Topic: Downfall of Brick and Mortar stores

RESEARCH ARTICLE

THE DEMAND EFFECTS OF PRODUCT RECOMMENDATION NETWORKS: AN EMPIRICAL ANALYSIS OF NETWORK

DIVERSITY AND STABILITY1

Zhijie Lin School of Business, Nanjing University, 22 Hankou Road,

Nanjing 210093, CHINA {[email protected]}

Khim-Yong Goh and Cheng-Suang Heng School of Computing, National University of Singapore, 15 Computing Drive,

Singapore 117418 SINGAPORE {[email protected]} {[email protected]}

With the increasing popularity of product recommendation networks in e-commerce, researchers and practi- tioners are eager to understand how they can strategically manage product assortments through the manipu- lation of such networks to drive demand. We examine product recommendation networks in e-commerce to investigate how the demand of a product is influenced by product network attributes in terms of network diversity and network stability. We also examine whether the demand of a product is influenced by both the incoming network and the outgoing network, and if the effects differ between co-view and co-purchase recommendation networks. Using data from Tmall.com for four product categories, we apply linear panel data models to examine the impact of network diversity and network stability on product demand, controlling for relevant factors at the individual product, pricing, product network, product category, and time unit levels. Importantly, we account for implicit demand correlation (i.e., substitution and complementarity) and potential simultaneity of demand and network structures. We unravel several important findings. First, a 1% increase in the category diversity of the incoming (outgoing) co-purchase network of a product is associated with a 0.011% (0.012%) increase (decrease) in the product’s demand. Second, a 1% increase in the stability of the outgoing co-purchase network is associated with a 0.012% decrease in demand. Third, the demand effects of network diversity and stability are both stronger in the co-purchase network, compared to their insignificant effects in the co-view network. Thus, this research provides theoretical contributions in terms of the economic effects of product recommendation networks through its focus on network diversity and stability in incoming/ outgoing and co-view/co-purchase networks. We also provide notable implications for recommendation-based product marketing and recommendation systems design.

1

Keywords: Recommendation system, product network, electronic commerce, network diversity, network stability, demand modeling, econometric analysis

1Sulin Ba was the accepting senior editor for this paper. Gal Oestreicher-Singer served as the associate editor.

The appendices for this paper are located in the “Online Supplements” section of the MIS Quarterly’s website (http://www.misq.org).

MIS Quarterly Vol. 41 No. 2, pp. 397-426/June 2017 397

Lin et al./Demand Effects of Product Recommendation Networks

Introduction

Electronic commerce (e-commerce) has burgeoned tremen- dously over the years. According to eMarketer, global e-commerce sales will rapidly increase from $1.316 trillion in 2014 to $2.489 trillion in 2018 (eMarketer 2014). Capitalizing on its popularity, online retailers have attempted to replicate the “diaper and beer” colocation practice in grocery stores (Srikant and Agrawal 1996). Specifically, on most e-commerce sites, each product is featured on its own designated webpage, but on each of these product pages, retailers often utilize some recommender systems2 to explicitly recommend additional relevant products, which might be of interest to consumers, either to help them find the most suitable products or to cross-sell other goods. Hence, a visible directed product network is created whereby products (i.e., network nodes) are explicitly connected by hyperlinks (i.e., network ties). If one analogizes the process of browsing an e-commerce website to walking the aisles of a physical store, the “aisle structure” of the e-commerce website will be defined by this graph of interconnected products, whereas recommending additional products on a webpage is likened to placing additional products on a neighboring shelf. Perhaps the best-known examples of recommender systems are the co- view and co-purchase recommendations on Amazon.com, where recommended products are listed under the titles “Customers who viewed this item also viewed” and “Customers who bought this item also bought” respectively. Figure 1 illustrates how a directed product network is formed. Similar recommender systems have also been adopted by various e-commerce websites (e.g., Alibaba.com, Walmart. com, and Tmall.com). This phenomenon has generated research investigating the economic impact (e.g., product sales impact) of such systems, especially from the network perspective (Carmi et al. 2017; Oestreicher-Singer et al. 2013; Oestreicher-Singer and Sundararajan 2012a, 2012b). However, critical research gaps, which motivate our study, still prevail.

First, our study aims to fill the research gap in existing prod- uct network literature by underscoring the demand effects of network diversity.3 Driven by marketing theories on product assortment and variety (e.g., Baumol and Ide 1956; Hoch et al. 1999; Ton and Raman 2010) and theories on consumer

heterogeneity (e.g., Allenby et al. 1998), we conjecture a critical role of network diversity. Prior studies have docu- mented that consumers seek product variety because they are more likely to find an ideal match with greater varieties and larger assortments (Baumol and Ide 1956). When consumers’ preferences are uncertain, variety offers them option value (Reibstein et al. 1975) and, in general, consumers possess intrinsic motivation for variety seeking (Van Trijp et al. 1996). Likewise, marketing theories on consumer hetero- geneity have also reiterated that consumers exhibit hetero- geneity in terms of their preferences for product categories (Allenby et al. 1998; Kamakura et al. 1996; Osborne 2011). Thus, a more diverse set of categories could increase the chance of meeting a consumer’s purchase or consumption needs. In essence, both theoretical angles have converged on the potentially important role of network diversity, to suggest that the offering of a diverse set of product recommendations may impact consumer online product search experiences and eventually influence their purchase decisions.

In addition, we make a subtle but important conceptual distinction between network diversity and assortative mixing (Oestreicher-Singer and Sundararajan 2012a, 2012b). Al- though a few studies have adopted the variable of assortative mixing to capture a relative aspect of product category similarity in a network and investigated its impact on a focal product’s demand, our network diversity variable is substan- tially different from assortative mixing. Conceptually, assor- tative mixing measures the relative extent of product category similarity between that of the focal product and its network neighbors, whereas network diversity captures the absolute extent of differences among the focal product’s network neighbors per se. Of greater importance, network diversity is more relevant to our context as compared to assortative mixing because assortative mixing only reflects the category similarity between a focal product and its neighbors, but does not reveal or suggest a categorically diverse product set to consumers.4 Importantly, the demand impact of network diversity, beyond assortative mixing, remains unknown.5

2In this study, we are interested in online product networks that are created by the use of recommender systems. Thus, we use the terms recommender system, recommendation system, recommendation network, and product network interchangeably.

3We explore network diversity through our focus on the diversity of product categories in the network.

4To illustrate, assume the focal product is in category A, and the five products in the recommendation network are in categories: (1) A, A, B, B and B, or (2) A, A, B, C and D. Assortative mixing does not differentiate between these two cases but only shows that the category similarity is 2/5 = 0.4 (i.e., category difference is 3/5 = 0.6), whereas our measure of network diversity can differentiate between the two cases by revealing there are two categories in the first case but four categories in the second case.

5For a robustness check, we also included assortative mixing variables in our empirical model (see Table 9, Column (8)) to demonstrate the distinct impacts of network diversity from assortative mixing on product demand. Notably, the estimated model coefficients of network diversity and stability remain consistent with the inclusion of assortative mixing variables.

398 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Figure 1. Amazon Product Recommendation Network

Second, we seek to enrich the product network literature by studying the unexplored demand effects of network stability.6 As mentioned above, product variety may enhance consumer product search experiences, but offering a large variety is often costly (Borle et al. 2005). The literature on product assortment has suggested strategies such as product assort- ment change or assortment rotation (Caro and Martínez-de- Albéeniz 2009), but its potential impact on product demand is ambivalent. On one hand, retailers can change the assort- ment (i.e., replace old products with new ones) so as to inject an air of freshness and novelty to consumers. For instance, the H&M annual report (H&M 2007) informs that,

New items every day make the stores interesting and lively. Having a number of collections each season also means that the store changes its appearance often. The aim is that customers should always be able to find something new and exciting.

On the other hand, the literature on product assortment reduc- tion also suggests that consumers may be averse to changes in assortment because they may have difficulties in finding their preferred items (Broniarczyk et al. 1998). The constant changes in recommendation updates and network connections would trigger low product network stability. Moreover, psychologically, consumers are likely to experience a higher level of perceived product search efforts when they are faced

with constant changes in product assortment (Shugan 1980), thereby potentially affecting their eventual purchases. Given the potential contention between these two camps of thought, elucidating the effects of network stability is critical to providing an in-depth theoretical understanding and valuable practical implications.

Third, past research on product recommendation networks has overlooked the importance of distinguishing and comparing the impacts of two types of recommendation networks: co- view and co-purchase recommendation networks. Although co-view and co-purchase recommendation systems have coexisted on most e-commerce sites (e.g., Amazon.com, Tmall.com) to offer information on additional related products to assist consumers in their purchase decisions, there exists a fundamental difference between them. Co-view recommen- dations communicate information about other consumers’ e-commerce website browsing behaviors across different products, while co-purchase recommendations communicate information about other consumers’ purchases of products bought typically in the same shopping session. Studies on consumer information seeking and processing have not only reported that consumers are in need of product information to form their quality and value perceptions to make purchase decisions (Nelson 1970; Punj and Staelin 1983), but also discovered that consumers may factor the sources or types of information into considerations (Goh et al. 2013; Rao and Sieben 1992; Westbrook and Fornell 1979). Given the prevalence of their coexistence in e-commerce and differences in information between co-view and co-purchase recommen- dations, it is surprising that their distinction has yet to receive

6We investigate network stability by focusing on the extent of updates of recommendations (analogously, product connections in the network).

MIS Quarterly Vol. 41 No. 2/June 2017 399

Lin et al./Demand Effects of Product Recommendation Networks

any attention from product network research, be it in infor- mation systems (Carmi et al. 2017; Oestreicher-Singer and Sundararajan 2012a) or in marketing (Oestreicher-Singer et al. 2013; Oestreicher-Singer and Sundararajan 2012b). Emphatically, our study aims to make such a contribution.

Fourth, our study hopes to achieve a more holistic and comprehensive understanding of product networks by exam- ining the effects of both incoming and outgoing networks on product demand. Although a paucity of studies have hinted at the role of network directionality, most prior studies unfor- tunately combined incoming and outgoing networks without differentiation to develop an integrated factor (e.g., PageRank) such that they were unable to separate the demand impacts of the two networks (Oestreicher-Singer et al. 2013; Oestreicher-Singer and Sundararajan 2012a). Other prior studies only compared some network metrics (separated into incoming and outgoing networks) between social and product networks on Youtube.com rather than estimating the econo- mic impacts of such networks (Goldenberg et al. 2012). Specifically, e-commerce product network links are directed and, at times, asymmetric. The existence of a recommen- dation link from product A to product B does not guarantee the availability of a reversed link from product B to product A. Hence, each focal product typically has a set of recom- mendation links initiating from its incoming network and another set of recommendation links connected to its outgoing network (see Figure 2 for an illustration). The direction of network links has played a central role in exerting network influences on some economic behaviors or outcomes. For instance, in the social network context, Hinz et al. (2014) claimed that it is the direction of social ties that determines the social influence of early adopters on prospective adopters’ product adoption decisions, and helps firms identify influ- ential consumers to target their marketing. In the product network context, Carmi et al. (2017) reported that the direc- tion of product links is important as it directs consumers’ attention and generates traffic to a product. Given the central role of network link directions, it is thus important to study the impacts of both incoming and outgoing networks for a more complete and comprehensive network analysis.

Based on the aforementioned research gaps, the objective of this research is threefold. First, we operationalize the product network in a more complete manner by observing both the incoming and outgoing networks of a focal product to inves- tigate and differentiate between the impacts of the incoming and outgoing networks. Second, we further separate the product network into two types of networks (co-view and co- purchase recommendation networks) to examine and differ- entiate between the impacts of these two different recom- mendation mechanisms. Third, based on the above opera- tionalizations, we propose and examine the economic impact

of two important network factors (i.e., network diversity and network stability). We aim to provide significant insights on the economic impact of product recommendation networks through these investigations. In essence, our research questions are:

(1) How is the demand of a product influenced by product network attributes in terms of network diversity and network stability?

(2) Is the demand of a product influenced by both the incoming network and outgoing network?

(3) How do the diversity and stability effects differ between two types of recommendation networks (co-view and co- purchase)?

To answer our research questions, we collect product recom- mendations and transactions data from four different online stores on Tmall.com selling digital cameras, personal com- puters, mobile phones, and cosmetic products. Our econo- metric specification models products’ daily demand as a function of network factors, controlling for relevant factors at the individual product, pricing, product network, product category, and time unit levels. In particular, our identification strategy for the economic impact of network structures is to identify and control for the implicit demand correlations (i.e., substitution and complementarity effects) and to account for potential simultaneity between network structures and product demand. We perform robustness checks to validate the con- sistency of our findings in the presence of numerous control variables, potential simultaneity, collinearity, serial correla- tion, price endogeneity, and across differences in variable operationalizations, time frames, product categories, and data samples.

We find several notable results. First, both a product’s incoming network and outgoing network have significant relationships with the product’s demand, albeit differently. Second, the category diversity in the incoming network (of co-purchases) is positively associated with the demand of the focal product, while that for the outgoing network (of co- purchases) exhibits a negative association. Third, the stability of the co-purchase outgoing network is negatively associated with product demand. Fourth, the co-purchase recommen- dation network has a stronger association with product demand than the co-view recommendation network for both network diversity and stability.

This research has important contributions. First, our study pioneers the quantification of the economic effects of both the incoming and outgoing product networks. Second, our research is a leading empirical effort to document both posi-

400 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Figure 2. Incoming Network and Outgoing Network

tive and negative economic effects of product recommen- dations on individual product demand.7 Third, our research serves as a novel attempt to propose and validate the network diversity and network stability effects of product recommen- dation networks. Fourth, we shed light on the differential effects of the co-view and co-purchase product networks. Finally, our study also provides implications for online retailers’ recommendation-based product marketing strategies and the design of recommendation systems.

Literature Review

Our study is most related to the stream of marketing literature that examines the impact of product assortment and variety on consumer decision making. Consumers care about product assortment and variety because these essentially determine whether consumers can find their ideal products (Baumol and Ide 1956). Consequently, firms try to offer assortment and variety, and strive to manipulate them to influence consumer purchases (Rooderkerk et al. 2013). Existing empirical studies have mainly examined the impacts of two major manipulation strategies by firms in the traditional retail con- text. First, firms strategize to influence consumer purchases or product sales by altering the size or variety of assortment, a move that has implications on the network diversity of product networks. Interestingly, prior studies examining this strategy have reported mixed results (e.g., Hoch et al. 1999; Kuksov and Villas-Boas 2010; Ton and Raman 2010). On

one hand, increasing the size or variety of assortment may increase the choice set and the probability that consumers will find what they want (Ton and Raman 2010). On the other hand, increasing the size or variety of assortment may result in higher product search and evaluation costs for consumers (Kuksov and Villas-Boas 2010), thus causing consumers to avoid making a choice. Second, firms strategize to influence consumer purchases or product sales through assortment rotations or updates (e.g., Caro and Martínez-de-Albéeniz 2009), a shift that has implications on the product network stability. In sum, however, both of these assortment strategies have hardly received any attention in the e-commerce context, hence warranting our scrutiny here through the lens of network diversity and stability.

The management of product assortment and variety is impor- tant not only in the traditional retail context, but also increasingly in e-commerce (Borle et al. 2005), especially when recommender systems are widely available for online retailers to adopt and help determine what products to explicitly offer to consumers. Recommender systems are a specific type of information filtering technique that auto- matically provides recommendations for items (e.g., music, book, or movie) that might be of interest to a consumer (Yi et al. 2011). According to Adomavicius and Tuzhilin (2005), recommendation algorithms are usually classified into three main categories: (1) content-based approach (recommenda- tions are generated based on the characteristics of an item), (2) collaborative filtering approach (recommendations are generated based on the consumer’s social environment), and (3) hybrid approach (combining the prior two methods). Recommender systems are increasingly used in various online communities (Sahoo et al. 2011). Particularly, they have been widely adopted by online retailers (e.g., Amazon.com, Tmall. com) to recommend related products to consumers when they are viewing or searching for a product (Huang et al. 2007).

7While Oestreicher-Singer and Sundararajan (2012a) reported that the network influence on a product category is associated with both increases and decreases in relative revenue of books depending on their popularity within the category, the authors did not conduct their analysis at the individual product level.

MIS Quarterly Vol. 41 No. 2/June 2017 401

Lin et al./Demand Effects of Product Recommendation Networks

The popularity of recommender systems has attracted some academic interest. Early research work, mostly in the data mining field, focused on developing and evaluating various recommendation algorithms (Aciar et al. 2007; Herlocker et al. 2004; Iaquinta et al. 2008). Among numerous algorithms, the collaborative filtering approach, which determines recom- mendations by the levels of similarity of preferences of other consumers, is the most popular in e-commerce settings (Schafer et al. 1999). Although significant efforts have been made to these system design related issues, there have been only a few studies investigating the economic impact (e.g., sales impact) of recommender systems. Specifically, Ander- son (2006) and Brynjolfsson et al. (2006) reported that recommendations help consumers discover new products and thus increase sales diversity, whereas Mooney and Roy (2000) argued that recommendations only reinforce the posi- tion of already-popular products and instead reduce diversity. To explain the existence of these two opposite anecdotal views, Fleder and Hosanagar (2009) found that it is possible for individual-consumer-level diversity to increase because recommendations can guide each individual consumer to new products, but aggregate-market-level diversity to decrease because recommendations often guide similar users toward the same products. In addition to the sales diversity impact of recommender systems, Pathak et al. (2010) also showed that recommendations can directly increase the sales of recom- mended products, and also their prices.

As recommendations form a visible product network, some researchers were particularly interested in the impact of recommender systems from the network perspective. How- ever, the limited attention given to product networks and the paucity of studies on it is surprising. The handful of studies on product networks analyze networks of World Wide Web sites (Katona and Sarvary 2008), blogs (Mayzlin and Yoganarasimhan 2012), news reports (Dellarocas et al. 2010), and videos (Goldenberg et al. 2012). However, these studies did not examine the impact on actual product demand. In the e-commerce contexts, there are some studies investigating the network impact on actual product demand that are more relevant to our research. However, these studies only focused on the co-purchase network, and did not differentiate the demand impacts of incoming and outgoing networks. For instance, Carmi et al. (2017) identified the spread of exo- genous demand shocks generated by book reviews featured on The Oprah Winfrey Show and published in The New York Times through the incoming co-purchase recommendation networks on Amazon.com. Oestreicher-Singer and Sunda- rarajan (2012a), in the spirit of the PageRank algorithm (Brin and Page 1998) which combined incoming and outgoing networks without differentiation to develop an integrated factor, associated the average influence of the network centrality on each book category with the inequality in the

distribution of its revenue on Amazon.com. Using similar data, Oestreicher-Singer and Sundararajan (2012b) focused only on the incoming network and showed that the explicit visibility of a co-purchase relationship could lead to a three- fold amplification of the influence that complementary products have on each other’s demand. Moreover, Oestreicher-Singer et al. (2013), again based on the PageRank algorithm, quantified the value of a product to the firm by decomposing the revenue of each product into the intrinsic value portion (i.e., self-generated by the product) and the extrinsic value portion (i.e., driven by the recommendation links pointing from other products to the focal product).

Although these product network studies have started exploring the impact of network centrality, the important roles of network diversity and network stability have not yet been examined. The only relevant studies are those in the field of social networks. Social network diversity refers to the diversity among a social member’s network neighbors in aspects such as gender, age, education, and work experience (Jehn et al. 1999). Most past studies focused on the asso- ciation between network diversity and work performance. They found that individuals who have a more diverse network are more productive than their peers (Reagans and Zuckerman 2001), would receive higher performance ratings and compen- sation (Cross and Cummings 2004), are more likely to be recognized as top performers (Burt 2000), and obtain more economic opportunities (Eagle et al. 2010). In addition to work performance, several studies also examined outcomes such as knowledge sharing (Cummings 2004), health condi- tion (Barefoot et al. 2005), and online content propagation (Yoganarasimhan 2012).

Additionally, the stability of an individual’s social network is typically defined as the extent of overlap of network neigh- bors or connections over time (Cummings and Higgins 2006). Ghose et al. (2011) studied the stability of an individual’s social network on individual behavior in a mobile Internet setting. They found that users with high network stability have a low intrinsic tendency to engage in content usage and generation on the mobile Internet. Moreover, Tucker (2011) examined the effects of instability in social networks on network externalities and on the rate of adoption of a video- calling system. She identified that the aggregate effect of network externalities on adoption is augmented by commu- nication network instability, due to individuals’ uncertainty of future communication networks.

In essence, our research differs from related prior studies by examining the impact of product networks through accounting for both incoming and outgoing networks, and differentiating between co-view and co-purchase networks to study the diversity effect and stability effect in product networks, so as to contribute to a unique comprehensive analysis.

402 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Theoretical Foundations

In this section, we discuss some relevant theoretical founda- tions for our research questions and objectives that can guide our subsequent empirical analyses in this paper. However, for this pioneering product network study examining the econo- mic effects of network diversity and stability in incoming/ outgoing and co-view/co-purchase networks, we opted to refrain from stating definitive research hypotheses because the current state of relevant theoretical development to support such hypotheses is inadequate (i.e., in terms of a convincing theory to support the demand effect of outgoing links vis-à-vis incoming ones). In essence, while we provide some theo- retical foundations to guide our data analyses, this paper aims to generate primarily empirical insights to shed light on our proposed research questions.

Network Diversity and Product Demand

Consumers exhibit heterogeneity in terms of their needs for products or categories (Allenby et al. 1998; Kamakura et al. 1996; Osborne 2011). Hence, in e-commerce contexts, consumers may search for different product categories (which can be substitutes or complements to the focal product) that suit their own needs. As such, if the category diversity of a focal product’s incoming network increases, this incoming network will bring forth a larger group of consumers or visitors. As incoming network products provide visible con- nections to the focal product, the focal product will be accessible by this larger group of potential consumers from the incoming network. Consequently, the exposure of the focal product will be increased and the demand may increase due to this heightened exposure (Carmi et al. 2017). More- over, past studies have reported that products that are connected in a network tend to have sales correlations. For instance, Oestreicher-Singer and Sundararajan (2012b) studied the network of books on Amazon.com and identified incremental correlations in book sales attributable to the product network’s visibility.

However, the impact of the outgoing network category diversity on the demand of the focal product may differ sub- stantially. When browsing the webpage of the focal product, consumers will be exposed to the associated recommendations (i.e., outgoing network products) on the same page. In the case of recommended substitute products in the outgoing links, this is likely to curtail the demand of the focal product. For recommended products that are complements to the focal one, increased category diversity in the outgoing network typically may also imply demand curtailment of the focal product. This is because increased exposure to multiple related categories of a product may lead consumers to seek

variety in purchases across categories in a single purchase instance (Menon and Kahn 1995; Van Trijp et al. 1996), rather than to fixate on the purchase of the focal product.

In addition, recommendations in the outgoing network may distract consumers’ attention, regardless of whether the out- going network products are substitutes or complements to the focal product, or even when they are irrelevant to consumers’ search goals. Based on the theory of stimulus complexity (Berlyne 1960), webpage complexity is defined as the level of diversity of information about the stimulus. This has been documented to distract consumers (Deng and Poole 2010; Nadkarni and Gupta 2007). In line with the arguments for demand curtailment and webpage complexity, the increase in the category diversity of the outgoing network is likely to shift consumers’ demand or attention away from the focal product, and consequently lower the chance of buying the focal product.

Network Stability and Product Demand

Besides the diversity of a focal product’s network, the stability of the network may also affect product demand. Various e-commerce studies have documented that most consumers do not purchase during their first visit to an online store (Johnson et al. 2003; Moe and Fader 2004), especially for the purchases of big ticket items such as digital cameras and personal computers. In other words, consumers often make more than one visit before deciding on the final purchase.8 Across multiple visits to an e-commerce website, consumers will be exposed to product assortment changes through the recommendation link updates, hence they will invariably notice the extent of stability (and instability) in the product network. Such changes in product assortment often lead to heightened perceptions of product search efforts (Broniarczyk et al. 1998).

When consumers cultivate habits in performing online searches (Moe and Yang 2009), they often develop a general reluctance to subsequently switch to other new search mech- anisms to reduce uncertainty in search outcomes (Shugan 1980). Thus, it is reasonable to assume that consumers will search for products by following those previous recommenda- tion links in subsequent visits or by repeating the same process that they are familiar with after the initial visits. Again, this suggests that consumers are likely to perceive the

8We further consulted AliResearch, which is under the Alibaba Group and is in charge of the data analytics research for the entire group including companies such as Tmall.com and Taobao.com. AliResearch’s observations reveal that consumers usually make multiple visits before committing to their final purchase, regardless of the prices of products.

MIS Quarterly Vol. 41 No. 2/June 2017 403

Lin et al./Demand Effects of Product Recommendation Networks

changes (if any) in recommendations and thus the extent of stability (and instability) in the product network.9

For a focal product, the stability of its incoming network will render those original incoming links to persist throughout consumers’ subsequent visits. It would hence be easier for consumers to surf along similar incoming network links to locate the same focal product identified during previous visits so as to commit to the final purchase. This suggests a likely positive relationship between incoming network stability and the focal product’s demand. This positive relationship could be further fuelled through psychological effects as well. Analogous to the offline changes in product assortment which heighten consumer perceptions of product search efforts (Broniarczyk et al. 1998), frequent changes in incoming network links (i.e., instability) may signal more tedious search efforts (Hoque and Lohse 1999) in relocating the focal product to purchase, due to the large number of products offered online. Arguably, higher incoming network stability can lower consumers’ perceived search efforts to induce purchase, and thus increase the demand of the focal product.

However, for a focal product, the stability of its outgoing network would likely exhibit a negative relationship with its product demand. First, let us examine the case whereby the outgoing network links comprise substitute products. As the purpose of the product recommendation systems is to help consumers find the right or ideal product (Schafer et al. 2001), consumers are likely to identify more desirable (compared to the focal product) substitute products in the focal product’s outgoing network (i.e., recommendation list) if there is high stability in the outgoing network (i.e., less updates in the recommendations). When the focal product consistently points to the same recommended products over time, con- sumers may be led into believing that these recommended products in the outgoing network are indeed widely or unanimously preferred by other consumers. Consequently, they may follow others’ preferences or behaviors (Duan et al. 2009) to purchase the recommended substitute products in the outgoing network instead of the focal product. Hence, the likelihood of purchasing the focal product will be reduced. In contrast, if the stability of the outgoing network decreases, this may signal to consumers that there is no consensus on the potential substitute products. Hence, they will more likely stick to their intended purchase of the focal products. In sum,

all these hint at a likely negative relationship between the outgoing network stability on the focal product’s demand, assuming the outgoing links comprise substitute products.10

Now, let us examine the case whereby the outgoing network links comprise complementary products. After all, another important objective of recommendation systems is to suggest additional complementary products to increase cross-selling opportunities (Schafer et al. 2001). Consumers may choose to co-purchase multiple product categories in a single or repeated shopping trips. Thus, recommendation systems often aim to propose complementary products to the focal product for consumers’ co-purchase. As such, if the stability of a focal product’s outgoing network increases, there would be fewer new options for consumers’ co-purchase of the focal product and the outgoing network products. Consequently, the demand of the focal product may decrease or remain stable at best. Thus, aggregating across both cases of substi- tute and complementary products in the outgoing network, it is still an open empirical question as to whether there is a negative demand relationship with outgoing network stability.

Co-View and Co-Purchase Networks and Product Demand

Co-view and co-purchase networks may have different effects on the demand of a focal product. It has been widely docu- mented that consumers make purchase decisions by observing other consumers’ preferences and behaviors (Yang and Allenby 2003). Co-view product recommendations contain information about other consumers’ choice set and online product search or browsing behaviors. In contrast, co- purchase product recommendations indicate other consumers’ actual purchase behaviors. As a result, the demand effects of co-view and co-purchase networks may differ, given that consumers do factor into consideration the different types or sources of information during their decision making process (Goh et al. 2013; Rao and Sieben 1992; Westbrook and Fornell 1979).

The relative effects of network diversity and network stability between co-view and co-purchase networks may be more complicated. Consumers are more likely to assess substitutes when they are still selecting their most preferred products

9This assertion is reasonable because a recommended product is typically displayed on a webpage using some short textual description coupled with a relatively salient and clear image which is usually prominent enough to facilitate consumers’ recall of products through recommendation links (see, for example, Figure 1). Consequently, consumers may be less inclined to use other search mechanisms (e.g., Google or even in-store search engine using textual inputs) in subsequent visits since they may not recall exact product names or model numbers.

10The aforementioned reasoning does not apply to the case of complementary products because competition does not exist between the focal product and its complements due to their different product functionalities. Thus, consumers may still purchase the focal product regardless of whether a specific complement is suitable or worth purchasing. However, the focal product and its substitutes face competition due to the same product functionalities. Hence, whether a substitute is preferred or not has a direct demand impact on the focal product.

404 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

(Roberts and Lattin 1991), but are more likely to evaluate complementary products when they are browsing or in an exploratory search mode (Schafer et al. 2001). This is why it is often believed that retailers attempt to utilize co-view recommendation systems to recommend substitutes to help consumers find the ideal product, but utilize co-purchase recommendation systems to suggest additional comple- mentary products to increase cross-selling opportunities (Schafer et al. 2001).

In other words, manipulating category diversity in the net- work, which is analogous to altering the variety of product assortment to offer a greater number of different product categories (Hoch et al. 1999), should be more relevant and important for consumers’ co-purchase of different comple- mentary product categories. In contrast, manipulating network stability, which is about replacing current products with new or better ones (Caro and Martínez-de-Albéeniz 2009), should be more relevant and important to help con- sumers identify and assess new substitute alternatives (Anderson 2006; Brynjolfsson et al. 2006). Based on this line of reasoning, the demand effect of network diversity may be stronger in the co-purchase network (relative to the co-view network) whereas the demand effect of network stability may be stronger in the co-view network (relative to the co- purchase network).

Empirical Methodology and Analysis

Data Description

Our dataset for empirical analysis is from Tmall.com (www.tmall.com), a China-based business-to-consumer (B2C) e-commerce platform under the Alibaba Group, which launched the largest initial public offering in U.S. history on November 19, 2014. Tmall was launched in April 2008, but was separated from Taobao’s consumer-to-consumer (C2C) marketplace to become an independent business in June 2011. Tmall consists of various online stores, currently featuring more than 70,000 multinational and Chinese brands from more than 50,000 merchants (Tmall 2015), and is the most visited B2C online retail website in China (Alexa 2015).

Our main dataset includes information of all products in an online flagship store which exclusively sells Nikon digital cameras and various associated components (e.g., lenses and batteries).11 Consistent with the retailer’s categorization, all products are grouped into six categories, namely (1) acces-

sory, (2) battery, (3) compact camera, (4) flash, (5) lens, and (6) single lens reflex (SLR) camera.12 Each product is fea- tured on its own designated webpage, including all relevant information (e.g., price, inventory, and consumer reviews) (see Figure 3 for an illustration) and its detailed transaction records (see Figure 4 for an illustration). Moreover, on each product webpage, Tmall also utilizes two different product recommendation systems. Recommendations can be listed under sections with the headings “Customers who viewed this item also viewed” (i.e., co-view, see Figure 5 for an illus- tration), and “Customers who bought this item also bought” (i.e., co-purchase, see Figure 6 for an illustration). The algorithm Tmall uses to provide recommendations is based on the collaborative filtering approach. Specifically, co-view (co-purchase) recommendation systems first identify the group of consumers who have viewed (purchased) a focal product. Co-view (co-purchase) recommendation systems then identify what other products these consumers also viewed (purchased) subsequently in order to provide the co- view (co-purchase) recommendations on the focal product’s page. Thus, recommendation links jointly form two networks (i.e., co-view network and co-purchase network) of all the products in the store.13 Noteworthy, these recommendations are updated in real time and are based solely on products from within a single focal store and do not include products across other stores. Therefore, this scheme of recommendation minimizes confounds from outside the focal store (e.g., competition from products outside the focal store) and provides us an unambiguous setting to investigate the impacts of within-store network factors on product demand.

In order to conduct our empirical analysis, we need to capture all products and their corresponding network structure. We collect data on product information and product network structure on a daily basis (12:00 a.m.). Consequently, our dataset consists of three parts: (1) daily snapshots of product- related information (e.g., price and consumer reviews), (2) daily snapshots of product network structure,14 and (3) detailed individual product transaction records with sales quantity. Our chief dataset on the digital camera category has 257 products in total across 184 days from May to December 2012.

11For robustness checks, we also collected data from another three stores selling personal computers, mobile phones, and cosmetic products. The substantive results across all four products were consistent.

12Compact cameras and SLR cameras are categorized separately because compact cameras are highly standardized whereas SLR cameras usually need extra lenses and flashes for different customizations.

13Tmall co-view and co-purchase recommendation lists have nonoverlapping constituent items, indicating that products in the co-view network are different from those in the co-purchase network. In Tmall, co-view recom- mendations are listed in a row just above the co-purchase ones at the bottom of a product’s webpage.

14The appendix illustrates snapshots of the product network structures using a graph.

MIS Quarterly Vol. 41 No. 2/June 2017 405

Lin et al./Demand Effects of Product Recommendation Networks

Figure 3. Tmall Product Webpage

Figure 4. Tmall Product Transaction Record

406 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Figure 5. Tmall Co-View Product Recommendation

Figure 6. Tmall Co-Purchase Product Recommendation

Empirical Model

Based on our dataset, we operationalize all relevant variables at the product-day level. Let subscript i denote each indi- vidual product in the online store, and subscript t denote each time period (daily). The dependent variable in this study is product i’s daily sales quantity, QUANit, measured as the total quantity of product i sold on day t. Next, our independent variables include network diversity and network stability. Network diversity is measured as the unique number of product categories in product i’s network on day t. Network stability is measured as the percentage of overlap of network connections in product i’s network across two days (snapshots of t and t+1), that is, the number of network connections that still existed on day t+1 over the number of original network connections on day t. Our research differs from prior studies in two aspects. First, we simultaneously examine the demand impact of both a focal product’s incoming and outgoing network. Second, we incorporate and differentiate the demand impact of two different recommendation mechanisms (i.e., co-view and co-purchase). Therefore, our final set of independent variables include incoming co-view network diversity (ID_CVit), incoming co-purchase network diversity

(ID_CPit), outgoing co-view network diversity (OD_CVit), outgoing co-purchase network diversity (OD_CPit), incoming co-view network stability (IS_CVit), incoming co-purchase network stability (IS_CPit), outgoing co-view network sta- bility (OS_CVit), and outgoing co-purchase network stability (OS_CPit). Finally, our control variables are gathered from those identified in our literature review and from the available information in our dataset. Specifically, we include control variables at the individual product, pricing, product network, product category, and time unit levels: (1) product list price (inclusive of discounts if any) (LPit), (2) volume of product reviews (PVit), (3) rating of product reviews (PRit), (4) past monthly sales quantity15 (PSit), (5) inventory

16 (INit), (6) num-

15This indicates the sales quantity of product i during the past month prior to day t.

16This indicates the available quantity of product i for sale on day t.

MIS Quarterly Vol. 41 No. 2/June 2017 407

Lin et al./Demand Effects of Product Recommendation Networks

ber of webpage bookmarks17 (BMit), (7) product vintage 18

(VTit), (8) network factors at the focal product level, that is, in-degree and out-degree co-view and co-purchase network centrality19 (IC_CVit, IC_CPit, OC_CVit, OC_CPit), (9) network factors at the entire network level, e.g., clustering coefficient20 (NW_CTit), (10) average sales quantity of incoming and outgoing co-view and co-purchase network products21 (IQ_CVit, IQ_CPit, OQ_CVit, OQ_CPit), (11) average sales quantity of outside goods22 in terms of competing products (OG_Q_Cit), search products (OG_Q_Sit) and experience products (OG_Q_Eit), (12) product category dummies (Ci), and (13) time dummies at the daily level (Tt). Control variables (1) to (6) are included because they are all displayed as the major product attributes on each product’s webpage on Tmall, and thus are likely to influence consumers’ product evaluations and purchase decisions subsequently. Control variable (7) is included because new products, compared to old products, are likely to attract more attention and purchases from consumers. Next, we include control variables (8) to (10) to capture the potential effects of other network factors at the focal product and entire network levels, apart from the network attributes of our focal interest (i.e., diversity and stability). Moreover, control variables (11) are included to control for alternative products that consumers may purchase other than Nikon digital cameras. Finally, control variables (12) and (13) are used to capture the fixed effects at the product category and time unit levels.

Noteworthy, we only observe the daily snapshots of product information and network structure. Thus, in order to derive more accurate measurement of variables, we measure our product-related variables (e.g., price) and network variables (e.g., network diversity) in time period t using the average values in the current (t) and the next (t+1) time periods.23

We model the influence of network diversity and network stability on product demand. We specify the dependent variable in logarithmic form.24 The panel-level linear model25

is specified in Equation (1):

(1)

( )ln _ _ _ _ _ _ _ _

_ _ _

_ _ _ _

QUAN ID CV ID CP OD CV OD CP

IS CV IS CP OS CV OS CP

LP PV PR PS IN BM

VT IC CV IC CP OC CV

OC CP NW CT IQ CV IQ CP

OQ

it it it it it

it it it it

it it it it it it

it it it it

it it it it

= + + + + + + + + + + + + + + + + + + + + + +

β β β β β β β β β β β β β β β β β β β β β β β

1 2 3 4

5 6 7 8

9 10 11 12 13 14

15 16 17 18

19 20 21 22

23 _ _ _ _

_ _ _ _

CV OQ CP OG Q C

OG Q S OG Q E C T it it it

it it i t i it

+ + + + + + + +

β β β β γ μ α ε

24 25

26 27

where αi captures unobserved product-specific effects. βs, γ, and μ are the model coefficients, and εit indicates the residual random error term. Table 1 reports the descriptive statistics and the correlation matrix is reported in Table 2.

Implicit Demand Correlation

We first address potential concerns over two products in the store having implicit demand correlation (i.e., substitution or complementarity) regardless of visible network connections being present. As such, the demand of a focal product would have been driven jointly by two factors: (1) implicit demand correlation and (2) explicit network structure. In order to distinguish the impact of network structure on product demand, we have to account for implicit demand correlation. Oestreicher-Singer and Sundararajan (2012b) provide stra- tegies to identify implicit demand correlation (albeit in terms of complementarity only). Specifically, for each focal prod-

17This indicates the total number of product i’s webpage bookmarked by consumers on day t.

18Although there is no information on a product’s first available date in the Tmall store, we resourcefully collected this information from Amazon by identifying the same product, and use the number of available days of product i up to day t as a measure for product vintage.

19 This measures the number of products in i’s product network on day t.

20In our model estimations, we also include many other network factors at the entire network level, such as betweenness centrality, closeness centrality, eigenvector centrality, and density, but they are not significant. Thus, we eventually exclude them from our model specification.

21This measures the average sales quantity across all products in i’s product network on day t. We also include other attributes of products in i’s product network including average list price, average product review volume, and rating. However, they are not significant and thus are excluded.

22This measures the average sales quantity of Canon digital cameras for competing products, notebook computers for search products, and facial skincare products for experience products. These data are collected from other Tmall stores. We use these three measures as proxies to control for the general extent of outside goods that consumers may purchase.

23Essentially, this captures the average value of a variable on day t, as we capture the snapshots of product information and network structure at 12:00 a.m. of t (the beginning of day t) and 12:00 a.m. of t+1 (the end of day t), and then derive the average of these two measures. We also report robustness checks on the sensitivity of this operationalization by using the actual day t values and find consistent results.

24Empirical analyses and model estimations often fit better with economic variables specified in logarithms (Wooldridge 2006, pp. 197-200). We add one to the variable to avoid logarithms of zeroes.

25We also examined potential nonlinear (i.e., quadratic) relationships between the focal network factors and product demand but practically found no such significant relationships.

408 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Table 1. Descriptive Statistics

Variable Mean Std. Dev. Min Max

QUAN (Sales quantity) 0.283 6.791 0.000 680.000

ID_CV (Incoming network diversity, co-view) 0.888 0.584 0.000 6.000

ID_CP (Incoming network diversity, co-purchase) 1.180 1.250 0.000 6.000

OD_CV (Outgoing network diversity, co-view) 1.090 0.446 0.000 4.000

OD_CP (Outgoing network diversity, co-purchase) 1.519 1.133 0.000 5.000

IS_CV (Incoming network stability, co-view) 0.787 0.307 0.000 1.000

IS_CP (Incoming network stability, co-purchase) 0.711 0.380 0.000 1.000

OS_CV (Outgoing network stability, co-view) 0.757 0.321 0.000 1.000

OS_CP (Outgoing network stability, co-purchase) 0.727 0.378 0.000 1.000

LP (Product list price) 4,245.943 9,811.471 1.000 69,100.000

PV (Product review volume) 3.875 30.804 0.000 634.000

PR (Product review rating) 1.537 2.276 0.000 5.000

PS (Product past monthly sales quantity) 1.754 12.378 0.000 320.000

IN (Product inventory) 84.742 357.640 1.000 9,765.000

BM (Number of product webpage bookmarks) 59.267 598.451 0.000 9,783.000

VT (Product vintage) 1,835.306 1,549.886 0.000 4,095.000

IC_CV (In-degree centrality, co-view) 3.712 4.830 0.000 57.500

IC_CP (In-degree centrality, co-purchase) 3.570 7.427 0.000 87.000

OC_CV (Out-degree centrality, co-view) 3.707 0.931 0.000 4.000

OC_CP (Out-degree centrality, co-purchase) 3.559 2.032 0.000 5.000

NW_CT (Network clustering coefficient) 0.317 0.229 0.000 1.333

IQ_CV (Incoming network product sales quantity, co-view) 0.052 0.652 0.000 57.500

IQ_CP (Incoming network product sales quantity, co-purchase) 0.079 0.641 0.000 43.167

OQ_CV (Outgoing network product sales quantity, co-view) 0.269 0.658 0.000 14.125

OQ_CP (Outgoing network product sales quantity, co-purchase) 0.335 1.210 0.000 70.125

OG_Q_C (Sales quantity of outside goods, competing brand) 0.056 0.487 0.000 6.333

OG_Q_S (Sales quantity of outside goods, search product) 3.200 2.154 1.000 12.571

OG_Q_E (Sales quantity of outside goods, experience product) 1.691 2.145 0.000 20.000

Notes: Number of observations = 41,379; Number of products = 257; Number of days = 184. All variables are at the product-day level.

Table 2. Correlations

Variable 1 2 3 4 5 6 7 8 9

1 QUAN (Sales quantity) -

2 ID_CV (Incoming network diversity, co-view) -0.027 -

3 ID_CP (Incoming network diversity, co-purchase) 0.000 0.478 -

4 OD_CV (Outgoing network diversity, co-view) -0.068 0.212 0.030 -

5 OD_CP (Outgoing network diversity, co- purchase)

-0.024 0.201 0.472 0.294 -

6 IS_CV (Incoming network stability, co-view) 0.020 -0.184 -0.081 0.060 0.066 -

7 IS_CP (Incoming network stability, co-purchase) 0.011 -0.143 -0.247 0.035 -0.025 0.146 -

8 OS_CV (Outgoing network stability, co-view) 0.013 0.030 0.032 -0.242 -0.009 0.193 0.100 -

9 OS_CP (Outgoing network stability, co-purchase) -0.000 -0.014 -0.097 -0.226 -0.303 0.026 0.112 0.434 -

Notes: Only major variables are reported. The correlations of these variables with other variables are generally small.

MIS Quarterly Vol. 41 No. 2/June 2017 409

Lin et al./Demand Effects of Product Recommendation Networks

uct, their study identified a set of products which have implicit demand correlation with the focal product, and then controlled for the demand of this set of products as their main identification strategy. Unfortunately, all strategies used in Oestreicher-Singer and Sundararajan (2012b) were based on a strong assumption, that is, complementarity is the only factor that drives the implicit demand correlation.

To better address this issue, we propose a different strategy which seeks to capture implicit demand correlation in terms of both product substitution and complementarity. Our stra- tegy goes beyond the assumption used in Oestreicher-Singer and Sundararajan (2012b) to allow a pair of products to be substitutable between, complementary to, or independent of each other. Specifically, we use the cross-product category price elasticity to determine the extent of substitution or com- plementary relationships (Leeflang and Parreño-Selva 2012; Manchanda et al. 1999; Niraj et al. 2008). Ideally, substitu- tion and complementarity effects should be estimated at the product level. However, past research (Berry 1994; Song and Chintagunta 2006) has outlined that estimating at the product level would be problematic due to the issue of “parameters explosion” (i.e., too many parameters to estimate due to the huge number of products when we specify a full model to account for product demand interdependency). We thus obtain cross-category level price elasticities to determine the extent of substitution or complementary relationships between each pair of products. Specifically, we first group all products into six different product categories as discussed previously based on the retailer’s classification according to the product functionalities, and consider all products within a product category as substitutes for one another. This approach is in line with the definition of substitutes proposed by Henderson and Quandt (1958, p. 29): “two commodities are substitutes if both can satisfy the same need.” This approach has been adopted in past research (Mulhern and Leone 1991; Walters 1991). After the grouping, we aggregate all product charac- teristic variables (e.g., price) from the product level to the category level.26

Let subscript m = 1, 2, 3, 4, 5, and 6 denote each product category, and subscript t denote each day. The dependent variable is category sales quantity,27 QUANmt, measured as the total quantity of category m sold on day t. The explanatory variables are all the major category-level characteristic vari-

ables aggregated from the product level, including: (1) list price (LPmt), (2) volume of product reviews (PVmt), (3) rating of product reviews (PRmt), (4) past monthly sales quantity (PSmt), (5) inventory (INmt), and (6) number of webpage book- marks (BMmt). The descriptive statistics of all of these category-level variables are reported in the appendix.

To obtain cross-category price elasticities, we specify a model including the above six variables for each product category m, where we estimate a panel-level linear model shown in Equation (2):

(2) ( )ln QUAN LP PV PR

PS IN BM T

mt mn nt n

n

mn nt mn nt n

n

n

n

mn nt mn nt mn nt n

n

n

n

n

n

t m mt

= + +

+ + + + + +

=

=

=

=

=

=

=

=

=

=

=

=

 



α β χ

δ ε ϕ μ ω ξ

1

6

1

6

1

6

1

6

1

6

1

6

where Tt denotes time dummies at the daily level, ωm captures unobserved category-specific effects, αmn, βmn, χmn, δmn, εmn, φmn, and μ are the model coefficients, and ξmt denotes the residual random error term.

The estimated model coefficients based on data from each category of multiple Nikon stores28 are summarized in the appendix. We compute the category-level price elasticities and summarize them in Table 3. As discussed above, we use the estimated cross-category price elasticities to determine the potential substitution or complementary relationship between each pair of products.29 Specifically, let Eij denote the price elasticity of product i’s demand with respect to the price of product j. Suppose product i is in category m and product j is in category n, we use the estimated category-level elasticity (i.e., entry (m, n) in Table 3, the price elasticity of category m’s demand with respect to the price of category n) as a proxy for Eij. However, if both products i and j are in the same category (i.e., m = n), we follow the above discussion to consider that products i and j are substitutes, and use the absolute value of own-category price elasticity30 (i.e., Eij = |entry (m, n)|) to indicate the substitution effect of product j on product i. As such, we can identify the underlying substitute or complementary relationship between each pair of products

26Each category-level characteristic variable is the weighted sum of the product-level characteristic variable. Weight for product i on day t is the sales quantity of product i on day t over the total sales quantity of all products in the same category on day t.

27We use sales quantity to maintain consistency with the product demand specification in Equation (1).

28Using data from multiple stores provides us with a larger model estimation sample for more precise estimates.

29In principle, our approach allows for specifying a random coefficient struc- ture for αmn, the price parameter, to allow for heterogeneity in price sensitivity within a product category. This relaxes the restriction that price elasticities for all products within the same category are identical.

30Theoretically, own-price elasticity of demand is always negative. However, the cross-price elasticity between two substitutes should be positive. Thus, we use the absolute value of own-price elasticity.

410 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Table 3. Price Elasticity Matrix

Category m Category n

(1) Accessory

(2) Battery

(3) Compact Camera

(4) Flash

(5) Lens

(6) SLR Camera

(1) Accessory -0.127*** (0.047)

-0.089** (0.041)

-0.055*** (0.008)

0.043*** (0.008)

0.071*** (0.016)

0.006 (0.025)

(2) Battery 0.245***

(0.076) -0.163* (0.094)

0.032* (0.018)

0.005 (0.012)

0.151*** (0.012)

0.121* (0.073)

(3) Compact camera -0.209 (0.318)

-0.183*** (0.070)

-0.469*** (0.135)

0.050** (0.023)

-0.043 (0.030)

-0.034 (0.024)

(4) Flash -0.796* (0.458)

-0.323** (0.153)

-0.338*** (0.092)

-0.703*** (0.077)

-0.015 (0.148)

-0.144 (0.125)

(5) Lens 0.183***

(0.024) -0.012 (0.051)

0.043** (0.021)

-0.007 (0.007)

-0.832*** (0.060)

0.013 (0.079)

(6) SLR camera 0.293

(0.325) 0.240

(0.173) -0.080 (0.145)

-0.022 (0.021)

0.094 (0.155)

-0.881*** (0.267)

Notes: Entry (m, n) represents the elasticity of category m with respect to the price of category n. Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01.

based on this strategy.31 Noteworthy, our approach allows for the implicit demand correlation in terms of both substitution and complementarity between the focal product and all of the other products in the same store, regardless of a visible net- work connection being present. Thus, for each focal product, we treat all other products in the same store as potential substitutes or complements. We consider product j to be a substitute (complement) to focal product i if Eij > 0 (Eij < 0).

Let s = 1, 2, 3,…, S denote a substitute product for focal product i. As different substitutes may have different extent of substitution effects on product i’s demand, and price elasticity Eis represents the level of substitution between products i and s, we weigh each substitute s’s sales quantity on day t (QUANst) according to Eis, and then sum these weighted demands32 to be the influence of substitution effect

on product i on day t, denoted as SEit, shown in Equation (3) below. Likewise, let c = 1, 2, 3,…, C denote a comple- mentary product for focal product i. As different comple- ments may have different extent of complementarity effects on product i’s demand, and price elasticity Eic represents the level of complementarity between products i and c, we weigh each complement c’s sales quantity on day t (QUANct) according to Eic, and then sum these weighted demands to be the influence of complementarity effect on product i on day t, denoted as CEit, shown in Equation (4) below.

(3)SE E QUAN for all s with Eit is s

s S

st is= ∗ > =

=

 1

0,

(4)CE E QUAN for all c with Eit ic ct ic c

c C

= ∗ < =

=

 , 0 1

Essentially, we include these two variables SEit and CEit 33 in

Equation (1) to account for the substitution and comple- mentarity effects of other products that may have intrinsic demand correlations through product usage or exogenous shocks, or may appear as visible network links, or are cur- rently invisible but potentially future network links (Carmi et al. 2017; Oestreicher-Singer and Sundararajan 2012b) which may all implicitly drive the focal product’s demand.

31We report robustness checks on the sensitivity of this operationalization by assigning zeros to all the insignificant estimated elasticities and find consistent results with the main findings reported.

32Our approach of weighing the demand of a product (be it a substitute or a complement) according to the cross-price elasticity (i.e., the extent of substi- tution or complementarity) is similar to the approach used in Oestreicher- Singer and Sundararajan (2012b) which weighed the demand of a product (complement only) according to its probability of being linked to (the extent of complementarity to) the focal product. Noteworthy, our use of price elas- ticities allows us to better identify both substitution and complementary relationships (based on signs of cross-price elasticities) whereas the use of the probability of link formation assumes the absence of substitution relationship. Furthermore, we report robustness checks on the sensitivity of this opera- tionalization by using the summation of actual demands without price elasticities as weights and find consistent results with the main findings reported.

33The correlations of SE and CE with other sales quantity related variables associated with the network links (i.e., IQ_CV, IQ_CP, OQ_CV, OQ_CP) are generally below 0.1, suggesting that both the SE and CE variables capture implicit demand factors beyond those reflected in the recommended product links.

MIS Quarterly Vol. 41 No. 2/June 2017 411

Lin et al./Demand Effects of Product Recommendation Networks

Results

We first estimate a random effects (RE) model of product sales quantity on all the control variables. As reported in Table 4, Column (1), product own attributes (e.g., review volume, review rating, past monthly sales, inventory, and product vintage) have the expected relationships with product sales quantity. Various network-level control variables (e.g., degree centrality, clustering coefficient, and sales quantity of the connected products) have significant relations with product sales quantity as well. Moreover, sales of outside goods in terms of competing products and search products have a negative relation with focal product sales, whereas sales of outside goods in terms of experience products have a positive relationship. More importantly, the substitution effect (SE) and the complementarity effect (CE) coefficients are statistically significant and with the expected signs, implying that a higher substitution effect decreases the demand of a product, whereas a higher complementarity effect instead increases it. These significant effects of SE and CE attest to the effectiveness of our approach in accounting for the implicit demand correlations. These results suggest that our set of control variables have good explanatory power.

Beyond these control variables, we then estimate a RE model including the independent variables (network diversity and network stability) of only the incoming network. As reported in Table 4, Column (2), the coefficients of co-view and co- purchase network diversity (ID_CV, ID_CP) and co-view net- work stability (IS_CV) are statistically significant. Similarly, we estimate a RE model including the independent variables of only the outgoing network and summarize the results in Table 4, Column (3). We find that three model coefficients for outgoing network variables (OD_CP, OS_CV, OS_CP) are statistically significant. These results suggest the importance of accounting for both incoming and outgoing networks when assessing the demand impacts of recommendation networks.34

We next estimate a RE model including the independent variables (incoming and outgoing network diversity and network stability) of only the co-view network. As reported in Table 4, Column (4), the coefficients of incoming network diversity (ID_CV) and incoming network stability (IS_CV) are statistically significant. However, these variables become insignificant in subsequent model estimations which include co-purchase network related variables. This highlights the importance of investigating both co-view and co-purchase recommendation networks for a more reliable and compre- hensive examination.

Likewise, we also estimate a RE model including the indepen- dent variables of only the co-purchase network. As reported in Table 4, Column (5), the coefficients of the incoming network diversity (ID_CP), outgoing network diversity (OD_CP), and outgoing network stability (OS_CP) are all highly significant and with the expected signs. By further comparing model fit statistics between the co-view RE model in Column (4) (R2 = 0.4914) and the co-purchase RE model in Column (5) (R2 = 0.4921), we find that co-purchase network factors, relative to co-view ones, explain more variations in product demand.

Next, we further estimate a full RE model including both incoming and outgoing network variables, and both co-view and co-purchase network variables. Table 4, Column (6), summarizes the results. Three independent variables are statistically significant. First, incoming co-purchase network diversity (ID_CP) is positive and significant, suggesting an expected positive relation with the focal product’s demand. Second, outgoing co-purchase network diversity (OD_CP) is negative and significant, suggesting an expected negative relation with the focal product’s demand. Last, outgoing co- purchase network stability (OS_CP) also has an expected negative relation with product demand.

In addition to the RE model, we further estimate a full fixed effects (FE) model of product sales quantity on all the explan- atory variables and summarize the results in Table 4, Column (7). The R2 of the FE model (R2 = 0.1081) is much lower than that of the full RE model (R2 = 0.4934) in Column (6), sug- gesting that RE model has a better model fit than the FE one. Importantly, the Hausman test (χ2 = 69.52, p = 1.00) suggests that the RE model estimates are not inconsistent, implying the appropriateness of the RE model over the FE one. Thus, we consider the RE model in Column (6) as our preferred model specification.

To address the concern of potential simultaneity between network structures (i.e., diversity and stability) and product demand, we report two additional model estimations. Speci- fically, we estimate the one-day and seven-day lagged effect of network structure on current product sales quantity. The network structure in the past should not have been influenced by the current product sales quantity. This helps rule out the potential simultaneity concern. The estimation results are reported in Table 5, Columns (2) and (3). For brevity, from this point onward, we only report the major variables of interest. For ease of reference, Table 5, Column (1), presents the results from Table 4, Column (6). Evidently, those previously identified three significant network diversity and stability estimates are all consistent with those in Column (1).

We note that Oestreicher-Singer and Sundararajan (2012b) excluded the products for which all incoming network links

34This observation has to be qualified in the case of bidirectional links (i.e., incoming links are the same as outgoing links). Specifically, in Table 5, Columns (4) to (6) present results on various models that exclude bidirectional links.

412 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Table 4. Model Estimation Results

Variable

(1) RE:

Control

(2) RE:

Incoming

(3) RE:

Outgoing

(4) RE:

Co-view

(5) RE: Co-

purchase

(6) RE: Full

(7) FE: Full

ID_CV -0.006** -0.006** -0.001 0.002 (Incoming network diversity, co-view) (0.002) (0.003) (0.003) (0.003) ID_CP 0.006*** 0.008*** 0.010*** 0.000 (Incoming network diversity, co- purchase)

(0.001) (0.001) (0.001) (0.002)

OD_CV -0.001 -0.000 -0.003 -0.004 (Outgoing network diversity, co-view) (0.004) (0.004) (0.004) (0.004) OD_CP -0.005*** -0.008*** -0.008*** -0.002 (Outgoing network diversity, co- purchase)

(0.002) (0.002) (0.002) (0.002)

IS_CV 0.007* 0.006* 0.005 0.003 (Incoming network stability, co-view) (0.003) (0.003) (0.003) (0.003) IS_CP -0.000 0.002 -0.000 -0.002 (Incoming network stability, co- purchase)

(0.003) (0.003) (0.003) (0.003)

OS_CV 0.007* 0.000 0.006 0.007* (Outgoing network stability, co-view) (0.004) (0.003) (0.004) (0.004) OS_CP -0.017*** -0.015*** -0.017*** -0.013*** (Outgoing network stability, co- purchase)

(0.003) (0.003) (0.003) (0.003)

LP 0.000 0.000 0.000 0.000 0.000 -0.000 -0.000* (Product list price) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) PV 0.002*** 0.002*** 0.001*** 0.001*** 0.002*** 0.002*** -0.000 (Product review volume) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) PR 0.004*** 0.003*** 0.003*** 0.004*** 0.003*** 0.003*** -0.001 (Product review rating) (0.001) (0.001) (0.001) (0.000) (0.001) (0.001) (0.001) PS 0.012*** 0.012*** 0.012*** 0.012*** 0.012*** 0.012*** 0.008*** (Product past monthly sales quantity) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) IN 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** (Product inventory) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) BM -0.000*** -0.000*** -0.000*** -0.000*** -0.000*** -0.000*** -0.000*** (Number of product webpage bookmarks)

(0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)

VT -0.000* -0.000*** -0.000** -0.000** -0.000*** -0.000** -0.000 (Product vintage) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) IC_CV -0.000 -0.001** -0.001*** -0.001* -0.000 (In-degree centrality, co-view) (0.000) (0.000) (0.000) (0.000) (0.000) IC_CP -0.001*** -0.002*** -0.002*** -0.002*** -0.002*** (In-degree centrality, co-purchase) (0.000) (0.000) (0.000) (0.000) (0.000) OC_CV -0.010*** -0.010*** -0.008*** -0.009*** 0.001 (Out-degree centrality, co-view) (0.001) (0.002) (0.002) (0.002) (0.002) OC_CP 0.000 0.001 0.001 0.001 -0.000 (Out-degree centrality, co-purchase) (0.001) (0.001) (0.001) (0.001) (0.001) NW_CT 0.011** 0.011*** 0.011*** 0.011*** 0.012*** 0.012*** 0.009** (Network clustering coefficient) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004)

MIS Quarterly Vol. 41 No. 2/June 2017 413

Lin et al./Demand Effects of Product Recommendation Networks

Table 4. Model Estimation Results (Continued)

Variable

(1) RE:

Control

(2) RE:

Incoming

(3) RE:

Outgoing

(4) RE:

Co-view

(5) RE: Co-

purchase

(6) RE: Full

(7) FE: Full

IQ_CV 0.002 0.002 0.001 0.002 0.002 (Incoming network product sales quantity, co-view)

(0.002) (0.002) (0.002) (0.002) (0.001)

IQ_CP 0.005*** 0.005*** 0.004** 0.004** 0.004** (Incoming network product sales quantity, co-purchase)

(0.002) (0.002) (0.002) (0.002) (0.002)

OQ_CV 0.006*** 0.006*** 0.005*** 0.008*** 0.003* (Outgoing network product sales quantity, co-view)

(0.002) (0.002) (0.002) (0.002) (0.002)

OQ_CP 0.003*** 0.004*** 0.003*** 0.003*** 0.003*** (Outgoing network product sales quantity, co-purchase)

(0.001) (0.001) (0.001) (0.001) (0.001)

OG_Q_C -0.045*** -0.045*** -0.044*** -0.043*** -0.047*** -0.045*** 0.069*** (Sales quantity of outside goods, competing brand)

(0.017) (0.017) (0.017) (0.017) (0.017) (0.017) (0.006)

OG_Q_S -0.143*** -0.139*** -0.143*** -0.146*** -0.132*** -0.139*** 0.022*** (Sales quantity of outside goods, search product)

(0.027) (0.027) (0.027) (0.027) (0.027) (0.027) (0.005)

OG_Q_E 0.035*** 0.034*** 0.035*** 0.036*** 0.032*** 0.034*** 0.013*** (Sales quantity of outside goods, experience product)

(0.006) (0.006) (0.006) (0.006) (0.006) (0.006) (0.002)

SE -0.005*** -0.005*** -0.005*** -0.005*** -0.005*** -0.005*** -0.004*** (Substitution effect) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) CE 0.009*** 0.008*** 0.009*** 0.009*** 0.009*** 0.009*** 0.007*** (Complementarity effect) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001)

Constant 0.326*** 0.276*** 0.323*** 0.319*** 0.276*** 0.323*** -0.159

(0.046) (0.046) (0.046) (0.046) (0.046) (0.046) (0.113) Category & time dummies -included- -included- -included- -included- -included- -included- -include- Number of observations 41,379 41,379 41,379 41,379 41,379 41,379 41,379 Hausman test χ² = 69.52, p = 1.00 R² 0.4924 0.4918 0.4917 0.4914 0.4921 0.4934 0.1081

Notes: Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01.

are bidirectional for each day to control for potential simul- taneity. We use a similar approach by excluding products for which all incoming network links are bidirectional (100% bidirectionality, i.e., the focal product has outgoing network links terminating at all its incoming network products). More than 17% of our observations are eliminated by this opera- tionalization. The estimation results based on this reduced sample are summarized in Table 5, Column (4). The results are consistent with those in Column (1). To be more conser- vative, we further exclude products for which more than 50%, and even more than 10%, of the incoming network links are bidirectional (i.e., the focal product has outgoing network links terminating at more than 50%, and even more than 10%,

of its incoming network products respectively). These two operationalizations reduce the observations by about 46% and 80% respectively. The model estimates are reported in Table 5, Columns (5) and (6) respectively. The estimated signs and significance are still consistent with those in Column (1). We further lower the cut-off values of the bidirectionality percent- age, and find that the model estimates maintain their consis- tency across estimation samples. In sum, our results here do significantly alleviate concerns of potential simultaneity.

Additionally, to further address the concern of simultaneity, we conduct a series of Granger causality tests (Granger 1969) to check whether current product network structure (i.e., diver-

414 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Table 5. Simultaneity Checks

Variable (1)

Preferred (2)

Lag 1 (3)

Lag 7

(4) Reduce

bidirectionality (= 100%)

(5) Reduce

bidirectionality (> 50%)

(6) Reduce

bidirectionality (> 10%)

ID_CV -0.001 -0.004 -0.010*** -0.001 0.003 0.007

(Incoming, diversity, co-view) (0.003) (0.003) (0.002) (0.003) (0.004) (0.013)

ID_CP 0.010*** 0.007*** 0.009*** 0.011*** 0.014*** 0.020**

(Incoming, diversity, co-purchase) (0.001) (0.001) (0.001) (0.002) (0.002) (0.009)

OD_CV -0.003 0.000 -0.004 -0.002 -0.003 0.016

(Outgoing, diversity, co-view) (0.004) (0.003) (0.003) (0.005) (0.006) (0.013)

OD_CP -0.008*** -0.004*** -0.005*** -0.009*** -0.015*** -0.030***

(Outgoing, diversity, co-purchase) (0.002) (0.001) (0.001) (0.002) (0.003) (0.008)

IS_CV 0.005 -0.004 0.001 0.006 0.010* 0.008

(Incoming, stability, co-view) (0.003) (0.004) (0.004) (0.004) (0.005) (0.013)

IS_CP -0.000 0.000 0.005* -0.000 -0.001 -0.004

(Incoming, stability, co-purchase) (0.003) (0.003) (0.003) (0.003) (0.005) (0.013)

OS_CV 0.006 0.009** 0.004 0.006 0.012* 0.022

(Outgoing, stability, co-view) (0.004) (0.004) (0.004) (0.004) (0.006) (0.013)

OS_CP -0.017*** -0.012*** -0.009*** -0.018*** -0.028*** -0.059***

(Outgoing, stability, co-purchase) (0.003) (0.003) (0.003) (0.004) (0.005) (0.014)

Constant 0.323*** 0.325*** 0.345*** 0.353*** 0.479*** 1.492***

(0.046) (0.046) (0.047) (0.054) (0.077) (0.180)

Control variables -included- -included- -included- -included- -included- -included-

Number of observations 41,379 40,836 38,927 34,070 22,279 8,200

Sample reduced 17.664% 46.159% 80.183%

R² 0.4934 0.4948 0.5021 0.5027 0.5340 0.6269

Notes: Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01.

sity and stability) would have been affected by past product sales performance (i.e., sales quantity). Specifically, we use different time lag levels from one to seven days to test whether product sales quantity Granger-causes each of the network diversity and stability variables. The test results show that current product network structures have not been affected by past sales quantity (i.e., simultaneity would not be a concern in this study). Thus, this lends further credence to the effects of network diversity and stability on product demand. We summarize the Granger causality test results in the appendix.

Last, to further corroborate the influence of network diversity and network stability on product demand, we adapt the procedure of the shuffle test (Anagnostopoulos et al. 2008) to a linear model in order to rule out a correlational interpre- tation of our results and establish more confidence for a causal interpretation. This test is based on the rationale that if causal influence of network diversity and stability does not play a role, even though a product’s demand could depend on

the neighboring products in the recommendations network, the timing of such dependency should be independent of the neighboring products. Thus, for each focal product, we ran- domly shuffle its network diversity and stability values35 over the entire sample period for each day, while keeping all the other measures (e.g., degree centrality, price) unchanged. Next, we reestimate the demand model in Equation (1) using the shuffled product network’s new diversity and stability measures. We thus obtain the estimated network diversity and stability parameters before and after the shuffling, and then test for the structural difference across these two sets of model parameters based on the Chow test. We perform five different randomizations for the shuffling, and all the Chow test results show that the two sets of parameters are signifi- cantly different, thus providing some evidence and basis of

35Essentially, this is analogous to randomly shuffling different sets of network links over time, since network diversity and stability values are calculated based on these links.

MIS Quarterly Vol. 41 No. 2/June 2017 415

Lin et al./Demand Effects of Product Recommendation Networks

Table 6. Elasticities of Demand

Network diversity ID_CV ID_CP OD_CV OD_CP

Elasticity - 0.011*** - -0.012***

Network stability IS_CV IS_CP OS_CV OS_CP

Elasticity - - - -0.012***

Notes: *p < 0.1, **p < 0.05, ***p < 0.01.

Table 7. Standardized Regression Coefficients

Network diversity ID_CV ID_CP OD_CV OD_CP

Std. coefficients - 0.044*** - -0.032***

Network stability IS_CV IS_CP OS_CV OS_CP

Std. coefficients - - - -0.023***

Notes: *p < 0.1, **p < 0.05, ***p < 0.01.

confidence for a causal interpretation on the influence of network diversity and network stability on product demand. The detailed process and results of the shuffle test are sum- marized in the appendix.

In summary, after accounting for the implicit demand correla- tion and potential simultaneity, we identify three significant effects of product network structures (i.e., ID_CP, OD_CP, OS_CP). We further report the elasticities for these three significant factors in Table 6 based on the preferred model. First, incoming co-purchase network diversity (ID_CP, elasticity = 0.011, p < 0.01) has a positive relation with the focal product’s demand. Second, outgoing co-purchase network diversity (OD_CP, elasticity = -0.012, p < 0.01) has a negative relation with the focal product’s demand. Next, outgoing co-purchase network stability (OS_CP, elasticity = -0.012, p < 0.01) has a negative relation with the focal prod- uct’s demand. However, incoming network stability (IS_CV and IS_CP) has no significant relation with product demand.

Finally, we compare the relative demand effects of network diversity and network stability between co-view and co- purchase networks. The estimated coefficients and elasticities from our main model indicate that the network diversity and network stability results are significant only in the co- purchase network but not in the co-view network. This shows the more influential role of the co-purchase network over that of the co-view network for both network diversity and net- work stability. As an additional test, we report in Table 7 the standardized regression coefficients by standardizing all the variables in our model (Darlington 1990) to have a more comparable interpretation. Similar to earlier results, network diversity and network stability variables are significant only

in the co-purchase network as well. Thus, both sets of results indicate that network diversity’s relation with product demand is significant only in the co-purchase networks, and thus stronger than that of the insignificant relation in the co-view networks. Network stability’s negative relation with product demand is also significant only in the co-purchase networks. Contrary to expectation, this relation is not weaker than the insignificant relation between network stability and product demand in the co-view networks.

Robustness Checks

We further corroborate our findings by checking its robust- ness in multiple ways. We first report robustness checks on the sensitivity of our operationalizations by constructing the substitution and complementarity effects (SE and CE) in Equations (3) and (4) in different ways. For ease of reference, Table 8, Column (1), presents the results from our preferred model in Table 4, Column (6).

First, some of the estimated cross-category price elasticities in Table 3 are not significant, which might imply that two products or categories do not have demand correlations. Thus, we replace the values of all the insignificant price elasticities with zeros for the constructions of SE and CE. The results in Table 8, Column (2) are almost identical to those in Column (1).

Second, we construct SE and CE using the sum of actual demands without using the estimated price elasticities as weights. Results shown in Table 8, Column (3) are generally consistent with those reported in Column (1).

416 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Table 8. Robustness Checks (1)

Variable (1)

Preferred

(2) Significant elasticity

only

(3) Without elasticity weight

(4) 1st-stage

uncertainty

ID_CV -0.001 -0.001 0.002 -0.001

(Incoming, diversity, co-view) (0.003) (0.003) (0.002) (0.003)

ID_CP 0.010*** 0.010*** 0.008*** 0.010***

(Incoming, diversity, co-purchase) (0.001) (0.001) (0.001) (0.001)

OD_CV -0.003 -0.003 -0.001 -0.002

(Outgoing, diversity, co-view) (0.004) (0.004) (0.003) (0.004)

OD_CP -0.008*** -0.008*** -0.005*** -0.007***

(Outgoing, diversity, co-purchase) (0.002) (0.002) (0.001) (0.002)

IS_CV 0.005 0.005 0.003 0.005

(Incoming, stability, co-view) (0.003) (0.003) (0.003) (0.003)

IS_CP -0.000 0.000 0.002 0.001

(Incoming, stability, co-purchase) (0.003) (0.003) (0.002) (0.003)

OS_CV 0.006 0.005 0.006* 0.005

(Outgoing, stability, co-view) (0.004) (0.004) (0.003) (0.004)

OS_CP -0.017*** -0.017*** -0.012*** -0.017***

(Outgoing, stability, co-purchase) (0.003) (0.003) (0.003) (0.003)

Constant 0.323*** 0.284*** 8.403*** 0.145***

(0.046) (0.046) (0.068) (0.046)

Control variables -included- -included- -included- -included-

Number of observations 41,379 41,379 41,379 41,379

R² 0.4934 0.4924 0.6620 0.4888

Notes: Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01.

Third, we address a potential concern over the construction of SE and CE in the main estimation (i.e., second stage) that uses the estimated price elasticities from the previous stage (i.e., first stage). We take into account the first-stage uncertainty of estimated coefficients (as reflected in standard errors) in the second-stage estimation. Specifically, in Equations (3) and (4) where SE and CE are computed, we further include the first-stage estimated standard errors as weights for the elasticity estimates. We then estimate a RE model with SE and CE replaced by the first-stage uncertainty weighted substitution effect and complementarity effect. The results are summarized in Table 8, Column (4). As indicated, the network diversity and network stability estimates are consistent with the results of our preferred specification in Column (1).

In addition to the robustness checks on the sensitivity of different constructions of SE and CE, we also check the robustness of our findings in many other ways. For ease of reference, Table 9, Column (1), and Table 10, Column (1)

also present the results from our preferred model in Table 4, Column (6).

First, one may be concerned that the changes in a focal product’s network (leading to instability in a network) may introduce different products into this focal product’s network (i.e., increasing the diversity of the network). Thus, the potential collinearity between network diversity and network stability would be a concern. However, Table 2 shows that the correlations between network diversity and network stability are generally small, which suggests that collinearity would not be a concern in this study. Nevertheless, we still perform mean-subtracted centralization and standardization to all the independent variables. We estimate Equation (1) based on these variables and summarize the results in Table 9, Columns (2) and (3) respectively. The results are consistent with those from our preferred specification.

Second, to account for the existence of potential serial correlation, we estimate a RE model with a first-order auto-

MIS Quarterly Vol. 41 No. 2/June 2017 417

Lin et al./Demand Effects of Product Recommendation Networks

Table 9. Robustness Checks (2)

Variable (1)

Preferred

(2) Mean-

centering

(3) Standardi-

zation

(4) Serial

correlation

(5) Price

endogeneity

(6) Stability

3-day

(7) Stability

7-day

(8) Assortative

mixing

ID_CV -0.001 -0.001 -0.001 0.001 -0.001 -0.002 -0.003 0.007**

(Incoming, diversity, co-view) (0.003) (0.003) (0.002) (0.003) (0.004) (0.003) (0.003) (0.003)

ID_CP 0.010*** 0.010*** 0.012*** 0.005*** 0.012*** 0.009*** 0.009*** 0.010***

(Incoming, diversity, co- purchase)

(0.001) (0.001) (0.002) (0.002) (0.002) (0.001) (0.001) (0.001)

OD_CV -0.003 -0.003 -0.001 -0.004 -0.007 -0.000 0.000 -0.003

(Outgoing, diversity, co-view) (0.004) (0.004) (0.002) (0.004) (0.005) (0.004) (0.004) (0.004)

OD_CP -0.008*** -0.008*** -0.009*** -0.005** -0.008*** -0.007*** -0.007*** -0.009***

(Outgoing, diversity, co- purchase)

(0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) (0.002)

IS_CV 0.005 0.005 0.002 0.005 0.007 0.005* 0.003 0.005

(Incoming, stability, co-view) (0.003) (0.003) (0.001) (0.003) (0.005) (0.003) (0.003) (0.003)

IS_CP -0.000 -0.000 -0.000 -0.002 -0.000 0.000 0.001 -0.001

(Incoming, stability, co- purchase)

(0.003) (0.003) (0.001) (0.003) (0.004) (0.003) (0.003) (0.003)

OS_CV 0.006 0.006 0.002 0.006 0.005 0.011*** 0.010*** 0.006

(Outgoing, stability, co-view) (0.004) (0.004) (0.001) (0.004) (0.005) (0.004) (0.004) (0.004)

OS_CP -0.017*** -0.017*** -0.006*** -0.014*** -0.020*** -0.011*** -0.008** -0.017***

(Outgoing, stability, co- purchase)

(0.003) (0.003) (0.001) (0.003) (0.004) (0.003) (0.003) (0.003)

IAM_CV -0.016***

(Incoming, assortative mixing, co-view)

(0.004)

IAM_CP -0.002

(Incoming, assortative mixing, co-purchase)

(0.003)

OAM_CV 0.003

(Outgoing, assortative mixing, co-view)

(0.007)

OAM_CP -0.010**

(Outgoing, assortative mixing, co-purchase)

(0.004)

Constant 0.323*** 0.314*** 0.314*** 0.293*** 0.396*** 0.321*** 0.310*** 0.322***

(0.046) (0.046) (0.046) (0.046) (0.061) (0.046) (0.048) (0.046)

Control variables -included- -included- -included- -included- -included- -included- -included- -included-

Number of observations 41,379 41,379 41,379 41,379 29,172 40,931 40,030 41,379

R² 0.4934 0.4934 0.4934 0.4752 0.5034 0.4946 0.4959 0.4937

Notes: Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01.

418 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

regressive (AR1) disturbance structure. As indicated in Table 9, Column (4), the model estimates under an AR1 structure are generally consistent with those of the preferred one in Column (1).

Third, we check whether our findings are robust after ac- counting for price endogeneity. Specifically, we treat product list price (LP) as an endogeneous variable and use the instrumental variables estimation method. The choice of instrumental variable is the average list price of the same product on the previous day from three additional Tmall stores which also exclusively sell Nikon products. The price from another Nikon store would have high correlation with the price of the same product in the focal store, since retailers would typically set comparable prices for the same product. However, prices from another store on the previous day are unlikely to shift the current demand of products in the focal store. Hence, we believe price from alternative stores on the previous day could serve as a reasonable instrument. We perform RE two-stage least-squares estimation and Table 9, Column (5), summarizes the estimated model coefficients which are relatively consistent with those in Column (1).

Fourth, we check the robustness of our findings across dif- ferences in variable operationalizations. Specifically, instead of using the overlap of connections in a network across two immediate consecutive days as the measure for network stability, we use the overlap of connections today and three days ago as an alternative measure. Similar results based on this new measure are summarized in Table 9, Column (6). Interestingly, in terms of the magnitude of elasticity, the effect of OS_CP drops from 0.012 in the preferred model to 0.008. We further use the overlap of connections today and seven days ago as the stability measure and report the results in Table 9, Column (7). The effect of OS_CP further diminishes (elasticity magnitude = 0.005). This suggests that network stability based on a consecutive day-to-day basis has a larger effect on product demand, compared to that based on a longer time period.

Fifth, past research (e.g., Oestreicher-Singer and Sundararajan 2012b) has reported the impact of assortative mixing (i.e., the percentage of networked products that are from the same category with the focal product) on product demand. We provide a comparison between network diversity and assor- tative mixing by including the assortative mixing variables in our preferred model for estimation. The results in Table 9, Column (8), show that assortative mixing has a negative impact on product demand, which is consistent with prior research findings. More importantly, our significant network diversity estimates indicate that, in addition to the category similarity between the focal product and the networked prod- ucts (as reflected in assortative mixing), the difference among

those networked products themselves (as measured by net- work diversity) has distinct associations with product demand.

Sixth, we use a product’s daily market share within a category as an alternative measure for the dependent variable. The model estimates based on this new measure, which are sum- marized in Table 10, Column (2), report similar findings. Moreover, instead of using the average values of a variable on day t and day t+1 as the measure for this variable on day t, we use the actual day t values. Consistent results using this operationalization are shown in Table 10, Column (3).

Seventh, we also control for the potential effect of shopping seasonality on product demand by further including a dummy variable in our empirical model to differentiate between weekdays and non-working days (i.e., weekends and public holidays). The results summarized in Table 10, Column (4) are consistent with those in Column (1).

Eighth, we check the robustness of our findings across timing differences. We use the weekly time frame instead of the daily time frame by computing the average values across all days in each week for our model variables. The estimates based on the weekly time frame are reported in Table 10, Column (5). The findings under a weekly time frame are similar to those using a daily level of analysis.

Ninth, we check whether our findings are robust across other product categories. We collected data from another three Tmall stores that sell personal computers (PCs), mobile phones, and cosmetic products.36 The results in Table 10, Columns (6) to (8) report findings that are similar to those using the digital camera category. Interestingly, a comparison of the main estimates across all four categories shows that the demand effect of product networks is more salient (as evi- denced by the larger number of significant estimates, and the generally larger magnitudes of the significant estimates) for experience products (i.e., cosmetic products) than for search products (i.e., cameras, PCs, mobile phones). Unlike search products such as cameras, PCs, and mobile phones, which have highly standardized product attributes for consumers to evaluate the quality, experience products are harder to assess prior to purchase or consumption. Thus, we believe that as compared to the case of search products, consumers may rely more on recommendation systems for experience products such as cosmetic products in this research and books examined in prior studies (e.g., Oestreicher-Singer and Sundararajan 2012a, 2012b). Furthermore, we also conduct similar analyses based on all six subcategories (i.e., acces- sory, battery, compact camera, flash, lens, SLR camera) of the

36The appendix presents the descriptive statistics for these three product categories.

MIS Quarterly Vol. 41 No. 2/June 2017 419

Lin et al./Demand Effects of Product Recommendation Networks

Table 10. Robustness Checks (3)

Variable (1)

Preferred

(2) Market share

(3) Actual

day (4)

Seasonality (5)

Weekly (6) PC

(7) Mobile phone

(8) Cosmetic

ID_CV -0.001 -0.158 -0.003 -0.001 0.018 0.002 0.004 0.011**

(Incoming, diversity, co-view) (0.003) (0.107) (0.002) (0.003) (0.011) (0.003) (0.006) (0.004)

ID_CP 0.010*** 0.479*** 0.007*** 0.010*** 0.055*** 0.009*** 0.010* 0.016***

(Incoming, diversity, co- purchase)

(0.001) (0.060) (0.001) (0.001) (0.006) (0.003) (0.006) (0.003)

OD_CV -0.003 0.312** -0.001 -0.003 -0.048*** -0.006 -0.015* -0.029***

(Outgoing, diversity, co-view) (0.004) (0.152) (0.003) (0.004) (0.018) (0.007) (0.009) (0.005)

OD_CP -0.008*** -0.227*** -0.005*** -0.008*** -0.038*** -0.014*** -0.017*** -0.020***

(Outgoing, diversity, co- purchase)

(0.002) (0.068) (0.001) (0.002) (0.007) (0.005) (0.006) (0.003)

IS_CV 0.005 0.033 0.003 0.005 0.031 0.011 -0.009 0.004

(Incoming, stability, co-view) (0.003) (0.144) (0.004) (0.003) (0.021) (0.012) (0.012) (0.007)

IS_CP -0.000 0.075 0.001 -0.000 0.042** 0.016* 0.003 0.013**

(Incoming, stability, co- purchase)

(0.003) (0.119) (0.003) (0.003) (0.017) (0.009) (0.010) (0.006)

OS_CV 0.006 0.372** 0.006 0.006 0.057** 0.024* 0.016 0.044***

(Outgoing, stability, co-view) (0.004) (0.159) (0.004) (0.004) (0.023) (0.014) (0.014) (0.008)

OS_CP -0.017*** -0.413*** -0.018*** -0.017*** -0.108*** -0.028*** -0.021** -0.025***

(Outgoing, stability, co- purchase)

(0.003) (0.132) (0.004) (0.003) (0.020) (0.009) (0.010) (0.006)

Constant 0.323*** 6.491*** 0.312*** 0.071 -0.009 0.147** 0.251 0.007

(0.046) (1.919) (0.046) (0.075) (0.026) (0.059) (0.218) (0.080)

Control variables -included- -included- -included- -included- -included- -included- -included- -included-

Number of observations 41,379 41,379 41,379 41,379 7,800 15,616 17,195 53,061

R² 0.4934 0.4029 0.4937 0.4934 0.6508 0.4565 0.7366 0.4757

Notes: Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01.

main sample of digital cameras, and report the results in the appendix. The results show that the product network’s demand effect is more influential for the SLR camera sub- category, as compared to that in all the other subcategories. This may be because SLR cameras are mostly expensive products and have many relevant complementary products. Consumers thus tend to rely more on recommendation systems to carefully search for the best camera and ideal complements.

Finally, some may be concerned that network stability would be influenced by the time since it was available in the market (i.e., product vintage, VT) and by past sales (i.e., past monthly sales quantity, PS). We thus construct and include the inter- action terms between network stability and product vintage, and also the interaction terms between network stability and past monthly sales quantity. Even after accounting for such interaction effects, we find that all of the main focal estimates

of network diversity and stability remain consistent with our preferred model estimates (see the appendix for full results).

In summary, we are confident that all of the various checks indicate robustness and consistency of our findings in the presence of numerous control variables, potential simul- taneity, collinearity, serial correlation, price endogeneity, and across differences in variable operationalizations, time frames, product categories and data samples.

Discussion and Contribution

Discussion of Findings

Our study, which investigates the effect of network diversity and network stability on product demand in an e-commerce

420 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

setting, has several notable findings. First, we empirically show that the diversity and stability of a product’s network are associated with the sales quantity of the product. In terms of elasticities, a 1% increase in the category diversity of the incoming (outgoing) co-purchase network of a product is associated with a 0.011% (0.012%) increase (decrease) in the product’s sales quantity. Moreover, a 1% increase in the stability of the outgoing co-purchase network of a product is associated with a 0.012% decrease in sales quantity. In terms of absolute values, based on the mean sales quantity and network factors, a one unit increase in the category diversity of the incoming (outgoing) co-purchase network of a product is associated with a 0.0026 (0.0029) unit increase (decrease) in the product's sales quantity. A one unit increase in the stability of the outgoing co-purchase network of a product is associated with a 0.0047 unit decrease in sales quantity. These results show that, similar to product assortment management in the offline retail context (Borle et al. 2005), strategic arrangements of online product assortment through the use of recommendation systems also have significant demand effects.

Second, by operationalizing a product’s network into incoming network and outgoing network for investigation, we identify that the demand of a product is influenced by both the incoming and outgoing networks of the product. Our main results show that the incoming network influences product demand only through diversity effects whereas the outgoing network influences it through both diversity and stability effects. This implies that the outgoing network links of a product in an e-commerce site play a more influential role than those of the incoming network, which highlights the importance of taking into account both directions of links in product network research, and reveals the deficiency of prior studies that mostly focused only on the incoming network impact. With respect to the insignificant result of incoming network stability, this might be attributed to two competing influences which could have cancelled out the effects of each other. On one hand, instability of the incoming network (i.e., frequent changes) may heighten the actual or perceived dif- ficulties in product search efforts (Hoque and Lohse 1999) so as to reduce the likelihood of purchase of the focal product. On the other hand, instability of the incoming network may help to create an illusion of updates, which often introduce new and novel product offerings. By imbuing consumers with a sense of newness and novelty, this raises product awareness and overall satisfaction during consumers’ explora- tion process (Goldenberg et al. 2012), and eventually triggers higher product demand. These two antagonistic influences plausibly resulted in the insignificant results of incoming network stability.

Third, by further differentiating a product’s network into co- view and co-purchase networks, we surprisingly find that only the co-purchase network has a significant relation with product demand, even though co-purchase recommendations are listed right below those of co-view at the same bottom section of a webpage on Tmall.37 Specifically, the co- purchase network diversity and stability exhibit significant relations with product demand, whereas those for the co-view network are insignificant. Thus, we report that the co- purchase network has a more influential relation with product demand than the co-view network for both network diversity and stability. This indeed provides support to our arguments that co-view and co-purchase recommendations essentially present two different sets of information, and consumers do factor into consideration the source or type of information when making their purchase decisions, as suggested by prior research on consumer information seeking and processing (Rao and Sieben 1992; Westbrook and Fornell 1979). As a result, network effects on demand are only significant in the co-purchase network but not in the co-view network. A possible reason is, while co-view recommendations contain information about other consumers’ online product search behavior, searching, or browsing, a certain product does not always lead to the consumer’s eventual purchase of the product. Conversely, co-purchase product recommendations indicate other consumers’ actual purchase behavior. Thus, from the consumers’ perspective, observational learning through co-purchase information may be more salient and persuasive than co-view information in driving product demand.

Theoretical Contributions

Our study offers important theoretical contributions in the following ways. First, prior product network research has made attempts to analyze the economic impact (e.g., sales impact) of product network structures (Carmi et al. 2017; Oestreicher-Singer et al. 2013; Oestreicher-Singer and Sundararajan 2012a, 2012b). Unfortunately, past studies have not been able to incorporate and differentiate the demand impacts of incoming and outgoing networks. Our study advances existing product network literature by adopting a “bidirectional” view to validate the effect of both incoming and outgoing networks. For instance, the impact of network degree centrality is investigated in both our study and the

37This result shows that the order or location effect of co-view and co- purchase recommendations should not be an explanation of our findings here. Instead, consumers do seem to saliently distinguish between peers’ behavioral differences and/or information captured in actual purchases versus those of online views of products or webpage.

MIS Quarterly Vol. 41 No. 2/June 2017 421

Lin et al./Demand Effects of Product Recommendation Networks

study by Oestreicher-Singer and Sundararajan (2012b). Although both studies identify a similar negative effect of in- degree centrality, our estimation results, based on our pre- ferred model of both incoming and outgoing networks, find that out-degree centrality, OC_CV (elasticity = -0.035), actually has a stronger relation with demand compared to that of in-degree centrality (elasticity of IC_CV = -0.002, elasticity of IC_CP = -0.008). The above results and discussions of our main findings on network diversity and network stability effects have lent credence to the assertion that prior studies that focused exclusively on the incoming network in product network research may have derived incomplete or even erroneous conclusions. In essence, our more comprehensive examination sheds new insights and suggests a revisit of prior works and findings.

Second, various studies on product recommendation systems (Carmi et al. 2017; Oestreicher-Singer et al. 2013; Oestreicher-Singer and Sundararajan 2012b; Pathak et al. 2010) have identified some positive economic influences of product recommendation and reported its benefits (e.g., recommendation systems lead to sales increase). However, in addition to the positive effects, our study also unravels the potential negative economic effects of product recommen- dations resulting from outgoing co-purchase network diversity and stability effects. Our research findings thus challenge conventional wisdom by elucidating possible detrimental effects of recommendation systems on product demand.

Third, in addition to the centrality effect on demand identified in prior product network literature (Goldenberg et al. 2012; Oestreicher-Singer et al. 2013; Oestreicher-Singer and Sundararajan 2012a, 2012b), our research is the first to study and validate the diversity effect and stability effect of product networks. Of greater importance, by comparing the effects between the prior investigated degree centrality factors and our proposed diversity and stability factors, we further find that the effects of all the significant degree centrality factors (except OC_CV), including IC_CV (elasticity = -0.002) and IC_CP (elasticity = -0.008), are smaller than the effects of network diversity, ID_CP (elasticity = 0.011) and OD_CP (elasticity = -0.012), and network stability, OS_CP (elasticity = -0.012). Our findings thus suggest that network diversity and stability are more influential attributes than degree cen- trality in driving demand in product recommendation networks.

Moreover, by identifying the demand effects of product assortment management in an e-commerce setting, our study has also enriched prior marketing literature on product assortment management in the traditional retail context (e.g., Caro and Martínez-de-Albéeniz 2009; Hoch et al. 1999;

Kuksov and Villas-Boas 2010; Ton and Raman 2010). Additionally, we have made an important conceptual distinc- tion between network diversity and assortative mixing (Oestreicher-Singer and Sundararajan 2012a, 2012b), and empirically demonstrated their distinct effects on product demand.

Fourth, by juxtaposing the role of co-view recommendation networks beside that of co-purchase recommendation net- works, we unravel the contention and intricacies between the two. Our findings suggest that co-purchase networks are significantly related to product demand through diversity and stability effects, whereas the effects of co-view network diversity and stability on product demand are insignificant. In other words, the relations of network diversity and stability with demand are both stronger in the co-purchase networks, compared to their insignificant relations in the co-view networks. The differential and even contrasting results of these two types of networks suggest that consumers not only respond to the diversity and stability of product recommen- dations, but also factor the mechanism of recommendations into consideration. Our findings thus complement and enrich prior research to provide a more comprehensive under- standing of the economic effect of product networks.

Finally, our research is one of the rare studies to examine product network impacts using actual demand information from e-commerce retailers. Prior research only used some demand proxies such as sales rank on Amazon.com for empirical analysis. Although some studies have documented that “actual demand” can be obtained by a log-linear trans- formation of sales rank (Brynjolfsson et al. 2003; Goolsbee and Chevalier 2002), the precision of research findings and especially the quantitative insights may have already been compromised via this transformation. Our empirical analysis in this paper can thus provide more accurate and reliable conclusions by virtue of the accuracy of our demand data.

Practical Implications

Our study also provides important practical implications for e-commerce retailers to drive product sales using recommen- dation systems. First, retailers can take advantage of the diversity effect to drive product sales. According to our results, the category diversity in the incoming co-purchase network of a product can increase the product’s demand. Retailers could thus configure co-purchase recommendation systems to have more diverse incoming links for a product in terms of more heterogeneous product categories. Conversely, the category diversity in the outgoing co-purchase network has a negative relation with product demand. This suggests

422 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

that retailers should not blindly follow their intuition to recommend as many categories as possible on a product’s webpage in order to drive consumers’ cross-buying behavior. On the other hand, retailers are also not advised to limit the recommendations to just one product category as this might impede consumers’ product search and discovery.38 Thus, we suggest that retailers should maintain the category diversity of outgoing links in the co-purchase network at a strategic level. In fact, we suspect there may exist a potential trade-off between cross-selling opportunities and consumers’ product search experience.

Second, e-commerce retailers can also capitalize on the network stability effect to influence product demand. Speci- fically, the negative stability effect of the outgoing co- purchase network suggests that retailers should periodically change or refresh the constituent items in the outgoing co- purchase recommendation list in order to positively influence the demand of the focal product.

Third, comparing the effects between co-view and co- purchase recommendation systems, only co-purchase recom- mendation systems exhibit a significant effect in driving product demand. Accordingly, retailers could allocate a higher percentage of the limited webpage space to more saliently display co-purchase recommendations to achieve better sales performance on e-commerce websites.

Finally, our study also offers implications for the design of product recommendation systems. As existing product recommendation systems are mainly automated (i.e., auto- mated algorithm execution to generate recommendations) (Sarwar et al. 2000), e-commerce platform or retail site operators may provide interfaces and access to individual brand owners for them to implement the above suggested strategies to drive product sales online. Ideally, recommen- dation system designers could further develop and integrate these functionalities based on the above insights for retailers’ implementation.

Conclusion

While this research has highlighted several notable findings and important contributions, we acknowledge some limita- tions. First, while our empirical strategy attempts to control for various observed product attributes such as price, product reviews, and inventory, as well as potential simultaneity and unobserved implicit demand correlations through accounting for both substitution and complementarity effects of other

network-linked and unlinked products, these approaches may not have fully controlled for all potential sources of endo- geneity bias. We thus do not make absolute causality claims (despite our additional Granger causality test and shuffle test) of the impacts of network diversity and network stability on product demand since our sample dataset is, after all, an observational one from a single e-commerce website. How- ever, as argued by Sundararajan et al. (2013), the lack of a causal mechanism in network studies should not preclude the usefulness and contribution of predictive modeling based on correlation solely. Notwithstanding, we conduct many addi- tional checks and tests such as the Granger causality test and shuffle test whose results all attest to the econometric rigor of our findings and give some credence for a causal interpre- tation of findings. Second, due to the lack of information on Tmall, our product vintage measure is derived from Amazon. While it serves as a reasonable proxy, we are not able to absolutely rule out any potential biases introduced by using data from a different e-commerce site. For instance, a par- ticular product may be released earlier in the United States compared to China, and thus the product vintage variable may deviate from its true value on Tmall. However, we expect that this limitation due to product vintage deviations can be mitigated by our inclusion of time dummies and time-varying product review ratings in the empirical model. Third, our research findings are largely applicable to products that may induce multiple search or shopping visits by consumers before a final purchase (e.g., relatively expensive or high- involvement products such as cameras and PCs). For prod- ucts that do not need multiple search or shopping efforts, consumers may not be able to perceive the change (i.e., instability) of the product network, and thus the effect of network stability on demand may be reduced.

Moving forward, we present potential avenues for future research. Tmall is a platform-based e-commerce site which only provides an online marketplace to facilitate the business between retailers and consumers, without its own investments and operations in product procurement and shipping logistics. Thus, the results reported in this study may be more generali- zable to similar e-commerce platforms such as eBay. How- ever, some e-commerce sites, such as Amazon, are mainly self-operated throughout the entire e-commerce value chain (e.g., procuring, owning, and delivering most of the products offered on the platform via its own business units). Future studies could examine to what extent the research findings based on Tmall are applicable to other e-commerce sites with different business models. Another meaningful extension to this research is to investigate the spillover effects of online word-of-mouth (WOM). Due to the use of recommendation systems, WOM information of different products is now commonly connected in networks within a retail site. It would be interesting to study the demand impact of a prod- uct’s WOM on other connected products in the network. This38We greatly appreciate this insight from an anonymous reviewer.

MIS Quarterly Vol. 41 No. 2/June 2017 423

Lin et al./Demand Effects of Product Recommendation Networks

would have important implications for retailers’ e-commerce product colocation practices. Alternatively, it would be fruitful to examine the interplay between these automated recommendations (co-view and co-purchase) and the consumer-generated WOM recommendations. One could thus identify whether these two recommendation mechanisms are substitutable, complementary, or independent in driving product demand. The findings would have important implications for online retailers’ marketing strategies and the design of e-commerce websites.

Acknowledgments

We thank the senior editor, associate editor, the anonymous reviewers, the seminar and conference participants at the sixth China India Insights Conference 2014 and the Workshop on Analytics for Business, Consumer and Social Insights (BCSI 2013) for their valuable comments and suggestions. This research is partially supported by the National Natural Science Foundation of China, Project Grant 71502079 and the Singapore Ministry of Education, Project Grants R-253-000-097-112 and R-253-000-105-112.

References

Aciar, S., Zhang, D., Simoff, S., and Debenham, J. 2007. “Informed Recommender: Basing Recommendations on Consumer Product Reviews,” IEEE Intelligent Systems (22:3), pp. 39-47.

Adomavicius, G., and Tuzhilin, A. 2005. “Toward the Next Generation of Recommender Systems: A Survey of the State-of- the-Art and Possible Extensions,” IEEE Transactions on Knowledge and Data Engineering (17:6), pp. 734-749.

Alexa. 2015. “Top Sites in China,” http://www.alexa.com/topsites/ countries/CN; retrieved June 24, 2015.

Allenby, G. M., Arora, N., and Ginter, J. L. 1998. “On the Heterogeneity of Demand,” Journal of Marketing Research (35:3), pp. 384-389.

Anagnostopoulos, A., Kumar, R., and Mahdian, M. 2008. “Influence and Correlation in Social Networks,” in Proceedings of the 14th ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining, Las Vegas, NV, pp. 7-15.

Anderson, C. 2006. The Long Tail, New York: Hyperion Books. Barefoot, J. C., Grønbæk, M., Jensen, G., Schnohr, P., and

Prescott, E. 2005. “Social Network Diversity and Risks of Ischemic Heart Disease and Total Mortality: Findings from the Copenhagen City Heart Study,” American Journal of Epide- miology (161:10), pp. 960-967.

Baumol, W. J., and Ide, E. A. 1956. “Variety in Retailing,” Management Science (3:1), pp. 93-101.

Berlyne, D. E. 1960. Conflict, Arousal, and Curiosity, New York: McGraw-Hill.

Berry, S. T. 1994. “Estimating Discrete-Choice Models of Product Differentiation,” RAND Journal of Economics (25:2), pp. 242-262.

Borle, S., Boatwright, P., Kadane, J. B., Nunes, J. C., and Shmueli, G. 2005. “The Effect of Product Assortment Changes on Customer Retention,” Marketing Science (24:4), pp. 616-622.

Brin, S., and Page, L. 1998. “The Anatomy of a Large-Scale Hypertextual Web Search Engine,” Computer Networks and ISDN Systems (30:1-7), pp. 107-117.

Broniarczyk, S. M., Hoyer, W. D., and McAlister, L. 1998. “Consumers’ Perceptions of the Assortment Offered in a Grocery Category: The Impact of Item Reduction,” Journal of Marketing Research (35:2), pp. 166-176.

Brynjolfsson, E., Hu, Y. J., and Smith, M. D. 2003. “Consumer Surplus in the Digital Economy: Estimating the Value of Increased Product Variety at Online Booksellers,” Management Science (49:11), pp. 1580-1596.

Brynjolfsson, E., Hu, Y., and Smith, M. D. 2006. “From Niches to Riches: Anatomy of the Long Tail,” Sloan Management Review (47:4), pp. 67-71.

Burt, R. S. 2000. “The Network Structure of Social Capital,” in Research in Organizational Behavior (Vol. 22), B. Staw and R. Sutton (eds.), New York: JAI Press, pp. 345-423.

Carmi, E., Oestreicher-Singer, G., Stettner, U., and Sundararajan, A. 2017. “Is Oprah Contagious? The Depth of Diffusion of Demand Shocks in a Product Market,” MIS Quarterly (41), forthcoming.

Caro, F., and Martínez-de-Albéeniz, V. 2009. “The Effect of Assortment Rotation on Consumer Choice and Its Impact on Competition,” in Consumer-Driven Demand and Operations Management Models, New York: Springer, pp. 63-79.

Cross, R., and Cummings, J. N. 2004. “Tie and Network Correlates of Individual Performance in Knowledge-Intensive Work,” Academy of Management Journal (47:6), pp. 928-937.

Cummings, J. N. 2004. “Work Groups, Structural Diversity, and Knowledge Sharing in a Global Organization,” Management Science (50:3), pp. 352-364.

Cummings, J. N., and Higgins, M. C. 2006. “Relational Instability at the Network Core: Support Dynamics in Developmental Net- works,” Social Networks (28:1), pp. 38-55.

Darlington, R. B. 1990. Regression and Linear Models, New York: McGraw-Hill.

Dellarocas, C., Katona, Z., and Rand, W. 2010. “Media, Aggre- gators and the Link Economy: Strategic Hyperlink Formation in Content Networks,” NET Institute Working Paper #10-13.

Deng, L., and Poole, M. S. 2010. “Affect in Web Interfaces: A Study of the Impacts of Web Page Visual Complexity and Order,” MIS Quarterly (34:4), pp. 711-730.

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.

Eagle, N., Macy, M., and Claxton, R. 2010. “Network Diversity and Economic Development,” Science (328:5981), pp. 1029-1031.

eMarketer. 2014. “Retail Sales Worldwide Will Top $22 Trillion This Year,” http://www.emarketer.com/Article/Retail-Sales- Worldwide-Will-Top-22-Trillion-This-Year/1011765; retrieved June 24, 2015.

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.

424 MIS Quarterly Vol. 41 No. 2/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Ghose, A., Han, S., and Iyengar, R. 2011. “Network Stability and Social Contagion on the Mobile Internet,” in Proceedings of the 32nd International Conference on Information Systems, Shanghai.

Goh, K.-Y., Heng, C.-S., and Lin, Z. 2013. “Social Media Brand Community and Consumer Behavior: Quantifying the Relative Impact of User- and Marketer-Generated Content,” Information Systems Research (24:1), pp. 88-107.

Goldenberg, J., Oestreicher-Singer, G., and Reichman, S. 2012. “The Quest for Content: How User-Generated Links Can Facilitate Online Exploration,” Journal of Marketing Research (49:4), pp. 452-468.

Goolsbee, A., and Chevalier, J. A. 2002. “Measuring Prices and Price Competition Online: Amazon and Barnes and Noble,” Quantitative Marketing and Economics (1:2), pp. 203-222.

Granger, C. W. J. 1969. “Investigating Causal Relations by Econo- metric Models and Cross-Spectral Methods,” Econometrica (37:3), pp. 424-438.

H&M. 2007. “Annual Report 2007,” http://about.hm.com/content/ dam/hm/about/documents/en/Annual%20Report/Annual_ Report_2007_en.pdf; retrieved June 24, 2015.

Henderson, J. M., and Quandt, R. E. 1958. Microeconomic Theory: A Mathematical Approach, New York: McGraw-Hill Book.

Herlocker, J. L., Konstan, J. A., Terveen, L. G., and Riedl, J. T. 2004. “Evaluating Collaborative Filtering Recommender Sys- tems,” ACM Transactions on Information Systems (22:1), pp. 5-53.

Hinz, O., Schulze, C., and Takac, C. 2014. “New Product Adoption in Social Networks: Why Direction Matters,” Journal of Busi- ness Research (67:1), pp. 2836-2844.

Hoch, S. J., Bradlow, E. T., and Wansink, B. 1999. “The Variety of an Assortment,” Marketing Science (18:4), pp. 527-546.

Hoque, A. Y., and Lohse, G. L. 1999. “An Information Search Cost Perspective for Designing Interfaces for Electronic Commerce,” Journal of Marketing Research (36:3), pp. 387-394.

Huang, Z., Zeng, D. D., and Chen, H. 2007. “Analyzing Consumer-Product Graphs: Empirical Findings and Applications in Recommender Systems,” Management Science (53:7), pp. 1146-1164.

Iaquinta, L., de Gemmis, M., Lops, P., Semeraro, G., Filannino, M., and Molino, P. 2008. “Introducing Serendipity in a Content- Based Recommender System,” in Proceedings of the 8th

International Conference on Hybrid Intelligent Systems, Barcelona, Spain, pp. 168-173.

Jehn, K. A., Northcraft, G. B., and Neale, M. A. 1999. “Why Differences Make a Difference: A Field Study of Diversity, Conflict and Performance in Workgroups,” Administrative Science Quarterly (44:4), pp. 741-763.

Johnson, E. J., Bellman, S., and Lohse, G. L. 2003. “Cognitive Lock-in and the Power Law of Practice,” Journal of Marketing (67:2), pp. 62-75.

Kamakura, W. A., Kim, B. D., and Lee, J. 1996. “Modeling Preference and Structural Heterogeneity in Consumer Choice,” Marketing Science (15:2), pp. 152-172.

Katona, Z., and Sarvary, M. 2008. “Network Formation and the Structure of the Commercial World Wide Web,” Marketing Science (27:5), pp. 764-778.

Kuksov, D., and Villas-Boas, J. M. 2010. “When More Alterna- tives Lead to Less Choice,” Marketing Science (29:3), pp. 507-524.

Leeflang, P. S. H., and Parreño-Selva, J. 2012. “Cross-Category Demand Effects of Price Promotions,” Journal of the Academy of Marketing Science (40:4), pp. 1-15.

Manchanda, P., Ansari, A., and Gupta, S. 1999. “The ‘Shopping Basket’: A Model for Multicategory Purchase Incidence Decisions,” Marketing Science (18:2), pp. 95-114.

Mayzlin, D., and Yoganarasimhan, H. 2012. “Link to Success: How Blogs Build an Audience by Promoting Rivals,” Management Science (58:9), pp. 1651-1668.

Menon, S., and Kahn, B. E. 1995. “The Impact of Context on Variety Seeking in Product Choices,” Journal of Consumer Research (22:3), pp. 285-295.

Moe, W. W., and Fader, P. S. 2004. “Dynamic Conversion Behavior at E-Commerce Sites,” Management Science (50:3), pp. 326-335.

Moe, W. W., and Yang, S. 2009. “Inertial Disruption: The Impact of a New Competitive Entrant on Online Consumer Search,” Journal of Marketing (73:1), pp. 109-121.

Mooney, R. J., and Roy, L. 2000. “Content-Based Book Recom- mending Using Learning for Text Categorization,” in Proceedings of the 5th ACM Conference on Digital Libraries, San Antonio, TX, pp. 195-204.

Mulhern, F. J., and Leone, R. P. 1991. “Implicit Price Bundling of Retail Products: A Multiproduct Approach to Maximizing Store Profitability,” Journal of Marketing (55:4), pp. 63-76.

Nadkarni, S., and Gupta, R. 2007. “A Task-Based Model of Perceived Website Complexity,” MIS Quarterly (31:3), pp. 501-524.

Nelson, P. 1970. “Information and Consumer Behavior,” Journal of Political Economy (78:2), pp. 311-329.

Niraj, R., Padmanabhan, V., and Seetharaman, P. B. 2008. “A Cross-Category Model of Households’ Incidence and Quantity Decisions,” Marketing Science (27:2), pp. 225-235.

Oestreicher-Singer, G., Libai, B., Sivan, L., Carmi, E., and Yassin, O. 2013. “The Network Value of Products,” Journal of Marketing (77:3), pp. 1-14.

Oestreicher-Singer, G., and Sundararajan, A. 2012a. “Recommen- dation Networks and the Long Tail of Electronic Commerce,” MIS Quarterly (36:1), pp. 65-83.

Oestreicher-Singer, G., and Sundararajan, A. 2012b. “The Visible Hand? Demand Effects of Recommendation Networks in Elec- tronic Markets,” Management Science (58:11), pp. 1963-1981.

Osborne, M. 2011. “Consumer Learning, Switching Costs, and Heterogeneity: A Structural Examination,” Quantitative Mar- keting and Economics (9:1), pp. 25-70.

Pathak, B., Garfinkel, R., Gopal, R. D., Venkatesan, R., and Yin, F. 2010. “Empirical Analysis of the Impact of Recommender Systems on Sales,” Journal of Management Information Systems (27:2), pp. 159-188.

Punj, G. N., and Staelin, R. 1983. “A Model of Consumer Infor- mation Search Behavior for New Automobiles,” Journal of Consumer Research (9:4), pp. 366-380.

Rao, A. R., and Sieben, W. A. 1992. “The Effect of Prior Knowl- edge on Price Acceptability and the Type of Information Examined,” Journal of Consumer Research (19:2), pp. 256-270.

MIS Quarterly Vol. 41 No. 2/June 2017 425

Lin et al./Demand Effects of Product Recommendation Networks

Reagans, R., and Zuckerman, E. W. 2001. “Networks, Diversity, and Productivity: The Social Capital of Corporate R&D Teams,” Organization Science (12:4), pp. 502-517.

Reibstein, D. J., Youngblood, S. A., and Fromkin, H. L. 1975. “Number of Choices and Perceived Decision Freedom as a Determinant of Satisfaction and Consumer Behavior,” Journal of Applied Psychology (60:4), pp. 434-437.

Roberts, J. H., and Lattin, J. M. 1991. “Development and Testing of a Model of Consideration Set Composition,” Journal of Marketing Research (28:4), pp. 429-440.

Rooderkerk, R. P., van Heerde, H. J., and Bijmolt, T. H. 2013. “Optimizing Retail Assortments,” Marketing Science (32:5), pp. 699-715.

Sahoo, N., Krishnan, R., Duncan, G., and Callan, J. 2011. “The Halo Effect in Multicomponent Ratings and its Implications for Recommender Systems: The Case of Yahoo! Movies,” Informa- tion Systems Research (23:1), pp. 231-246.

Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. 2000. “Analysis of Recommendation Algorithms for E-Commerce,” in Proceedings of the Second ACM Conference on Electronic Com- merce, Minneapolis, MN, pp. 158-167.

Schafer, J. B., Konstan, J., and Riedi, J. 1999. “Recommender Systems in E-commerce,” in Proceedings of the First ACM Conference on Electronic Commerce, Denver, CO, pp. 158-166.

Schafer, J. B., Konstan, J. A., and Riedl, J. 2001. “E-Commerce Recommendation Applications,” Data Mining and Knowledge Discovery (5:1), pp. 115-153.

Shugan, S. M. 1980. “The Cost of Thinking,” Journal of Consumer Research (7:2), pp. 99-111.

Song, I., and Chintagunta, P. K. 2006. “Measuring Cross-Category Price Effects with Aggregate Store Data,” Management Science (52:10), pp. 1594-1609.

Srikant, R., and Agrawal, R. 1996. “Mining Quantitative Associa- tion Rules in Large Relational Tables,” in Proceedings of the ACM SIGMOD International Conference on Management of Data, New York, pp. 1-12.

Sundararajan, A., Provost, F., Oestreicher-Singer, G., and Aral, S. 2013. “Information in Digital, Economic, and Social Networks,” Information Systems Research (24:4), pp. 883-905.

Tmall. 2015. “About Us,” http://news.alibaba.com/specials/ aboutalibaba/aligroup/index.html; retrieved June 24, 2015.

Ton, Z., and Raman, A. 2010. “The Effect of Product Variety and Inventory Levels on Retail Store Sales: A Longitudinal Study,” Production and Operations Management (19:5), pp. 546-560.

Tucker, C. 2011. “Network Stability, Network Externalities and Technology Adoption,” Working Paper 17246, National Bureau of Economic Research.

Van Trijp, H. C. M., Hoyer, W. D., and Inman, J. J. 1996. “Why Switch? Product Category: Level Explanations for True Variety-Seeking Behavior,” Journal of Marketing Research (33:3), pp. 281-292.

Walters, R. G. 1991. “Assessing the Impact of Retail Price Promo- tions on Product Substitution, Complementary Purchase, and Interstore Sales Displacement,” Journal of Marketing (55:2), pp. 17-28.

Westbrook, R. A., and Fornell, C. 1979. “Patterns of Information Source Usage among Durable Goods Buyers,” Journal of Marketing Research (16:3), pp. 303-312.

Wooldridge, J. M. 2006. Introductory Econometrics: A Modern Approach (3rd ed.). Mason, OH: Thomson Southwestern.

Yang, S., and Allenby, G. M. 2003. “Modeling Interdependent Consumer Preferences,” Journal of Marketing Research (40:3), pp. 282-294.

Yi, P., Wenli, Z., and Gang, K. 2011. “A Hybrid Movie Recom- mender System,” paper presented at the 23rd Workshop on Information Systems and Economics, Shanghai.

Yoganarasimhan, H. 2012. “Impact of Social Network Structure on Content Propagation: A Study Using Youtube Data,” Quantit- ative Marketing and Economics (10:1), pp. 111-150.

About the Authors

Zhijie Lin is an assistant professor in the Department of Marketing and Electronic Business, School of Business at the Nanjing University. He received his Ph.D. degree in Information Systems from the National University of Singapore. His research interests focus on electronic commerce and social media platforms. He has published papers in journals such as Information Systems Research, Journal of Management Information Systems, and Decision Support Systems, as well as in conferences such as the International Conference on Information Systems.

Khim-Yong Goh is an associate professor of Information Systems and Vice Dean (Corporate Relations) in the School of Computing at the National University of Singapore. He received his Ph.D. degree in Business Administration (Marketing: Economics and Quantitative Methods) from the University of Chicago, Booth School of Business. His research interests focus on marketing and advertising in digital media platforms, consumer and firm behaviors in markets with network and social interaction effects, competitive product, pricing, and promotional strategies in IT-mediated markets. He has published papers in journals such as Management Science, Journal of Marketing Research, Information Systems Research, IEEE Transactions on Engineering Management, Journal of the Association for Information Systems, and in conferences such as the International Conference on Information Systems.

Cheng-Suang Heng is an associate professor of Information Systems in the School of Computing at the National University of Singapore. He received his Ph.D. degree in Organization, Technol- ogy and Entrepreneurship from Stanford University’s Management Science and Engineering Department. His research interests focus on organization strategies, with emphasis on e-commerce and social media. He has published papers in journals such as Information Systems Research, Journal of Management Information Systems, Journal of the Association for Information Systems, and in confer- ences such as the International Conference on Information Systems.

426 MIS Quarterly Vol. 41 No. 2/June 2017

RESEARCH ARTICLE

THE DEMAND EFFECTS OF PRODUCT RECOMMENDATION NETWORKS: AN EMPIRICAL ANALYSIS OF NETWORK

DIVERSITY AND STABILITY Zhijie Lin

School of Business, Nanjing University, 22 Hankou Road,

Nanjing 210093, CHINA {[email protected]}

Khim-Yong Goh and Cheng-Suang Heng School of Computing, National University of Singapore, 15 Computing Drive,

Singapore 117418 SINGAPORE {[email protected]} {[email protected]}

Appendix

Product Network Structures and Metrics

Figures A1 and A2 present the co-view and co-purchase networks at the midpoint (September 1, 2012) of the sample period. Tables A1 and A2 summarize the corresponding network metrics.

MIS Quarterly Vol. 41 No. 2—Appendices/June 2017 A1

Lin et al./Demand Effects of Product Recommendation Networks

Figure A1. Co-View Network Structure

Table A1. Co-View Network Metrics

Metric Mean Std. Dev. Min Max

In-degree centrality 4.000 5.597 0.000 37.000

Out-degree centrality 4.000 0.000 4.000 4.000

Betweenness centrality 565.644 1,549.084 0.000 17,502.777

Closeness centrality 0.009 0.029 0.001 0.143

Eigenvector centrality 0.004 0.012 0.000 0.074

PageRank 1.000 0.634 0.587 5.365

Clustering coefficient 0.452 0.300 0.000 1.000

Notes: Number of nodes = 225; Number of edges = 900; Graph density = 0.018.

A2 MIS Quarterly Vol. 41 No. 2—Appendices/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Figure A2. Co-Purchase Network Structure

Table A2. Co-Purchase Network Metrics

Metric Mean Std. Dev. Min Max

In-degree centrality 4.942 9.298 0.000 72.000

Out-degree centrality 4.942 0.269 3.000 5.000

Betweenness centrality 386.347 1,098.046 0.000 9,634.566

Closeness centrality 0.002 0.000 0.001 0.002

Eigenvector centrality 0.004 0.004 0.000 0.034

PageRank 1.000 0.879 0.509 7.415

Clustering coefficient 0.174 0.156 0.000 0.650

Notes: Number of nodes = 225; Number of edges = 1,125; Graph density = 0.022.

MIS Quarterly Vol. 41 No. 2—Appendices/June 2017 A3

Lin et al./Demand Effects of Product Recommendation Networks

Descriptive Statistics (Product Category)

Table A3 presents the descriptive statistics of product categories included in our main sample of digital cameras.

Table A3. Descriptive Statistics (Product Category)

Variable Mean Std. Dev. Min Max

QUAN1 (Sales quantity: accessory) 33.863 112.720 0.000 2,200.000

QUAN2 (Sales quantity: battery) 3.349 5.898 0.000 36.000

QUAN3 (Sales quantity: compact camera) 2.313 4.967 0.000 50.000

QUAN4 (Sales quantity: flash) 0.056 0.275 0.000 2.000

QUAN5 (Sales quantity: lens) 1.019 2.105 0.000 26.000

QUAN6 (Sales quantity: SLR camera) 4.514 6.857 0.000 73.000

LP1 (List price: accessory) 302.980 355.428 1.000 2,550.000

LP2 (List price: battery) 282.417 139.319 48.000 978.000

LP3 (List price: compact camera) 1,319.572 389.374 376.000 2,770.000

LP4 (List price: flash) 2,572.396 386.164 848.000 3,578.000

LP5 (List price: lens) 6,551.921 3,943.045 499.000 13,024.500

LP6 (List price: SLR camera) 6,396.445 4,291.084 2,546.000 15,087.420

PV1 (Review volume: accessory) 18.502 19.071 0.000 102.500

PV2 (Review volume: battery) 79.662 157.054 0.000 639.500

PV3 (Review volume: compact camera) 30.967 54.525 0.000 275.500

PV4 (Review volume: flash) 1.000 0.989 0.000 6.000

PV5 (Review volume: lens) 7.964 18.202 0.000 147.000

PV6 (Review volume: SLR camera) 179.917 179.716 0.000 634.000

PR1 (Review rating: accessory) 3.463 1.770 0.000 5.000

PR2 (Review rating: battery) 3.800 1.145 0.000 5.000

PR3 (Review rating: compact camera) 3.323 1.380 0.000 5.000

PR4 (Review rating: flash) 1.991 1.756 0.000 5.000

PR5 (Review rating: lens) 2.691 1.725 0.000 5.000

PR6 (Review rating: SLR camera) 3.710 1.480 0.000 5.000

PS1 (Past monthly sales quantity: accessory) 19.956 22.639 0.000 100.000

PS2 (Past monthly sales quantity: battery) 32.896 60.624 0.000 311.933

PS3 (Past monthly sales quantity: compact camera) 12.070 20.042 0.000 92.833

PS4 (Past monthly sales quantity: flash) 0.426 0.856 0.000 7.500

PS5 (Past monthly sales quantity: lens) 3.206 5.577 0.000 35.000

PS6 (Past monthly sales quantity: SLR camera) 63.832 72.856 0.000 314.000

IN1 (Inventory: accessory) 2,867.024 4,661.136 8.500 38,857.000

IN2 (Inventory: battery) 259.014 269.346 1.500 1,290.500

IN3 (Inventory: compact camera) 523.199 504.865 7.000 3,511.500

IN4 (Inventory: flash) 209.769 306.949 4.500 2,994.500

IN5 (Inventory: lens) 66.008 97.170 2.500 612.500

IN6 (Inventory: SLR camera) 490.337 583.760 1.000 7,053.500

BM1 (Number of bookmarks: accessory) 23.005 32.749 0.000 185.213

BM2 (Number of bookmarks: battery) 152.111 286.563 0.000 1,248.500

BM3 (Number of bookmarks: compact camera) 118.589 158.925 1.000 737.500

BM4 (Number of bookmarks: flash) 15.569 8.891 4.000 76.500

BM5 (Number of bookmarks: lens) 66.406 93.617 1.000 514.000

BM6 (Number of bookmarks: SLR camera) 4,105.217 3,733.216 0.000 12,114.000

Note: Number of observations = 533.

A4 MIS Quarterly Vol. 41 No. 2—Appendices/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Model Estimation Results (Product Category)

Table A4 presents the product category level model estimation results. The Hausman test results suggest that a random effects model should be chosen over a fixed effects model for all the six categories.

Table A4. Model Estimation Results (Product Category)

Variable

(1) Accessory

(2) Battery

(3) Compact camera

(4) Flash

(5) Lens

(6) SLR

camera

Fixed Effects

Random Effects

Fixed Effects

Random Effects

Fixed Effects

Random Effects

Fixed Effects

Random Effects

Fixed Effects

Random Effects

Fixed Effects

Random Effects

LP1 -0.249 -0.420*** -0.137 -0.294** -0.144* -0.180*** 0.087** 0.142*** 0.227** 0.234*** 0.186* 0.021

(List price: accessory)

(0.164) (0.155) (0.089) (0.135) (0.055) (0.028) (0.023) (0.028) (0.066) (0.053) (0.076) (0.083)

LP2 0.728** 0.867*** -0.766 -0.577* 0.115 0.112* 0.085 0.018 0.568*** 0.534*** 0.377 0.429*

(List price: battery) (0.201) (0.269) (0.373) (0.334) (0.051) (0.062) (0.057) (0.043) (0.045) (0.041) (0.294) (0.259)

LP3 -0.166 -0.159 -0.136 -0.139*** -0.356** -0.355*** 0.036 0.038** -0.036 -0.033 -0.037 -0.026

(List price: compact camera)

(0.203) (0.241) (0.059) (0.053) (0.105) (0.102) (0.015) (0.017) (0.023) (0.023) (0.028) (0.018)

LP4 -0.060 -0.310* -0.026 -0.126** -0.214*** -0.132*** -0.272* -0.273*** 0.005 -0.006 0.001 -0.056

(List price: flash) (0.159) (0.178) (0.090) (0.059) (0.025) (0.036) (0.105) (0.030) (0.070) (0.058) (0.092) (0.048)

LP5 0.028** 0.028*** -0.002 -0.002 0.005** 0.006** -0.000 -0.001 -0.127*** -0.127*** -0.001 0.002

(List price: lens) (0.007) (0.004) (0.007) (0.008) (0.001) (0.003) (0.001) (0.001) (0.009) (0.009) (0.012) (0.012)

LP6 0.029 0.046 0.021* 0.037 0.001 -0.013 -0.000 -0.003 0.019 0.015 -0.124** -0.138***

(List price: SLR camera)

(0.040) (0.051) (0.009) (0.027) (0.026) (0.023) (0.005) (0.003) (0.030) (0.024) (0.038) (0.042)

Control variables -included- -included- -included- -included- -included- -included-

Number of observations

533 533 533 533 533 533

Hausman test χ2 = 17.66,

p = 1.00 χ2 = 29.34,

p = 1.00 χ2 = 7.56, p = 1.00

χ2 = 73.57, p = 1.00

χ2 = 2.13, p = 1.00

χ2 = 17.52, p = 1.00

R² 0.7670 0.9123 0.5415 0.8889 0.7822 0.8674 0.3349 0.7551 0.8160 0.8663 0.8610 0.9035

Notes: Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01. All the six price variables were divided by 1,000 before estimations. For brevity, only price estimates are reported.

Granger Causality Tests

Table A5 presents the results of Granger causality tests based on different time lag levels from 1 day to 7 days.

Table A5. Granger Causality Test

No. of

lags

ID_CV ID_CP OD_CV OD_CP IS_CV IS_CP OS_CV OS_CP

χ2 p χ2 p χ2 p χ2 p χ2 p χ2 p χ2 p χ2 p 1 0.638 0.425 0.024 0.876 2.585 0.108 0.511 0.475 2.359 0.125 2.731 0.098 2.404 0.121 3.054 0.081

2 2.902 0.234 0.653 0.722 3.159 0.206 0.570 0.752 3.130 0.209 3.814 0.148 3.335 0.189 4.997 0.082

3 1.934 0.586 0.977 0.807 1.658 0.646 1.392 0.707 4.759 0.190 4.980 0.173 4.240 0.237 6.169 0.104

4 2.253 0.689 2.715 0.607 3.913 0.418 4.265 0.371 8.189 0.085 7.910 0.095 4.472 0.346 6.735 0.151

5 3.626 0.604 4.631 0.463 4.154 0.528 4.011 0.548 7.954 0.159 8.630 0.125 4.349 0.500 7.496 0.186

6 3.623 0.728 5.855 0.440 3.511 0.742 4.183 0.652 7.534 0.274 8.785 0.186 5.460 0.486 8.700 0.191

7 3.700 0.814 7.405 0.388 3.633 0.821 4.300 0.745 8.251 0.311 10.569 0.159 5.640 0.582 9.051 0.249

MIS Quarterly Vol. 41 No. 2—Appendices/June 2017 A5

Lin et al./Demand Effects of Product Recommendation Networks

Shuffle Tests

There are three major steps for the shuffle test. First, we compute all the diversity and stability variables based on our main model and operationalization. Thus, we will have a set of diversity and stability values for each product in each time period before shuffling. Second, for each focal product, we then randomly shuffle those diversity and stability values over all its time periods while keeping all the other measures (e.g., degree centrality, price, etc.) unchanged. Thus, we will have a randomized set of diversity and stability values for each product in each time period after shuffling. We note that this step is equivalent to randomly shuffling different sets of connected links (i.e., with no change in any single link within a set) for a focal product over various time periods, and keeping the control variables unchanged in a time period. Finally, we estimate Equation (1) separately based on the above two sets of before-shuffling and after-shuffling measures, and perform the Chow test to test for structural differences across these two sets of model parameters before and after shuffling.

Table A6 presents the results of a set of shuffle tests based on different seed values of 1, 2, 3, 4, and 5 used to generate a sequence of random numbers for shuffling. The five sets of Chow test results all indicate that the estimated parameters before and after shuffling are significantly different, thus providing some evidence for the causal effect of network diversity and stability on product demand.

Table A6. Shuffle Test

Variable

(1) Before

shuffling (Preferred)

(2) After shuffling

(Seed = 1)

(3) After shuffling

(Seed = 2)

(4) After shuffling

(Seed = 3)

(5) After shuffling

(Seed = 4)

(6) After shuffling

(Seed = 5)

ID_CV -0.001 0.002 0.007*** 0.004 0.007*** 0.002

(Incoming, diversity, co- view)

(0.003) (0.002) (0.002) (0.002) (0.002) (0.002)

ID_CP 0.010*** 0.008*** 0.010*** 0.006*** 0.009*** 0.006***

(Incoming, diversity, co- purchase)

(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)

OD_CV -0.003 -0.007*** -0.012*** -0.010*** -0.009*** -0.007***

(Outgoing, diversity, co- view)

(0.004) (0.003) (0.003) (0.003) (0.003) (0.003)

OD_CP -0.008*** -0.003** -0.003*** -0.002** -0.002** -0.001

(Outgoing, diversity, co- purchase)

(0.002) (0.001) (0.001) (0.001) (0.001) (0.001)

IS_CV 0.005 0.006* 0.008** 0.004 0.005 0.006*

(Incoming, stability, co- view)

(0.003) (0.003) (0.003) (0.003) (0.003) (0.003)

IS_CP -0.000 0.004 0.009*** 0.004 0.005* 0.007**

(Incoming, stability, co- purchase)

(0.003) (0.003) (0.003) (0.003) (0.003) (0.003)

OS_CV 0.006 -0.003 -0.001 -0.006 -0.000 -0.004

(Outgoing, stability, co- view)

(0.004) (0.004) (0.004) (0.004) (0.004) (0.004)

OS_CP -0.017*** -0.003 -0.007** -0.001 -0.004 -0.003

(Outgoing, stability, co- purchase)

(0.003) (0.003) (0.003) (0.003) (0.003) (0.003)

Constant 0.323*** 0.324*** 0.322*** 0.327*** 0.322*** 0.323***

(0.046) (0.046) (0.046) (0.046) (0.046) (0.046)

Control variables -included- -included- -included- -included- -included- -included-

Number of observations 41,379 41,379 41,379 41,379 41,379 41,379

Chow test F = 1.99, p = 0.04

F = 2.89, p = 0.00

F = 3.09, p = 0.00

F = 2.64, p = 0.01

F = 2.67, p = 0.01

R² 0.4934 0.4931 0.4939 0.4930 0.4935 0.4929

Notes: Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01.

A6 MIS Quarterly Vol. 41 No. 2—Appendices/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Descriptive Statistics (Personal Computers, Mobile Phones, Cosmetic Products)

Tables A7 to A9 present the major descriptive statistics for product categories of personal computers, mobile phones and cosmetic products, respectively.

Table A7. Descriptive Statistics (Personal Computers)

Variable Mean Std. Dev. Min Max

QUAN (Sales quantity) 1.085 31.368 0.000 2,206.000

ID_CV (Incoming network diversity, co-view) 1.361 1.619 0.000 15.000

ID_CP (Incoming network diversity, co-purchase) 1.673 2.191 0.000 17.500

OD_CV (Outgoing network diversity, co-view) 1.345 0.682 0.000 4.000

OD_CP (Outgoing network diversity, co-purchase) 1.557 0.914 0.000 5.000

IS_CV (Incoming network stability, co-view) 0.864 0.239 0.000 1.000

IS_CP (Incoming network stability, co-purchase) 0.718 0.344 0.000 1.000

OS_CV (Outgoing network stability, co-view) 0.842 0.238 0.000 1.000

OS_CP (Outgoing network stability, co-purchase) 0.719 0.354 0.000 1.000

LP (Product list price) 6,173.236 5,915.509 1.000 30,999.000

PV (Product review volume) 28.040 131.437 0.000 1,695.500

PR (Product review rating) 3.067 2.284 0.000 5.000

PS (Product past monthly sales quantity) 9.196 32.995 0.000 486.500

IN (Product inventory) 602.221 7,352.986 1.000 99,994.500

BM (Number of product webpage bookmarks) 185.879 767.755 0.000 10,972.500

Notes: Number of observations = 15,616; Number of products = 228; Number of days = 142.

Table A8. Descriptive Statistics (Mobile Phones)

Variable Mean Std. Dev. Min Max

QUAN (Sales quantity) 3.799 15.923 0.000 482.000

ID_CV (Incoming network diversity, co-view) 1.215 0.910 0.000 7.500

ID_CP (Incoming network diversity, co-purchase) 1.408 1.198 0.000 8.000

OD_CV (Outgoing network diversity, co-view) 1.332 0.558 0.000 4.000

OD_CP (Outgoing network diversity, co-purchase) 1.500 0.758 0.000 4.500

IS_CV (Incoming network stability, co-view) 0.823 0.270 0.000 1.000

IS_CP (Incoming network stability, co-purchase) 0.708 0.331 0.000 1.000

OS_CV (Outgoing network stability, co-view) 0.799 0.283 0.000 1.000

OS_CP (Outgoing network stability, co-purchase) 0.703 0.366 0.000 1.000

LP (Product list price) 749.580 976.975 1.000 5,280.000

PV (Product review volume) 230.909 960.165 0.000 8,211.500

PR (Product review rating) 2.729 2.348 0.000 5.000

PS (Product past monthly sales quantity) 101.751 384.877 0.000 3,100.000

IN (Product inventory) 927.980 8,523.731 1.000 99,999.500

BM (Number of product webpage bookmarks) 607.238 1,858.128 0.000 13,447.000

Notes: Number of observations = 17,195; Number of products = 238; Number of days = 138.

MIS Quarterly Vol. 41 No. 2—Appendices/June 2017 A7

Lin et al./Demand Effects of Product Recommendation Networks

Table A9. Descriptive Statistics (Cosmetic Products)

Variable Mean Std. Dev. Min Max

QUAN (Sales quantity) 1.107 6.597 0.000 272.000

ID_CV (Incoming network diversity, co-view) 1.153 0.763 0.000 8.000

ID_CP (Incoming network diversity, co-purchase) 1.529 1.171 0.000 9.000

OD_CV (Outgoing network diversity, co-view) 1.364 0.604 0.000 4.000

OD_CP (Outgoing network diversity, co-purchase) 1.856 0.923 0.000 5.000

IS_CV (Incoming network stability, co-view) 0.783 0.294 0.000 1.000

IS_CP (Incoming network stability, co-purchase) 0.679 0.344 0.000 1.000

OS_CV (Outgoing network stability, co-view) 0.747 0.314 0.000 1.000

OS_CP (Outgoing network stability, co-purchase) 0.657 0.393 0.000 1.000

LP (Product list price) 65.734 53.465 1.000 499.500

PV (Product review volume) 46.613 166.842 0.000 2,700.000

PR (Product review rating) 3.880 1.815 0.000 5.000

PS (Product past monthly sales quantity) 14.478 52.794 0.000 1,169.500

IN (Product inventory) 110.757 433.525 1.000 9,978.000

BM (Number of product webpage bookmarks) 67.528 248.560 0.000 4,050.000

Notes: Number of observations = 53,061; Number of products = 722; Number of days = 111.

A8 MIS Quarterly Vol. 41 No. 2—Appendices/June 2017

Lin et al./Demand Effects of Product Recommendation Networks

Estimation Results Based on Subcategories of the Main Sample of Digital Cameras

Table A10 presents the model estimation results based on all the six subcategories of our main sample of digital cameras. The results show that the product network’s impact is more influential for the SLR camera subcategory, compared to all other subcategories. Noteworthy, we include all the six subcategories, rather than only the SLR camera subcategory, in our empirical analysis because the product network may connect across different categories. Thus, it is more complete and accurate to analyze the entire network (and thus all the subcategories).

Table A10. Model Estimation Results (Subcategories)

Variable

(1)

Preferred

(All)

(2)

Subcategory

(Accessory)

(3)

Subcategory

(Battery)

(4)

Subcategory

(Compact

camera)

(5)

Subcategory

(Flash)

(6)

Subcategory

(Lens)

(7)

Subcategory

(SLR camera)

ID_CV -0.001 -0.002 0.000 -0.003* 0.002 -0.001*** 0.025***

(Incoming, diversity, co-view) (0.003) (0.003) (0.007) (0.002) (0.004) (0.000) (0.006)

ID_CP 0.010*** -0.002 -0.001 -0.000 0.003 -0.000 0.019***

(Incoming, diversity, co-

purchase) (0.001) (0.002) (0.004) (0.001) (0.005) (0.000) (0.003)

OD_CV -0.003 0.005 0.001 0.006 0.001 0.000 -0.000

(Outgoing, diversity, co-view) (0.004) (0.004) (0.008) (0.003) (0.005) (0.001) (0.013)

OD_CP -0.008*** 0.001 0.002 -0.001 0.003 0.000 -0.008***

(Outgoing, diversity, co-

purchase) (0.002) (0.002) (0.005) (0.001) (0.003) (0.000) (0.003)

IS_CV 0.005 0.002 0.002 -0.001 0.004 0.000 -0.003

(Incoming, stability, co-view) (0.003) (0.004) (0.009) (0.002) (0.006) (0.001) (0.009)

IS_CP -0.000 -0.004 0.009 -0.001 -0.002 0.000 0.006

(Incoming, stability, co-

purchase) (0.003) (0.004) (0.008) (0.002) (0.003) (0.000) (0.007)

OS_CV 0.006 0.006 0.012 0.000 0.003 -0.000 0.011

(Outgoing, stability, co-view) (0.004) (0.004) (0.011) (0.003) (0.006) (0.001) (0.011)

OS_CP -0.017*** -0.006 -0.006 0.001 -0.000 0.001 -0.021***

(Outgoing, stability, co-

purchase) (0.003) (0.004) (0.008) (0.002) (0.005) (0.001) (0.006)

Constant 0.323*** 0.000 0.000 0.000 0.000 0.000 0.320***

(0.046) (0.000) (0.000) (0.000) (0.000) (0.000) (0.063)

Control variables -included- -included- -included- -included- -included- -included- -included-

Number of observations 41,379 19,260 2,147 3,069 907 11,095 4,901

R² 0.4934 0.7472 0.9218 0.9528 0.0171 0.9656 0.8586

Notes: Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01.

MIS Quarterly Vol. 41 No. 2—Appendices/June 2017 A9

Lin et al./Demand Effects of Product Recommendation Networks

Interaction Effect

Table A11 presents the results of the preferred model with the inclusion of interaction terms between network stability and product vintage, and between network stability and product past monthly sales quantity. As seen in Column (2), there exist significant interaction effects between network stability and product vintage (IS_CV * VT), and between network stability and past monthly sales quantity (IS_CV * PS, IS_CP * PS, OS_CP * PS). However, the focal estimates for network diversity and stability are similar to those in Column (1).

Table A11. Interaction Effect

Variable (1)

Preferred (2)

Interaction

ID_CV -0.001 0.001

(Incoming, diversity, co-view) (0.003) (0.003)

ID_CP 0.010*** 0.006***

(Incoming, diversity, co-purchase) (0.001) (0.001)

OD_CV -0.003 -0.002

(Outgoing, diversity, co-view) (0.004) (0.004)

OD_CP -0.008*** -0.005***

(Outgoing, diversity, co-purchase) (0.002) (0.002)

IS_CV 0.005 0.007

(Incoming, stability, co-view) (0.003) (0.005)

IS_CP -0.000 -0.001

(Incoming, stability, co-purchase) (0.003) (0.004)

OS_CV 0.006 0.006

(Outgoing, stability, co-view) (0.004) (0.006)

OS_CP -0.017*** -0.011**

(Outgoing, stability, co-purchase) (0.003) (0.005)

IS_CV * VT -0.000***

(0.000)

IS_CP * VT 0.000

(0.000)

OS_CV * VT -0.000

(0.000)

OS_CP * VT 0.000

(0.000)

IS_CV * PS 0.013***

(0.001)

IS_CP * PS -0.006***

(0.000)

OS_CV * PS -0.000

(0.000)

OS_CP * PS -0.004***

(0.000)

Constant 0.323*** 0.319***

(0.046) (0.046)

Control variables -included- -included-

Number of observations 41,379 41,379

R² 0.4934 0.5053

Notes: Standard errors in parentheses. *p < 0.1, **p < 0.05, ***p < 0.01. VT: product vintage. PS: product past monthly sales quantity.

A10 MIS Quarterly Vol. 41 No. 2—Appendices/June 2017

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.