The Effect of Customers' Social Media Participation on ... · tiveness of ﬁrms’ social media...

This article was downloaded by: [130.233.35.243] On: 21 August 2016, At: 03:29Publisher: Institute for Operations Research and the Management Sciences (INFORMS)INFORMS is located in Maryland, USA

Information Systems Research

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

The Effect of Customers' Social Media Participation onCustomer Visit Frequency and Profitability: An EmpiricalInvestigationRishika Rishika, Ashish Kumar, Ramkumar Janakiraman, Ram Bezawada,

To cite this article:Rishika Rishika, Ashish Kumar, Ramkumar Janakiraman, Ram Bezawada, (2013) The Effect of Customers' Social MediaParticipation on Customer Visit Frequency and Profitability: An Empirical Investigation. Information Systems Research24(1):108-127. http://dx.doi.org/10.1287/isre.1120.0460

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 © 2013, 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/isre.1120.0460

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

http://www.informs.org

Information Systems ResearchVol. 24, No. 1, March 2013, pp. 108–127ISSN 1047-7047 (print) � ISSN 1526-5536 (online) http://dx.doi.org/10.1287/isre.1120.0460

© 2013 INFORMS

The Effect of Customers’ Social Media Participationon Customer Visit Frequency and Profitability:

An Empirical InvestigationRishika Rishika

Mays Business School, Texas A&M University, College Station, Texas 77843,[email protected]

Ashish KumarDepartment of Marketing, Aalto University School of Business, FI-00076 Aalto, Finland,

[email protected]

Ramkumar JanakiramanMays Business School, Texas A&M University, College Station, Texas 77843,

[email protected]

Ram BezawadaSchool of Management, State University of New York, Buffalo, New York 14260,

[email protected]

In this study we examine the effect of customers’ participation in a firm’s social media efforts on the intensityof the relationship between the firm and its customers as captured by customers’ visit frequency. We further

hypothesize and test for the moderating roles of social media activity and customer characteristics on thelink between social media participation and the intensity of customer-firm relationship. Importantly, we alsoquantify the impact of social media participation on customer profitability. We assemble a novel data set thatcombines customers’ social media participation data with individual customer level transaction data. To accountfor endogeneity that could arise because of customer self-selection, we utilize the propensity score matchingtechnique in combination with difference in differences analysis. Our results suggest that customer participationin a firm’s social media efforts leads to an increase in the frequency of customer visits. We find that thisparticipation effect is greater when there are high levels of activity in the social media site and for customerswho exhibit a strong patronage with the firm, buy premium products, and exhibit lower levels of buying focusand deal sensitivity. We find that the above set of results holds for customer profitability as well. We discusstheoretical implications of our results and offer prescriptions for managers on how to engage customers viasocial media. Our study emphasizes the need for managers to integrate knowledge from customers’ transactionalrelationship with their social media participation to better serve customers and create sustainable business value.

Key words : social media marketing; social media participation; customer-firm relationship; shopping visitfrequency; customer profitability; propensity score matching; quasi-experiment; difference-in-differences

History : Chris Dellarocas, Senior Editor; Pei-Yu Chen, Associate Editor. This paper was received on January12, 2012, and was with the authors 4 months for 2 revisions. Published online in Articles in AdvanceDecember 20, 2012.

1. IntroductionSocial media has changed the way individuals inter-act with each other and is redefining the way firmsconnect with their customers. With 75% of Inter-net users participating in some form of social media(Forrester Research 2008), companies are increasinglyinvesting in the creation of social media platformsand strategies to augment customer-firm interactions.A recent article in ZDNet (2011) reports that 69%of small business owners use some form of socialmedia (e.g., Facebook, Twitter) and about 78% ofthem plan to increase their budgets devoted to thismedium. Social media enables firms to converseabout their offerings with customers interactively

while also providing them the opportunity to shareinformation about products (Godes and Mayzlin 2004,Agarwal et al. 2008). Thus, social media has providedfirms—both large and small—with a new tool for cus-tomer engagement.

Despite the eagerness on the part of firms to em-brace social media to connect with customers, thereis also much skepticism about its efficacy. For exam-ple, a recent IBM report (2011, p. 3) mentions that“social media is no longer the adorable baby everyonewants to hold, but the angst-filled adolescent—stillimmature yet no longer cute—who inspires mixedfeelings.” The doubts about the effectiveness of socialmedia arise because the link between firms’ social

108

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

Rishika et al.: The Effect of Customers’ Social Media Participation on Customer Visit Frequency and ProfitabilityInformation Systems Research 24(1), pp. 108–127, © 2013 INFORMS 109

media efforts and the return on their investment hasnot been established. In particular, no study, usingindividual level data to our knowledge, has connectedcustomers’ social media participation with their actualbehavior. Recent studies in the broad area of socialmedia have primarily focused on the effectiveness ofuser generated content (UGC) in stimulating productsales (Godes and Mayzlin 2004, Dellarocas et al. 2007,Forman et al. 2008, Zhu and Zhang 2010), encourag-ing content diffusion (Susarla et al. 2012), and foster-ing acceptance of recommendations (Ho et al. 2011).However, from the perspective of a firm that is con-templating social media investments, it is imperativeto know if such expenditures help in furthering theintensity of customer-firm relationships and in cre-ating sustainable firm value. In this regard, becausemost of the extant studies use aggregate data, theyare unable to uncover critical individual level differ-ences that are essential for obtaining insights at thecustomer level. Thus, the objective of this paper is tosystematically examine the effect of customers’ par-ticipation in firm initiated social media efforts on thebusiness value generated thereof for the firm. For thispurpose, we choose to work with customers’ shop-ping visit frequency as the focal variable because itcaptures the intensity of relationship between a firmand a focal customer and serves as a key driverof customer lifetime value (Venkatesan and Kumar2004). Furthermore, we also examine the impact ofcustomer social media participation on customer prof-itability because that directly contributes to a firm’sbottom line.

To study the central question concerning the effec-tiveness of firms’ social media strategy, we first exam-ine the main effect of customers’ participation insocial media on the intensity of relationship betweencustomers and the firm (as measured by customers’frequency of shopping visits). We then hypothesizeand test for the moderating role of overall activity inthe social media site and a set of customer charac-teristics. Because contribution margin from customersis another key component of the customer lifetimevalue, we next complement our core findings relatedto visit frequency by examining the effect of socialmedia participation on customer profits. In order toaccomplish our objectives, we assemble a novel dataset in which we observe customer participation deci-sions in a firm initiated social media site and con-nect this to individual purchase data for the same setof customers. We believe that our study is the firstto establish the link between customers’ social mediaparticipation and the transaction driven firm valuegenerated thereof using actual behavioral data.

However, in our research context, it is importantto account for endogeneity issues arising because ofcustomer self-selection. For instance, customers who

have an affinity toward the firm are also more likelyto join its social media site and have an intensive rela-tionship with the firm. To control for this self-selectionissue, we employ the combination of propensityscore matching (PSM) technique and difference-in-differences (DID) analysis. In particular, we utilizea quasi-experimental research design with two dis-tinct groups of customers: (1) a “treatment” groupthat comprises customers who participate in the firminitiated social media site and whose behaviors wewish to analyze and (2) a “control” group consist-ing of customers who are not part of or never par-ticipate in the firm’s social media site. We use PSMin combination with DID analysis because it helpsmimic a random experimental design, thus produc-ing an unbiased estimate of the “treatment effect,”i.e., the impact of customer social media participation(Dehejia and Wahba 1999, Huang et al. 2012). We thenextend the DID modeling framework to examine howthe social media participation effect varies across cus-tomers depending on their level of relationship withthe firm.

Our study makes three substantial contributionsto the social media literature. First, we attempt touncover the direct benefits that accrue to firms asa consequence of their social media efforts. Ourapproach is different from the extant research (e.g.,Dholakia and Durham 2010, Algesheimer et al. 2010)in that we use actual behavioral data to estab-lish the causal link between customers’ social mediaparticipation and their buying behavior. Moreover,unlike the above studies, we explicitly account forendogeneity issues arising because of self-selection.Second, no study to our knowledge has integratedcustomers’ social media participation behavior withtheir transaction relationship with the firm and stud-ied how customer-firm relationships in online andoffline domains may interact to create value for thefirm. This study helps fill this gap in the literature.Third, our research also examines how the effect ofsocial media participation behavior varies across dif-ferent types of customers. By focusing on transac-tion based customer characteristics, we offer insightson when participation becomes more salient. More-over, we quantify the economic benefits that resultfor firms due to customers’ social media participa-tion. Managers can use these results to understandhow customer relationships cultivated in the onlinesocial media environment can be leveraged to opti-mize return on investment.

2. Hypotheses DevelopmentThe focus of the study is on examining how socialmedia efforts by the firm help transform the inten-sity of customer-firm relationships. Accordingly, we

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

Rishika et al.: The Effect of Customers’ Social Media Participation on Customer Visit Frequency and Profitability110 Information Systems Research 24(1), pp. 108–127, © 2013 INFORMS

measure this through a key characteristic representingcustomers’ transactional relationship with the firm,frequency of visits. We believe that customer visit(or purchase) frequency captures the strength ofcustomer-firm relationship because prior research hasconsidered visit frequency to be an important indi-cator of the relationship that is also positively linkedto customer lifetime value (e.g., Venkatesan andKumar 2004).

In examining the impact of social media partici-pation on the intensity of customer-firm relationship,we focus on two sets of factors that can influencethis relationship: (a) overall social media activity and(b) customer characteristics. Social media has becomea preferred platform for cultivating meaningful rela-tionships between firms and their customers. How-ever, from a managerial perspective, it is crucial tounderstand the moderating factors that make cus-tomers’ social media participation more or less salient.In particular, we investigate how the level of activ-ity on the social media site in terms of number ofmessages and comments posted by the firm and cus-tomers moderates the effect of social media participa-tion. Furthermore, managers often segment customersbased on their transaction characteristics for bettertargeting and optimal allocation of resources. Forbusiness executives, it is important to know on whichcustomer segments they must spend greater resourcesin order to build and strengthen relationships (Rigbyet al. 2002). Thus, an understanding of how the impactof social media participation differs across customersegments can enable firms to develop better customerrelationship management (CRM) initiatives. We drawon CRM literature to identify key customer character-istics (i.e., purchase amount, buying focus, deal sensi-tivity, and premium product purchases) that can helpin ascertaining the differential response among differ-ent customer segments and aid in managerial decisionmaking.

2.1. Customers’ Participation in Firm HostedSocial Media

By participating in the firm initiated social mediasite, customers can strengthen their relationship withthe firm through many processes. First, the socialmedia platform provides an opportunity for cus-tomers to voice their opinions and share experiencesabout the firm with other customers as well as thefirm, thereby creating social interactions (Agarwalet al. 2009). In the words of Kietzman et al. (2011,p. 241), “social media employ mobile and web-basedtechnologies to create highly interactive platforms viawhich individuals and communities share, co-create,discuss, and modify user-generated content.” Thus,this ability of social media to facilitate greater inter-actions between consumers and the firm can help inestablishing a deeper connection between them.

Second, participation in a firm’s social media siteprovides an easy access to a social network consistingof other customers’ opinions and experiences. Anactive network with message postings from othercustomers regarding their experiences with the firmand its products creates the feeling of a brand/firmcommunity that will positively influence a focalcustomer’s intensity of relationship with the firm.Research suggests that relationships among fellowcustomers, who are an integral part of a community,help in the fostering of firm loyalty, thereby inten-sifying the relationship with the firm (McAlexanderet al. 2002).

Third, participating in a firm hosted social mediasite provides customers easy access to information onfirms’ offerings and other related messages. Throughthese postings, customers can infer the firm’s levelof involvement with its customer base and its com-mitment toward enhancing their experiences. Suchincreased interactions can lead to better customeridentification with the firm and create greater trustand customer loyalty (Algesheimer et al. 2005). Thus,firm hosted social media opens a new channel ofcommunication through which customers can connectand communicate directly with the firm and othercustomers that will positively affect the intensity ofcustomer-firm relationship. Based on the above dis-cussion, we hypothesize the following:

Hypothesis 1 (H1). Participation in firm hosted socialmedia by a focal customer will have a positive impact onthe intensity of the customer-firm relationship.

2.2. Interaction Effect with Social Media ActivityCustomers who choose to engage with the firmthrough its social media site gain easier access tomessages from both the firm and other customers.We believe that the impact of social media partic-ipation on customer-firm relationship will be influ-enced by the level of activity in the social mediasite in terms of the number of messages posted bythe firm and other customers. First, because one ofthe main purposes of the firm hosted social mediasite is to create a platform for communicating withits customers and building relationships with them,a firm must maintain a high level of activity on thesocial media site. A more active and vibrant com-munity with regular new message postings will cre-ate trust and allow customers to infer the level ofa firm’s relationship commitment and bolster cus-tomers’ bond with the firm. Extant research suggeststhat commitment and trust in a relationship are pos-itively associated with a party’s intentions to con-tinue the relationship with the other party (Morganand Hunt 1994). Upon creating social media sites(e.g., Facebook pages, Twitter accounts), if firms failto interact regularly with their clientele through this

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


new medium, it could create customer skepticismand result in less favorable behavioral outcomes. Sec-ond, a higher number of postings by other customersreflects the level of customers’ involvement with thefirm. Individuals often wish to validate their choicesin an environment that further strengthens their ownaffinity with the firm (Zhu et al. 2009). In the contextof a social media site, by scrutinizing other customers’behavior and comments, focal customers can feelmore confident about their own choices, thus impact-ing their intensity of relationship with the firm.1

Next, a mere exposure effect may also exist thatmay make the firm more accessible to the cus-tomer (Janiszewski 1993). As customers are exposedto a greater number of messages from the firm andother customers about its products, they becomemore favorably disposed toward the firm. Thus, anactive social media page with regular new messages/postings can help customers form more positive atti-tudes toward the firm and strengthen the customer-firm relationship. Hence, we hypothesize that theimpact of customers’ social media participation on theintensity of their relationship with the firm will begreater for a higher level of activity (e.g., messages,comments) on the social media site.

Hypothesis 2 (H2). The impact of participation in firmhosted social media on the intensity of the customer-firmrelationship will be greater for a higher level of messagepostings in the social media site.

2.3. Interaction Effects withCustomer Characteristics

Purchase Amount. Although participation in a firmhosted social media site will have a positive effecton customer-firm relationship for all customers (asargued for H1), we expect that customers who have astronger transaction relationship with the firm (beforeparticipation in social media), as measured by theirspending levels, will likely experience greater ben-efits from association with the social media site.This is likely to happen through two mechanisms.First, extant literature in CRM suggests that exchangecharacteristics such as average transaction amountare associated with greater satisfaction (Bolton 1998)and commitment that results in better behavioral

1 We note that these arguments are valid only if most customer com-ments are positive. We also wish to point out that in our context, allof the messages that originate from the firm are positive and morethan 90% of customer comments are also positive in nature. Thisis likely because this is a firm-hosted social media site where mostof the messages originate from the firm and customer commentsare in response to these messages and generally take the form ofpositive comments or general queries. This is unlike a third partywebsite where customers may post product reviews and recom-mendations that can be either positive or negative. We thank ananonymous reviewer for seeking clarification on this issue.

outcomes, which strengthens the relationship betweenthe firm and the consumer (Reinartz and Kumar2003). Thus, customers who spend a higher amountwith the firm are likely to be more committedand more relationship oriented than are customerswho spend less with the firm (Crosby et al. 1990).Because they have a greater affinity toward the firm,customers with a larger pecuniary investment willexhibit a greater response to the firm’s relationshipbuilding communications through social media thanwill customers with lower financial investments. Sec-ond, prior research also shows that customers withhigher spending levels, due to their greater finan-cial investments with the firm, feel that they aremore important to the firm; therefore, they are moredemanding and expect greater relationship invest-ments on the part of the firm (Wangenheim andBayón 2007). Thus, high value customers are likely tovalue firms’ relationship investments in social mediamore than low value customers are. Hence, we expectresponse to social media participation to be greater forcustomers with higher spending levels as comparedto customers with lower spending levels. Hence, wehypothesize the following:

Hypothesis 3 (H3). The impact of participation in firmhosted social media on the intensity of the customer-firmrelationship will be greater for customers with a larger pur-chase amount.

Focus of Buying. Buying focus refers to the breadthof interaction between a customer and the focal firm,and extant research considers buying focus to be animportant indicator of relationship status (Verhoefand Donkers 2005). High buying focus typicallyentails limited customer-firm interaction, wherein afocal customer tends to buy exclusively from a singleproduct category and thereby does not broaden thescope of interactions with the firm. Research suggeststhat lower customer satisfaction with the offerings ofthe firm can manifest itself as highly focused purchas-ing behavior (Verhoef et al. 2001). Lower customersatisfaction has been linked to less favorable out-comes in several studies (e.g., Anderson and Sullivan1993, Bolton 1998). On the other hand, less focusedbuying (purchasing across a wide range of productsand categories) increases customers’ familiarity withthe firm and can lead to greater satisfaction withthe firm. For these less focused buyers, resources onthe social media site can prove to be particularly valu-able. Because the social media site provides exposureto information on new and different products, lessfocused customers can reap greater benefits from par-ticipation in social media because they tend to pur-chase across multiple product categories. Thus, weexpect that the impact of social media participationon the intensity of customer-firm relationship will be

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


lower for customers having a greater buying focus.Therefore,

Hypothesis 4 (H4). The impact of participation in firmhosted social media on the intensity of the customer-firmrelationship will be lower for customers who have a greaterfocus in buying.

Deal Sensitivity. An important customer characteris-tic in the context of customer-firm relationship is theextent to which a focal customer looks to buy items onpromotion. Extant research suggests that customerswho are more deal prone tend to search more forproducts from competing sellers and tend to maketheir purchases with the seller who offers the lowestprice (Ailawadi et al. 2001). Thus, deal prone con-sumers are likely to be less committed to the firm andhave reduced interest in developing a stronger rela-tionship with the firm. These customers may visit thefirm’s social media site only to search for coupons orpromotions. In such a case, we would expect firms’relationship building efforts in the form of engagingin interactive communications with their customersthrough social media to elicit a greater response fromless deal prone consumers who may appreciate firms’relationship investments more.2 Thus, we argue thatonline social media participation would strengthenthe customer-firm relationship for customers who areless deal prone:

Hypothesis 5 (H5). The impact of participation in firmhosted social media on the intensity of the customer-firmrelationship will be greater for customers with lower dealsensitivity.

Share of Premium Products. Customers who purchasethe firm’s most valuable offerings are also consid-ered to be its most lucrative customers, and firmsoften invest considerable amount of resources in nur-turing relationships with them (Reinartz and Kumar2000). In addition, premium products (because oftheir higher price tags) are often high involvementpurchases, which in turn are associated with greaterlevels of uncertainty and perceived risk (Pavlou et al.2007).3 Customers with a larger proportion of high

2 We note that in case the firm’s social media site is mostly devotedto offering coupons or promotions, we would expect the oppositeeffect where deal prone customers would shop with the firm more.However, our focus is on understanding the impact of firms’ rela-tionship building efforts through social media, where firms engagein a variety of communications with their consumers (both productand non-product related), which is the strategy used by the firmunder study. We thank an anonymous reviewer for bringing ourattention to this issue.3 We note that uncertainty can be due to either seller or prod-uct uncertainty. Our focus here is on product related uncertaintybecause consumers perceive a high risk associated with high-involvement purchases because of their higher uncertainty (Pavlouet al. 2007). We thank an anonymous reviewer for helping us clarifythe issue.

ticket purchases are likely to have greater switchingcosts and thus may be more inclined to continue therelationship with their current firm and attempt togain preferential status with it. Moreover, interactionswith the firm through the online social media canhelp such customers achieve these objectives throughobtaining a greater degree of personalization andinformation for premium products in real time. More-over, they may also utilize social media to gather cuesabout premium products from other customers of thesocial media site to mitigate uncertainty (Pavlou et al.2007). Thus, participation in social media is likely tostrengthen the intensity of customer-firm relationshipfor customers who purchase a greater share of pre-mium products:

Hypothesis 6 (H6). The impact of participation infirm hosted social media on the intensity of the customer-firm relationship will be greater for customers who have agreater share of premium product purchases.

3. Research SettingThe data set for this study comes from a largespecialty firm that owns multiple stores and sellswine and like products in the northeastern UnitedStates.

3.1. The Firm’s Customer EngagementInitiative via Social Media

In August 2009, the focal firm, for the first time, cre-ated a webpage on a major social networking website(e.g., Facebook) with the objective of connecting withits customers through social media.4 Since then, it hasmade conscious efforts to continue to invest in socialmedia by regularly updating the site through mes-sage postings and by encouraging customers to visitthe site and become involved by posting, comment-ing, and other such activities. The messages postedby the firm on the above site differ both in their con-tent and characteristics. In particular, the site is usedby the firm to build store equity, foster better cus-tomer relationships, communicate to customers aboutvarious events/activities, and disseminate informa-tion about its products and promotions. Examplesof equity and relationship building messages include“we are currently touring wineries in Argentina toget you the best wines” and “we are constantly striv-ing to improve your shopping experiences,” respec-tively. Messages that inform customers about variousevents usually relate to community based activities(e.g., 5K runs for charity). Other message postingscan be solely product (e.g., “how to select a good

4 The confidentiality arrangements with the firm preclude us fromdisclosing the name of the firm. We also cannot furnish details thatallow identification of the social networking website.

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Chardonnay”) or promotion (e.g., “buy one bottle atregular price and get another free”) related.5

When the firm launched the social media site, italso undertook a marketing campaign to inform itsclientele about these developments. As a result, exist-ing as well as potential customers began participat-ing by signing up and/or following the firm on itssocial media site. However, they differ with respectto their timing of participation, with some customersparticipating soon after the site was launched whereasothers did so at a later date. No incentives (eithermonetary or promotional) were offered, and cus-tomers signed up of their own volition.

3.2. Social Media Participation andCustomer Transaction Data

A key and notable contribution of our study is thatwe combine data related to customers’ participationin the focal firm’s social media site with their actualpurchases. This data gathering process is not straight-forward and presents a rather onerous task given thenature and idiosyncrasies associated with this kind ofwork. We adopted the following multi-step approachto identify customers who participate (e.g., becomea “fan” of the firm on Facebook) in the firms’ socialmedia site. We first begin with a survey that wassent to the firm’s customer base in early 2011. Thissurvey that was sent to 5,000 customers featured asection devoted to social media wherein customerswere specifically asked if they participate in the firm’ssocial media site in question (e.g., the firm’s Facebookpage).6

From the survey, we identified 845 customers whoactively participate in the firm’s social media site (e.g.,Facebook). Working in conjunction with the firm, wethen matched the above customers with actual par-ticipants of the specific social media site in question.This was possible because the firm was tracking itsparticipants on this site (e.g., Facebook) since theinception of its program. For this purpose, completenames were used and whenever there were multi-ple matches they were discarded and only uniqueexact names were retained. Additionally, these nameswere also matched with the transaction/purchasedata of the retailer and duplicate matches wereonce again eliminated. To be further sure about thisprocess, demographic and residence/location infor-mation obtained from the surveys for the respec-tive customers was utilized and this was crosstabulated with the corresponding information avail-able from the firm’s transaction data. Specifically,information pertaining to customers’ gender, age,

5 We note that no one particular type of posting dominates.6 The survey had 1,249 valid questionnaires returned, representinga response rate of 24.98%.

and race along with their residence address wasused. Only those customers whose demographicsand residence location/address matched closely wereretained. The process described above was doneentirely by the retailer, and we were furnished withthe data set containing unique codes assigned tocustomers upon masking/de-identifying all sensitiveinformation (e.g., names and residence location). Thecustomers’ participation in social media can be linkedto their transaction data using these unique customercodes. This process was adopted to ensure completecustomer confidentiality.

We use data from January 2008 to March 2011 toconduct our empirical analysis. As will be evidentin §4, we also need reasonable numbers of customerpurchases and sufficient data period post customerparticipation to reliably estimate our model param-eters. Consequently, we do not consider customerswho joined the social media site during the lastsix months of the data time period.7 The criteriadescribed above left us with 394 customers whom weinclude for further analysis.

3.3. Variable OperationalizationThe dependent variable for our analysis is customers’intensity of relationship with the firm. For our pur-poses, following precedence (e.g., Venkatesan andKumar 2004) we use visit frequency as a measure forthis variable. Accordingly, we analyze if customers’social media participation affects their frequency ofvisits and operationalize it as the number of timesa customer visits the store in a given time period(denoted by freq). With respect to operationalizationof the independent variables, customers’ participation(CParT) is a binary variable that takes the value 1if the customer is a participant in the firm’s socialmedia site (e.g., becomes a “fan” on the firm’s Face-book page) at the given time period and 0 otherwise.8

We note that different customers participate in thefirm’s social media site at different points of time dur-ing the estimation period. Social media activity rele-vant to a focal customer is measured by the numberof messages posted by the firm and all other cus-tomers. Because recent postings may be more impact-ful than older ones, we formulate a stock variable ofpostings (PostingsStock) in the following manner (e.g.,Tellis and Weiss 1995):

PostingsStockt =�Postingst+41−�5PostingsStockt−10 (1)

7 Because the customer participation period spans the time periodAugust 2009–March 2011, we consider only those customers whojoined during the time period August 2009–September 2010.8 In our case, customer participation takes the form of becominga fan of the firm, liking the firm, etc. Our sample customers donot “unparticipate” once they decide to participate/take part in thefirm’s social media site.

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


PostingsStock is the number of messages postedexcluding that of the focal customer at time t and�∈ 40115 is the decay coefficient.9 Purchase amount(PurAmt) is the spending incurred by the customeracross all products in a given time period. Buyingfocus (BuyFocus) represents how broad or narrow thebuying pattern is for a particular customer. In ourcase, because we consider wine purchases, follow-ing current literature in this area (Huerta et al. 1998),we use wine color type to capture this variable. Thecolor is a predominant characteristic for wines andserves as an important element for their classifica-tion because it depends on the grape variety (Jensenet al. 2008). Wine is classified broadly into three colortypes: white, red, and rosé. For each customer, dur-ing the calibration period,10 we calculate her spend-ing for each of these wine color types and designatethat color as the principal one on which her spendingis the most. The buying focus (calculated during theestimation period) then is defined as a dummy vari-able that takes the value 1 if the customer buys theprincipal color type at a given time and 0 otherwise.

To calculate the share of premium products (Premi-umShare), we first determine which products are pre-mium. For this purpose, we calculate the mean priceper unit ($/ml) for all the products and plot the distri-bution. On finding that this is distributed normal, wedesignate those products as being premium that areone standard deviation above the mean. The share ofpremium products bought by a focal customer is thenobtained as the ratio of the number of premium to thetotal number of products purchased by the customer.Deal sensitivity (Deal) designates how sensitive a cus-tomer is toward deals and is defined as the proportionof products the customer buys on promotion, i.e., thenumber of products bought on promotion divided bythe total number of products purchased in a giventime period.

We also include in our model several control vari-ables such as customer loyalty and demographics.We follow prior literature and measure loyalty (Loy-alty) as the share of wallet or the percentage of totalcustomer spending at the focal retailer in a giventime period (e.g., Ailawadi et al. 2008). To opera-tionalize this variable, we need to know the total cus-tomer spending on wine products across all stores(and not just the focal retailer). In order to computethis, we rely on the survey information wherein cus-tomers were specifically asked to report how much

9 The decay coefficient � accounts for the possibility that recentmessages would matter more than the earlier postings. To avoidcomputational burden, we do not estimate �; instead, we use a gridsearch procedure and find that 0.83 provides the best fit.10 We use 12 months of data prior to January 2008 as the calibra-tion period to calculate/initialize some of the model variables inthis case.

they spend on wine products in a typical year. Loyaltyis then defined as the ratio of the expenditures atthe focal retailer to the (self-reported) total spend-ing incurred on such categories in each time period.11

In case of demographics, we include the actual agein years (Age), the logarithm of customers’ income indollars (Income), and dummy variables for Gender andRace which take the value 1 if the customer is maleand white, respectively, and 0 otherwise.

4. Self-Selection Issue andEndogeneity

It is possible that customer specific unobserved fac-tors may jointly determine customers’ decision to par-ticipate in the firm initiated social media site and theirintensity of relationship with the firm. For example,customers’ positive predisposition toward the firmand its products (beyond customer loyalty that wecontrol for) may simultaneously influence both theirwillingness to participate in the social media site andtheir intensity of relationship with the firm. Thus,there could be endogeneity issues arising because ofself-selection that need to be considered for accurateidentification of the social media participation effect.

To resolve the above issues and establish a causallink between customers’ social media participationand their intensity of relationship with the firm, asrecommended by current literature (e.g., Huang et al.2012, Girma and Gorg 2007, Blundell and Dias 2000),we utilize propensity score matching in combinationwith difference-in-differences analysis. A DID estima-tor represents the difference in pre and post partici-pation differences between two groups of customers,the treatment group (i.e., customers who participate)and the control group (i.e., customers who do notparticipate). Note that the difference in the pre andpost behavior of participating customers alone maynot produce accurate results because of the presenceof extraneous factors that might affect behavior (e.g.,frequency of visits) in one period but not the other.Thus, a better research design entails using as ref-erence or benchmark customers who do not partici-pate in the firm initiated social media (i.e., the controlgroup). Comparing the behavior of the treatment andcontrol groups in the pre and post participation peri-ods helps to control for the influence of extraneousfactors. The challenge then is to compose the controlgroup that comprises customers who are very similar

11 In our case, we need to allocate the yearly (self-reported) totalexpenditures to specific time periods/shopping trips. We use his-torical spending patterns covering two full years (January 2006–December 2007) prior to our estimation time period to computeaverage weights, which are then used to assign expenditures,because spending may not be even or consistent across months.

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


to the treatment group customers (participating cus-tomers) in all respects but for the fact that they havenot participated in the firm hosted social media, incontrast to the treatment customers who have par-ticipated. The difference in behavior between thetwo groups (i.e., treatment and control) can then beattributed to social media participation.

In order to reliably construct the aforementionedcontrol group, we use PSM. This is because match-ing through propensity scores substantially improvesthe similarity between the treatment and the con-trol group because customers who are like treatmentcustomers are given a greater weight for inclusionin the control group, thereby improving the infer-ence related to DID analysis (Stewart and Swaffield2004, Abadie et al. 2010). In other words, the con-trol group (obtained through PSM) that is used as areference or comparison in the DID estimation willbe as similar as possible to the treatment group onthe observed characteristics, which helps to eliminatetemporally invariant estimation bias (Smith and Todd2005). Moreover, PSM helps us mimic a randomizedexperimental setup (Rubin 2006).

To summarize, we use the PSM procedure to con-struct a matched pair set of customers, one from the“treatment” group—comprising customers who par-ticipate in the firm initiated social media site andwhose behaviors we wish to analyze—and anotherfrom the “control” group—consisting of customerswho are not part of or never participate in the firm’ssocial media site. Subsequently, we employ the DIDestimation on the resulting matched sample to esti-mate the impact of social media participation. We notethat using PSM in combination with DID analysis, asdocumented by prior literature, considerably reducesthe bias resulting from both observable and unob-servable variables and helps improve the consistencyof the estimates (Stewart and Swaffield 2004, Caoet al. 2011). In addition, it improves the reliability ofthe estimation process (Cao et al. 2011) and is alsomore efficient as compared to other approaches (e.g.,instrument variables) because it obviates the need forstrong exclusion restrictions (e.g., Girma and Gorg2007, Blundell and Dias 2000).

We note that although instrument variable tech-nique is commonly used to correct for selection bias,it is not well suited for our purpose because find-ing good instruments for social media participation isdifficult and using weak instruments may only exac-erbate the problem (Bound et al. 1995). Furthermore,even if valid instruments were to be found, the esti-mation process is still fraught with perils because thesocial media participation enters as a discrete vari-able in the analysis, which can result in poorly identi-fied (Wooldridge 2002) and/or inconsistent estimates(Angrist and Pischke 2009).

4.1. Propensity Score MatchingThe objective of PSM is to select treatment andcontrol group customers who resemble each otherin all relevant characteristics before social mediaparticipation (the treatment), thereby creating a sta-tistical equivalence between the two groups by bal-ancing them on observables (Rosenbaum and Rubin1983, Brynjolfsson et al. 2010). For PSM to be feasible,data should be available for both the treatment andcontrol customers before the intervention (i.e., socialmedia participation). We utilize 82 weeks of data per-taining to customer characteristics for both groupsprior to the launch of the social media site. We fol-low recommendations of recent literature and use thesame data sources for both groups (Heckman et al.1997). We model customer participation decision as afunction of customer specific transaction and demo-graphic characteristics.

In the interest of exposition, we discuss the salientpoints here and provide the detailed description,including the steps we follow for conducting PSM—which are based on current literature (e.g., Caliendoand Kopeinig 2008, Guo and Frasier 2010)—in the(online) appendix (available at http://dx.doi.org/10.1287/isre.1120.0460). We begin by checking if thevariables used in the model are balanced between thetreatment and control customers using social mediaparticipation as the grouping variable through appro-priate bivariate tests and include them for furtheranalysis if they are significantly different across thetwo groups.

We then conduct the propensity score analysisusing a logistic model formulation because the pro-gram intervention (social media participation) isbinary.12 These results along with the fit statistics arereported in Table 1. We then perform the process ofmatching the treatment and control customers usingthe estimated propensity scores. Although manymatching algorithms are available, we use the optimalpair matching technique, wherein each customer whoparticipates is matched to another customer who doesnot participate. This procedure substantially improvesthe power of the analysis and reduces bias that canoccur when multiple (potentially) dissimilar treat-ment and control customers are matched (Guo andFrasier 2010, Huang et al. 2012). Additionally, optimalpair matching also enables us to treat each distinctmatched pair comprising treatment and control cus-tomers separately, thus allowing us to control for anyresidual variability emanating from social media par-ticipation effects (Huang et al. 2012). As a result of theabove, we are able to satisfactorily match 287 treat-ment customers to equivalent control customers, allof whom are included for model estimation.

12 We also fit a probit model for social media participation and theresults were similar.

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 1 Logit Regression Model of Customers’ Social MediaParticipation

Variable Parameter Std. error

PurAmt 000075∗∗∗ 000028BuyFocus −300751∗∗∗ 003975Deal 006113∗∗ 002941PremiumShare 000981∗∗ 000409Loyalty 304004∗∗∗ 102452Age −000685∗∗∗ 000112Gender 002053∗ 001117Income 001510 002322Race 004478∗ 002625Constant 303035 207626Log likelihood −30105350

∗p≤0010, ∗∗p≤0005, ∗∗∗p≤0001.

Next, we check to see if the underlying assumptionsof the PSM process hold. We check if there is sub-stantial overlap in the characteristics of the customerswho participate in the social media and those whodo not (i.e., common support condition). For this pur-pose, as recommended by Lechner (2002), we performa visual analysis of the propensity score distributionsthrough box-plots and histograms (Figure 1) and findevidence for the existence of common support.

To further ensure the quality of the PSM proce-dure, we check if the covariates are balanced betweenthe treatment and control groups in the pre and postmatching condition, the results of which are reportedin Table 2. As can be seen from the table, the variablesthat are used for PSM are not significantly differentacross the two groups of customers post matching,implying statistical balance.

Finally, we conduct sensitivity analysis to checkfor “hidden” bias—the presence of unobserved fac-tors that may affect the assignment into treatment(social media participation) and outcome (inten-sity of customer-firm relationship) simultaneously(Rosenbaum 2002). Essentially, this involves inves-tigating whether some unobservable factors overlyinfluence the treatment effect of customers’ socialmedia participation. The analysis based on the meth-ods expounded in the literature (e.g., Rosenbaum2002, Becker and Caliendo 2007) shows that “hidden”bias is not a serious concern in our case.

We report the overall summary statistics, whichare obtained by averaging the corresponding valuesacross the treatment and the matched control groups.With respect to number of postings, we find that boththe firm and its customers actively post messages onthe social media site. The average of the stock of thepostings variable (see Equation (1)) is 58.23. Althoughwe do not explicitly code the valence of the postings,as mentioned earlier, we find that all of the messagesinitiated by the retailer are positive and that morethan 90% of postings made by the customers are posi-tive in nature. Customers on an average spend $61.05

Figure 1 Distribution of Propensity Score Before and After Matching

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Box plot

Smooth histogram plot

TG CG_O CG_M

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.95

10

15

20

25

30

35

TGCG_OCG_M

Notes. The figure shows the box plots and smooth versions of histograms ofpropensity score distributions for following groups: TG—Treatment Group,CG_O—Control Group before Matching, CG_M—Control Group after Match-ing. Before matching, the distributions for treatment and control groups arequite different; however, after matching they are almost identical, providingevidence of common support. Furthermore, we also conduct a Kolmogorov-Smirnov (KS) test to compare the two distributions. The p-value of the KS-test between TG and CG_O is 0.0000 whereas p-value of the KS-test betweenTG and CG_M is 0.9377, which provide evidence of similarity of propensityscore distributions between TG and CG_M and hence the existence of com-mon support between the two distributions.

per time period. The large amount is a reflection ofthe nature of category (wine) and also that customersbuy multiple items on a trip. Premium selling prod-ucts constitute 6.69% of the total purchases across thecustomers. Customers seem to be buying substantialamounts on deal, which could be due to the frequencyof promotions that are offered. Moreover, they alsoexhibit considerable buying focus. Likewise, they arealso loyal to the firm (as measured by share of wallet)plausibly because the focal firm is one of the domi-nant players in the region. As for demographics, we

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 2 Summary Statistics and Covariate Comparison Before and After Matching

Control group

Treatment group Before matching After matching

Mean Mean Mean diff. t stat. Var. ratio Mean Mean diff t stat. Var. ratio

PurAmt 5905052 7208310 −1303259 −300462 006454 6206005 −300953 −100772 009442BuyFocus 005504 007690 −002186 −1002615 008730 006257 −000753 −102859 009354Deal 006405 006744 −000339 −107948 101000 006615 −000210 −009731 009828PremiumShare 000635 000769 −000134 −005611 101302 000701 −000066 −003045 100298Loyalty 008256 008353 −000097 −402823 008025 008296 −000040 −100153 009915Age 5004550 5601339 −506789 −701329 100922 5104079 −009529 −106229 100084Gender 004408 004017 000391 108698 100306 004342 000066 001152 100033Income 1009990 1009675 000315 107515 100545 1009828 000162 002455 100100Race 008737 008605 000132 106724 008974 008715 000022 002014 009857

Notes. Mean difference (Mean Diff.) for each covariate is calculated by subtracting the mean of control group from the mean of treatment group. The t statistics(t stat.) for these differences in mean are also reported. The variance ratio (Var. Ratio) for each covariate is calculated as a ratio of variance of treatment groupto variance of control group. The covariates are imbalanced and hence matching would be compulsory if mean difference is away from zero and variance ratiois away from one. The mean differences and variance ratios are close to 0 and 1, respectively, after matching, indicating the improvement in covariate balancingdue to matching.

find the mean age to be about 51 years, reflectingthe nature of the category (because 21 years is thelegal minimum for alcoholic purchases), and the meanincome is $59,334; 43.75% of the sample is male and87.26% of the customers in the sample are white.

5. Econometric Specification5.1. DID Model: Main Effect of Social Media

ParticipationFor each matched pair comprising treatment and con-trol customers (see §4.1), the logarithm of frequencyof visits is modeled as13

log4freqijt5 = �0j +�1 TreatDij +�2 CParTijt +�3 TreatDij

×CParTijt +äXij +�it0 (2)

In Equation (2), i denotes a matched pair of cus-tomers, j denotes a treatment group or a controlgroup customer, and t denotes the time period.TreatDij is the treatment dummy variable (thatequals 1 if the customer is in the treatment groupand 0 if the customer is in the control group),and CParTijt is a dummy variable denoting socialmedia participation, which takes the value 0 and 1for periods prior to and post participation, respec-tively, for customers belonging to the matched pair i.Xij represents a vector of demographic and behav-ioral control variables with ä being their correspond-ing estimated coefficients. �0j are the customer-specificfixed effects that capture the differences in baselinerelationship intensity and enable us to control forunobserved heterogeneity across customers. Note that

13 In the very few cases where the frequency is zero, we add asmall value to it (order of 0.001) so as to enable the logarithmictransformation.

in the above formulation consisting of matched treat-ment and control group customers, we use monthlydata that spans both pre and post launch time periodsof the social media site (January 2008–March 2011),which results in time-series data that are then stackedfor estimation. The main parameter of interest is �3,which captures the change in the frequency of visitsfor treatment customers post participation relative tocontrol customers who do not participate in firm ini-tiated social media.

5.2. DDD Approach: Moderating Effects ofHypothesized Variables

We now describe an alternative version of the modelto the one presented earlier (in Equation (2)) thatenables us to test our hypotheses discussed in §2.Specifically, we segment customers into high andlow levels for each of the moderating variables(i.e., social media activity, purchase amount, buyingfocus, deal sensitivity, and share of premium prod-ucts). To accomplish this, consistent with the meth-ods expounded in the current literature (e.g., Tuckeret al. 2012), we use a median split based on the val-ues of the above variables and divide them into highand low levels/types. Thus, for a certain variable(e.g., purchase amount) customers will be classifiedas belonging to the high level if they score greaterthan the median on that variable and low level oth-erwise. To make sure that the above process does notconfound the estimation time period, we use the cal-ibration period for this purpose.

The above approach allows us to test when thetreatment effect (i.e., the social media participationeffect) is more effective. Conducting such an analysismay be especially pertinent for managerial decisionmaking because it would help in understanding whenand for which customer segments social media efforts

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


yield greater returns.14 Therefore, to investigate theeffect of social media participation at different levelsof the moderating variables based on prior literature(e.g., Gruber 1994), we use the following formulation:

log4freqijt5

=�0j +�1TreatDij +�2CParTijt +�3TreatDij ×CParTijt

+�4TreatDij ×CParTijt ×PostingsStockij

+�5TreatDij ×CParTijt ×PurAmtij +�6TreatDij

×CParTijt ×BuyFocusij +�7TreatDij ×CParTijt

×Dealij +�8TreatDij ×CParTijt ×PremiumShareij

+ìXij +�it0 (3)

In Equation (3), the variables have the samemeaning as in Equation (2).15 As in the previouscase, we include customer specific fixed effects anddemographic/behavioral controls. Additionally, theequation includes three way interactions betweentreatment, social media participation, and each of themoderating variables (they take on the value of 1for high levels of the variable and 0 otherwise). Forinstance, TreatDij ×CParTijt×PurAmtij is the three wayinteraction between the dummies of treatment, par-ticipation, and the high level of purchase amount.Because of the presence of third level interactionsin Equation (3), the above framework results in adifference-in-difference-in-differences (DDD) specifi-cation. The key parameters of interest are the coeffi-cients associated with the three way interaction terms(�4 −�85. They capture the effect of social media par-ticipation on visit frequency for customers with highlevel of the corresponding variable (e.g., purchaseamount) relative to those customers with a low levelof the same variable.

6. Results6.1. DID Results for Visit FrequencyWe display the raw difference-in-differences valuesfor frequency of visits for the treatment and controlgroups before and after the social media site waslaunched by the firm (Table 3). As can be seen fromthe table, the mean visit frequency increase is notsignificant for the control group (2.486 post launchversus 2.417 pre-launch), but this increase is sig-nificant for the treatment group (2.762 post launchversus 2.540 pre-launch). Moreover, there is no sig-nificant difference in the frequency of visits between

14 We thank the AE for the suggestion.15 Although we include the various relevant two way interactionvariables in the model, for the sake of exposition, we do not presentand discuss the results related to these two way interaction vari-ables. More information is available from the authors upon request.

Table 3 Difference in Differences: Frequency of Visits of Treatmentvs. Control Groups

Group monthly mean

Treatment Control Difference t-test

After social media launch 207618 204855 002763 304652Before social media launch 205403 204169 001234 009886Difference 002215 000686 001529t-test 301764 105892

Notes. There is no significant difference in frequency of visits between treat-ment and control groups before the launch of social media; however, thedifference becomes significant after the social media launch. Because thesedifferences are taken for the matched sample, the significance of the dif-ference in post launch period can be attributed to customers’ social mediaparticipation.

the treatment and control group customers before thesocial media site is launched by the firm, whereas thedifference is significant post launch. This lends primafacie credence to the fact that participation in the firminitiated social media impacts the behavior of par-ticipating consumers. We investigate the above moreformally using the DID analysis next.

To benchmark the results and statistical fit of ourDID model, we estimate a series of alternative models(Table 4). Model (1) is the basic DID model withoutany control variables. As part of Model (2), we includedemographic variables as controls, and Model (3) con-tains behavioral variables in addition to demograph-ics as controls. Model (4) is an augmented version ofModel (3) that adds customer specific fixed effects.As is evident from Table 4, the results are substan-tively similar across the models. The fit statistic (R25increases from Model (1) to Model (4), which lendsface validity to the results.

We find that the parameter corresponding to thetreatment effect of social media participation (TreatD×

CParT) is consistently positive and significant acrossthe different models (for example, in Model (4) thisestimate is 0.0509). This suggests that participation inthe firm initiated social media positively influencescustomers’ intensity of relationship with the firm interms of their frequency of visits. Thus, we find sup-port for H1—the central hypothesis of our study. Inorder to assess the magnitude of the effect, we com-pute the elasticity measure using the correspondingparameter estimate from Model (4), the best fittingDID model. We find that elasticity of participation onvisit frequency to be 5.204,16 which implies that cus-tomers who participate in the firm’s social media visitthe firm about 5.2% more frequently compared to cus-tomers who do not participate.

16 Because TreatD×CParT is a dummy variable and the regres-sion equation is semi-logarithmic, the elasticity is given by 100×

6exp8�̂3 −4005Var4�̂3559−17, where �̂3 is the estimate of �3 andVar4�̂35 is its estimated variance (Kennedy 1981, Halvorsen andPalmquist 1980).

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 4 Impact of Social Media Participation on Customer Frequency of Visits

Model (4)Model (2) Model (3) Control for demographic Model (5)Control for Control for and behavioral variables Three-way difference

Model (1) demographic demographic and and customer model with customerNo controls variables behavioral variables fixed effects fixed effects (DDD)

Variables Estimate SD Estimate SD Estimate SD Estimate SD Estimate SD

TreatD 000508 000307 000383 000409 000327 000243 000502 002151 000342 002185CParT 000279∗ 000161 000853∗ 000464 001260∗ 000714 001542∗ 000906 001541∗ 000865TreatD×CParT 000619∗∗∗ 000201 000583∗∗∗ 000202 000566∗∗∗ 000218 000509∗∗∗ 000180 000500∗∗∗ 000129TreatD×CParT×PostingsStock 000002∗∗ 000001TreatD×CParT×PurAmt 000005∗∗∗ 000002TreatD×CParT×BuyFocus −000405∗∗∗ 000071TreatD×CParT×Deal −000436∗∗∗ 000162TreatD×CParT×PremiumShare 000837∗∗ 000345PostingsStock 000016∗∗ 000008 000020∗∗ 000008 000019∗∗ 000009PurAmt 000002∗∗∗ 000001 000003∗∗∗ 000001 000002∗∗∗ 000001BuyFocus −002177 001490 −000552 000520 −000533 000546Deal −000986 000676 −000301 000428 −000376 000472PremiumShare 001144∗∗ 000537 000861∗∗∗ 000286 000663∗∗ 000298Loyalty 001141∗∗ 000478 000055∗∗∗ 000019 000100∗∗∗ 000024Age 000018 000020 000019 000018 000018 000115 000024 000115Gender −000155 000154 −000103 000153 −000170 003743 −000280 003751Income 000021 000034 000019 000034 000282 002436 000049 003254Race −000190 000425 −000087 000420 −000404 003228 −000319 002444Intercept 008795∗∗ 001130 006022∗∗∗ 000583 000192∗∗∗ 000065

Customer fixed effects No No No Yes YesR-squared 0.1821 0.2359 0.3375 0.4527 0.5545

∗p≤0010, ∗∗p≤0005, ∗∗∗p≤0001.

6.2. DDD Analysis of High vs. Low TypeCustomers

We now focus on the results of the DDD model, whichis concerned with the effect of social media participa-tion for high and low levels of the moderating vari-ables. We report these results as Model (5) in Table 4.Because the DDD model also includes the two-wayinteraction of the treatment effect, the results can beinterpreted as the relative impact of the correspondinglevel of the moderating variable on visit frequency. H2suggests that the effect of customer participation onintensity of relationship will be greater for higher lev-els of activity or message postings in the social mediasite. We find evidence to support this (�4 =000002;p<0001). As per hypotheses H3, the impact of cus-tomer participation in firms’ social media is amplifiedfor customers with high purchase amount levels. H4suggests that the effect of social media participation islower for customers with a greater buying focus. Wefind support for both these hypothesized interactioneffects (�5 =000005; p<0001 for purchase amount and�6 =−000405; p<0001 for buying focus).

We find that the effect of social media partic-ipation is greater for customers with lower dealsensitivity (�7 =−0004363p<0001). Thus, we find sup-port for H5. Finally, we find that the effect ofsocial media participation is higher for customerswhose share of premium product purchases is greater

(�8 =0008373p<0001), thereby supporting H6. Withrespect to the control variables, we note that to be con-sistent with the DDD analysis, we report the resultswhen the variables were used in levels (similar resultswere obtained when actual values of the variableswere used). We find that higher level of activity in thesocial media site is associated with greater intensityof relationship between the firm and its customers.Moreover, customers with higher levels of spendingand greater shares of premium products visit the firmmore frequently, as do more loyal customers. We findno significant effects for buying focus, deal sensitiv-ity, and the demographic variables (i.e., age, genderand race) plausibly because of better covariate balancethat was achieved for these variables.

7. Economic Impact of CustomerSocial Media Participation

7.1. Profitability AnalysisIn this section, we analyze the benefits that can berealized by firms as a result of their foray into socialmedia. Specifically, we begin by documenting theincremental profits of customers who participate inthe social media (i.e., treatment group customers)vis-à-vis those who do not participate (i.e., controlgroup customers). For this purpose, we compute the

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 5 DID: Profitability of Treatment vs. Control Group

Groupmonthly mean ($)

Treatment Control Difference t-test

After social media launch 2502607 2306989 105618 303972Before social media launch 2202549 2200354 002195 102184Difference 300058 106635 103423t-test 209672 105924

Notes. There is no significant difference in total customer profits (in dollarsper month) between treatment and control group before the launch of socialmedia site; however the difference becomes significant post launch. Becausethese differences are taken for the matched sample, the significance of thedifference in post period can be attributed to the customers’ social mediaparticipation.

aggregated per period net total profits separately forcustomers in both groups before and after the launchof the social media site by the firm, the results ofwhich are reported in Table 5. Note that this is pos-sible in our case because we have data pertaining toboth retailer prices and costs at the individual prod-uct level. The values reported in the table are themean total profits across all the products purchasedby the respective group of customers. On comparingthe table values, we once again find that there is asubstantial increase in total profits for the treatmentgroup after the launch of the social media site ($25.261

Table 6 Impact of Social Media Participation on Customer Profitability

Model (4)Model (2) Model (3) Control for demographic Model (5)Control for Control for and behavioral variables Three-way difference

Model (1) demographic demographic and and customer model with customerNo controls variables behavioral variables fixed effects fixed effects (DDD)

Variables Estimate SD Estimate SD Estimate SD Estimate SD Estimate SD

TreatD 000098 000482 000108 000546 000609 000479 000247 002572 000927 002811CParT 000726∗ 000394 000878∗ 000462 000868∗ 000235 000225∗ 000125 000334∗ 000187TreatD×CParT 000578∗∗∗ 000318 000569∗∗∗ 000183 000561∗∗∗ 000187 000550∗∗∗ 000213 000505∗∗∗ 000186TreatD×CParT×PostingsStock 000007∗∗ 000003TreatD×CParT×PurAmt 000003∗∗∗ 000001TreatD×CParT×BuyFocus −000170∗∗∗ 000060TreatD×CParT×Deal −000110∗∗ 000052TreatD×CParT×PremiumShare 000315∗∗∗ 000109PostingsStock 000003∗∗ 000001 000004∗∗ 000002 000009∗ 000005PurAmt 000027∗∗∗ 000003 000028∗∗∗ 000003 000027∗∗∗ 000003BuyFocus −000975 000632 −001625 001221 −001278 000793Deal −001005 000931 −001491 001012 −001653 001096PremiumShare 000538∗∗∗ 000107 000389∗∗ 000179 000880∗∗∗ 000129Loyalty 202030∗∗∗ 000548 201367∗∗∗ 000640 201304∗∗∗ 000641Age 000003 000017 000001 000009 000221 000137 000213 000137Gender 001026 000947 000087 000175 009551 006475 009593 006481Income 000077 000356 000247 000179 000585 003859 00088 003887Race 001769 001258 000187 000482 004572 002912 005013 004920Intercept 300928∗∗∗ 009891 208112∗∗∗ 008107 109887∗∗∗ 002094

Customer fixed effects No No No Yes YesR-squared 0.2168 0.2864 0.4526 0.5407 0.6246

∗p≤0010, ∗∗p≤0005, ∗∗∗p≤0001.

pre launch versus $22.255 post launch). In contrast,there is no significant difference in total profits before($22.035) and after ($23.698) the launch of the socialmedia site for the control group. This suggests thatcustomer participation in social media improves firmprofitability.

To investigate the above issue more formally, weutilize the DID framework of Equation (2) with cus-tomer profits as the dependent variable and report theresults in Table 6. Like earlier, we estimate a series ofmodels ranging from the basic DID model sans anycontrol variables (Model 1) to the one that includesboth behavioral and demographic controls along withcustomer fixed effects (Model 4). We once again findthat the parameter related to the treatment effect ofsocial media participation (TreatD×CParT) is positiveand significant across all the different model specifi-cations. For instance, in Model (4) the parameter esti-mate is 0.0550, which corresponds to profit elasticityof 0.0563. This implies that customers who partici-pate in the firm’s social media contribute more to thefirm’s bottom line. We also estimate the DDD modelthat includes the three way interaction effects betweenthe treatment variable, customer social media par-ticipation, and the indicator variable for the highlevel of the moderating variables (Equation (3)). Wefind that the results have the same directionality andsimilar significance as reported for visit frequency

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


(i.e., Model 5 in Table 4). Thus, the above analysisestablishes that social media participation has a pos-itive impact on profitability (a key metric for firms)in a manner akin to intensity of relationship as mea-sured by visit frequency.

7.2. Segment Level AnalysisThe results from the DDD analysis establish that therelative impact of social media participation differsacross different customer segments (i.e., high versuslow type). However, from a managerial perspective,it might be more interesting to understand the actualmagnitude of the treatment effect for the two seg-ments. In other words, managers may be interested inunderstanding not only whether the impact of socialmedia is greater for high type versus low type cus-tomers but also the absolute magnitude of the impactfor each of these segments. This kind of segmentlevel analysis can help in understanding which seg-ments are more lucrative and should be allocatedgreater marketing dollars. To investigate this issue,we divide the sample data set into two subsamples,with one consisting of customers with high levels andthe other with low levels of the moderating variable,respectively (we do this for each of the moderatingvariables). We utilize the same process for dividingcustomers into high and low levels/types as was usedfor the DDD analysis (i.e., median split analysis asdescribed earlier in the paper). We then conduct theDID analysis for each customer segment to distin-guish the impact of social media participation at thesegment level for both visit frequency as well as prof-its (Model 4 formulation as presented in Equation (2)).

We provide the results of the above analysis inTable 7. In the interest of exposition, we present themain results concerning the treatment effect only. Wefind that the effect of participation for the high seg-ment is positive and significant in case of frequencyof visits for all the moderating variables. Interest-ingly, we also find that the impact of social mediaparticipation is still positive and significant for lowsegment customers. In other words, social media

Table 7 DID Estimates for Segment (High vs. Low) Level Analysis

Postings stock Purchase amount Buying focus Deal sensitivity Premium share

Variable High Low High Low High Low High Low High Low

Frequency of visitsTreatD 000718 000323 000612 000403 000104 000896 000078 000921 000912 000098CParT 001884 001186 002216∗ 000834∗ 001219∗ 001867∗ 001383∗ 001843∗ 002154∗ 000932∗

TreatD×CParT 000642∗∗∗ 000463∗∗ 000523∗∗∗ 000484∗∗∗ 000135∗∗∗ 000881∗∗∗ 000099∗∗∗ 000911∗∗∗ 000894∗∗∗ 000116∗∗∗

ProfitabilityTreatD 000307 000162 000462 000008 000026 000446 000099 000325 000492 000014CParT 000318 000133 000403∗ 000052∗ 000028∗ 000427∗ 000121∗ 000328∗ 000428∗ 000021∗

TreatD×CParT 000787∗∗∗ 000294∗∗∗ 000628∗∗∗ 000487∗∗∗ 000116∗∗∗ 000987∗∗∗ 000262∗∗∗ 000844∗∗∗ 000908∗∗∗ 000203∗∗∗

∗p≤0010, ∗∗p≤0005, ∗∗∗p≤0001.

participation influences the behavior of not only highlevel/segment customers but also those belonging tothe low level/segment with respect to their inten-sity of relationship with the firm. We obtain sim-ilar results for the customer profitability model aswell. We also represent these results pictorially in Fig-ure 2, wherein the four panels succinctly establish thatalthough social media participation matters for bothhigh and low type customers, the differences betweenthe two are more pronounced for certain moderatingvariables. Such an exercise can be useful in quantify-ing the benefits that can accrue to firms as a resultof their social media initiatives at the individual cus-tomer segment level and can also help discern therelative impact for each variable.

8. Robustness Checks8.1. Inclusion of Additional

Variables for MatchingIn addition to the original nine variables that areused for the main analysis, we check the validityof our results by using an augmented set compris-ing three additional variables: Internet risk aversion,technology comfortness, and social networking atti-tude. We choose to include these variables, whichare obtained from the survey for both participantsand nonparticipants, because they may be related tocustomers’ propensity to participate in firm initiatedsocial media activities. For example, customers whoscore high on social networking attitude may be morelikely to join the firm’s social media. The pertinentdetails including a description of the relevant vari-ables and sources used to construct them are pro-vided in the online appendix. Note that we do notinclude the above mentioned variables in the mainanalysis because it is recommended that they be col-lected prior to customers’ social media participation(Caliendo and Kopeinig 2008), whereas the survey(that is the source of these variables) was adminis-tered at a later date. We find that the results pertain-ing to the effect of social media participation using the

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Figure 2 DID Estimate Plots for Segment Analysis (High vs. Low Groups)

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

High

Low

Customer segments

Diff

eren

ce in

diff

eren

ce e

stim

ate

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

High

Low

Customer segments

Diff

eren

ce in

diff

eren

ce e

stim

ate

Frequency of visit

Profitability

Purchase amount

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

High

Low

Customer segments

Diff

eren

ce in

diff

eren

ce e

stim

ate

Deal sensitivity

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

High

Low

Customer segments

Diff

eren

ce in

diff

eren

ce e

stim

ate

Premium share

Buy focus

augmented set are substantively similar to the resultsthat we presented earlier. We summarize these resultsin the appendix.

8.2. Matching with Alternative TechniquesAs mentioned previously in §4.2, we use the opti-mal pair matching for our analysis. However, wealso verify the robustness of our results by employ-ing alternative matching methods. Specifically, weuse greedy matching (nearest neighborhood with-out calipers) and Mahalanobis distance (Dehejia andWahba 2002) and find that the results with these alter-native matching techniques are very similar to thoseobtained from the optimal pair matching method. Weprovide some pertinent results of the above alterna-tive matching techniques in the online appendix.

8.3. Robustness of the DID AnalysisWe conduct a series of robustness checks to confirmthe results of the DID analysis. For our case, we usedata consisting of 18 months pre and post launch ofthe social media site by the firm. To make sure that the

results are not because of some unobservable factorspresent during certain periods and are robust to thespecification of time windows, we conduct the follow-ing analysis. We run the DID models (for both visit fre-quency and profitability) wherein we restrict the posttreatment period (i.e., data post social media launch)to a time frame consisting of 3, 6, 9, 12, and 15 months,the results of which are reported in Table 8. We findthat the DID estimates are positive and significantin each of the above cases, indicating that customerparticipation in social media positively impacts visitfrequency and profitability. Moreover, the parameterestimate of the treatment effect gradually convergeswith the estimate at 15 months being close to thatobtained overall (i.e., using 18 months). For illustra-tion purposes, we also plot the DID estimates obtainedfrom the different time windows in the above analy-sis along with the 95% confidence interval in Figure 3.As can be seen from the graph, as expected, these esti-mates seem to be stabilizing with time.

To further check that the results of our analysisare not simply a consequence of spurious correlation

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 8 Results of DID Analysis with Different Time Windows for Treatment Periods

Treatment window (post launch period)

15 months 12 months 9 months 6 months 3 months

Variable Visits Profitability Visits Profitability Visits Profitability Visits Profitability Visits Profitability

TreatD 000297 000484 000298∗ 000485∗ 000302∗ 000488∗ 000305∗∗ 000492∗∗ 000317∗∗∗ 000502∗∗

40002235 40003205 40001735 40002895 40001635 40002615 40001235 40002215 40000235 40002215CParT 000572∗ 000281∗ 000702∗ 000440∗ 000768∗ 000523∗ 000832∗ 000594∗ 000877 000655

40003225 40001705 40004015 40002295 40004215 40002995 40004215 40003195 40006215 40005195TreatD×CParT 000501∗∗∗ 000534∗∗∗ 000478∗∗∗ 000498∗∗∗ 000439∗∗∗ 000457∗∗∗ 000371∗∗∗ 000391∗∗∗ 000267∗∗∗ 000283∗∗∗

40000325 40000285 40000335 40000305 40000355 40000325 40000395 40000355 40000485 40000445

∗p≤0010, ∗∗p≤0005, ∗∗∗p≤0001.

and/or due to model misspecification, we run a seriesof “phantom” regressions (Chetty et al. 2009). Forthese regressions, we consider only the data period(18 months) prior to the actual launch of the socialmedia site by the firm (i.e., from January 2008 toAugust 2009) and designate a random month fromthis period as the time of the launch of the site whencustomers could participate. For example, if we desig-nate the end of the ninth month as the time when thesocial media site went into effect, then months 10–18constitute the “treatment” period and months 1–9make up the “nontreatment” period. We run a seriesof DID regressions with different time periods (third,sixth, and ninth month) randomly designated as themonth when the social media site was launched, theresults of which are reported in Table 9. We find thatparameter estimate corresponding to the treatmenteffects is statistically indistinguishable from zero inall the cases. The above analysis with “placebo” esti-mates suggests that our results can be attributed tocustomers’ social media participation.

Because our data consist observations over time, apotential concern is the presence of serial correlation.

Figure 3 95% Confidence Interval for DID Estimates with Different Treatment Windows

2 4 6 8 10 12 14 16 180.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

Time window

Diff

eren

ce in

diff

eren

ce e

stim

ate

Upper 95%CIEstimateLower 95%CI

2 4 6 8 10 12 14 16 180.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

Time window

Diff

eren

ce in

diff

eren

ce e

stim

ate

Frequency of visits Profitability

To make sure that the above does not pose a largeconcern for the model results, we adopt the fol-lowing method based on Bertrand et al. (2004). Weaggregate the data set, by ignoring the time seriesaspect of it, so that we end up with two distinctdata points—one each for pre and post participationperiods—for each customer in the matched pair. Onreestimating the DID models using this aggregateddata set, we find that our results are substantially sim-ilar (Table 10). We additionally replicate our regres-sions using Newey-West corrected standard errorsand also perform Dickey-Fuller tests to make surethat our dependent and independent variables are sta-tionary. Finally, we also reestimate our DID modelswith random effects in lieu of customer specific fixedeffects and find that the results are very similar.

8.4. Cohort EffectIt is plausible that early participants affect the behav-ior of later participants. In our data set, because weobserve when the customers become participants ofthe firm initiated social media site, we classify theminto their respective cohorts (early or late participants)

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 9 DID Analysis for “Phantom” Regressions

“Phantom” treatment window

3 months 6 months 9 months

Variable Visits Profitability Visits Profitability Visits Profitability

TreatD −000049 000050 −000023 000053 −000014 00002340000395 40000365 40000325 40000505 40000305 40000275

CParT 000489 000189 000486 000195 000481∗∗ 000184∗∗∗

40003235 40001305 40003245 40001385 40002265 40000245

TreatD× 000029 000078 000029 000068 000033 000020CParT 40000555 40000515 40000465 40000425 40000425 40000395

∗p≤0010, ∗∗p≤0005, ∗∗∗p≤0001.

Table 10 Aggregate DID Results

Visit frequency Profitability

Variable Estimate SD Estimate SD

TreatD 000002 000020 000001 000019CParT 000498∗ 000280 000178∗ 000099TreatD×CParT 000049∗∗∗ 000018 000069∗∗∗ 000026PostingsStock 000012∗∗ 000006 000002∗∗ 000001PurAmt 000002∗∗∗ 3.33E−05 000001∗∗∗ 2.13E−05BuyFocus −000044 000054 −000128∗∗ 000050Deal −000092∗ 000050 −000016 000047PremiumShare 000753∗∗∗ 000264 001049∗∗ 000518Loyalty 000009∗∗∗ 000003 000304∗∗∗ 000116Age 000001 000001 000009 000008Gender −000030 000035 000042 000034Income 000002 000002 000004 000002Race −000021 000022 000024 000020Customer fixed effects Yes YesR-squared 0.3557 0.4694

∗p≤0010, ∗∗p≤0005, ∗∗∗p≤0001.

using a time based median split. We then reestimateour models separately for each cohort and find thesubstantive results to be similar.

8.5. Alternative Operationalization of VariablesWe check the sensitivity of the model results to dif-ferent operationalization of the variables. Specifically,we use the total quantity purchased by customers forpurchase amount. Similarly, we considered two stan-dard deviations above the mean as the cut-off for cus-tomers’ purchase of premium share products. We alsoconsidered transformation of the pertinent variablessuch as logarithm of purchase amount, age, and post-ings. We find that the results are very similar in allthe above cases.

9. Discussion, Conclusion, andDirections for Future Research

In this paper, we study the impact of customer partic-ipation in a firm’s social media site on an importantindicator of the intensity of customer-firm relation-ship, customer shopping visit, or purchase frequency.We further examine the moderating roles of social

media activity and customer characteristics. In addi-tion, we quantify the effect of customer social mediaparticipation on a key metric that contributes to firms’bottom line, i.e., customer profitability. We draw onCRM and social media literatures to formulate ourhypotheses and test them using a novel individuallevel transaction data set that is complemented withdata on social media participation for the same set ofcustomers. We rely on a quasi-experimental setup thatuses PSM in combination with DID analysis to ruleout customer self-selection issues concerning socialmedia participation and the intensity of customer-firm relationship as measured by frequency of visits.Our results show that customer engagement throughsocial media increases customers’ shopping visits.Our results also suggest that this effect is greater for agreater level of activity in the firm hosted social mediasite. Furthermore, we find that the effect is greaterfor customers who have higher levels of spendingand share of premium products and lower levels ofbuying focus and deal sensitivity. We find that theseresults hold for customer profitability as well.

9.1. Theoretical ImplicationsOur results extend findings from social media andCRM literatures while also contributing to them inseveral ways. First, although many studies haveattempted to analyze the impact of online socialmedia, they have primarily considered the influenceof online reviews on product sales (Chevalier andMayzlin 2006) or new product diffusion (Susarla et al.2012). Our study adds to this burgeoning literatureby analyzing the role of social media in driving busi-ness value for firms by linking customers’ socialmedia participation to the additional value created forthe firm.

Second, we view a firm’s social media efforts asan extension of its relationship building efforts withits customers. A significant stream of literature inCRM attempts to understand the drivers of customerequity and bring accountability to relationship man-agement expenditures (Rust et al. 2004). This researchhas been primarily limited to studying the impact offirms’ offline relationship building expenditures andcustomer characteristics. We add to this research andprovide insights into how firms’ online social mediaactivities may influence customer visit (or purchase)frequency, a key driver of customer lifetime value(Venkatesan and Kumar 2004).

Third, our study takes a holistic look at socialmedia participation and adds new insights to grow-ing research that considers the impact of communityparticipation on customer behavior. For example,Dholakia and Durham (2010) conduct surveys andfind that customers who became fans of a firm onFacebook reported increased spending with the firm

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


as compared to non-fans. Although these findings areinteresting, the study relies mainly on self-reporteddata and does not establish a causal connectionbetween Facebook participation and customer behav-ior. Another study, that of Algesheimer et al. (2010),reports that participation in the eBay communityleads to a decline in the amount of money spent withthe firm. Although this research establishes the effectof participation in an online community, these find-ings do not help understand how firms’ social mediaefforts can transform a customer-firm relationship,which is the focus of this study. Furthermore, ourstudy also helps tie customers’ social media participa-tion to their transaction relationship with the firm andthus provides managerial guidelines for better deci-sion making. Finally, the few studies that examine theimpact of social media tend to be aggregate in nature(e.g., Mudambi and Schuff 2010). Although such stud-ies help provide an overview, analyzing the abovephenomenon at the individual level and adopting acustomer centric view enriches the extant social medialiterature while providing novel insights into behav-ior subsequent to customer social media participation.

9.2. Managerial ImplicationsThe results of our study yield several important man-agerial implications. As customers spend more timeon social media sites, managers are devoting moreresources to them. However, in turbulent times whenfirms are under increased pressure to justify all theirpromotional budgets, social media spending needs tobe more financially accountable. Thus, our researchhelps managers gain a better understanding of returnon investment accruing to social media initiatives.Based on our findings, we offer the following recom-mendations for practice.

Nurture Customer Relationships Through Social Media.Our results show that firms’ online social media activ-ities help strengthen the bond between the customerand the firm and contribute to financial performancein the long run by increasing customer visit frequen-cies and profitability. Although our findings suggest apositive effect of customer social media participationon customer visit frequency and profitability, man-agers must be cautious and not interpret this result toimply that simply creating, say, a Facebook page andinviting customers to become fans will lead to posi-tive customer outcomes. They must carefully manageand devise opportunities to create and nurture rela-tionships with customers through the social medium.For example, maintaining a user-friendly social mediasite interface, providing regular updates about events,sending personalized messages to individual cus-tomers, and encouraging member contributions cancreate interactive communication that can enhancefirm equity (Agarwal et al. 2008).

Be Wary; Not All Customers are Created Equal. Ourresults also provide actionable recommendations formanagers in terms of segmenting and targeting cus-tomers based on customer purchase histories andinteractions with the firm. Our results suggest thatmanagers need to understand that not all customerswill have a similar response to the firm’s social mediaefforts. It is vital that managers integrate their knowl-edge about customers from both offline transactionsand online social media sources in order to betterserve them.

Create Product Based Social Media Forums. Besidesfinding a differential response of social media partic-ipation across different types of customers, we alsofind an interaction effect between social media par-ticipation and customers’ shopping baskets. Thus,managers can increase the response to social mediaparticipation by creating specialized subcommunitiesor discussion forums for customers looking for pre-mium and unique products.

In summary, we believe the results of the studyprovide unique theoretical and managerial insightsinto the relationship between a firm’s social mediaefforts and firm performance. However, the study suf-fers from some limitations that can be addressed byfuture research. First, we are able to provide resultsfrom only one firm. It would be interesting to studywhether the results are generalizable for other com-panies. Furthermore, checking for generalizability ofresults across different forms of social media wouldalso be an interesting addition to this research. Sec-ond, although we studied the impact of social mediapostings, due to data limitations we were not ableto examine how the valence of messages affects cus-tomers’ purchase behavior. In particular, we note thatmost of the message postings in our data are pos-itive, which limits us from studying the impact ofvalence of messages. Next, future research could alsoextend the current findings by studying whether dif-ferent types of message postings by firm (e.g., productor price related) will elicit a different response fromcustomers. Finally, the results of this study also suf-fer from the limitation that they are specific to thesocial media strategy employed by the firm. Futureresearch can address this by examining the impactof different social media strategies across multiple,disparate firms.

Electronic CompanionAn electronic companion to this paper is available aspart of the online version at http://dx.doi.org/10.1287/isre.1120.0460.

AcknowledgmentsThe authors thank Professor Arun Jain and the ResearchGroup in Integrated Marketing (RIM) at the School of

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Management, SUNY-Buffalo, for accesses to the data andresearch assistance. They thank the participants of the ISRspecial issue workshop at University of Maryland and theINFORMS 2012 Marketing Science Conference. They alsothank the Senior Editor, the Associate Editor, and the twoanonymous reviewers for valuable comments and sugges-tions. The authors contributed equally to the article and theauthor ordering is reverse alphabetical. All errors are theauthors’ own.

ReferencesAbadie A, Diamond A, Hainmueller J (2010) Synthetic control

methods for comparative case studies: Estimating the effect ofCalifornia’s Tobacco Control Program. J. Amer. Statist. Assoc.105(490):493–505.

Agarwal R, Gupta AK, Kraut R (2008) The interplay between digitaland social networks. Inform. Systems Res. 19(3):243–252.

Agarwal R, Animesh A, Prasad K (2009) Social interactions and the“digital divide”: Explaining variations in Internet use. Inform.Systems Res. 20(2):277–294.

Ailawadi KL, Neslin SA, Gedenk K (2001) Pursuing the value-conscious consumer: Store brands versus national brand pro-motions. J. Marketing 65(1):71–89.

Ailawadi KL, Pauwels K, Steenkamp J-BEM (2008) Private-label useand store loyalty. J. Marketing 72(6):9–30.

Algesheimer R, Dholakia UM, Hermann A (2005) The social influ-ence of brand community: Evidence from European car clubs.J. Marketing 69(3):19–34.

Algesheimer R, Borle S, Dholakia U, Singh SS (2010) Impact ofcustomer community participation on customer behaviors:An empirical investigation. Marketing Sci. 29(4):756–769.

Anderson EW, Sullivan MW (1993) The antecedents and con-sequences of customer satisfaction for firms. Marketing Sci.12(2):125–143.

Angrist JD, Pischke JS (2009) Mostly Harmless Econometrics: AnEmpiricist’s Companion (Princeton University Press, Prince-ton, NJ).

Becker SO, Caliendo M (2007) Sensitivity analysis for average treat-ment effect. Stata J. 7(1):71–83.

Bertrand M, Duflo E, Mullainathan S (2004) How much shouldwe trust differences-in-differences estimates? Quart. J. Econom.119(1):249–275.

Blundell R, Dias MC (2000) Evaluation methods for non-experimental data. Fiscal Stud. 21(4):427–468.

Bolton RN (1998) A dynamic model of the duration of the cus-tomer’s relationship with a continuous service provider: Therole of satisfaction. Marketing Sci. 17(1):45–65.

Bound J, Jaeger D, Baker R (1995) Problems with instrumental vari-ables estimation when the correlation between the instrumentsand the endogenous explanatory variable is weak. J. Amer.Statist. Assoc. 90(430):443–450.

Brynjolfsson E, Hu Y(J), Smith MD (2010) The longer tail: Thechanging shape of Amazon’s sales distribution curve. http://dx.doi.org/10.2139/ssrn.1679991.

Caliendo M, Kopeinig S (2008) Some practical guidance for theimplementation of propensity score matching. J. Econom. Sur-veys 22(1):31–72.

Cao Z, Song X, Hamilton L (2011) Difference-in-difference andinstrumental variables approaches: An alternative and com-plement to propensity score matching in estimating treatmenteffects. Thomson Reuters; CER Issue Brief, http://healthcare.thomsonreuters.com/cer/assets/DifferenceCERIssueBrief.pdf.

Chetty R, Looney A, Kroft K (2009) Salience and taxation: Theoryand evidence. Amer. Econom. Rev. 99(4):1145–1177

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

Crosby LA, Evans KR, Cowles D (1990) Relationship quality in ser-vices selling: An interpersonal influence perspective. J. Market-ing 54(3):68–81.

Dehejia RH, Wahba S (1999) Causal effects in non-experimentalstudies: Reevaluating the evaluation of training programs.J. Amer. Statist. Assoc. 94(448):1053–1062.

Dehejia RH, Wahba S (2002) Propensity score-matching meth-ods for non-experimental causal studies. Rev. Econom. Statist.84(1):151–161.

Dellarocas C, Zhang X, Awad NF (2007) Exploring the value ofonline product reviews in forecasting sales: The case of motionpictures. J. Interactive Marketing. 21(4):23–45.

Dholakia UM, Durham E (2010) One café chain’s Facebook experi-ment. Harvard Bus. Rev. 88(3):26.

DiPrete T, Gangl M (2004) Assessing bias in the estimation of causaleffects: Rosenbaum bounds on matching estimators and instru-mental variables estimation with imperfect instruments. Sociol.Methodology 34(1):271–310.

Forman C, Ghose A, Wiesenfeld B (2008) Examining the relation-ship between reviews and sales: The role of reviewer iden-tity discloser in electronic markets. Inform. Systems Res. 19(3):291–313.

Forrester Research (2008) The Growth of Social Technology Adoption(Forrester Research, Cambridge, MA).

Girma S, Gorg H (2007) Evaluating the foreign ownership wagepremium using a difference-in-differences matching approach.J. Internat. Econom. 72(1):97–112.

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

Gruber J (1994) The consumption smoothing benefits of unemploy-ment insurance. Amer. Econom. Rev. 87(1):192–205

Guo S, Frasier MW (2010) Propensity Score Analysis: Statistical Meth-ods and Applications (Sage, Menlo Park, CA).

Halvorsen R, Palmquist R (1980) The interpretation of dummyvariables in semilogarithmic equations. Amer. Econom. Rev.70(3):474–475.

Heckman J, Ichimura H, Todd P (1997) Matching as an econometricevaluation estimator: Evidence from evaluating a job trainingprogramme. Rev. Econom. Stud. 64(4):605–654.

Ho SY, Bodoff D, Tam KY (2011) Timing of adaptive Web person-alization and its effects on online consumer behavior. Inform.Systems Res. 22(3):660–679.

Huang Q, Nijs VR, Hansen K, Anderson ET (2012) Wal-Mart’simpact on supplier profit. J. Marketing Res. 49(2):131–143.

Huerta MRS, Masoud T, Alonso GL (1998) Wine differentiationaccording to color using conventional parameters and volatilecomponents. J. Food Composition Anal. 11(4):363–374.

IBM Report (2011) 11 Key Marketing Trends for 2011. Unica/IBM,New York, NY.

Janiszewski C (1993) Pre-attentive mere exposure effects. J. Con-sumer Res. 20(3):376–392.

Jensen JS, Demiray S, Egego M, Meyer AS (2008) Prediction of winecolor attributes from the phenolic profiles of red grapes (Vitisvinifera). J. Agricultural Food Chemistry 56(3):1105–1115.

Kennedy BW (1981) Variance component estimation and predictionof breeding values. Canadian J. Genetics Cytology 23(4):565–578.

Kietzmann JH, Hermkens K, McCarthy IP, Silvestre BS (2011) Socialmedia? Get serious! Understanding the functional buildingblocks of social media. Bus. Horizons 54(3):241–251.

Lechner M (2002) Some practical issues in the evaluation of het-erogeneous labor market programmes by matching methods.J. Royal Statist. Soc. A 165(1):59–82.

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


McAlexander JH, Schouten JW, Koenig HF (2002) Building brandcommunity. J. Marketing 66(1):38–54.

Morgan RM, Hunt SD (1994) The commitment-trust theory of rela-tionship marketing. J. Marketing 58(3):20–38.

Mudambi SM, Schuff D (2010) What makes a helpful online review?A study of customer reviews on Amazon.com. MIS Quart.34(1):185–200.

Pavlou P, Liang H, Xue Y (2007) Uncertainty and mitigating uncer-tainty in online exchange relationships: A principal–agent per-spective. MIS Quart. 31(1):105–131.

Reinartz WJ, Kumar V (2000) On the profitability of long-life cus-tomers in a noncontractual setting: An empirical investigationand implications for marketing. J. Marketing 64(4):17–35.

Reinartz WJ, Kumar V (2003) The impact of customer relation-ship characteristics on profitable lifetime duration. J. Marketing.67(1):77–99.

Rigby DK, Reichheld FR, Schefter P (2002) Avoid the four perils ofCRM. Harvard Bus. Rev. 7(2):101–109.

Rosenbaum PR (2002) Observational Studies (Springer, New York).Rosenbaum PR, Rubin DB (1983) The central role of the propen-

sity score in observational studies for causal effects. Biometrika70(1):41–55.

Rubin DB (2006) Matched Sampling for Causal Effects (CambridgeUniversity Press, New York).

Rust RT, Lemon KN, Zeithaml VA (2004) Return on marketing:Using customer equity to focus marketing strategy. J. Marketing68(1):109–127.

Smith JA, Todd PE (2005) Does matching overcome LaLonde’s cri-tique of nonexperimental estimators? J. Econometrics 125(1–2):305–353.

Stewart MB, Swaffield JK (2004) The other margin: Do minimumwages cause working hours adjustments for low-wage work-ers? Working paper, University of Warwick, UK.

Susarla A, Oh J-H, Tan Y (2012) Social networks and the diffusionof user-generated content: Evidence from YouTube. Inform. Sys-tems Res. 23(1):23–41.

Tellis GJ, Weiss D (1995) Does TV advertising really affect sales?J. Advertising 24(3):1–12.

Tucker C, Zhang J, Zhu T (2012) Days on market and homesales. Working paper. MIT Sloan School of Management,Cambridge, MA.

Venkatesan R, Kumar V (2004) A customer lifetime value frame-work for customer selection and resource allocation strategy.J. Marketing 68(1):106–125.

Verhoef PC, Donkers B (2005) The effect of acquisition channelson customer loyalty and cross-buying. J. Interactive Marketing19(2):31–43.

Verhoef PC, Franses PH, Hoekstra JC (2001) The impact of satisfac-tion and payment equity on cross-buying: A dynamic modelfor a multi-service provider. J. Retailing 77(3):359–378.

Wangenheim FV, Bayón T (2007) The chain from customer satisfac-tion via word-of-mouth referrals to new customer acquisition.J. Acad. Marketing Sci. 35(2):233–249.

Wooldridge JM (2002) Econometric Analysis of Cross Section and PanelData (MIT Press, Cambridge, MA).

ZDNet (2011) Two thirds of small business owners use Facebookfor marketing. Available at http://www.zdnet.com/blog/facebook/two-thirds-of-small-business-owners-use-facebook-for-marketing/1529. Accessed October 26, 2012.

Zhu F, Zhang X (2010) Impact of online consumer reviews on sales:The moderating role of product and consumer characteristics.J. Marketing 74(2):133–148.

Zhu L, Benbasat I, Jiang Z (2009) Let’s shop online together: Anempirical investigation of collaborative online shopping sup-port. Inform. Systems Res. 21(4):872–891.

Dow

nloa

ded

from

info

rms.

org

by [

130.

233.

35.2

43]

on 2

1 A

ugus

t 201

6, a

t 03:

29 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

The Effect of Customers' Social Media Participation on ... · tiveness of ﬁrms’ social media...

Documents

Transcript of The Effect of Customers' Social Media Participation on ... · tiveness of ﬁrms’ social media...