Rishika, R., Kumar, A., Janakiraman, R., & Bezawada, R. (2013) - The Effect of Customers' Social

Information Systems Research
Vol. 24, No. 1, March 2013, pp. 108–127

ISSN 1047-7047 (print) ISSN 1526-5536 (online) http://dx.doi.org/10.1287/isre.1120.0460
© 2013 INFORMS
The Effect of Customers’ Social Media Participation

on Customer Visit Frequency and Profitability:
An Empirical Investigation
Rishika Rishika
Mays Business School, Texas A&M University, College Station, Texas 77843,
[email protected]
Ashish Kumar
Department of Marketing, Aalto University School of Business, FI-00076 Aalto, Finland,
[email protected]
Ramkumar Janakiraman
Mays Business School, Texas A&M University, College Station, Texas 77843,
[email protected]
Ram Bezawada
School of Management, State University of New York, Buffalo, New York 14260,
[email protected]
I n this study we examine the effect of customers’ participation in a firm’s social media efforts on the intensity
of 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 the
link between social media participation and the intensity of customer-firm relationship. Importantly, we also
quantify the impact of social media participation on customer profitability. We assemble a novel data set that
combines customers’ social media participation data with individual customer level transaction data. To account
for endogeneity that could arise because of customer self-selection, we utilize the propensity score matching
technique in combination with difference in differences analysis. Our results suggest that customer participation
in a firm’s social media efforts leads to an increase in the frequency of customer visits. We find that this
participation effect is greater when there are high levels of activity in the social media site and for customers
who exhibit a strong patronage with the firm, buy premium products, and exhibit lower levels of buying focus
and deal sensitivity. We find that the above set of results holds for customer profitability as well. We discuss
theoretical implications of our results and offer prescriptions for managers on how to engage customers via
social media. Our study emphasizes the need for managers to integrate knowledge from customers’ transactional
relationship 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 visit
frequency; 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 January
12, 2012, and was with the authors 4 months for 2 revisions. Published online in Articles in Advance
December 20, 2012.
1. Introduction while also providing them the opportunity to share

Social media has changed the way individuals inter- information about products (Godes and Mayzlin 2004,
act with each other and is redefining the way firms Agarwal et al. 2008). Thus, social media has provided
connect with their customers. With 75% of Inter- firms—both large and small—with a new tool for cus-
net users participating in some form of social media tomer engagement.
(Forrester Research 2008), companies are increasingly Despite the eagerness on the part of firms to em-
investing in the creation of social media platforms brace social media to connect with customers, there
and strategies to augment customer-firm interactions. is also much skepticism about its efficacy. For exam-
A recent article in ZDNet (2011) reports that 69% ple, a recent IBM report (2011, p. 3) mentions that
of small business owners use some form of social “social media is no longer the adorable baby everyone
media (e.g., Facebook, Twitter) and about 78% of wants to hold, but the angst-filled adolescent—still
them plan to increase their budgets devoted to this immature yet no longer cute—who inspires mixed
medium. Social media enables firms to converse feelings.” The doubts about the effectiveness of social
about their offerings with customers interactively media arise because the link between firms’ social
108
Rishika et al.: The Effect of Customers’ Social Media Participation on Customer Visit Frequency and Profitability
Information Systems Research 24(1), pp. 108–127, © 2013 INFORMS 109
media efforts and the return on their investment has have an affinity toward the firm are also more likely
not been established. In particular, no study, using to join its social media site and have an intensive rela-
individual level data to our knowledge, has connected tionship with the firm. To control for this self-selection
customers’ social media participation with their actual issue, we employ the combination of propensity
behavior. Recent studies in the broad area of social score matching (PSM) technique and difference-in-
media have primarily focused on the effectiveness of differences (DID) analysis. In particular, we utilize
user generated content (UGC) in stimulating product a quasi-experimental research design with two dis-
sales (Godes and Mayzlin 2004, Dellarocas et al. 2007, tinct groups of customers: (1) a “treatment” group
Forman et al. 2008, Zhu and Zhang 2010), encourag- that comprises customers who participate in the firm
ing content diffusion (Susarla et al. 2012), and foster- initiated social media site and whose behaviors we
ing acceptance of recommendations (Ho et al. 2011). wish to analyze and (2) a “control” group consist-
However, from the perspective of a firm that is con- ing of customers who are not part of or never par-
templating social media investments, it is imperative ticipate in the firm’s social media site. We use PSM
to know if such expenditures help in furthering the in combination with DID analysis because it helps
intensity of customer-firm relationships and in cre- mimic a random experimental design, thus produc-
ating sustainable firm value. In this regard, because ing an unbiased estimate of the “treatment effect,”
most of the extant studies use aggregate data, they i.e., the impact of customer social media participation
are unable to uncover critical individual level differ- (Dehejia and Wahba 1999, Huang et al. 2012). We then
ences that are essential for obtaining insights at the extend the DID modeling framework to examine how
customer level. Thus, the objective of this paper is to the social media participation effect varies across cus-
systematically examine the effect of customers’ par- tomers depending on their level of relationship with
ticipation in firm initiated social media efforts on the the firm.
business value generated thereof for the firm. For this Our study makes three substantial contributions
purpose, we choose to work with customers’ shop- to the social media literature. First, we attempt to
ping visit frequency as the focal variable because it uncover the direct benefits that accrue to firms as
captures the intensity of relationship between a firm a consequence of their social media efforts. Our
and a focal customer and serves as a key driver approach is different from the extant research (e.g.,
of customer lifetime value (Venkatesan and Kumar Dholakia and Durham 2010, Algesheimer et al. 2010)
2004). Furthermore, we also examine the impact of in that we use actual behavioral data to estab-
customer social media participation on customer prof- lish the causal link between customers’ social media
itability because that directly contributes to a firm’s participation and their buying behavior. Moreover,
bottom line. unlike the above studies, we explicitly account for
To study the central question concerning the effec- endogeneity issues arising because of self-selection.
tiveness of firms’ social media strategy, we first exam- Second, no study to our knowledge has integrated
ine the main effect of customers’ participation in customers’ social media participation behavior with
social media on the intensity of relationship between their transaction relationship with the firm and stud-
customers and the firm (as measured by customers’ ied how customer-firm relationships in online and
frequency of shopping visits). We then hypothesize offline domains may interact to create value for the
and test for the moderating role of overall activity in firm. This study helps fill this gap in the literature.
the social media site and a set of customer charac- Third, our research also examines how the effect of
teristics. Because contribution margin from customers social media participation behavior varies across dif-
is another key component of the customer lifetime ferent types of customers. By focusing on transac-
value, we next complement our core findings related tion based customer characteristics, we offer insights
to visit frequency by examining the effect of social on when participation becomes more salient. More-
media participation on customer profits. In order to over, we quantify the economic benefits that result
accomplish our objectives, we assemble a novel data for firms due to customers’ social media participa-
set in which we observe customer participation deci- tion. Managers can use these results to understand
sions in a firm initiated social media site and con- how customer relationships cultivated in the online
nect this to individual purchase data for the same set social media environment can be leveraged to opti-
of customers. We believe that our study is the first mize return on investment.
to establish the link between customers’ social media
participation and the transaction driven firm value
generated thereof using actual behavioral data. 2. Hypotheses Development
However, in our research context, it is important The focus of the study is on examining how social
to account for endogeneity issues arising because of media efforts by the firm help transform the inten-
customer self-selection. For instance, customers who sity of customer-firm relationships. Accordingly, we
110 Information Systems Research 24(1), pp. 108–127, © 2013 INFORMS
measure this through a key characteristic representing Second, participation in a firm’s social media site
customers’ transactional relationship with the firm, provides an easy access to a social network consisting
frequency of visits. We believe that customer visit of other customers’ opinions and experiences. An
(or purchase) frequency captures the strength of active network with message postings from other
customer-firm relationship because prior research has customers regarding their experiences with the firm
considered visit frequency to be an important indi- and its products creates the feeling of a brand/firm
cator of the relationship that is also positively linked community that will positively influence a focal
to customer lifetime value (e.g., Venkatesan and customer’s intensity of relationship with the firm.
Kumar 2004). Research suggests that relationships among fellow
In examining the impact of social media partici- customers, who are an integral part of a community,
pation on the intensity of customer-firm relationship, help in the fostering of firm loyalty, thereby inten-
we focus on two sets of factors that can influence sifying the relationship with the firm (McAlexander
this relationship: (a) overall social media activity and et al. 2002).
(b) customer characteristics. Social media has become Third, participating in a firm hosted social media
a preferred platform for cultivating meaningful rela- site provides customers easy access to information on
tionships between firms and their customers. How- firms’ offerings and other related messages. Through
ever, from a managerial perspective, it is crucial to these postings, customers can infer the firm’s level
understand the moderating factors that make cus- of involvement with its customer base and its com-
tomers’ social media participation more or less salient. mitment toward enhancing their experiences. Such
In particular, we investigate how the level of activ- increased interactions can lead to better customer
ity on the social media site in terms of number of identification with the firm and create greater trust
messages and comments posted by the firm and cus- and customer loyalty (Algesheimer et al. 2005). Thus,
tomers moderates the effect of social media participa- firm hosted social media opens a new channel of
tion. Furthermore, managers often segment customers communication through which customers can connect
based on their transaction characteristics for better and communicate directly with the firm and other
targeting and optimal allocation of resources. For customers that will positively affect the intensity of
business executives, it is important to know on which customer-firm relationship. Based on the above dis-
customer segments they must spend greater resources cussion, we hypothesize the following:
in order to build and strengthen relationships (Rigby
et al. 2002). Thus, an understanding of how the impact Hypothesis 1 (H1). Participation in firm hosted social
of social media participation differs across customer media by a focal customer will have a positive impact on
segments can enable firms to develop better customer the intensity of the customer-firm relationship.
relationship management (CRM) initiatives. We draw
2.2. Interaction Effect with Social Media Activity
on CRM literature to identify key customer character-
Customers who choose to engage with the firm
istics (i.e., purchase amount, buying focus, deal sensi-
through its social media site gain easier access to
tivity, and premium product purchases) that can help
in ascertaining the differential response among differ- messages from both the firm and other customers.
ent customer segments and aid in managerial decision We believe that the impact of social media partic-
making. ipation on customer-firm relationship will be influ-
enced by the level of activity in the social media
2.1. Customers’ Participation in Firm Hosted site in terms of the number of messages posted by
Social Media the firm and other customers. First, because one of
By participating in the firm initiated social media the main purposes of the firm hosted social media
site, customers can strengthen their relationship with site is to create a platform for communicating with
the firm through many processes. First, the social its customers and building relationships with them,
media platform provides an opportunity for cus- a firm must maintain a high level of activity on the
tomers to voice their opinions and share experiences social media site. A more active and vibrant com-
about the firm with other customers as well as the munity with regular new message postings will cre-
firm, thereby creating social interactions (Agarwal ate trust and allow customers to infer the level of
et al. 2009). In the words of Kietzman et al. (2011, a firm’s relationship commitment and bolster cus-
p. 241), “social media employ mobile and web-based tomers’ bond with the firm. Extant research suggests
technologies to create highly interactive platforms via that commitment and trust in a relationship are pos-
which individuals and communities share, co-create, itively associated with a party’s intentions to con-
discuss, and modify user-generated content.” Thus, tinue the relationship with the other party (Morgan
this ability of social media to facilitate greater inter- and Hunt 1994). Upon creating social media sites
actions between consumers and the firm can help in (e.g., Facebook pages, Twitter accounts), if firms fail
establishing a deeper connection between them. to interact regularly with their clientele through this
new medium, it could create customer skepticism outcomes, which strengthens the relationship between
and result in less favorable behavioral outcomes. Sec- the firm and the consumer (Reinartz and Kumar
ond, a higher number of postings by other customers 2003). Thus, customers who spend a higher amount
reflects the level of customers’ involvement with the with the firm are likely to be more committed
firm. Individuals often wish to validate their choices and more relationship oriented than are customers
in an environment that further strengthens their own who spend less with the firm (Crosby et al. 1990).
affinity with the firm (Zhu et al. 2009). In the context Because they have a greater affinity toward the firm,
of a social media site, by scrutinizing other customers’ customers with a larger pecuniary investment will
behavior and comments, focal customers can feel exhibit a greater response to the firm’s relationship
more confident about their own choices, thus impact- building communications through social media than
ing their intensity of relationship with the firm.1 will customers with lower financial investments. Sec-
Next, a mere exposure effect may also exist that ond, prior research also shows that customers with
may make the firm more accessible to the cus- higher spending levels, due to their greater finan-
tomer (Janiszewski 1993). As customers are exposed cial investments with the firm, feel that they are
to a greater number of messages from the firm and more important to the firm; therefore, they are more
other customers about its products, they become demanding and expect greater relationship invest-
more favorably disposed toward the firm. Thus, an ments on the part of the firm (Wangenheim and
active social media page with regular new messages/ Bayón 2007). Thus, high value customers are likely to
postings can help customers form more positive atti- value firms’ relationship investments in social media
tudes toward the firm and strengthen the customer- more than low value customers are. Hence, we expect
firm relationship. Hence, we hypothesize that the response to social media participation to be greater for
impact of customers’ social media participation on the customers with higher spending levels as compared
intensity of their relationship with the firm will be to customers with lower spending levels. Hence, we
greater for a higher level of activity (e.g., messages, hypothesize the following:
comments) on the social media site. Hypothesis 3 (H3). The impact of participation in firm
Hypothesis 2 (H2). The impact of participation in firm hosted social media on the intensity of the customer-firm
hosted social media on the intensity of the customer-firm relationship will be greater for customers with a larger pur-
relationship will be greater for a higher level of message chase amount.
postings in the social media site. Focus of Buying. Buying focus refers to the breadth
of interaction between a customer and the focal firm,
2.3. Interaction Effects with
and extant research considers buying focus to be an
Customer Characteristics
important indicator of relationship status (Verhoef
Purchase Amount. Although participation in a firm
and Donkers 2005). High buying focus typically
hosted social media site will have a positive effect
entails limited customer-firm interaction, wherein a
on customer-firm relationship for all customers (as
focal customer tends to buy exclusively from a single
argued for H1), we expect that customers who have a
product category and thereby does not broaden the
stronger transaction relationship with the firm (before
scope of interactions with the firm. Research suggests
participation in social media), as measured by their
that lower customer satisfaction with the offerings of
spending levels, will likely experience greater ben-
the firm can manifest itself as highly focused purchas-
efits from association with the social media site.
ing behavior (Verhoef et al. 2001). Lower customer
This is likely to happen through two mechanisms.
satisfaction has been linked to less favorable out-
First, extant literature in CRM suggests that exchange
comes in several studies (e.g., Anderson and Sullivan
characteristics such as average transaction amount
1993, Bolton 1998). On the other hand, less focused
are associated with greater satisfaction (Bolton 1998)
buying (purchasing across a wide range of products
and commitment that results in better behavioral
and categories) increases customers’ familiarity with
the firm and can lead to greater satisfaction with
1
We note that these arguments are valid only if most customer com- the firm. For these less focused buyers, resources on
ments are positive. We also wish to point out that in our context, all
of the messages that originate from the firm are positive and more the social media site can prove to be particularly valu-
than 90% of customer comments are also positive in nature. This able. Because the social media site provides exposure
is likely because this is a firm-hosted social media site where most to information on new and different products, less
of the messages originate from the firm and customer comments focused customers can reap greater benefits from par-
are in response to these messages and generally take the form of ticipation in social media because they tend to pur-
positive comments or general queries. This is unlike a third party
website where customers may post product reviews and recom- chase across multiple product categories. Thus, we
mendations that can be either positive or negative. We thank an expect that the impact of social media participation
anonymous reviewer for seeking clarification on this issue. on the intensity of customer-firm relationship will be
lower for customers having a greater buying focus. ticket purchases are likely to have greater switching
Therefore, costs and thus may be more inclined to continue the
Hypothesis 4 (H4). The impact of participation in firm relationship with their current firm and attempt to
hosted social media on the intensity of the customer-firm gain preferential status with it. Moreover, interactions
relationship will be lower for customers who have a greater with the firm through the online social media can
focus in buying. help such customers achieve these objectives through
obtaining a greater degree of personalization and
Deal Sensitivity. An important customer characteris- information for premium products in real time. More-
tic in the context of customer-firm relationship is the over, they may also utilize social media to gather cues
extent to which a focal customer looks to buy items on about premium products from other customers of the
promotion. Extant research suggests that customers social media site to mitigate uncertainty (Pavlou et al.
who are more deal prone tend to search more for 2007). Thus, participation in social media is likely to
products from competing sellers and tend to make strengthen the intensity of customer-firm relationship
their purchases with the seller who offers the lowest for customers who purchase a greater share of pre-
price (Ailawadi et al. 2001). Thus, deal prone con-
mium products:
sumers are likely to be less committed to the firm and
have reduced interest in developing a stronger rela- Hypothesis 6 (H6). The impact of participation in
tionship with the firm. These customers may visit the firm hosted social media on the intensity of the customer-
firm’s social media site only to search for coupons or firm relationship will be greater for customers who have a
promotions. In such a case, we would expect firms’ greater share of premium product purchases.
relationship building efforts in the form of engaging
in interactive communications with their customers
through social media to elicit a greater response from 3. Research Setting
less deal prone consumers who may appreciate firms’ The data set for this study comes from a large
relationship investments more.2 Thus, we argue that specialty firm that owns multiple stores and sells
online social media participation would strengthen wine and like products in the northeastern United
the customer-firm relationship for customers who are States.
less deal prone:
3.1. The Firm’s Customer Engagement
Hypothesis 5 (H5). The impact of participation in firm Initiative via Social Media
hosted social media on the intensity of the customer-firm In August 2009, the focal firm, for the first time, cre-
relationship will be greater for customers with lower deal ated a webpage on a major social networking website
sensitivity. (e.g., Facebook) with the objective of connecting with
Share of Premium Products. Customers who purchase its customers through social media.4 Since then, it has
the firm’s most valuable offerings are also consid- made conscious efforts to continue to invest in social
ered to be its most lucrative customers, and firms media by regularly updating the site through mes-
often invest considerable amount of resources in nur- sage postings and by encouraging customers to visit
turing relationships with them (Reinartz and Kumar the site and become involved by posting, comment-
2000). In addition, premium products (because of ing, and other such activities. The messages posted
their higher price tags) are often high involvement by the firm on the above site differ both in their con-
purchases, which in turn are associated with greater tent and characteristics. In particular, the site is used
levels of uncertainty and perceived risk (Pavlou et al. by the firm to build store equity, foster better cus-
2007).3 Customers with a larger proportion of high tomer relationships, communicate to customers about
various events/activities, and disseminate informa-
2
We note that in case the firm’s social media site is mostly devoted tion about its products and promotions. Examples
to offering coupons or promotions, we would expect the opposite of equity and relationship building messages include
effect where deal prone customers would shop with the firm more. “we are currently touring wineries in Argentina to
However, our focus is on understanding the impact of firms’ rela-
get you the best wines” and “we are constantly striv-
tionship building efforts through social media, where firms engage
in a variety of communications with their consumers (both product ing to improve your shopping experiences,” respec-
and non-product related), which is the strategy used by the firm tively. Messages that inform customers about various
under study. We thank an anonymous reviewer for bringing our events usually relate to community based activities
attention to this issue. (e.g., 5K runs for charity). Other message postings
3
We note that uncertainty can be due to either seller or prod- can be solely product (e.g., “how to select a good
uct uncertainty. Our focus here is on product related uncertainty
because consumers perceive a high risk associated with high-
4
involvement purchases because of their higher uncertainty (Pavlou The confidentiality arrangements with the firm preclude us from
et al. 2007). We thank an anonymous reviewer for helping us clarify disclosing the name of the firm. We also cannot furnish details that
the issue. allow identification of the social networking website.
Chardonnay”) or promotion (e.g., “buy one bottle at and race along with their residence address was
regular price and get another free”) related.5 used. Only those customers whose demographics
When the firm launched the social media site, it and residence location/address matched closely were
also undertook a marketing campaign to inform its retained. The process described above was done
clientele about these developments. As a result, exist- entirely by the retailer, and we were furnished with
ing as well as potential customers began participat- the data set containing unique codes assigned to
ing by signing up and/or following the firm on its customers upon masking/de-identifying all sensitive
social media site. However, they differ with respect information (e.g., names and residence location). The
to their timing of participation, with some customers customers’ participation in social media can be linked
participating soon after the site was launched whereas to their transaction data using these unique customer
others did so at a later date. No incentives (either codes. This process was adopted to ensure complete
monetary or promotional) were offered, and cus- customer confidentiality.
tomers signed up of their own volition. We use data from January 2008 to March 2011 to
conduct our empirical analysis. As will be evident
3.2. Social Media Participation and in §4, we also need reasonable numbers of customer
Customer Transaction Data purchases and sufficient data period post customer
A key and notable contribution of our study is that participation to reliably estimate our model param-
we combine data related to customers’ participation eters. Consequently, we do not consider customers
in the focal firm’s social media site with their actual who joined the social media site during the last
purchases. This data gathering process is not straight- six months of the data time period.7 The criteria
forward and presents a rather onerous task given the described above left us with 394 customers whom we
nature and idiosyncrasies associated with this kind of include for further analysis.
work. We adopted the following multi-step approach
to identify customers who participate (e.g., become 3.3. Variable Operationalization
a “fan” of the firm on Facebook) in the firms’ social The dependent variable for our analysis is customers’
media site. We first begin with a survey that was intensity of relationship with the firm. For our pur-
sent to the firm’s customer base in early 2011. This poses, following precedence (e.g., Venkatesan and
survey that was sent to 5,000 customers featured a Kumar 2004) we use visit frequency as a measure for
section devoted to social media wherein customers this variable. Accordingly, we analyze if customers’
were specifically asked if they participate in the firm’s social media participation affects their frequency of
social media site in question (e.g., the firm’s Facebook visits and operationalize it as the number of times
page).6 a customer visits the store in a given time period
From the survey, we identified 845 customers who (denoted by freq). With respect to operationalization
actively participate in the firm’s social media site (e.g., of the independent variables, customers’ participation
Facebook). Working in conjunction with the firm, we (CParT) is a binary variable that takes the value 1
then matched the above customers with actual par- if the customer is a participant in the firm’s social
ticipants of the specific social media site in question. media site (e.g., becomes a “fan” on the firm’s Face-
This was possible because the firm was tracking its book page) at the given time period and 0 otherwise.8
participants on this site (e.g., Facebook) since the We note that different customers participate in the
inception of its program. For this purpose, complete firm’s social media site at different points of time dur-
names were used and whenever there were multi- ing the estimation period. Social media activity rele-
ple matches they were discarded and only unique vant to a focal customer is measured by the number
exact names were retained. Additionally, these names of messages posted by the firm and all other cus-
were also matched with the transaction/purchase tomers. Because recent postings may be more impact-
data of the retailer and duplicate matches were ful than older ones, we formulate a stock variable of
once again eliminated. To be further sure about this postings (PostingsStock) in the following manner (e.g.,
process, demographic and residence/location infor- Tellis and Weiss 1995):
mation obtained from the surveys for the respec-
tive customers was utilized and this was cross PostingsStock t = Postingst +41−5PostingsStock t−1 0 (1)
tabulated with the corresponding information avail-
able from the firm’s transaction data. Specifically, 7
Because the customer participation period spans the time period
information pertaining to customers’ gender, age, August 2009–March 2011, we consider only those customers who
joined during the time period August 2009–September 2010.
8
In our case, customer participation takes the form of becoming
5
We note that no one particular type of posting dominates. a fan of the firm, liking the firm, etc. Our sample customers do
6
The survey had 1,249 valid questionnaires returned, representing not “unparticipate” once they decide to participate/take part in the
a response rate of 24.98%. firm’s social media site.
PostingsStock is the number of messages posted they spend on wine products in a typical year. Loyalty
excluding that of the focal customer at time t and is then defined as the ratio of the expenditures at
∈ 40115 is the decay coefficient.9 Purchase amount the focal retailer to the (self-reported) total spend-
(PurAmt) is the spending incurred by the customer ing incurred on such categories in each time period.11
across all products in a given time period. Buying In case of demographics, we include the actual age
focus (BuyFocus) represents how broad or narrow the in years (Age), the logarithm of customers’ income in
buying pattern is for a particular customer. In our dollars (Income), and dummy variables for Gender and
case, because we consider wine purchases, follow- Race which take the value 1 if the customer is male
ing current literature in this area (Huerta et al. 1998), and white, respectively, and 0 otherwise.
we use wine color type to capture this variable. The
color is a predominant characteristic for wines and
serves as an important element for their classifica- 4. Self-Selection Issue and
tion because it depends on the grape variety (Jensen Endogeneity
et al. 2008). Wine is classified broadly into three color It is possible that customer specific unobserved fac-
types: white, red, and rosé. For each customer, dur- tors may jointly determine customers’ decision to par-
ing the calibration period,10 we calculate her spend- ticipate in the firm initiated social media site and their
ing for each of these wine color types and designate intensity of relationship with the firm. For example,
that color as the principal one on which her spending customers’ positive predisposition toward the firm
is the most. The buying focus (calculated during the and its products (beyond customer loyalty that we
estimation period) then is defined as a dummy vari- control for) may simultaneously influence both their
able that takes the value 1 if the customer buys the willingness to participate in the social media site and
principal color type at a given time and 0 otherwise. their intensity of relationship with the firm. Thus,
To calculate the share of premium products (Premi- there could be endogeneity issues arising because of
umShare), we first determine which products are pre- self-selection that need to be considered for accurate
mium. For this purpose, we calculate the mean price identification of the social media participation effect.
per unit ($/ml) for all the products and plot the distri- To resolve the above issues and establish a causal
bution. On finding that this is distributed normal, we
link between customers’ social media participation
designate those products as being premium that are
and their intensity of relationship with the firm, as
one standard deviation above the mean. The share of
recommended by current literature (e.g., Huang et al.
premium products bought by a focal customer is then
2012, Girma and Gorg 2007, Blundell and Dias 2000),
obtained as the ratio of the number of premium to the
we utilize propensity score matching in combination
total number of products purchased by the customer.
with difference-in-differences analysis. A DID estima-
Deal sensitivity (Deal) designates how sensitive a cus-
tor represents the difference in pre and post partici-
tomer is toward deals and is defined as the proportion
pation differences between two groups of customers,
of products the customer buys on promotion, i.e., the
the treatment group (i.e., customers who participate)
number of products bought on promotion divided by
the total number of products purchased in a given and the control group (i.e., customers who do not
time period. participate). Note that the difference in the pre and
We also include in our model several control vari- post behavior of participating customers alone may
ables such as customer loyalty and demographics. not produce accurate results because of the presence
We follow prior literature and measure loyalty (Loy- of extraneous factors that might affect behavior (e.g.,
alty) as the share of wallet or the percentage of total frequency of visits) in one period but not the other.
customer spending at the focal retailer in a given Thus, a better research design entails using as ref-
time period (e.g., Ailawadi et al. 2008). To opera- erence or benchmark customers who do not partici-
tionalize this variable, we need to know the total cus- pate in the firm initiated social media (i.e., the control
tomer spending on wine products across all stores group). Comparing the behavior of the treatment and
(and not just the focal retailer). In order to compute control groups in the pre and post participation peri-
this, we rely on the survey information wherein cus- ods helps to control for the influence of extraneous
tomers were specifically asked to report how much factors. The challenge then is to compose the control
group that comprises customers who are very similar
9
The decay coefficient accounts for the possibility that recent
messages would matter more than the earlier postings. To avoid 11
In our case, we need to allocate the yearly (self-reported) total
computational burden, we do not estimate ; instead, we use a grid expenditures to specific time periods/shopping trips. We use his-
search procedure and find that 0.83 provides the best fit. torical spending patterns covering two full years (January 2006–
10
We use 12 months of data prior to January 2008 as the calibra- December 2007) prior to our estimation time period to compute
tion period to calculate/initialize some of the model variables in average weights, which are then used to assign expenditures,
this case. because spending may not be even or consistent across months.
to the treatment group customers (participating cus- 4.1. Propensity Score Matching
tomers) in all respects but for the fact that they have The objective of PSM is to select treatment and
not participated in the firm hosted social media, in control group customers who resemble each other
contrast to the treatment customers who have par- in all relevant characteristics before social media
ticipated. The difference in behavior between the participation (the treatment), thereby creating a sta-
two groups (i.e., treatment and control) can then be tistical equivalence between the two groups by bal-
attributed to social media participation. ancing them on observables (Rosenbaum and Rubin
In order to reliably construct the aforementioned 1983, Brynjolfsson et al. 2010). For PSM to be feasible,
control group, we use PSM. This is because match- data should be available for both the treatment and
ing through propensity scores substantially improves control customers before the intervention (i.e., social
the similarity between the treatment and the con- media participation). We utilize 82 weeks of data per-
trol group because customers who are like treatment taining to customer characteristics for both groups
customers are given a greater weight for inclusion prior to the launch of the social media site. We fol-
in the control group, thereby improving the infer- low recommendations of recent literature and use the
ence related to DID analysis (Stewart and Swaffield same data sources for both groups (Heckman et al.
2004, Abadie et al. 2010). In other words, the con- 1997). We model customer participation decision as a
trol group (obtained through PSM) that is used as a function of customer specific transaction and demo-
reference or comparison in the DID estimation will graphic characteristics.
be as similar as possible to the treatment group on In the interest of exposition, we discuss the salient
the observed characteristics, which helps to eliminate points here and provide the detailed description,
temporally invariant estimation bias (Smith and Todd including the steps we follow for conducting PSM—
2005). Moreover, PSM helps us mimic a randomized which are based on current literature (e.g., Caliendo
experimental setup (Rubin 2006). and Kopeinig 2008, Guo and Frasier 2010)—in the
To summarize, we use the PSM procedure to con- (online) appendix (available at http://dx.doi.org/
struct a matched pair set of customers, one from the 10.1287/isre.1120.0460). We begin by checking if the
“treatment” group—comprising customers who par- variables used in the model are balanced between the
ticipate in the firm initiated social media site and treatment and control customers using social media
whose behaviors we wish to analyze—and another participation as the grouping variable through appro-
from the “control” group—consisting of customers priate bivariate tests and include them for further
who are not part of or never participate in the firm’s analysis if they are significantly different across the
social media site. Subsequently, we employ the DID two groups.
estimation on the resulting matched sample to esti- We then conduct the propensity score analysis
using a logistic model formulation because the pro-
mate the impact of social media participation. We note
gram intervention (social media participation) is
that using PSM in combination with DID analysis, as
binary.12 These results along with the fit statistics are
documented by prior literature, considerably reduces
reported in Table 1. We then perform the process of
the bias resulting from both observable and unob-
matching the treatment and control customers using
servable variables and helps improve the consistency
the estimated propensity scores. Although many
of the estimates (Stewart and Swaffield 2004, Cao
matching algorithms are available, we use the optimal
et al. 2011). In addition, it improves the reliability of
pair matching technique, wherein each customer who
the estimation process (Cao et al. 2011) and is also
participates is matched to another customer who does
more efficient as compared to other approaches (e.g.,
not participate. This procedure substantially improves
instrument variables) because it obviates the need for
the power of the analysis and reduces bias that can
strong exclusion restrictions (e.g., Girma and Gorg
occur when multiple (potentially) dissimilar treat-
2007, Blundell and Dias 2000). ment and control customers are matched (Guo and
We note that although instrument variable tech- Frasier 2010, Huang et al. 2012). Additionally, optimal
nique is commonly used to correct for selection bias, pair matching also enables us to treat each distinct
it is not well suited for our purpose because find- matched pair comprising treatment and control cus-
ing good instruments for social media participation is tomers separately, thus allowing us to control for any
difficult and using weak instruments may only exac- residual variability emanating from social media par-
erbate the problem (Bound et al. 1995). Furthermore, ticipation effects (Huang et al. 2012). As a result of the
even if valid instruments were to be found, the esti- above, we are able to satisfactorily match 287 treat-
mation process is still fraught with perils because the ment customers to equivalent control customers, all
social media participation enters as a discrete vari- of whom are included for model estimation.
able in the analysis, which can result in poorly identi-
fied (Wooldridge 2002) and/or inconsistent estimates 12
We also fit a probit model for social media participation and the
(Angrist and Pischke 2009). results were similar.
Table 1 Logit Regression Model of Customers’ Social Media Figure 1 Distribution of Propensity Score Before and After Matching
Participation
Box plot
Variable Parameter Std. error
0.9
PurAmt 000075∗∗∗ 000028
0.8
BuyFocus −300751∗∗∗ 003975
Deal 006113∗∗ 002941 0.7
PremiumShare 000981∗∗ 000409
Loyalty 304004∗∗∗ 102452 0.6
Age −000685∗∗∗ 000112
Gender 002053∗ 001117 0.5
Income 001510 002322
0.4
Race 004478∗ 002625
Constant 303035 207626 0.3
Log likelihood −30105350
0.2
∗
p ≤ 0010, ∗∗ p ≤ 0005, ∗∗∗ p ≤ 0001.
0.1
Next, we check to see if the underlying assumptions TG CG_O CG_M
of the PSM process hold. We check if there is sub-
stantial overlap in the characteristics of the customers Smooth histogram plot
who participate in the social media and those who 35
do not (i.e., common support condition). For this pur-
pose, as recommended by Lechner (2002), we perform
30
a visual analysis of the propensity score distributions
through box-plots and histograms (Figure 1) and find
evidence for the existence of common support. 25
To further ensure the quality of the PSM proce-
dure, we check if the covariates are balanced between 20
the treatment and control groups in the pre and post
matching condition, the results of which are reported 15
in Table 2. As can be seen from the table, the variables
that are used for PSM are not significantly different
10 TG
across the two groups of customers post matching, CG_O
implying statistical balance. CG_M
Finally, we conduct sensitivity analysis to check 5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
for “hidden” bias—the presence of unobserved fac-
Notes. The figure shows the box plots and smooth versions of histograms of
tors that may affect the assignment into treatment propensity score distributions for following groups: TG—Treatment Group,
(social media participation) and outcome (inten- CG_O—Control Group before Matching, CG_M—Control Group after Match-
sity of customer-firm relationship) simultaneously ing. Before matching, the distributions for treatment and control groups are
(Rosenbaum 2002). Essentially, this involves inves- quite different; however, after matching they are almost identical, providing
tigating whether some unobservable factors overly evidence of common support. Furthermore, we also conduct a Kolmogorov-
Smirnov (KS) test to compare the two distributions. The p-value of the KS-
influence the treatment effect of customers’ social
test between TG and CG_O is 0.0000 whereas p-value of the KS-test between
media participation. The analysis based on the meth- TG and CG_M is 0.9377, which provide evidence of similarity of propensity
ods expounded in the literature (e.g., Rosenbaum score distributions between TG and CG_M and hence the existence of com-
2002, Becker and Caliendo 2007) shows that “hidden” mon support between the two distributions.
bias is not a serious concern in our case.
We report the overall summary statistics, which
are obtained by averaging the corresponding values per time period. The large amount is a reflection of
across the treatment and the matched control groups. the nature of category (wine) and also that customers
With respect to number of postings, we find that both buy multiple items on a trip. Premium selling prod-
the firm and its customers actively post messages on ucts constitute 6.69% of the total purchases across the
the social media site. The average of the stock of the customers. Customers seem to be buying substantial
postings variable (see Equation (1)) is 58.23. Although amounts on deal, which could be due to the frequency
we do not explicitly code the valence of the postings, of promotions that are offered. Moreover, they also
as mentioned earlier, we find that all of the messages exhibit considerable buying focus. Likewise, they are
initiated by the retailer are positive and that more also loyal to the firm (as measured by share of wallet)
than 90% of postings made by the customers are posi- plausibly because the focal firm is one of the domi-
tive in nature. Customers on an average spend $61.05 nant players in the region. As for demographics, we
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 009442
BuyFocus 005504 007690 −002186 −1002615 008730 006257 −000753 −102859 009354
Deal 006405 006744 −000339 −107948 101000 006615 −000210 −009731 009828
PremiumShare 000635 000769 −000134 −005611 101302 000701 −000066 −003045 100298
Loyalty 008256 008353 −000097 −402823 008025 008296 −000040 −100153 009915
Age 5004550 5601339 −506789 −701329 100922 5104079 −009529 −106229 100084
Gender 004408 004017 000391 108698 100306 004342 000066 001152 100033
Income 1009990 1009675 000315 107515 100545 1009828 000162 002455 100100
Race 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 group
to variance of control group. The covariates are imbalanced and hence matching would be compulsory if mean difference is away from zero and variance ratio
is away from one. The mean differences and variance ratios are close to 0 and 1, respectively, after matching, indicating the improvement in covariate balancing
due to matching.
find the mean age to be about 51 years, reflecting in the above formulation consisting of matched treat-
the nature of the category (because 21 years is the ment and control group customers, we use monthly
legal minimum for alcoholic purchases), and the mean data that spans both pre and post launch time periods
income is $59,334; 43.75% of the sample is male and of the social media site (January 2008–March 2011),
87.26% of the customers in the sample are white. which results in time-series data that are then stacked
for estimation. The main parameter of interest is 3 ,
5. Econometric Specification which captures the change in the frequency of visits
for treatment customers post participation relative to
5.1. DID Model: Main Effect of Social Media control customers who do not participate in firm ini-
Participation tiated social media.
For each matched pair comprising treatment and con-
trol customers (see §4.1), the logarithm of frequency 5.2. DDD Approach: Moderating Effects of
of visits is modeled as13 Hypothesized Variables
We now describe an alternative version of the model
log4freqijt 5 = 0j +1 TreatDij +2 CParT ijt +3 TreatDij to the one presented earlier (in Equation (2)) that
×CParT ijt +äXij +it 0 (2) enables us to test our hypotheses discussed in §2.
Specifically, we segment customers into high and
In Equation (2), i denotes a matched pair of cus- low levels for each of the moderating variables
tomers, j denotes a treatment group or a control (i.e., social media activity, purchase amount, buying
group customer, and t denotes the time period. focus, deal sensitivity, and share of premium prod-
TreatDij is the treatment dummy variable (that ucts). To accomplish this, consistent with the meth-
equals 1 if the customer is in the treatment group ods expounded in the current literature (e.g., Tucker
and 0 if the customer is in the control group), et al. 2012), we use a median split based on the val-
and CParTijt is a dummy variable denoting social ues of the above variables and divide them into high
media participation, which takes the value 0 and 1 and low levels/types. Thus, for a certain variable
for periods prior to and post participation, respec- (e.g., purchase amount) customers will be classified
tively, for customers belonging to the matched pair i. as belonging to the high level if they score greater
Xij represents a vector of demographic and behav- than the median on that variable and low level oth-
ioral control variables with ä being their correspond- erwise. To make sure that the above process does not
ing estimated coefficients. 0j are the customer-specific confound the estimation time period, we use the cal-
fixed effects that capture the differences in baseline ibration period for this purpose.
relationship intensity and enable us to control for The above approach allows us to test when the
unobserved heterogeneity across customers. Note that treatment effect (i.e., the social media participation
effect) is more effective. Conducting such an analysis
13
In the very few cases where the frequency is zero, we add a may be especially pertinent for managerial decision
small value to it (order of 0.001) so as to enable the logarithmic making because it would help in understanding when
transformation. and for which customer segments social media efforts
yield greater returns.14 Therefore, to investigate the Table 3 Difference in Differences: Frequency of Visits of Treatment
effect of social media participation at different levels vs. Control Groups
of the moderating variables based on prior literature Group monthly mean
(e.g., Gruber 1994), we use the following formulation:
Treatment Control Difference t-test
log4freqijt 5 After social media launch 207618 204855 002763 304652
= 0j +1 TreatDij +2 CParT ijt +3 TreatDij ×CParT ijt Before social media launch 205403 204169 001234 009886
Difference 002215 000686 001529
+4 TreatDij ×CParT ijt ×PostingsStock ij t-test 301764 105892
+5 TreatDij ×CParT ijt ×PurAmtij +6 TreatDij Notes. There is no significant difference in frequency of visits between treat-
ment and control groups before the launch of social media; however, the
×CP arTijt ×BuyFocusij +7 TreatDij ×CParT ijt difference becomes significant after the social media launch. Because these
differences are taken for the matched sample, the significance of the dif-
×Dealij +8 TreatDij ×CP arTijt ×PremiumShareij ference in post launch period can be attributed to customers’ social media
participation.
+ìXij +it 0 (3)
In Equation (3), the variables have the same the treatment and control group customers before the
meaning as in Equation (2).15 As in the previous social media site is launched by the firm, whereas the
case, we include customer specific fixed effects and difference is significant post launch. This lends prima
demographic/behavioral controls. Additionally, the facie credence to the fact that participation in the firm
equation includes three way interactions between initiated social media impacts the behavior of par-
treatment, social media participation, and each of the ticipating consumers. We investigate the above more
moderating variables (they take on the value of 1 formally using the DID analysis next.
for high levels of the variable and 0 otherwise). For To benchmark the results and statistical fit of our
instance, TreatDij ×CParT ijt ×PurAmtij is the three way DID model, we estimate a series of alternative models
interaction between the dummies of treatment, par- (Table 4). Model (1) is the basic DID model without
any control variables. As part of Model (2), we include
ticipation, and the high level of purchase amount.
demographic variables as controls, and Model (3) con-
Because of the presence of third level interactions
tains behavioral variables in addition to demograph-
in Equation (3), the above framework results in a
ics as controls. Model (4) is an augmented version of
difference-in-difference-in-differences (DDD) specifi-
Model (3) that adds customer specific fixed effects.
cation. The key parameters of interest are the coeffi-
As is evident from Table 4, the results are substan-
cients associated with the three way interaction terms
tively similar across the models. The fit statistic (R2 5
(4 −8 5. They capture the effect of social media par-
increases from Model (1) to Model (4), which lends
ticipation on visit frequency for customers with high
face validity to the results.
level of the corresponding variable (e.g., purchase
We find that the parameter corresponding to the
amount) relative to those customers with a low level
treatment effect of social media participation (TreatD×
of the same variable.
CParT) is consistently positive and significant across
the different models (for example, in Model (4) this
6. Results estimate is 0.0509). This suggests that participation in
the firm initiated social media positively influences
6.1. DID Results for Visit Frequency
customers’ intensity of relationship with the firm in
We display the raw difference-in-differences values
terms of their frequency of visits. Thus, we find sup-
for frequency of visits for the treatment and control
port for H1—the central hypothesis of our study. In
groups before and after the social media site was
order to assess the magnitude of the effect, we com-
launched by the firm (Table 3). As can be seen from
pute the elasticity measure using the corresponding
the table, the mean visit frequency increase is not
parameter estimate from Model (4), the best fitting
significant for the control group (2.486 post launch DID model. We find that elasticity of participation on
versus 2.417 pre-launch), but this increase is sig- visit frequency to be 5.204,16 which implies that cus-
nificant for the treatment group (2.762 post launch tomers who participate in the firm’s social media visit
versus 2.540 pre-launch). Moreover, there is no sig- the firm about 5.2% more frequently compared to cus-
nificant difference in the frequency of visits between tomers who do not participate.
14
We thank the AE for the suggestion. 16
Because TreatD×CParT is a dummy variable and the regres-
15
Although we include the various relevant two way interaction sion equation is semi-logarithmic, the elasticity is given by 100×
variables in the model, for the sake of exposition, we do not present 6exp8ˆ 3 −4005Var4ˆ 3 559−17, where ˆ 3 is the estimate of 3 and
and discuss the results related to these two way interaction vari- Var4ˆ 3 5 is its estimated variance (Kennedy 1981, Halvorsen and
ables. More information is available from the authors upon request. Palmquist 1980).
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 customer
No 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 002185
CParT 000279∗ 000161 000853∗ 000464 001260∗ 000714 001542∗ 000906 001541∗ 000865
TreatD×CParT 000619∗∗∗ 000201 000583∗∗∗ 000202 000566∗∗∗ 000218 000509∗∗∗ 000180 000500∗∗∗ 000129
TreatD×CParT×PostingsStock 000002∗∗ 000001
TreatD×CParT×PurAmt 000005∗∗∗ 000002
TreatD×CParT×BuyFocus −000405∗∗∗ 000071
TreatD×CParT×Deal −000436∗∗∗ 000162
TreatD×CParT×PremiumShare 000837∗∗ 000345
PostingsStock 000016∗∗ 000008 000020∗∗ 000008 000019∗∗ 000009
PurAmt 000002∗∗∗ 000001 000003∗∗∗ 000001 000002∗∗∗ 000001
BuyFocus −002177 001490 −000552 000520 −000533 000546
Deal −000986 000676 −000301 000428 −000376 000472
PremiumShare 001144∗∗ 000537 000861∗∗∗ 000286 000663∗∗ 000298
Loyalty 001141∗∗ 000478 000055∗∗∗ 000019 000100∗∗∗ 000024
Age 000018 000020 000019 000018 000018 000115 000024 000115
Gender −000155 000154 −000103 000153 −000170 003743 −000280 003751
Income 000021 000034 000019 000034 000282 002436 000049 003254
Race −000190 000425 −000087 000420 −000404 003228 −000319 002444
Intercept 008795∗∗ 001130 006022∗∗∗ 000583 000192∗∗∗ 000065
Customer fixed effects No No No Yes Yes
R-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 Type (8 = 0008373p < 0001), thereby supporting H6. With
Customers respect to the control variables, we note that to be con-
We now focus on the results of the DDD model, which sistent with the DDD analysis, we report the results
is concerned with the effect of social media participa- when the variables were used in levels (similar results
tion for high and low levels of the moderating vari- were obtained when actual values of the variables
ables. We report these results as Model (5) in Table 4. were used). We find that higher level of activity in the
Because the DDD model also includes the two-way social media site is associated with greater intensity
interaction of the treatment effect, the results can be of relationship between the firm and its customers.
interpreted as the relative impact of the corresponding Moreover, customers with higher levels of spending
level of the moderating variable on visit frequency. H2 and greater shares of premium products visit the firm
suggests that the effect of customer participation on more frequently, as do more loyal customers. We find
intensity of relationship will be greater for higher lev- no significant effects for buying focus, deal sensitiv-
els of activity or message postings in the social media ity, and the demographic variables (i.e., age, gender
site. We find evidence to support this (4 = 000002; and race) plausibly because of better covariate balance
p < 0001). As per hypotheses H3, the impact of cus- that was achieved for these variables.
tomer participation in firms’ social media is amplified
for customers with high purchase amount levels. H4
suggests that the effect of social media participation is 7. Economic Impact of Customer
lower for customers with a greater buying focus. We Social Media Participation
find support for both these hypothesized interaction
effects (5 = 000005; p < 0001 for purchase amount and 7.1. Profitability Analysis
6 = −000405; p < 0001 for buying focus). In this section, we analyze the benefits that can be
We find that the effect of social media partic- realized by firms as a result of their foray into social
ipation is greater for customers with lower deal media. Specifically, we begin by documenting the
sensitivity (7 = −0004363p < 0001). Thus, we find sup- incremental profits of customers who participate in
port for H5. Finally, we find that the effect of the social media (i.e., treatment group customers)
social media participation is higher for customers vis-à-vis those who do not participate (i.e., control
whose share of premium product purchases is greater group customers). For this purpose, we compute the
Table 5 DID: Profitability of Treatment vs. Control Group pre launch versus $22.255 post launch). In contrast,
there is no significant difference in total profits before
Group
monthly mean ($) ($22.035) and after ($23.698) the launch of the social
media site for the control group. This suggests that
Treatment Control Difference t-test customer participation in social media improves firm
After social media launch 2502607 2306989 105618 303972 profitability.
Before social media launch 2202549 2200354 002195 102184 To investigate the above issue more formally, we
Difference 300058 106635 103423 utilize the DID framework of Equation (2) with cus-
t-test 209672 105924 tomer profits as the dependent variable and report the
Notes. There is no significant difference in total customer profits (in dollars results in Table 6. Like earlier, we estimate a series of
per month) between treatment and control group before the launch of social models ranging from the basic DID model sans any
media site; however the difference becomes significant post launch. Because control variables (Model 1) to the one that includes
these differences are taken for the matched sample, the significance of the both behavioral and demographic controls along with
difference in post period can be attributed to the customers’ social media
customer fixed effects (Model 4). We once again find
participation.
that the parameter related to the treatment effect of
social media participation (TreatD×CParT) is positive
aggregated per period net total profits separately for and significant across all the different model specifi-
customers in both groups before and after the launch cations. For instance, in Model (4) the parameter esti-
of the social media site by the firm, the results of mate is 0.0550, which corresponds to profit elasticity
which are reported in Table 5. Note that this is pos- of 0.0563. This implies that customers who partici-
sible in our case because we have data pertaining to pate in the firm’s social media contribute more to the
both retailer prices and costs at the individual prod- firm’s bottom line. We also estimate the DDD model
uct level. The values reported in the table are the that includes the three way interaction effects between
mean total profits across all the products purchased the treatment variable, customer social media par-
by the respective group of customers. On comparing ticipation, and the indicator variable for the high
the table values, we once again find that there is a level of the moderating variables (Equation (3)). We
substantial increase in total profits for the treatment find that the results have the same directionality and
group after the launch of the social media site ($25.261 similar significance as reported for visit frequency
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 customer
No 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 002811
CParT 000726∗ 000394 000878∗ 000462 000868∗ 000235 000225∗ 000125 000334∗ 000187
TreatD×CParT 000578∗∗∗ 000318 000569∗∗∗ 000183 000561∗∗∗ 000187 000550∗∗∗ 000213 000505∗∗∗ 000186
TreatD×CParT×PostingsStock 000007∗∗ 000003
TreatD×CParT×PurAmt 000003∗∗∗ 000001
TreatD×CParT×BuyFocus −000170∗∗∗ 000060
TreatD×CParT×Deal −000110∗∗ 000052
TreatD×CParT×PremiumShare 000315∗∗∗ 000109
PostingsStock 000003∗∗ 000001 000004∗∗ 000002 000009∗ 000005
PurAmt 000027∗∗∗ 000003 000028∗∗∗ 000003 000027∗∗∗ 000003
BuyFocus −000975 000632 −001625 001221 −001278 000793
Deal −001005 000931 −001491 001012 −001653 001096
PremiumShare 000538∗∗∗ 000107 000389∗∗ 000179 000880∗∗∗ 000129
Loyalty 202030∗∗∗ 000548 201367∗∗∗ 000640 201304∗∗∗ 000641
Age 000003 000017 000001 000009 000221 000137 000213 000137
Gender 001026 000947 000087 000175 009551 006475 009593 006481
Income 000077 000356 000247 000179 000585 003859 00088 003887
Race 001769 001258 000187 000482 004572 002912 005013 004920
Intercept 300928∗∗∗ 009891 208112∗∗∗ 008107 109887∗∗∗ 002094
Customer fixed effects No No No Yes Yes
R-squared 0.2168 0.2864 0.4526 0.5407 0.6246
∗
p ≤ 0010, ∗∗ p ≤ 0005, ∗∗∗ p ≤ 0001.
(i.e., Model 5 in Table 4). Thus, the above analysis participation influences the behavior of not only high
establishes that social media participation has a pos- level/segment customers but also those belonging to
itive impact on profitability (a key metric for firms) the low level/segment with respect to their inten-
in a manner akin to intensity of relationship as mea- sity of relationship with the firm. We obtain sim-
sured by visit frequency. ilar results for the customer profitability model as
well. We also represent these results pictorially in Fig-
7.2. Segment Level Analysis ure 2, wherein the four panels succinctly establish that
The results from the DDD analysis establish that the although social media participation matters for both
relative impact of social media participation differs high and low type customers, the differences between
across different customer segments (i.e., high versus the two are more pronounced for certain moderating
low type). However, from a managerial perspective, variables. Such an exercise can be useful in quantify-
it might be more interesting to understand the actual ing the benefits that can accrue to firms as a result
magnitude of the treatment effect for the two seg- of their social media initiatives at the individual cus-
ments. In other words, managers may be interested in tomer segment level and can also help discern the
understanding not only whether the impact of social relative impact for each variable.
media is greater for high type versus low type cus-
tomers but also the absolute magnitude of the impact
for each of these segments. This kind of segment
8. Robustness Checks
level analysis can help in understanding which seg- 8.1. Inclusion of Additional
ments are more lucrative and should be allocated Variables for Matching
greater marketing dollars. To investigate this issue, In addition to the original nine variables that are
we divide the sample data set into two subsamples, used for the main analysis, we check the validity
with one consisting of customers with high levels and of our results by using an augmented set compris-
the other with low levels of the moderating variable, ing three additional variables: Internet risk aversion,
respectively (we do this for each of the moderating technology comfortness, and social networking atti-
variables). We utilize the same process for dividing tude. We choose to include these variables, which
customers into high and low levels/types as was used are obtained from the survey for both participants
for the DDD analysis (i.e., median split analysis as and nonparticipants, because they may be related to
described earlier in the paper). We then conduct the customers’ propensity to participate in firm initiated
DID analysis for each customer segment to distin- social media activities. For example, customers who
guish the impact of social media participation at the score high on social networking attitude may be more
segment level for both visit frequency as well as prof- likely to join the firm’s social media. The pertinent
its (Model 4 formulation as presented in Equation (2)). details including a description of the relevant vari-
We provide the results of the above analysis in ables and sources used to construct them are pro-
Table 7. In the interest of exposition, we present the vided in the online appendix. Note that we do not
main results concerning the treatment effect only. We include the above mentioned variables in the main
find that the effect of participation for the high seg- analysis because it is recommended that they be col-
ment is positive and significant in case of frequency lected prior to customers’ social media participation
of visits for all the moderating variables. Interest- (Caliendo and Kopeinig 2008), whereas the survey
ingly, we also find that the impact of social media (that is the source of these variables) was adminis-
participation is still positive and significant for low tered at a later date. We find that the results pertain-
segment customers. In other words, social media ing to the effect of social media participation using the
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 visits
TreatD 000718 000323 000612 000403 000104 000896 000078 000921 000912 000098
CParT 001884 001186 002216∗ 000834∗ 001219∗ 001867∗ 001383∗ 001843∗ 002154∗ 000932∗
TreatD×CParT 000642∗∗∗ 000463∗∗ 000523∗∗∗ 000484∗∗∗ 000135∗∗∗ 000881∗∗∗ 000099∗∗∗ 000911∗∗∗ 000894∗∗∗ 000116∗∗∗
Profitability
TreatD 000307 000162 000462 000008 000026 000446 000099 000325 000492 000014
CParT 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.
Figure 2 DID Estimate Plots for Segment Analysis (High vs. Low Groups)
Purchase amount Buy focus
0.09 Frequency of visit

0.09
Profitability
Difference in difference estimate

0.08 Low
0.08
0.07
0.07
0.06 0.06
Low
0.05 0.05
High
0.04 0.04
0.03 0.03
0.02 High
0.02
0.01
0.01
Customer segments Customer segments
Deal sensitivity Premium share
0.09 0.09

0.08 Low 0.08 High
0.07 0.07
0.06 0.06
0.05 0.05
0.04 0.04
0.03 High 0.03 Low

0.02 0.02
0.01 0.01
Customer segments Customer segments
augmented set are substantively similar to the results results are not because of some unobservable factors
that we presented earlier. We summarize these results present during certain periods and are robust to the
in the appendix. specification of time windows, we conduct the follow-
ing analysis. We run the DID models (for both visit fre-
8.2. Matching with Alternative Techniques quency and profitability) wherein we restrict the post
As mentioned previously in §4.2, we use the opti- treatment period (i.e., data post social media launch)
mal pair matching for our analysis. However, we to a time frame consisting of 3, 6, 9, 12, and 15 months,
also verify the robustness of our results by employ- the results of which are reported in Table 8. We find
ing alternative matching methods. Specifically, we that the DID estimates are positive and significant
use greedy matching (nearest neighborhood with- in each of the above cases, indicating that customer
out calipers) and Mahalanobis distance (Dehejia and participation in social media positively impacts visit
Wahba 2002) and find that the results with these alter- frequency and profitability. Moreover, the parameter
native matching techniques are very similar to those estimate of the treatment effect gradually converges
obtained from the optimal pair matching method. We with the estimate at 15 months being close to that
provide some pertinent results of the above alterna- obtained overall (i.e., using 18 months). For illustra-
tive matching techniques in the online appendix. tion purposes, we also plot the DID estimates obtained
from the different time windows in the above analy-
8.3. Robustness of the DID Analysis sis along with the 95% confidence interval in Figure 3.
We conduct a series of robustness checks to confirm As can be seen from the graph, as expected, these esti-
the results of the DID analysis. For our case, we use mates seem to be stabilizing with time.
data consisting of 18 months pre and post launch of To further check that the results of our analysis
the social media site by the firm. To make sure that the are not simply a consequence of spurious correlation
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 40002215
CParT 000572∗ 000281∗ 000702∗ 000440∗ 000768∗ 000523∗ 000832∗ 000594∗ 000877 000655
40003225 40001705 40004015 40002295 40004215 40002995 40004215 40003195 40006215 40005195
TreatD×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 series To make sure that the above does not pose a large
of “phantom” regressions (Chetty et al. 2009). For concern for the model results, we adopt the fol-
these regressions, we consider only the data period lowing method based on Bertrand et al. (2004). We
(18 months) prior to the actual launch of the social aggregate the data set, by ignoring the time series
media site by the firm (i.e., from January 2008 to aspect of it, so that we end up with two distinct
August 2009) and designate a random month from data points—one each for pre and post participation
this period as the time of the launch of the site when periods—for each customer in the matched pair. On
customers could participate. For example, if we desig- reestimating the DID models using this aggregated
nate the end of the ninth month as the time when the data set, we find that our results are substantially sim-
social media site went into effect, then months 10–18 ilar (Table 10). We additionally replicate our regres-
constitute the “treatment” period and months 1–9 sions using Newey-West corrected standard errors
make up the “nontreatment” period. We run a series and also perform Dickey-Fuller tests to make sure
of DID regressions with different time periods (third, that our dependent and independent variables are sta-
sixth, and ninth month) randomly designated as the tionary. Finally, we also reestimate our DID models
month when the social media site was launched, the with random effects in lieu of customer specific fixed
results of which are reported in Table 9. We find that effects and find that the results are very similar.
parameter estimate corresponding to the treatment
effects is statistically indistinguishable from zero in 8.4. Cohort Effect
all the cases. The above analysis with “placebo” esti- It is plausible that early participants affect the behav-
mates suggests that our results can be attributed to ior of later participants. In our data set, because we
customers’ social media participation. observe when the customers become participants of
Because our data consist observations over time, a the firm initiated social media site, we classify them
potential concern is the presence of serial correlation. into their respective cohorts (early or late participants)
Figure 3 95% Confidence Interval for DID Estimates with Different Treatment Windows
Frequency of visits Profitability

0.08 0.08
Upper 95%CI
Estimate
0.07 0.07
Lower 95%CI
0.06 0.06
0.05 0.05
0.04 0.04
0.03 0.03
0.02 0.02
0.01 0.01
2 4 6 8 10 12 14 16 18 2 4 6 8 10 12 14 16 18
Time window Time window
Table 9 DID Analysis for “Phantom” Regressions media activity and customer characteristics. In addi-
“Phantom” treatment window
tion, we quantify the effect of customer social media
participation on a key metric that contributes to firms’
3 months 6 months 9 months
bottom line, i.e., customer profitability. We draw on
Variable Visits Profitability Visits Profitability Visits Profitability CRM and social media literatures to formulate our
TreatD −000049 000050 −000023 000053 −000014 000023 hypotheses and test them using a novel individual
40000395 40000365 40000325 40000505 40000305 40000275 level transaction data set that is complemented with
CParT 000489 000189 000486 000195 000481∗∗ 000184∗∗∗ data on social media participation for the same set of
40003235 40001305 40003245 40001385 40002265 40000245 customers. We rely on a quasi-experimental setup that
TreatD× 000029 000078 000029 000068 000033 000020 uses PSM in combination with DID analysis to rule
CParT 40000555 40000515 40000465 40000425 40000425 40000395 out customer self-selection issues concerning social
∗
p ≤ 0010, ∗∗ p ≤ 0005, ∗∗∗ p ≤ 0001. media participation and the intensity of customer-
firm relationship as measured by frequency of visits.
Table 10 Aggregate DID Results Our results show that customer engagement through
social media increases customers’ shopping visits.
Visit frequency Profitability
Our results also suggest that this effect is greater for a
Variable Estimate SD Estimate SD greater level of activity in the firm hosted social media
TreatD 000002 000020 000001 000019
site. Furthermore, we find that the effect is greater
CParT 000498∗ 000280 000178∗ 000099 for customers who have higher levels of spending
TreatD×CParT 000049∗∗∗ 000018 000069∗∗∗ 000026 and share of premium products and lower levels of
PostingsStock 000012∗∗ 000006 000002∗∗ 000001 buying focus and deal sensitivity. We find that these
PurAmt 000002∗∗∗ 3.33E−05 000001∗∗∗ 2.13E−05
results hold for customer profitability as well.
BuyFocus −000044 000054 −000128∗∗ 000050
Deal −000092∗ 000050 −000016 000047
PremiumShare 000753∗∗∗ 000264 001049∗∗ 000518 9.1. Theoretical Implications
Loyalty 000009∗∗∗ 000003 000304∗∗∗ 000116 Our results extend findings from social media and
Age 000001 000001 000009 000008 CRM literatures while also contributing to them in
Gender −000030 000035 000042 000034 several ways. First, although many studies have
Income 000002 000002 000004 000002
attempted to analyze the impact of online social
Race −000021 000022 000024 000020
Customer fixed effects Yes Yes media, they have primarily considered the influence
R-squared 0.3557 0.4694 of online reviews on product sales (Chevalier and
∗
Mayzlin 2006) or new product diffusion (Susarla et al.
p ≤ 0010, ∗∗ p ≤ 0005, ∗∗∗ p ≤ 0001.
2012). Our study adds to this burgeoning literature
by analyzing the role of social media in driving busi-
using a time based median split. We then reestimate ness value for firms by linking customers’ social
our models separately for each cohort and find the media participation to the additional value created for
substantive results to be similar. the firm.
Second, we view a firm’s social media efforts as
8.5. Alternative Operationalization of Variables
an extension of its relationship building efforts with
We check the sensitivity of the model results to dif-
ferent operationalization of the variables. Specifically, its customers. A significant stream of literature in
we use the total quantity purchased by customers for CRM attempts to understand the drivers of customer
purchase amount. Similarly, we considered two stan- equity and bring accountability to relationship man-
dard deviations above the mean as the cut-off for cus- agement expenditures (Rust et al. 2004). This research
tomers’ purchase of premium share products. We also has been primarily limited to studying the impact of
considered transformation of the pertinent variables firms’ offline relationship building expenditures and
such as logarithm of purchase amount, age, and post- customer characteristics. We add to this research and
ings. We find that the results are very similar in all provide insights into how firms’ online social media
the above cases. activities may influence customer visit (or purchase)
frequency, a key driver of customer lifetime value
(Venkatesan and Kumar 2004).
9. Discussion, Conclusion, and Third, our study takes a holistic look at social
Directions for Future Research media participation and adds new insights to grow-
In this paper, we study the impact of customer partic- ing research that considers the impact of community
ipation in a firm’s social media site on an important participation on customer behavior. For example,
indicator of the intensity of customer-firm relation- Dholakia and Durham (2010) conduct surveys and
ship, customer shopping visit, or purchase frequency. find that customers who became fans of a firm on
We further examine the moderating roles of social Facebook reported increased spending with the firm
as compared to non-fans. Although these findings are Be Wary; Not All Customers are Created Equal. Our
interesting, the study relies mainly on self-reported results also provide actionable recommendations for
data and does not establish a causal connection managers in terms of segmenting and targeting cus-
between Facebook participation and customer behav- tomers based on customer purchase histories and
ior. Another study, that of Algesheimer et al. (2010), interactions with the firm. Our results suggest that
reports that participation in the eBay community managers need to understand that not all customers
leads to a decline in the amount of money spent with will have a similar response to the firm’s social media
the firm. Although this research establishes the effect efforts. It is vital that managers integrate their knowl-
of participation in an online community, these find- edge about customers from both offline transactions
ings do not help understand how firms’ social media and online social media sources in order to better
efforts can transform a customer-firm relationship, serve them.
which is the focus of this study. Furthermore, our Create Product Based Social Media Forums. Besides
study also helps tie customers’ social media participa- finding a differential response of social media partic-
tion to their transaction relationship with the firm and ipation across different types of customers, we also
thus provides managerial guidelines for better deci- find an interaction effect between social media par-
sion making. Finally, the few studies that examine the ticipation and customers’ shopping baskets. Thus,
impact of social media tend to be aggregate in nature managers can increase the response to social media
(e.g., Mudambi and Schuff 2010). Although such stud- participation by creating specialized subcommunities
ies help provide an overview, analyzing the above or discussion forums for customers looking for pre-
phenomenon at the individual level and adopting a mium and unique products.
customer centric view enriches the extant social media In summary, we believe the results of the study
literature while providing novel insights into behav- provide unique theoretical and managerial insights
ior subsequent to customer social media participation. into the relationship between a firm’s social media
efforts and firm performance. However, the study suf-
9.2. Managerial Implications fers from some limitations that can be addressed by
The results of our study yield several important man- future research. First, we are able to provide results
agerial implications. As customers spend more time from only one firm. It would be interesting to study
on social media sites, managers are devoting more whether the results are generalizable for other com-
resources to them. However, in turbulent times when panies. Furthermore, checking for generalizability of
firms are under increased pressure to justify all their results across different forms of social media would
promotional budgets, social media spending needs to also be an interesting addition to this research. Sec-
be more financially accountable. Thus, our research ond, although we studied the impact of social media
helps managers gain a better understanding of return postings, due to data limitations we were not able
on investment accruing to social media initiatives. to examine how the valence of messages affects cus-
Based on our findings, we offer the following recom- tomers’ purchase behavior. In particular, we note that
mendations for practice. most of the message postings in our data are pos-
Nurture Customer Relationships Through Social Media. itive, which limits us from studying the impact of
Our results show that firms’ online social media activ- valence of messages. Next, future research could also
ities help strengthen the bond between the customer extend the current findings by studying whether dif-
and the firm and contribute to financial performance ferent types of message postings by firm (e.g., product
in the long run by increasing customer visit frequen- or price related) will elicit a different response from
cies and profitability. Although our findings suggest a customers. Finally, the results of this study also suf-
positive effect of customer social media participation fer from the limitation that they are specific to the
on customer visit frequency and profitability, man- social media strategy employed by the firm. Future
agers must be cautious and not interpret this result to research can address this by examining the impact
imply that simply creating, say, a Facebook page and of different social media strategies across multiple,
inviting customers to become fans will lead to posi- disparate firms.
tive customer outcomes. They must carefully manage
and devise opportunities to create and nurture rela- Electronic Companion
tionships with customers through the social medium. An electronic companion to this paper is available as
For example, maintaining a user-friendly social media part of the online version at http://dx.doi.org/10.1287/
site interface, providing regular updates about events, isre.1120.0460.
sending personalized messages to individual cus-
tomers, and encouraging member contributions can Acknowledgments
create interactive communication that can enhance The authors thank Professor Arun Jain and the Research
firm equity (Agarwal et al. 2008). Group in Integrated Marketing (RIM) at the School of
Management, SUNY-Buffalo, for accesses to the data and Chevalier JA, Mayzlin D (2006) The effect of word of mouth
research assistance. They thank the participants of the ISR on sales: Online book reviews. J. Marketing Res. 43(August):
special issue workshop at University of Maryland and the 345–354.
INFORMS 2012 Marketing Science Conference. They also Crosby LA, Evans KR, Cowles D (1990) Relationship quality in ser-
thank the Senior Editor, the Associate Editor, and the two vices selling: An interpersonal influence perspective. J. Market-
ing 54(3):68–81.
anonymous reviewers for valuable comments and sugges-
Dehejia RH, Wahba S (1999) Causal effects in non-experimental
tions. The authors contributed equally to the article and the
studies: Reevaluating the evaluation of training programs.
author ordering is reverse alphabetical. All errors are the J. Amer. Statist. Assoc. 94(448):1053–1062.
authors’ own. Dehejia RH, Wahba S (2002) Propensity score-matching meth-
ods for non-experimental causal studies. Rev. Econom. Statist.
84(1):151–161.
References Dellarocas C, Zhang X, Awad NF (2007) Exploring the value of
Abadie A, Diamond A, Hainmueller J (2010) Synthetic control online product reviews in forecasting sales: The case of motion
methods for comparative case studies: Estimating the effect of pictures. J. Interactive Marketing. 21(4):23–45.
California’s Tobacco Control Program. J. Amer. Statist. Assoc. Dholakia UM, Durham E (2010) One café chain’s Facebook experi-
105(490):493–505. ment. Harvard Bus. Rev. 88(3):26.
Agarwal R, Gupta AK, Kraut R (2008) The interplay between digital DiPrete T, Gangl M (2004) Assessing bias in the estimation of causal
and social networks. Inform. Systems Res. 19(3):243–252. effects: Rosenbaum bounds on matching estimators and instru-
Agarwal R, Animesh A, Prasad K (2009) Social interactions and the mental variables estimation with imperfect instruments. Sociol.
“digital divide”: Explaining variations in Internet use. Inform. Methodology 34(1):271–310.
Systems Res. 20(2):277–294. Forman C, Ghose A, Wiesenfeld B (2008) Examining the relation-
Ailawadi KL, Neslin SA, Gedenk K (2001) Pursuing the value- ship between reviews and sales: The role of reviewer iden-
conscious consumer: Store brands versus national brand pro- tity discloser in electronic markets. Inform. Systems Res. 19(3):
motions. J. Marketing 65(1):71–89. 291–313.
Ailawadi KL, Pauwels K, Steenkamp J-BEM (2008) Private-label use Forrester Research (2008) The Growth of Social Technology Adoption
and store loyalty. J. Marketing 72(6):9–30. (Forrester Research, Cambridge, MA).
Algesheimer R, Dholakia UM, Hermann A (2005) The social influ- Girma S, Gorg H (2007) Evaluating the foreign ownership wage
ence of brand community: Evidence from European car clubs. premium using a difference-in-differences matching approach.
J. Marketing 69(3):19–34. J. Internat. Econom. 72(1):97–112.
Algesheimer R, Borle S, Dholakia U, Singh SS (2010) Impact of Godes D, Mayzlin D (2004) Using online conversations to study
customer community participation on customer behaviors: word-of-mouth communication. Marketing Sci. 23(4):545–560.
An empirical investigation. Marketing Sci. 29(4):756–769. Gruber J (1994) The consumption smoothing benefits of unemploy-
Anderson EW, Sullivan MW (1993) The antecedents and con- ment insurance. Amer. Econom. Rev. 87(1):192–205
sequences of customer satisfaction for firms. Marketing Sci. Guo S, Frasier MW (2010) Propensity Score Analysis: Statistical Meth-
12(2):125–143. ods and Applications (Sage, Menlo Park, CA).
Angrist JD, Pischke JS (2009) Mostly Harmless Econometrics: An Halvorsen R, Palmquist R (1980) The interpretation of dummy
Empiricist’s Companion (Princeton University Press, Prince- variables in semilogarithmic equations. Amer. Econom. Rev.
ton, NJ).
70(3):474–475.
Becker SO, Caliendo M (2007) Sensitivity analysis for average treat-
Heckman J, Ichimura H, Todd P (1997) Matching as an econometric
ment effect. Stata J. 7(1):71–83.
evaluation estimator: Evidence from evaluating a job training
Bertrand M, Duflo E, Mullainathan S (2004) How much should programme. Rev. Econom. Stud. 64(4):605–654.
we trust differences-in-differences estimates? Quart. J. Econom.
Ho SY, Bodoff D, Tam KY (2011) Timing of adaptive Web person-
119(1):249–275.
alization and its effects on online consumer behavior. Inform.
Blundell R, Dias MC (2000) Evaluation methods for non- Systems Res. 22(3):660–679.
experimental data. Fiscal Stud. 21(4):427–468.
Huang Q, Nijs VR, Hansen K, Anderson ET (2012) Wal-Mart’s
Bolton RN (1998) A dynamic model of the duration of the cus- impact on supplier profit. J. Marketing Res. 49(2):131–143.
tomer’s relationship with a continuous service provider: The
Huerta MRS, Masoud T, Alonso GL (1998) Wine differentiation
role of satisfaction. Marketing Sci. 17(1):45–65.
according to color using conventional parameters and volatile
Bound J, Jaeger D, Baker R (1995) Problems with instrumental vari- components. J. Food Composition Anal. 11(4):363–374.
ables estimation when the correlation between the instruments
and the endogenous explanatory variable is weak. J. Amer. IBM Report (2011) 11 Key Marketing Trends for 2011. Unica/IBM,
Statist. Assoc. 90(430):443–450. New York, NY.
Brynjolfsson E, Hu Y(J), Smith MD (2010) The longer tail: The Janiszewski C (1993) Pre-attentive mere exposure effects. J. Con-
changing shape of Amazon’s sales distribution curve. http:// sumer Res. 20(3):376–392.
dx.doi.org/10.2139/ssrn.1679991. Jensen JS, Demiray S, Egego M, Meyer AS (2008) Prediction of wine
Caliendo M, Kopeinig S (2008) Some practical guidance for the color attributes from the phenolic profiles of red grapes (Vitis
implementation of propensity score matching. J. Econom. Sur- vinifera). J. Agricultural Food Chemistry 56(3):1105–1115.
veys 22(1):31–72. Kennedy BW (1981) Variance component estimation and prediction
Cao Z, Song X, Hamilton L (2011) Difference-in-difference and of breeding values. Canadian J. Genetics Cytology 23(4):565–578.
instrumental variables approaches: An alternative and com- Kietzmann JH, Hermkens K, McCarthy IP, Silvestre BS (2011) Social
plement to propensity score matching in estimating treatment media? Get serious! Understanding the functional building
effects. Thomson Reuters; CER Issue Brief, http://healthcare blocks of social media. Bus. Horizons 54(3):241–251.
.thomsonreuters.com/cer/assets/DifferenceCERIssueBrief.pdf. Lechner M (2002) Some practical issues in the evaluation of het-
Chetty R, Looney A, Kroft K (2009) Salience and taxation: Theory erogeneous labor market programmes by matching methods.
and evidence. Amer. Econom. Rev. 99(4):1145–1177 J. Royal Statist. Soc. A 165(1):59–82.
McAlexander JH, Schouten JW, Koenig HF (2002) Building brand Susarla A, Oh J-H, Tan Y (2012) Social networks and the diffusion
community. J. Marketing 66(1):38–54. of user-generated content: Evidence from YouTube. Inform. Sys-
Morgan RM, Hunt SD (1994) The commitment-trust theory of rela- tems Res. 23(1):23–41.
tionship marketing. J. Marketing 58(3):20–38. Tellis GJ, Weiss D (1995) Does TV advertising really affect sales?
Mudambi SM, Schuff D (2010) What makes a helpful online review? J. Advertising 24(3):1–12.
A study of customer reviews on Amazon.com. MIS Quart. Tucker C, Zhang J, Zhu T (2012) Days on market and home
34(1):185–200. sales. Working paper. MIT Sloan School of Management,
Pavlou P, Liang H, Xue Y (2007) Uncertainty and mitigating uncer- Cambridge, MA.
tainty in online exchange relationships: A principal–agent per- Venkatesan R, Kumar V (2004) A customer lifetime value frame-
spective. MIS Quart. 31(1):105–131. work for customer selection and resource allocation strategy.
Reinartz WJ, Kumar V (2000) On the profitability of long-life cus- J. Marketing 68(1):106–125.
tomers in a noncontractual setting: An empirical investigation
Verhoef PC, Donkers B (2005) The effect of acquisition channels
and implications for marketing. J. Marketing 64(4):17–35.
on customer loyalty and cross-buying. J. Interactive Marketing
Reinartz WJ, Kumar V (2003) The impact of customer relation- 19(2):31–43.
ship characteristics on profitable lifetime duration. J. Marketing.
67(1):77–99. Verhoef PC, Franses PH, Hoekstra JC (2001) The impact of satisfac-
tion and payment equity on cross-buying: A dynamic model
Rigby DK, Reichheld FR, Schefter P (2002) Avoid the four perils of
for a multi-service provider. J. Retailing 77(3):359–378.
CRM. Harvard Bus. Rev. 7(2):101–109.
Rosenbaum PR (2002) Observational Studies (Springer, New York). Wangenheim FV, Bayón T (2007) The chain from customer satisfac-
tion via word-of-mouth referrals to new customer acquisition.
Rosenbaum PR, Rubin DB (1983) The central role of the propen-
J. Acad. Marketing Sci. 35(2):233–249.
sity score in observational studies for causal effects. Biometrika
70(1):41–55. Wooldridge JM (2002) Econometric Analysis of Cross Section and Panel
Rubin DB (2006) Matched Sampling for Causal Effects (Cambridge Data (MIT Press, Cambridge, MA).
University Press, New York). ZDNet (2011) Two thirds of small business owners use Facebook
Rust RT, Lemon KN, Zeithaml VA (2004) Return on marketing: for marketing. Available at http://www.zdnet.com/blog/
Using customer equity to focus marketing strategy. J. Marketing facebook/two-thirds-of-small-business-owners-use-facebook
68(1):109–127. -for-marketing/1529. Accessed October 26, 2012.
Smith JA, Todd PE (2005) Does matching overcome LaLonde’s cri- Zhu F, Zhang X (2010) Impact of online consumer reviews on sales:
tique of nonexperimental estimators? J. Econometrics 125(1–2): The moderating role of product and consumer characteristics.
305–353. J. Marketing 74(2):133–148.
Stewart MB, Swaffield JK (2004) The other margin: Do minimum Zhu L, Benbasat I, Jiang Z (2009) Let’s shop online together: An
wages cause working hours adjustments for low-wage work- empirical investigation of collaborative online shopping sup-
ers? Working paper, University of Warwick, UK. port. Inform. Systems Res. 21(4):872–891.

Rishika, R., Kumar, A., Janakiraman, R., & Bezawada, R. (2013) - The Effect of Customers' Social

Uploaded by

Copyright:

Available Formats

Rishika, R., Kumar, A., Janakiraman, R., & Bezawada, R. (2013) - The Effect of Customers' Social

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Rishika, R., Kumar, A., Janakiraman, R., & Bezawada, R. (2013) - The Effect of Customers' Social

Uploaded by

Copyright:

Available Formats

Information Systems Research

Vol. 24, No. 1, March 2013, pp. 108–127

The Effect of Customers’ Social Media Participation

1. Introduction while also providing them the opportunity to share

Treatment group Before matching After matching

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

Variables Estimate SD Estimate SD Estimate SD Estimate SD Estimate SD

Table 6 Impact of Social Media Participation on Customer Profitability

Variables Estimate SD Estimate SD Estimate SD Estimate SD Estimate SD

Purchase amount Buy focus

0.09 Frequency of visit

Difference in difference estimate

Customer segments Customer segments

Deal sensitivity Premium share

Difference in difference estimate

0.03 High 0.03 Low

Customer segments Customer segments

Treatment window (post launch period)

15 months 12 months 9 months 6 months 3 months

Frequency of visits Profitability

You might also like