Topic: Downfall of Brick and Mortar stores
RESEARCH NOTE
HERDING AND SOCIAL MEDIA WORD-OF-MOUTH: EVIDENCE FROM GROUPON1
Xitong Li Department of Information Systems and Operations Management, HEC Paris, 1 Rue de la Libération,
Jouy en Jousas Cedex, 78351, FRANCE {lix@hec.fr}
Lynn Wu The Wharton School, University of Pennsylvania, 3733 Spruce Street,
Philadelphia, PA 19104-6340 U.S.A. {wulynn@upenn.edu}
Modern online retailing practices provide consumers with new types of real-time information that can poten- tially increase demand. In particular, showing sales information to a customer can increase certainty about product quality, inducing consumers to herd. This effect can be particularly salient for experience goods due to their quality being inherently highly uncertain. Social media word-of-mouth (WOM) can increase product awareness as product information spreads via social media, increasing demand directly while amplifying existing quality signals such as past sales. This study examines the mechanisms behind the strategy of facili- tating herding and the strategy of integrating social media platforms to understand the potential complemen- tarities between the two strategies. We conduct empirical analysis using data from Groupon.com, which sells goods in a fast cycle format of “daily deals.” We find that facilitating herding and integrating social media platforms are complements that generate sales, supporting the idea that it is beneficial to combine the two strategies on social media platforms. Furthermore, we find that herding is more salient for experience goods, consistent with our hypothesized mechanisms, while the effect of social media WOM is similar for experience goods and search goods.
Keywords: Herding, word-of-mouth, social media, interaction effect, complementarity
Introduction 1
With advances in online marketing technology, retailers have acquired two types of strategies to provide product informa- tion and increase product awareness, which may increase demand individually or collectively. One strategy is to dis- play past sales information to potential consumers. For example, Amazon.com employs this strategy by showing a product’s sales rank in its description. The other strategy is to integrate social media platforms (e.g., Facebook) into the sales experience, allowing people to “like” products on the platform. Social media integration appears to be common
among large retailers such as Walmart.com and BestBuy.com as well as many smaller retailers with limited marketing resources. Often, these retailers integrate social media icons (e.g., Facebook’s Like button) directly on the product page, allowing users to share the products on social media plat- forms. Researchers find the strategy of showing past sales can affect consumers’ interest in the products and their pur- chase choices by reducing uncertainty about product quality (Tucker and Zhang 2011), inducing consumers to herd (Zhang and Liu 2012). Researchers find that the strategy of inte- grating social media platforms can influence consumers’ product adoption by creating online word-of-mouth (WOM) which can simultaneously decrease quality uncertainty and increase product awareness (Aral and Walker 2011; Della- rocas 2003).
1Paul Pavlou was the accepting senior editor for this paper. Gal Oestreicher- Singer served as the associate editor.
DOI: 10.25300/MISQ/2018/14108 MIS Quarterly Vol. 42 No. 4, pp. 1331-1351/December 2018 1331
Li & Wu/Herding and Social Media Word-of-Mouth
Although prior studies separately document the effectiveness of each individual strategy, few examine their potential interactive relationship when both strategies are implemented together. For example, Amazon.com implements the stra- tegies of showing past sales and integrating social media plat- forms together on a product page. Showing past sales infor- mation is particularly useful when consumers are uncertain about product quality and use the prior purchase decisions of others as a signal for quality and product fit. Social media WOM can not only increase product awareness through information diffusion via social media but also reduce quality uncertainty. Thus, the two strategies (showing past sales and integrating social media platforms) can be mutually rein- forcing for at least two reasons: (1) the awareness effect of social media WOM can magnify the effect of showing past sales information by exposing more people to the information, and (2) past sales information and social media WOM could provide independent (and likely different) signals of product quality. Thus their combination may reinforce each other, leading to a greater reduction in uncertainty about product quality than either strategy could accomplish on its own (Kirmani and Rao 2000).2 Therefore, we expect that showing past sales information and integrating social media platforms are complements, especially when consumers harbor substan- tial uncertainty about product quality.
Furthermore, we uncover the underlying mechanism for each strategy, which can also explain their interaction effects. Prior studies on the effect of showing past sales often assume that it could reduce quality uncertainty, but they do not expli- citly validate this assumption, especially for newer marketing formats such as daily deals (Duan et al. 2009; Tucker and Zhang 2011). Examining the underlying mechanism would help us understand the boundary conditions under which the strategy of showing past sales could be effective and under which it could not. On the other hand, prior studies on social media platforms largely recognize that social media WOM can increase product awareness (Aral and Walker 2011; Chen et al. 2015; Liu 2006), but do not sufficiently examine the implications of social media WOM as a way to reduce quality uncertainty (e.g., “liking” a product on Facebook can be con- sidered as an implicit social media endorsement (Li 2018)). Thus, by exploring the primary mechanisms through which showing past sales and integrating social media affect future sales respectively, we can better understand why the two strategies could be complements and how to maximize the return on their effects. To examine their underlying mech-
anisms, we exploit the inherent variation in quality uncer- tainty for a variety of products. Specifically, we compare experience goods for which quality is difficult to ascertain prior to consumption with search goods for which quality is easier to ascertain (Hong et al. 2014; Nelson 1974). If the primary mechanism is to reduce quality uncertainty, we expect the effect to be larger for experience goods than for search goods. However, if the primary mechanism is to increase product awareness, the effect would not differ between the two types of goods.
Our data comes from Groupon.com, a daily deal site that provides short-term sales events of deeply discounted vouchers (Dholakia 2011). While daily deal retailing is potentially interesting due to its market size ($5.5 billion in 2016), rapid growth (23% annually in 2013)3 and limited prior research coverage,4 it is also a particularly useful setting to investigate information diffusion effects. The short cycle of daily deals allows rapid information diffusion processes to be explored without being confounded by factors such as adver- tising or changes in the underlying product quality. Groupon and other daily deal sites are also much more aggressive in their use of social media and the distribution of extremely recent past sales information (placing this information prominently in the product description) than other retailers, making this platform a good test bed for examining how these strategies perform both individually and together. Further- more, daily deal sites simultaneously sell a large variety of goods with wide-ranging degrees of uncertainty in product quality. For instance, restaurants and salons are typically classified as “experience goods” with a high uncertainty in quality (see Weathers et al. 2007), whereas the quality of consumer electronics items (e.g., iPads) is well understood by most potential buyers, leaving relatively greater certainty about product quality (“search good”). This cross-sectional variation in quality uncertainty is useful in distinguishing the effect of observing past sales from that of social media WOM.
Using hourly panel data about Groupon deals, we find that past cumulative sales and Facebook-mediated WOM both positively affect future sales after controlling for deal-fixed effects, time-fixed effects, and linear and nonlinear time trends that mimic product diffusion. We find robust evidence of complementarity between herding and Facebook-mediated WOM. The economic significance of the complementarity is also substantial: products receiving an average number of Facebook likes can boost the effect of prior cumulative sales
2In addition, both strategies may be subject to declining returns—past sales may signal oversubscription (Eyster et al. 2014; Horton 2014) which deters purchase. Similarly, the incentive to make a social media post may be reduced if there are already many similar posts (Lobel et al. 2017), thus providing an endogenous limitation on WOM.
3The Statistics Portal, http://www.statista.com/statistics/264797/projected-us- consumer-spending-on-daily-deal-websites/.
4There are a few recent exceptions (Li 2018; Li and Wu 2013; Luo et al. 2014; Wu et al. 2015).
1332 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
on next-hour sales by more than 22%, compared to those with zero Facebook likes. The effect on sales can further com- pound over time as these signals grow stronger because of the complementarities. To examine the underlying mechanisms, we classify products into experience goods and search goods using three different methods: surveys, data-driven methods based on product reviews, and product pricing. These methods complement each other in addressing the weaknesses of each individual approach alone. We show that the effect of past sales is stronger for experience goods than for search goods, while there is no difference in the effects of Facebook likes on the two product categories. This result is consistent with our hypothesis that the display of past sales data primarily reduces quality uncertainty while social media WOM can primarily increase product awareness. The com- plementarity that we observe appears to result from the two different mechanisms reinforcing each other such that the joint use of the strategies increases the effects compared to implementing them separately.
Related Literature
As daily deals have become a popular marketing tool, it is important to understand how they can spur demand. Prior studies examine daily deals in the framework of coupons (Kumar and Rajan 2012). Luo et al. (2014) examine how the popularity of daily deals affects the two-stage decision pro- cess of buying the deals. Our work differs from prior studies by examining a relatively unique aspect of daily deals that leverage both strategies of facilitating herding and of using social media integration, thus allowing the exploration of their interaction effects on demand. Because of the variety of products offered on daily deals, we compare experience goods with search goods to determine the mechanisms behind these strategies and to explain the effectiveness of daily deals apart from traditional marketing vehicles.
The theoretical work on herding shows that agents make decisions by observing the actions of earlier actors (Banerjee 1992; Bikhchandani et al. 1992). If prior decisions have con- verged to a single choice, subsequent agents would simply follow that choice regardless of their own private information. The empirical evidence of herding has been abundantly docu- mented in recent works. For example, in the context of micro-financing, both Herzenstein et al. (2011) and Zhang and Liu (2012) find that lenders tend to herd on the converged choices made by previous lenders.
WOM, a well-established construct in the literature (Della- rocas et al. 2007; Godes and Mayzlin 2004), is shown to increase product awareness (Liu 2006). Godes and Mayzlin
(2009) find that WOM transmitted through acquaintances is most effective in driving product sales. Social media WOM has also been compared to other traditional marketing chan- nels. Trusov et al. (2009) find that social media has a strong carryover effect on the growth of memberships in social media platforms. Aral and Walker (2011) show social media WOM can significantly increase the adoption of a free application on Facebook. Our work is most related to the work of Chen et al. (2011), which leverages a policy change in a major retailer and documents herding and WOM that arise from recommen- dation systems. Our study differs because we focus on social media WOM. A key difference is that social media WOM encourages users to visit the product site while recommender- based WOM only works after the site is visited. Furthermore, we also explore the underlying mechanism behind social media WOM and how it differs from herding to explain their complementary relationship, a phenomenon that is under- explored in the literature.
Hypothesis Development
Herding Through Showing Past Sales
The economics literature (Banerjee 1992; Bikhchandani et al. 1992) suggests that people with imperfect private information about a decision would learn and update their beliefs by observing the prior decisions of others. This process is called observational learning (Cai et al. 2009; Zhang 2010). When the prior decisions of others converge, the economics litera- ture shows that people favor going with the converging deci- sion over their own private information, leading them to herd (Herzenstein et al. 2011; Zhang and Liu 2012). In a business context, because past sales (prior consumers’ purchase deci- sions) can serve as a positive signal for quality, a product with higher past sales would receive more sales from subsequent consumers, resulting in a positive relationship between past sales and future sales (Tucker and Zhang 2011). While there are other mechanisms that can also generate a positive rela- tionship, in this paper we use the term herding to refer to the positive effect of past sales on future sales that is triggered through the process of observational learning.
Daily-deal sites can facilitate consumer herding by promi- nently showing the cumulative voucher sales of each deal in real-time (see Figure 1, where the past sales are highlighted by a box). The information of past cumulative sales is a useful signal to reduce quality uncertainty. As Dholakia (2011) shows, more than 80% of those who bought a daily deal are customers new to the participating business and they
MIS Quarterly Vol. 42 No. 4/December 2018 1333
Li & Wu/Herding and Social Media Word-of-Mouth
Figure 1. Screenshot of a Typical Deal Featured by Groupon.com (The total number of voucher sales and the buttons for Facebook likes and Twitter are circled.)
would find the information about past sales particularly useful for making purchase decisions (Tucker and Zhang 2011). For example, suppose two restaurant deals have identical charac- teristics. All else being equal, new customers are more likely to choose the one with higher past sales because that data is seen as indicative of a higher product quality.
Social Media Word-of-Mouth
WOM refers to the dissemination of information from one person to another. It can be measured by volume (e.g., the number of messages transmitted) or valence (e.g., the senti- ment of the disseminated information). The valence of WOM can affect sales through conveying sentiments, whereas the volume of WOM can affect sales through increasing product awareness. Although both valence and volume can influence product sales (Chevalier and Mayzlin 2006; Chintagunta et al. 2010; Dellarocas et al. 2007), we focus on volume because valence in our setting is constant; each click on the Facebook “Like” button disseminates the same message and thus iden- tical valence.
While many WOM studies have focused on traditional online channels such as recommender systems or forums (Fleder and Hosanagar 2009; Hosanagar et al. 2014; Xiao and Benbasat
2007), we focus on WOM spread through social media because many retailers, including those promoting daily deals, have started using social media to promote products (see Figures 2 and 3). For example, after clicking the Facebook “Like” button on a product page, the activity could be dis- played on friends’ newsfeeds (see Figure 2).5 Upon seeing the message, users can follow the link to visit the product page. Twitter uses a similar mechanism to disseminate deal information (see Figure 3). In contrast to traditional online WOM (e.g., online reviews) where users pull information by first visiting the product page, social media can push product information to new users who have not yet visited the product page. This push method can provide an extra channel to generate awareness.
The awareness effect can be further enhanced because social media spreads product information through social ties. According to a recent statistic, the average Facebook user has 229 Facebook friends.6 This suggests that a Facebook like
5Facebook may algorithmically select the friends who will see a user’s post, with the purpose of enhancing the effect of Facebook likes and to reduce annoyance.
6See Embrace Disruptionpr at http://embracedisruption.com/2013/01/08/an- average-facebook-user-has-229-friends-100-social-stats-from-2012/.
1334 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
Figure 2. A Groupon Deal Is Shared on a Facebook News Feed
Figure 3. A Groupon Deal is Shared on Twitter
could potentially push a deal to at least some of these friends who may then purchase the deal.7 In addition to increasing
product awareness, social media WOM can also reduce quality uncertainty, because consumers often choose to endorse superior products by sharing a deal via Facebook as a way to enhance their self-image or identity (Akerlof and Kranton 2000; Berger and Heath 2007; Wojnicki and Godes 2008). Furthermore, observing multiple social media posts
7The actual number of posts shown depends on the Facebook Feeds Algorithm which is not publicly disclosed.
MIS Quarterly Vol. 42 No. 4/December 2018 1335
Li & Wu/Herding and Social Media Word-of-Mouth
about the same product or seeing the actual number of Face- book likes associated with the product8 can reduce quality uncertainty, as can past sales information. Yet the effect of social media WOM could also be different from that of past sales information. Social media endorsements come from friends who may not have bought the product (a non- purchase-based signal from social ties), whereas past sales information is a purchase-based signal from anonymous strangers (Muchnik et al. 2013). As both mechanisms (increasing product awareness and reducing quality uncer- tainty) can help increase product sales, we expect deals shared by more Facebook likes to receive more sales in subsequent time periods.
Interaction Effect
Because online retailers can potentially implement the stra- tegies of showing past sales and integrating social media platforms together or separately, it is important to examine how the two strategies could interact with each other. If the primary effect of social media WOM is to increase product awareness, social media could complement the strategy of showing past sales aimed at primarily reducing quality uncertainty. For uninformed buyers, as is often the case with Groupon buyers (Dholakia 2011), external signals, such as past sales information, can reduce quality uncertainty for potential buyers and thus increase their likelihood of buying (Tucker and Zhang 2011). By using social media WOM to increase product awareness among potential consumers first and then using high past sales to reduce quality uncertainty for these consumers, the two strategies can work together as complements to generate more sales than either strategy could on its own.
On the other hand, if reducing quality uncertainty is the primary mechanism by which social media WOM increases sales, it could also complement the strategy of showing past sales, albeit in different ways. An endorsement via social media can reduce quality uncertainty because the information embedded in social media WOM comes from one’s social ties that are more likely to have similar tastes or know about a person’s idiosyncratic preferences (McPherson et al. 2001). This type of information may differ from that of past sales. Whereas past sales information comes from anonymous consumers who have purchased the products, social media WOM are endorsements from social ties who may not necessarily have purchased them (Li 2018). Accordingly, the anonymous but purchase-based information (past sales) and
the social but non-purchase-based information (social media WOM) can reinforce each other to further reduce quality uncertainty. Upon seeing high past sales that reduce the general uncertainty about product quality, endorsements from social ties could further signal a good product fit that is speci- fic to the person, creating complementarities that magnify the effect of each individual strategy. Therefore, we expect social media WOM and showing past sales to be complements.
Hypothesis (H1): There is a positive interaction effect between past cumulative sales and past cumulative social media WOM in generating incremental sales in the subsequent time period.
Experience Goods Versus Search Goods
The literature suggests that products can be classified into experience and search goods based on how difficult it is to ascertain product quality before consumption (Hong et al. 2014; Nelson 1974). Experience goods are those whose product quality is difficult to ascertain before consumption, such as restaurants, spas, massages, and cleaning services. Thus, the descriptive information on the product page is often insufficient to ascertain the quality of experience goods, and consumers must seek external quality signals (e.g., past sales information) for further assessment. By contrast, the product quality of search goods (e.g., photo frames) is easier to ascer- tain before consumption. The description on the product page often has sufficient information for consumers to assess product quality, and additional signals would not necessarily be helpful. Thus, we expect showing past sales is more useful in reducing quality uncertainty for experience goods than for search goods.
Hypothesis (H2): The effect of past cumulative sales on the next-period sales is greater for experi- ence goods than for search goods.
On the other hand, the volume of social media WOM (e.g., Facebook likes) represents the cumulative endorsements received (Li 2018) and accordingly, it can also serve as an external signal to reduce quality uncertainty. Comparing the effect of social media WOM between experience and search goods can help determine how social media affect sales. If the primary mechanism driving the effect of social media WOM is to reduce quality uncertainty (rather than to increase product awareness), we expect social media WOM to have a larger impact for experience goods than for search goods. However, if the primary mechanism driving the effect of social media WOM is to increase product awareness, its effect would not differ between the two product categories, because increasing awareness for a product should not depend on its
8It is possible for some browsers to display the total number of likes, albeit the font is very small.
1336 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
experience or search attributes. To test the primary mech- anisms, we hypothesize that
Hypothesis (H3): The effect of past cumulative social media WOM on the next-period sales is greater for experience goods than for search goods.
Data and Empirical Model
We use data from Groupon.com, one of the largest daily-deal sites, to test the hypotheses. Groupon provides a unique setting because it implements both of the strategies: showing past sales and integrating social media platforms. Every day, it features a single deal on the main page for each local market. Figure 1 shows a screenshot of a typical featured deal. The deal site shows the characteristics of the deal, including a brief description, vendor, discounted voucher price, and percentage discount. The total number of vouchers sold is prominently displayed in real-time, allowing shoppers to observe prior purchasing decisions. The Facebook “Like” and Twitter “Share” buttons are displayed below the sales information, allowing shoppers to share the deal on social media. The number of tweets is not shown on the deal page, but the number of Facebook likes is occasionally observable although it is much less visible than other information on the page and is not displayed at all in Internet Explorer, which is the dominant browser at the time of our data collection. Because the number of Facebook likes is either invisible or easily neglected due to its lack of visual prominence (Nisbett and Ross 1980), we assume that the primary channel for Face- book likes to affect future sales is through sharing the deal on social media as opposed to observing the actual number of social media endorsements. Our subsequent analysis also suggests that Facebook likes differ from past sales in their effects on the next-period sales, consistent with this assumption.
Data Collection
We used Cameleon Web Wrapper (Firat et al. 2000) to crawl hourly sales information and deal characteristics, including discounted voucher price, original/face value, and the product category. To collect social media activities, we used Face- book’s and Twitter’s open APIs to collect the number of likes and tweets associated with a Groupon deal for each hour of the day. We sampled six metropolitan areas in the United States, including East Coast (Boston, New York City), Cen- tral (Chicago, Houston), and West Coast (Los Angeles, San Francisco). Our data collection period spanned July 1, 2011,
to September 27, 2011, when Groupon stopped displaying the number of sales for each deal.9 Accordingly, the data set included 526 featured deals. For each deal, we collected hourly data on the number of voucher sales and the number of Facebook likes and tweets during the first day when the deal was featured. We focus on analyzing the data on the first day because, in most cases, a Groupon deal was only featured during the first day of the sale and the majority of the sales occurred on that day. Twenty-six deals (4.9%) had errors and were removed from the dataset. Because the distribution of voucher sales is highly skewed toward the right (that is, a few deals had extremely high sales), we also removed deals with final sales above the 95th percentile. In total, we have an unbalanced panel data set consisting of 463 deals for 24 hourly periods.
Table 1 presents the descriptive statistics about deal characteristics. On average, the discounted voucher price is $110.66 with an average discount rate of 57%. The average number of vouchers sold during the first day is 647, gener- ating $71,419 in revenue for an average featured Groupon deal. The average number of Facebook likes and tweets received during the first day is 78.46 and 8.64, respectively.
Empirical Model
We denote the natural log of cumulative voucher sales of deal i up to the tth hour of the day by Yi,t, where t = 1,2,…,24. Because the cumulative sales are prominently displayed in real-time on Groupon.com and directly observed by con- sumers, we capture the past sales information using the one- hour lag of cumulative sales Yi,t-1 which reflects the aggregate purchases before the tth hour. We use the natural log of incre- mental sales during the tth hour as the dependent variable, denoted by yi,t. In keeping with the estimation specification suggested by Duan et al. (2009) and Zhang and Liu (2012), we estimate the effect of past sales, Yi,t-1, on the next-period sales, yi,t, after controlling for deal and time fixed effects as well as other time-varying variables. The natural log of the cumulative number of Facebook likes or tweets associated with deal i up to the tth hour is denoted by FBi,t and TWi,t, respectively.10 We use the one-hour lag of cumulative Face- book likes and tweets (FBi,t-1, TWi,t-1) to avoid the potential simultaneity between sales and social media WOM in the same period. In the log-log estimation specification, we esti-
9See “AllThingsD” at http://allthingsd.com/20111010/groupon-makes-it- less-possible-to-track-how-well-it-is-doing/.
10In the data set, some deals have zero cumulative sales, Facebook likes, or tweets at certain time points. To include those observations, we add 1 to the cumulative amounts before taking the natural log transformations.
MIS Quarterly Vol. 42 No. 4/December 2018 1337
Li & Wu/Herding and Social Media Word-of-Mouth
Table 1. Descriptive Statistics
Variable N Mean s.d. Min. Max.
Voucher price 463 110.36 364.88 2 2999
Original value 463 283.57 862.04 5 7900
Discount rate 463 57.37 10.91 37.3 95.0
Total voucher sales 434 647.15 643.70 0 3157
Total Facebook likes 434 78.46 104.53 0 782
Total tweets 434 8.64 23.85 0 460
Notes: The descriptive statistics of total sales, total Facebook likes and tweets are based on 434 deals for which we collected the observations at the end of the day.
mate the effect of past cumulative sales (β1) and social media WOM (β2, β3). Thus, our base specification is in Equation 1:
yi,t = α1 +β1 Yi,t-1 + β2 FBi,t-1 + β3 TWi,t-1 + μi + νt + εi,t (1)
Product fixed effects (μi) control for deal-specific hetero- geneity that is invariant over time, including any observable (voucher price and location) and unobservable characteristics (inherent quality of the good). In particular, the unobservable characteristics of a deal are likely to be time-invariant and thus captured by μi, because we focus on sales during the first 24 hours (i.e., the first day) and it is unlikely that any inherent quality of the deal would change during the first day. We use the hour-fixed effects (νt) to eliminate temporal variations such as Internet use and shopping activity over the day. Table 2 presents the overall mean and standard deviation of the time-varying variables in the data set.
The literature on product diffusion and social contagion shows that the adoption rate increases as the user base grows and then decreases when approaching saturation (Young 2009). This nonlinear shape of sales over time may cause a spurious relationship between past cumulative sales and future sales. Prior studies (Carare 2012; Duan et al. 2009) address this issue by including the linear and quadratic terms of product age. In our context, product age is operationalized as the number of hours (t) that has elapsed since the deal was first announced. To allow for some heterogeneity in the diffusion process but maintain enough degrees of freedom, we follow the strategy suggested by Duan et al. (2009) to allow the coefficients of the deal age and its quadratic term to vary across the six different metropolitan areas (indexed by j), as shown in Equation 2.
yij,t = α1 + β1Yij,t-1 + β2 FBij,t-1 + β3 TWij,t-1 + γj,1t + γj,2t 2
+ μi + νt + εij,t (2)
To explore the interaction effect between herding and social media WOM, we include the interaction terms between past
cumulative sales and Facebook likes and between past cumu- lative sales and tweets, as shown in Equation 3. The coeffi- cients of the interaction terms η1 and η2 estimate the interaction effects of showing past sales with Facebook- mediated WOM and Twitter-mediated WOM, respectively.
yij,t = α1 + β1Yij,t-1 + β2 FBij,t-1 + β3 TWij,t-1 + η1Yij,t-1FBij,t-1 + η2Yij,t-1TWij,t-1 + γj,1t + γj,2t
2+ μi + νt + εij,t (3)
Classifying Experience Goods and Search Goods
While a binary classification between experience goods and search goods is used in the prior literature, a continuous spectrum is more appropriate (Nelson 1981) because products often have both experience and search attributes (Caves and Williamson 1985). However, allocating products along this spectrum is not trivial. We use three methods to perform the classification. First, we use the traditional surveys that ask participants to rate products. It is labor intensive (Weathers et al. 2007), especially when there are many products to be assessed (our sample has about 500 deals). Second, we employ an automatic data-driven method that uses online reviews to assess the experience and search attributes of products (Hong et al. 2014). Compared to surveys, the data- driven method is algorithmic, scalable, and less costly to implement. However, it can only be used for products with a sufficient number of online reviews, whereas surveys can be used for any product. Finally, we use product pricing to classify products (Smith 1990) because the risk of choosing a wrong product is higher the more expensive the good is. The price of a product is easy to measure, but it is only a rough proxy for the experience attributes. Although all three methods have their own strengths and weaknesses, we imple- ment all of them to reduce the idiosyncrasies associated with any particular method, especially if the key findings do not vary across methods.
1338 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
Table 2. Pearson Correlation among Time-Varying Variables
Variable Mean s.d. 1 2 3
1. Log of Incremental sales: yij,t 2.43 1.49 1
2. Log of past cumulative sales: Yij,t-1 4.24 2.10 0.749 1
3. Log of past cumulative Facebook likes: FBij,t-1 2.39 1.66 0.548 0.806 1
4. Log of past cumulative tweets: TWij,t-1 1.54 0.71 0.307 0.457 0.494
Survey-Based Assessment
We administered an online survey in April 2016 and recruited participants on Amazon Mechanical Turk to classify whether a good possesses more experience attributes or search attributes. Each participant was asked to assess 20 randomly selected deals. For each deal, the participants answered four questions about each good using a seven-point Likert-type scale, as suggested in the literature (Hong and Pavlou 2014; Weathers et al. 2007). The first two questions assess the experience components. Specifically, we ask the users to rate (1) “It is important for me to experience this product to evaluate how well it will perform,” and (2) “It is important for me to try this product to evaluate how well it will perform.” The other two questions ask about the search components: (3) “I can adequately evaluate this product using only provided information about the product’s attributes and features,” and (4) “I can evaluate the quality of this product simply by reading information about the product.” The difference between the answers to the first two questions (experience-based measurement) and the answers to the latter two questions (search-based measurement) reflects a parti- cipant’s subjective assessment about the degree to which a product possesses experience or search attributes. To avoid any anchoring effect, we ensured the sequence of the four questions was random and each product was assessed by at least 15 different participants. As we expected, the assess- ments of the experience components and the search com- ponents are negatively correlated with acceptable inter-rater reliability (Krippendorff 2004) and internal consistency (Cronbach 1951).11 Therefore, we use the average assessment to measure the extent to which a deal has experience attributes. The distribution based on the average assessments is unimodal (mean = 0 28, s.d. = 0.92, median = 0.27). To make a clear classification, we define deals with experience scores higher than the 75th percentile as experience goods and those lower than the 25th percentile as search goods.
Data-Driven Method Based on Online Reviews
Hong et al. (2014) develop an automatic data-driven method for classifying experience goods using online reviews. Because review ratings for search goods converge faster than they do for experience goods (Hong et al. 2014), we can measure the experience attributes of a good by how fast the ratings converge. Specifically, the variance of the ratings of search goods would decrease with each additional review, whereas the variance of the ratings of experience goods would remain relatively constant. According to Hong et al., a large positive estimation between rating variance and the number of reviews indicates that the ratings are slow to converge, exhibiting more experience attributes than search attributes. Using this insight, we collected 67,063 unique online reviews from Yelp.com for 374 deals in our sample. Following the practice suggested by Hong et al., we retained deals with at least 25 Yelp reviews, a number that is sufficient to calculate the variance. We also dropped the first five reviews for each deal to eliminate potential biases from review manipulations. For each deal, we estimate the relationship between the review variance for the n previous reviews and the sequence of the review n. Then, as mentioned above, we classify deals as experience goods when the estimated slope is above the 75th percentile and as search goods when it is below the 25th
percentile.
Price-Based Classification
Smith (1990) shows price can be used to classify experience goods because they are more likely to have higher prices. As the financial stake rises with the cost of a good, outside signals (e.g., past sales) are increasingly likely to have an effect on the purchase decisions for expensive goods. Ac- cordingly, we also classify goods as expensive if the price is above the 75th percentile ($69) and as low-priced if the price is below the 25th percentile ($20).
11Krippendorff’s alpha > 0.82 and Cronbach’s alpha > 0.76.
MIS Quarterly Vol. 42 No. 4/December 2018 1339
Li & Wu/Herding and Social Media Word-of-Mouth
Results
Main Effects of Past Sales and Social Media WOM
We estimate the main effects of past sales and social media WOM using deal and hour fixed effects with and without city- specific product diffusion effects. All standard errors are clustered at the deal level. Table 3 reports the results. Columns 1–3 show that past sales, Facebook likes and Twitter tweets all have a positive effect on sales when estimated individually. Column 4 shows their effects when estimated together in a single model and Column 5 reports the estimates when linear and quadratic city-specific product diffusion trends are included. We find that the effects of past cumu- lative sales and Facebook likes are both positive, but the effect of tweets is not statistically significant. According to the estimates in Column 5, an increase of 10% in past sales, on average, is associated with an increase of 4.06% in sales in the next hour, while an increase of 10% in past Facebook likes is associated with an increase of 1.12% in sales in the next hour. Because the dependent variable is hourly sales, the estimated effects are substantial as past sales and Facebook likes accumulate over time.
The lack of a significant effect for tweets is both interesting and useful, especially given the significant and consistent effect of Facebook likes. While it may result from the tran- sient nature of tweets, the insignificant effect of tweets can help reduce certain omitted variable bias. If unobserved confounding factors are correlated, social media activities (such as promotional activities or market condition), tweets (TWij,t-1) could be used to control for these unobserved factors. If the correlation between the unobserved confounding factor and TWij,t-1 is strong, then the insignificance of the effect of tweets suggests that these factors are unlikely to be a concern. However, if the correlation is weak, controlling for TWij,t-1 is not sufficient and we use the dynamic generalized method of moments (GMM) and instrumental variable method to address this concern in the robustness checks.
Complementarity
We use the model specified in Equation 3 to estimate the interaction effect between past sales and social media WOM.12 The independent variables in the interaction terms
are demeaned. Using deal and time fixed effects, we show that past sales and Facebook-mediated WOM have a signi- ficantly positive interaction effect on future sales (Table 4), supporting H1. To gauge the economic significance of this complementarity, we consider the relationship between Yij,t-1 and yij,t for a deal with an average hourly number of Facebook likes (the mean of FBij,t-1 is 2.39). According to the estimates reported in Column 4, we find that having the lowest number of Facebook likes would decrease the estimated coefficient of past sales from 0.429 to 0.334 for an average deal, a decrease of about 22%.13
If an unobserved time-varying confounding factor is corre- lated with past sales (Yij,t-1) or Facebook likes (FBij,t-1) but is only weakly correlated with tweets (TWij,t-1), our results in Table 4 could be biased. To address this concern, we estimate the model using GMM. GMM instruments the endogenous independent variables using their lags, while addressing the fixed effects using first-differencing (Arellano and Bover 1995; Blundell and Bond 1998). GMM has been widely used in the information systems and marketing literature to address the endogeneity of independent variables in panel data analysis (Aral et al. 2012; Bardhan et al. 2013; Burtch et al. 2013; Zhang and Liu 2012), particularly when regressors are a form of the lagged dependent variable, such as in our case. In the context of estimating observational learning, Zhang and Liu (2012) also use GMM to estimate the effect of past sales on future sales in a peer-to-peer lending market.
In this study, we follow the guidelines by Roodman (2009a) and use the Arellano–Bond/Blundell–Bover two-step robust system GMM estimation with orthogonal deviations and small sample correction. In the main analysis, we use a fixed num- ber of lags (that is, from the third to the fifth lags) as the GMM instruments.14 Table 5 reports the estimates from the system GMM, including a number of tests that we conduct to check its validity. For example, Column 2 shows that the p- value of the Hansen test of over-identification restrictions is 0.12, and the p-value of the difference-in-Hansen test of the system GMM instruments is 0.94. Neither the Hansen test of
12 We also include the current hourly WOM in the estimation model as a robustness check and the results remain similar in the size of the effect and statistical significance.
13The marginal effect of a 10% increase in past sales is 4.29% (i.e., hourly incremental sales increase by 4.29%). To calculate the marginal effect when FB likes = 1 (we cannot use FB = 0 because it is undefined in logs), we calculate predicted sales (ŷ0) using the regression estimates in Column 4 and assume the values of all other independent variables are at their respective means. Then, we calculate the predicted value again (ŷ1) at Y = 1.1 * mean(Y), that is, a 10% increase in past sales above the average. We then take the percentage difference between ŷ1 and ŷ0 as the marginal effect of past sales when the number of FB likes is small.
14We also use other sets of lags (e.g., from the 11th to the 19th lags) as the GMM instruments and get similar results.
1340 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
Table 3. Fixed-Effect Estimation for the Direct Effects
(1) (2) (3) (4) (5)
Past cumulative sales: Yij,t-1 0.436***
(0.026) 0.387***
(0.031) 0.406***
(0.029)
Past cumulative Facebook likes: FBij,t-1 0.224***
(0.026) 0.130***
(0.025) 0.112***
(0.024)
Past cumulative tweets: TWij,t-1 0.079**
(0.037) 0.046
(0.028) 0.020
(0.028)
Include linear and quadratic time trends? No No No No Yes
Deal fixed effects Yes Yes Yes Yes Yes
Hour fixed effects Yes Yes Yes Yes Yes
Number of observations 9,781 9,781 9,781 9,781 9,781
Number of clusters 463 463 463 463 463
R² 0.78 0.53 0.45 0.79 0.79
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
Table 4. Interaction Effects between Herding and Social Media WOM
(1) (2) (3) (4)
Past cumulative sales: Yij,t-1 0.406***
(0.029) 0.431***
(0.028) 0.406***
(0.029) 0.429***
(0.028)
Past cumulative Facebook likes: FBij,t-1 0.112***
(0.024) 0.065***
(0.024) 0.103***
(0.024) 0.065***
(0.024)
Past cumulative tweets: TWij,t-1 0.020
(0.028) -0.023 (0.025)
-0.005 (0.029)
-0.027 (0.026)
Yij,t-1 × FBij,t-1 0.037***
(0.004) 0.035***
(0.005)
Yij,t-1 × TWij,t-1 0.041***
(0.008) 0.009
(0.008)
Include linear and quadratic time trends? Yes Yes Yes Yes
Deal fixed effects Yes Yes Yes Yes
Hour fixed effects Yes Yes Yes Yes
Number of observations 9781 9781 9781 9781
Number of clusters 463 463 463 463
R² 0.79 0.80 0.79 0.80
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
MIS Quarterly Vol. 42 No. 4/December 2018 1341
Li & Wu/Herding and Social Media Word-of-Mouth
Table 5. Results from System GMM Estimation
(1) (2)
Past cumulative sales: Yij,t-1 0.594***
(0.037) 0.723***
(0.016)
Past cumulative Facebook likes: FBij,t-1 0.112***
(0.031) 0.032
(0.020)
Past cumulative tweets: TWij,t-1 0.052
(0.051) 0.009
(0.028)
Yij,t-1 × FBij,t-1 0.062***
(0.007)
Yij,t-1 × TWij,t-1 -0.005 (0.016)
Include linear and quadratic time trends? Yes Yes
Deal fixed effects Yes Yes
Hour fixed effects Yes Yes
Number of observations 9781 9781
Number of clusters 463 463
Number of instruments 264 418
Hansen test of over-identification restrictions (p-value) 0.11 0.12
Difference-in-Hansen test of the system GMM instruments (p-value) 0.93 0.94
Arellano–Bond test for AR(2) in differences (p-value) 0.13 0.13
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
over-identification restrictions nor the difference-in-Hansen test of the system GMM instruments rejects the null that the instruments are uncorrelated with the error term, ensuring the validity of the instruments used in the GMM estimation. The number of instruments used in the estimation is 418, less than the number of clusters (463), suggesting that instrument proliferation is not an issue (Roodman 2009b). The Arellano– Bond test for AR(2) autocorrelation in first differences sug- gests that serial correlation is not a concern in the first- differencing equations (p = 0.13). In sum, all the tests pass the standard criteria of using GMM, suggesting that our GMM estimates are reliable (Roodman 2007, 2009a, 2009b). The estimates in Column 1 confirm that past sales and Face- book likes both have significantly positive effects on sales, while the effect of tweets is minimal. The estimates in Column 2 show that there is a positive and significant inter- action effect between past sales and Facebook likes, suggesting that they are complements. According to the estimates in Column 2, an increase of 10% in past cumulative sales data is associated with an increase of 7.23% in hourly sales, but if there were no interaction with Facebook WOM it would be associated with an increase of only 5.75%. This suggests that the complementarity is responsible for an increase of 25.7% in the effect of past sales (from 5.75% to 7.23%).
A particular concern is the potential endogeneity of past sales Yij,t-1. While the system GMM is specifically designed to ad- dress issues of using regressors that are lags of the dependent variable, we use external instrumental variables to instrument Yij,t-1 as another robustness check. Specifically, we instrument the cumulative sales of deal i, in city j, at hour t-1 (i.e., Yij,t-1) using the cumulative sales of a different deal, in the same city j, at the same hour t-1, but on the next day, that is, Y(i+1) j,t-1. The exclusion restriction of the validity is satisfied, because the sales of the different deal on the next day should not directly affect the sales of the focal deal today. The relevance condition of the validity lies in the plausible correlation be- tween Yij,t-1 and Y(i+1) j,t-1, which stems from two aspects. First, from the consumers’ side, the shopping behaviors for buying Groupons in a specific city should be relatively stable in the short run. For example, users may visit Groupon.com at a similar time each day according to their working/leisure schedules (for example, lunch breaks, coffee breaks, etc.). This type of behavior should be relatively stable over two consecutive days, which would result in a high correlation between Yij,t-1 and Y(i+1) j,t-1. Second, from Groupon’s side, it is plausible that the deals featured by Groupon in a specific city are related to each other. For example, Groupon managers may feature two similar deals during two consecutive days, especially if those deals are seasonal, thereby leading to a
1342 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
Table 6. Results from 2SLS Estimation
(1) (2)
Past cumulative sales: Yij,t-1 0.624***
(0.232) 0.696***
(0.216)
Past cumulative Facebook likes: FBij,t-1 0.060
(0.059) 0.0002
(0.057)
Past cumulative tweets: TWij,t-1 0.030
(0.030) -0.005 (0.034)
Yij,t-1 × FBij,t-1 0.040***
(0.011)
Yij,t-1 × TWij,t-1 -0.001 (0.015)
Include linear and quadratic time trends? Yes Yes
Deal fixed effects Yes Yes
Hour fixed effects Yes Yes
Number of observations 8555 8555
Number of clusters 443 443
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
positive correlation between Yij,t-1 and Y(i+1) j,t-1. 15 Empirically,
we find a positive and statistically significant correlations between Yij,t-1 and Y(i+1) j,t-1 in the first-stage analysis after controlling for deal and hourly time fixed effects, and linear and quadratic time trends specific to each city (p < 0.001). The first-stage F-statistic is 11.74, above the critical value of 10, supporting the relevance condition. Then, we use 2SLS to estimate the data and report the results in Table 6 which are qualitatively similar to the previous results in Tables 4 and 5.
The fact that showing past sales is complementary to Face- book but not Twitter (as shown in Tables 4–6) addresses certain types of potential endogeneity issues, such as the possibility that “good” products naturally receive more social media WOM and more sales. While it is easy to believe that both social media WOM and past sales are endogenous, it is more difficult to argue how the presence of both strategies of showing past sales and of integrating social media platforms
yields higher performance than one without the other (as well as higher performance when neither is used), as is implied by the interaction term; see Bresnahan et al. (2002) and Bryn- jolfsson and Milgrom (2013) for more discussion of endo- geneity in the complementarities framework.
Differential Effects between Experience Goods and Search Goods
To uncover the mechanisms underlying the effects of showing past sales and social media WOM, we explore how they differ between experience goods and search goods (Hong et al. 2014; Hong and Pavlou 2014; Nelson 1974). As explained in earlier, we compute an experience score for each product using three different methods: a survey-based method, a data- driven method using online reviews, and a price-based method. We then classify a product as an experience good if the computed experience score is above the 75th percentile and as a search good if the score is below the 25th percentile. Finally, we estimate the effect of past sales and social media WOM for experience goods and for search goods separately, and test whether the effects between the two product cate- gories are different.
Table 7 reports the results from the survey-based classifica- tion method. The effect of past sales on the next-hour sales is 0.457 for experience goods and 0.324 for search goods, respectively. Their difference is statistically significant (diff. = 0.133, s.e. = 0.057, p < 0.05). However, the effect for Face-
15To ensure that this type of interest in Groupon is not due to direct email services from Groupon’s promotional service, we graph the interest in “Groupon” on Google trends in the United States over a 1-week period. Google trends capture the interests on the search term on “Groupon” which comes from Google’s search engine directly. Thus, it would not capture visits to Groupon through direct email or social media because users would simply follow the links embedded in the email or social media posts as opposed to searching for it on Google. The cyclic nature of the graph indicates similar shopping behaviors at the same time each day. To ensure that it is not a general pattern about any search term, we also graphed the interest in “winemaking,” which also exhibits some cyclic pattern but not nearly as prominent as the pattern for Groupon.
MIS Quarterly Vol. 42 No. 4/December 2018 1343
Li & Wu/Herding and Social Media Word-of-Mouth
Table 7. Differential Effects between Experience Goods and Search Goods Using Human Survey Classification Method
Human’s Evaluation
(1) Experience
(2) Search
(3) Diff.
Past cumulative sales: Yij,t-1 0.457***
(0.042) 0.324***
(0.039) 0.133**
(0.057)
Past cumulative Facebook likes: FBij,t-1 0.107***
(0.036) 0.148***
(0.038) -0.041 (0.052)
Past cumulative tweets: TWij,t-1 0.074
(0.058) 0.023
(0.061) 0.051
(0.084)
Include linear and quadratic time trends? Yes Yes –
Deal fixed effects Yes Yes –
Hour fixed effects Yes Yes –
Number of observations 2379 2451 –
Number of clusters 113 117 –
R² 0.82 0.77 –
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
book likes between the two types of goods is not statistically significant (0.107 versus 0.148, diff. = -0.041, s.e. = 0.052, p = 0.43). Table 8 reports similar results from the automatic data-driven classification method. The estimated effects of Yij,t-1 for experience goods and search goods are 0.397 and 0.248, respectively, and their difference is statistically signi- ficant (diff. = 0.149, s.e. = 0.067, p < 0.05). Like the results in Table 7, we find there is no significant difference in the effects of Facebook likes (0.136 versus 0.139, diff. = -0.003, s.e. = 0.056, p = 0.96). Finally, Table 9 reports the results of using the product price as a proxy for experience attributes of a product. We observe that the effect of past sales for high- priced goods is greater than for low-priced goods, and that the difference is statistically significant (diff. = 0.193, s.e. = 0.064, p < 0.01). The difference in the effects of Facebook likes is barely significant and only at the 10% level (0.052 versus 0.148, diff. = -0.096, s.e. = 0.056, p = 0.09). The sign of the difference is opposite of our expectation in that it shows that the effect of Facebook likes is smaller for high- priced goods (proxy for experience goods) than for low-priced goods (proxy for search goods). These results do not seem to support that the effect of Facebook likes comes from reducing quality uncertainty. Nevertheless, we note that product price is only a rough proxy for the experience and search attributes in a product.
To summarize, the results in Tables 7–9 consistently show that the effect of past sales is greater for experience goods than for search goods, supporting H2. This suggests that reducing quality uncertainty is the key mechanism through which past sales can affect future sales. On the other hand,
we do not find supportive evidence for H3. The lack of a sig- nificant difference for Facebook likes suggests that reducing quality uncertainty is not the primary mechanism for social media endorsements to affect future sales. Instead, the effect of Facebook-mediated WOM is likely one of increasing product awareness, because the ability to increase product awareness should be independent of the experience attributes of a product. These mechanisms also explain the complemen- tarities between past sales and social media WOM (H1). With social media WOM acting to increase product awareness for potential consumers and with past sales acting to ascertain product quality for these consumers, the two strategies com- plement each other to generate sales that are additional to what either strategy could achieve alone.
Additional Robustness Checks and Alternative Explanations
Serial Correlation
Serial correlation could be a concern as we estimate the effect of past cumulative sales on future sales in the next hour. To ensure serial correlation does not seriously bias our estimates, we explicitly model a first-order autoregressive relationship εij,t = ρεij,t-1 + eij,t, where ρ captures the first-order autoregres- sion AR(1) and eij,t is a random component, and estimate the data using the fixed-effect AR(1) model. While the results indicate some degree of serial correlation in the data (esti- mated ρ = 0.3), the results reported in Tables 10 and 4 are
1344 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
Table 8. Differential Effects between Experience Goods and Search Goods Using Automatic Data- Driven Classification Method
Yelp Review Rating
(1) Experience
(2) Search
(3) Diff.
Past cumulative sales: Yij,t-1 0.397***
(0.036) 0.248***
(0.056) 0.149**
(0.067)
Past cumulative Facebook likes: FBij,t-1 0.136***
(0.037) 0.139***
(0.042) -0.003 (0.056)
Past cumulative tweets: TWij,t-1 -0.008 (0.088)
-0.067 (0.061)
0.059 (0.107)
Include linear and quadratic time trends? Yes Yes –
Deal fixed effects Yes Yes –
Hour fixed effects Yes Yes –
Number of observations 1707 1648 –
Number of clusters 77 81 –
R² 0.79 0.76 –
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
Table 9. Differential Effects for High- Versus Low-Priced Goods
(1) High-priced
(2) Low-priced
(3) Diff.
Past cumulative sales: Yij,t-1 0.435***
(0.032) 0.242***
(0.056) 0.193***
(0.064)
Past cumulative Facebook likes: FBij,t-1 0.052
(0.040) 0.148***
(0.039) -0.096* (0.056)
Past cumulative tweets: TWij,t-1 -0.003 (0.054)
0.093** (0.045)
-0.096 (0.070)
Include linear and quadratic time trends? Yes Yes –
Deal fixed effects Yes Yes –
Hour fixed effects Yes Yes –
Number of observations 2255 2016 –
Number of clusters 107 96 –
R² 0.73 0.78 –
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
MIS Quarterly Vol. 42 No. 4/December 2018 1345
Li & Wu/Herding and Social Media Word-of-Mouth
Table 10. Robustness Checks Addressing Serial Correlation
(1) (2) (3) (4)
Past cumulative sales: Yij,t-1 0.202***
(0.020) 0.243***
(0.020) 0.207***
(0.020) 0.242***
(0.020)
Past cumulative Facebook likes: FBij,t-1 0.121***
(0.016) 0.082***
(0.016) 0.113***
(0.016) 0.083***
(0.016)
Past cumulative tweets: TWij,t-1 0.015
(0.024) -0.023 (0.024)
-0.013 (0.025)
-0.028 (0.024)
Yij,t-1 × FBij,t-1 0.036*** (0.004)
0.034*** (0.004)
Yij,t-1 × TWij,t-1 0.041*** (0.008)
0.009 (0.008)
Estimated ρ 0.30 0.29 0.30 0.29 Include linear and quadratic time trends? Yes Yes Yes Yes Deal fixed effects Yes Yes Yes Yes Hour fixed effects Yes Yes Yes Yes Number of observations 9318 9318 9318 9318 Number of clusters 461 461 461 461 R² 0.65 0.70 0.66 0.70
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
qualitatively similar: both past cumulative sales and Face- book likes have positive effects on the next-period sales while tweets do not. The interaction effect between past cumulative sales and Facebook likes is also positive and statistically significant, suggesting they remain complements using the fixed-effect AR(1) model estimation.
We also estimate the fixed-effect AR(1) model for experience and for search goods (Table 11). Overall, we find that the effect of past cumulative sales is stronger for experience goods than for search goods, similar to the earlier results in Tables 7–9. Therefore, Tables 10 and 11 suggest that serial correlation is not a serious concern.
Tipped Deals
At the time of the data collection, deep discount in a Groupon deal is only applied after meeting a predetermined threshold. When it is not met, the deal is off and buyers are not charged. Thus, the pressure to tip a deal over the threshold by buying into it, especially as the threshold is approached as past sales mount, could also explain why past cumulative sales can affect the next-period sales. However, most deals were typically tipped around 7:30 a.m. and 90% of deals were tipped before 10 a.m., suggesting the threshold effect would not apply most of the time. To ensure sales motivated by attempts to tip the deal are not driving our results, we only analyze post-tipped sales, as shown in Table 12. The results are similar to earlier results.
Because the average number of Facebook likes per hour after deals are tipped is 3.06 and the estimated coefficient of the interaction term is 0.014, having the lowest number of Facebook likes would decrease the effect of past sales from 0.152 to 0.109 (=0.152 – 3.06 × 0.014), accounting for a reduction of 28%. Compared to Table 4, the interaction effect after deals are tipped is not only more economically signi- ficant but greater in the size of the estimated effects than earlier results when all deals are included.
Alternative Explanations for What Drives Herding
Although we control for deal-specific time-invariant hetero- geneity, hourly time trends, and nonlinear shapes of product diffusion, we further rule out several alternative explanations that could expose biases in our estimates. Social pressure would have been a concern if adoption choices are observable and verifiable (e.g., fashion items). However, the products in our context are largely personal experience/service goods (such as restaurant meals, spas, massages, and cleaning services), occurring in private settings not likely noticed by friends or peers. Consuming such products is less observable or verifiable, and thus social pressure is less likely to affect quality assessment.16
16For vouchers that friends may want to use together, like paintball/spa, social pressure or a network effect may be relevant. While we acknowledge that it is difficult to completely eliminate this possible confounder in this study, the mechanism checks reported earlier help alleviate this concern.
1346 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
Table 11. Differential Effects for Experience and Search Goods Addressing Serial Correlation
Human’s Evaluation Yelp Review Rating Product Price
(1) Experience
(2) Search
(3) Experience
(4) Search
(5) High-priced
(6) Low-priced
Past cumulative sales: Yij,t-1 0.310***
(0.039) 0.105** (0.041)
0.201*** (0.046)
0.042 (0.056)
0.313*** (0.034)
0.060 (0.049)
Past cumulative Facebook likes: FBij,t-1 0.110***
(0.031) 0.175*** (0.033)
0.178*** (0.039)
0.134*** (0.035)
0.069** (0.032)
0.176*** (0.033)
Past cumulative tweets: TWij,t-1 0.023
(0.063) 0.032 (0.051)
-0.055 (0.091)
-0.074 (0.050)
-0.025 (0.059)
0.106** (0.044)
Include linear and quadratic time trends?
Yes Yes Yes Yes Yes Yes
Deal fixed effects Yes Yes Yes Yes Yes Yes
Hour fixed effects Yes Yes Yes Yes Yes Yes
Number of observations 2266 2334 1630 1567 2148 1920
Number of clusters 113 117 77 81 105 96
R² 0.72 0.59 0.65 0.64 0.62 0.69
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
Table 12. Results on Data After Deals Are Tipped
(1) (2) (3) (4)
Past cumulative sales: Yij,t-1 0.152***
(0.043) 0.152***
(0.042) 0.152***
(0.043) 0.152***
(0.042)
Past cumulative Facebook likes: FBij,t-1 0.065**
(0.028) 0.052*
(0.030) 0.065**
(0.028) 0.050*
(0.030)
Past cumulative tweets: TWij,t-1 0.004
(0.028) -0.009 (0.028)
-0.008 (0.038)
0.003 (0.038)
Yij,t-1 × FBij,t-1 0.013*
(0.008) 0.014*
(0.008)
Yij,t-1 × TWij,t-1 0.007
(0.014) -0.008 (0.014)
Include linear and quadratic time trends? Yes Yes Yes Yes
Deal fixed effects Yes Yes Yes Yes
Hour fixed effects Yes Yes Yes Yes
Number of observations 7159 7159 7159 7159
Number of clusters 456 456 456 456
R² 0.39 0.40 0.39 0.41
Notes: *p < 0.10, **p < 0.05, ***p < 0.01
The positioning or saliency effect can also create biases. Consumers often choose the prominently displayed products (Cai et al. 2009), especially when the entire choice set is not observable. Instead of using cumulative sales to infer product quality, consumers may choose a product simply because of its saliency (Cai et al. 2009). For instance, when CNET.com prominently displayed the most popular software by sorting
products according to the number of downloads, then saliency, as opposed to uncertainty reduction, explains why the most popular software becomes even more popular over time. To avoid saliency-related confounding factors, we pur- posefully collect only featured deals, which are always placed in the same location (thus occurring the same position) on Groupon’s webpage.
MIS Quarterly Vol. 42 No. 4/December 2018 1347
Li & Wu/Herding and Social Media Word-of-Mouth
Overcrowding can also inhibit herding. Facing a capacity constraint, high existing sales may indicate overcrowding, which can reduce product quality, especially when the expiration date for the deal is imminent (Eyster et al. 2014; Horton 2014). Providing past sales information may also fail to trigger a herding response when consumer preferences are dissimilar to those of typical mass-market consumers. In that instance, having high sales would work in reverse and dis- courage niche consumers. However, these scenarios would in fact reduce our power to find the positive effect of showing past sales.
Alternative Explanations for Facebook-Mediated WOM
Online or offline promotions (such as TV advertising) could simultaneously affect sales and social media endorsements (Facebook likes), and accordingly they could bias the estima- tion for Facebook likes. Although our fine-grained measure- ment at the hourly level can substantially reduce the omitted variable biases (Bakshy and Eckles 2013; Bakshy et al. 2011), we address these potential biases in several ways. First, our baseline model includes hourly dummies that control for tem- poral shocks, such as other online or offline promotions that have occurred during a given time of the day. Second, we model the product diffusion process by including city-specific linear and quadratic time trends that control for the nonlinear shape in product diffusion (Carare 2012; Duan et al. 2009) for each local market as well as any other unobserved factors that are nonlinear in time. Third, we include the cumulative num- ber of tweets as a time-varying control. Because offline promotions and other related unobservable confounds are correlated with an increase in social media activities, including the number of tweets in the model could potentially control for these confounds, especially if the correlation is strong. If the number of Facebook likes continues to have a positive effect on sales after controlling for tweets, it is likely that the effect comes from Facebook activities as opposed to unobservable offline promotions. Furthermore, we have instrumented Facebook likes using the internal lags in the system GMM framework, which is well-suited to handle endogenous regressors in panel data. The estimation using external instrumental variables based on a product sold in the same city on the next day also does not change our main results. These robustness checks using different specifica- tions enhance the confidence that our results do not depend on a single method. Finally, we exploit the complementarities framework to address the potential endogeneity in social media WOM. While it is easy to believe that both social media WOM and past sales are endogenous, it is more difficult to explain how using both strategies yields higher performance than one without the other (as well as higher performance when neither is used). This effect is further
enhanced when the positive interaction is only with one type of social media but not the other.
Implications and Discussion
This study yields several theoretical and managerial impli- cations. Due to limited data availability, prior empirical research measures the effectiveness of either the strategies that facilitate herding or the strategies that promote traditional online WOM (e.g., online reviews about products), but rarely is social media WOM examined in conjunction with herding. As social media’s popularity increases, this study provides empirical evidence that online WOM via social media can have both a direct effect on product sales and a positive interaction effect when implemented with other marketing strategies. In particular, we find social media WOM can complement the strategy of showing past sales to further increase sales.
By exploiting the differences between experience goods and search goods, we find the underlying mechanisms behind the observed complementarities. We find that showing past sales can primarily reduce quality uncertainty, while social media WOM can primarily increase product awareness. The com- plementarities come from having social media to generate product awareness to many potential customers and having past sales information to reduce quality uncertainty for these customers. Together, they can rapidly generate greater sales than could either strategy on its own, a finding that can be useful for firms to reduce inventory, especially if imple- mented early in the sales cycle. Furthermore, the economic significance for the complementarities is substantial, especially as the effects compound over time.
Our full specification model (e.g., equation 3) includes prod- uct fixed effects which control for deal-specific heterogeneity that are invariant over time, such as voucher price and unobservable product quality. Our specification model also includes hour fixed effects which control for hourly variations over the first day of the deals, such as consumers’ Internet use as well as Groupon’s cyclical hourly marketing activities. In addition, the full specification model controls the city-specific linear and quadratic time trends which can address effects from product diffusion and possible market saturation in each local market. Finally, the number of tweets per hour is used to control for other potential social media-related confounding factors that are sufficiently correlated with tweets. However, we acknowledge that there still may be unobserved time- varying confounders that we cannot control for due to the lack of randomized experiments, and thus our estimation may still suffer from some omitted variable bias. To address this limitation, we use the system GMM, which relies on the lags of the regressors as internal IVs, and 2SLS estimation, which
1348 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
relies on external IVs. These results are similar to the earlier results. This is consistent with the earlier findings that show an 80% reduction in error (from values generated in ran- domized experiments) when fine-grained measurements with high dimensionality are used (Bakshy and Eckles 2013; Bakshy et al. 2011). Thus, these fine-grained measurements can provide reasonable inferences in settings where ran- domized experiments are not possible. Nevertheless, we acknowledge that our study, based on archival data, is limited in causal identification. Future research could use experi- mental methods to verify our findings and produce a higher degree of causal inference.
Although these findings suggest that the strategy of showing past sales is beneficial and it is easy to implement technically, they can also incur substantial indirect costs of revealing sensitive financial information about the products and the sellers’ true revenues. Perhaps due to this concern, beginning in September 2011, Groupon.com stopped displaying precise sales volume. While revealing it is undesirable in some cases, online retailers could still consider several alternative stra- tegies that can facilitate herding in similar ways. First, online retailers can consider using the stepwise approximation of sales volume. Groupon has since changed from showing the exact sales to showing a bucket system, indicated as “over 100 bought,” “over 500 bought,” etc. Second, online retailers can consider showing sales rank. Amazon has long displayed the sales information on each product’s webpage, indicated as “Amazon Best Sellers Rank.” Sales rank can be further refined for a specific product category or a specific time period. Finally, online retailers can consider developing a normalized sales index. Google Trends only discloses the search volume index of a search phrase that shows the relative intensity of the search without revealing the precise search volume. Online retailers can similarly develop a normalized sales index. These alternative strategies could allow online retailers to take advantage of showing a proxy for sales to generate herding while preserving sensitive financial information.
This study finds that neither the direct nor the interaction effect existed for Twitter, suggesting that differences in the platform might account for the absence. Ultimately, the effectiveness of a social media platform depends on many specific contextual factors, including how consumers interact with the platform. For example, Facebook posts contain a picture of the deals, whereas tweets do not. Tweets may also be more transient relative to Facebook posts in part because Facebook users and Twitter users are different in terms of their demographics and usage patterns.17 Facebook friends mostly have mutual social ties (two-way links), whereas
Twitter followers often have just one-way connections with the people they follow.18 These differences may explain why Facebook-mediated WOM has significant effects in our setting and Twitter-mediated WOM does not. Thus it is important for marketing managers to understand the specific characteristics of a social media platform to maximize the effectiveness of the associated WOM especially when these effects could be amplified due to their complementary relationship with herding strategies.
Acknowledgments
We are grateful for funding support from the HEC Foundation, ANR-11-IDEX-003/Labex ECODEC (No. ANR-11-LABX-0047), and the Wharton Dean’s Research Fund.
References
Akerlof, G. A., and Kranton, R. E. 2000. “Economics and Iden- tity,” The Quarterly Journal of Economics (115:3), pp. 715-753.
Aral, S., Brynjolfsson, E., and Van Alstyne, M. 2012. “Informa- tion, Technology, and Information Worker Productivity,” Information Systems Research (23:3), pp. 849-867.
Aral, S., and Walker, D. 2011. “Creating Social Contagion through Viral Product Design: A Randomized Trial of Peer Influence in Networks,” Management Science (57:9), pp. 1623-1639.
Arellano, M., and Bover, O. 1995. “Another Look at the Instru- mental Variable Estimation of Error-Components Models,” Journal of Econometrics (68:1), pp. 29-51.
Bakshy, E., and Eckles, D. 2013. “Uncertainty in Online Experi- ments with Dependent Data: An Evaluation of Bootstrap Methods,” in Proceedings of the 19th ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining, New York: ACM Press, pp. 1303-1311.
Bakshy, E., Hofman, J. M., Mason, W. A., and Watts, D. J. 2011. “Everyone's an Influencer: Quantifying Influence on Twitter,” in Proceedings of the 4th ACM International Conference on Web Search and Data Mining, New York: ACM Press, pp. 65-74.
Banerjee, A. V. 1992. “A Simple Model of Herd Behavior,” The Quarterly Journal of Economics (107:3), pp. 797-817.
Bardhan, I., Krishnan, V., and Lin, S. 2013. “Research Note: Business Value of Information Technology: Testing the Inter- action Effect of It and R&D on Tobin’s Q,” Information Systems Research (24:4), pp. 1147-1161.
Berger, J., and Heath, C. 2007. “Where Consumers Diverge from Others: Identity Signaling and Product Domains,” Journal of Consumer Research (34:2), pp. 121-134.
Bikhchandani, S., Hirshleifer, D., and Welch, I. 1992. “A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades,” Journal of Political Economy (100:5), pp. 992-1026.
17https://www.bloomberg.com/gadfly/articles/2016-02-12/social-studies- comparing-twitter-with-facebook-in-charts. 18http://www.diffen.com/difference/Facebook_vs_Twitter.
MIS Quarterly Vol. 42 No. 4/December 2018 1349
Li & Wu/Herding and Social Media Word-of-Mouth
Blundell, R., and Bond, S. 1998. “Initial Conditions and Moment Restrictions in Dynamic Panel Data Models,” Journal of Econometrics (87:1), pp. 115-143.
Bresnahan, T. F., Brynjolfsson, E., and Hitt, L. M. 2002. “Informa- tion Technology, Workplace Organization, and the Demand for Skilled Labor: Firm-Level Evidence,” The Quarterly Journal of Economics (117:1), pp. 339-376.
Brynjolfsson, E., and Milgrom, P. 2013. “Complementarity in Organizations,” The Handbook of Organizational Economics, R. Gibbons and J. Roberts (eds.), Princeton, NJ: Princeton Univer- sity Press, pp. 11-55.
Burtch, G., Ghose, A., and Wattal, S. 2013. “An Empirical Exam- ination of the Antecedents and Consequences of Contribution Patterns in Crowd-Funded Markets,” Information Systems Research (24:3), pp. 499-519.
Cai, H., Chen, Y., and Fang, H. 2009. “Observational Learning: Evidence from a Randomized Natural Field Experiment,” American Economics Review (99:3), pp. 864-882.
Carare, O. 2012. “The Impact of Bestseller Rank on Demand: Evidence from the App Market,” International Economic Review (53:3), pp. 717-742.
Caves, R. E., and Williamson, P. J. 1985. “What Is Product Differ- entiation, Really?,” The Journal of Industrial Economics (34:2), pp. 113-132.
Chen, H., De, P., and Hu, Y. J. 2015. “IT-Enabled Broadcasting in Social Media: An Empirical Study of Artists’ Activities and Music Sales,” Information Systems Research (26:3), pp. 513-531.
Chen, Y., Wang, Q., and Xie, J. 2011. “Online Social Interactions: A Natural Experiment on Word of Mouth Versus Observational Learning,” Journal of Marketing Research (48:2), pp. 238-254.
Chevalier, J. A., and Mayzlin, D. 2006. “The Effect of Word of Mouth on Sales: Online Book Reviews,” Journal of Marketing Research (43:3), pp. 345-354.
Chintagunta, P. K., Gopinath, S., and Venkataraman, S. 2010. “The Effects of Online User Reviews on Movie Box Office Perfor- mance: Accounting for Sequential Rollout and Aggregation across Local Markets,” Marketing Science (29:5), pp. 944-957.
Cronbach, L. J. 1951. “Coefficient Alpha and the Internal Structure of Tests,” Psychometrika (16:3), pp. 297-334.
Dellarocas, C. 2003. “The Digitization of Word of Mouth: Promise and Challenges of Online Feedback Mechanisms,” Management Science (49:10), pp. 1407-1424.
Dellarocas, C., Zhang, X. M., and Awad, N. F. 2007. “Exploring the Value of Online Product Reviews in Forecasting Sales: The Case of Motion Pictures,” Journal of Interactive Marketing (21:4), pp. 23-45.
Dholakia, U. 2011. “How Businesses Fare with Daily Deals: A Multi-Site Analysis of Groupon, Livingsocial, Opentable, Travelzoo, and Buywithme Promotions,” available at SSRN: http://ssrn.com/abstract=1863466.
Duan, W., Gu, B., and Whinston, A. B. 2009. “Informational Cascades and Software Adoption on the Internet: An Empirical Investigation,” MIS Quarterly (33:1), pp. 23-48.
Eyster, E., Galeotti, A., Kartik, N., and Rabin, M. 2014. “Con- gested Observational Learning,” Games and Economic Behavior (87), pp. 519-538.
Firat, A., Madnick, S., and Siegel, M. 2000. “The Cameleon Web Wrapper Engine,” in Proceedings of the VLDB2000 Workshop on Technologies for E-Services, Cairo, Egypt. pp. 1-9.
Fleder, D., and Hosanagar, K. 2009. “Blockbuster Culture's Next Rise or Fall: The Impact of Recommender Systems on Sales Diversity,” Management Science (55:5), pp. 697-712.
Godes, D., and Mayzlin, D. 2004. “Using Online Conversations to Study Word-of-Mouth Communication,” Marketing Science (23:4), pp. 545-560.
Godes, D., and Mayzlin, D. 2009. “Firm-Created Word-of-Mouth Communication: Evidence from a Field Test,” Marketing Science (28:4), pp. 721-739.
Herzenstein, M., Dholakia, U. M., and Andrews, R. L. 2011. “Strategic Herding Behavior in Peer-to-Peer Loan Auctions,” Journal of Interactive Marketing (25:1), pp. 27-36.
Hong, Y., Chen, P.-Y., and Hitt, L. M. 2014. “Measuring Product Type with Dynamics of Online Product Review Variances: A Theoretical Model and the Empirical Applications,” NET Institute Working Paper 14-03.
Hong, Y., and Pavlou, P. A. 2014. “Product Fit Uncertainty in Online Markets: Nature, Effects, and Antecedents,” Information Systems Research (25:2), pp. 328-344.
Horton, J. J. 2014. “Misdirected Search Effort in a Matching Market: Causes, Consequences and a Partial Solution,” in Proceedings of the 15th ACM Conference on Economics and Computation, New York: ACM Press, pp. 357-357.
Hosanagar, K., Fleder, D., Lee, D., and Buja, A. 2014. “Will the Global Village Fracture into Tribes? Recommender Systems and Their Effects on Consumer Fragmentation,” Management Science (60:4), pp. 805-823.
Kirmani, A., and Rao, A. R. 2000. “No Pain, No Gain: A Critical Review of the Literature on Signaling Unobservable Product Quality,” Journal of Marketing (64:2), pp. 66-79.
Krippendorff, K. 2004. “Reliability in Content Analysis,” Human Communication Research (30:3), pp. 411-433.
Kumar, V., and Rajan, B. 2012. “Social Coupons as a Marketing Strategy: A Multifaceted Perspective,” Journal of the Academy of Marketing Science (40:1), pp. 120-136.
Li, X. 2016. “Could Deal Promotion Improve Merchants’ Online Reputations? The Moderating Role of Prior Reviews,” Journal of Management Information Systems (33:1), pp. 171-201.
Li, X. 2018. “Impact of Average Rating on Social Media Endorse- ment: The Moderating Role of Rating Dispersion and Discount Threshold,” Information Systems Research (Forthcoming).
Li, X., and Wu, L. 2013. “Measuring Effects of Observational Learning and Social-Network Word-of-Mouth (WOM) on the Sales of Daily-Deal Vouchers,” in Proceedings of the 46th Hawaii International Conference on System Sciences, pp. 2908-2917.
Liu, Y. 2006. “Word-of-Mouth for Movies: Its Dynamics and Impact on Box Office Revenue,” Journal of Marketing (70:3), pp. 74-89.
Lobel, I., Sadler, E., and Varshney, L. R. 2017. “Customer Referral Incentives and Social Media,” Management Science (63:10), pp. 3514-3529.
Luo, X., Andrews, M., Song, Y., and Aspara, J. 2014. “Group- Buying Deal Popularity,” Journal of Marketing (78:2), pp. 20-33.
1350 MIS Quarterly Vol. 42 No. 4/December 2018
Li & Wu/Herding and Social Media Word-of-Mouth
McPherson, M., Smith-Lovin, L., and Cook, J. M. 2001. “Birds of a Feather: Homophily in Social Networks,” Annual Review of Sociology (27:1), pp. 415-444.
Muchnik, L., Aral, S., and Taylor, S. J. 2013. “Social Influence Bias: A Randomized Experiment,” Science (341:6146), pp. 647-651.
Nelson, P. 1974. “Advertising as Information,” The Journal of Political Economy (82:4), pp. 729-754.
Nelson, P. J. 1981. “Consumer Information and Advertising,” in Economics of Information. Springer, pp. 42-77.
Nisbett, R. E., and Ross, L. 1980. Human Inference: Strategies and Shortcomings of Social Judgment, Englewood Cliffs, NJ: Prentice-Hall.
Roodman, D. 2009a. “How to Do Xtabond2: An Introduction to Difference and System Gmm in Stata,” The Stata Journal (9:1), pp. 86-136.
Roodman, D. 2009b. “A Note on the Theme of Too Many Instruments,” Oxford Bulletin of Economics and Statistics (71:1), pp. 135-158.
Smith, D. B. 1990. “The Economics of Information: An Empirical Approach to Nelson's Search-Experience Framework,” Journal of Public Policy & Marketing (9), pp. 111-128.
Trusov, M., Bucklin, R. E., and Pauwels, K. 2009. “Effects of Word-of-Mouth Versus Traditional Marketing: Findings from an Internet Social Networking Site,” Journal of Marketing (73:5), pp. 90-102.
Tucker, C., and Zhang, J. 2011. “How Does Popularity Information Affect Choices? A Field Experiment,” Management Science (57:5), pp. 828-842.
Weathers, D., Sharma, S., and Wood, S. L. 2007. “Effects of On- line Communication Practices on Consumer Perceptions of Performance Uncertainty for Search and Experience Goods,” Journal of Retailing (83:4), pp. 393-401.
Wojnicki, A., and Godes, D. 2008. “Word-of-Mouth as Self- Enhancement,” Harvard Business School Marketing Research Paper No. 06-01.
Wu, J., Shi, M., and Hu, M. 2015. “Threshold Effects in Online Group Buying,” Management Science (61:9), pp. 2025-2040.
Xiao, B., and Benbasat, I. 2007. “E-Commerce Product Recom- mendation Agents: Use, Characteristics, and Impact,” MIS Quarterly (31:1), pp. 137-209.
Young, H. P. 2009. “Innovation Diffusion in Heterogeneous Populations: Contagion, Social Influence, and Social Learning,” The American Economic Review (99:5), pp. 1899-1924.
Zhang, J. 2010. “The Sound of Silence: Observational Learning in the Us Kidney Market,” Marketing Science (29:2), pp. 315-335.
Zhang, J., and Liu, P. 2012. “Rational Herding in Microloan Markets,” Management Science (58:5), pp. 892-912.
About the Authors
Xitong Li is an assistant professor in the Department of Information Systems and Operations Management, HEC Paris, France. His research interests include the economic and social impacts of using online data/information, and innovative technologies using online data and services. His research has appeared, or is forthcoming, in Information Systems Research, MIS Quarterly, Journal of Manage- ment Information Systems, ACM Transactions on Internet Tech- nology, ACM Transactions on Multimedia Computing Communica- tions and Applications, IEEE Transactions on Engineering Management, IEEE Transactions on Automation Science and Engineering, IEEE Transactions on Systems, Man & Cybernetics, IEEE Communications Magazine, and other leading international journals. He was a finalist for the OCIS Best Paper Award at the 77th Academy of Management Annual Meeting in 2017 and won the Best Paper Award at the 46th Hawaii International Conference on System Sciences in 2013. He received his Ph.D. in management from the MIT Sloan School of Management and his Ph.D. in engineering from Tsinghua University.
Lynn Wu is an assistant professor at the Wharton School, Univer- sity of Pennsylvania. She studies how emerging information tech- nologies, such as enterprise social media and big data analytics, affect productivity, innovation, and organization practices. Her research has been published in Information Systems Research, Management Science, and MIS Quarterly. She was a finalist for the Best Paper Award at the Conference on Information Systems and Technology in 2017, and received the Best Published Paper award at ISS/Information System Research in 2014, the Best Publication award at the Americas Conference on Information Systems in 2014, the Best Paper Award at the 46th Hawaii International Conference on System Sciences in 2013, and the Best Paper Award at the Inter- national Conference on Information Systems in 2008. She received her Ph.D. in management from the MIT Sloan School of Management.
MIS Quarterly Vol. 42 No. 4/December 2018 1351
Copyright of MIS Quarterly is the property of MIS Quarterly and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.