversations to studyword-of-mouth communication. Marketing Sci. 23(4):545–560.

Godes D, Mayzlin D (2009) Firm-created word-of-mouth com-munication: Evidence from a field test. Marketing Sci. 28(4):721–739.

Godes D, Silva J (2012) Sequential and temporal dynamics of onlineopinion. Marketing Sci. 31(3):448–473.

Goffman E (1959) The Presentation of Self in Everyday Life (Double-day, Oxford, UK).

Gopinath S, Thomas JS, Krishnamurthi L (2014) Investigating therelationship between the content of online word of mouth,advertising, and brand performance. Marketing Sci. 33(2):241–258.

Hennig-Thurau T, Gwinner KP, Walsh G, Gremler DD (2004) Elec-tronic word-of-mouth via consumer-opinion platforms: Whatmotivates consumers to articulate themselves on the Internet?J. Interactive Marketing 18(1):38–52.

Hirschman AO (1970) Exit, Voice and Loyalty (Harvard UniversityPress, Cambridge, MA).

Jansen BJ, Chang M, Sobel K, Chowdury A (2009) Twitter power:Tweets as electronic word of mouth. J. Amer. Soc. Inform. Sci.Tech. 60(11):2169–2188.

Knox G, van Oest R (2014) Customer complaints and recovery effec-tiveness: A customer base approach. J. Marketing 78(5):42–57.

Kumar V, Bhaskaran V, Mirchandani R, Shah M (2013) Practiceprize winner—Creating a measurable social media marketingstrategy: Increasing the value and ROI of intangibles and tan-gibles for Hokey Pokey. Marketing Sci. 32(2):194–212.

Li S, Liechty JC, Montgomery AL (2013) Modeling category view-ership of Web users with multivariate count models. Workingpaper, Indiana University, Bloomington.

Li S, Sun B, Montgomery AL (2011) Cross-selling the right prod-uct to the right customer at the right time. J. Marketing Res.48(4):683–700.

Li X, Hitt LM (2008) Self-selection and information role of onlineproduct reviews. Inform. Systems Res. 19(4):456–474.

Liu Y (2006) Word-of-mouth for movies: Its dynamics and impacton box office revenue. J. Marketing 70(3):74–89.

Lovett MJ, Peres R, Shachar R (2013) On brands and word ofmouth. J. Marketing Res. 50(4):427–444.

Luo A, Kumar V (2013) Recovering hidden buyer-seller rela-tionship states to measure the return on marketing invest-ment in business-to-business markets. J. Marketing Res. 50(1):143–160.

Manchanda P, Rossi PE, Chintagunta PK (2004) Response model-ing with nonrandom marketing-mix variables. J. Marketing Res.41(4):467–478.

Moe WW, Schweidel DA (2012) Online product opinion: Incidence,evaluation and evolution. Marketing Sci. 31(3):372–386.

Moe WW, Trusov M (2011) The value of social dynamics in onlineproduct ratings forums. J. Marketing Res. 48(3):444–456.

Montgomery AL, Li S, Srinivasan K, Liechty JC (2004) Modelingonline browsing and path analysis using clickstream data. Mar-keting Sci. 23(4):579–595.

Moon S, Kamakura WA, Ledolter J (2007) Estimating promo-tion response when competitive promotions are unobservable.J. Marketing Res. 44(3):503–515.

Netzer O, Lattin JM, Srinivasan V (2008) A hidden Markov model ofcustomer relationship dynamics. Marketing Sci. 27(2):185–204.

Newton MA, Raftery AE (1994) Approximate Bayesian inferenceby the weighted likelihood bootstrap (with discussion). J. Roy.Statist. Soc. Ser. B 56(1):3–48.

Richins ML (1984) Word of mouth communication as negativeinformation. Adv. Consumer Res. 11(1):697–702.

Rust RT, Chung TS (2006) Marketing models of service and rela-tionships. Marketing Sci. 25(6):560–580.

Schlosser AE (2005) Posting versus lurking: Communicating in amultiple audience context. J. Consumer Res. 32(2):260–265.

Schweidel D, Moe WW (2014) Listening in on social media: A jointmodel of sentiment and venue format choice. J. Marketing Res.51(4):387–402.

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Smith AN, Fischer E, Chen Y (2012) How does brand-related user-generated content differ across YouTube, Facebook, and Twit-ter? J. Interactive Marketing 26(2):102–113.

Sonnier GP, McAlister L, Rutz OJ (2011) A dynamic model of theeffect of online communications on firm sales. Marketing Sci.30(4):702–716.

Stephen AT, Galak J (2012) The effects of traditional and socialearned media on sales: A study of a microlending marketplace.J. Marketing Res. 49(5):624–639.

Tirunillai S, Tellis GJ (2012) Does chatter really matter? Dynamics ofuser-generated content and stock performance. Marketing Sci.31(2):198–215.

Tirunillai S, Tellis GJ (2014) Mining marketing meaning from onlinechatter: Strategic brand analysis of big data using latent Dirich-let allocation. J. Marketing Res. 51(4):463–479.

Toubia O, Stephen AT (2013) Intrinsic vs. image-related utility insocial media: Why do people contribute content to Twitter?Marketing Sci. 32(3):368–392.

Trusov M, Bodapati AV, Bucklin RE (2010) Determining influen-tial users in Internet social networks. J. Marketing Res. 47(4):643–658.

Trusov M, Bucklin RE, Pauwels K (2009) Effects of word-of-mouthversus traditional marketing: Findings from an Internet socialnetworking site. J. Marketing 73(5):90–102.

Wu F, Huberman B (2008) How public opinion forms. PapadimitrouC, Zhang S, eds. Internet and Network Economics, Lecture NotesComput. Sci., Vol. 5385 (Springer, Berlin Heidelberg), 334–341.

Yang S, Hu M, Winer RS, Assael H, Chen X (2012) An empiricalstudy of word-of-mouth generation and consumption. Market-ing Sci. 31(6):952–963.

Ying Y, Feinberg F, Wedel M (2006) Leveraging missing ratingsto improve online recommendation systems. J. Marketing Res.43(3):355–365.

Zeithaml VA, Leonard BL, Parasuraman A (1993) The nature anddeterminants of customer expectations of service. J. Acad. Mar-keting Sci. 21(1):1–12.

Dow

nloa

ded

from

info

rms.

org

by [

207.

41.1

89.7

4] o

n 29

Sep

tem

ber

2015

, at 0

8:12

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

mba china | business school china | executive mba - the...

Documents